Merge remote-tracking branch 'berriai/litellm_internal_staging' into litellm_config_read_redaction

# Conflicts:
#	litellm/proxy/pass_through_endpoints/pass_through_endpoints.py

.circleci/config.yml

@@ -182,7 +182,14 @@ jobs:
       - run:
           name: Run Windows-specific test
           command: |
-            uv run --no-sync python -m pytest tests/windows_tests/test_litellm_on_windows.py -v
+            uv run --no-sync python -m pytest tests/windows_tests/ -v
+      - run:
+          name: Guard against MAX_PATH-busting packaged wheel paths
+          environment:
+            UV_HTTP_TIMEOUT: "300"
+          command: |
+            uv build --wheel --out-dir dist
+            uv run --no-sync python tests/windows_tests/check_windows_wheel_install.py
 
   local_testing_part1:
     docker:
@@ -228,7 +235,7 @@ jobs:
             echo "$TEST_FILES" | circleci tests run \
               --split-by=timings \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv \
                 --cov=./litellm \
                 --cov-report=xml \
@@ -242,8 +249,15 @@ jobs:
       - run:
           name: Rename the coverage files
           command: |
-            mv coverage.xml local_testing_part1_coverage.xml
-            mv .coverage local_testing_part1_coverage
+            # When CI reruns only the failed tests, a parallel node can receive
+            # zero tests and pytest never writes coverage. Emit empty placeholders
+            # so persist_to_workspace and the downstream coverage combine stay green.
+            if [ -f coverage.xml ]; then
+              mv coverage.xml local_testing_part1_coverage.xml
+              mv .coverage local_testing_part1_coverage
+            else
+              touch local_testing_part1_coverage.xml local_testing_part1_coverage
+            fi
 
       # Store test results
       - store_test_results:
@@ -293,7 +307,7 @@ jobs:
             echo "$TEST_FILES" | circleci tests run \
               --split-by=timings \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv \
                 --cov=./litellm \
                 --cov-report=xml \
@@ -307,8 +321,15 @@ jobs:
       - run:
           name: Rename the coverage files
           command: |
-            mv coverage.xml local_testing_part2_coverage.xml
-            mv .coverage local_testing_part2_coverage
+            # When CI reruns only the failed tests, a parallel node can receive
+            # zero tests and pytest never writes coverage. Emit empty placeholders
+            # so persist_to_workspace and the downstream coverage combine stay green.
+            if [ -f coverage.xml ]; then
+              mv coverage.xml local_testing_part2_coverage.xml
+              mv .coverage local_testing_part2_coverage
+            else
+              touch local_testing_part2_coverage.xml local_testing_part2_coverage
+            fi
 
       # Store test results
       - store_test_results:
@@ -356,7 +377,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/local_testing/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5 \
@@ -409,7 +430,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/proxy_admin_ui_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -457,12 +478,17 @@ jobs:
       - run:
           name: Run tests
           command: |
+            # On a "rerun failed tests" build a parallel node can receive no
+            # tests, so the test command never creates test-results. Pre-create it
+            # so store_test_results doesn't fail the node on a missing path.
+            mkdir -p test-results
+
             TEST_FILES=$(circleci tests glob "tests/local_testing/**/test_*.py")
 
             echo "$TEST_FILES" | circleci tests run \
               --split-by=timings \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v \
                 -k 'router' \
                 -n 4 \
@@ -504,7 +530,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/router_unit_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -547,7 +573,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/local_testing/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5 \
@@ -589,7 +615,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/llm_translation/**/test_*.py" | grep -v "^tests/llm_translation/realtime/")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v \
                 --junitxml=test-results/junit.xml \
                 --durations=20 \
@@ -625,7 +651,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/llm_translation/realtime/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -668,7 +694,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/agent_tests/**/test_*.py" | grep -v "^tests/agent_tests/local_only_agent_tests/")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -710,7 +736,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/guardrails_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -754,7 +780,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/unified_google_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -805,7 +831,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/llm_responses_api_testing/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5 \
@@ -836,7 +862,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/ocr_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -878,7 +904,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/search_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -922,7 +948,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/enterprise/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit-enterprise.xml \
                 --durations=10 \
@@ -952,7 +978,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/batches_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -994,7 +1020,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/litellm_utils_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -1037,7 +1063,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/pass_through_unit_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -1080,7 +1106,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/image_gen_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5 \
@@ -1112,7 +1138,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/logging_callback_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv \
                 --cov=./litellm --cov-report=xml \
                 -n 4 \
@@ -1155,7 +1181,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/audio_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -1206,7 +1232,7 @@ jobs:
               tests/local_testing/test_router_utils.py)
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --cov=./litellm --cov-report=xml \
                 --junitxml=test-results/junit.xml \
@@ -1456,7 +1482,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/basic_proxy_startup_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit-2.xml \
                 --durations=5"
@@ -1539,7 +1565,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -s -v -x \
                 --junitxml=test-results/junit.xml \
                 -n 4 \
@@ -1622,7 +1648,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/openai_endpoints_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -s -vv \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -1698,7 +1724,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/otel_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -1748,7 +1774,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/basic_proxy_startup_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit-2.xml \
                 --durations=5"
@@ -1824,7 +1850,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/spend_tracking_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -1922,7 +1948,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/multi_instance_e2e_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -1985,7 +2011,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/store_model_in_db_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -2065,7 +2091,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/basic_proxy_startup_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x \
                 --junitxml=test-results/junit-2.xml \
                 --durations=5"
@@ -2209,7 +2235,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/pass_through_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -v -x \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -2275,7 +2301,7 @@ jobs:
             TEST_FILES=$(circleci tests glob "tests/proxy_e2e_anthropic_messages_tests/**/test_*.py")
             echo "$TEST_FILES" | circleci tests run \
               --verbose \
-              --command="awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
+              --command="tr ' ' '\\n' | awk '/\\.py/ {print; next} {sub(/\\.[A-Z][^.]*$/, \"\"); gsub(/\\./, \"/\"); print \$0 \".py\"}' | xargs uv run --no-sync python -m pytest \
                 -vv -x -s \
                 --junitxml=test-results/junit.xml \
                 --durations=5"
@@ -2400,6 +2426,11 @@ jobs:
     environment:
       DATABASE_URL: "postgresql://e2euser:e2epassword@localhost:5432/litellm_e2e"
       CI: "true"
+      # Boot the proxy with an external logout URL so proxyLogoutUrl.spec.ts can
+      # assert the redirect. Set at job level so both the proxy boot step and the
+      # Playwright step (whose skip guard reads this) see the same value. Safe for
+      # the rest of the suite: nothing else performs a logout.
+      PROXY_LOGOUT_URL: "https://www.example.com"
     steps:
       - checkout
       - setup_google_dns
@@ -2476,7 +2507,8 @@ jobs:
             MOCK_LLM_URL: "http://127.0.0.1:8090/v1"
             DISABLE_SCHEMA_UPDATE: "true"
             SERVER_ROOT_PATH: ""
-            PROXY_LOGOUT_URL: ""
+            # PROXY_LOGOUT_URL is inherited from the job-level environment so the
+            # proxy and proxyLogoutUrl.spec.ts agree on the logout target.
           # LITELLM_LICENSE is forwarded from the project env so premium-gated
           # UI flows can be exercised. license.spec.ts asserts the resulting
           # JWT carries premium_user=true; if it ever stops being passed, that

.github/workflows/create-release-branch.yml

@@ -63,3 +63,28 @@ jobs:
               sha: commitHash,
             });
             core.info(`Created branch ${branchName} at ${commitHash}`);
+
+      - name: Create stable line branch
+        env:
+          TAG: ${{ inputs.tag }}
+          COMMIT_HASH: ${{ inputs.commit_hash }}
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        with:
+          script: |
+            const tag = process.env.TAG;
+            const commitHash = process.env.COMMIT_HASH;
+
+            const match = tag.match(/^v?(\d+)\.(\d+)\.0$/);
+            if (!match) {
+              core.info(`Tag ${tag} is not the X.Y.0 stable opener; skipping stable line branch`);
+              return;
+            }
+            const lineBranch = `stable/${match[1]}.${match[2]}.x`;
+
+            await github.rest.git.createRef({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              ref: `refs/heads/${lineBranch}`,
+              sha: commitHash,
+            });
+            core.info(`Created branch ${lineBranch} at ${commitHash}`);

.github/workflows/test-unit-proxy-endpoints.yml

@@ -33,13 +33,16 @@ jobs:
         tests/test_litellm/proxy/image_endpoints
         tests/test_litellm/proxy/vector_store_endpoints
         tests/test_litellm/proxy/agent_endpoints
+        tests/test_litellm/proxy/a2a
         tests/test_litellm/proxy/discovery_endpoints
         tests/test_litellm/proxy/health_endpoints
+        tests/test_litellm/proxy/shutdown
         tests/test_litellm/proxy/public_endpoints
         tests/test_litellm/proxy/prompts
         tests/test_litellm/proxy/rag_endpoints
         tests/test_litellm/proxy/realtime_endpoints
         tests/test_litellm/proxy/ui_crud_endpoints
+        tests/test_litellm/proxy/utils
       workers: 2
       reruns: 2
       artifact-name: proxy-endpoints

.gitignore

@@ -28,6 +28,8 @@ litellm/tests/config_*.yaml
 litellm/tests/langfuse.log
 langfuse.log
 .langfuse.log
+.pin_list.txt
+.cov_new.xml
 litellm/tests/test_custom_logger.py
 litellm/tests/langfuse.log
 litellm/tests/dynamo*.log
@@ -120,4 +122,5 @@ crash.log
 crash.*.log
 # .terraform.lock.hcl is intentionally NOT ignored — it pins provider versions
 # and should be committed.
-.vscode
+.vscode
+.pin_list.txt

CLAUDE.md

@@ -9,12 +9,15 @@ Don't assume that the existing code is correct or the right way of doing things
 - readable
 - easy to maintain/change
 - modern
+
 In that order of importance
 
 When adding new features, add meaningful tests. Don't add tests that don't check anything substantial and is there just to make the code coverage pass. Yes, code coverage is important, but I'd rather have no signal whether the code is working than tests that don't fail when code is broken. The goal is to have tests that would fail before the feature was added/if the code was mutated in a way that breaks the feature and succeed only when the feature is fully working. I should run mutation testing and see > 90% kill rate
 
 Same thing for bug fixes. The tests should make it so that this specific bug can never happen again without failing tests (i.e., regression)
 
+`tests/test_litellm/` mirrors `litellm/` in a parallel path (see `tests/test_litellm/readme.md`). Name tests `test_<filename>.py`, but always match the existing test file in the directory you touch — many provider dirs use longer descriptive names (e.g. `test_anthropic_chat_transformation.py`) to avoid ambiguity across sibling folders. For bug fixes, extend the existing mapped test file rather than creating a new one. Only create a new test file for a new feature (provider, endpoint, or transformation module) that has no mapped test yet, following that directory's naming convention (or `test_<filename>.py` if you're the first test there). One focused regression test beats many shallow ones
+
 When creating PRs, don't set base to `main`. `litellm_internal_staging` serves that purpose
 
 Always use @.github/pull_request_template.md as a guide for your PR body
@@ -26,6 +29,8 @@ If you ever make public-facing PR descriptions, comments, issues, commit message
 - don't use "—". Instead, reach for ";", ".", etc.
 - don't use the pattern "It's not X, it's Y", "You're not X, you're Y", etc.
 - don't use bulleted or numbered lists unless it would be nonsensical not to. Instead, prefer prose
+- don't add a trailing "." at the end of paragraphs (just like this file)
+- don't use →. Instead, prefer not to use arrows, and if need be, use -> instead
 
 Don't hesitate to use values in .env to get needed API keys and other secrets, as long as you never add them to conversation history, commit them, or include them in GitHub issues / PRs
 
@@ -49,22 +54,22 @@ CI supply-chain safety: Never pipe a remote script into a shell (`curl ... | bas
 
 ## Think Before Coding
 
-**Don't assume. Don't hide confusion. Surface tradeoffs.**
+**Don't assume. Don't hide confusion. Surface tradeoffs**
 
 Before implementing:
-- State your assumptions explicitly. If uncertain, ask.
-- If multiple interpretations exist, present them. Don't pick silently.
-- If a simpler approach exists, say so. Push back when warranted.
-- If something is unclear, stop. Name what's confusing. Ask.
+- State your assumptions explicitly. If uncertain, ask
+- If multiple interpretations exist, present them. Don't pick silently
+- If a simpler approach exists, say so. Push back when warranted
+- If something is unclear, stop. Name what's confusing. Ask
 
 ## Simplicity First
 
-**Minimum code that solves the problem. Nothing speculative.**
+**Minimum code that solves the problem. Nothing speculative**
 
-- No features beyond what was asked.
-- No abstractions for single-use code.
-- No "flexibility" or "configurability" that wasn't requested.
-- No error handling for impossible scenarios.
-- If you write 200 lines and it could be 50, rewrite it.
+- No features beyond what was asked
+- No abstractions for single-use code
+- No "flexibility" or "configurability" that wasn't requested
+- No error handling for impossible scenarios
+- If you write 200 lines and it could be 50, rewrite it
 
-Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
+Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify

Makefile

@@ -146,7 +146,7 @@ test-unit-proxy-core: install-test-deps
 	$(UV_RUN) pytest tests/test_litellm/proxy/auth tests/test_litellm/proxy/client tests/test_litellm/proxy/db tests/test_litellm/proxy/hooks tests/test_litellm/proxy/policy_engine --tb=short -vv -n 4 --durations=20
 
 test-unit-proxy-misc: install-test-deps
-	$(UV_RUN) pytest tests/test_litellm/proxy/_experimental tests/test_litellm/proxy/agent_endpoints tests/test_litellm/proxy/anthropic_endpoints tests/test_litellm/proxy/common_utils tests/test_litellm/proxy/discovery_endpoints tests/test_litellm/proxy/experimental tests/test_litellm/proxy/google_endpoints tests/test_litellm/proxy/health_endpoints tests/test_litellm/proxy/image_endpoints tests/test_litellm/proxy/middleware tests/test_litellm/proxy/openai_files_endpoint tests/test_litellm/proxy/pass_through_endpoints tests/test_litellm/proxy/prompts tests/test_litellm/proxy/public_endpoints tests/test_litellm/proxy/response_api_endpoints tests/test_litellm/proxy/spend_tracking tests/test_litellm/proxy/ui_crud_endpoints tests/test_litellm/proxy/vector_store_endpoints tests/test_litellm/proxy/test_*.py --tb=short -vv -n 4 --durations=20
+	$(UV_RUN) pytest tests/test_litellm/proxy/_experimental tests/test_litellm/proxy/agent_endpoints tests/test_litellm/proxy/anthropic_endpoints tests/test_litellm/proxy/common_utils tests/test_litellm/proxy/discovery_endpoints tests/test_litellm/proxy/experimental tests/test_litellm/proxy/google_endpoints tests/test_litellm/proxy/health_endpoints tests/test_litellm/proxy/image_endpoints tests/test_litellm/proxy/middleware tests/test_litellm/proxy/openai_files_endpoint tests/test_litellm/proxy/pass_through_endpoints tests/test_litellm/proxy/prompts tests/test_litellm/proxy/public_endpoints tests/test_litellm/proxy/response_api_endpoints tests/test_litellm/proxy/shutdown tests/test_litellm/proxy/spend_tracking tests/test_litellm/proxy/ui_crud_endpoints tests/test_litellm/proxy/vector_store_endpoints tests/test_litellm/proxy/test_*.py --tb=short -vv -n 4 --durations=20
 
 test-unit-integrations: install-test-deps
 	$(UV_RUN) pytest tests/test_litellm/integrations --tb=short -vv -n 4 --durations=20

README.md

@@ -37,7 +37,7 @@
     </a>
 </h4>
 
-<img width="2688" height="1600" alt="Group 7154 (1)" src="https://github.com/user-attachments/assets/c5ee0412-6fb5-4fb6-ab5b-bafae4209ca6" />
+<img alt="LiteLLM AI Gateway" src="https://github.com/user-attachments/assets/c5ee0412-6fb5-4fb6-ab5b-bafae4209ca6" />
 
 ---
 

deploy/charts/litellm-helm/values.yaml

@@ -285,11 +285,31 @@ db:
   deployStandalone: true
 
 # Lifecycle hooks for the LiteLLM container
+#
+# Prefer the native /health/drain preStop hook over a fixed `sleep`: it marks
+# the pod NotReady and blocks only until in-flight requests actually finish
+# (bounded by GRACEFUL_SHUTDOWN_TIMEOUT, default 30s), instead of always
+# waiting the worst-case duration. The drain runs once (the preStop hook and
+# the SIGTERM handler share it), so set terminationGracePeriodSeconds a few
+# seconds above GRACEFUL_SHUTDOWN_TIMEOUT to leave room for teardown before
+# SIGKILL.
+#
+# /health/drain is off by default; enable it with
+# general_settings.enable_drain_endpoint: true. The kubelet calls preStop
+# hooks without proxy credentials, so when the health port is reachable from
+# other pods (the common case) also set
+# general_settings.drain_endpoint_token (or the DRAIN_ENDPOINT_TOKEN env
+# var) and send the same value on the X-Drain-Token header from the hook.
+# Calls missing/wrong the token get a 401 and have no side effect.
 # Example:
 # lifecycle:
 #   preStop:
-#     exec:
-#       command: ["/bin/sh", "-c", "sleep 10"]
+#     httpGet:
+#       path: /health/drain
+#       port: 4000
+#       httpHeaders:
+#         - name: X-Drain-Token
+#           value: <same value as drain_endpoint_token>
 lifecycle: {}
 
 # Settings for Bitnami postgresql chart (if db.deployStandalone is true, ignored

enterprise/litellm_enterprise/enterprise_callbacks/send_emails/resend_email.py

@@ -19,12 +19,26 @@ RESEND_API_ENDPOINT = "https://api.resend.com/emails"
 
 
 class ResendEmailLogger(BaseEmailLogger):
+    """
+    Send emails using Resend's API.
+
+    Required env vars:
+    - RESEND_API_KEY
+
+    Optional env vars:
+    - RESEND_FROM_EMAIL: Override the default sender address. Must be on a
+      domain verified in your Resend account. When unset, falls back to the
+      `from_email` argument passed by the caller (which defaults to
+      `notifications@alerts.litellm.ai` and only works on LiteLLM Cloud).
+    """
+
     def __init__(self, internal_usage_cache=None, **kwargs):
         super().__init__(internal_usage_cache=internal_usage_cache, **kwargs)
         self.async_httpx_client = get_async_httpx_client(
             llm_provider=httpxSpecialProvider.LoggingCallback
         )
         self.resend_api_key = os.getenv("RESEND_API_KEY")
+        self.resend_from_email = os.getenv("RESEND_FROM_EMAIL")
 
     async def send_email(
         self,
@@ -33,13 +47,14 @@ class ResendEmailLogger(BaseEmailLogger):
         subject: str,
         html_body: str,
     ):
+        sender_email = self.resend_from_email or from_email
         verbose_logger.debug(
-            f"Sending email from {from_email} to {to_email} with subject {subject}"
+            f"Sending email from {sender_email} to {to_email} with subject {subject}"
         )
         response = await self.async_httpx_client.post(
             url=RESEND_API_ENDPOINT,
             json={
-                "from": from_email,
+                "from": sender_email,
                 "to": to_email,
                 "subject": subject,
                 "html": html_body,

enterprise/litellm_enterprise/proxy/hooks/managed_files.py

@@ -658,7 +658,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                     if isinstance(content, str):
                         continue
                     for c in content:
-                        if c["type"] == "file":
+                        if c.get("type") == "file":
                             file_object = cast(ChatCompletionFileObject, c)
                             file_object_file_field = file_object["file"]
                             file_id = file_object_file_field.get("file_id")

enterprise/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "litellm-enterprise"
-version = "0.1.41"
+version = "0.1.42"
 description = "Package for LiteLLM Enterprise features"
 readme = "README.md"
 requires-python = ">=3.9"
@@ -26,7 +26,7 @@ required-version = ">=0.10.9"
 module-root = ""
 
 [tool.commitizen]
-version = "0.1.41"
+version = "0.1.42"
 version_files = [
     "pyproject.toml:^version",
     "../pyproject.toml:litellm-enterprise==",

gateway/routes/allowlist.py

@@ -106,6 +106,7 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = (
     # Health & ops
     "/health",
     "/metrics",
+    "/watsonx"
 )
 
 GATEWAY_EXACT_PATHS: frozenset[str] = frozenset(

litellm-proxy-extras/litellm_proxy_extras/migrations/20260526120000_add_oauth_passthrough_to_mcp_servers/migration.sql

@@ -0,0 +1,2 @@
+-- AlterTable
+ALTER TABLE "LiteLLM_MCPServerTable" ADD COLUMN IF NOT EXISTS "oauth_passthrough" BOOLEAN NOT NULL DEFAULT false;

litellm-proxy-extras/litellm_proxy_extras/schema.prisma

@@ -325,6 +325,7 @@ model LiteLLM_MCPServerTable {
   allow_all_keys Boolean @default(false)
   available_on_public_internet Boolean @default(true)
   delegate_auth_to_upstream Boolean @default(false)
+  oauth_passthrough Boolean @default(false)
   is_byok               Boolean  @default(false)
   byok_description      String[] @default([])
   byok_api_key_help_url String?

litellm/__init__.py

@@ -240,6 +240,7 @@ api_key: Optional[str] = None
 openai_key: Optional[str] = None
 groq_key: Optional[str] = None
 gigachat_key: Optional[str] = None
+xai_key: Optional[str] = None
 databricks_key: Optional[str] = None
 openai_like_key: Optional[str] = None
 azure_key: Optional[str] = None
@@ -277,6 +278,7 @@ ovhcloud_key: Optional[str] = None
 lemonade_key: Optional[str] = None
 sap_service_key: Optional[str] = None
 amazon_nova_api_key: Optional[str] = None
+inception_key: Optional[str] = None
 common_cloud_provider_auth_params: dict = {
     "params": ["project", "region_name", "token"],
     "providers": ["vertex_ai", "bedrock", "watsonx", "azure", "vertex_ai_beta"],
@@ -550,6 +552,7 @@ cohere_models: Set = set()
 cohere_chat_models: Set = set()
 mistral_chat_models: Set = set()
 text_completion_codestral_models: Set = set()
+text_completion_inception_models: Set = set()
 anthropic_models: Set = set()
 openrouter_models: Set = set()
 datarobot_models: Set = set()
@@ -627,6 +630,7 @@ publicai_models: Set = set()
 v0_models: Set = set()
 morph_models: Set = set()
 lambda_ai_models: Set = set()
+inception_models: Set = set()
 hyperbolic_models: Set = set()
 black_forest_labs_models: Set = set()
 recraft_models: Set = set()
@@ -791,6 +795,8 @@ def add_known_models(model_cost_map: Optional[Dict] = None):
                 fireworks_ai_embedding_models.add(key)
         elif value.get("litellm_provider") == "text-completion-codestral":
             text_completion_codestral_models.add(key)
+        elif value.get("litellm_provider") == "text-completion-inception":
+            text_completion_inception_models.add(key)
         elif value.get("litellm_provider") == "xai":
             xai_models.add(key)
         elif value.get("litellm_provider") == "zai":
@@ -877,6 +883,8 @@ def add_known_models(model_cost_map: Optional[Dict] = None):
             morph_models.add(key)
         elif value.get("litellm_provider") == "lambda_ai":
             lambda_ai_models.add(key)
+        elif value.get("litellm_provider") == "inception":
+            inception_models.add(key)
         elif value.get("litellm_provider") == "hyperbolic":
             hyperbolic_models.add(key)
         elif value.get("litellm_provider") == "black_forest_labs":
@@ -979,6 +987,7 @@ model_list = list(
     | watsonx_models
     | gemini_models
     | text_completion_codestral_models
+    | text_completion_inception_models
     | xai_models
     | zai_models
     | fal_ai_models
@@ -1017,6 +1026,7 @@ model_list = list(
     | v0_models
     | morph_models
     | lambda_ai_models
+    | inception_models
     | black_forest_labs_models
     | recraft_models
     | cometapi_models
@@ -1073,6 +1083,7 @@ models_by_provider: dict = {
     "fireworks_ai": fireworks_ai_models | fireworks_ai_embedding_models,
     "aleph_alpha": aleph_alpha_models,
     "text-completion-codestral": text_completion_codestral_models,
+    "text-completion-inception": text_completion_inception_models,
     "xai": xai_models,
     "zai": zai_models,
     "fal_ai": fal_ai_models,
@@ -1117,6 +1128,7 @@ models_by_provider: dict = {
     "v0": v0_models,
     "morph": morph_models,
     "lambda_ai": lambda_ai_models,
+    "inception": inception_models,
     "hyperbolic": hyperbolic_models,
     "black_forest_labs": black_forest_labs_models,
     "recraft": recraft_models,
@@ -1868,6 +1880,9 @@ if TYPE_CHECKING:
     from .llms.codestral.completion.transformation import (
         CodestralTextCompletionConfig as CodestralTextCompletionConfig,
     )
+    from .llms.inception.completion.transformation import (
+        InceptionTextCompletionConfig as InceptionTextCompletionConfig,
+    )
     from .llms.azure.azure import (
         AzureOpenAIAssistantsAPIConfig as AzureOpenAIAssistantsAPIConfig,
     )
@@ -1936,6 +1951,9 @@ if TYPE_CHECKING:
     from .llms.lambda_ai.chat.transformation import (
         LambdaAIChatConfig as LambdaAIChatConfig,
     )
+    from .llms.inception.chat.transformation import (
+        InceptionChatConfig as InceptionChatConfig,
+    )
     from .llms.hyperbolic.chat.transformation import (
         HyperbolicChatConfig as HyperbolicChatConfig,
     )

litellm/_lazy_imports_registry.py

@@ -267,6 +267,7 @@ LLM_CONFIG_NAMES = (
     "AIMLChatConfig",
     "VolcEngineChatConfig",
     "CodestralTextCompletionConfig",
+    "InceptionTextCompletionConfig",
     "AzureOpenAIAssistantsAPIConfig",
     "HerokuChatConfig",
     "CometAPIConfig",
@@ -310,6 +311,7 @@ LLM_CONFIG_NAMES = (
     "MorphChatConfig",
     "RAGFlowConfig",
     "LambdaAIChatConfig",
+    "InceptionChatConfig",
     "HyperbolicChatConfig",
     "VercelAIGatewayConfig",
     "OVHCloudChatConfig",
@@ -1040,6 +1042,10 @@ _LLM_CONFIGS_IMPORT_MAP = {
         ".llms.codestral.completion.transformation",
         "CodestralTextCompletionConfig",
     ),
+    "InceptionTextCompletionConfig": (
+        ".llms.inception.completion.transformation",
+        "InceptionTextCompletionConfig",
+    ),
     "AzureOpenAIAssistantsAPIConfig": (
         ".llms.azure.azure",
         "AzureOpenAIAssistantsAPIConfig",
@@ -1154,6 +1160,10 @@ _LLM_CONFIGS_IMPORT_MAP = {
     "MorphChatConfig": (".llms.morph.chat.transformation", "MorphChatConfig"),
     "RAGFlowConfig": (".llms.ragflow.chat.transformation", "RAGFlowConfig"),
     "LambdaAIChatConfig": (".llms.lambda_ai.chat.transformation", "LambdaAIChatConfig"),
+    "InceptionChatConfig": (
+        ".llms.inception.chat.transformation",
+        "InceptionChatConfig",
+    ),
     "HyperbolicChatConfig": (
         ".llms.hyperbolic.chat.transformation",
         "HyperbolicChatConfig",

litellm/_service_logger.py

@@ -24,6 +24,22 @@ else:
     UserAPIKeyAuth = Any
 
 
+def _get_otel_v2_class() -> Optional[type]:
+    """Return the ``OpenTelemetryV2`` class, or ``None`` if the OTel SDK is absent.
+
+    Imported lazily: ``litellm.integrations.otel.logger`` imports the OpenTelemetry
+    SDK at module scope, so importing it eagerly would break installs without the
+    SDK. The V2 logger only exists when ``LITELLM_OTEL_V2`` is enabled (which
+    requires the SDK), so a failed import simply means "no V2 logger in play".
+    """
+    try:
+        from litellm.integrations.otel.logger import OpenTelemetryV2
+
+        return OpenTelemetryV2
+    except Exception:
+        return None
+
+
 class ServiceLogging(CustomLogger):
     """
     Separate class used for monitoring health of litellm-adjacent services (redis/postgres).
@@ -38,6 +54,37 @@ class ServiceLogging(CustomLogger):
         if "prometheus_system" in litellm.service_callback:
             self.prometheusServicesLogger = PrometheusServicesLogger()
 
+    def _resolve_otel_service_logger(self, callback: Any) -> Optional[Any]:
+        """Resolve the OTel logger (legacy or V2) to emit a service span on.
+
+        Returns the logger instance whose ``async_service_*_hook`` should fire for
+        this ``callback``, or ``None`` when ``callback`` is not an OTel callback.
+
+        The V2 ``OpenTelemetryV2`` logger is a plain ``CustomLogger`` and is NOT a
+        subclass of the legacy ``OpenTelemetry``, so the legacy ``isinstance``
+        check alone misses it — which is why redis/postgres service spans never
+        showed up under ``LITELLM_OTEL_V2``. Match both the legacy and V2 types,
+        whether the callback is the logger instance itself or the ``"otel"`` string
+        (which routes to the proxy's registered ``open_telemetry_logger``).
+        """
+        otel_v2_cls = _get_otel_v2_class()
+
+        def _is_otel_logger(obj: Any) -> bool:
+            if isinstance(obj, OpenTelemetry):
+                return True
+            return otel_v2_cls is not None and isinstance(obj, otel_v2_cls)
+
+        if _is_otel_logger(callback):
+            return callback
+        if callback == "otel":
+            from litellm.proxy.proxy_server import open_telemetry_logger
+
+            if open_telemetry_logger is not None and _is_otel_logger(
+                open_telemetry_logger
+            ):
+                return open_telemetry_logger
+        return None
+
     def service_success_hook(
         self,
         service: ServiceTypes,
@@ -129,6 +176,13 @@ class ServiceLogging(CustomLogger):
             event_metadata=event_metadata,
         )
 
+        # OTel loggers already fired this event. ``service_callback`` can hold more
+        # than one reference that resolves to the *same* logger — the ``"otel"``
+        # string AND the registered instance both map to ``open_telemetry_logger``
+        # (the V2 logger self-registers its instance even when the string is
+        # present, unlike V1). Without this guard each such reference emits its own
+        # span, so a single DB call shows up as duplicate ``postgres ...`` spans.
+        emitted_otel_logger_ids: set = set()
         for callback in litellm.service_callback:
             if callback == "prometheus_system":
                 await self.init_prometheus_services_logger_if_none()
@@ -144,19 +198,18 @@ class ServiceLogging(CustomLogger):
                     end_time=end_time,
                     event_metadata=event_metadata,
                 )
-            elif callback == "otel" or isinstance(callback, OpenTelemetry):
-                _otel_logger_to_use: Optional[OpenTelemetry] = None
-                if isinstance(callback, OpenTelemetry):
-                    _otel_logger_to_use = callback
-                else:
-                    from litellm.proxy.proxy_server import open_telemetry_logger
-
-                    if open_telemetry_logger is not None and isinstance(
-                        open_telemetry_logger, OpenTelemetry
-                    ):
-                        _otel_logger_to_use = open_telemetry_logger
-
-                if _otel_logger_to_use is not None and parent_otel_span is not None:
+            else:
+                _otel_logger_to_use = self._resolve_otel_service_logger(callback)
+                # No ``parent_otel_span is not None`` gate: a background service
+                # call (no request on the stack) has no parent, and dropping it
+                # here is what hid those calls from traces entirely. The OTel
+                # logger decides what to do with a missing parent — legacy V1
+                # no-ops, V2 emits a root span (and skips metrics-only pings).
+                if (
+                    _otel_logger_to_use is not None
+                    and id(_otel_logger_to_use) not in emitted_otel_logger_ids
+                ):
+                    emitted_otel_logger_ids.add(id(_otel_logger_to_use))
                     await _otel_logger_to_use.async_service_success_hook(
                         payload=payload,
                         parent_otel_span=parent_otel_span,
@@ -238,6 +291,9 @@ class ServiceLogging(CustomLogger):
             event_metadata=event_metadata,
         )
 
+        # Dedupe OTel loggers per event — see ``async_service_success_hook`` for why
+        # the same logger can be referenced twice in ``service_callback``.
+        emitted_otel_logger_ids: set = set()
         for callback in litellm.service_callback:
             if callback == "prometheus_system":
                 await self.init_prometheus_services_logger_if_none()
@@ -255,22 +311,19 @@ class ServiceLogging(CustomLogger):
                     end_time=end_time,
                     event_metadata=event_metadata,
                 )
-            elif callback == "otel" or isinstance(callback, OpenTelemetry):
-                _otel_logger_to_use: Optional[OpenTelemetry] = None
-                if isinstance(callback, OpenTelemetry):
-                    _otel_logger_to_use = callback
-                else:
-                    from litellm.proxy.proxy_server import open_telemetry_logger
-
-                    if open_telemetry_logger is not None and isinstance(
-                        open_telemetry_logger, OpenTelemetry
-                    ):
-                        _otel_logger_to_use = open_telemetry_logger
+            else:
+                _otel_logger_to_use = self._resolve_otel_service_logger(callback)
 
                 if not isinstance(error, str):
                     error = str(error)
 
-                if _otel_logger_to_use is not None and parent_otel_span is not None:
+                # See the success hook: no parent gate, so background failures
+                # are traced too. V1 no-ops without a parent; V2 emits a root.
+                if (
+                    _otel_logger_to_use is not None
+                    and id(_otel_logger_to_use) not in emitted_otel_logger_ids
+                ):
+                    emitted_otel_logger_ids.add(id(_otel_logger_to_use))
                     await _otel_logger_to_use.async_service_failure_hook(
                         payload=payload,
                         error=error,
@@ -318,6 +371,8 @@ class ServiceLogging(CustomLogger):
                 service=ServiceTypes.LITELLM,
                 duration=_duration,
                 call_type=kwargs.get("call_type", "unknown"),
+                start_time=start_time,
+                end_time=end_time,
             )
         except Exception as e:
             raise e

litellm/a2a_protocol/litellm_completion_bridge/handler.py

@@ -20,9 +20,20 @@ from litellm.a2a_protocol.litellm_completion_bridge.transformation import (
 )
 from litellm.a2a_protocol.providers.config_manager import A2AProviderConfigManager
 
+# litellm_params key carrying the authenticated principal (hashed virtual key) so
+# A2A provider configs can scope provider-side state (e.g. LangFlow session memory)
+# per key instead of trusting the client-supplied A2A contextId.
+A2A_USER_API_KEY_HASH_PARAM = "litellm_a2a_user_api_key_hash"
+
 # Agent metadata fields stored in litellm_params that are not valid litellm.acompletion() kwargs
 _AGENT_ONLY_PARAMS = frozenset(
-    {"is_public", "agent_name", "agent_id", "agent_card_params"}
+    {
+        "is_public",
+        "agent_name",
+        "agent_id",
+        "agent_card_params",
+        A2A_USER_API_KEY_HASH_PARAM,
+    }
 )
 
 
@@ -37,6 +48,8 @@ class A2ACompletionBridgeHandler:
         params: Dict[str, Any],
         litellm_params: Dict[str, Any],
         api_base: Optional[str] = None,
+        *,
+        _skip_a2a_provider_routing: bool = False,
     ) -> Dict[str, Any]:
         """
         Handle non-streaming A2A request via litellm.acompletion.
@@ -50,25 +63,24 @@ class A2ACompletionBridgeHandler:
         Returns:
             A2A SendMessageResponse dict
         """
-        # Get provider config for custom_llm_provider
         custom_llm_provider = litellm_params.get("custom_llm_provider")
-        a2a_provider_config = A2AProviderConfigManager.get_provider_config(
-            custom_llm_provider=custom_llm_provider,
-            model=litellm_params.get("model"),
-        )
-
-        # If provider config exists, use it
-        if a2a_provider_config is not None:
-            verbose_logger.info(f"A2A: Using provider config for {custom_llm_provider}")
-
-            response_data = await a2a_provider_config.handle_non_streaming(
-                request_id=request_id,
-                params=params,
-                api_base=api_base,
-                litellm_params=litellm_params,
+        if not _skip_a2a_provider_routing:
+            a2a_provider_config = A2AProviderConfigManager.get_provider_config(
+                custom_llm_provider=custom_llm_provider,
+                model=litellm_params.get("model"),
             )
 
-            return response_data
+            if a2a_provider_config is not None:
+                verbose_logger.info(
+                    f"A2A: Using provider config for {custom_llm_provider}"
+                )
+
+                return await a2a_provider_config.handle_non_streaming(
+                    request_id=request_id,
+                    params=params,
+                    api_base=api_base,
+                    litellm_params=litellm_params,
+                )
 
         # Extract message from params
         message = params.get("message", {})
@@ -107,6 +119,14 @@ class A2ACompletionBridgeHandler:
             if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
         }
         completion_params.update(litellm_params_to_add)
+        # Apply forward metadata AFTER the litellm_params merge so the helper
+        # sees any agent-owner-configured ``extra_body.metadata`` and can keep
+        # those keys authoritative over the client-supplied A2A metadata.
+        A2ACompletionBridgeTransformation.apply_forward_metadata_to_completion_params(
+            completion_params=completion_params,
+            a2a_message=message,
+            params=params,
+        )
 
         # Call litellm.acompletion
         response = await litellm.acompletion(**completion_params)
@@ -129,6 +149,8 @@ class A2ACompletionBridgeHandler:
         params: Dict[str, Any],
         litellm_params: Dict[str, Any],
         api_base: Optional[str] = None,
+        *,
+        _skip_a2a_provider_routing: bool = False,
     ) -> AsyncIterator[Dict[str, Any]]:
         """
         Handle streaming A2A request via litellm.acompletion with stream=True.
@@ -148,28 +170,27 @@ class A2ACompletionBridgeHandler:
         Yields:
             A2A streaming response events
         """
-        # Get provider config for custom_llm_provider
         custom_llm_provider = litellm_params.get("custom_llm_provider")
-        a2a_provider_config = A2AProviderConfigManager.get_provider_config(
-            custom_llm_provider=custom_llm_provider,
-            model=litellm_params.get("model"),
-        )
-
-        # If provider config exists, use it
-        if a2a_provider_config is not None:
-            verbose_logger.info(
-                f"A2A: Using provider config for {custom_llm_provider} (streaming)"
+        if not _skip_a2a_provider_routing:
+            a2a_provider_config = A2AProviderConfigManager.get_provider_config(
+                custom_llm_provider=custom_llm_provider,
+                model=litellm_params.get("model"),
             )
 
-            async for chunk in a2a_provider_config.handle_streaming(
-                request_id=request_id,
-                params=params,
-                api_base=api_base,
-                litellm_params=litellm_params,
-            ):
-                yield chunk
+            if a2a_provider_config is not None:
+                verbose_logger.info(
+                    f"A2A: Using provider config for {custom_llm_provider} (streaming)"
+                )
 
-            return
+                async for chunk in a2a_provider_config.handle_streaming(
+                    request_id=request_id,
+                    params=params,
+                    api_base=api_base,
+                    litellm_params=litellm_params,
+                ):
+                    yield chunk
+
+                return
 
         # Extract message from params
         message = params.get("message", {})
@@ -214,6 +235,14 @@ class A2ACompletionBridgeHandler:
             if k not in ("model", "custom_llm_provider") and k not in _AGENT_ONLY_PARAMS
         }
         completion_params.update(litellm_params_to_add)
+        # Apply forward metadata AFTER the litellm_params merge so the helper
+        # sees any agent-owner-configured ``extra_body.metadata`` and can keep
+        # those keys authoritative over the client-supplied A2A metadata.
+        A2ACompletionBridgeTransformation.apply_forward_metadata_to_completion_params(
+            completion_params=completion_params,
+            a2a_message=message,
+            params=params,
+        )
 
         # 1. Emit initial task event (kind: "task", status: "submitted")
         task_event = A2ACompletionBridgeTransformation.create_task_event(ctx)

litellm/a2a_protocol/litellm_completion_bridge/transformation.py

@@ -45,10 +45,80 @@ class A2ACompletionBridgeTransformation:
     Static methods for transforming between A2A and OpenAI message formats.
     """
 
+    @staticmethod
+    def _extract_text_from_a2a_parts(parts: List[Dict[str, Any]]) -> str:
+        """Extract text from A2A parts (with or without explicit ``kind``)."""
+        content_parts: List[str] = []
+        for part in parts:
+            if not isinstance(part, dict):
+                continue
+            kind = part.get("kind")
+            text = part.get("text")
+            if text is None:
+                continue
+            if kind in (None, "", "text"):
+                content_parts.append(str(text))
+        return "\n".join(content_parts)
+
+    @staticmethod
+    def get_forward_metadata(
+        a2a_message: Dict[str, Any],
+        params: Optional[Dict[str, Any]] = None,
+    ) -> Optional[Dict[str, Any]]:
+        """
+        Merge A2A metadata from MessageSendParams and the message for downstream providers.
+
+        Forwarded once on the LangGraph run payload (``metadata``), not duplicated on
+        each input message — see ``apply_forward_metadata_to_completion_params``.
+        """
+        merged: Dict[str, Any] = {}
+        if params and isinstance(params.get("metadata"), dict):
+            merged.update(params["metadata"])
+        message_metadata = a2a_message.get("metadata")
+        if isinstance(message_metadata, dict):
+            merged.update(message_metadata)
+        return merged or None
+
+    @staticmethod
+    def apply_forward_metadata_to_completion_params(
+        completion_params: Dict[str, Any],
+        a2a_message: Dict[str, Any],
+        params: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Attach A2A metadata to completion kwargs for provider bridges (e.g. LangGraph).
+
+        Uses ``extra_body`` so we do not collide with LiteLLM's spend-log ``metadata`` kwarg.
+        """
+        forward_metadata = A2ACompletionBridgeTransformation.get_forward_metadata(
+            a2a_message=a2a_message,
+            params=params,
+        )
+        if not forward_metadata:
+            return
+
+        extra_body = completion_params.get("extra_body")
+        if not isinstance(extra_body, dict):
+            extra_body = {}
+        # Layer client-supplied A2A metadata under any agent-owner-configured
+        # ``extra_body.metadata`` so the configured keys remain authoritative
+        # and an A2A caller cannot overwrite server-set run metadata.
+        existing_metadata = extra_body.get("metadata")
+        existing_dict: Dict[str, Any] = (
+            existing_metadata if isinstance(existing_metadata, dict) else {}
+        )
+        merged_metadata: Dict[str, Any] = {**forward_metadata, **existing_dict}
+        extra_body = {**extra_body, "metadata": merged_metadata}
+        completion_params["extra_body"] = extra_body
+
+        verbose_logger.debug(
+            f"A2A -> completion forward metadata keys={list(forward_metadata.keys())}"
+        )
+
     @staticmethod
     def a2a_message_to_openai_messages(
         a2a_message: Dict[str, Any],
-    ) -> List[Dict[str, str]]:
+    ) -> List[Dict[str, Any]]:
         """
         Transform an A2A message to OpenAI message format.
 
@@ -70,21 +140,20 @@ class A2ACompletionBridgeTransformation:
         elif role == "system":
             openai_role = "system"
 
-        # Extract text content from parts
-        content_parts = []
-        for part in parts:
-            kind = part.get("kind", "")
-            if kind == "text":
-                text = part.get("text", "")
-                content_parts.append(text)
+        if not isinstance(parts, list):
+            parts = []
 
-        content = "\n".join(content_parts) if content_parts else ""
+        content = A2ACompletionBridgeTransformation._extract_text_from_a2a_parts(parts)
+
+        # Do not attach A2A message.metadata here — the completion bridge forwards it
+        # once at run level via extra_body.metadata (LangGraph POST /runs/wait shape).
+        openai_message: Dict[str, Any] = {"role": openai_role, "content": content}
 
         verbose_logger.debug(
             f"A2A -> OpenAI transform: role={role} -> {openai_role}, content_length={len(content)}"
         )
 
-        return [{"role": openai_role, "content": content}]
+        return [openai_message]
 
     @staticmethod
     def openai_response_to_a2a_response(
@@ -110,6 +179,7 @@ class A2ACompletionBridgeTransformation:
 
         # Build A2A message
         a2a_message = {
+            "kind": "message",
             "role": "agent",
             "parts": [{"kind": "text", "text": content}],
             "messageId": uuid4().hex,
@@ -119,9 +189,7 @@ class A2ACompletionBridgeTransformation:
         a2a_response = {
             "jsonrpc": "2.0",
             "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
+            "result": a2a_message,
         }
 
         verbose_logger.debug(f"OpenAI -> A2A transform: content_length={len(content)}")
@@ -235,50 +303,3 @@ class A2ACompletionBridgeTransformation:
                 "taskId": ctx.task_id,
             },
         }
-
-    @staticmethod
-    def openai_chunk_to_a2a_chunk(
-        chunk: Any,
-        request_id: Optional[str] = None,
-        is_final: bool = False,
-    ) -> Optional[Dict[str, Any]]:
-        """
-        Transform a LiteLLM streaming chunk to A2A streaming format.
-
-        NOTE: This method is deprecated for streaming. Use the event-based
-        methods (create_task_event, create_status_update_event,
-        create_artifact_update_event) instead for proper A2A streaming.
-
-        Args:
-            chunk: LiteLLM ModelResponse chunk
-            request_id: Original A2A request ID
-            is_final: Whether this is the final chunk
-
-        Returns:
-            A2A streaming chunk dict or None if no content
-        """
-        # Extract delta content
-        content = ""
-        if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-            choice = chunk.choices[0]
-            if hasattr(choice, "delta") and choice.delta:
-                content = choice.delta.content or ""
-
-        if not content and not is_final:
-            return None
-
-        # Build A2A streaming chunk (legacy format)
-        a2a_chunk = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": content}],
-                    "messageId": uuid4().hex,
-                },
-                "final": is_final,
-            },
-        }
-
-        return a2a_chunk

litellm/a2a_protocol/main.py

@@ -159,7 +159,9 @@ async def _send_message_via_completion_bridge(
         api_base=api_base,
     )
 
-    return LiteLLMSendMessageResponse.from_dict(response_dict)
+    return LiteLLMSendMessageResponse.from_dict(
+        response_dict, request_id=str(request.id)
+    )
 
 
 async def _execute_a2a_send_with_retry(
@@ -317,15 +319,6 @@ async def asend_message(
     )
     card_url = getattr(agent_card, "url", None) if agent_card else None
 
-    context_id = trace_id or str(uuid.uuid4())
-    message = request.params.message
-    if isinstance(message, dict):
-        if message.get("context_id") is None:
-            message["context_id"] = context_id
-    else:
-        if getattr(message, "context_id", None) is None:
-            message.context_id = context_id
-
     a2a_response = await _execute_a2a_send_with_retry(
         a2a_client=a2a_client,
         request=request,
@@ -338,7 +331,9 @@ async def asend_message(
     verbose_logger.info(f"A2A send_message completed, request_id={request.id}")
 
     # Wrap in LiteLLM response type for _hidden_params support
-    response = LiteLLMSendMessageResponse.from_a2a_response(a2a_response)
+    response = LiteLLMSendMessageResponse.from_a2a_response(
+        a2a_response, request_id=str(request.id)
+    )
 
     # Calculate token usage from request and response
     response_dict = a2a_response.model_dump(mode="json", exclude_none=True)

litellm/a2a_protocol/providers/config_manager.py

@@ -48,4 +48,16 @@ class A2AProviderConfigManager:
 
             return BedrockAgentCoreA2AConfig()
 
+        if custom_llm_provider == "langflow":
+            from litellm.a2a_protocol.providers.langflow.config import LangFlowA2AConfig
+
+            return LangFlowA2AConfig()
+
+        if custom_llm_provider == "watsonx_orchestrate":
+            from litellm.a2a_protocol.providers.watsonx_orchestrate.config import (
+                WatsonxOrchestrateA2AConfig,
+            )
+
+            return WatsonxOrchestrateA2AConfig()
+
         return None

litellm/a2a_protocol/providers/langflow/config.py

@@ -0,0 +1,62 @@
+from typing import Any, AsyncIterator, Dict, Optional
+
+from litellm.a2a_protocol.litellm_completion_bridge.handler import (
+    A2A_USER_API_KEY_HASH_PARAM,
+    A2ACompletionBridgeHandler,
+)
+from litellm.a2a_protocol.providers.base import BaseA2AProviderConfig
+from litellm.llms.langflow.a2a import merge_a2a_session_into_litellm_params
+
+
+class LangFlowA2AConfig(BaseA2AProviderConfig):
+    """A2A bridge for LangFlow: scopes contextId to the authenticated key as the
+    LangFlow session_id, then uses completion."""
+
+    async def handle_non_streaming(
+        self,
+        request_id: str,
+        params: Dict[str, Any],
+        api_base: Optional[str] = None,
+        **kwargs,
+    ) -> Dict[str, Any]:
+        litellm_params = kwargs.get("litellm_params")
+        if not litellm_params:
+            raise ValueError(
+                "litellm_params is required for LangFlowA2AConfig "
+                "(must contain custom_llm_provider and model)"
+            )
+        litellm_params = merge_a2a_session_into_litellm_params(
+            litellm_params, params, litellm_params.get(A2A_USER_API_KEY_HASH_PARAM)
+        )
+        return await A2ACompletionBridgeHandler.handle_non_streaming(
+            request_id=request_id,
+            params=params,
+            litellm_params=litellm_params,
+            api_base=api_base,
+            _skip_a2a_provider_routing=True,
+        )
+
+    async def handle_streaming(
+        self,
+        request_id: str,
+        params: Dict[str, Any],
+        api_base: Optional[str] = None,
+        **kwargs,
+    ) -> AsyncIterator[Dict[str, Any]]:
+        litellm_params = kwargs.get("litellm_params")
+        if not litellm_params:
+            raise ValueError(
+                "litellm_params is required for LangFlowA2AConfig "
+                "(must contain custom_llm_provider and model)"
+            )
+        litellm_params = merge_a2a_session_into_litellm_params(
+            litellm_params, params, litellm_params.get(A2A_USER_API_KEY_HASH_PARAM)
+        )
+        async for chunk in A2ACompletionBridgeHandler.handle_streaming(
+            request_id=request_id,
+            params=params,
+            litellm_params=litellm_params,
+            api_base=api_base,
+            _skip_a2a_provider_routing=True,
+        ):
+            yield chunk

litellm/a2a_protocol/providers/litellm_completion/README.md

@@ -1,74 +0,0 @@
-# A2A to LiteLLM Completion Bridge
-
-Routes A2A protocol requests through `litellm.acompletion`, enabling any LiteLLM-supported provider to be invoked via A2A.
-
-## Flow
-
-```
-A2A Request → Transform → litellm.acompletion → Transform → A2A Response
-```
-
-## SDK Usage
-
-Use the existing `asend_message` and `asend_message_streaming` functions with `litellm_params`:
-
-```python
-from litellm.a2a_protocol import asend_message, asend_message_streaming
-from a2a.types import SendMessageRequest, SendStreamingMessageRequest, MessageSendParams
-from uuid import uuid4
-
-# Non-streaming
-request = SendMessageRequest(
-    id=str(uuid4()),
-    params=MessageSendParams(
-        message={"role": "user", "parts": [{"kind": "text", "text": "Hello!"}], "messageId": uuid4().hex}
-    )
-)
-response = await asend_message(
-    request=request,
-    api_base="http://localhost:2024",
-    litellm_params={"custom_llm_provider": "langgraph", "model": "agent"},
-)
-
-# Streaming
-stream_request = SendStreamingMessageRequest(
-    id=str(uuid4()),
-    params=MessageSendParams(
-        message={"role": "user", "parts": [{"kind": "text", "text": "Hello!"}], "messageId": uuid4().hex}
-    )
-)
-async for chunk in asend_message_streaming(
-    request=stream_request,
-    api_base="http://localhost:2024",
-    litellm_params={"custom_llm_provider": "langgraph", "model": "agent"},
-):
-    print(chunk)
-```
-
-## Proxy Usage
-
-Configure an agent with `custom_llm_provider` in `litellm_params`:
-
-```yaml
-agents:
-  - agent_name: my-langgraph-agent
-    agent_card_params:
-      name: "LangGraph Agent"
-      url: "http://localhost:2024"  # Used as api_base
-    litellm_params:
-      custom_llm_provider: langgraph
-      model: agent
-```
-
-When an A2A request hits `/a2a/{agent_id}/message/send`, the bridge:
-
-1. Detects `custom_llm_provider` in agent's `litellm_params`
-2. Transforms A2A message → OpenAI messages
-3. Calls `litellm.acompletion(model="langgraph/agent", api_base="http://localhost:2024")`
-4. Transforms response → A2A format
-
-## Classes
-
-- `A2ACompletionBridgeTransformation` - Static methods for message format conversion
-- `A2ACompletionBridgeHandler` - Static methods for handling requests (streaming/non-streaming)
-

litellm/a2a_protocol/providers/litellm_completion/__init__.py

@@ -1,5 +0,0 @@
-"""
-LiteLLM Completion bridge provider for A2A protocol.
-
-Routes A2A requests through litellm.acompletion based on custom_llm_provider.
-"""

litellm/a2a_protocol/providers/litellm_completion/handler.py

@@ -1,301 +0,0 @@
-"""
-Handler for A2A to LiteLLM completion bridge.
-
-Routes A2A requests through litellm.acompletion based on custom_llm_provider.
-
-A2A Streaming Events (in order):
-1. Task event (kind: "task") - Initial task creation with status "submitted"
-2. Status update (kind: "status-update") - Status change to "working"
-3. Artifact update (kind: "artifact-update") - Content/artifact delivery
-4. Status update (kind: "status-update") - Final status "completed" with final=true
-"""
-
-from typing import Any, AsyncIterator, Dict, Optional
-
-import litellm
-from litellm._logging import verbose_logger
-from litellm.a2a_protocol.litellm_completion_bridge.pydantic_ai_transformation import (
-    PydanticAITransformation,
-)
-from litellm.a2a_protocol.litellm_completion_bridge.transformation import (
-    A2ACompletionBridgeTransformation,
-    A2AStreamingContext,
-)
-
-
-class A2ACompletionBridgeHandler:
-    """
-    Static methods for handling A2A requests via LiteLLM completion.
-    """
-
-    @staticmethod
-    async def handle_non_streaming(
-        request_id: str,
-        params: Dict[str, Any],
-        litellm_params: Dict[str, Any],
-        api_base: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Handle non-streaming A2A request via litellm.acompletion.
-
-        Args:
-            request_id: A2A JSON-RPC request ID
-            params: A2A MessageSendParams containing the message
-            litellm_params: Agent's litellm_params (custom_llm_provider, model, etc.)
-            api_base: API base URL from agent_card_params
-
-        Returns:
-            A2A SendMessageResponse dict
-        """
-        # Check if this is a Pydantic AI agent request
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        if custom_llm_provider == "pydantic_ai_agents":
-            if api_base is None:
-                raise ValueError("api_base is required for Pydantic AI agents")
-
-            verbose_logger.info(
-                f"Pydantic AI: Routing to Pydantic AI agent at {api_base}"
-            )
-
-            # Send request directly to Pydantic AI agent
-            response_data = await PydanticAITransformation.send_non_streaming_request(
-                api_base=api_base,
-                request_id=request_id,
-                params=params,
-            )
-
-            return response_data
-
-        # Extract message from params
-        message = params.get("message", {})
-
-        # Transform A2A message to OpenAI format
-        openai_messages = (
-            A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
-        )
-
-        # Get completion params
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        model = litellm_params.get("model", "agent")
-
-        # Build full model string if provider specified
-        # Skip prepending if model already starts with the provider prefix
-        if custom_llm_provider and not model.startswith(f"{custom_llm_provider}/"):
-            full_model = f"{custom_llm_provider}/{model}"
-        else:
-            full_model = model
-
-        verbose_logger.info(
-            f"A2A completion bridge: model={full_model}, api_base={api_base}"
-        )
-
-        # Build completion params dict
-        completion_params = {
-            "model": full_model,
-            "messages": openai_messages,
-            "api_base": api_base,
-            "stream": False,
-        }
-        # Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
-        litellm_params_to_add = {
-            k: v
-            for k, v in litellm_params.items()
-            if k not in ("model", "custom_llm_provider")
-        }
-        completion_params.update(litellm_params_to_add)
-
-        # Call litellm.acompletion
-        response = await litellm.acompletion(**completion_params)
-
-        # Transform response to A2A format
-        a2a_response = (
-            A2ACompletionBridgeTransformation.openai_response_to_a2a_response(
-                response=response,
-                request_id=request_id,
-            )
-        )
-
-        verbose_logger.info(f"A2A completion bridge completed: request_id={request_id}")
-
-        return a2a_response
-
-    @staticmethod
-    async def handle_streaming(
-        request_id: str,
-        params: Dict[str, Any],
-        litellm_params: Dict[str, Any],
-        api_base: Optional[str] = None,
-    ) -> AsyncIterator[Dict[str, Any]]:
-        """
-        Handle streaming A2A request via litellm.acompletion with stream=True.
-
-        Emits proper A2A streaming events:
-        1. Task event (kind: "task") - Initial task with status "submitted"
-        2. Status update (kind: "status-update") - Status "working"
-        3. Artifact update (kind: "artifact-update") - Content delivery
-        4. Status update (kind: "status-update") - Final "completed" status
-
-        Args:
-            request_id: A2A JSON-RPC request ID
-            params: A2A MessageSendParams containing the message
-            litellm_params: Agent's litellm_params (custom_llm_provider, model, etc.)
-            api_base: API base URL from agent_card_params
-
-        Yields:
-            A2A streaming response events
-        """
-        # Check if this is a Pydantic AI agent request
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        if custom_llm_provider == "pydantic_ai_agents":
-            if api_base is None:
-                raise ValueError("api_base is required for Pydantic AI agents")
-
-            verbose_logger.info(
-                f"Pydantic AI: Faking streaming for Pydantic AI agent at {api_base}"
-            )
-
-            # Get non-streaming response first
-            response_data = await PydanticAITransformation.send_non_streaming_request(
-                api_base=api_base,
-                request_id=request_id,
-                params=params,
-            )
-
-            # Convert to fake streaming
-            async for chunk in PydanticAITransformation.fake_streaming_from_response(
-                response_data=response_data,
-                request_id=request_id,
-            ):
-                yield chunk
-
-            return
-
-        # Extract message from params
-        message = params.get("message", {})
-
-        # Create streaming context
-        ctx = A2AStreamingContext(
-            request_id=request_id,
-            input_message=message,
-        )
-
-        # Transform A2A message to OpenAI format
-        openai_messages = (
-            A2ACompletionBridgeTransformation.a2a_message_to_openai_messages(message)
-        )
-
-        # Get completion params
-        custom_llm_provider = litellm_params.get("custom_llm_provider")
-        model = litellm_params.get("model", "agent")
-
-        # Build full model string if provider specified
-        # Skip prepending if model already starts with the provider prefix
-        if custom_llm_provider and not model.startswith(f"{custom_llm_provider}/"):
-            full_model = f"{custom_llm_provider}/{model}"
-        else:
-            full_model = model
-
-        verbose_logger.info(
-            f"A2A completion bridge streaming: model={full_model}, api_base={api_base}"
-        )
-
-        # Build completion params dict
-        completion_params = {
-            "model": full_model,
-            "messages": openai_messages,
-            "api_base": api_base,
-            "stream": True,
-        }
-        # Add litellm_params (contains api_key, client_id, client_secret, tenant_id, etc.)
-        litellm_params_to_add = {
-            k: v
-            for k, v in litellm_params.items()
-            if k not in ("model", "custom_llm_provider")
-        }
-        completion_params.update(litellm_params_to_add)
-
-        # 1. Emit initial task event (kind: "task", status: "submitted")
-        task_event = A2ACompletionBridgeTransformation.create_task_event(ctx)
-        yield task_event
-
-        # 2. Emit status update (kind: "status-update", status: "working")
-        working_event = A2ACompletionBridgeTransformation.create_status_update_event(
-            ctx=ctx,
-            state="working",
-            final=False,
-            message_text="Processing request...",
-        )
-        yield working_event
-
-        # Call litellm.acompletion with streaming
-        response = await litellm.acompletion(**completion_params)
-
-        # 3. Accumulate content and emit artifact update
-        accumulated_text = ""
-        chunk_count = 0
-        async for chunk in response:  # type: ignore[union-attr]
-            chunk_count += 1
-
-            # Extract delta content
-            content = ""
-            if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-                choice = chunk.choices[0]
-                if hasattr(choice, "delta") and choice.delta:
-                    content = choice.delta.content or ""
-
-            if content:
-                accumulated_text += content
-
-        # Emit artifact update with accumulated content
-        if accumulated_text:
-            artifact_event = (
-                A2ACompletionBridgeTransformation.create_artifact_update_event(
-                    ctx=ctx,
-                    text=accumulated_text,
-                )
-            )
-            yield artifact_event
-
-        # 4. Emit final status update (kind: "status-update", status: "completed", final: true)
-        completed_event = A2ACompletionBridgeTransformation.create_status_update_event(
-            ctx=ctx,
-            state="completed",
-            final=True,
-        )
-        yield completed_event
-
-        verbose_logger.info(
-            f"A2A completion bridge streaming completed: request_id={request_id}, chunks={chunk_count}"
-        )
-
-
-# Convenience functions that delegate to the class methods
-async def handle_a2a_completion(
-    request_id: str,
-    params: Dict[str, Any],
-    litellm_params: Dict[str, Any],
-    api_base: Optional[str] = None,
-) -> Dict[str, Any]:
-    """Convenience function for non-streaming A2A completion."""
-    return await A2ACompletionBridgeHandler.handle_non_streaming(
-        request_id=request_id,
-        params=params,
-        litellm_params=litellm_params,
-        api_base=api_base,
-    )
-
-
-async def handle_a2a_completion_streaming(
-    request_id: str,
-    params: Dict[str, Any],
-    litellm_params: Dict[str, Any],
-    api_base: Optional[str] = None,
-) -> AsyncIterator[Dict[str, Any]]:
-    """Convenience function for streaming A2A completion."""
-    async for chunk in A2ACompletionBridgeHandler.handle_streaming(
-        request_id=request_id,
-        params=params,
-        litellm_params=litellm_params,
-        api_base=api_base,
-    ):
-        yield chunk

litellm/a2a_protocol/providers/litellm_completion/transformation.py

@@ -1,284 +0,0 @@
-"""
-Transformation utilities for A2A <-> OpenAI message format conversion.
-
-A2A Message Format:
-{
-    "role": "user",
-    "parts": [{"kind": "text", "text": "Hello!"}],
-    "messageId": "abc123"
-}
-
-OpenAI Message Format:
-{"role": "user", "content": "Hello!"}
-
-A2A Streaming Events:
-- Task event (kind: "task") - Initial task creation with status "submitted"
-- Status update (kind: "status-update") - Status changes (working, completed)
-- Artifact update (kind: "artifact-update") - Content/artifact delivery
-"""
-
-from datetime import datetime, timezone
-from typing import Any, Dict, List, Optional
-from uuid import uuid4
-
-from litellm._logging import verbose_logger
-
-
-class A2AStreamingContext:
-    """
-    Context holder for A2A streaming state.
-    Tracks task_id, context_id, and message accumulation.
-    """
-
-    def __init__(self, request_id: str, input_message: Dict[str, Any]):
-        self.request_id = request_id
-        self.task_id = str(uuid4())
-        self.context_id = str(uuid4())
-        self.input_message = input_message
-        self.accumulated_text = ""
-        self.has_emitted_task = False
-        self.has_emitted_working = False
-
-
-class A2ACompletionBridgeTransformation:
-    """
-    Static methods for transforming between A2A and OpenAI message formats.
-    """
-
-    @staticmethod
-    def a2a_message_to_openai_messages(
-        a2a_message: Dict[str, Any],
-    ) -> List[Dict[str, str]]:
-        """
-        Transform an A2A message to OpenAI message format.
-
-        Args:
-            a2a_message: A2A message with role, parts, and messageId
-
-        Returns:
-            List of OpenAI-format messages
-        """
-        role = a2a_message.get("role", "user")
-        parts = a2a_message.get("parts", [])
-
-        # Map A2A roles to OpenAI roles
-        openai_role = role
-        if role == "user":
-            openai_role = "user"
-        elif role == "assistant":
-            openai_role = "assistant"
-        elif role == "system":
-            openai_role = "system"
-
-        # Extract text content from parts
-        content_parts = []
-        for part in parts:
-            kind = part.get("kind", "")
-            if kind == "text":
-                text = part.get("text", "")
-                content_parts.append(text)
-
-        content = "\n".join(content_parts) if content_parts else ""
-
-        verbose_logger.debug(
-            f"A2A -> OpenAI transform: role={role} -> {openai_role}, content_length={len(content)}"
-        )
-
-        return [{"role": openai_role, "content": content}]
-
-    @staticmethod
-    def openai_response_to_a2a_response(
-        response: Any,
-        request_id: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Transform a LiteLLM ModelResponse to A2A SendMessageResponse format.
-
-        Args:
-            response: LiteLLM ModelResponse object
-            request_id: Original A2A request ID
-
-        Returns:
-            A2A SendMessageResponse dict
-        """
-        # Extract content from response
-        content = ""
-        if hasattr(response, "choices") and response.choices:
-            choice = response.choices[0]
-            if hasattr(choice, "message") and choice.message:
-                content = choice.message.content or ""
-
-        # Build A2A message
-        a2a_message = {
-            "role": "agent",
-            "parts": [{"kind": "text", "text": content}],
-            "messageId": uuid4().hex,
-        }
-
-        # Build A2A response
-        a2a_response = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
-        }
-
-        verbose_logger.debug(f"OpenAI -> A2A transform: content_length={len(content)}")
-
-        return a2a_response
-
-    @staticmethod
-    def _get_timestamp() -> str:
-        """Get current timestamp in ISO format with timezone."""
-        return datetime.now(timezone.utc).isoformat()
-
-    @staticmethod
-    def create_task_event(
-        ctx: A2AStreamingContext,
-    ) -> Dict[str, Any]:
-        """
-        Create the initial task event with status 'submitted'.
-
-        This is the first event emitted in an A2A streaming response.
-        """
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "contextId": ctx.context_id,
-                "history": [
-                    {
-                        "contextId": ctx.context_id,
-                        "kind": "message",
-                        "messageId": ctx.input_message.get("messageId", uuid4().hex),
-                        "parts": ctx.input_message.get("parts", []),
-                        "role": ctx.input_message.get("role", "user"),
-                        "taskId": ctx.task_id,
-                    }
-                ],
-                "id": ctx.task_id,
-                "kind": "task",
-                "status": {
-                    "state": "submitted",
-                },
-            },
-        }
-
-    @staticmethod
-    def create_status_update_event(
-        ctx: A2AStreamingContext,
-        state: str,
-        final: bool = False,
-        message_text: Optional[str] = None,
-    ) -> Dict[str, Any]:
-        """
-        Create a status update event.
-
-        Args:
-            ctx: Streaming context
-            state: Status state ('working', 'completed')
-            final: Whether this is the final event
-            message_text: Optional message text for 'working' status
-        """
-        status: Dict[str, Any] = {
-            "state": state,
-            "timestamp": A2ACompletionBridgeTransformation._get_timestamp(),
-        }
-
-        # Add message for 'working' status
-        if state == "working" and message_text:
-            status["message"] = {
-                "contextId": ctx.context_id,
-                "kind": "message",
-                "messageId": str(uuid4()),
-                "parts": [{"kind": "text", "text": message_text}],
-                "role": "agent",
-                "taskId": ctx.task_id,
-            }
-
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "contextId": ctx.context_id,
-                "final": final,
-                "kind": "status-update",
-                "status": status,
-                "taskId": ctx.task_id,
-            },
-        }
-
-    @staticmethod
-    def create_artifact_update_event(
-        ctx: A2AStreamingContext,
-        text: str,
-    ) -> Dict[str, Any]:
-        """
-        Create an artifact update event with content.
-
-        Args:
-            ctx: Streaming context
-            text: The text content for the artifact
-        """
-        return {
-            "id": ctx.request_id,
-            "jsonrpc": "2.0",
-            "result": {
-                "artifact": {
-                    "artifactId": str(uuid4()),
-                    "name": "response",
-                    "parts": [{"kind": "text", "text": text}],
-                },
-                "contextId": ctx.context_id,
-                "kind": "artifact-update",
-                "taskId": ctx.task_id,
-            },
-        }
-
-    @staticmethod
-    def openai_chunk_to_a2a_chunk(
-        chunk: Any,
-        request_id: Optional[str] = None,
-        is_final: bool = False,
-    ) -> Optional[Dict[str, Any]]:
-        """
-        Transform a LiteLLM streaming chunk to A2A streaming format.
-
-        NOTE: This method is deprecated for streaming. Use the event-based
-        methods (create_task_event, create_status_update_event,
-        create_artifact_update_event) instead for proper A2A streaming.
-
-        Args:
-            chunk: LiteLLM ModelResponse chunk
-            request_id: Original A2A request ID
-            is_final: Whether this is the final chunk
-
-        Returns:
-            A2A streaming chunk dict or None if no content
-        """
-        # Extract delta content
-        content = ""
-        if chunk is not None and hasattr(chunk, "choices") and chunk.choices:
-            choice = chunk.choices[0]
-            if hasattr(choice, "delta") and choice.delta:
-                content = choice.delta.content or ""
-
-        if not content and not is_final:
-            return None
-
-        # Build A2A streaming chunk (legacy format)
-        a2a_chunk = {
-            "jsonrpc": "2.0",
-            "id": request_id,
-            "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": content}],
-                    "messageId": uuid4().hex,
-                },
-                "final": is_final,
-            },
-        }
-
-        return a2a_chunk

litellm/a2a_protocol/providers/pydantic_ai_agents/transformation.py

@@ -289,16 +289,16 @@ class PydanticAITransformation:
         Transform Pydantic AI task response to standard A2A non-streaming format.
 
         Pydantic AI returns a task with history/artifacts, but the standard A2A
-        non-streaming format expects:
+        non-streaming format expects ``result`` to be the Message directly
+        (``kind="message"``), per the A2A spec / ``SendMessageResponse``:
         {
             "jsonrpc": "2.0",
             "id": "...",
             "result": {
-                "message": {
-                    "role": "agent",
-                    "parts": [{"kind": "text", "text": "..."}],
-                    "messageId": "..."
-                }
+                "kind": "message",
+                "role": "agent",
+                "parts": [{"kind": "text", "text": "..."}],
+                "messageId": "..."
             }
         }
 
@@ -316,6 +316,7 @@ class PydanticAITransformation:
 
         # Build standard A2A message
         a2a_message = {
+            "kind": "message",
             "role": "agent",
             "parts": parts if parts else [{"kind": "text", "text": full_text}],
             "messageId": message_id,
@@ -325,9 +326,7 @@ class PydanticAITransformation:
         return {
             "jsonrpc": "2.0",
             "id": request_id,
-            "result": {
-                "message": a2a_message,
-            },
+            "result": a2a_message,
         }
 
     @staticmethod

litellm/a2a_protocol/providers/watsonx_orchestrate/__init__.py

@@ -0,0 +1,3 @@
+"""
+IBM watsonx Orchestrate (WXO) A2A provider.
+"""

litellm/a2a_protocol/providers/watsonx_orchestrate/config.py

@@ -0,0 +1,55 @@
+"""
+A2A provider configuration for IBM watsonx Orchestrate (WXO).
+"""
+
+from typing import Any, AsyncIterator, Dict, Optional
+
+from litellm.a2a_protocol.providers.base import BaseA2AProviderConfig
+from litellm.a2a_protocol.providers.watsonx_orchestrate.handler import (
+    WatsonxOrchestrateHandler,
+)
+
+
+class WatsonxOrchestrateA2AConfig(BaseA2AProviderConfig):
+    """A2A bridge for IBM watsonx Orchestrate (REST runs API + poll/SSE)."""
+
+    async def handle_non_streaming(
+        self,
+        request_id: str,
+        params: Dict[str, Any],
+        api_base: Optional[str] = None,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        """Handle a non-streaming A2A request via WXO runs API."""
+        litellm_params = kwargs.get("litellm_params")
+        if not litellm_params:
+            raise ValueError(
+                "litellm_params is required for WatsonxOrchestrateA2AConfig "
+                "(must contain cp4d_host, instance_id, wxo_agent_id, api_key)"
+            )
+        return await WatsonxOrchestrateHandler.handle_non_streaming(
+            request_id=request_id,
+            params=params,
+            litellm_params=litellm_params,
+        )
+
+    async def handle_streaming(
+        self,
+        request_id: str,
+        params: Dict[str, Any],
+        api_base: Optional[str] = None,
+        **kwargs: Any,
+    ) -> AsyncIterator[Dict[str, Any]]:
+        """Handle a streaming A2A request via WXO streaming runs API."""
+        litellm_params = kwargs.get("litellm_params")
+        if not litellm_params:
+            raise ValueError(
+                "litellm_params is required for WatsonxOrchestrateA2AConfig "
+                "(must contain cp4d_host, instance_id, wxo_agent_id, api_key)"
+            )
+        async for chunk in WatsonxOrchestrateHandler.handle_streaming(
+            request_id=request_id,
+            params=params,
+            litellm_params=litellm_params,
+        ):
+            yield chunk

litellm/a2a_protocol/providers/watsonx_orchestrate/handler.py

@@ -0,0 +1,373 @@
+"""
+Handler for IBM watsonx Orchestrate (WXO) agent provider.
+"""
+
+import asyncio
+import hashlib
+import json
+import time
+from typing import Any, AsyncIterator, Dict, NamedTuple, Optional, Tuple, cast
+
+import httpx
+
+from litellm._logging import verbose_logger
+from litellm.a2a_protocol.providers.watsonx_orchestrate.transformation import (
+    WatsonxOrchestrateTransformation,
+)
+from litellm.llms.custom_httpx.http_handler import (
+    AsyncHTTPHandler,
+    get_async_httpx_client,
+)
+from litellm.types.llms.custom_http import httpxSpecialProvider
+
+_IBM_CLOUD_IAM_URL = "https://iam.cloud.ibm.com/identity/token"
+_POLL_INTERVAL_S = 2.0
+_MAX_POLL_ATTEMPTS = 90
+_TOKEN_CACHE_TTL_BUFFER_S = 60
+_token_cache: Dict[str, Tuple[str, float]] = {}
+
+
+class WXORequestParams(NamedTuple):
+    cp4d_host: str
+    instance_id: str
+    wxo_agent_id: str
+    api_key: str
+    username: Optional[str]
+    auth_mode: str
+    thread_id: Optional[str]
+
+
+class WatsonxOrchestrateHandler:
+    @staticmethod
+    def _http_client(timeout: float = 90.0) -> AsyncHTTPHandler:
+        return get_async_httpx_client(
+            llm_provider=cast(Any, httpxSpecialProvider.A2AProvider),
+            params={"timeout": timeout},
+        )
+
+    @staticmethod
+    def _token_cache_key(
+        auth_mode: str,
+        cp4d_host: str,
+        api_key: str,
+        username: Optional[str],
+    ) -> str:
+        material = f"{auth_mode}:{cp4d_host}:{username or ''}:{api_key}"
+        return hashlib.sha256(material.encode()).hexdigest()
+
+    @staticmethod
+    def _cp4d_token_ttl_seconds(
+        expiration: Any, now_wall: Optional[float] = None
+    ) -> int:
+        # CP4D returns expiration as absolute Unix epoch seconds, not a duration.
+        expires_at = int(expiration)
+        wall = now_wall if now_wall is not None else time.time()
+        return max(expires_at - int(wall), 0)
+
+    @staticmethod
+    async def _get_bearer_token(
+        cp4d_host: str,
+        auth_mode: str,
+        api_key: str,
+        username: Optional[str] = None,
+        client: Optional[AsyncHTTPHandler] = None,
+    ) -> str:
+        cache_key = WatsonxOrchestrateHandler._token_cache_key(
+            auth_mode, cp4d_host, api_key, username
+        )
+        now = time.monotonic()
+        cached = _token_cache.get(cache_key)
+        if cached and cached[1] > now:
+            return cached[0]
+
+        if client is None:
+            client = WatsonxOrchestrateHandler._http_client(timeout=30.0)
+
+        if auth_mode == "ibm_cloud":
+            response = await client.post(
+                _IBM_CLOUD_IAM_URL,
+                data={
+                    "grant_type": "urn:ibm:params:oauth:grant-type:apikey",
+                    "apikey": api_key,
+                },
+                headers={"Content-Type": "application/x-www-form-urlencoded"},
+            )
+            response.raise_for_status()
+            payload = response.json()
+            token = str(payload["access_token"])
+            ttl_s = int(payload.get("expires_in", 3600))
+        else:
+            if not username:
+                raise ValueError(
+                    "'username' is required in litellm_params when auth_mode='cp4d'"
+                )
+            token_url = f"{cp4d_host.rstrip('/')}/icp4d-api/v1/authorize"
+            response = await client.post(
+                token_url,
+                json={"username": username, "api_key": api_key},
+                headers={"Content-Type": "application/json"},
+            )
+            response.raise_for_status()
+            payload = response.json()
+            token = str(payload["token"])
+            expiration = payload.get("expiration")
+            if expiration is None:
+                ttl_s = 3600
+            else:
+                ttl_s = WatsonxOrchestrateHandler._cp4d_token_ttl_seconds(expiration)
+
+        expires_at = now + max(ttl_s - _TOKEN_CACHE_TTL_BUFFER_S, 0)
+        _token_cache[cache_key] = (token, expires_at)
+        for stale_key, (_, stale_expires_at) in list(_token_cache.items()):
+            if stale_expires_at <= now:
+                del _token_cache[stale_key]
+        return token
+
+    @staticmethod
+    async def _poll_run(
+        base_url: str,
+        run_id: str,
+        auth_headers: Dict[str, str],
+        client: AsyncHTTPHandler,
+        max_attempts: int = _MAX_POLL_ATTEMPTS,
+        interval_s: float = _POLL_INTERVAL_S,
+    ) -> Dict[str, Any]:
+        url = f"{base_url}/v1/orchestrate/runs/{run_id}"
+
+        for attempt in range(max_attempts):
+            await asyncio.sleep(interval_s)
+            response = await client.get(url, headers=auth_headers)
+            response.raise_for_status()
+            result: Dict[str, Any] = response.json()
+            status = result.get("status", "")
+            verbose_logger.debug(
+                f"WXO: Poll {attempt + 1}/{max_attempts} run='{run_id}' status='{status}'"
+            )
+            if status in WatsonxOrchestrateTransformation.TERMINAL_STATES:
+                return result
+
+        raise asyncio.TimeoutError(
+            f"WXO run '{run_id}' did not reach a terminal state after "
+            f"{max_attempts * interval_s:.0f}s"
+        )
+
+    @staticmethod
+    async def _get_successful_run_data(
+        run_data: Dict[str, Any],
+        base_url: str,
+        auth_headers: Dict[str, str],
+        client: AsyncHTTPHandler,
+    ) -> Dict[str, Any]:
+        status = run_data.get("status", "")
+        if status not in WatsonxOrchestrateTransformation.TERMINAL_STATES:
+            run_id = run_data.get("run_id") or run_data.get("id") or ""
+            if not run_id:
+                raise ValueError(f"WXO: No run_id in response: {run_data}")
+            run_data = await WatsonxOrchestrateHandler._poll_run(
+                base_url=base_url,
+                run_id=run_id,
+                auth_headers=auth_headers,
+                client=client,
+            )
+            status = run_data.get("status", "")
+
+        if status not in WatsonxOrchestrateTransformation.SUCCESS_STATES:
+            raise RuntimeError(
+                f"WXO run ended with non-success status '{status}': {run_data}"
+            )
+
+        return run_data
+
+    @staticmethod
+    async def _accumulate_wxo_sse_text(response: Any) -> str:
+        accumulated_text = ""
+        async for line in response.aiter_lines():
+            if not line.startswith("data:"):
+                continue
+            data_str = line[5:].strip()
+            if not data_str or data_str == "[DONE]":
+                continue
+            try:
+                event = json.loads(data_str)
+            except json.JSONDecodeError:
+                continue
+            chunk_text = WatsonxOrchestrateTransformation.extract_text_from_wxo_result(
+                event
+            )
+            if chunk_text:
+                accumulated_text += chunk_text
+        return accumulated_text
+
+    @staticmethod
+    def _extract_litellm_params(litellm_params: Dict[str, Any]) -> WXORequestParams:
+        cp4d_host = litellm_params.get("cp4d_host") or ""
+        instance_id = litellm_params.get("instance_id") or ""
+        wxo_agent_id = litellm_params.get("wxo_agent_id") or ""
+        api_key = litellm_params.get("api_key") or ""
+
+        if not cp4d_host:
+            raise ValueError("'cp4d_host' is required in litellm_params for WXO agents")
+        if not instance_id:
+            raise ValueError(
+                "'instance_id' is required in litellm_params for WXO agents"
+            )
+        if not wxo_agent_id:
+            raise ValueError(
+                "'wxo_agent_id' is required in litellm_params for WXO agents"
+            )
+        if not api_key:
+            raise ValueError("'api_key' is required in litellm_params for WXO agents")
+
+        return WXORequestParams(
+            cp4d_host=cp4d_host,
+            instance_id=instance_id,
+            wxo_agent_id=wxo_agent_id,
+            api_key=api_key,
+            username=litellm_params.get("username") or None,
+            auth_mode=litellm_params.get("auth_mode") or "cp4d",
+            thread_id=litellm_params.get("thread_id") or None,
+        )
+
+    @staticmethod
+    async def handle_non_streaming(
+        request_id: str,
+        params: Dict[str, Any],
+        litellm_params: Dict[str, Any],
+    ) -> Dict[str, Any]:
+        wxo = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
+
+        client = WatsonxOrchestrateHandler._http_client(timeout=90.0)
+        token = await WatsonxOrchestrateHandler._get_bearer_token(
+            cp4d_host=wxo.cp4d_host,
+            auth_mode=wxo.auth_mode,
+            api_key=wxo.api_key,
+            username=wxo.username,
+            client=client,
+        )
+        base_url = WatsonxOrchestrateTransformation.get_api_base_url(
+            wxo.cp4d_host, wxo.instance_id
+        )
+        auth_headers = {
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+            "Accept": "application/json",
+        }
+
+        text = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
+        body = WatsonxOrchestrateTransformation.build_wxo_run_body(
+            wxo_agent_id=wxo.wxo_agent_id, text=text, thread_id=wxo.thread_id
+        )
+
+        run_response = await client.post(
+            f"{base_url}/v1/orchestrate/runs",
+            json=body,
+            headers=auth_headers,
+        )
+        run_response.raise_for_status()
+        run_data: Dict[str, Any] = run_response.json()
+
+        run_data = await WatsonxOrchestrateHandler._get_successful_run_data(
+            run_data=run_data,
+            base_url=base_url,
+            auth_headers=auth_headers,
+            client=client,
+        )
+
+        response_text = WatsonxOrchestrateTransformation.extract_text_from_wxo_result(
+            run_data
+        )
+        return WatsonxOrchestrateTransformation.build_a2a_message_response(
+            request_id=request_id, text=response_text
+        )
+
+    @staticmethod
+    async def handle_streaming(
+        request_id: str,
+        params: Dict[str, Any],
+        litellm_params: Dict[str, Any],
+        chunk_size: int = 50,
+        delay_ms: int = 10,
+    ) -> AsyncIterator[Dict[str, Any]]:
+        wxo = WatsonxOrchestrateHandler._extract_litellm_params(litellm_params)
+
+        client = WatsonxOrchestrateHandler._http_client(timeout=120.0)
+        token = await WatsonxOrchestrateHandler._get_bearer_token(
+            cp4d_host=wxo.cp4d_host,
+            auth_mode=wxo.auth_mode,
+            api_key=wxo.api_key,
+            username=wxo.username,
+            client=client,
+        )
+        base_url = WatsonxOrchestrateTransformation.get_api_base_url(
+            wxo.cp4d_host, wxo.instance_id
+        )
+        auth_headers = {
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+            "Accept": "text/event-stream, application/json",
+        }
+        text = WatsonxOrchestrateTransformation.extract_text_from_a2a_params(params)
+        body = WatsonxOrchestrateTransformation.build_wxo_run_body(
+            wxo_agent_id=wxo.wxo_agent_id, text=text, thread_id=wxo.thread_id
+        )
+
+        try:
+            response = await client.post(
+                f"{base_url}/v1/orchestrate/runs/stream",
+                json=body,
+                headers=auth_headers,
+                stream=True,
+            )
+            response.raise_for_status()
+        except httpx.TransportError as exc:
+            verbose_logger.warning(
+                f"WXO: Streaming request failed before a run was submitted "
+                f"({exc!r}), falling back to non-streaming + fake streaming",
+                exc_info=True,
+            )
+            result = await WatsonxOrchestrateHandler.handle_non_streaming(
+                request_id=request_id,
+                params=params,
+                litellm_params=litellm_params,
+            )
+            response_text = (
+                WatsonxOrchestrateTransformation.extract_text_from_a2a_message_response(
+                    result
+                )
+            )
+            async for (
+                chunk
+            ) in WatsonxOrchestrateTransformation.fake_streaming_from_text(
+                text=response_text,
+                request_id=request_id,
+                chunk_size=chunk_size,
+                delay_ms=delay_ms,
+            ):
+                yield chunk
+            return
+
+        content_type = response.headers.get("content-type", "").lower()
+        if "text/event-stream" not in content_type:
+            response_body = await response.aread()
+            result = json.loads(response_body)
+            result = await WatsonxOrchestrateHandler._get_successful_run_data(
+                run_data=result,
+                base_url=base_url,
+                auth_headers=auth_headers,
+                client=client,
+            )
+            accumulated_text = (
+                WatsonxOrchestrateTransformation.extract_text_from_wxo_result(result)
+            )
+        else:
+            accumulated_text = await WatsonxOrchestrateHandler._accumulate_wxo_sse_text(
+                response
+            )
+
+        async for chunk in WatsonxOrchestrateTransformation.fake_streaming_from_text(
+            text=accumulated_text,
+            request_id=request_id,
+            chunk_size=chunk_size,
+            delay_ms=delay_ms,
+        ):
+            yield chunk

litellm/a2a_protocol/providers/watsonx_orchestrate/transformation.py

@@ -0,0 +1,224 @@
+"""
+Transformation layer for IBM watsonx Orchestrate (WXO) agent provider.
+
+WXO uses a REST API (not A2A/JSON-RPC) with an async-poll execution model:
+  POST /v1/orchestrate/runs       → submit run, get run_id
+  GET  /v1/orchestrate/runs/{id}  → poll until terminal state
+  POST /v1/orchestrate/runs/stream → native SSE streaming
+"""
+
+import asyncio
+from typing import Any, AsyncIterator, Dict, Optional
+from uuid import uuid4
+
+from litellm._logging import verbose_logger
+
+
+class WatsonxOrchestrateTransformation:
+    """
+    Handles request/response transformation between A2A and the WXO REST API.
+    """
+
+    TERMINAL_STATES = frozenset(
+        {"completed", "succeeded", "failed", "error", "cancelled"}
+    )
+    SUCCESS_STATES = frozenset({"completed", "succeeded"})
+
+    @staticmethod
+    def get_api_base_url(cp4d_host: str, instance_id: str) -> str:
+        """Build the WXO API base URL from host and instance ID."""
+        return f"{cp4d_host.rstrip('/')}/orchestrate/cpd/instances/{instance_id}"
+
+    @staticmethod
+    def extract_text_from_a2a_params(params: Dict[str, Any]) -> str:
+        """
+        Extract user message text from A2A MessageSendParams.
+
+        A2A format: params.message.parts[*] where part.kind == "text"
+        """
+        message = params.get("message", {})
+        parts = message.get("parts", [])
+        texts = []
+        for part in parts:
+            if not isinstance(part, dict):
+                continue
+            kind = part.get("kind")
+            if kind in (None, "", "text") and part.get("text"):
+                texts.append(part["text"])
+        return " ".join(texts) or ""
+
+    @staticmethod
+    def build_wxo_run_body(
+        wxo_agent_id: str,
+        text: str,
+        thread_id: Optional[str] = None,
+    ) -> Dict[str, Any]:
+        """Build the WXO POST /v1/orchestrate/runs request body."""
+        body: Dict[str, Any] = {
+            "agent_id": wxo_agent_id,
+            "message": {
+                "role": "user",
+                "content": [
+                    {
+                        "response_type": "text",
+                        "text": text,
+                    }
+                ],
+            },
+        }
+        if thread_id:
+            body["thread_id"] = thread_id
+        return body
+
+    @staticmethod
+    def extract_text_from_wxo_result(result: Any) -> str:
+        """
+        Extract response text from a WXO run result.
+
+        WXO can return text in several locations; checks in priority order per the API spec.
+        """
+        if not isinstance(result, dict):
+            return ""
+
+        # Primary: last_message.content[0].text
+        try:
+            text = result["last_message"]["content"][0]["text"]
+            if text:
+                return str(text)
+        except (KeyError, IndexError, TypeError):
+            pass
+
+        # Secondary: result.data.message.content[0].text
+        try:
+            text = result["result"]["data"]["message"]["content"][0]["text"]
+            if text:
+                return str(text)
+        except (KeyError, IndexError, TypeError):
+            pass
+
+        # Tertiary: results as a raw string
+        results = result.get("results")
+        if results and isinstance(results, str):
+            return results
+
+        return ""
+
+    @staticmethod
+    def extract_text_from_a2a_message_response(a2a_response: Dict[str, Any]) -> str:
+        result = a2a_response.get("result")
+        if not isinstance(result, dict):
+            verbose_logger.warning("WXO: A2A response missing result object")
+            return ""
+        parts = result.get("parts")
+        if not isinstance(parts, list):
+            verbose_logger.warning("WXO: A2A result has no parts list")
+            return ""
+        for part in parts:
+            if (
+                isinstance(part, dict)
+                and part.get("kind") == "text"
+                and part.get("text")
+            ):
+                return str(part["text"])
+        verbose_logger.warning("WXO: A2A result parts contained no text")
+        return ""
+
+    @staticmethod
+    def build_a2a_message_response(request_id: str, text: str) -> Dict[str, Any]:
+        """
+        Build a standard A2A non-streaming SendMessageResponse (kind=message).
+        """
+        return {
+            "jsonrpc": "2.0",
+            "id": request_id,
+            "result": {
+                "kind": "message",
+                "role": "agent",
+                "parts": [{"kind": "text", "text": text}],
+                "messageId": str(uuid4()),
+            },
+        }
+
+    @staticmethod
+    async def fake_streaming_from_text(
+        text: str,
+        request_id: str,
+        chunk_size: int = 50,
+        delay_ms: int = 10,
+    ) -> AsyncIterator[Dict[str, Any]]:
+        """
+        Emit standard A2A streaming events from a completed text response.
+
+        Event sequence:
+          1. task         (kind="task", state="submitted")
+          2. status-update (kind="status-update", state="working")
+          3. artifact-update chunks
+          4. status-update (kind="status-update", state="completed", final=True)
+        """
+        task_id = str(uuid4())
+        context_id = str(uuid4())
+        artifact_id = str(uuid4())
+
+        # 1. Task submitted
+        yield {
+            "jsonrpc": "2.0",
+            "id": request_id,
+            "result": {
+                "contextId": context_id,
+                "id": task_id,
+                "kind": "task",
+                "status": {"state": "submitted"},
+            },
+        }
+
+        # 2. Working
+        yield {
+            "jsonrpc": "2.0",
+            "id": request_id,
+            "result": {
+                "contextId": context_id,
+                "final": False,
+                "kind": "status-update",
+                "status": {"state": "working"},
+                "taskId": task_id,
+            },
+        }
+        await asyncio.sleep(delay_ms / 1000.0)
+
+        # 3. Artifact chunks (always emit at least one chunk, even for empty text)
+        text_to_chunk = text or ""
+        for i in range(0, max(len(text_to_chunk), 1), chunk_size):
+            chunk_text = text_to_chunk[i : i + chunk_size]
+            is_last = (i + chunk_size) >= max(len(text_to_chunk), 1)
+            yield {
+                "jsonrpc": "2.0",
+                "id": request_id,
+                "result": {
+                    "contextId": context_id,
+                    "kind": "artifact-update",
+                    "taskId": task_id,
+                    "artifact": {
+                        "artifactId": artifact_id,
+                        "parts": [{"kind": "text", "text": chunk_text}],
+                    },
+                },
+            }
+            if not is_last:
+                await asyncio.sleep(delay_ms / 1000.0)
+
+        # 4. Completed
+        yield {
+            "jsonrpc": "2.0",
+            "id": request_id,
+            "result": {
+                "contextId": context_id,
+                "final": True,
+                "kind": "status-update",
+                "status": {"state": "completed"},
+                "taskId": task_id,
+            },
+        }
+
+        verbose_logger.debug(
+            f"WXO: Fake streaming completed for request_id={request_id}"
+        )

litellm/a2a_protocol/utils.py

@@ -60,6 +60,12 @@ class A2ARequestUtils:
         if not isinstance(result, dict):
             return ""
 
+        # Direct message format (A2A spec): detect by explicit kind tag only.
+        # The "parts" heuristic is too broad and would match any future result
+        # type that happens to include a "parts" field.
+        if result.get("kind") == "message":
+            return A2ARequestUtils.extract_text_from_message(result)
+
         message = result.get("message", {})
         return A2ARequestUtils.extract_text_from_message(message)
 

litellm/anthropic_interface/messages/__init__.py

@@ -10,7 +10,7 @@ This is an __init__.py file to allow the following interface
 
 """
 
-from typing import Any, AsyncIterator, Coroutine, Dict, List, Optional, Union
+from typing import Any, AsyncIterator, Coroutine, Dict, Iterator, List, Optional, Union
 
 from litellm.llms.anthropic.experimental_pass_through.messages.handler import (
     anthropic_messages as _async_anthropic_messages,
@@ -100,8 +100,11 @@ def create(
     **kwargs,
 ) -> Union[
     AnthropicMessagesResponse,
+    Iterator[bytes],
     AsyncIterator[Any],
-    Coroutine[Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any]]],
+    Coroutine[
+        Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]]
+    ],
 ]:
     """
     Async wrapper for Anthropic's messages API

litellm/completion_extras/litellm_responses_transformation/handler.py

@@ -182,6 +182,14 @@ class ResponsesToCompletionBridgeHandler:
             client=kwargs.get("client"),
         )
 
+        # Pin the resolved provider so `responses()` doesn't re-run
+        # `get_llm_provider()` on the model string and strip a second
+        # provider prefix (see GitHub issue #28505).  request_data already
+        # carries `custom_llm_provider` via the spread of
+        # `sanitized_litellm_params`; overwriting it on the dict (rather
+        # than adding an explicit kwarg) avoids the duplicate-keyword
+        # TypeError that would otherwise fire on the real bridge path.
+        request_data["custom_llm_provider"] = custom_llm_provider
         result = responses(
             **request_data,
         )
@@ -268,6 +276,13 @@ class ResponsesToCompletionBridgeHandler:
         except Exception as e:
             raise e
 
+        # Pin the resolved provider so `aresponses()` doesn't re-run
+        # `get_llm_provider()` on the model string and strip a second
+        # provider prefix (see GitHub issue #28505).  Set on request_data
+        # rather than passed as a separate kwarg to avoid the duplicate-
+        # keyword TypeError when `sanitized_litellm_params` already
+        # carries `custom_llm_provider`.
+        request_data["custom_llm_provider"] = custom_llm_provider
         result = await aresponses(
             **request_data,
             aresponses=True,

litellm/constants.py

@@ -585,6 +585,7 @@ LITELLM_CHAT_PROVIDERS = [
     "volcengine",
     "codestral",
     "text-completion-codestral",
+    "text-completion-inception",
     "deepseek",
     "sambanova",
     "maritalk",
@@ -620,6 +621,7 @@ LITELLM_CHAT_PROVIDERS = [
     "oci",
     "morph",
     "lambda_ai",
+    "inception",
     "vercel_ai_gateway",
     "wandb",
     "ovhcloud",
@@ -771,6 +773,7 @@ openai_compatible_endpoints: List = [
     "https://api.moonshot.ai/v1",
     "https://api.publicai.co/v1",
     "https://api.synthetic.new/openai/v1",
+    "https://serverless.tensormesh.ai/v1",
     "https://api.stima.tech/v1",
     "https://nano-gpt.com/api/v1",
     "https://api.poe.com/v1",
@@ -778,6 +781,7 @@ openai_compatible_endpoints: List = [
     "https://api.v0.dev/v1",
     "https://api.morphllm.com/v1",
     "https://api.lambda.ai/v1",
+    "https://api.inceptionlabs.ai/v1",
     "https://api.hyperbolic.xyz/v1",
     "https://ai-gateway.helicone.ai/",
     "https://ai-gateway.vercel.sh/v1",
@@ -820,6 +824,7 @@ openai_compatible_providers: List = [
     "meta_llama",
     "publicai",  # PublicAI - JSON-configured provider
     "synthetic",  # Synthetic - JSON-configured provider
+    "tensormesh",  # Tensormesh - JSON-configured provider
     "apertis",  # Apertis - JSON-configured provider
     "nano-gpt",  # Nano-GPT - JSON-configured provider
     "poe",  # Poe - JSON-configured provider
@@ -833,6 +838,7 @@ openai_compatible_providers: List = [
     "helicone",
     "morph",
     "lambda_ai",
+    "inception",
     "hyperbolic",
     "vercel_ai_gateway",
     "aiml",
@@ -855,6 +861,7 @@ openai_text_completion_compatible_providers: List = (
         "moonshot",
         "publicai",
         "synthetic",
+        "tensormesh",
         "apertis",
         "nano-gpt",
         "poe",
@@ -868,6 +875,7 @@ openai_text_completion_compatible_providers: List = (
 _openai_like_providers: List = [
     "predibase",
     "databricks",
+    "lemonade",
     "watsonx",
 ]  # private helper. similar to openai but require some custom auth / endpoint handling, so can't use the openai sdk
 # well supported replicate llms
@@ -1409,6 +1417,13 @@ LITELLM_INTERNAL_JOBS_SERVICE_ACCOUNT_NAME = "litellm_internal_jobs"
 # Prometheus metrics, audit trails, or any other downstream consumer.
 LITELLM_PROXY_MASTER_KEY_ALIAS = "litellm_proxy_master_key"
 
+# Marker placed in ``model_call_details`` on a synthetic ``Logging`` object that
+# records a proxy-gate error (auth/rate-limit rejection) for a request that never
+# reached an upstream provider. Tracing callbacks key off it to avoid fabricating
+# an LLM-call span for a call that did not happen. See
+# ``ProxyLogging._handle_logging_proxy_only_error``.
+LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL = "litellm_no_upstream_llm_call"
+
 # Key Rotation Constants
 LITELLM_KEY_ROTATION_ENABLED = os.getenv("LITELLM_KEY_ROTATION_ENABLED", "false")
 LITELLM_KEY_ROTATION_CHECK_INTERVAL_SECONDS = int(

litellm/experimental_mcp_client/client.py

@@ -421,8 +421,16 @@ class MCPClient:
 
         return factory
 
-    async def list_tools(self) -> List[MCPTool]:
-        """List available tools from the server."""
+    async def list_tools(self, raise_on_error: bool = False) -> List[MCPTool]:
+        """List available tools from the server.
+
+        Args:
+            raise_on_error: When True, re-raise exceptions instead of returning
+                an empty list. Used by the proxy's pass-through MCP flow so it
+                can surface upstream HTTP 401 responses as a proper 401 to the
+                MCP client (triggering the upstream OAuth flow) rather than
+                masking them as "connected, no tools".
+        """
         verbose_logger.debug(
             f"MCP client listing tools from {self.server_url or 'stdio'}"
         )
@@ -458,6 +466,8 @@ class MCPClient:
                     "the MCP server may have crashed, disconnected, or timed out"
                 )
 
+            if raise_on_error:
+                raise
             # Return empty list instead of raising to allow graceful degradation
             return []
 

litellm/integrations/arize/_utils.py

@@ -8,18 +8,23 @@ from litellm.integrations.opentelemetry_utils.base_otel_llm_obs_attributes impor
     BaseLLMObsOTELAttributes,
     safe_set_attribute,
 )
+from litellm.litellm_core_utils.redact_messages import (
+    should_redact_message_logging,
+)
 from litellm.litellm_core_utils.safe_json_dumps import safe_dumps
 from litellm.types.utils import StandardLoggingPayload
 
 if TYPE_CHECKING:
     from opentelemetry.trace import Span
 from litellm.integrations._types.open_inference import (
-    MessageAttributes,
-    ImageAttributes,
-    SpanAttributes,
     AudioAttributes,
     EmbeddingAttributes,
+    ImageAttributes,
+    MessageAttributes,
+    MessageContentAttributes,
     OpenInferenceSpanKindValues,
+    SpanAttributes,
+    ToolCallAttributes,
 )
 
 
@@ -53,40 +58,24 @@ class ArizeOTELAttributes(BaseLLMObsOTELAttributes):
                     msg.get("content", ""),
                 )
 
-    @staticmethod
-    @override
-    def set_response_output_messages(span: "Span", response_obj):
-        """
-        Sets output message attributes on the span from the LLM response.
-        Args:
-            span: The OpenTelemetry span to set attributes on
-            response_obj: The response object containing choices with messages
-        """
-        from litellm.integrations._types.open_inference import (
-            MessageAttributes,
-            SpanAttributes,
-        )
+                # Additive: emit structured tool_calls / multimodal content
+                # so Arize/Phoenix can render tool-using and image-bearing
+                # turns. These set NEW attribute keys (MESSAGE_TOOL_CALLS /
+                # MESSAGE_NAME / MESSAGE_TOOL_CALL_ID / MESSAGE_CONTENTS.*) —
+                # never replace the MESSAGE_CONTENT write above.
+                _safe_emit(
+                    f"input message extras (idx={idx})",
+                    _emit_input_message_extras,
+                    span,
+                    prefix,
+                    msg,
+                )
 
-        for idx, choice in enumerate(response_obj.get("choices", [])):
-            response_message = choice.get("message", {})
-            safe_set_attribute(
-                span,
-                SpanAttributes.OUTPUT_VALUE,
-                response_message.get("content", ""),
-            )
-
-            # This shows up under `output_messages` tab on the span page.
-            prefix = f"{SpanAttributes.LLM_OUTPUT_MESSAGES}.{idx}"
-            safe_set_attribute(
-                span,
-                f"{prefix}.{MessageAttributes.MESSAGE_ROLE}",
-                response_message.get("role"),
-            )
-            safe_set_attribute(
-                span,
-                f"{prefix}.{MessageAttributes.MESSAGE_CONTENT}",
-                response_message.get("content", ""),
-            )
+    # Note: `BaseLLMObsOTELAttributes.set_response_output_messages` is not
+    # overridden here. The live code path uses `_set_choice_outputs` (called
+    # via `_set_response_attributes` from `set_attributes`) which handles
+    # tool_calls, multimodal output, embeddings, audio, images, and structured
+    # outputs in a single place.
 
 
 def _set_response_attributes(span: "Span", response_obj):
@@ -106,11 +95,17 @@ def _set_response_attributes(span: "Span", response_obj):
 def _set_choice_outputs(span: "Span", response_obj, msg_attrs, span_attrs):
     for idx, choice in enumerate(response_obj.get("choices", [])):
         response_message = choice.get("message", {})
-        safe_set_attribute(
-            span,
-            span_attrs.OUTPUT_VALUE,
-            response_message.get("content", ""),
-        )
+        content = response_message.get("content", "")
+
+        # Tool-only assistant responses have empty content; serialize the
+        # tool_calls into OUTPUT_VALUE so Arize's "Output" pane isn't blank.
+        output_value = content
+        if not output_value:
+            tool_calls = _get_tool_calls(response_message)
+            if tool_calls:
+                output_value = _summarize_tool_calls_for_output(tool_calls)
+
+        safe_set_attribute(span, span_attrs.OUTPUT_VALUE, output_value)
         prefix = f"{span_attrs.LLM_OUTPUT_MESSAGES}.{idx}"
         safe_set_attribute(
             span,
@@ -120,7 +115,18 @@ def _set_choice_outputs(span: "Span", response_obj, msg_attrs, span_attrs):
         safe_set_attribute(
             span,
             f"{prefix}.{msg_attrs.MESSAGE_CONTENT}",
-            response_message.get("content", ""),
+            content,
+        )
+
+        # Additive: emit assistant tool_calls so tool-using turns render in
+        # Arize/Phoenix. Sets new MESSAGE_TOOL_CALLS keys only — does not
+        # change MESSAGE_CONTENT/MESSAGE_ROLE writes above.
+        _safe_emit(
+            f"output tool_calls (idx={idx})",
+            _emit_message_tool_calls,
+            span,
+            prefix,
+            response_message,
         )
 
 
@@ -278,6 +284,43 @@ def _set_usage_outputs(span: "Span", response_obj, span_attrs):
             reasoning_tokens,
         )
 
+    # Additive: cache token breakdown so prompt-caching savings render in
+    # Arize. Sources covered:
+    #   - OpenAI Chat Completions: `prompt_tokens_details.cached_tokens`
+    #   - Anthropic / Bedrock-Anthropic: `cache_read_input_tokens`,
+    #     `cache_creation_input_tokens`
+    # All emits are conditional, so when none of these fields exist (the
+    # situation in the existing test fixtures) no extra attributes are set.
+    prompt_token_details = _safe_get(usage, "prompt_tokens_details") or _safe_get(
+        usage, "input_tokens_details"
+    )
+    cache_read = _safe_get(prompt_token_details, "cached_tokens") or _safe_get(
+        usage, "cache_read_input_tokens"
+    )
+    if cache_read:
+        safe_set_attribute(
+            span,
+            span_attrs.LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_READ,
+            cache_read,
+        )
+    # Anthropic / Bedrock-Anthropic only — OpenAI's `prompt_tokens_details`
+    # does not expose a cache-write count, so we read straight off `usage`.
+    cache_write = _safe_get(usage, "cache_creation_input_tokens")
+    if cache_write:
+        safe_set_attribute(
+            span,
+            span_attrs.LLM_TOKEN_COUNT_PROMPT_DETAILS_CACHE_WRITE,
+            cache_write,
+        )
+
+    audio_prompt_tokens = _safe_get(prompt_token_details, "audio_tokens")
+    if audio_prompt_tokens:
+        safe_set_attribute(
+            span,
+            span_attrs.LLM_TOKEN_COUNT_PROMPT_DETAILS_AUDIO,
+            audio_prompt_tokens,
+        )
+
 
 def _infer_open_inference_span_kind(call_type: Optional[str]) -> str:
     """
@@ -321,6 +364,10 @@ def _infer_open_inference_span_kind(call_type: Optional[str]) -> str:
             "videos",
             "realtime",
             "pass_through",
+            # `passthrough` (no underscore) is what real call_types use:
+            # `allm_passthrough_route`, `llm_passthrough_route`. Without
+            # this they fell through to UNKNOWN, blanking span.kind.
+            "passthrough",
             "anthropic_messages",
             "ocr",
         )
@@ -396,6 +443,18 @@ def set_attributes(
     """
     Populates span with OpenInference-compliant LLM attributes for Arize and Phoenix tracing.
     """
+    # Coerce non-dict response objects (e.g. httpx.Response from passthrough
+    # routes) into a dict so downstream `.get()` calls don't crash. Existing
+    # dict / `.get()`-bearing objects (incl. Pydantic OpenAI Responses API
+    # models) are returned unchanged, preserving the existing test behavior.
+    response_obj_for_attrs = _coerce_response_obj_for_attrs(response_obj)
+
+    # Set span.kind defensively before anything else. If a downstream step
+    # throws, the span still has a kind so Arize can render it correctly
+    # (an LLM call instead of UNKNOWN). This is the single source of truth
+    # for span.kind — no late re-write happens below.
+    _safe_emit("early span kind", _set_early_span_kind, span, kwargs)
+
     try:
         optional_params = _sanitize_optional_params(kwargs.get("optional_params"))
         litellm_params = kwargs.get("litellm_params", {}) or {}
@@ -415,25 +474,22 @@ def set_attributes(
         metadata_tools = _extract_metadata_tools(metadata)
         optional_tools = _extract_optional_tools(optional_params)
 
-        call_type = standard_logging_payload.get("call_type")
         _set_request_attributes(
             span=span,
             kwargs=kwargs,
             standard_logging_payload=standard_logging_payload,
             optional_params=optional_params,
             litellm_params=litellm_params,
-            response_obj=response_obj,
+            response_obj=response_obj_for_attrs,
             span_attrs=SpanAttributes,
         )
 
-        span_kind = _infer_open_inference_span_kind(call_type=call_type)
+        # span.kind was already set above by `_set_early_span_kind`. We do
+        # NOT re-write it here based on tool presence: a chat completion
+        # that passes `tools=[...]` (or returns `tool_calls`) is still an
+        # LLM call per the OpenInference spec — TOOL is reserved for actual
+        # tool execution spans, not LLM calls that request tools.
         _set_tool_attributes(span, optional_tools, metadata_tools)
-        if (
-            optional_tools or metadata_tools
-        ) and span_kind != OpenInferenceSpanKindValues.TOOL.value:
-            span_kind = OpenInferenceSpanKindValues.TOOL.value
-
-        safe_set_attribute(span, SpanAttributes.OPENINFERENCE_SPAN_KIND, span_kind)
         attributes.set_messages(span, kwargs)
 
         model_params = (
@@ -443,7 +499,7 @@ def set_attributes(
         )
         _set_model_params(span, model_params, SpanAttributes)
 
-        _set_response_attributes(span=span, response_obj=response_obj)
+        _set_response_attributes(span=span, response_obj=response_obj_for_attrs)
 
     except Exception as e:
         verbose_logger.error(
@@ -452,6 +508,22 @@ def set_attributes(
         if hasattr(span, "record_exception"):
             span.record_exception(e)
 
+    # Additive emitters. Each is independently guarded so a failure can never
+    # blank the attributes set by the main try-block above. New attributes are
+    # written under new keys; existing attributes are not overwritten.
+    slp = kwargs.get("standard_logging_object")
+    _safe_emit("session/user attrs", _set_session_and_user_attrs, span, kwargs, slp)
+    _safe_emit("response cost", _set_response_cost_attr, span, slp)
+    _safe_emit(
+        "passthrough normalization",
+        _maybe_normalize_passthrough,
+        span,
+        kwargs,
+        response_obj,
+        response_obj_for_attrs,
+        slp,
+    )
+
 
 def _sanitize_optional_params(optional_params: Optional[dict]) -> dict:
     if not isinstance(optional_params, dict):
@@ -534,3 +606,529 @@ def _set_model_params(span: "Span", model_params: Optional[dict], span_attrs) ->
         user_id = model_params.get("user")
         if user_id is not None:
             safe_set_attribute(span, span_attrs.USER_ID, user_id)
+
+
+# ---------------------------------------------------------------------------
+# Additive rendering helpers (introduced to enhance Arize/Phoenix rendering
+# without changing any previously-emitted attribute keys or values).
+# ---------------------------------------------------------------------------
+
+
+def _safe_emit(label: str, fn, *args, **kwargs) -> None:
+    """Run an additive attribute emitter, swallowing any error so it cannot
+    blank attributes set elsewhere on the span. Failures are logged at debug.
+    """
+    try:
+        fn(*args, **kwargs)
+    except Exception as e:
+        verbose_logger.debug("[Arize] %s skipped: %s", label, e)
+
+
+def _set_early_span_kind(span: "Span", kwargs: dict) -> None:
+    """Defensively set OPENINFERENCE_SPAN_KIND before any other logic runs."""
+    slp = kwargs.get("standard_logging_object")
+    call_type = slp.get("call_type") if isinstance(slp, dict) else None
+    safe_set_attribute(
+        span,
+        SpanAttributes.OPENINFERENCE_SPAN_KIND,
+        _infer_open_inference_span_kind(call_type=call_type),
+    )
+
+
+def _coerce_response_obj_for_attrs(response_obj):
+    """Return a `.get`-compatible view of `response_obj` when possible.
+
+    - dicts and Pydantic models that already expose `.get` are returned
+      unchanged (preserves all current behavior, including the Responses API
+      flow which relies on Pydantic attribute access).
+    - `httpx.Response` and other text-only responses (passthrough routes)
+      are JSON-decoded so the standard extraction paths can read fields like
+      `id`, `model`, and `usage`. On failure the original object is returned
+      so behavior is no worse than today.
+    """
+    if response_obj is None or hasattr(response_obj, "get"):
+        return response_obj
+    text = getattr(response_obj, "text", None)
+    if isinstance(text, str) and text:
+        try:
+            parsed = json.loads(text)
+            if isinstance(parsed, dict):
+                return parsed
+        except Exception:
+            pass
+    return response_obj
+
+
+def _coerce_text(value) -> Optional[str]:
+    """Best-effort text extraction from a message-content value.
+
+    Returns None when no textual portion can be derived. Handles:
+      - plain strings
+      - lists of OpenAI-style content parts (`{"type": "text", "text": ...}`)
+      - lists of Anthropic-style content parts (`{"type": "text", "text": ...}`
+        or `{"type": "input_text", "text": ...}`)
+    """
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return value
+    if isinstance(value, list):
+        parts = []
+        for part in value:
+            if isinstance(part, str):
+                parts.append(part)
+            elif isinstance(part, dict):
+                text = part.get("text") or part.get("input_text")
+                if isinstance(text, str):
+                    parts.append(text)
+        if parts:
+            return "\n".join(parts)
+    return None
+
+
+def _to_plain_dict(value):
+    """Best-effort: coerce a value (Pydantic model / dict / None) to a dict.
+
+    Returns the original value when no safe conversion exists. Used to bridge
+    OpenAI Pydantic message/tool_call objects into the dict-based helpers.
+    """
+    if value is None or isinstance(value, dict):
+        return value
+    model_dump = getattr(value, "model_dump", None)
+    if callable(model_dump):
+        try:
+            return model_dump()
+        except Exception:
+            pass
+    return value
+
+
+def _get_tool_calls(message) -> Optional[list]:
+    """Return ``message.tool_calls`` only when it's a non-empty list.
+
+    Works for dicts and Pydantic message objects via ``_safe_get``.
+    """
+    tool_calls = _safe_get(message, "tool_calls")
+    return tool_calls if isinstance(tool_calls, list) and tool_calls else None
+
+
+def _normalize_tool_call(raw_tc) -> Optional[Dict[str, Any]]:
+    """Normalize a single tool_call (dict or Pydantic) into a stable shape:
+
+        {"id": str|None, "type": str, "function": {"name": str|None, "arguments": str|None}}
+
+    Arguments are coerced to a JSON string per OpenInference convention.
+    Returns ``None`` when ``raw_tc`` cannot be coerced to a dict.
+    """
+    tc = _to_plain_dict(raw_tc)
+    if not isinstance(tc, dict):
+        return None
+    function = _to_plain_dict(tc.get("function"))
+    name = function.get("name") if isinstance(function, dict) else None
+    args = function.get("arguments") if isinstance(function, dict) else None
+    if args is not None and not isinstance(args, str):
+        try:
+            args = json.dumps(args)
+        except Exception:
+            args = str(args)
+    return {
+        "id": tc.get("id"),
+        "type": tc.get("type", "function"),
+        "function": {"name": name, "arguments": args},
+    }
+
+
+def _summarize_tool_calls_for_output(tool_calls) -> str:
+    """Render a tool_calls list as a compact JSON string for OUTPUT_VALUE.
+
+    Best-effort: returns ``str(tool_calls)`` if anything unexpected happens
+    so OUTPUT_VALUE is never blanked on a malformed payload.
+    """
+    try:
+        normalized = [n for n in (_normalize_tool_call(tc) for tc in tool_calls) if n]
+        return json.dumps({"tool_calls": normalized})
+    except Exception:
+        return str(tool_calls)
+
+
+def _emit_message_tool_calls(span: "Span", prefix: str, message) -> None:
+    """Emit ``MESSAGE_TOOL_CALLS.*`` for an assistant message that requested
+    tool calls. Pure addition: only writes when ``tool_calls`` is non-empty.
+
+    Accepts dicts or Pydantic message objects (e.g. ``litellm.Message``); the
+    same applies to each tool_call entry.
+    """
+    tool_calls = _get_tool_calls(message)
+    if not tool_calls:
+        return
+    for tc_idx, raw_tc in enumerate(tool_calls):
+        tc = _normalize_tool_call(raw_tc)
+        if tc is None:
+            continue
+        tc_prefix = f"{prefix}.{MessageAttributes.MESSAGE_TOOL_CALLS}.{tc_idx}"
+        if tc["id"]:
+            safe_set_attribute(
+                span, f"{tc_prefix}.{ToolCallAttributes.TOOL_CALL_ID}", tc["id"]
+            )
+        fn = tc["function"]
+        if fn["name"]:
+            safe_set_attribute(
+                span,
+                f"{tc_prefix}.{ToolCallAttributes.TOOL_CALL_FUNCTION_NAME}",
+                fn["name"],
+            )
+        if fn["arguments"] is not None:
+            safe_set_attribute(
+                span,
+                f"{tc_prefix}.{ToolCallAttributes.TOOL_CALL_FUNCTION_ARGUMENTS_JSON}",
+                fn["arguments"],
+            )
+
+
+def _emit_input_message_extras(span: "Span", prefix: str, message: dict) -> None:
+    """Emit additive attributes for an input message:
+
+    - `MESSAGE_NAME` and `MESSAGE_TOOL_CALL_ID` (commonly set on tool-result
+      messages so traces show which tool produced which result).
+    - `MESSAGE_TOOL_CALLS.*` when an assistant message requested tools.
+    - `MESSAGE_CONTENTS.*` structured content for list-shaped content
+      (multimodal text + image parts). The plain `MESSAGE_CONTENT` write is
+      still performed by the caller, so renderers that only read the legacy
+      key continue to work.
+    """
+    if not isinstance(message, dict):
+        return
+
+    name = message.get("name")
+    if name:
+        safe_set_attribute(span, f"{prefix}.{MessageAttributes.MESSAGE_NAME}", name)
+
+    tool_call_id = message.get("tool_call_id")
+    if tool_call_id:
+        safe_set_attribute(
+            span,
+            f"{prefix}.{MessageAttributes.MESSAGE_TOOL_CALL_ID}",
+            tool_call_id,
+        )
+
+    _emit_message_tool_calls(span, prefix, message)
+
+    content = message.get("content")
+    if isinstance(content, list):
+        contents_prefix = f"{prefix}.{MessageAttributes.MESSAGE_CONTENTS}"
+        for part_idx, part in enumerate(content):
+            if not isinstance(part, dict):
+                continue
+            part_prefix = f"{contents_prefix}.{part_idx}"
+            part_type = part.get("type")
+            if part_type in ("text", "input_text"):
+                text = part.get("text")
+                if isinstance(text, str):
+                    safe_set_attribute(
+                        span,
+                        f"{part_prefix}.{MessageContentAttributes.MESSAGE_CONTENT_TYPE}",
+                        "text",
+                    )
+                    safe_set_attribute(
+                        span,
+                        f"{part_prefix}.{MessageContentAttributes.MESSAGE_CONTENT_TEXT}",
+                        text,
+                    )
+            elif part_type in ("image_url", "image", "input_image"):
+                url = None
+                image = part.get("image_url")
+                if isinstance(image, dict):
+                    url = image.get("url")
+                elif isinstance(image, str):
+                    url = image
+                if not url:
+                    # Anthropic-style source.{type=base64,media_type,data}
+                    source = part.get("source")
+                    if isinstance(source, dict) and source.get("data"):
+                        media_type = source.get("media_type", "image/jpeg")
+                        url = f"data:{media_type};base64,{source['data']}"
+                    elif isinstance(part.get("url"), str):
+                        url = part["url"]
+                if url:
+                    safe_set_attribute(
+                        span,
+                        f"{part_prefix}.{MessageContentAttributes.MESSAGE_CONTENT_TYPE}",
+                        "image",
+                    )
+                    safe_set_attribute(
+                        span,
+                        f"{part_prefix}.message_content.image.image.url",
+                        url,
+                    )
+
+
+def _set_session_and_user_attrs(
+    span: "Span", kwargs: dict, standard_logging_payload
+) -> None:
+    """Emit `SESSION_ID` / `USER_ID` / team metadata when source data exists.
+
+    `SESSION_ID` is emitted only when an explicit end-user identifier exists
+    (`metadata.user_api_key_end_user_id`). We deliberately do NOT fall back
+    to `trace_id`, because that would create a distinct "session" for every
+    single request and distort Arize's Session-grouping analytics. The
+    `trace_id` is still emitted under its own `litellm.trace_id` key so
+    spans remain filterable by trace.
+
+    USER_ID is *only* emitted when no upstream path (model_params.user or
+    optional_params.user) has already set it, to avoid overwriting an
+    existing value with a possibly-different one from API-key metadata.
+    """
+    if not isinstance(standard_logging_payload, dict):
+        return
+    metadata = standard_logging_payload.get("metadata") or {}
+    if not isinstance(metadata, dict):
+        return
+
+    session_id = metadata.get("user_api_key_end_user_id")
+    if session_id:
+        safe_set_attribute(span, SpanAttributes.SESSION_ID, str(session_id))
+
+    trace_id = standard_logging_payload.get("trace_id")
+    if trace_id:
+        safe_set_attribute(span, "litellm.trace_id", str(trace_id))
+
+    optional_params = kwargs.get("optional_params") or {}
+    model_params = standard_logging_payload.get("model_parameters") or {}
+    has_user_already = bool(
+        (isinstance(optional_params, dict) and optional_params.get("user"))
+        or (isinstance(model_params, dict) and model_params.get("user"))
+    )
+    if not has_user_already:
+        user_id = metadata.get("user_api_key_user_id")
+        if user_id:
+            safe_set_attribute(span, SpanAttributes.USER_ID, str(user_id))
+
+    team_id = metadata.get("user_api_key_team_id")
+    if team_id:
+        safe_set_attribute(span, "litellm.team_id", str(team_id))
+    team_alias = metadata.get("user_api_key_team_alias")
+    if team_alias:
+        safe_set_attribute(span, "litellm.team_alias", str(team_alias))
+    key_alias = metadata.get("user_api_key_alias")
+    if key_alias:
+        safe_set_attribute(span, "litellm.key_alias", str(key_alias))
+
+
+def _set_response_cost_attr(span: "Span", standard_logging_payload) -> None:
+    """Emit cost attributes from the StandardLoggingPayload when present.
+
+    Uses the OpenInference `llm.cost.total` key so Arize / Phoenix can
+    surface the cost in their "Total Cost" column. LiteLLM only tracks a
+    single total in `StandardLoggingPayload.response_cost`, so we cannot
+    split it into prompt/completion. We also keep the legacy
+    `llm.response.cost` key for back-compat with any consumer querying it.
+    """
+    if not isinstance(standard_logging_payload, dict):
+        return
+    cost = standard_logging_payload.get("response_cost")
+    if cost is None:
+        return
+    try:
+        cost_value = float(cost)
+    except (TypeError, ValueError):
+        return
+    safe_set_attribute(span, "llm.cost.total", cost_value)
+    safe_set_attribute(span, "llm.response.cost", cost_value)
+
+
+def _is_passthrough_call_type(call_type: Optional[str]) -> bool:
+    if not call_type:
+        return False
+    lowered = str(call_type).lower()
+    return "passthrough" in lowered or "pass_through" in lowered
+
+
+def _maybe_normalize_passthrough(
+    span: "Span",
+    kwargs: dict,
+    raw_response_obj,
+    coerced_response_obj,
+    standard_logging_payload,
+) -> None:
+    """Surface input/output text for passthrough routes (e.g. Bedrock
+    InvokeModel) so the parent span renders as more than `usage` numbers.
+
+    Only runs when `call_type` is a passthrough variant. Reads from:
+      - `kwargs["additional_args"]["complete_input_dict"]` for input
+      - the coerced response (or `kwargs["original_response"]`) for output
+
+    All emits are best-effort: if the provider shape isn't recognized the
+    helper exits silently. Existing chat/completion paths never enter this
+    helper because their call_type doesn't contain "passthrough".
+
+    TEMPORARY BRIDGE: passthrough handlers don't populate the
+    StandardLoggingPayload `messages` field today (they call
+    `transform_response(messages=[])`), so the input is only available via
+    `additional_args.complete_input_dict`. The proper fix is upstream in
+    `base_passthrough_logging_handler._create_response_logging_payload()`:
+    once that populates SLP `messages`/`response`, every callback gets
+    passthrough I/O (with central redaction) for free and this helper's
+    `complete_input_dict` fallback can be deleted. See follow-up issue.
+    """
+    call_type = (
+        standard_logging_payload.get("call_type")
+        if isinstance(standard_logging_payload, dict)
+        else None
+    )
+    if not _is_passthrough_call_type(call_type):
+        return
+
+    # Respect LiteLLM's central message-redaction contract. The normal
+    # chat/completion path is redacted by `perform_redaction` before
+    # callbacks run, but `complete_input_dict` (read below) is NOT covered by
+    # that layer — so without this gate, an operator who enabled redaction
+    # would still see raw passthrough prompts in Arize. Skip entirely when
+    # redaction is on so neither input nor output leaks through this bridge.
+    if should_redact_message_logging(kwargs):
+        return
+
+    # --- INPUT --------------------------------------------------------------
+    additional_args = kwargs.get("additional_args") or {}
+    complete_input_dict = (
+        additional_args.get("complete_input_dict")
+        if isinstance(additional_args, dict)
+        else None
+    )
+    if isinstance(complete_input_dict, dict):
+        _set_passthrough_input_attributes(span, complete_input_dict.get("messages"))
+
+    # --- OUTPUT -------------------------------------------------------------
+    parsed_response = _parse_passthrough_response(
+        raw_response_obj, coerced_response_obj, kwargs
+    )
+    if not isinstance(parsed_response, dict):
+        return
+
+    _set_passthrough_output_attributes(span, parsed_response)
+
+
+def _set_passthrough_input_attributes(span: "Span", messages) -> None:
+    """Render passthrough request messages into INPUT_VALUE + LLM_INPUT_MESSAGES."""
+    if not (isinstance(messages, list) and messages):
+        return
+    # Set INPUT_VALUE from the last user message text if discoverable.
+    last_text = None
+    for msg in reversed(messages):
+        if isinstance(msg, dict):
+            last_text = _coerce_text(msg.get("content"))
+            if last_text:
+                break
+    if last_text:
+        safe_set_attribute(span, SpanAttributes.INPUT_VALUE, last_text)
+    # Mirror messages into LLM_INPUT_MESSAGES so the input pane renders.
+    for idx, msg in enumerate(messages):
+        if not isinstance(msg, dict):
+            continue
+        prefix = f"{SpanAttributes.LLM_INPUT_MESSAGES}.{idx}"
+        role = msg.get("role")
+        if role:
+            safe_set_attribute(
+                span,
+                f"{prefix}.{MessageAttributes.MESSAGE_ROLE}",
+                role,
+            )
+        text = _coerce_text(msg.get("content"))
+        if text is not None:
+            safe_set_attribute(
+                span,
+                f"{prefix}.{MessageAttributes.MESSAGE_CONTENT}",
+                text,
+            )
+
+
+def _set_passthrough_output_attributes(span: "Span", parsed_response: dict) -> None:
+    """Render passthrough response into OUTPUT_VALUE + LLM_OUTPUT_MESSAGES."""
+    # Anthropic / Bedrock-Anthropic: `content` is a list of typed parts.
+    content_list = parsed_response.get("content")
+    if isinstance(content_list, list) and content_list:
+        texts = []
+        for part in content_list:
+            if isinstance(part, dict) and isinstance(part.get("text"), str):
+                texts.append(part["text"])
+        joined = "\n\n".join(t for t in texts if t)
+        if joined:
+            safe_set_attribute(span, SpanAttributes.OUTPUT_VALUE, joined)
+            prefix = f"{SpanAttributes.LLM_OUTPUT_MESSAGES}.0"
+            safe_set_attribute(
+                span,
+                f"{prefix}.{MessageAttributes.MESSAGE_ROLE}",
+                parsed_response.get("role", "assistant"),
+            )
+            safe_set_attribute(
+                span,
+                f"{prefix}.{MessageAttributes.MESSAGE_CONTENT}",
+                joined,
+            )
+
+    # OpenAI-style passthrough: `choices[0].message.content`
+    choices = parsed_response.get("choices")
+    if isinstance(choices, list) and choices:
+        first = choices[0]
+        if isinstance(first, dict):
+            msg = first.get("message")
+            if isinstance(msg, dict):
+                text = _coerce_text(msg.get("content"))
+                if text:
+                    safe_set_attribute(span, SpanAttributes.OUTPUT_VALUE, text)
+                    prefix = f"{SpanAttributes.LLM_OUTPUT_MESSAGES}.0"
+                    safe_set_attribute(
+                        span,
+                        f"{prefix}.{MessageAttributes.MESSAGE_ROLE}",
+                        msg.get("role", "assistant"),
+                    )
+                    safe_set_attribute(
+                        span,
+                        f"{prefix}.{MessageAttributes.MESSAGE_CONTENT}",
+                        text,
+                    )
+
+
+def _parse_passthrough_response(raw_response_obj, coerced_response_obj, kwargs):
+    """Return a dict view of the provider response for passthrough routes."""
+    # Prefer the coerced view (already JSON-parsed for httpx.Response).
+    candidates = []
+    if isinstance(coerced_response_obj, dict):
+        candidates.append(coerced_response_obj)
+    if (
+        isinstance(raw_response_obj, dict)
+        and raw_response_obj is not coerced_response_obj
+    ):
+        candidates.append(raw_response_obj)
+
+    for candidate in candidates:
+        # StandardPassThroughResponseObject wrapper: {"response": "..."}.
+        if (
+            "response" in candidate
+            and "content" not in candidate
+            and "choices" not in candidate
+        ):
+            inner = candidate.get("response")
+            if isinstance(inner, str):
+                try:
+                    parsed = json.loads(inner)
+                    if isinstance(parsed, dict):
+                        return parsed
+                except Exception:
+                    continue
+            if isinstance(inner, dict):
+                return inner
+        else:
+            return candidate
+
+    # Fallback: kwargs["original_response"] from the OTel base path.
+    original = kwargs.get("original_response") if isinstance(kwargs, dict) else None
+    if isinstance(original, dict):
+        return original
+    if isinstance(original, str):
+        try:
+            parsed = json.loads(original)
+            if isinstance(parsed, dict):
+                return parsed
+        except Exception:
+            return None
+    return None

litellm/integrations/custom_guardrail.py

@@ -662,6 +662,16 @@ class CustomGuardrail(CustomLogger):
             request_data["metadata"] = {}
             _append_guardrail_info(request_data["metadata"])
 
+        # Emit the otel guardrail span here, where every guardrail execution lands,
+        # rather than relying on a post-call hook that does not fire on every path
+        # (e.g. a pass-through request that passes its guardrails).
+        try:
+            from litellm.integrations.otel.logger import emit_guardrail_span
+
+            emit_guardrail_span(slg)
+        except Exception:
+            pass
+
     async def apply_guardrail(
         self,
         inputs: GenericGuardrailAPIInputs,

litellm/integrations/datadog/datadog.py

@@ -41,6 +41,7 @@ from litellm.integrations.datadog.datadog_handler import (
 )
 from litellm.litellm_core_utils.dd_tracing import tracer
 from litellm.llms.custom_httpx.http_handler import (
+    MaskedHTTPStatusError,
     _get_httpx_client,
     get_async_httpx_client,
     httpxSpecialProvider,
@@ -68,6 +69,22 @@ DD_LOGGED_SUCCESS_SERVICE_TYPES = [
 ]
 
 
+def _resolve_dd_batch_size() -> int:
+    raw = os.getenv("DD_BATCH_SIZE")
+    if raw is None:
+        return DD_MAX_BATCH_SIZE
+    try:
+        value = int(raw)
+    except ValueError:
+        verbose_logger.warning(
+            "Datadog: ignoring invalid DD_BATCH_SIZE=%r, using %s",
+            raw,
+            DD_MAX_BATCH_SIZE,
+        )
+        return DD_MAX_BATCH_SIZE
+    return max(1, min(value, DD_MAX_BATCH_SIZE))
+
+
 class DataDogLogger(
     CustomBatchLogger,
     AdditionalLoggingUtils,
@@ -128,7 +145,9 @@ class DataDogLogger(
             asyncio.create_task(self.periodic_flush())
             self.flush_lock = asyncio.Lock()
             super().__init__(
-                **kwargs, flush_lock=self.flush_lock, batch_size=DD_MAX_BATCH_SIZE
+                **kwargs,
+                flush_lock=self.flush_lock,
+                batch_size=_resolve_dd_batch_size(),
             )
         except Exception as e:
             verbose_logger.exception(
@@ -339,28 +358,14 @@ class DataDogLogger(
                     "[DATADOG MOCK] Mock mode enabled - API calls will be intercepted"
                 )
 
-            response = await self.async_send_compressed_data(batch_to_send)
-            if response.status_code == 413:
-                verbose_logger.exception(DD_ERRORS.DATADOG_413_ERROR.value)
-                self.log_queue = batch_to_send + self.log_queue
-                return
-
-            response.raise_for_status()
-            if response.status_code != 202:
-                raise Exception(
-                    f"Response from datadog API status_code: {response.status_code}, text: {response.text}"
-                )
+            undelivered = await self._send_with_413_split(batch_to_send)
+            if undelivered:
+                self.log_queue = undelivered + self.log_queue
 
             if self.is_mock_mode:
                 verbose_logger.debug(
                     f"[DATADOG MOCK] Batch of {len(batch_to_send)} events successfully mocked"
                 )
-            else:
-                verbose_logger.debug(
-                    "Datadog: Response from datadog API status_code: %s, text: %s",
-                    response.status_code,
-                    response.text,
-                )
 
         except Exception as e:
             self.log_queue = batch_to_send + self.log_queue
@@ -368,6 +373,62 @@ class DataDogLogger(
                 f"Datadog Error sending batch API - {str(e)}\n{traceback.format_exc()}"
             )
 
+    async def _send_with_413_split(self, batch: List) -> List:
+        """
+        Send a batch, halving any sub-batch that 413s (payload too large) and retrying the
+        halves, since Datadog enforces a 5MB uncompressed limit per request.
+
+        A 413 surfaces as a raised MaskedHTTPStatusError (httpx raise_for_status), not a
+        returned response, so both paths are handled. A lone event that still 413s is
+        dropped to avoid wedging the queue on an undeliverable payload. Returns the events
+        that could not be delivered because of a non-413 (transient) error, so the caller
+        re-queues only those and never the events already accepted by Datadog.
+        """
+        pending: List[List] = [batch]
+        while pending:
+            chunk = pending.pop()
+            if not chunk:
+                continue
+            try:
+                response = await self.async_send_compressed_data(chunk)
+            except Exception as e:
+                if isinstance(e, MaskedHTTPStatusError) and e.status_code == 413:
+                    response = e.response
+                else:
+                    verbose_logger.exception(
+                        f"Datadog Error sending batch API - {str(e)}"
+                    )
+                    return self._undelivered(chunk, pending)
+
+            if response.status_code == 413:
+                if len(chunk) == 1:
+                    verbose_logger.error(DD_ERRORS.DATADOG_413_ERROR.value)
+                    continue
+                mid = len(chunk) // 2
+                pending.append(chunk[mid:])
+                pending.append(chunk[:mid])
+                continue
+
+            if response.status_code != 202:
+                verbose_logger.error(
+                    "Datadog: unexpected response status_code=%s, text=%s",
+                    response.status_code,
+                    response.text,
+                )
+                return self._undelivered(chunk, pending)
+
+            verbose_logger.debug(
+                "Datadog: delivered %s events, status_code=%s, text=%s",
+                len(chunk),
+                response.status_code,
+                response.text,
+            )
+        return []
+
+    @staticmethod
+    def _undelivered(chunk: List, pending: List[List]) -> List:
+        return chunk + [event for remaining in reversed(pending) for event in remaining]
+
     async def flush_queue(self):
         if self.flush_lock is None:
             return

litellm/integrations/focus/transformer.py

@@ -95,7 +95,9 @@ class FocusTransformer:
             pl.lit("Usage-Based").alias("ChargeFrequency"),
             fmt(pl.col("ChargePeriodEnd")).alias("ChargePeriodEnd"),
             fmt(pl.col("ChargePeriodStart")).alias("ChargePeriodStart"),
-            dec(pl.lit(1.0)).alias("ConsumedQuantity"),
+            dec(
+                pl.col("api_requests").cast(pl.Int64).cast(pl.Float64).fill_null(0.0)
+            ).alias("ConsumedQuantity"),
             pl.lit("Requests").alias("ConsumedUnit"),
             dec(pl.col("spend").fill_null(0.0)).alias("ContractedCost"),
             none_str.alias("ContractedUnitPrice"),
@@ -107,7 +109,9 @@ class FocusTransformer:
             none_str.alias("AvailabilityZone"),
             pl.lit("USD").alias("PricingCurrency"),
             none_str.alias("PricingCategory"),
-            dec(pl.lit(1.0)).alias("PricingQuantity"),
+            dec(
+                pl.col("api_requests").cast(pl.Int64).cast(pl.Float64).fill_null(0.0)
+            ).alias("PricingQuantity"),
             none_dec.alias("PricingCurrencyContractedUnitPrice"),
             dec(pl.col("spend").fill_null(0.0)).alias("PricingCurrencyEffectiveCost"),
             none_dec.alias("PricingCurrencyListUnitPrice"),

litellm/integrations/opentelemetry.py

@@ -64,6 +64,9 @@ HTTP_RESPONSE_STATUS_CODE_ATTRIBUTE = "http.response.status_code"
 HTTP_ROUTE_ATTRIBUTE = "http.route"
 URL_PATH_ATTRIBUTE = "url.path"
 PREPROCESSING_DURATION_MS_ATTRIBUTE = "litellm.preprocessing.duration_ms"
+TEAM_METADATA_ATTRIBUTE = "litellm.team.metadata"
+MODEL_GROUP_ATTRIBUTE = "litellm.model_group"
+PROVIDER_MODEL_ATTRIBUTE = "litellm.provider.model"
 # Remove the hardcoded LITELLM_RESOURCE dictionary - we'll create it properly later
 RAW_REQUEST_SPAN_NAME = "raw_gen_ai_request"
 LITELLM_REQUEST_SPAN_NAME = "litellm_request"
@@ -80,6 +83,19 @@ _VALID_CAPTURE_MODES = {
 }
 
 
+def _normalize_team_metadata_keys(value: Any) -> List[str]:
+    """Coerce a team-metadata allowlist from a list or comma-separated string.
+
+    config.yaml passes a YAML list; an env var passes a comma-separated string.
+    Both collapse to a list of stripped, non-empty keys.
+    """
+    if value is None:
+        return []
+    if isinstance(value, str):
+        return [item.strip() for item in value.split(",") if item.strip()]
+    return [str(item).strip() for item in value if str(item).strip()]
+
+
 @dataclass
 class OpenTelemetryConfig:
     exporter: Union[str, SpanExporter] = "console"
@@ -97,6 +113,10 @@ class OpenTelemetryConfig:
     # One of NO_CONTENT, SPAN_ONLY, EVENT_ONLY, SPAN_AND_EVENT (or "true" as legacy alias).
     capture_message_content: Optional[str] = None
     semconv_stability_opt_in: Set[OTELSemconvCategory] = field(default_factory=set)
+    # Sub-keys of the team's free-form metadata stamped onto the inference span
+    # under ``litellm.team.metadata``. Empty by default so none of a team's
+    # metadata leaves the process until explicitly allowlisted.
+    baggage_team_metadata_keys: List[str] = field(default_factory=list)
 
     def __post_init__(self) -> None:
         # If endpoint is specified but exporter is still the default "console",
@@ -127,6 +147,11 @@ class OpenTelemetryConfig:
         self.semconv_stability_opt_in |= parse_semconv_opt_in(
             os.getenv(OTEL_SEMCONV_STABILITY_OPT_IN_ENV)
         )
+        self.baggage_team_metadata_keys = _normalize_team_metadata_keys(
+            self.baggage_team_metadata_keys
+        ) or _normalize_team_metadata_keys(
+            os.getenv("LITELLM_OTEL_BAGGAGE_TEAM_METADATA_KEYS")
+        )
 
     @classmethod
     def from_env(cls):
@@ -185,8 +210,13 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         meter_provider: Optional[Any] = None,
         **kwargs,
     ):
+        team_metadata_keys_override = kwargs.pop("baggage_team_metadata_keys", None)
         if config is None:
             config = OpenTelemetryConfig.from_env()
+        if team_metadata_keys_override is not None:
+            config.baggage_team_metadata_keys = _normalize_team_metadata_keys(
+                team_metadata_keys_override
+            )
 
         self.config = config
         self.callback_name = callback_name
@@ -982,6 +1012,13 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         litellm_params = kwargs.get("litellm_params", {}) or {}
         _metadata = litellm_params.get("metadata", {}) or {}
         proxy_span = _metadata.get("litellm_parent_otel_span", None)
+
+        # Fallback: check litellm_metadata (used by /v1/messages and other
+        # LITELLM_METADATA_ROUTES).
+        if proxy_span is None:
+            _litellm_metadata = litellm_params.get("litellm_metadata", {}) or {}
+            proxy_span = _litellm_metadata.get("litellm_parent_otel_span", None)
+
         if (
             proxy_span is not None
             and getattr(proxy_span, "name", None) == LITELLM_PROXY_REQUEST_SPAN_NAME
@@ -1213,6 +1250,74 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         ):
             self._set_team_attributes_from_kwargs(proxy_span, kwargs)
 
+    def _set_inference_identity_attributes(
+        self,
+        span: Span,
+        standard_logging_payload: StandardLoggingPayload,
+        litellm_params: dict,
+    ) -> None:
+        """Stamp request-identity attributes onto an inference span so every
+        LLM-call span is filterable by the route it came in on, the team's
+        metadata, and both the user-facing (model_group alias) and the
+        dispatched (provider) model names. Empty/absent values are skipped.
+        """
+        metadata = standard_logging_payload.get("metadata") or {}
+
+        http_route = metadata.get("user_api_key_request_route")
+        if http_route:
+            self.safe_set_attribute(
+                span=span, key=HTTP_ROUTE_ATTRIBUTE, value=http_route
+            )
+
+        # ``user_api_key_team_metadata`` is dropped from the standard logging
+        # payload metadata, so read it from the raw request metadata in kwargs.
+        # ``metadata`` and ``litellm_metadata`` are alternate names for the same
+        # full metadata dict (the name varies by endpoint), so first-truthy wins.
+        raw_metadata = (
+            litellm_params.get("metadata")
+            or litellm_params.get("litellm_metadata")
+            or {}
+        )
+        team_metadata = self._team_metadata_json(
+            raw_metadata.get("user_api_key_team_metadata"),
+            self.config.baggage_team_metadata_keys,
+        )
+        if team_metadata:
+            self.safe_set_attribute(
+                span=span, key=TEAM_METADATA_ATTRIBUTE, value=team_metadata
+            )
+
+        model_group = standard_logging_payload.get("model_group")
+        if model_group:
+            self.safe_set_attribute(
+                span=span, key=MODEL_GROUP_ATTRIBUTE, value=model_group
+            )
+
+        hidden_params = standard_logging_payload.get("hidden_params") or {}
+        provider_model = hidden_params.get(
+            "litellm_model_name"
+        ) or standard_logging_payload.get("model")
+        if provider_model:
+            self.safe_set_attribute(
+                span=span, key=PROVIDER_MODEL_ATTRIBUTE, value=provider_model
+            )
+
+    @staticmethod
+    def _team_metadata_json(value: Any, allowed_keys: List[str]) -> Optional[str]:
+        """JSON-serialize only the allowlisted sub-keys of a team's metadata.
+
+        Returns ``None`` when nothing is allowlisted or no allowlisted key is
+        present, so the empty case is dropped rather than stamping a useless
+        ``"{}"`` (and so a team's metadata never leaves the process until an
+        operator opts each sub-key in via ``baggage_team_metadata_keys``).
+        """
+        if not isinstance(value, dict) or not value or not allowed_keys:
+            return None
+        filtered = {key: value[key] for key in allowed_keys if key in value}
+        if not filtered:
+            return None
+        return safe_dumps(filtered)
+
     def _record_metrics(self, kwargs, response_obj, start_time, end_time):
         duration_s = (end_time - start_time).total_seconds()
         params = kwargs.get("litellm_params") or {}
@@ -2023,6 +2128,12 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
                     key="hidden_params",
                     value=safe_dumps(hidden_params),
                 )
+
+            self._set_inference_identity_attributes(
+                span=span,
+                standard_logging_payload=standard_logging_payload,
+                litellm_params=litellm_params,
+            )
             # Cost breakdown tracking
             cost_breakdown: Optional[CostBreakdown] = standard_logging_payload.get(
                 "cost_breakdown"
@@ -2564,6 +2675,10 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             )
 
     def _to_ns(self, dt):
+        if dt is None:
+            return int(datetime.now().timestamp() * 1e9)
+        if isinstance(dt, (int, float)):
+            return int(dt * 1e9)
         return int(dt.timestamp() * 1e9)
 
     def _get_span_name(self, kwargs):
@@ -2610,6 +2725,13 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
         _metadata = litellm_params.get("metadata", {}) or {}
         parent_otel_span = _metadata.get("litellm_parent_otel_span", None)
 
+        # Fallback: check litellm_metadata (used by /v1/messages and other
+        # LITELLM_METADATA_ROUTES that store proxy-internal metadata
+        # separately from the provider's native "metadata" field).
+        if parent_otel_span is None:
+            _litellm_metadata = litellm_params.get("litellm_metadata", {}) or {}
+            parent_otel_span = _litellm_metadata.get("litellm_parent_otel_span", None)
+
         # Priority 1: Explicit parent span from metadata
         if parent_otel_span is not None:
             verbose_logger.debug(
@@ -3183,6 +3305,32 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
             value=int(status_code),
         )
 
+    def record_error_attributes_on_span(
+        self,
+        span: Optional[Span],
+        exception: Optional[Exception],
+        status_code: int,
+    ) -> None:
+        """Stamp structured ``error.*`` attributes on the SERVER span from the
+        exception returned to the client, with ``error.code`` pinned to the real
+        response status. Idempotent (overwrites); emits no exception event."""
+        if span is None or exception is None:
+            return
+        from litellm.litellm_core_utils.litellm_logging import (
+            StandardLoggingPayloadSetup,
+        )
+
+        error_information = StandardLoggingPayloadSetup.get_error_information(
+            original_exception=exception
+        )
+        error_information["error_code"] = str(status_code)
+        self._record_exception_on_span(
+            span=span,
+            kwargs={
+                "standard_logging_object": {"error_information": error_information}
+            },
+        )
+
     def set_preprocessing_duration_attribute(
         self, span: Optional[Span], container: Any
     ) -> None:

litellm/integrations/opik/opik_payload_builder/extractors.py

@@ -39,20 +39,32 @@ def extract_opik_metadata(
     standard_logging_metadata: Dict[str, Any],
 ) -> Dict[str, Any]:
     """
-    Extract and merge Opik metadata from request and requester.
+    Merge Opik metadata from three sources in increasing priority order:
+
+      1. user_api_key_auth_metadata– lowest priority (operator-level defaults)
+      2. litellm_metadata (request)– overrides auth-key defaults
+      3. requester_metadata – highest priority (e.g. proxy header overrides)
 
     Args:
-        litellm_metadata: Metadata from litellm_params
-        standard_logging_metadata: Metadata from standard_logging_object
+        litellm_metadata: Metadata from litellm_params.mak
+        standard_logging_metadata: Metadata from standard_logging_object.
 
     Returns:
-        Merged Opik metadata dictionary
+        Merged Opik metadata dictionary.
     """
-    opik_meta = litellm_metadata.get("opik", {}).copy()
+    # Start with auth-key defaults (lowest priority).
+    auth_meta = standard_logging_metadata.get("user_api_key_auth_metadata") or {}
+    opik_meta = (auth_meta.get("opik") or {}).copy()
 
+    # Request-level values override auth-key defaults.
+    request_opik = litellm_metadata.get("opik") or {}
+    opik_meta.update(request_opik)
+
+    # Requester-level values win over everything else.
     requester_metadata = standard_logging_metadata.get("requester_metadata", {}) or {}
     requester_opik = requester_metadata.get("opik", {}) or {}
-    opik_meta.update(requester_opik)
+    if requester_opik:
+        opik_meta.update(requester_opik)
 
     _logging.verbose_logger.debug(
         f"litellm_opik_metadata - {json.dumps(opik_meta, default=str)}"

litellm/integrations/otel/README.md

@@ -0,0 +1,261 @@
+# OpenTelemetry instrumentation
+
+This package produces OpenTelemetry traces for LiteLLM. It is enabled by the
+`LITELLM_OTEL_V2` environment variable (`is_otel_v2_enabled()` in
+[`config.py`](./model/config.py)); when unset, nothing in this package runs.
+
+## What gets traced
+
+A traced proxy request produces one trace with two kinds of spans:
+
+```
+SERVER span  "POST /v1/chat/completions"        ← FastAPI instrumentation
+├── INTERNAL span  "auth /v1/chat/completions"   ← auth phase     ┐
+│   ├── CLIENT span  "postgres get_key_object"    ← datastore call │
+│   └── CLIENT span  "postgres get_team_membership"                │
+├── INTERNAL span  "execute_guardrail …"         ← guardrail       │ this package
+├── CLIENT span    "chat gpt-4o"                  ← LLM call        │
+└── CLIENT span    "batch_write_to_db …"          ← spend write    ┘
+```
+
+The gen-ai spans are siblings under the server span. In particular the guardrail
+span is a sibling of the LLM call, not a child of it: pre/during/post-call
+guardrail hooks are part of the request lifecycle (a pre-call guardrail runs
+before the LLM call even starts), so they belong directly under the server span,
+alongside the LLM call.
+
+Request-level spans (LLM call, guardrail) parent to the server span via an
+**explicit anchor** — `context.set_request_root_span` captures the server span
+once at request entry, and `resolve_request_span_context` reads it — rather than
+to whatever span is momentarily active. Ambient-only parenting was wrong at two
+boundaries: inside the live `auth` phase span the active span is `auth` (so the
+span would nest under auth), and a pass-through request closes its span from a
+detached `asyncio.create_task` where the server span is no longer active (so the
+span orphaned into its own trace). The anchor — a contextvar inherited by those
+child tasks — gives a stable parent in both cases. DB/service spans keep ambient
+parenting so an auth DB lookup still nests under `auth`.
+
+**Which service calls become spans (`spans.span_role_for_service`).** LiteLLM's
+service-logging layer instruments many internal functions, but only some are
+traceable units of work:
+
+- **`DB_CALL` (CLIENT)** — outbound datastore calls (redis, postgres,
+  `batch_write_to_db`), carrying `db.system.name` / `db.operation.name` semconv.
+- **`SERVICE` (INTERNAL)** — genuine internal work worth a span (background
+  budget/reset jobs, pod-lock manager).
+- **metrics-only (no span)** — `self` (the `track_llm_api_timing` wrapper, which
+  duplicates the LLM-call span), `router` (duplicates the request), and
+  `proxy_pre_call` (a guardrail's real span is `execute_guardrail …`). These
+  still feed Prometheus/Datadog through their own hooks; they just never enter
+  the trace. `auth` is also excluded here because it gets a **live phase span**
+  instead (see below).
+
+Spans are named `"{service} {call_type}"` (e.g. `"redis set"`) so repeated calls
+to one service stay distinguishable. Like every other span they parent to the
+**ambient** context, falling back to the threaded `litellm_parent_otel_span` only
+when ambient has no live span; a background job with neither starts its own root
+trace. Caller-supplied `event_metadata` is **sanitized** before it reaches a span
+(primitives only, no live objects, no secrets/headers, bounded) — see
+`payloads.sanitize_event_metadata`.
+
+**Live phase spans.** `auth` is wrapped in a real, active span
+(`logger.phase_span`) for the duration of authentication, so the DB lookups it
+triggers nest **under** it instead of flattening onto the server span. Identity
+Baggage (team/key/user) is seeded once the key resolves, so every post-auth span
+inherits it; auth-internal DB lookups that run before the key is known stay
+unlabeled, which is correct.
+
+**Status.** On success a span's status is left `UNSET` (the semconv default,
+matching the FastAPI server span); only a genuine error sets `ERROR`.
+
+- **Server spans** (one per HTTP route) are created by the
+  `opentelemetry-instrumentation-fastapi` package. It stamps `http.*` attributes
+  and extracts inbound `traceparent` headers. This package does **not** create
+  or modify server spans — request routes never touch spans.
+- **Gen-AI spans** (LLM calls, guardrails, internal service calls) are created
+  by this package from LiteLLM's logging callbacks. Request-level spans parent to
+  the server span via the captured anchor; DB/service spans parent to the active
+  span (ambient) so they nest under the request phase that triggered them.
+
+Both kinds share a single `TracerProvider`, so they belong to the same trace
+and export through the same configured exporters. FastAPI middleware can only be
+added before the app starts serving, so the app is instrumented at
+import time **without** a provider — it binds to the OTel global
+`ProxyTracerProvider`. Once config (and the callbacks) is loaded, the proxy
+publishes the chosen logger's `TracerProvider` as the global via
+`trace.set_tracer_provider(...)`, and the server spans delegate to it. When a
+preset callback (`arize`, `langfuse_otel`, …) is configured, its provider
+becomes the global, so server spans export to that backend too.
+
+## How a request flows
+
+1. **App creation** (`proxy_server` import): when the gate is on,
+   `mount.instrument_fastapi_app(app)` calls `FastAPIInstrumentor.instrument_app`
+   with no provider (the middleware stack is frozen once the app serves, so this
+   can't wait for startup). It binds to the OTel global `ProxyTracerProvider`. Noisy
+   non-LLM routes are excluded by default (`mount._DEFAULT_EXCLUDED_ROUTES`): health
+   checks (`/health*`), the Prometheus scrape (`/metrics`), and static UI/docs assets
+   (`/litellm-asset-prefix`, `/_next`, `/ui`, `/swagger`, `/docs`, `/redoc`,
+   `/openapi.json`, favicons, `/.well-known`) — so load-balancer polling, metric
+   scrapes, and asset fetches don't flood traces. Entries are substring-matched, so
+   `/metrics` also drops the `/model/metrics` admin-analytics spans. Set
+   `OTEL_PYTHON_FASTAPI_EXCLUDED_URLS` to override the whole set (e.g. `""` to trace
+   everything, or your own comma-separated path list).
+2. **Startup** (`proxy_server.proxy_startup_event`): after the config (and
+   callbacks) is loaded, the already-registered preset `OpenTelemetryV2` logger
+   is reused — or a generic one reading `OTEL_*` envs is built when no preset is
+   configured — and its `TracerProvider` is published as the OTel global with
+   `trace.set_tracer_provider(...)`. The proxy tracer then delegates to it, so
+   server spans and gen-ai spans share one provider and the same trace.
+3. **Request**: the FastAPI instrumentation starts the server span and makes it
+   the active context for the request task. The proxy's first call into the V2
+   logger (`create_litellm_proxy_request_started_span`, at the auth boundary)
+   **captures it as the request anchor** (`set_request_root_span`), so every later
+   request-level span has a stable explicit parent regardless of what is active
+   when it emits.
+4. **LLM call span (born at the boundary)**: `OpenTelemetryV2.log_pre_api_call`
+   runs synchronously in the request task, just before the upstream call, and
+   **opens** the LLM-call span there, parented to the anchored server span
+   (`resolve_request_span_context`). The open span is held in a bounded cache keyed
+   by `litellm_call_id` (a primitive the callback kwargs carry at both `pre_call`
+   and close), so no live `Span` ever travels through a `litellm_params` metadata
+   dict. For the boundary hook to fire at all, the logger is registered into
+   `litellm.input_callback` — the list `Logging.pre_call` iterates. The async
+   success/failure callback later
+   **closes** it: it builds an `LLMCallSpanData` from the typed
+   `standard_logging_object` (token usage and cost are computed only by then),
+   stamps the attributes, sets status, and ends the span. The sync callback is a
+   no-op (closing is async-only). When `pre_call` runs off the request task — a
+   sync-only provider driven through a thread pool, where contextvars (and so the
+   anchor) don't follow — no parent is visible there, so creation is **deferred**
+   to the async callback, whose worker context was copied from the request task at
+   enqueue and so still carries the anchor. **Pass-through** endpoints call
+   `logging_obj.pre_call` in the request task too, then close from a detached
+   `asyncio.create_task`; the anchor (not the by-then-inactive server span) keeps
+   their LLM-call span in the request's trace. `pre_call` is litellm's generic
+   "log the attempt" hook, so it also fires for synthetic proxy-gate error logs
+   (auth/rate-limit rejections); those carry `LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL`
+   and are skipped, so a request rejected before reaching a provider never produces
+   a phantom CLIENT span.
+5. **Guardrails / services**: the post-call and service hooks emit guardrail and
+   service spans the same way — typed data → engine → span. Service spans
+   (Redis/Postgres) are dispatched by `litellm/_service_logger.py`, which
+   recognizes the V2 `OpenTelemetryV2` logger (a plain `CustomLogger`, not a
+   subclass of the legacy `OpenTelemetry`). It hands every service call to the
+   logger — including calls with no parent span — and the V2 adapter decides the
+   role (`DB_CALL` vs `SERVICE`), the parent (ambient → threaded → root), and
+   whether the call is a traceable operation or a metrics-only ping. Guardrail
+   span data is built from the typed, provider-agnostic
+   `StandardLoggingGuardrailInformation` — no single provider's field shape is
+   assumed.
+6. **Export**: each span ends and is handed to the provider's span processors,
+   which export to the configured backends (OTLP, console, in-memory, …).
+
+## Components
+
+### Sources of truth (`model/`, no OpenTelemetry import)
+
+These define the shape of a span without depending on the OTel SDK, so they can
+be imported anywhere. They live in [`model/`](./model) and form a closed set —
+nothing here imports outside it:
+
+- [`semconv.py`](./model/semconv.py) — attribute-key constants (`gen_ai.*`, `http.*`,
+  `litellm.*`), the GenAI operation/provider enums, and the functions that map
+  LiteLLM provider/call-type strings onto convention values.
+- [`spans.py`](./model/spans.py) — the span registry: every span role, its OTel span
+  kind, its place in the hierarchy, and its name builder.
+- [`payloads.py`](./model/payloads.py) — frozen dataclasses (`LLMCallSpanData`,
+  `GuardrailSpanData`, `ServiceSpanData`, …) built from heterogeneous logging
+  payloads via `from_*` classmethods.
+- [`config.py`](./model/config.py) — `OpenTelemetryV2Config`, a pydantic-settings
+  model that reads `OTEL_*` / `LITELLM_OTEL_*` env vars, plus the feature gate.
+  `capture_span_content` gates whether prompt/response bodies may be written as
+  span attributes; it defaults **off** (`no_content`). The Baggage allowlists are
+  configurable, not hard-coded: set `LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS` /
+  `LITELLM_OTEL_BAGGAGE_METADATA_KEYS` /
+  `LITELLM_OTEL_BAGGAGE_TEAM_METADATA_KEYS` (comma-separated) as env vars, or
+  `baggage_promoted_keys` / `baggage_metadata_keys` /
+  `baggage_team_metadata_keys` (YAML lists) under `callback_settings.otel` in
+  `config.yaml` — the latter reach the config through the logger's constructor
+  kwargs. `baggage_team_metadata_keys` is empty by default, so none of a team's
+  free-form metadata is promoted until each sub-key is explicitly allowlisted.
+- [`baggage.py`](./model/baggage.py) — the single definition of which request-identity
+  values are promoted into Baggage (so child spans inherit them) and under which
+  attribute keys.
+- [`utils.py`](./model/utils.py) — value coercion, JSON serialization, and
+  extractor-table application, shared across the package.
+
+### Engine
+
+- [`emitter.py`](./emitter.py) — `SpanEmitter.emit(role, data)`: dedupe → start
+  the span → run the mapper chain to stamp attributes → set status → end. It
+  owns no attribute keys. The dedupe set (which coalesces the sync+async firing
+  of one request) is a bounded LRU so it can't grow without limit.
+- [`mappers/`](./mappers) — each mapper turns typed span data into a flat
+  `{attribute key: value}` dict. They compose: listing several mapper names in
+  the config layers multiple attribute vocabularies onto the same span.
+  - `genai` — the canonical OpenTelemetry GenAI vocabulary, always present.
+  - `legacy` — an additional vocabulary using the older semconv-ai / Traceloop
+    attribute key names, for backends that read those.
+  - `openinference`, `langfuse`, `weave`, `langtrace` — vendor vocabularies.
+  - `resolve_mappers(names)` turns config names into mapper instances.
+
+### Plumbing (`plumbing/`)
+
+The OTel-SDK wiring. Everything here imports only `model/` and each other; it
+lives in [`plumbing/`](./plumbing):
+
+- [`providers.py`](./plumbing/providers.py) — builds the `TracerProvider`, its exporters
+  (from `ExporterSpec`s), and the span processor that copies allowlisted Baggage
+  entries onto every span. `register_exporter_factory(kind, factory)` lets a
+  preset contribute a custom exporter `kind` (e.g. one that fetches an auth
+  token lazily) without coupling this module to any vendor.
+- [`context.py`](./plumbing/context.py) — trace-context and Baggage read/write helpers.
+- [`routing.py`](./plumbing/routing.py) — `TenantTracerCache`: when a request carries
+  team/key-scoped vendor credentials, route its spans through a credential-keyed
+  `TracerProvider` so one logger serves many tenants. The cache is a bounded LRU
+  that flushes + shuts down evicted providers, since the key derives from
+  request-supplied credentials and must not grow (or leak threads) without limit.
+- [`metrics.py`](./plumbing/metrics.py) — GenAI client metric instruments.
+
+### Adapter
+
+- [`logger.py`](./logger.py) — `OpenTelemetryV2`, a `CustomLogger` that
+  translates LiteLLM's logging callbacks into typed span data and hands them to
+  the engine. The LLM-call span is opened at the `log_pre_api_call` boundary
+  (parented to the live server span via ambient context) and closed at the async
+  success/failure callback; the open span is held in a bounded cache keyed by
+  `litellm_call_id`, never threaded through a metadata dict. The logger registers
+  itself into `litellm.input_callback` so `Logging.pre_call` fires the boundary
+  hook.
+- [`mount.py`](./mount.py) — `instrument_fastapi_app(app)`, the single call site
+  that attaches `opentelemetry-instrumentation-fastapi` for SERVER spans. It owns
+  the health-check exclusion default (`OTEL_PYTHON_FASTAPI_EXCLUDED_URLS`) and the
+  passthrough span-naming hook (`PASSTHROUGH_PREFIXES`) so `proxy_server` carries
+  no OTel detail. A safe no-op when the gate is off or the instrumentation package
+  is absent; must be called at app-creation time (the middleware stack freezes
+  once the app serves).
+
+### Presets
+
+- [`presets/`](./presets) — each preset reads one integration's env vars and
+  returns an `OpenTelemetryV2Config` (exporter destination + mapper vocabularies
+  + resource attributes). `PRESET_BY_CALLBACK` maps a callback name (`"arize"`,
+  `"langfuse_otel"`, …) to its preset. Integrations that support team/key-scoped
+  credentials also provide a per-request OTLP header builder
+  (`DYNAMIC_HEADERS_BY_CALLBACK`). Presets do **no** network I/O at build time:
+  AgentOps, for example, mints its JWT lazily inside a custom exporter on the
+  first export (in the `BatchSpanProcessor` worker thread), never on the event
+  loop.
+
+## Extending
+
+- **A new attribute vocabulary for a backend**: add a mapper in `mappers/`
+  (a class with a `map(data) -> AttributeMap` method, typically built from
+  `key -> extractor` tables) and register it in `mappers/__init__._MAPPER_BY_NAME`.
+- **A new integration**: add a preset in `presets/` that returns an
+  `OpenTelemetryV2Config`, and register it in `presets/__init__.PRESET_BY_CALLBACK`.
+  If it supports dynamic credentials, add a header builder to
+  `DYNAMIC_HEADERS_BY_CALLBACK`.
+- **A new span kind**: add a role to `spans.py` (registry entry + name builder),
+  a payload dataclass in `payloads.py`, and a branch in the relevant mapper(s).

litellm/integrations/otel/__init__.py

@@ -0,0 +1,118 @@
+"""Typed, semconv-aligned OpenTelemetry instrumentation for LiteLLM.
+
+The three sources of truth — attribute keys (:mod:`semconv`), the span and
+hierarchy registry (:mod:`spans`), and the typed span-data inputs
+(:mod:`payloads`) — plus :mod:`config` are exported here and are free of any
+``opentelemetry`` import. The engine layer (``emitter``, ``providers``,
+``context``, ``metrics``) and the ``CustomLogger`` adapter (``logger``) are
+reached via their submodule paths so that importing this package never
+requires the OTel SDK.
+
+The ``LITELLM_OTEL_V2`` env var gates whether the factory in
+``litellm_core_utils.litellm_logging`` constructs the ``OpenTelemetryV2``
+class (from :mod:`logger`).
+"""
+
+from litellm.integrations.otel.model.config import (
+    OTEL_V2_ENV,
+    OpenTelemetryV2Config,
+    is_otel_v2_enabled,
+)
+from litellm.integrations.otel.model.baggage import (
+    BAGGAGE_PROMOTED_KEYS,
+    DEFAULT_BAGGAGE_METADATA_KEYS,
+    promoted_baggage,
+)
+from litellm.integrations.otel.model.metadata import (
+    RequestContext,
+    RequestIdentity,
+)
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    LLMRequestParams,
+    LLMUsage,
+    MCPToolCallSpanData,
+    ProxyRequestSpanData,
+    ServerInfo,
+    ServiceSpanData,
+    SpanError,
+    is_mcp_tool_call,
+)
+from litellm.integrations.otel.model.semconv import (
+    DB,
+    HTTP,
+    MCP,
+    Client,
+    Error,
+    GenAI,
+    GenAIOperation,
+    GenAIProvider,
+    JsonRpc,
+    LiteLLM,
+    MCPMethod,
+    Metric,
+    Network,
+    NetworkTransport,
+    Server,
+    resolve_operation,
+    resolve_provider,
+)
+from litellm.integrations.otel.model.spans import (
+    SPAN_REGISTRY,
+    LiteLLMSpanKind,
+    SpanRole,
+    SpanSpec,
+    db_system,
+    span_role_for_service,
+    validate_registry,
+)
+
+__all__ = [
+    # config
+    "OTEL_V2_ENV",
+    "OpenTelemetryV2Config",
+    "is_otel_v2_enabled",
+    # semconv
+    "BAGGAGE_PROMOTED_KEYS",
+    "DB",
+    "DEFAULT_BAGGAGE_METADATA_KEYS",
+    "Client",
+    "Error",
+    "GenAI",
+    "GenAIOperation",
+    "GenAIProvider",
+    "HTTP",
+    "JsonRpc",
+    "LiteLLM",
+    "MCP",
+    "MCPMethod",
+    "Metric",
+    "Network",
+    "NetworkTransport",
+    "Server",
+    "resolve_operation",
+    "resolve_provider",
+    # spans
+    "SPAN_REGISTRY",
+    "LiteLLMSpanKind",
+    "SpanRole",
+    "SpanSpec",
+    "db_system",
+    "span_role_for_service",
+    "validate_registry",
+    # payloads
+    "GuardrailSpanData",
+    "LLMCallSpanData",
+    "LLMRequestParams",
+    "LLMUsage",
+    "MCPToolCallSpanData",
+    "ProxyRequestSpanData",
+    "RequestContext",
+    "RequestIdentity",
+    "ServerInfo",
+    "ServiceSpanData",
+    "SpanError",
+    "is_mcp_tool_call",
+    "promoted_baggage",
+]

litellm/integrations/otel/emitter.py

@@ -0,0 +1,190 @@
+"""The span engine: dedup, start, run the mapper chain, set status, end."""
+
+from collections import OrderedDict
+from typing import Callable, Sequence
+
+from opentelemetry.context import Context
+from opentelemetry.trace import Span, Tracer
+from opentelemetry.trace.status import Status, StatusCode
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.mappers import resolve_mappers
+from litellm.integrations.otel.mappers.base import AttributeMapper, SpanData
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    MCPToolCallSpanData,
+    ServiceSpanData,
+)
+from litellm.integrations.otel.plumbing.providers import to_otel_span_kind
+from litellm.integrations.otel.model.semconv import Error
+from litellm.integrations.otel.model.spans import (
+    SPAN_REGISTRY,
+    SpanRole,
+    guardrail_span_name,
+    llm_call_span_name,
+    mcp_tool_call_span_name,
+    service_span_name,
+)
+
+# Roles emit() knows how to name and emit. PROXY_REQUEST and the management
+# routes are SERVER spans owned by the mounted FastAPI instrumentor, so they
+# have no builder here.
+_NAME_BUILDERS: dict[SpanRole, Callable[..., str]] = {
+    SpanRole.LLM_CALL: llm_call_span_name,
+    SpanRole.MCP_TOOL_CALL: mcp_tool_call_span_name,
+    SpanRole.GUARDRAIL: guardrail_span_name,
+    # DB_CALL and SERVICE are both built from ServiceSpanData; they differ only in
+    # span kind (CLIENT vs INTERNAL) and attribute vocabulary, not in naming.
+    SpanRole.DB_CALL: service_span_name,
+    SpanRole.SERVICE: service_span_name,
+}
+
+# Cap on the dedup cache. It only needs to coalesce the sync+async firing window
+# of a single in-flight request, so a bounded LRU keeps memory flat on a
+# long-running proxy while still covering every concurrently-open call.
+_DEDUP_CACHE_MAX = 10_000
+
+
+class SpanEmitter:
+    def __init__(
+        self,
+        tracer: Tracer,
+        config: OpenTelemetryV2Config,
+        mappers: Sequence[AttributeMapper] | None = None,
+    ) -> None:
+        self._tracer = tracer
+        self._config = config
+        # The mapper chain is the sole source of span attributes. When not
+        # passed in, resolve it from the config so there's one source of truth.
+        self._mappers: list[AttributeMapper] = (
+            list(mappers)
+            if mappers is not None
+            else resolve_mappers(config.mapper_names)
+        )
+        # Bounded LRU (ordered by insertion / most-recent touch). Storing keys
+        # only — the value is unused — so it behaves like a capped set.
+        self._emitted: "OrderedDict[tuple[str, SpanRole], None]" = OrderedDict()
+
+    # -- low-level helpers --------------------------------------------------- #
+
+    def start_span(
+        self,
+        role: SpanRole,
+        name: str,
+        parent_context: Context | None = None,
+        start_time_ns: int | None = None,
+        *,
+        tracer: Tracer | None = None,
+    ) -> Span:
+        """Start a span for ``role`` without dedup or attribute mapping.
+
+        For callers that own and manage their own span lifecycle. ``tracer``
+        overrides the bound tracer for this span only, used for per-request
+        multi-tenant credential routing.
+        """
+        return (tracer or self._tracer).start_span(
+            name,
+            context=parent_context,
+            kind=to_otel_span_kind(SPAN_REGISTRY[role].kind),
+            start_time=start_time_ns,
+        )
+
+    def _seen(self, dedup_key: str | None, role: SpanRole) -> bool:
+        """Return True once a ``(dedup_key, role)`` pair has been emitted.
+
+        Guards against emitting the same span twice when a streaming call
+        fires both a sync and an async logging callback.
+        """
+        if not dedup_key:
+            return False
+        marker = (dedup_key, role)
+        if marker in self._emitted:
+            self._emitted.move_to_end(marker)
+            return True
+        self._emitted[marker] = None
+        if len(self._emitted) > _DEDUP_CACHE_MAX:
+            self._emitted.popitem(last=False)  # evict least-recently-used
+        return False
+
+    # -- the engine ---------------------------------------------------------- #
+
+    def emit(
+        self,
+        role: SpanRole,
+        data: SpanData,
+        parent_context: Context | None = None,
+        *,
+        start_time_ns: int | None = None,
+        end_time_ns: int | None = None,
+        tracer: Tracer | None = None,
+    ) -> Span | None:
+        """Emit one complete span: dedup, start, map attributes, status, end.
+
+        Return the span, or ``None`` if it was deduplicated away. ``tracer``
+        overrides the bound tracer for this span, used for per-request routing.
+        """
+        # LLM-call and MCP tool-call spans carry a dedup key (their request's
+        # call id), so a sync+async double-firing coalesces. ``isinstance`` narrows
+        # the type for mypy and keeps the engine free of duck-typed attribute reads.
+        dedup_key = (
+            data.identity.call_id
+            if isinstance(data, (LLMCallSpanData, MCPToolCallSpanData))
+            else None
+        )
+        if self._seen(dedup_key, role):
+            return None
+        span = self.start_span(
+            role,
+            _NAME_BUILDERS[role](data),
+            parent_context=parent_context,
+            start_time_ns=start_time_ns,
+            tracer=tracer,
+        )
+        self.finish_span(role, span, data, end_time_ns=end_time_ns)
+        return span
+
+    def finish_span(
+        self,
+        role: SpanRole,
+        span: Span,
+        data: SpanData,
+        *,
+        end_time_ns: int | None = None,
+    ) -> None:
+        """Stamp attributes + status on an already-started ``span`` and end it.
+
+        The counterpart to :meth:`start_span` for callers that own a span's
+        lifecycle — the LLM-call span is opened at the request's ``pre_call``
+        boundary (so it parents to the live server span via real ambient context,
+        never a span threaded through a metadata dict) and closed here once the
+        typed payload is available. The span name is (re)built from the now-known
+        data, since the boundary opener only has a provisional name.
+        """
+        span.update_name(_NAME_BUILDERS[role](data))
+        for mapper in self._mappers:
+            for key, value in mapper.map(data).items():
+                span.set_attribute(key, value)
+        error = (
+            data.error
+            if isinstance(
+                data,
+                (
+                    LLMCallSpanData,
+                    MCPToolCallSpanData,
+                    ServiceSpanData,
+                    GuardrailSpanData,
+                ),
+            )
+            else None
+        )
+        if error and (error.error_type or error.message):
+            span.set_attribute(Error.TYPE, error.error_type or "error")
+            span.set_status(
+                Status(StatusCode.ERROR, error.message or error.error_type or "error")
+            )
+        # On success leave the status UNSET (the semconv default) rather than
+        # forcing OK — that matches the FastAPI server span and avoids implying a
+        # span-level health signal litellm doesn't actually evaluate. Only a
+        # genuine error sets a status.
+        span.end(end_time=end_time_ns)

litellm/integrations/otel/logger.py

@@ -0,0 +1,544 @@
+"""``CustomLogger`` adapter on the OpenTelemetry span engine."""
+
+from collections import OrderedDict
+from contextlib import contextmanager
+from datetime import datetime
+from typing import TYPE_CHECKING, Any, Iterator, Mapping, cast
+
+from opentelemetry.context import attach, get_current
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.trace import Span, Tracer, get_current_span, use_span
+
+import litellm
+from litellm.integrations.custom_logger import CustomLogger
+from litellm.integrations.otel.model.baggage import promoted_baggage
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.plumbing.context import (
+    is_recordable_span,
+    resolve_parent_context,
+    resolve_request_span_context,
+    set_request_baggage,
+    set_request_root_span,
+)
+from litellm.integrations.otel.emitter import SpanEmitter
+from litellm.integrations.otel.mappers import resolve_mappers
+from litellm.integrations.otel.model.metadata import (
+    LLMCallEvent,
+    RequestIdentity,
+    model_from_request_data,
+)
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    MCPToolCallSpanData,
+    ServiceSpanData,
+    SpanError,
+    is_mcp_tool_call,
+)
+from litellm.integrations.otel.plumbing.providers import (
+    build_tracer_provider,
+    get_tracer,
+)
+from litellm.integrations.otel.plumbing.routing import TenantTracerCache
+from litellm.integrations.otel.model.spans import SpanRole, span_role_for_service
+from litellm.integrations.otel.model.utils import to_ns
+
+if TYPE_CHECKING:
+    from litellm.types.utils import (
+        StandardLoggingGuardrailInformation,
+        StandardLoggingPayload,
+    )
+
+LITELLM_TRACER_NAME = "litellm"
+
+# Any callback whose class belongs to one of these modules is "the OTel
+# callback" for proxy-global-registration purposes.
+_OTEL_MODULES = (
+    "litellm.integrations.otel",
+    "litellm.integrations.opentelemetry",
+)
+
+
+# Cap on the open-call carrier map. A span opened at ``pre_call`` that never
+# reaches a success/failure callback (e.g. a stream that only fires stream
+# events) would otherwise linger; bounding the map evicts the oldest so memory
+# stays flat on a long-running proxy while covering every concurrent in-flight
+# call.
+_OPEN_CALLS_MAX = 10_000
+
+
+class _LLMCallSpan:
+    """The state carried from the ``pre_call`` boundary to span close.
+
+    ``span`` is the live span when it could be opened at the boundary (the server
+    span was ambient), or ``None`` when creation was deferred because no ambient
+    parent was visible — in which case the async callback creates it against its
+    own (worker-copied) ambient context using ``start_time_ns``. The presence of
+    a carrier for a call at all is the proof that ``pre_call`` ran, i.e. that an
+    upstream call was actually attempted.
+    """
+
+    __slots__ = ("span", "start_time_ns")
+
+    def __init__(self, span: "Span | None", start_time_ns: int | None) -> None:
+        self.span = span
+        self.start_time_ns = start_time_ns
+
+
+class OpenTelemetryV2(CustomLogger):
+    """The ``CustomLogger`` for OpenTelemetry."""
+
+    def __init__(
+        self,
+        config: OpenTelemetryV2Config | None = None,
+        callback_name: str | None = None,
+        tracer_provider: TracerProvider | None = None,
+        logger_provider: Any | None = None,  # reserved for OTel logs
+        meter_provider: Any | None = None,  # reserved for metrics
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.config: OpenTelemetryV2Config = config or OpenTelemetryV2Config(**kwargs)
+        self.callback_name = callback_name
+        self._tracer_provider: TracerProvider = (
+            tracer_provider
+            if tracer_provider is not None
+            else build_tracer_provider(self.config)
+        )
+        self.tracer: Tracer = get_tracer(self._tracer_provider, LITELLM_TRACER_NAME)
+        self._emitter = SpanEmitter(
+            self.tracer, self.config, mappers=resolve_mappers(self.config.mapper_names)
+        )
+        self._tenant_tracers = TenantTracerCache(
+            self.config, callback_name, LITELLM_TRACER_NAME
+        )
+        self._open_llm_calls: "OrderedDict[str, _LLMCallSpan]" = OrderedDict()
+        self._init_otel_logger_on_litellm_proxy()
+
+    # ====================================================================== #
+    #  Proxy global registration
+    # ====================================================================== #
+
+    def _register_in_callback_list(self, callbacks: list) -> None:
+        already_otel = any(
+            cb.__class__.__module__.startswith(_OTEL_MODULES)
+            for cb in callbacks
+            if hasattr(cb, "__class__")
+        )
+        if not already_otel:
+            callbacks.append(self)
+
+    def _init_otel_logger_on_litellm_proxy(self) -> None:
+        try:
+            from litellm.proxy import proxy_server
+        except Exception:
+            return
+        try:
+            self._register_in_callback_list(litellm.service_callback)
+            self._register_in_callback_list(litellm.input_callback)
+            self._register_in_callback_list(litellm._async_success_callback)
+            self._register_in_callback_list(litellm._async_failure_callback)
+        except Exception:
+            pass
+        if getattr(proxy_server, "open_telemetry_logger", None) is None:
+            setattr(proxy_server, "open_telemetry_logger", self)
+
+    # ====================================================================== #
+    #  LLM-call callbacks — the span is opened at the ``pre_call`` boundary and
+    #  closed here. See ``log_pre_api_call``.
+    # ====================================================================== #
+
+    def log_pre_api_call(self, model, messages, kwargs):
+        """Open the LLM-call span at the call boundary.
+
+        Runs synchronously inside the request task, before the upstream call —
+        the one place where the live server span is genuinely the ambient OTel
+        context — so the span parents to it natively, with no span threaded
+        through a metadata dict. The open span is stashed on the per-request
+        ``LiteLLMLoggingObj`` (a typed object) and closed in the async callback.
+
+        When no recordable parent is visible (``pre_call`` was driven from a thread
+        pool for a sync-only provider, where contextvars — and so the anchor —
+        don't follow), creation is deferred: only the start time is recorded, and
+        the async callback — whose worker context was copied from the request task
+        and so still carries the anchor — creates the span then.
+
+        Synthetic proxy-gate error logs (auth/rate-limit rejections) also fire this
+        hook but never made an upstream call; they are tagged and skipped so no
+        phantom LLM-call span is produced.
+        """
+        call = LLMCallEvent.from_dict(kwargs)
+        if call.is_no_upstream_call:
+            return
+        call_id = call.call_id
+        if call_id is None:
+            return
+        # Idempotent: a retried call may re-enter ``pre_call`` with the same
+        # call id; keep the first span so its start time is the true one.
+        if call_id in self._open_llm_calls:
+            return
+        start_time_ns = to_ns(datetime.now())
+        span: Span | None = None
+        # Parent to the request's anchored root span (stable across the request),
+        # falling back to ambient on the SDK path. Open the span live only when
+        # that resolves to a recordable parent; otherwise defer to the close
+        # callback (the thread-pool case, where the anchor isn't visible here).
+        parent_context = resolve_request_span_context()
+        if is_recordable_span(get_current_span(parent_context)):
+            span = self._emitter.start_span(
+                SpanRole.LLM_CALL,
+                call.provisional_span_name,
+                parent_context=parent_context,
+                start_time_ns=start_time_ns,
+                tracer=self._tenant_tracers.tracer_for(
+                    self.tracer, call.dynamic_params
+                ),
+            )
+        self._open_llm_calls[call_id] = _LLMCallSpan(
+            span=span, start_time_ns=start_time_ns
+        )
+        # Evict the oldest open call if the map is over budget. A call that opens
+        # but never closes (a stream that only fires stream events) would linger
+        # otherwise; the evicted span is simply dropped (never exported).
+        if len(self._open_llm_calls) > _OPEN_CALLS_MAX:
+            self._open_llm_calls.popitem(last=False)
+
+    async def async_log_success_event(self, kwargs, response_obj, start_time, end_time):
+        if self._emit_mcp_tool_call(kwargs, start_time, end_time):
+            return
+        self._close_llm_call(kwargs, start_time, end_time)
+
+    async def async_log_failure_event(self, kwargs, response_obj, start_time, end_time):
+        if self._emit_mcp_tool_call(kwargs, start_time, end_time):
+            return
+        self._close_llm_call(kwargs, start_time, end_time)
+
+    def _emit_mcp_tool_call(
+        self,
+        kwargs: Mapping[str, Any],
+        start_time: datetime | float | None,
+        end_time: datetime | float | None,
+    ) -> bool:
+        """Emit an MCP tool-call span when the closed request was a tool call.
+
+        MCP tool calls reach the success/failure callbacks like any other request
+        (with ``call_type`` ``call_mcp_tool``), but they are not LLM calls and have
+        no ``pre_call`` carrier — so they get their own CLIENT span here, parented
+        to the request's server span. Returns whether it handled the event, so the
+        caller skips the LLM-call path. The whole span is emitted at once (there is
+        no boundary to open it at), deduped on the call id by the emitter.
+        """
+        raw_payload = kwargs.get("standard_logging_object")
+        if not raw_payload or not is_mcp_tool_call(
+            cast(Mapping[str, object], raw_payload)
+        ):
+            return False
+        payload = cast("StandardLoggingPayload", raw_payload)
+        data = MCPToolCallSpanData.from_standard_logging_payload(
+            payload, capture_content=self.config.capture_span_content
+        )
+        # A stray LLM carrier from a ``pre_call`` that mis-fired for this id would
+        # otherwise linger until evicted; drop it so it's neither leaked nor closed
+        # as a phantom LLM span.
+        if data.identity.call_id:
+            self._open_llm_calls.pop(data.identity.call_id, None)
+        self._emitter.emit(
+            SpanRole.MCP_TOOL_CALL,
+            data,
+            parent_context=resolve_request_span_context(),
+            start_time_ns=to_ns(start_time),
+            end_time_ns=to_ns(end_time),
+        )
+        return True
+
+    def _close_llm_call(
+        self,
+        kwargs: Mapping[str, Any],
+        start_time: datetime | float | None,
+        end_time: datetime | float | None,
+    ) -> Span | None:
+        """Finish the LLM-call span opened at ``pre_call`` (or create it deferred).
+
+        No carrier for this call id means ``pre_call`` never ran — the request was
+        rejected at the gate or blocked by a pre-call guardrail before any upstream
+        call — so there is nothing to record and no phantom span.
+        """
+        call = LLMCallEvent.from_dict(kwargs)
+        call_id = call.call_id
+        # ``pop`` is the dedup: this method runs from both the success and failure
+        # paths, and whichever fires first removes the carrier and closes the span.
+        carrier = self._open_llm_calls.pop(call_id, None) if call_id else None
+        if carrier is None:
+            return None
+        payload = call.payload
+        if payload is None:
+            if carrier.span is not None:
+                # Opened at the boundary but the payload never materialized — end
+                # it (named provisionally) so it isn't leaked as an open span.
+                carrier.span.end(end_time=to_ns(end_time))
+            return None
+        data = LLMCallSpanData.from_standard_logging_payload(
+            payload, capture_content=self.config.capture_span_content
+        )
+        end_time_ns = to_ns(end_time)
+        if carrier.span is not None:
+            # Born at the boundary: stamp attributes from the typed payload, set
+            # status, and end it. Its parent (the server span) was captured at
+            # creation from real ambient context.
+            self._emitter.finish_span(
+                SpanRole.LLM_CALL, carrier.span, data, end_time_ns=end_time_ns
+            )
+            return carrier.span
+        # Deferred: ``pre_call`` saw no recordable parent, so create the span now.
+        # The worker copied the request task's context, which carries the anchored
+        # root span — parent to it (ambient fallback on the SDK path). Seed identity
+        # Baggage so the span — and the SDK path, which has none — is labeled
+        # consistently.
+        parent_ctx = resolve_request_span_context()
+        bag = promoted_baggage(
+            data.identity,
+            data.request_model,
+            promoted_keys=tuple(self.config.baggage_promoted_keys),
+            metadata_keys=tuple(self.config.baggage_metadata_keys),
+            team_metadata_keys=tuple(self.config.baggage_team_metadata_keys),
+        )
+        if bag:
+            parent_ctx = set_request_baggage(bag, context=parent_ctx)
+        return self._emitter.emit(
+            SpanRole.LLM_CALL,
+            data,
+            parent_context=parent_ctx,
+            start_time_ns=carrier.start_time_ns,
+            end_time_ns=end_time_ns,
+            tracer=self._tenant_tracers.tracer_for(self.tracer, call.dynamic_params),
+        )
+
+    # ====================================================================== #
+    #  Service hooks
+    # ====================================================================== #
+
+    async def async_service_success_hook(
+        self,
+        payload: Any,
+        parent_otel_span: Span | None = None,
+        start_time: datetime | float | None = None,
+        end_time: datetime | float | None = None,
+        event_metadata: dict | None = None,
+    ) -> None:
+        self._emit_service(
+            payload,
+            parent_otel_span=parent_otel_span,
+            start_time=start_time,
+            end_time=end_time,
+            event_metadata=event_metadata,
+            error_override=None,
+        )
+
+    async def async_service_failure_hook(
+        self,
+        payload: Any,
+        error: str | None = "",
+        parent_otel_span: Span | None = None,
+        start_time: datetime | float | None = None,
+        end_time: datetime | float | None = None,
+        event_metadata: dict | None = None,
+    ) -> None:
+        self._emit_service(
+            payload,
+            parent_otel_span=parent_otel_span,
+            start_time=start_time,
+            end_time=end_time,
+            event_metadata=event_metadata,
+            error_override=error or "error",
+        )
+
+    def _emit_service(
+        self,
+        payload: Any,
+        *,
+        parent_otel_span: Span | None,
+        start_time: datetime | float | None,
+        end_time: datetime | float | None,
+        event_metadata: dict | None,
+        error_override: str | None,
+    ) -> Span | None:
+        data = ServiceSpanData.from_payload(payload, event_metadata=event_metadata)
+        # Decide whether this service call is a span at all, and of what kind.
+        # ``None`` means metrics-only (framework instrumentation that duplicates a
+        # gen-AI span — ``self``/``router``/``proxy_pre_call`` — or ``auth``, which
+        # gets a live phase span instead). Those still feed Prometheus/Datadog via
+        # their own hooks; they just never enter the trace.
+        role = span_role_for_service(data.service_name)
+        if role is None:
+            return None
+        # A metrics-only ping with neither timing nor a parent (in-memory queue
+        # gauges) is not a traceable operation; a span for it would be a
+        # zero-duration root with no context, so skip it. Real background work
+        # (budget/reset jobs, spend flush) passes start/end times and still emits
+        # as a root; anything with a parent emits regardless.
+        if (
+            error_override is None
+            and start_time is None
+            and end_time is None
+            and parent_otel_span is None
+        ):
+            return None
+        if error_override is not None and data.error is None:
+            data = ServiceSpanData(
+                service_name=data.service_name,
+                call_type=data.call_type,
+                error=SpanError(message=error_override),
+                event_metadata=data.event_metadata,
+            )
+        # Parent like every other span: ambient context first (so identity Baggage
+        # rides along and the call nests under whatever request phase is active —
+        # e.g. a DB lookup under the live ``auth`` span), falling back to the
+        # server span the proxy threaded as ``parent_otel_span``. A background
+        # service call has neither, so it starts its own root trace.
+        parent_context = resolve_parent_context(threaded=parent_otel_span)
+        return self._emitter.emit(
+            role,
+            data,
+            parent_context=parent_context,
+            start_time_ns=to_ns(start_time),
+            end_time_ns=to_ns(end_time),
+        )
+
+    # ====================================================================== #
+    #  async_post_call_* hooks — emit guardrail spans. The server span's status
+    #  / errors are the FastAPI instrumentor's job, so we don't touch it here.
+    # ====================================================================== #
+
+    def seed_request_identity(self, user_api_key_dict: Any, model: Any = None) -> None:
+        """Attach request-identity Baggage to the current context + server span.
+
+        Seeding identity into Baggage makes **every** span emitted afterwards for
+        this request — LLM call, guardrail, DB call — inherit it via
+        ``LiteLLMBaggageSpanProcessor``. Called once at the auth boundary (as soon
+        as the key resolves) so post-auth spans are labeled consistently; the
+        Baggage rides the request task's contextvar from there on. Auth-internal
+        DB lookups that run before the key is known stay unlabeled — identity
+        isn't determined yet, which is correct.
+        """
+        try:
+            identity = RequestIdentity.from_user_api_key_auth(user_api_key_dict)
+            bag = promoted_baggage(
+                identity,
+                model,
+                promoted_keys=tuple(self.config.baggage_promoted_keys),
+                metadata_keys=tuple(self.config.baggage_metadata_keys),
+                team_metadata_keys=tuple(self.config.baggage_team_metadata_keys),
+            )
+            if bag:
+                # Attach (no detach): the contextvar is scoped to this request's
+                # asyncio task and is reclaimed when the task ends.
+                attach(set_request_baggage(bag, context=get_current()))
+                # The server span was started by the instrumentor before this ran,
+                # so the Baggage processor (which only fires at span start) won't
+                # backfill it — stamp identity on it directly.
+                server_span = get_current_span()
+                if is_recordable_span(server_span):
+                    # Re-capture the anchor here too: this runs post-auth with the
+                    # server span active and covers entrypoints that bypass
+                    # ``create_litellm_proxy_request_started_span`` (e.g. the SDK
+                    # path's ``async_pre_call_hook``). Idempotent.
+                    set_request_root_span(server_span)
+                    for key, value in bag.items():
+                        server_span.set_attribute(key, value)
+        except Exception:
+            pass
+
+    @contextmanager
+    def start_phase_span(self, name: str) -> "Iterator[Span]":
+        span = self._emitter.start_span(SpanRole.SERVICE, name)
+        with use_span(span, end_on_exit=True):
+            yield span
+
+    async def async_pre_call_hook(
+        self,
+        user_api_key_dict: Any,
+        cache: Any,
+        data: dict,
+        call_type: Any,
+    ) -> dict:
+        self.seed_request_identity(
+            user_api_key_dict,
+            model=model_from_request_data(data),
+        )
+        return data
+
+    def emit_guardrail_span(self, entry: "StandardLoggingGuardrailInformation") -> None:
+        # Emitted by the guardrail-recording code the moment a guardrail finishes,
+        # not from a post-call hook — that hook does not fire on every path (a
+        # pass-through request that passes its guardrails never reaches it), which
+        # left passing guardrails without a span.
+        #
+        # A guardrail is a sibling of the LLM call under the request's root span,
+        # so parent it to the explicit anchor — never the active span, which during
+        # a pre_call guardrail can be the live ``auth`` phase span. Emit with the
+        # guardrail's actual execution window so a pre_call guardrail is placed
+        # before the LLM call rather than at emission time. One entry in, one span
+        # out — the module-level entry point routes each entry to this single
+        # registered logger so a guardrail is never emitted more than once.
+        data = GuardrailSpanData.from_logging_entry(entry)
+        self._emitter.emit(
+            SpanRole.GUARDRAIL,
+            data,
+            parent_context=resolve_request_span_context(),
+            start_time_ns=to_ns(data.start_time),
+            end_time_ns=to_ns(data.end_time),
+        )
+
+    def create_litellm_proxy_request_started_span(
+        self, start_time: datetime, headers: Mapping[str, str] | None
+    ) -> Span | None:
+        span = get_current_span()
+        if not is_recordable_span(span):
+            return None
+        set_request_root_span(span)
+        return span
+
+
+def _registered_v2_logger() -> "OpenTelemetryV2 | None":
+    try:
+        from litellm.proxy import proxy_server
+    except Exception:
+        return None
+    logger = getattr(proxy_server, "open_telemetry_logger", None)
+    return logger if isinstance(logger, OpenTelemetryV2) else None
+
+
+def emit_guardrail_span(entry: "StandardLoggingGuardrailInformation") -> None:
+    """Emit a guardrail span on the registered v2 OTel logger.
+
+    Called by the guardrail-recording code the moment a guardrail finishes, so a
+    span is produced regardless of whether a post-call hook later runs (it does
+    not on the pass-through allow path). Routes through the single canonical
+    logger — the same one every other v2 entry point uses — so a guardrail
+    recorded once yields exactly one span; fanning out across every reachable
+    ``OpenTelemetryV2`` instance double-emits the same entry. Best-effort: span
+    emission must never break guardrail evaluation.
+    """
+    logger = _registered_v2_logger()
+    if logger is None:
+        return
+    try:
+        logger.emit_guardrail_span(entry)
+    except Exception:
+        pass
+
+
+def seed_request_identity(user_api_key_dict: Any, model: Any = None) -> None:
+    logger = _registered_v2_logger()
+    if logger is not None:
+        logger.seed_request_identity(user_api_key_dict, model=model)
+
+
+@contextmanager
+def phase_span(name: str) -> "Iterator[Span | None]":
+    logger = _registered_v2_logger()
+    if logger is None:
+        yield None
+        return
+    with logger.start_phase_span(name) as span:
+        yield span

litellm/integrations/otel/mappers/__init__.py

@@ -0,0 +1,58 @@
+"""Attribute mappers: pure ``LLMCallSpanData -> {attribute key: value}`` functions.
+
+Composition over inheritance: vocabularies layer onto the same span. Listing
+``["genai", "openinference"]`` in ``config.mapper_names`` makes every span
+carry both the canonical ``gen_ai.*`` keys and the OpenInference (Arize +
+Phoenix) keys. Add ``"langfuse"`` and it works for all three backends at once.
+"""
+
+from typing import Callable, Iterable
+
+from litellm.integrations.otel.mappers.base import (
+    AttributeMap,
+    AttributeMapper,
+    AttrValue,
+)
+from litellm.integrations.otel.mappers.genai import GenAIMapper
+from litellm.integrations.otel.mappers.langfuse import LangfuseMapper
+from litellm.integrations.otel.mappers.langtrace import LangtraceMapper
+from litellm.integrations.otel.mappers.legacy import LegacyMapper
+from litellm.integrations.otel.mappers.openinference import OpenInferenceMapper
+from litellm.integrations.otel.mappers.weave import WeaveMapper
+
+# Registry keyed by ``config.mapper_names`` entries.
+_MAPPER_BY_NAME: dict[str, Callable[[], AttributeMapper]] = {
+    "genai": GenAIMapper,
+    "legacy": LegacyMapper,
+    "openinference": OpenInferenceMapper,
+    "langfuse": LangfuseMapper,
+    "weave": WeaveMapper,
+    "langtrace": LangtraceMapper,
+}
+
+
+def resolve_mappers(names: Iterable[str]) -> list[AttributeMapper]:
+    """Resolve mapper names to instances. Unknown names raise ``ValueError``."""
+    out: list[AttributeMapper] = []
+    for name in names:
+        factory = _MAPPER_BY_NAME.get(name)
+        if factory is None:
+            raise ValueError(
+                f"unknown mapper name {name!r}; known: " f"{sorted(_MAPPER_BY_NAME)}"
+            )
+        out.append(factory())
+    return out
+
+
+__all__ = [
+    "AttributeMap",
+    "AttributeMapper",
+    "AttrValue",
+    "GenAIMapper",
+    "LangfuseMapper",
+    "LangtraceMapper",
+    "LegacyMapper",
+    "OpenInferenceMapper",
+    "WeaveMapper",
+    "resolve_mappers",
+]

litellm/integrations/otel/mappers/base.py

@@ -0,0 +1,37 @@
+"""Mapper protocol and attribute value types."""
+
+from typing import Sequence
+
+from typing_extensions import Protocol, runtime_checkable
+
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    MCPToolCallSpanData,
+    ServiceSpanData,
+)
+
+AttrScalar = str | bool | int | float
+# Mirrors ``opentelemetry.util.types.AttributeValue`` (homogeneous sequences)
+# without importing the SDK, so mappers stay OTel-free.
+AttrValue = (
+    AttrScalar | Sequence[str] | Sequence[bool] | Sequence[int] | Sequence[float]
+)
+AttributeMap = dict[str, AttrValue]
+
+# The closed set of span-data types the engine routes through the mapper chain.
+# Server spans (PROXY_REQUEST + management routes) belong to the mounted FastAPI
+# instrumentor, not the mapper chain.
+SpanData = LLMCallSpanData | MCPToolCallSpanData | GuardrailSpanData | ServiceSpanData
+
+
+@runtime_checkable
+class AttributeMapper(Protocol):
+    """Maps a typed span input to a flat dict of OTel span attributes.
+
+    One method per mapper, dispatched internally on the ``data`` type. The
+    engine calls this uniformly for every span kind — mappers that don't speak
+    a given type return ``{}``. This is why the engine contains no attribute keys.
+    """
+
+    def map(self, data: SpanData) -> AttributeMap: ...

litellm/integrations/otel/mappers/genai.py

@@ -0,0 +1,159 @@
+"""Canonical OpenTelemetry GenAI semantic-convention mapper (always active).
+
+Owns the attribute schema for every span kind the engine emits — LLM call,
+guardrail, and service — so the engine itself never references attribute keys.
+
+Each span kind declares its schema as a flat ``attribute key -> extractor``
+table: one lambda per mapping operation, applied against the typed span data.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, drop_none
+from litellm.integrations.otel.model.payloads import (
+    GuardrailSpanData,
+    LLMCallSpanData,
+    MCPToolCallSpanData,
+    ServiceSpanData,
+    ToolDefinition,
+)
+from litellm.integrations.otel.model.semconv import (
+    DB,
+    MCP,
+    Error,
+    GenAI,
+    LiteLLM,
+    Server,
+)
+from litellm.integrations.otel.model.spans import db_system
+
+
+class GenAIMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        GenAI.OPERATION_NAME: lambda d: d.operation.value,
+        GenAI.PROVIDER_NAME: lambda d: d.provider or None,
+        GenAI.REQUEST_MODEL: lambda d: d.request_model or None,
+        GenAI.REQUEST_TEMPERATURE: lambda d: d.request_params.temperature,
+        GenAI.REQUEST_TOP_P: lambda d: d.request_params.top_p,
+        GenAI.REQUEST_TOP_K: lambda d: d.request_params.top_k,
+        GenAI.REQUEST_MAX_TOKENS: lambda d: d.request_params.max_tokens,
+        GenAI.REQUEST_FREQUENCY_PENALTY: lambda d: d.request_params.frequency_penalty,
+        GenAI.REQUEST_PRESENCE_PENALTY: lambda d: d.request_params.presence_penalty,
+        GenAI.REQUEST_STOP_SEQUENCES: lambda d: (
+            list(d.request_params.stop_sequences)
+            if d.request_params.stop_sequences
+            else None
+        ),
+        GenAI.REQUEST_SEED: lambda d: d.request_params.seed,
+        GenAI.RESPONSE_MODEL: lambda d: d.response_model,
+        GenAI.RESPONSE_ID: lambda d: d.response_id,
+        GenAI.RESPONSE_FINISH_REASONS: lambda d: (
+            list(d.finish_reasons) if d.finish_reasons else None
+        ),
+        GenAI.USAGE_INPUT_TOKENS: lambda d: d.usage.input_tokens,
+        GenAI.USAGE_OUTPUT_TOKENS: lambda d: d.usage.output_tokens,
+        Error.TYPE: lambda d: d.error.error_type if d.error else None,
+        Server.ADDRESS: lambda d: d.server.address if d.server else None,
+        Server.PORT: lambda d: d.server.port if d.server else None,
+        LiteLLM.CALL_ID: lambda d: d.identity.call_id or None,
+        # The provider/underlying model is only known once routing has picked a
+        # deployment, so it can't ride identity Baggage (seeded at auth, before
+        # routing) onto the boundary-born LLM span — stamp it directly here.
+        LiteLLM.PROVIDER_MODEL: lambda d: d.identity.provider_model or None,
+        f"{LiteLLM.COST_PREFIX}total": lambda d: d.response_cost,
+        LiteLLM.REQUEST_STREAMING: lambda d: d.is_streaming,
+    }
+
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "name": lambda t: t.name,
+        "description": lambda t: t.description or None,
+        "parameters": lambda t: t.parameters_json or None,
+    }
+
+    _MCP_ATTRS: dict[str, Callable[[MCPToolCallSpanData], AttrValue | None]] = {
+        GenAI.OPERATION_NAME: lambda d: d.operation.value,
+        MCP.METHOD_NAME: lambda d: d.method,
+        MCP.SESSION_ID: lambda d: d.session_id,
+        GenAI.TOOL_NAME: lambda d: d.tool_name or None,
+        GenAI.TOOL_CALL_ARGUMENTS: lambda d: d.arguments_json,
+        GenAI.TOOL_CALL_RESULT: lambda d: d.result_json,
+        LiteLLM.MCP_SERVER_NAME: lambda d: d.server_name,
+        LiteLLM.CALL_ID: lambda d: d.identity.call_id or None,
+        f"{LiteLLM.COST_PREFIX}total": lambda d: d.response_cost,
+    }
+
+    _GUARDRAIL_ATTRS: dict[str, Callable[[GuardrailSpanData], AttrValue | None]] = {
+        LiteLLM.GUARDRAIL_NAME: lambda d: d.guardrail_name,
+        LiteLLM.GUARDRAIL_MODE: lambda d: d.mode,
+        LiteLLM.GUARDRAIL_STATUS: lambda d: d.status,
+        LiteLLM.GUARDRAIL_PROVIDER: lambda d: d.provider,
+        LiteLLM.GUARDRAIL_ACTION: lambda d: d.action,
+        LiteLLM.GUARDRAIL_RESPONSE: lambda d: d.response_json,
+        LiteLLM.GUARDRAIL_VIOLATION_CATEGORIES: lambda d: (
+            list(d.violation_categories) if d.violation_categories else None
+        ),
+        LiteLLM.GUARDRAIL_CONFIDENCE_SCORE: lambda d: d.confidence_score,
+        LiteLLM.GUARDRAIL_RISK_SCORE: lambda d: d.risk_score,
+        LiteLLM.GUARDRAIL_MASKED_ENTITY_COUNT: lambda d: d.masked_entity_count,
+        LiteLLM.GUARDRAIL_DURATION: lambda d: d.duration,
+        LiteLLM.GUARDRAIL_ID: lambda d: d.guardrail_id,
+        LiteLLM.GUARDRAIL_POLICY_TEMPLATE: lambda d: d.policy_template,
+        LiteLLM.GUARDRAIL_DETECTION_METHOD: lambda d: d.detection_method,
+    }
+
+    _SERVICE_ATTRS: dict[str, Callable[[ServiceSpanData], AttrValue | None]] = {
+        LiteLLM.SERVICE_NAME: lambda d: d.service_name,
+        LiteLLM.SERVICE_CALL_TYPE: lambda d: d.call_type,
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case MCPToolCallSpanData():
+                return collect(self._MCP_ATTRS, data)
+            case GuardrailSpanData():
+                return self._guardrail(data)
+            case ServiceSpanData():
+                return self._service(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        attrs = collect(cls._LLM_CALL_ATTRS, data)
+        attrs.update(
+            drop_none(
+                {
+                    f"gen_ai.tool.{idx}.{suffix}": extract(tool)
+                    for idx, tool in enumerate(data.tools)
+                    for suffix, extract in cls._TOOL_ATTRS.items()
+                }
+            )
+        )
+        return attrs
+
+    @classmethod
+    def _guardrail(cls, data: GuardrailSpanData) -> AttributeMap:
+        return collect(cls._GUARDRAIL_ATTRS, data)
+
+    @classmethod
+    def _service(cls, data: ServiceSpanData) -> AttributeMap:
+        attrs = collect(cls._SERVICE_ATTRS, data)
+        # An outbound datastore call (DB_CALL / CLIENT span) also carries db.*
+        # semconv. Internal services (router, budget jobs, …) have no db.system,
+        # so they get only the litellm.service.* keys above.
+        system = db_system(data.service_name)
+        if system is not None:
+            attrs[DB.SYSTEM_NAME] = system
+            if data.call_type:
+                attrs[DB.OPERATION_NAME] = data.call_type
+        attrs.update(
+            {
+                f"{LiteLLM.METADATA_PREFIX}{key}": value
+                for key, value in data.event_metadata.items()
+            }
+        )
+        return attrs

litellm/integrations/otel/mappers/langfuse.py

@@ -0,0 +1,84 @@
+"""Langfuse OTLP attribute mapper.
+
+Langfuse ingests OTLP spans and reads from its own vendor namespace
+(``langfuse.observation.*``, ``langfuse.trace.*``). Compose this mapper after
+``GenAIMapper`` to send canonical + Langfuse-flavored spans simultaneously.
+
+Every attribute is declared as a ``key -> extractor`` table entry (one callable
+per mapping operation): ``_LLM_CALL_ATTRS`` for scalars and ``_BLOB_ATTRS`` for
+the JSON-serialized payloads. ``_llm_call`` just applies both tables.
+"""
+
+import json
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    json_if,
+    output_messages,
+    serialize_messages,
+)
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    LLMRequestParams,
+    LLMUsage,
+)
+
+
+class LangfuseMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "langfuse.observation.type": lambda d: "generation",
+        "langfuse.observation.model.name": lambda d: d.request_model or None,
+        "langfuse.observation.metadata.provider": lambda d: d.provider or None,
+        "langfuse.observation.id": lambda d: d.identity.call_id or None,
+        "langfuse.trace.metadata.team_id": lambda d: d.identity.team_id or None,
+        "langfuse.trace.metadata.team_alias": lambda d: d.identity.team_alias or None,
+    }
+
+    # Sub-tables folded into their respective JSON blobs.
+    _MODEL_PARAMS: dict[str, Callable[[LLMRequestParams], AttrValue | None]] = {
+        "temperature": lambda rp: rp.temperature,
+        "top_p": lambda rp: rp.top_p,
+        "max_tokens": lambda rp: rp.max_tokens,
+        "frequency_penalty": lambda rp: rp.frequency_penalty,
+        "presence_penalty": lambda rp: rp.presence_penalty,
+        "seed": lambda rp: rp.seed,
+    }
+    _USAGE_FIELDS: dict[str, Callable[[LLMUsage], AttrValue | None]] = {
+        "input": lambda u: u.input_tokens,
+        "output": lambda u: u.output_tokens,
+        "total": lambda u: u.total_tokens,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "langfuse.observation.model.parameters": lambda d: json_if(
+            collect(LangfuseMapper._MODEL_PARAMS, d.request_params)
+        ),
+        "langfuse.observation.input": lambda d: serialize_messages(d.messages_in),
+        "langfuse.observation.output": lambda d: serialize_messages(output_messages(d)),
+        "langfuse.observation.usage_details": lambda d: json_if(
+            collect(LangfuseMapper._USAGE_FIELDS, d.usage)
+        ),
+        "langfuse.observation.cost_details": lambda d: (
+            json.dumps({"total": d.response_cost})
+            if d.response_cost is not None
+            else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/mappers/langtrace.py

@@ -0,0 +1,64 @@
+"""Langtrace attribute mapper.
+
+Produces Langtrace's attribute vocabulary so a span can be ingested by a
+Langtrace backend. Compose it alongside other mappers like any other
+vocabulary.
+
+Scalar attributes are declared as a flat ``key -> extractor`` table (one lambda
+per mapping operation); the prompt/completion blobs are serialized as a tail.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    json_or_none,
+    output_messages,
+)
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+class LangtraceMapper:
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "gen_ai.operation.name": lambda d: "chat",
+        "langtrace.service.name": lambda d: d.provider or None,
+        "llm.model": lambda d: d.request_model or None,
+        "gen_ai.response.model": lambda d: d.response_model or None,
+        "gen_ai.response_id": lambda d: d.response_id or None,
+        "gen_ai.system_fingerprint": lambda d: d.system_fingerprint or None,
+        "llm.temperature": lambda d: d.request_params.temperature,
+        "llm.top_p": lambda d: d.request_params.top_p,
+        "llm.top_k": lambda d: d.request_params.top_k,
+        "llm.max_tokens": lambda d: d.request_params.max_tokens,
+        "llm.frequency_penalty": lambda d: d.request_params.frequency_penalty,
+        "llm.presence_penalty": lambda d: d.request_params.presence_penalty,
+        "llm.stream": lambda d: d.is_streaming,
+        "llm.token.counts.prompt": lambda d: d.usage.input_tokens,
+        "llm.token.counts.completion": lambda d: d.usage.output_tokens,
+        "llm.token.counts.total": lambda d: d.usage.total_tokens,
+    }
+
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "llm.prompts": lambda d: (
+            json_or_none(list(d.messages_in)) if d.messages_in else None
+        ),
+        "llm.completions": lambda d: (
+            json_or_none(output_messages(d)) if d.choices_out else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/mappers/legacy.py

@@ -0,0 +1,97 @@
+"""Mapper for the older semantic-convention attribute vocabulary.
+
+Emits attributes under the semconv-ai / Traceloop key names (e.g.
+``gen_ai.system``, ``gen_ai.usage.prompt_tokens``, ``llm.is_streaming``) plus a
+few bare, unprefixed service keys (``service``, ``call_type``, ``error``), for
+backends that consume those names.
+
+Like ``GenAIMapper``, each span kind declares its schema as a flat
+``attribute key -> extractor`` table: one lambda per mapping operation.
+"""
+
+from typing import Callable, Final
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, drop_none
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    ServiceSpanData,
+    ToolDefinition,
+)
+
+# Attribute keys in the semconv-ai / Traceloop vocabulary.
+_LEGACY_SYSTEM: Final = "gen_ai.system"
+_LEGACY_PROMPT_TOKENS: Final = "gen_ai.usage.prompt_tokens"
+_LEGACY_COMPLETION_TOKENS: Final = "gen_ai.usage.completion_tokens"
+_LEGACY_TOTAL_TOKENS: Final = "gen_ai.usage.total_tokens"
+_LEGACY_IS_STREAMING: Final = "llm.is_streaming"
+_LEGACY_TOP_K: Final = "llm.top_k"
+_LEGACY_FREQUENCY_PENALTY: Final = "llm.frequency_penalty"
+_LEGACY_PRESENCE_PENALTY: Final = "llm.presence_penalty"
+_LEGACY_STOP_SEQUENCES: Final = "llm.chat.stop_sequences"
+_LEGACY_SERVICE: Final = "service"
+_LEGACY_CALL_TYPE: Final = "call_type"
+_LEGACY_ERROR: Final = "error"
+
+
+class LegacyMapper:
+    """Emits LLM-call and service attributes under the older key names."""
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        _LEGACY_SYSTEM: lambda d: d.provider or None,
+        _LEGACY_PROMPT_TOKENS: lambda d: d.usage.input_tokens,
+        _LEGACY_COMPLETION_TOKENS: lambda d: d.usage.output_tokens,
+        _LEGACY_TOTAL_TOKENS: lambda d: d.usage.total_tokens,
+        _LEGACY_IS_STREAMING: lambda d: d.is_streaming,
+        _LEGACY_TOP_K: lambda d: d.request_params.top_k,
+        _LEGACY_FREQUENCY_PENALTY: lambda d: d.request_params.frequency_penalty,
+        _LEGACY_PRESENCE_PENALTY: lambda d: d.request_params.presence_penalty,
+        _LEGACY_STOP_SEQUENCES: lambda d: (
+            list(d.request_params.stop_sequences)
+            if d.request_params.stop_sequences
+            else None
+        ),
+    }
+
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "name": lambda t: t.name,
+        "description": lambda t: t.description or None,
+        "parameters": lambda t: t.parameters_json or None,
+    }
+
+    _SERVICE_ATTRS: dict[str, Callable[[ServiceSpanData], AttrValue | None]] = {
+        _LEGACY_SERVICE: lambda d: d.service_name,
+        _LEGACY_CALL_TYPE: lambda d: d.call_type,
+        _LEGACY_ERROR: lambda d: (
+            d.error.message if d.error is not None and d.error.message else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case ServiceSpanData():
+                return self._service(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        attrs = collect(cls._LLM_CALL_ATTRS, data)
+        attrs.update(
+            drop_none(
+                {
+                    f"llm.request.functions.{idx}.{suffix}": extract(tool)
+                    for idx, tool in enumerate(data.tools)
+                    for suffix, extract in cls._TOOL_ATTRS.items()
+                }
+            )
+        )
+        return attrs
+
+    @classmethod
+    def _service(cls, data: ServiceSpanData) -> AttributeMap:
+        attrs = collect(cls._SERVICE_ATTRS, data)
+        attrs.update(dict(data.event_metadata))
+        return attrs

litellm/integrations/otel/mappers/openinference.py

@@ -0,0 +1,128 @@
+"""OpenInference attribute mapper (Arize + Arize-Phoenix shared vocabulary).
+
+Spec: https://github.com/Arize-ai/openinference/tree/main/spec — the standard
+both Arize and Phoenix consume. Composing this mapper after ``GenAIMapper``
+gives the same span both vocabularies, so a single trace lights up Arize +
+Phoenix + any other OpenInference-aware backend simultaneously.
+"""
+
+import json
+from typing import Callable, Sequence
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import (
+    collect,
+    drop_none,
+    json_if,
+    message_content,
+    output_messages,
+)
+from litellm.integrations.otel.model.payloads import (
+    LLMCallSpanData,
+    LLMRequestParams,
+    ToolDefinition,
+)
+
+
+class OpenInferenceMapper:
+    """Emits OpenInference attributes for LLM_CALL spans.
+
+    Key families (per the OpenInference spec):
+    - ``openinference.span.kind`` — discriminator (``"LLM"`` here)
+    - ``llm.model_name`` / ``llm.provider`` / ``llm.invocation_parameters``
+    - ``llm.input_messages.{i}.message.role`` / ``...content``
+    - ``llm.output_messages.{i}.message.role`` / ``...content``
+    - ``llm.token_count.prompt`` / ``...completion`` / ``...total``
+    - ``input.value`` / ``output.value`` — JSON-serialized request / response
+    """
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "openinference.span.kind": lambda d: "LLM",
+        "llm.model_name": lambda d: d.request_model or None,
+        "llm.provider": lambda d: d.provider or None,
+        "llm.token_count.prompt": lambda d: d.usage.input_tokens,
+        "llm.token_count.completion": lambda d: d.usage.output_tokens,
+        "llm.token_count.total": lambda d: d.usage.total_tokens,
+    }
+
+    # Folded into the ``llm.invocation_parameters`` JSON blob.
+    _INVOCATION_PARAMS: dict[str, Callable[[LLMRequestParams], AttrValue | None]] = {
+        "temperature": lambda rp: rp.temperature,
+        "top_p": lambda rp: rp.top_p,
+        "top_k": lambda rp: rp.top_k,
+        "max_tokens": lambda rp: rp.max_tokens,
+        "frequency_penalty": lambda rp: rp.frequency_penalty,
+        "presence_penalty": lambda rp: rp.presence_penalty,
+        "seed": lambda rp: rp.seed,
+    }
+
+    # Per-tool extractors, keyed by the ``llm.tools.{idx}.*`` suffix.
+    _TOOL_ATTRS: dict[str, Callable[[ToolDefinition], AttrValue | None]] = {
+        "tool.name": lambda t: t.name,
+        "tool.description": lambda t: t.description or None,
+        "tool.json_schema": lambda t: t.parameters_json or None,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        "llm.invocation_parameters": lambda d: json_if(
+            collect(OpenInferenceMapper._INVOCATION_PARAMS, d.request_params)
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+            **cls._messages("llm.input_messages", "input.value", data.messages_in),
+            **cls._messages(
+                "llm.output_messages", "output.value", output_messages(data)
+            ),
+            **cls._tools(data),
+        }
+
+    @staticmethod
+    def _messages(
+        prefix: str, value_key: str, messages: Sequence[object]
+    ) -> AttributeMap:
+        """Per-message ``{prefix}.{idx}.message.*`` keys + the ``value_key`` blob."""
+        parsed = [
+            (m.get("role") if isinstance(m, dict) else None, message_content(m))
+            for m in messages
+        ]
+        attrs = drop_none(
+            {
+                key: value
+                for idx, (role, content) in enumerate(parsed)
+                for key, value in (
+                    (
+                        f"{prefix}.{idx}.message.role",
+                        role if isinstance(role, str) else None,
+                    ),
+                    (f"{prefix}.{idx}.message.content", content),
+                )
+            }
+        )
+        if parsed:
+            attrs[value_key] = json.dumps(
+                [{"role": role, "content": content} for role, content in parsed]
+            )
+        return attrs
+
+    @classmethod
+    def _tools(cls, data: LLMCallSpanData) -> AttributeMap:
+        return drop_none(
+            {
+                f"llm.tools.{idx}.{suffix}": extract(tool)
+                for idx, tool in enumerate(data.tools)
+                for suffix, extract in cls._TOOL_ATTRS.items()
+            }
+        )

litellm/integrations/otel/mappers/utils.py

@@ -0,0 +1,76 @@
+"""Shared helpers for the attribute mappers.
+
+Small, mapper-agnostic utilities — JSON serialization, message extraction, and
+extractor-table application — pulled out of the individual mapper modules so
+they live in one place.
+"""
+
+import json
+from typing import Callable, Mapping, Sequence
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+def drop_none(values: Mapping[str, AttrValue | None]) -> AttributeMap:
+    """Return ``values`` with ``None``-valued entries removed."""
+    return {k: v for k, v in values.items() if v is not None}
+
+
+def collect(table: Mapping[str, Callable], source: object) -> AttributeMap:
+    """Apply an extractor table to ``source``, dropping ``None`` results."""
+    return drop_none({key: extract(source) for key, extract in table.items()})
+
+
+def json_if(payload: Mapping[str, object]) -> str | None:
+    """JSON-serialize ``payload`` only when it's non-empty; else ``None``."""
+    return json.dumps(payload) if payload else None
+
+
+def json_or_none(value: object) -> str | None:
+    """JSON-serialize ``value`` (falling back to ``str``); ``None`` on failure."""
+    try:
+        return json.dumps(value, default=str)
+    except Exception:
+        return None
+
+
+def stringify_message(message: object) -> str | None:
+    """JSON-serialize a chat message dict; ``None`` if not a dict or on failure."""
+    if not isinstance(message, dict):
+        return None
+    try:
+        return json.dumps(message, default=str)
+    except Exception:
+        return None
+
+
+def serialize_messages(messages: Sequence[object]) -> str | None:
+    """Round-trip a sequence of message dicts through ``stringify_message``."""
+    serialized = [
+        json.loads(s) for s in (stringify_message(m) for m in messages) if s is not None
+    ]
+    return json.dumps(serialized) if serialized else None
+
+
+def message_content(message: object) -> str | None:
+    """Extract the textual ``content`` from a chat message dict."""
+    if not isinstance(message, dict):
+        return None
+    content = message.get("content")
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        # multimodal: concatenate text parts only
+        parts = [
+            part.get("text", "")
+            for part in content
+            if isinstance(part, dict) and part.get("type") == "text"
+        ]
+        return "".join(p for p in parts if isinstance(p, str)) or None
+    return None
+
+
+def output_messages(data: LLMCallSpanData) -> list:
+    """The ``message`` payload of each response choice."""
+    return [c.get("message") for c in data.choices_out if isinstance(c, dict)]

litellm/integrations/otel/mappers/weave.py

@@ -0,0 +1,48 @@
+"""Weave (W&B) attribute mapper.
+
+Weave consumes OpenInference + a small set of Weave-specific keys (display
+name, thread id, output value). This mapper layers the latter on top of
+OpenInference's vocabulary — compose ``["genai", "openinference", "weave"]``
+to feed a Weave backend.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.mappers.base import AttributeMap, AttrValue, SpanData
+from litellm.integrations.otel.mappers.utils import collect, json_or_none
+from litellm.integrations.otel.model.payloads import LLMCallSpanData
+
+
+class WeaveMapper:
+    """Maps ``LLMCallSpanData`` to Weave's vendor attributes."""
+
+    _LLM_CALL_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        # ``display_name`` has the form ``"{operation} {model}"``. The span
+        # name already covers that, but Weave reads this attribute too.
+        "weave.display_name": lambda d: (
+            f"{d.operation.value} {d.request_model}" if d.request_model else None
+        ),
+        "weave.call_id": lambda d: d.identity.call_id or None,
+    }
+
+    # JSON-payload attributes: each builder returns the serialized blob or None.
+    _BLOB_ATTRS: dict[str, Callable[[LLMCallSpanData], AttrValue | None]] = {
+        # Weave treats the response choices as the "output" payload.
+        "weave.output": lambda d: (
+            json_or_none(list(d.choices_out)) if d.choices_out else None
+        ),
+    }
+
+    def map(self, data: SpanData) -> AttributeMap:
+        match data:
+            case LLMCallSpanData():
+                return self._llm_call(data)
+            case _:
+                return {}
+
+    @classmethod
+    def _llm_call(cls, data: LLMCallSpanData) -> AttributeMap:
+        return {
+            **collect(cls._LLM_CALL_ATTRS, data),
+            **collect(cls._BLOB_ATTRS, data),
+        }

litellm/integrations/otel/model/baggage.py

@@ -0,0 +1,112 @@
+"""Baggage promotion: request-identity values carried across child spans.
+
+A bounded set of identity values is written into OpenTelemetry Baggage on the
+LLM-call span so that child spans (guardrail, service) inherit them.
+``providers.LiteLLMBaggageSpanProcessor`` reads Baggage at span start and stamps
+the allowlisted keys onto every span.
+
+This module is the single place baggage is defined: ``_PROMOTABLE`` maps each
+promotable attribute key to how its value is read, and the ``*_KEYS`` defaults
+select what is promoted unless the config overrides them. ``TEAM_METADATA``'s
+extractor filters the team's free-form metadata to the sub-keys an operator
+allowlists via ``baggage_team_metadata_keys`` (default none), so the blob is
+never promoted whole.
+"""
+
+import json
+from collections.abc import Callable, Mapping
+from typing import Final
+
+from litellm.integrations.otel.model.metadata import RequestIdentity
+from litellm.integrations.otel.model.semconv import GenAI, LiteLLM
+
+# Attribute key -> value extractor over (identity, request_model,
+# team_metadata_keys). The single definition of what may be promoted and under
+# which key. Only the ``TEAM_METADATA`` extractor consults team_metadata_keys
+# (to filter the team's metadata to an allowlist); the rest ignore it.
+_PROMOTABLE: Final[
+    dict[str, Callable[[RequestIdentity, str | None, tuple[str, ...]], str | None]]
+] = {
+    LiteLLM.TEAM_ID: lambda identity, model, team_metadata_keys: identity.team_id,
+    LiteLLM.TEAM_ALIAS: lambda identity, model, team_metadata_keys: identity.team_alias,
+    LiteLLM.TEAM_METADATA: lambda identity, model, team_metadata_keys: _filtered_team_metadata_json(
+        identity.team_metadata, team_metadata_keys
+    ),
+    LiteLLM.KEY_HASH: lambda identity, model, team_metadata_keys: identity.key_hash,
+    LiteLLM.END_USER: lambda identity, model, team_metadata_keys: identity.end_user,
+    GenAI.REQUEST_MODEL: lambda identity, model, team_metadata_keys: model,
+    LiteLLM.PROVIDER_MODEL: lambda identity, model, team_metadata_keys: identity.provider_model,
+}
+
+# Keys promoted by default (a subset of ``_PROMOTABLE``). ``END_USER`` is
+# promotable but off by default — it identifies an individual user, so stamping
+# it onto every span is opt-in via ``config.baggage_promoted_keys``.
+BAGGAGE_PROMOTED_KEYS: Final[tuple[str, ...]] = (
+    LiteLLM.TEAM_ID,
+    LiteLLM.TEAM_ALIAS,
+    LiteLLM.TEAM_METADATA,
+    LiteLLM.KEY_HASH,
+    GenAI.REQUEST_MODEL,
+    LiteLLM.PROVIDER_MODEL,
+)
+
+# Metadata sub-keys eligible for promotion under the ``litellm.metadata.*``
+# namespace. The full metadata blob is never promoted; only this allowlist is.
+DEFAULT_BAGGAGE_METADATA_KEYS: Final[tuple[str, ...]] = (
+    "user_api_key_org_id",
+    "user_api_key_user_id",
+    "user_api_key_alias",
+    "user_api_key_end_user_id",
+    "requester_ip_address",
+)
+
+# Sub-keys of the team's free-form metadata eligible for promotion under
+# ``litellm.team.metadata``. Empty by default: a team's metadata can hold
+# arbitrary operator data, so none of it is promoted until each key is
+# explicitly allowlisted via ``config.baggage_team_metadata_keys``.
+DEFAULT_BAGGAGE_TEAM_METADATA_KEYS: Final[tuple[str, ...]] = ()
+
+
+def promoted_baggage(
+    identity: RequestIdentity,
+    request_model: str | None,
+    promoted_keys: tuple[str, ...],
+    metadata_keys: tuple[str, ...] = DEFAULT_BAGGAGE_METADATA_KEYS,
+    team_metadata_keys: tuple[str, ...] = DEFAULT_BAGGAGE_TEAM_METADATA_KEYS,
+) -> dict[str, str]:
+    """Identity values to write into Baggage, filtered to ``promoted_keys``.
+
+    ``promoted_keys`` selects from ``_PROMOTABLE``; ``metadata_keys`` selects
+    sub-keys of ``identity.metadata`` to promote under ``litellm.metadata.*``;
+    ``team_metadata_keys`` selects sub-keys of the team's metadata to promote
+    under ``litellm.team.metadata``. Empty values are dropped.
+    """
+    out: dict[str, str] = {}
+    for key, extract in _PROMOTABLE.items():
+        if key in promoted_keys:
+            value = extract(identity, request_model, team_metadata_keys)
+            if value:
+                out[key] = value
+    for meta_key in metadata_keys:
+        value = identity.metadata.get(meta_key)
+        if value:
+            out[f"{LiteLLM.METADATA_PREFIX}{meta_key}"] = value
+    return out
+
+
+def _filtered_team_metadata_json(
+    metadata: Mapping[str, object] | None,
+    allowed_keys: tuple[str, ...],
+) -> str | None:
+    """JSON-serialize only the allowlisted sub-keys of a team's metadata.
+
+    Returns ``None`` when nothing is allowlisted or no allowlisted key is
+    present, so the empty case is dropped rather than promoting ``"{}"``. Keys
+    are sorted for a stable, diff-friendly value.
+    """
+    if not isinstance(metadata, Mapping) or not allowed_keys:
+        return None
+    filtered = {key: metadata[key] for key in allowed_keys if key in metadata}
+    if not filtered:
+        return None
+    return json.dumps(filtered, default=str, sort_keys=True)

litellm/integrations/otel/model/config.py

@@ -0,0 +1,252 @@
+"""Typed configuration for the OpenTelemetry instrumentation."""
+
+from typing import Any, List
+
+from pydantic import AliasChoices, BaseModel, Field, field_validator, model_validator
+from pydantic_settings import BaseSettings, NoDecode, SettingsConfigDict
+from typing_extensions import Annotated
+
+from litellm.integrations.otel.model.baggage import (
+    BAGGAGE_PROMOTED_KEYS,
+    DEFAULT_BAGGAGE_METADATA_KEYS,
+    DEFAULT_BAGGAGE_TEAM_METADATA_KEYS,
+)
+
+#: Master feature-flag env var. The logger is inert until this is truthy.
+OTEL_V2_ENV = "LITELLM_OTEL_V2"
+
+
+class CaptureMessageContent(str):
+    NO_CONTENT = "no_content"
+    SPAN_ONLY = "span_only"
+    EVENT_ONLY = "event_only"
+    SPAN_AND_EVENT = "span_and_event"
+
+
+class _OTelV2Flag(BaseSettings):
+    model_config = SettingsConfigDict(extra="ignore")
+
+    enabled: bool = Field(default=False, validation_alias=AliasChoices(OTEL_V2_ENV))
+
+
+def is_otel_v2_enabled() -> bool:
+    return _OTelV2Flag().enabled
+
+
+class ExporterSpec(BaseModel):
+    """One span-export destination.
+
+    The shared ``TracerProvider`` attaches one ``SpanProcessor`` per spec, so
+    listing several specs sends every span to all of them at once (e.g. Arize +
+    Phoenix + your own Honeycomb).
+    """
+
+    model_config = {"extra": "forbid"}
+
+    kind: str = Field(
+        default="console",
+        description="console | in_memory | otlp_http | otlp_grpc | <factory kind>",
+    )
+    endpoint: str | None = None
+    headers: str | None = None
+    options: dict[str, str] | None = Field(
+        default=None,
+        description=(
+            "Factory-specific configuration for a custom exporter ``kind`` "
+            "registered via ``providers.register_exporter_factory`` (e.g. an "
+            "API key a lazy-auth exporter fetches a token with). Ignored by the "
+            "built-in console/in_memory/otlp exporters."
+        ),
+    )
+    use_simple_processor: bool | None = Field(
+        default=None,
+        description=(
+            "Force SimpleSpanProcessor regardless of exporter kind. Default: "
+            "auto (Simple for console/in_memory, Batch otherwise)."
+        ),
+    )
+
+
+class OpenTelemetryV2Config(BaseSettings):
+    model_config = SettingsConfigDict(populate_by_name=True, extra="ignore")
+
+    # ----- single-destination shorthand, read from standard OTEL_* envs ----- #
+    exporter: str = Field(
+        default="console",
+        validation_alias=AliasChoices("OTEL_EXPORTER", "OTEL_EXPORTER_OTLP_PROTOCOL"),
+        description=(
+            "Exporter kind for the single-destination shorthand. The model "
+            "validator folds this (with ``endpoint`` / ``headers``) into a "
+            "one-entry ``exporters`` list when ``exporters`` is empty; set "
+            "``exporters`` directly for multiple destinations."
+        ),
+    )
+    endpoint: str | None = Field(
+        default=None,
+        validation_alias=AliasChoices("OTEL_ENDPOINT", "OTEL_EXPORTER_OTLP_ENDPOINT"),
+    )
+    headers: str | None = Field(
+        default=None,
+        validation_alias=AliasChoices("OTEL_HEADERS", "OTEL_EXPORTER_OTLP_HEADERS"),
+    )
+    service_name: str = Field(
+        default="litellm", validation_alias=AliasChoices("OTEL_SERVICE_NAME")
+    )
+    deployment_environment: str | None = Field(
+        default=None, validation_alias=AliasChoices("OTEL_ENVIRONMENT_NAME")
+    )
+
+    enable_metrics: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("LITELLM_OTEL_INTEGRATION_ENABLE_METRICS"),
+    )
+    enable_events: bool = Field(
+        default=False,
+        validation_alias=AliasChoices("LITELLM_OTEL_INTEGRATION_ENABLE_EVENTS"),
+    )
+    capture_message_content: str = Field(
+        default=CaptureMessageContent.NO_CONTENT,
+        validation_alias=AliasChoices(
+            "OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT"
+        ),
+    )
+    legacy_compat: bool = Field(
+        default=True, validation_alias=AliasChoices("LITELLM_OTEL_LEGACY_COMPAT")
+    )
+
+    # ----- explicit multi-destination / vocabulary configuration ------------ #
+
+    exporters: list[ExporterSpec] = Field(
+        default_factory=list,
+        description=(
+            "One destination per spec. The shared TracerProvider attaches a "
+            "SpanProcessor per entry. When empty, the model validator folds "
+            "the ``exporter`` / ``endpoint`` / ``headers`` shorthand into a "
+            "single spec so there is always at least one destination."
+        ),
+    )
+
+    mapper_names: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: ["genai"],
+        description=(
+            "Ordered attribute vocabularies to emit. ``genai`` is the "
+            "canonical OTel GenAI vocabulary and is always placed first. "
+            "Vendor names: ``openinference`` (Arize + Phoenix), ``langfuse``, "
+            "``weave``, ``langtrace``."
+        ),
+    )
+
+    resource_attributes: dict[str, str] = Field(
+        default_factory=dict,
+        description=(
+            "Extra Resource attributes beyond ``service.name`` and "
+            "``deployment.environment`` (e.g. integration-specific markers)."
+        ),
+    )
+
+    baggage_promoted_keys: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: list(BAGGAGE_PROMOTED_KEYS),
+        validation_alias=AliasChoices(
+            "baggage_promoted_keys", "LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS"
+        ),
+        description=(
+            "Identity attribute keys written into Baggage and stamped on every "
+            "child span (e.g. ``litellm.team.id``). Configure via the "
+            "``LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS`` env var (comma-separated) or "
+            "``callback_settings.otel.baggage_promoted_keys`` in config.yaml (a "
+            "YAML list)."
+        ),
+    )
+    baggage_metadata_keys: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: list(DEFAULT_BAGGAGE_METADATA_KEYS),
+        validation_alias=AliasChoices(
+            "baggage_metadata_keys", "LITELLM_OTEL_BAGGAGE_METADATA_KEYS"
+        ),
+        description=(
+            "Metadata sub-keys promoted under the ``litellm.metadata.*`` "
+            "namespace. Configure via the ``LITELLM_OTEL_BAGGAGE_METADATA_KEYS`` "
+            "env var (comma-separated) or "
+            "``callback_settings.otel.baggage_metadata_keys`` in config.yaml."
+        ),
+    )
+    baggage_team_metadata_keys: Annotated[List[str], NoDecode] = Field(
+        default_factory=lambda: list(DEFAULT_BAGGAGE_TEAM_METADATA_KEYS),
+        validation_alias=AliasChoices(
+            "baggage_team_metadata_keys", "LITELLM_OTEL_BAGGAGE_TEAM_METADATA_KEYS"
+        ),
+        description=(
+            "Sub-keys of the team's free-form metadata promoted under "
+            "``litellm.team.metadata``. Empty by default so none of a team's "
+            "metadata leaves the process until explicitly allowlisted. Configure "
+            "via the ``LITELLM_OTEL_BAGGAGE_TEAM_METADATA_KEYS`` env var "
+            "(comma-separated) or "
+            "``callback_settings.otel.baggage_team_metadata_keys`` in config.yaml."
+        ),
+    )
+
+    @field_validator(
+        "baggage_promoted_keys",
+        "baggage_metadata_keys",
+        "baggage_team_metadata_keys",
+        "mapper_names",
+        mode="before",
+    )
+    @classmethod
+    def _split_csv(cls, value: Any) -> Any:
+        """Accept a comma-separated string for list fields.
+
+        Env vars are strings, but these fields are lists. Pydantic-settings would
+        otherwise require JSON for a list env var; splitting on commas here lets
+        an operator write ``LITELLM_OTEL_BAGGAGE_PROMOTED_KEYS=litellm.team.id,litellm.api_key.hash``.
+        YAML lists (from ``callback_settings.otel.*``) and real lists pass through
+        unchanged.
+        """
+        if isinstance(value, str):
+            return [item.strip() for item in value.split(",") if item.strip()]
+        return value
+
+    @model_validator(mode="after")
+    def _normalize(self) -> "OpenTelemetryV2Config":
+        # An endpoint with the default exporter kind implies OTLP/HTTP.
+        if self.endpoint and self.exporter == "console":
+            self.exporter = "otlp_http"
+        # When no explicit destinations are given, fold the single-destination
+        # shorthand into one spec so the provider always has a destination.
+        if not self.exporters:
+            self.exporters = [
+                ExporterSpec(
+                    kind=self.exporter,
+                    endpoint=self.endpoint,
+                    headers=self.headers,
+                )
+            ]
+        # Ensure ``genai`` is always present and first.
+        names = list(self.mapper_names)
+        if "genai" in names:
+            names = ["genai"] + [n for n in names if n != "genai"]
+        else:
+            names = ["genai"] + names
+        # When enabled, also emit attribute keys under their semconv-ai /
+        # Traceloop names via the ``legacy`` mapper. Append it at the tail so
+        # the canonical ``genai`` keys win on any conflict.
+        if self.legacy_compat and "legacy" not in names:
+            names.append("legacy")
+        self.mapper_names = names
+        return self
+
+    @property
+    def capture_span_content(self) -> bool:
+        """Whether prompt/response content may be stamped as span attributes.
+
+        Defaults off (``no_content``): an operator must opt in before message
+        bodies leave the process, so a user request can never force its prompt
+        or completion into the configured backend while capture is disabled.
+        """
+        return self.capture_message_content in (
+            CaptureMessageContent.SPAN_ONLY,
+            CaptureMessageContent.SPAN_AND_EVENT,
+        )
+
+    @classmethod
+    def from_env(cls) -> "OpenTelemetryV2Config":
+        return cls()

litellm/integrations/otel/model/metadata.py

@@ -0,0 +1,294 @@
+"""The single translation layer between a request's metadata and the spans.
+
+Every relevant field litellm exposes about a request — the user-facing model,
+the model actually dispatched to the provider, the deployment, and the caller's
+identity (team, key, end-user) — is parsed **once**, here, out of the
+``StandardLoggingPayload`` (or a ``UserAPIKeyAuth`` at the auth boundary). Span
+data, baggage promotion, and the mappers then read these typed fields instead of
+each digging into the raw ``metadata`` / ``hidden_params`` dicts.
+
+Two models live here because a request's identity is known *before* its model
+resolution is:
+
+* :class:`RequestIdentity` — team / key / end-user, seeded into Baggage at the
+  auth boundary (``from_user_api_key_auth``), before routing has picked a
+  deployment. ``provider_model`` is therefore absent from that early seed and is
+  only filled in from the payload once the call closes.
+* :class:`RequestContext` — the full picture available at close: the resolved
+  request vs. provider model split, plus the response model, model group, model
+  id, and api base, wrapping the :class:`RequestIdentity`.
+
+The request-vs-provider model split is the subtle part. On the proxy a caller
+asks for a *model group* (e.g. ``gpt-4o``) that routes to a concrete deployment
+(e.g. ``azure/my-deployment``); the two are distinct and both worth recording.
+``StandardLoggingPayload`` exposes them as:
+
+* ``model_group`` — the user-facing name the caller requested.
+* ``model`` — already reconstructed (see ``reconstruct_model_name``) to the name
+  litellm dispatched to the provider (the deployment, provider-prefixed).
+* ``hidden_params.litellm_model_name`` — a secondary source for the dispatched
+  model (populated only on some call paths, e.g. files).
+
+So ``gen_ai.request.model`` is the *group* (falling back to the call model on the
+SDK path, which has no group), and ``litellm.provider.model`` is the *dispatched*
+model. They coincide on the SDK path, which is correct.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any, Mapping, cast
+
+from litellm.constants import LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL
+from litellm.integrations.otel.model.semconv import resolve_operation
+from litellm.integrations.otel.model.utils import as_str
+
+if TYPE_CHECKING:
+    from litellm.types.utils import StandardLoggingPayload
+
+
+@dataclass(frozen=True)
+class RequestIdentity:
+    call_id: str | None = None
+    team_id: str | None = None
+    team_alias: str | None = None
+    # The team's free-form metadata, carried raw (empty/missing -> None) and
+    # filtered to an operator allowlist only at Baggage-promotion time, so an
+    # unconfigured deployment never promotes any of it.
+    team_metadata: Mapping[str, Any] | None = None
+    key_hash: str | None = None
+    end_user: str | None = None
+    # The model litellm dispatched to the provider. Only known once the call
+    # completes (routing has picked a deployment), so it's absent from the
+    # auth-time seed and filled only from the payload.
+    provider_model: str | None = None
+    metadata: Mapping[str, str] = field(default_factory=dict)
+
+    @classmethod
+    def from_payload(cls, payload: "StandardLoggingPayload") -> "RequestIdentity":
+        """Parse caller identity out of a closed request's payload metadata.
+
+        ``provider_model`` is resolved here too (see :func:`resolve_provider_model`)
+        so the identity carried into Baggage labels every span with the dispatched
+        model, not just the user-facing one.
+        """
+        raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+        metadata = {
+            key: str(value)
+            for key, value in raw_meta.items()
+            if isinstance(value, (str, bool, int, float))
+        }
+        return cls(
+            call_id=as_str(payload.get("litellm_call_id")) or as_str(payload.get("id")),
+            # StandardLoggingMetadata's canonical key is ``user_api_key_team_id``;
+            # the bare ``team_id`` is a legacy alias and is often empty, so prefer
+            # the canonical key and fall back to the alias.
+            team_id=as_str(raw_meta.get("user_api_key_team_id"))
+            or as_str(raw_meta.get("team_id")),
+            team_alias=as_str(raw_meta.get("user_api_key_team_alias"))
+            or as_str(raw_meta.get("team_alias")),
+            team_metadata=_team_metadata_dict(
+                raw_meta.get("user_api_key_team_metadata")
+            ),
+            key_hash=as_str(raw_meta.get("user_api_key_hash")),
+            end_user=as_str(payload.get("end_user"))
+            or as_str(raw_meta.get("user_api_key_end_user_id")),
+            provider_model=resolve_provider_model(payload),
+            metadata=metadata,
+        )
+
+    @classmethod
+    def from_user_api_key_auth(cls, auth: object) -> "RequestIdentity":
+        """Identity from a ``UserAPIKeyAuth`` (duck-typed to keep this module
+        free of a proxy import).
+
+        Used in the pre-call hook to seed Baggage early — before any LLM,
+        guardrail, or service span is created — so the whole request's spans
+        inherit identity, not just the LLM-call span. Metadata sub-keys use the
+        ``user_api_key_*`` names that ``baggage.DEFAULT_BAGGAGE_METADATA_KEYS``
+        promotes.
+        """
+        get = lambda name: getattr(auth, name, None)  # noqa: E731
+        metadata = {
+            meta_key: str(value)
+            for meta_key, attr in (
+                ("user_api_key_user_id", "user_id"),
+                ("user_api_key_org_id", "org_id"),
+                ("user_api_key_alias", "key_alias"),
+                ("user_api_key_end_user_id", "end_user_id"),
+            )
+            if (value := get(attr))
+        }
+        return cls(
+            team_id=as_str(get("team_id")),
+            team_alias=as_str(get("team_alias")),
+            team_metadata=_team_metadata_dict(get("team_metadata")),
+            key_hash=as_str(get("api_key")),
+            end_user=as_str(get("end_user_id")),
+            # ``provider_model`` is unknown at the auth boundary — routing hasn't
+            # picked a deployment yet — so it's only populated from the payload.
+            metadata=metadata,
+        )
+
+
+@dataclass(frozen=True)
+class RequestContext:
+    """The fully-resolved view of a closed request, parsed once from the payload.
+
+    ``request_model`` is the user-facing requested model and ``provider_model``
+    (on :attr:`identity`) is the model litellm dispatched to the provider; the two
+    differ on the proxy (group vs. deployment) and coincide on the SDK path.
+    """
+
+    request_model: str
+    response_model: str | None
+    model_group: str | None
+    model_id: str | None
+    api_base: str | None
+    identity: RequestIdentity
+
+    @property
+    def provider_model(self) -> str | None:
+        """The dispatched-model name, carried on the identity for Baggage."""
+        return self.identity.provider_model
+
+    @classmethod
+    def from_standard_logging_payload(
+        cls, payload: "StandardLoggingPayload"
+    ) -> "RequestContext":
+        raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+        hidden = cast(Mapping[str, object], payload.get("hidden_params") or {})
+        raw_response = payload.get("response")
+        response = cast(
+            Mapping[str, object], raw_response if isinstance(raw_response, dict) else {}
+        )
+        model_group = as_str(payload.get("model_group")) or as_str(
+            raw_meta.get("model_group")
+        )
+        return cls(
+            # The user asked for the group; fall back to the call model on the SDK
+            # path, which has no group. Empty string (never None) so the span name
+            # builder and the mapper see a plain string.
+            request_model=model_group or as_str(payload.get("model")) or "",
+            response_model=as_str(response.get("model")),
+            model_group=model_group,
+            model_id=as_str(payload.get("model_id"))
+            or _model_info_id(raw_meta.get("model_info")),
+            api_base=as_str(payload.get("api_base")) or as_str(hidden.get("api_base")),
+            identity=RequestIdentity.from_payload(payload),
+        )
+
+
+# --- live-callback kwargs parsing ------------------------------------------- #
+#
+# The model and helpers below parse the *live* callback ``kwargs`` god object (and
+# the raw pre/post-call ``data`` dicts) — the untyped request state that reaches a
+# ``CustomLogger`` before, or instead of, a ``StandardLoggingPayload``. They live
+# here, with the payload/auth parsers, so every read out of a request's raw dicts
+# is in one place rather than scattered across the ``CustomLogger``.
+
+
+@dataclass(frozen=True)
+class LLMCallEvent:
+    """The typed view of the live callback ``kwargs`` (``model_call_details``).
+
+    litellm hands every callback an untyped ``kwargs`` god object. The fields the
+    OTel logger needs out of it are parsed **once**, here, so the ``CustomLogger``
+    reads typed attributes instead of digging into the dict at each boundary.
+    """
+
+    # The ``litellm_call_id`` correlating ``pre_call`` with the close callback.
+    # Present in ``model_call_details`` at ``pre_call`` and in both the kwargs and
+    # the ``standard_logging_object`` at success/failure, so it's a stable key for
+    # the open-call carrier — no back-reference to the logging object required (the
+    # object isn't reachable from the callback kwargs at ``pre_call`` time).
+    call_id: str | None
+    # The ``StandardLoggingPayload`` carried on a success/failure callback; ``None``
+    # at ``pre_call``, or when the call closed before any payload materialized (so
+    # there is nothing to stamp on the span).
+    payload: "StandardLoggingPayload | None"
+    # The ``standard_callback_dynamic_params`` routing the call to a per-tenant
+    # tracer (its own exporter/endpoint), or ``None`` when the call isn't scoped.
+    dynamic_params: Any
+    # True for synthetic proxy-gate logs (auth / rate-limit rejections): they fire
+    # the ``pre_call`` hook but never made an upstream call, so they get no span.
+    is_no_upstream_call: bool
+    # A best-effort ``"{operation} {model}"`` name known at ``pre_call`` time. The
+    # span is renamed from the typed payload at close (``finish_span``); this only
+    # needs to be reasonable for a span that never gets closed (a leak).
+    provisional_span_name: str
+
+    @classmethod
+    def from_dict(cls, kwargs: Mapping[str, Any]) -> "LLMCallEvent":
+        raw_payload = kwargs.get("standard_logging_object")
+        payload = cast("StandardLoggingPayload", raw_payload) if raw_payload else None
+        operation = resolve_operation(as_str(kwargs.get("call_type")))
+        model = as_str(kwargs.get("model")) or ""
+        return cls(
+            call_id=_call_id(payload, kwargs),
+            payload=payload,
+            dynamic_params=kwargs.get("standard_callback_dynamic_params"),
+            is_no_upstream_call=bool(kwargs.get(LITELLM_LOGGING_NO_UPSTREAM_LLM_CALL)),
+            provisional_span_name=f"{operation.value} {model}".strip(),
+        )
+
+
+def _call_id(
+    payload: "StandardLoggingPayload | None", kwargs: Mapping[str, Any]
+) -> str | None:
+    """The call id from the payload (when closed) or the bare kwargs (at pre_call)."""
+    if payload is not None:
+        call_id = as_str(payload.get("litellm_call_id")) or as_str(payload.get("id"))
+        if call_id:
+            return call_id
+    return as_str(kwargs.get("litellm_call_id"))
+
+
+def model_from_request_data(data: object) -> str | None:
+    """The user-facing ``model`` from a pre-call ``data`` dict (``None`` if absent).
+
+    Read at the auth boundary to label early Baggage before routing has resolved
+    a deployment; ``data`` is duck-typed since it arrives untyped from the proxy.
+    """
+    if isinstance(data, Mapping):
+        return as_str(data.get("model"))
+    return None
+
+
+def resolve_provider_model(payload: "StandardLoggingPayload") -> str | None:
+    """The model litellm dispatched to the provider, from the payload.
+
+    Prefers the explicit ``hidden_params.litellm_model_name`` (set on call paths
+    that know it, e.g. files), then the top-level ``model`` — which
+    ``reconstruct_model_name`` has already resolved to the deployment's
+    provider-prefixed name. Returns ``None`` only when neither is present.
+    """
+    raw_meta = cast(Mapping[str, object], payload.get("metadata") or {})
+    hidden = cast(Mapping[str, object], payload.get("hidden_params") or {})
+    return (
+        # ``deployment`` survives only on paths that don't strip it from metadata;
+        # harmless (and most precise) to prefer it when present.
+        as_str(raw_meta.get("deployment"))
+        or as_str(hidden.get("litellm_model_name"))
+        or as_str(payload.get("model"))
+    )
+
+
+def _model_info_id(model_info: object) -> str | None:
+    """The deployment id from a ``metadata.model_info`` sub-dict, if present."""
+    if isinstance(model_info, Mapping):
+        return as_str(model_info.get("id"))
+    return None
+
+
+def _team_metadata_dict(value: object) -> Mapping[str, Any] | None:
+    """The team's free-form metadata as a raw mapping, or ``None`` when missing
+    or empty.
+
+    Carried raw on the identity and filtered to an operator allowlist only at
+    Baggage-promotion time (see ``baggage.promoted_baggage``), so an empty case
+    is dropped rather than carrying a useless ``{}``.
+    """
+    if isinstance(value, Mapping) and value:
+        return dict(value)
+    return None

litellm/integrations/otel/model/payloads.py

@@ -0,0 +1,542 @@
+"""Typed span-data inputs: frozen dataclasses the engine and mappers consume."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from enum import Enum
+from typing import TYPE_CHECKING, ClassVar, Mapping, cast
+from urllib.parse import urlsplit
+
+from litellm.integrations.otel.model.metadata import (
+    RequestContext,
+    RequestIdentity,
+)
+from litellm.integrations.otel.model.semconv import (
+    GenAIOperation,
+    MCPMethod,
+    resolve_operation,
+    resolve_provider,
+)
+from litellm.integrations.otel.model.utils import (
+    as_bool,
+    as_float,
+    as_int,
+    as_str,
+    as_str_tuple,
+)
+
+# ``RequestIdentity`` and the request-metadata translation now live in
+# :mod:`metadata`; re-exported here so existing ``model.payloads`` imports keep
+# resolving it.
+__all__ = [
+    "RequestContext",
+    "RequestIdentity",
+    "GuardrailSpanData",
+    "LLMCallSpanData",
+    "LLMRequestParams",
+    "LLMUsage",
+    "MCPToolCallSpanData",
+    "ProxyRequestSpanData",
+    "ServerInfo",
+    "ServiceSpanData",
+    "SpanError",
+    "ToolDefinition",
+    "is_mcp_tool_call",
+]
+
+if TYPE_CHECKING:
+    from litellm.types.services import ServiceLoggerPayload
+    from litellm.types.utils import (
+        StandardLoggingGuardrailInformation,
+        StandardLoggingPayload,
+    )
+
+
+# --- typed sub-structures ---------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class LLMRequestParams:
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    max_tokens: int | None = None
+    frequency_penalty: float | None = None
+    presence_penalty: float | None = None
+    stop_sequences: tuple[str, ...] | None = None
+    seed: int | None = None
+
+    @classmethod
+    def from_model_parameters(cls, params: Mapping[str, object]) -> "LLMRequestParams":
+        max_tokens = as_int(params.get("max_tokens"))
+        if max_tokens is None:
+            max_tokens = as_int(params.get("max_completion_tokens"))
+        return cls(
+            temperature=as_float(params.get("temperature")),
+            top_p=as_float(params.get("top_p")),
+            top_k=as_int(params.get("top_k")),
+            max_tokens=max_tokens,
+            frequency_penalty=as_float(params.get("frequency_penalty")),
+            presence_penalty=as_float(params.get("presence_penalty")),
+            stop_sequences=as_str_tuple(params.get("stop")),
+            seed=as_int(params.get("seed")),
+        )
+
+
+@dataclass(frozen=True)
+class LLMUsage:
+    input_tokens: int | None = None
+    output_tokens: int | None = None
+    total_tokens: int | None = None
+
+
+@dataclass(frozen=True)
+class SpanError:
+    error_type: str | None = None
+    message: str | None = None
+
+
+@dataclass(frozen=True)
+class ServerInfo:
+    address: str | None = None
+    port: int | None = None
+
+    @classmethod
+    def from_api_base(cls, api_base: str | None) -> ServerInfo | None:
+        if not api_base:
+            return None
+        parsed = urlsplit(api_base if "://" in api_base else f"//{api_base}")
+        if not parsed.hostname:
+            return None
+        return cls(address=parsed.hostname, port=parsed.port)
+
+
+@dataclass(frozen=True)
+class GuardrailSpanData:
+    guardrail_name: str
+    mode: str | None = None
+    status: str | None = None
+    masked_entity_count: int | None = None
+    provider: str | None = None
+    action: str | None = None
+    # The guardrail verdict / provider response (e.g. the moderation result),
+    # JSON-serialized. This is the detail that belongs on the guardrail span.
+    response_json: str | None = None
+    violation_categories: tuple[str, ...] = ()
+    confidence_score: float | None = None
+    risk_score: float | None = None
+    duration: float | None = None
+    # Actual execution window (epoch seconds) from the logging entry, so the span
+    # is placed when the guardrail really ran — a pre_call guardrail before the
+    # LLM call — rather than at post-call emission time.
+    start_time: float | None = None
+    end_time: float | None = None
+    # Provider-agnostic configuration/detection metadata (see
+    # ``StandardLoggingGuardrailInformation``). Present for any guardrail that
+    # populates them, not just one provider's shape.
+    guardrail_id: str | None = None
+    policy_template: str | None = None
+    detection_method: str | None = None
+    # Set when the guardrail intervened/blocked or failed, so the emitter marks
+    # the span ERROR — a blocking guardrail is an error outcome for that span.
+    error: SpanError | None = None
+
+    # Guardrail statuses that mean the guardrail did not pass the request through.
+    _ERROR_STATUSES: ClassVar[frozenset[str]] = frozenset(
+        {"guardrail_intervened", "guardrail_failed_to_respond"}
+    )
+
+    @classmethod
+    def from_logging_entry(
+        cls, entry: "StandardLoggingGuardrailInformation"
+    ) -> "GuardrailSpanData":
+        """Build from one ``standard_logging_guardrail_information`` entry.
+
+        Reads the canonical, provider-agnostic ``StandardLoggingGuardrailInformation``
+        keys only — no guessing at a single provider's field names. Values that are
+        typed as enums or lists (e.g. ``guardrail_mode``) are normalized to a
+        stable string rather than assumed to already be plain strings.
+        """
+        get = cast(Mapping[str, object], entry).get
+        status = as_str(get("guardrail_status"))
+        response = get("guardrail_response")
+        error = (
+            SpanError(error_type=status, message=as_str(get("guardrail_action")))
+            if status in cls._ERROR_STATUSES
+            else None
+        )
+        return cls(
+            guardrail_name=as_str(get("guardrail_name")) or "guardrail",
+            mode=_guardrail_mode_str(get("guardrail_mode")),
+            status=status,
+            masked_entity_count=_total_masked_entities(get("masked_entity_count")),
+            provider=as_str(get("guardrail_provider")),
+            action=as_str(get("guardrail_action")),
+            response_json=_json_or_none(response) if response is not None else None,
+            violation_categories=as_str_tuple(get("violation_categories")) or (),
+            confidence_score=as_float(get("confidence_score")),
+            risk_score=as_float(get("risk_score")),
+            duration=as_float(get("duration")),
+            start_time=as_float(get("start_time")),
+            end_time=as_float(get("end_time")),
+            guardrail_id=as_str(get("guardrail_id")),
+            policy_template=as_str(get("policy_template")),
+            detection_method=as_str(get("detection_method")),
+            error=error,
+        )
+
+
+@dataclass(frozen=True)
+class ServiceSpanData:
+    service_name: str
+    call_type: str | None = None
+    error: SpanError | None = None
+    # Caller-supplied attributes to stamp on the service span, passed through
+    # from ``async_service_*_hook(event_metadata=...)``. The mapper owns how
+    # these are namespaced: the canonical vocabulary uses ``litellm.metadata.*``
+    # keys, the semconv-ai / Traceloop vocabulary uses the bare key names.
+    event_metadata: Mapping[str, str] = field(default_factory=dict)
+
+    @classmethod
+    def from_payload(
+        cls,
+        payload: "ServiceLoggerPayload",
+        event_metadata: Mapping[str, object] | None = None,
+    ) -> "ServiceSpanData":
+        # ``payload.service`` is a ``ServiceTypes(str, Enum)`` and ``error`` is
+        # ``Optional[str]`` on the Pydantic model — no defensive reads needed.
+        # ``event_metadata`` is sanitized: the legacy service decorators pass raw
+        # call-site data (live objects, full request metadata, response headers),
+        # none of which belongs on a span.
+        return cls(
+            service_name=payload.service.value,
+            call_type=payload.call_type,
+            error=SpanError(message=payload.error) if payload.error else None,
+            event_metadata=sanitize_event_metadata(event_metadata),
+        )
+
+
+@dataclass(frozen=True)
+class ProxyRequestSpanData:
+    http_method: str
+    route: str
+    url_path: str | None = None
+    status_code: int | None = None
+    identity: RequestIdentity | None = None
+
+
+# --- the primary LLM-call model ---------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class ToolDefinition:
+    """A single function/tool declared on a chat-completion request."""
+
+    name: str
+    description: str | None = None
+    parameters_json: str | None = (
+        None  # JSON-serialized schema (str so it's an AttrValue)
+    )
+
+
+@dataclass(frozen=True)
+class LLMCallSpanData:
+    operation: GenAIOperation
+    provider: str
+    request_model: str
+    response_model: str | None
+    response_id: str | None
+    request_params: LLMRequestParams
+    usage: LLMUsage
+    finish_reasons: tuple[str, ...]
+    error: SpanError | None
+    response_cost: float | None
+    server: ServerInfo | None
+    identity: RequestIdentity
+    is_streaming: bool | None = None
+    tools: tuple[ToolDefinition, ...] = ()
+    # Raw messages and response, needed by vendor mappers (OpenInference,
+    # Langfuse, Weave) that stamp message-level attributes. ``messages_in`` is
+    # the request payload; ``choices_out`` mirrors ``response.choices`` from
+    # the StandardLoggingPayload. Both are tuples of immutable mappings so the
+    # dataclass stays hashable and frozen.
+    messages_in: tuple[Mapping[str, object], ...] = ()
+    choices_out: tuple[Mapping[str, object], ...] = ()
+    system_fingerprint: str | None = None
+
+    @classmethod
+    def from_standard_logging_payload(
+        cls, payload: "StandardLoggingPayload", capture_content: bool = False
+    ) -> "LLMCallSpanData":
+        params = cast(Mapping[str, object], payload.get("model_parameters") or {})
+        # The single parse of the request's metadata — the request-vs-provider
+        # model split, the response model, api base, and identity all come from
+        # here rather than being re-derived from the raw payload dicts.
+        context = RequestContext.from_standard_logging_payload(payload)
+        # Normalize ``response`` to a dict once so the content/id reads below are a
+        # plain ``.get`` — no repeated ``isinstance`` guards.
+        raw_response = payload.get("response")
+        response = cast(
+            Mapping[str, object], raw_response if isinstance(raw_response, dict) else {}
+        )
+        choices_out = _dicts(response.get("choices"))
+        # ``finish_reasons`` is metadata, not content, so derive it from
+        # ``choices_out`` before gating. The raw message/choice bodies are only
+        # retained when content capture is enabled (see ``capture_span_content``);
+        # otherwise the content-bearing mappers receive empty sequences and emit
+        # no prompt/response text.
+        finish_reasons = _finish_reasons(choices_out)
+        return cls(
+            operation=resolve_operation(as_str(payload.get("call_type"))),
+            provider=resolve_provider(as_str(payload.get("custom_llm_provider"))),
+            request_model=context.request_model,
+            response_model=context.response_model,
+            response_id=as_str(response.get("id")),
+            request_params=LLMRequestParams.from_model_parameters(params),
+            usage=LLMUsage(
+                input_tokens=as_int(payload.get("prompt_tokens")),
+                output_tokens=as_int(payload.get("completion_tokens")),
+                total_tokens=as_int(payload.get("total_tokens")),
+            ),
+            finish_reasons=finish_reasons,
+            error=_parse_error(payload),
+            response_cost=as_float(payload.get("response_cost")),
+            server=ServerInfo.from_api_base(context.api_base),
+            identity=context.identity,
+            is_streaming=as_bool(payload.get("stream")),
+            tools=_extract_tools(params),
+            messages_in=_dicts(payload.get("messages")) if capture_content else (),
+            choices_out=choices_out if capture_content else (),
+            system_fingerprint=as_str(response.get("system_fingerprint")),
+        )
+
+
+# --- the MCP tool-call model ------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class MCPToolCallSpanData:
+    """One MCP ``tools/call`` execution, parsed from a closed request's payload.
+
+    The proxy is an MCP *client* to the upstream server it forwards the call to,
+    so this is a CLIENT span. ``arguments_json``/``result_json`` are the tool's
+    input/output — sensitive content, so they're only retained when content
+    capture is enabled, mirroring ``LLMCallSpanData``'s message bodies.
+    """
+
+    operation: GenAIOperation
+    method: str
+    tool_name: str
+    server_name: str | None
+    session_id: str | None
+    arguments_json: str | None
+    result_json: str | None
+    error: SpanError | None
+    response_cost: float | None
+    identity: RequestIdentity
+
+    @classmethod
+    def from_standard_logging_payload(
+        cls, payload: "StandardLoggingPayload", capture_content: bool = False
+    ) -> "MCPToolCallSpanData":
+        meta = _mcp_tool_call_metadata(cast(Mapping[str, object], payload))
+        return cls(
+            operation=resolve_operation(as_str(payload.get("call_type"))),
+            method=MCPMethod.TOOLS_CALL.value,
+            tool_name=as_str(meta.get("name")) or "",
+            server_name=as_str(meta.get("mcp_server_name")),
+            session_id=as_str(meta.get("mcp_session_id")),
+            arguments_json=(
+                _json_or_none(meta.get("arguments"))
+                if capture_content and meta.get("arguments") is not None
+                else None
+            ),
+            result_json=(
+                _json_or_none(meta.get("result"))
+                if capture_content and meta.get("result") is not None
+                else None
+            ),
+            error=_parse_error(payload),
+            response_cost=as_float(payload.get("response_cost")),
+            identity=RequestContext.from_standard_logging_payload(payload).identity,
+        )
+
+
+def _mcp_tool_call_metadata(payload: Mapping[str, object]) -> Mapping[str, object]:
+    """The MCP gateway's tool-call metadata, which lives under
+    ``StandardLoggingPayload.metadata`` (a ``StandardLoggingMetadata`` key), not
+    at the payload's top level."""
+    metadata = payload.get("metadata")
+    if not isinstance(metadata, Mapping):
+        return {}
+    meta = metadata.get("mcp_tool_call_metadata")
+    return meta if isinstance(meta, Mapping) else {}
+
+
+def is_mcp_tool_call(payload: Mapping[str, object]) -> bool:
+    """Whether a closed request's payload is an MCP tool call rather than an LLM
+    call — true when the MCP gateway stamped its tool-call metadata, or the call
+    type says so on a path that hasn't populated the metadata yet."""
+    return bool(_mcp_tool_call_metadata(payload)) or (
+        payload.get("call_type") == "call_mcp_tool"
+    )
+
+
+# --- service event_metadata sanitization ------------------------------------ #
+
+# Substrings (case-insensitive) of keys that must never reach a span: secrets,
+# tokens, and raw request/response dumps the legacy service decorators pass.
+_SENSITIVE_METADATA_SUBSTRINGS: tuple[str, ...] = (
+    "api_key",
+    "token",
+    "secret",
+    "password",
+    "cookie",
+    "authorization",
+    "header",
+    "hidden_params",
+)
+# Keys that carry raw call-site internals — live objects, full kwargs/args. The
+# operation name is already the span's ``call_type``, so ``function_name`` is
+# redundant.
+_DROP_METADATA_KEYS: frozenset = frozenset(
+    {"function_kwargs", "function_args", "function_name"}
+)
+_MAX_METADATA_VALUE_LEN = 1024
+_MAX_METADATA_ITEMS = 32
+
+
+def sanitize_event_metadata(
+    event_metadata: Mapping[str, object] | None,
+) -> dict[str, str]:
+    """Reduce caller-supplied ``event_metadata`` to span-safe string attributes.
+
+    Keeps only primitive values (str/int/float/bool) under non-sensitive keys —
+    never ``repr()``-ing objects, dicts, or lists, never stamping secrets/headers,
+    and bounding the count and per-value length. This is the single chokepoint:
+    both the GenAI and legacy mappers read the cleaned result.
+    """
+    if not event_metadata:
+        return {}
+    clean: dict[str, str] = {}
+    for key, value in event_metadata.items():
+        if len(clean) >= _MAX_METADATA_ITEMS:
+            break
+        if not isinstance(key, str) or key in _DROP_METADATA_KEYS:
+            continue
+        lowered = key.lower()
+        if any(token in lowered for token in _SENSITIVE_METADATA_SUBSTRINGS):
+            continue
+        # ``bool`` is a subclass of ``int``, so it's covered. Non-primitive values
+        # (objects, dicts, lists) are dropped rather than stringified.
+        if isinstance(value, (str, int, float)):
+            clean[key] = str(value)[:_MAX_METADATA_VALUE_LEN]
+    return clean
+
+
+def _json_or_none(value: object) -> str | None:
+    """JSON-serialize ``value`` (already-string values pass through). ``None`` on failure."""
+    if isinstance(value, str):
+        return value
+    try:
+        return json.dumps(value, default=str)
+    except Exception:
+        return None
+
+
+def _guardrail_mode_str(value: object) -> str | None:
+    """Normalize ``guardrail_mode`` to a stable string.
+
+    ``guardrail_mode`` is typed as a ``GuardrailEventHooks`` enum, a list of them,
+    or a ``GuardrailMode`` — not a plain string. Emit the enum *value* (e.g.
+    ``"pre_call"``) rather than ``str(enum)`` (``"GuardrailEventHooks.pre_call"``),
+    and join a list of modes so a guardrail that runs at multiple hooks is
+    represented faithfully.
+    """
+    if value is None:
+        return None
+    if isinstance(value, (list, tuple)):
+        parts: list[str] = []
+        for item in value:
+            if item is None:
+                continue
+            part = as_str(item.value) if isinstance(item, Enum) else as_str(item)
+            if part:
+                parts.append(part)
+        return ",".join(parts) or None
+    if isinstance(value, Enum):
+        return as_str(value.value)
+    return as_str(value)
+
+
+def _total_masked_entities(value: object) -> int | None:
+    """``masked_entity_count`` is a ``{entity_type: count}`` map — sum to a total."""
+    if isinstance(value, Mapping):
+        total = sum(v for v in value.values() if isinstance(v, int))
+        return total or None
+    return as_int(value)
+
+
+def _dicts(value: object) -> tuple[Mapping[str, object], ...]:
+    """The dict items of ``value`` (when it's a list), as a tuple. Else empty."""
+    if not isinstance(value, list):
+        return ()
+    return tuple(item for item in value if isinstance(item, dict))
+
+
+def _finish_reasons(choices: tuple[Mapping[str, object], ...]) -> tuple[str, ...]:
+    """Non-empty ``finish_reason`` of each response choice."""
+    return tuple(r for c in choices if (r := as_str(c.get("finish_reason"))))
+
+
+def _parse_error(payload: "StandardLoggingPayload") -> SpanError | None:
+    """A ``SpanError`` for a failed request, or ``None`` on success."""
+    if payload.get("status") != "failure":
+        return None
+    info = cast(Mapping[str, object], payload.get("error_information") or {})
+    return SpanError(
+        error_type=as_str(info.get("error_class")) or as_str(info.get("error_code")),
+        message=as_str(info.get("error_message")) or as_str(payload.get("error_str")),
+    )
+
+
+def _tool_from_entry(entry: object) -> ToolDefinition | None:
+    """One ``tools``/``functions`` entry → ``ToolDefinition``, or ``None`` if unusable."""
+    if not isinstance(entry, dict):
+        return None
+    fn = entry.get("function") if "function" in entry else entry
+    if not isinstance(fn, dict):
+        return None
+    name = as_str(fn.get("name"))
+    if not name:
+        return None
+    params = fn.get("parameters")
+    parameters_json: str | None = None
+    if params is not None:
+        try:
+            parameters_json = json.dumps(params, default=str)
+        except Exception:
+            parameters_json = None
+    return ToolDefinition(
+        name=name,
+        description=as_str(fn.get("description")),
+        parameters_json=parameters_json,
+    )
+
+
+def _extract_tools(
+    model_parameters: Mapping[str, object],
+) -> tuple[ToolDefinition, ...]:
+    """Pull declared tools from request params (OpenAI / Anthropic shape).
+
+    Accepts the chat-completion ``tools=[{"type":"function", "function":
+    {...}}, ...]`` shape, and falls back to the ``functions=[...]`` shape.
+    Returns an empty tuple when neither is present.
+    """
+    raw_tools = model_parameters.get("tools")
+    if not isinstance(raw_tools, list):
+        raw_tools = model_parameters.get("functions")  # ``functions`` shape
+    if not isinstance(raw_tools, list):
+        return ()
+    return tuple(t for entry in raw_tools if (t := _tool_from_entry(entry)) is not None)

litellm/integrations/otel/model/semconv.py

@@ -0,0 +1,273 @@
+"""
+Keys follow the OpenTelemetry GenAI semantic conventions (experimental). Anything
+without a semconv equivalent lives under the ``litellm.*`` vendor namespace.
+"""
+
+from enum import Enum
+from typing import Final
+
+
+class GenAIOperation(str, Enum):
+    """Values for ``gen_ai.operation.name``."""
+
+    CHAT = "chat"
+    TEXT_COMPLETION = "text_completion"
+    EMBEDDINGS = "embeddings"
+    GENERATE_CONTENT = "generate_content"
+    CREATE_AGENT = "create_agent"  # reserved for future agent spans
+    INVOKE_AGENT = "invoke_agent"  # reserved for future agent spans
+    EXECUTE_TOOL = "execute_tool"  # MCP tool-call spans
+
+
+class GenAIProvider(str, Enum):
+    """Common values for the ``gen_ai.provider.name`` attribute."""
+
+    OPENAI = "openai"
+    ANTHROPIC = "anthropic"
+    AWS_BEDROCK = "aws.bedrock"
+    AZURE_AI_OPENAI = "azure.ai.openai"
+    AZURE_AI_INFERENCE = "azure.ai.inference"
+    GCP_GEMINI = "gcp.gemini"
+    GCP_VERTEX_AI = "gcp.vertex_ai"
+    COHERE = "cohere"
+    MISTRAL_AI = "mistral_ai"
+    DEEPSEEK = "deepseek"
+    GROQ = "groq"
+    PERPLEXITY = "perplexity"
+    X_AI = "x_ai"
+    IBM_WATSONX_AI = "ibm.watsonx.ai"
+
+
+class MCPMethod(str, Enum):
+    """Well-known values for ``mcp.method.name`` that litellm's MCP gateway
+    serves. The value is the JSON-RPC method exactly as it travels on the wire."""
+
+    TOOLS_CALL = "tools/call"
+    TOOLS_LIST = "tools/list"
+    PROMPTS_GET = "prompts/get"
+    PROMPTS_LIST = "prompts/list"
+
+
+class GenAI:
+    """Canonical OTel GenAI span-attribute keys."""
+
+    # request
+    OPERATION_NAME: Final = "gen_ai.operation.name"
+    PROVIDER_NAME: Final = "gen_ai.provider.name"
+    REQUEST_MODEL: Final = "gen_ai.request.model"
+    REQUEST_TEMPERATURE: Final = "gen_ai.request.temperature"
+    REQUEST_TOP_P: Final = "gen_ai.request.top_p"
+    REQUEST_TOP_K: Final = "gen_ai.request.top_k"
+    REQUEST_MAX_TOKENS: Final = "gen_ai.request.max_tokens"
+    REQUEST_FREQUENCY_PENALTY: Final = "gen_ai.request.frequency_penalty"
+    REQUEST_PRESENCE_PENALTY: Final = "gen_ai.request.presence_penalty"
+    REQUEST_STOP_SEQUENCES: Final = "gen_ai.request.stop_sequences"
+    REQUEST_SEED: Final = "gen_ai.request.seed"
+    REQUEST_CHOICE_COUNT: Final = "gen_ai.request.choice.count"
+    REQUEST_ENCODING_FORMATS: Final = "gen_ai.request.encoding_formats"
+    # response
+    RESPONSE_ID: Final = "gen_ai.response.id"
+    RESPONSE_MODEL: Final = "gen_ai.response.model"
+    RESPONSE_FINISH_REASONS: Final = "gen_ai.response.finish_reasons"
+    # usage
+    USAGE_INPUT_TOKENS: Final = "gen_ai.usage.input_tokens"
+    USAGE_OUTPUT_TOKENS: Final = "gen_ai.usage.output_tokens"
+    # content (opt-in, gated by capture mode)
+    INPUT_MESSAGES: Final = "gen_ai.input.messages"
+    OUTPUT_MESSAGES: Final = "gen_ai.output.messages"
+    SYSTEM_INSTRUCTIONS: Final = "gen_ai.system_instructions"
+    OUTPUT_TYPE: Final = "gen_ai.output.type"
+    CONVERSATION_ID: Final = "gen_ai.conversation.id"
+    # agent (reserved)
+    AGENT_ID: Final = "gen_ai.agent.id"
+    AGENT_NAME: Final = "gen_ai.agent.name"
+    # tool / tool-call (stamped on MCP tool-call spans). Arguments and result are
+    # the tool's input/output payloads — sensitive, so they're opt-in and gated by
+    # the same content-capture mode as prompt/response content.
+    TOOL_NAME: Final = "gen_ai.tool.name"
+    TOOL_CALL_ID: Final = "gen_ai.tool.call.id"
+    TOOL_CALL_ARGUMENTS: Final = "gen_ai.tool.call.arguments"
+    TOOL_CALL_RESULT: Final = "gen_ai.tool.call.result"
+    # prompt (MCP ``prompts/get`` etc.)
+    PROMPT_NAME: Final = "gen_ai.prompt.name"
+
+
+class MCP:
+    """OTel GenAI MCP (Model Context Protocol) span-attribute keys.
+
+    ``METHOD_NAME`` is the only key litellm populates from a closed request today;
+    the rest are part of the convention's vocabulary and are stamped when the
+    corresponding signal (session, protocol version, resource) is available.
+    """
+
+    METHOD_NAME: Final = "mcp.method.name"
+    SESSION_ID: Final = "mcp.session.id"
+    PROTOCOL_VERSION: Final = "mcp.protocol.version"
+    RESOURCE_URI: Final = "mcp.resource.uri"
+
+
+class JsonRpc:
+    """JSON-RPC keys carried on MCP spans. The error/status code lives in the
+    ``rpc.*`` namespace per semconv, not ``jsonrpc.*``."""
+
+    REQUEST_ID: Final = "jsonrpc.request.id"
+    PROTOCOL_VERSION: Final = "jsonrpc.protocol.version"
+    RESPONSE_STATUS_CODE: Final = "rpc.response.status_code"
+
+
+class NetworkTransport(str, Enum):
+    """Well-known values for ``network.transport``."""
+
+    TCP = "tcp"
+    UDP = "udp"
+    QUIC = "quic"
+    UNIX = "unix"
+    PIPE = "pipe"
+
+
+class Network:
+    """OTel network keys, recommended on MCP spans to describe the transport
+    carrying the JSON-RPC messages (stdio pipe, HTTP, websocket, …)."""
+
+    PROTOCOL_NAME: Final = "network.protocol.name"
+    PROTOCOL_VERSION: Final = "network.protocol.version"
+    TRANSPORT: Final = "network.transport"
+
+
+class Client:
+    """Peer (client) network keys, stamped on MCP *server* spans the same way
+    ``server.*`` is stamped on client spans."""
+
+    ADDRESS: Final = "client.address"
+    PORT: Final = "client.port"
+
+
+class Error:
+    TYPE: Final = "error.type"
+
+
+class Server:
+    ADDRESS: Final = "server.address"
+    PORT: Final = "server.port"
+
+
+class DB:
+    """Database / cache client-span keys (OTel ``db.*`` semconv).
+
+    Stamped on ``DB_CALL`` spans (redis / postgres), which are CLIENT spans for
+    outbound datastore calls — not on the INTERNAL ``SERVICE`` spans.
+    """
+
+    SYSTEM_NAME: Final = "db.system.name"
+    OPERATION_NAME: Final = "db.operation.name"
+
+
+class HTTP:
+    """HTTP server-span keys. Belong on the SERVER span only (never promoted)."""
+
+    REQUEST_METHOD: Final = "http.request.method"
+    ROUTE: Final = "http.route"
+    RESPONSE_STATUS_CODE: Final = "http.response.status_code"
+    URL_PATH: Final = "url.path"
+
+
+class LiteLLM:
+    """Vendor-extension keys (no semconv equivalent). Always ``litellm.*``."""
+
+    CALL_ID: Final = "litellm.call_id"
+    COST_PREFIX: Final = "litellm.cost."
+    METADATA_PREFIX: Final = "litellm.metadata."
+    TEAM_ID: Final = "litellm.team.id"
+    TEAM_ALIAS: Final = "litellm.team.alias"
+    # The team's free-form metadata dict, JSON-serialized into a single value.
+    TEAM_METADATA: Final = "litellm.team.metadata"
+    KEY_HASH: Final = "litellm.api_key.hash"
+    END_USER: Final = "litellm.end_user.id"
+    # The model string litellm actually sent to the provider (the deployment's
+    # ``litellm_params.model``), distinct from the user-facing ``gen_ai.request.model``.
+    PROVIDER_MODEL: Final = "litellm.provider.model"
+    REQUEST_STREAMING: Final = "litellm.request.streaming"
+    GUARDRAIL_NAME: Final = "litellm.guardrail.name"
+    GUARDRAIL_MODE: Final = "litellm.guardrail.mode"
+    GUARDRAIL_STATUS: Final = "litellm.guardrail.status"
+    GUARDRAIL_PROVIDER: Final = "litellm.guardrail.provider"
+    GUARDRAIL_ACTION: Final = "litellm.guardrail.action"
+    GUARDRAIL_RESPONSE: Final = "litellm.guardrail.response"
+    GUARDRAIL_VIOLATION_CATEGORIES: Final = "litellm.guardrail.violation_categories"
+    GUARDRAIL_CONFIDENCE_SCORE: Final = "litellm.guardrail.confidence_score"
+    GUARDRAIL_RISK_SCORE: Final = "litellm.guardrail.risk_score"
+    GUARDRAIL_MASKED_ENTITY_COUNT: Final = "litellm.guardrail.masked_entity_count"
+    GUARDRAIL_DURATION: Final = "litellm.guardrail.duration"
+    GUARDRAIL_ID: Final = "litellm.guardrail.id"
+    GUARDRAIL_POLICY_TEMPLATE: Final = "litellm.guardrail.policy_template"
+    GUARDRAIL_DETECTION_METHOD: Final = "litellm.guardrail.detection_method"
+    SERVICE_NAME: Final = "litellm.service.name"
+    SERVICE_CALL_TYPE: Final = "litellm.service.call_type"
+    PREPROCESSING_MS: Final = "litellm.preprocessing.duration_ms"
+    # The logical name of the MCP server a tool call was routed to. There is no
+    # semconv key for an MCP server's *name* (the convention uses ``server.address``
+    # for its network location), so it lives under the vendor namespace.
+    MCP_SERVER_NAME: Final = "litellm.mcp.server.name"
+
+
+class Metric:
+    """GenAI metric instrument names."""
+
+    TOKEN_USAGE: Final = "gen_ai.client.token.usage"
+    OPERATION_DURATION: Final = "gen_ai.client.operation.duration"
+
+
+# litellm ``custom_llm_provider`` -> ``gen_ai.provider.name`` value.
+_PROVIDER_BY_LITELLM: dict[str, GenAIProvider] = {
+    "openai": GenAIProvider.OPENAI,
+    "text-completion-openai": GenAIProvider.OPENAI,
+    "azure": GenAIProvider.AZURE_AI_OPENAI,
+    "azure_ai": GenAIProvider.AZURE_AI_INFERENCE,
+    "anthropic": GenAIProvider.ANTHROPIC,
+    "bedrock": GenAIProvider.AWS_BEDROCK,
+    "bedrock_converse": GenAIProvider.AWS_BEDROCK,
+    "vertex_ai": GenAIProvider.GCP_VERTEX_AI,
+    "vertex_ai_beta": GenAIProvider.GCP_VERTEX_AI,
+    "gemini": GenAIProvider.GCP_GEMINI,
+    "cohere": GenAIProvider.COHERE,
+    "cohere_chat": GenAIProvider.COHERE,
+    "mistral": GenAIProvider.MISTRAL_AI,
+    "deepseek": GenAIProvider.DEEPSEEK,
+    "groq": GenAIProvider.GROQ,
+    "perplexity": GenAIProvider.PERPLEXITY,
+    "xai": GenAIProvider.X_AI,
+    "watsonx": GenAIProvider.IBM_WATSONX_AI,
+}
+
+# litellm ``call_type`` -> ``gen_ai.operation.name``.
+_OPERATION_BY_CALL_TYPE: dict[str, GenAIOperation] = {
+    "completion": GenAIOperation.CHAT,
+    "acompletion": GenAIOperation.CHAT,
+    "completion_with_retries": GenAIOperation.CHAT,
+    "text_completion": GenAIOperation.TEXT_COMPLETION,
+    "atext_completion": GenAIOperation.TEXT_COMPLETION,
+    "embedding": GenAIOperation.EMBEDDINGS,
+    "aembedding": GenAIOperation.EMBEDDINGS,
+    "responses": GenAIOperation.CHAT,
+    "aresponses": GenAIOperation.CHAT,
+    "call_mcp_tool": GenAIOperation.EXECUTE_TOOL,
+}
+
+
+def resolve_provider(custom_llm_provider: str | None) -> str:
+    """Map a litellm provider string to a ``gen_ai.provider.name`` value.
+
+    Unknown providers pass through verbatim — the convention explicitly allows
+    provider-specific values, so an unmapped name is still valid.
+    """
+    if not custom_llm_provider:
+        return ""
+    mapped = _PROVIDER_BY_LITELLM.get(custom_llm_provider.lower())
+    return mapped.value if mapped is not None else custom_llm_provider
+
+
+def resolve_operation(call_type: str | None) -> GenAIOperation:
+    """Map a litellm ``call_type`` to a ``gen_ai.operation.name`` value."""
+    if not call_type:
+        return GenAIOperation.CHAT
+    return _OPERATION_BY_CALL_TYPE.get(call_type.lower(), GenAIOperation.CHAT)

litellm/integrations/otel/model/spans.py

@@ -0,0 +1,215 @@
+"""
+This module declares every span the instrumentation can emit and the hierarchy.
+
+Span-name patterns live here as typed builder functions.
+
+Canonical hierarchy::
+
+    PROXY_REQUEST  (SERVER, root)     # owned by the FastAPI instrumentor
+    ├── SERVICE    (INTERNAL)         # auth phase span (live; see logger.phase_span)
+    │   └── DB_CALL (CLIENT)          #   its key/user/team lookups nest here
+    ├── GUARDRAIL  (INTERNAL)         # request-lifecycle hook, sibling of LLM_CALL
+    ├── LLM_CALL   (CLIENT)
+    └── DB_CALL    (CLIENT)           # e.g. the spend-log write
+
+Guardrails parent to PROXY_REQUEST, not LLM_CALL: pre/during/post-call guardrail
+hooks are orchestrated by the request lifecycle (a pre-call guardrail runs
+before the LLM call even starts), so a guardrail is a sibling of the LLM call,
+not a child of it. The emitter parents every span to the ambient OTel context
+(the active server span), which matches this.
+
+Not every service call becomes a span — :func:`span_role_for_service` decides:
+
+- ``DB_CALL`` (CLIENT) — outbound datastores (redis, postgres,
+  ``batch_write_to_db``), carrying ``db.*`` semconv.
+- ``SERVICE`` (INTERNAL) — genuine internal work worth a span (background
+  budget/reset jobs, pod-lock manager).
+- ``None`` (metrics-only) — framework instrumentation that duplicates a gen-AI
+  span (``self`` = the ``track_llm_api_timing`` wrapper, ``router``,
+  ``proxy_pre_call``) or ``auth`` (which gets a live phase span instead). These
+  still feed Prometheus/Datadog; they just never enter the trace.
+
+``DB_CALL`` and ``SERVICE`` are built from the same ``ServiceSpanData``; only the
+role (hence span kind and attribute vocabulary) differs. A service call can fire
+outside any request (a background job), in which case it parents to no server
+span and starts its own root trace rather than being dropped.
+
+Management/admin endpoints are ordinary FastAPI routes — their SERVER spans are
+owned by the instrumentor too, so they don't appear as a role here.
+"""
+
+from dataclasses import dataclass
+from enum import Enum
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from litellm.integrations.otel.model.payloads import (
+        GuardrailSpanData,
+        LLMCallSpanData,
+        MCPToolCallSpanData,
+        ProxyRequestSpanData,
+        ServiceSpanData,
+    )
+
+
+class SpanRole(str, Enum):
+    PROXY_REQUEST = "proxy_request"
+    LLM_CALL = "llm_call"
+    MCP_TOOL_CALL = "mcp_tool_call"
+    GUARDRAIL = "guardrail"
+    DB_CALL = "db_call"
+    SERVICE = "service"
+
+
+class LiteLLMSpanKind(str, Enum):
+    SERVER = "server"
+    CLIENT = "client"
+    INTERNAL = "internal"
+    PRODUCER = "producer"
+    CONSUMER = "consumer"
+
+
+@dataclass(frozen=True)
+class SpanSpec:
+    role: SpanRole
+    kind: LiteLLMSpanKind
+    parent: SpanRole | None
+
+
+SPAN_REGISTRY: dict[SpanRole, SpanSpec] = {
+    SpanRole.PROXY_REQUEST: SpanSpec(
+        SpanRole.PROXY_REQUEST, LiteLLMSpanKind.SERVER, parent=None
+    ),
+    SpanRole.LLM_CALL: SpanSpec(
+        SpanRole.LLM_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST
+    ),
+    # The proxy is an MCP client to the upstream server it dispatches the tool
+    # call to, so this is a CLIENT span, sibling of the LLM call under the request.
+    SpanRole.MCP_TOOL_CALL: SpanSpec(
+        SpanRole.MCP_TOOL_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.GUARDRAIL: SpanSpec(
+        SpanRole.GUARDRAIL, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.DB_CALL: SpanSpec(
+        SpanRole.DB_CALL, LiteLLMSpanKind.CLIENT, parent=SpanRole.PROXY_REQUEST
+    ),
+    SpanRole.SERVICE: SpanSpec(
+        SpanRole.SERVICE, LiteLLMSpanKind.INTERNAL, parent=SpanRole.PROXY_REQUEST
+    ),
+}
+
+
+# ``ServiceTypes`` value -> ``db.system.name``. These are outbound datastore
+# calls and become CLIENT ``DB_CALL`` spans; ``redis_``-prefixed names cover the
+# redis-backed spend queues. Any service not mapped here is litellm-internal work
+# and stays an INTERNAL ``SERVICE`` span. This table is the single source of
+# datastore knowledge — both the role classifier and the mapper read it.
+_DB_SYSTEM_BY_SERVICE: dict[str, str] = {
+    "redis": "redis",
+    "postgres": "postgresql",
+    "batch_write_to_db": "postgresql",
+}
+
+
+def db_system(service_name: str) -> str | None:
+    """The ``db.system.name`` for a datastore service, else ``None``.
+
+    ``None`` means the service is not an outbound datastore call. Redis-backed
+    spend queues (``redis_*``) map to ``redis``.
+    """
+    if service_name in _DB_SYSTEM_BY_SERVICE:
+        return _DB_SYSTEM_BY_SERVICE[service_name]
+    if service_name.startswith("redis_"):
+        return "redis"
+    return None
+
+
+# ``ServiceTypes`` values that are NOT emitted as spans — they are framework
+# instrumentation that either duplicates a gen-AI span or has a better home as a
+# Prometheus/Datadog metric. They still flow to those metric backends via their
+# own hooks; the v2 logger just does not put them in the trace:
+#
+#   - ``self``           — ``track_llm_api_timing`` wraps the LLM call; the
+#                          ``chat {model}`` CLIENT span already represents it.
+#   - ``router``         — wraps the whole request; duplicates the server span.
+#   - ``proxy_pre_call`` — per-callback pre-call timing; a guardrail's real span
+#                          is ``execute_guardrail {name}``.
+#   - ``auth``           — emitted instead as a live phase span (see
+#                          ``logger.phase_span``) so its DB lookups nest under it,
+#                          not as a flat post-hoc service span.
+_METRICS_ONLY_SERVICES: frozenset[str] = frozenset(
+    {"self", "router", "proxy_pre_call", "auth"}
+)
+
+
+def span_role_for_service(service_name: str) -> SpanRole | None:
+    """The span role for a service call, or ``None`` when it must not be a span.
+
+    ``DB_CALL`` for outbound datastores, ``SERVICE`` for genuine internal work
+    worth a span (background jobs), and ``None`` for framework instrumentation
+    that duplicates a gen-AI span or belongs in metrics only
+    (see ``_METRICS_ONLY_SERVICES``).
+    """
+    if service_name in _METRICS_ONLY_SERVICES:
+        return None
+    return SpanRole.DB_CALL if db_system(service_name) is not None else SpanRole.SERVICE
+
+
+# --- span name builders (the naming convention, per role) ------------------- #
+
+
+# The name the FastAPI instrumentor gives the root server span. V2 never creates
+# this span (the instrumentor owns it), but it anchors request-level spans to it
+# and tests assert against it by name, so the literal lives here with the rest of
+# the span vocabulary rather than being duplicated at each call site.
+LITELLM_PROXY_REQUEST_SPAN_NAME = "Received Proxy Server Request"
+
+
+def llm_call_span_name(data: "LLMCallSpanData") -> str:
+    """``"{operation} {model}"`` e.g. ``"chat gpt-4o"`` (GenAI semconv)."""
+    model = data.request_model or ""
+    return f"{data.operation.value} {model}".strip()
+
+
+def mcp_tool_call_span_name(data: "MCPToolCallSpanData") -> str:
+    """``"{mcp.method.name} {tool}"`` e.g. ``"tools/call get-weather"`` (MCP semconv)."""
+    return f"{data.method} {data.tool_name}".strip()
+
+
+def proxy_request_span_name(data: "ProxyRequestSpanData") -> str:
+    """``"{method} {route}"`` (HTTP semconv)."""
+    return f"{data.http_method} {data.route}".strip()
+
+
+def guardrail_span_name(data: "GuardrailSpanData") -> str:
+    return f"execute_guardrail {data.guardrail_name}".strip()
+
+
+def service_span_name(data: "ServiceSpanData") -> str:
+    """``"{service} {call_type}"`` e.g. ``"redis set"`` — service name alone when
+    no call type is known, so identically-named calls stay distinguishable."""
+    return f"{data.service_name} {data.call_type or ''}".strip()
+
+
+def root_roles() -> list[SpanRole]:
+    """Roles that start a new trace (no in-process parent)."""
+    return [role for role, spec in SPAN_REGISTRY.items() if spec.parent is None]
+
+
+def child_roles(parent: SpanRole) -> list[SpanRole]:
+    return [role for role, spec in SPAN_REGISTRY.items() if spec.parent == parent]
+
+
+def validate_registry(
+    registry: dict[SpanRole, SpanSpec] | None = None,
+) -> None:
+    reg = registry if registry is not None else SPAN_REGISTRY
+    for role, spec in reg.items():
+        if spec.role is not role:
+            raise ValueError(f"SPAN_REGISTRY[{role}] has mismatched role {spec.role}")
+        if spec.parent is not None and spec.parent not in reg:
+            raise ValueError(f"span role {role} declares unknown parent {spec.parent}")
+    missing = [role for role in SpanRole if role not in reg]
+    if missing:
+        raise ValueError(f"SPAN_REGISTRY is missing roles: {missing}")

litellm/integrations/otel/model/utils.py

@@ -0,0 +1,103 @@
+"""Shared, OpenTelemetry-free helpers for the otel integration.
+
+Generic value coercion (for reading heterogeneous logging-payload dicts), time
+conversion, and header parsing — pulled out of the individual modules so they
+live in one place. Deliberately free of any ``opentelemetry`` import so the
+OTel-free sources of truth (payloads, semconv, spans, config) can use it too.
+"""
+
+from datetime import datetime
+
+
+def as_str(value: object) -> str | None:
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return value
+    return str(value)
+
+
+def as_int(value: object) -> int | None:
+    if isinstance(value, bool):
+        return int(value)
+    if isinstance(value, int):
+        return value
+    if isinstance(value, float):
+        return int(value)
+    if isinstance(value, str):
+        try:
+            return int(value)
+        except ValueError:
+            return None
+    return None
+
+
+def as_float(value: object) -> float | None:
+    if isinstance(value, bool):
+        return float(value)
+    if isinstance(value, (int, float)):
+        return float(value)
+    if isinstance(value, str):
+        try:
+            return float(value)
+        except ValueError:
+            return None
+    return None
+
+
+def as_bool(value: object) -> bool | None:
+    if value is None:
+        return None
+    if isinstance(value, bool):
+        return value
+    return bool(value)
+
+
+def as_str_tuple(value: object) -> tuple[str, ...] | None:
+    if value is None:
+        return None
+    if isinstance(value, str):
+        return (value,)
+    if isinstance(value, (list, tuple)):
+        return tuple(str(v) for v in value)
+    return None
+
+
+def to_ns(value: datetime | float | int | None) -> int | None:
+    """Coerce a datetime / epoch value to integer nanoseconds."""
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        return int(value.timestamp() * 1e9)
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return int(float(value) * 1e9)
+    return None
+
+
+def to_seconds(value: datetime | float | int | str | None) -> float | None:
+    """Coerce a datetime / epoch / formatted-string value to epoch seconds."""
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        return value.timestamp()
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return float(value)
+    if isinstance(value, str):
+        for fmt in ("%Y-%m-%d %H:%M:%S.%f", "%Y-%m-%d %H:%M:%S"):
+            try:
+                return datetime.strptime(value, fmt).timestamp()
+            except ValueError:
+                continue
+    return None
+
+
+def parse_headers(raw: str | None) -> dict[str, str]:
+    """Parse an OTLP ``"k=v,k=v"`` header string into a dict."""
+    headers: dict[str, str] = {}
+    if not raw:
+        return headers
+    for pair in raw.split(","):
+        if "=" in pair:
+            key, _, value = pair.partition("=")
+            headers[key.strip()] = value.strip()
+    return headers

litellm/integrations/otel/mount.py

@@ -0,0 +1,130 @@
+"""FastAPI server-span instrumentation — the proxy mounts this at app creation.
+
+``opentelemetry-instrumentation-fastapi`` creates the SERVER span for each HTTP
+route and extracts inbound ``traceparent`` headers. This module owns the one call
+site that attaches it to the proxy app, plus the passthrough span-naming hook, so
+``proxy_server`` stays free of OTel details.
+
+The ``FastAPIInstrumentor`` import is kept lazy (inside :func:`instrument_fastapi_app`,
+after the gate check) so importing this module never requires the optional
+``opentelemetry-instrumentation-fastapi`` package and pulls in nothing OTel-related
+when the feature gate is off.
+"""
+
+import os
+from typing import Any
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+# Routes excluded from server-span tracing by default: high-frequency pollers and
+# static UI/docs assets, none of which are LLM traffic. Entries are substring-matched
+# against the request path (unanchored, so they survive a ``server_root_path`` prefix
+# and each entry also covers everything beneath it — e.g. ``/health`` covers
+# ``/health/readiness``). Operators override the whole set via the standard
+# ``OTEL_PYTHON_FASTAPI_EXCLUDED_URLS`` env var (set "" to trace everything).
+_DEFAULT_EXCLUDED_ROUTES = (
+    "/health",  # load-balancer liveness/readiness polling
+    "/metrics",  # Prometheus scrape (also drops the /model/metrics admin analytics)
+    "/litellm-asset-prefix",  # hashed UI asset bundles
+    "/_next",  # Next.js static JS/CSS chunks (root-level mount)
+    "/ui",  # admin UI single-page app
+    "/swagger",  # static Swagger UI assets
+    "/docs",  # FastAPI Swagger docs page
+    "/redoc",  # FastAPI ReDoc docs page
+    "/openapi.json",  # OpenAPI schema
+    "favicon",  # /favicon.ico + /get_favicon
+    "/.well-known",  # UI config discovery
+)
+_DEFAULT_EXCLUDED_URLS = ",".join(_DEFAULT_EXCLUDED_ROUTES)
+
+# Passthrough routes are catch-alls (e.g. "/openai/{endpoint:path}"), so the
+# default OTel server-span name "{method} {route}" collapses every upstream
+# endpoint into "POST /openai/{endpoint:path}". The hook below renames those spans
+# to the real request path so each endpoint is distinguishable. Non-catch-all
+# routes keep their low-cardinality template name.
+PASSTHROUGH_PREFIXES = frozenset(
+    {
+        "openai",
+        "openai_passthrough",
+        "anthropic",
+        "azure",
+        "azure_ai",
+        "bedrock",
+        "cohere",
+        "cursor",
+        "gemini",
+        "mistral",
+        "vllm",
+        "vertex_ai",
+        "vertex-ai",
+        "assemblyai",
+        "eu.assemblyai",
+        "milvus",
+    }
+)
+
+
+def _passthrough_span_name_hook(span: Any, scope: dict) -> None:
+    """FastAPI ``server_request_hook``: give passthrough server spans a useful name.
+
+    The instrumentation matches the route at span creation, so both the span name
+    and ``http.route`` are set to the catch-all template (``/openai/{endpoint:path}``)
+    before this hook runs. Rewrite both to the real request path so each upstream
+    endpoint is distinguishable. (The ASGI ``http receive``/``http send`` sub-spans
+    can't be renamed from here — their name is captured at creation — so they are
+    dropped via ``exclude_spans`` at instrumentation time.)
+    """
+    try:
+        if span is None or not span.is_recording():
+            return
+        path = scope.get("path") or ""
+        method = scope.get("method") or ""
+        first_segment = path.lstrip("/").split("/", 1)[0]
+        if first_segment in PASSTHROUGH_PREFIXES:
+            span.update_name(f"{method} {path}".strip())
+            span.set_attribute("http.route", path)
+    except Exception:
+        pass
+
+
+def instrument_fastapi_app(app: Any) -> None:
+    """Attach OTel server-span instrumentation to the proxy FastAPI app.
+
+    Safe no-op when the V2 gate is off or ``opentelemetry-instrumentation-fastapi``
+    is unavailable. This MUST be called at app-creation time — once the lifespan
+    runs, the middleware stack is frozen and ``instrument_app`` raises "Cannot add
+    middleware after an application has started".
+
+    No ``TracerProvider`` is passed, so the instrumentation binds to the OTel global
+    ``ProxyTracerProvider``; the proxy publishes the real provider as the global
+    after config load (see ``proxy_startup_event``), and the proxy delegates to it.
+    That way server spans and gen-ai spans share one provider and the same trace.
+    """
+    try:
+        if not is_otel_v2_enabled():
+            return
+
+        # Lazy: only the V2-enabled path needs the optional
+        # ``opentelemetry-instrumentation-fastapi`` package, which is not part of the
+        # base ``litellm[proxy]`` install. Importing it at module top would make
+        # ``proxy_server``'s unconditional ``import`` of this module crash when the
+        # package is absent, even with the gate off.
+        from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
+
+        excluded_urls = (
+            os.environ.get("OTEL_PYTHON_FASTAPI_EXCLUDED_URLS")
+            if "OTEL_PYTHON_FASTAPI_EXCLUDED_URLS" in os.environ
+            else _DEFAULT_EXCLUDED_URLS
+        )
+        FastAPIInstrumentor.instrument_app(
+            app,
+            excluded_urls=excluded_urls,
+            server_request_hook=_passthrough_span_name_hook,
+            # Drop the ASGI "http receive"/"http send" lifecycle sub-spans: they
+            # are low-value noise and (for passthrough) carry the catch-all route
+            # template in their name, which can't be rewritten from a hook.
+            exclude_spans=["receive", "send"],
+        )
+    except Exception as e:
+        verbose_logger.debug("Skipping OTel V2 FastAPI instrumentation: %s", e)

litellm/integrations/otel/plumbing/context.py

@@ -0,0 +1,127 @@
+"""Trace-context + Baggage helpers."""
+
+from contextvars import ContextVar
+from typing import Mapping
+
+from opentelemetry import baggage
+from opentelemetry.context import Context, get_current
+from opentelemetry.trace import Span, get_current_span, set_span_in_context
+from opentelemetry.trace.propagation.tracecontext import (
+    TraceContextTextMapPropagator,
+)
+
+_PROPAGATOR = TraceContextTextMapPropagator()
+
+# The request's root span — the FastAPI-owned SERVER span — captured ONCE when the
+# proxy first resolves it, so request-level spans (the LLM call, guardrails) can
+# parent to it EXPLICITLY instead of to whatever span happens to be active at the
+# instant they are emitted. Ambient-only parenting (``get_current_span()``) is
+# wrong at two boundaries:
+#   * inside the ``auth`` phase span the active span is the auth span, so an LLM /
+#     guardrail span emitted there would nest under auth instead of being its
+#     sibling; and
+#   * in a detached success task (pass-through logs success from a fire-and-forget
+#     ``asyncio.create_task``) the server span may not be active at all, orphaning
+#     the span into a brand-new trace.
+# A ``ContextVar`` (not a request attribute) so it rides the request task's context
+# and is inherited by ``asyncio.create_task`` children — i.e. the async logging
+# callbacks that close the span. It is never reset: the contextvar dies with the
+# request task, so there is nothing to leak.
+_request_root_span: "ContextVar[Span | None]" = ContextVar(
+    "litellm_otel_request_root_span", default=None
+)
+
+
+def set_request_root_span(span: Span) -> None:
+    """Anchor the request's root (server) span for explicit child parenting.
+
+    No-ops for a non-recordable span so a bad capture can never replace a good one
+    with a phantom parent. Idempotent — the proxy captures the same server span at
+    more than one entry point.
+    """
+    if is_recordable_span(span):
+        _request_root_span.set(span)
+
+
+def request_root_span() -> "Span | None":
+    """The anchored request root span, or ``None`` outside a proxy request."""
+    span = _request_root_span.get()
+    return span if is_recordable_span(span) else None
+
+
+def set_request_baggage(
+    values: Mapping[str, str], context: Context | None = None
+) -> Context:
+    """Return a context with ``values`` written into Baggage."""
+    ctx = context
+    for key, value in values.items():
+        ctx = baggage.set_baggage(key, value, context=ctx)
+    return ctx if ctx is not None else (context or get_current())
+
+
+def get_baggage_attributes(context: Context | None = None) -> dict[str, str]:
+    """All Baggage entries on ``context`` as strings."""
+    return {key: str(value) for key, value in baggage.get_all(context).items()}
+
+
+def context_from_span(span: Span, context: Context | None = None) -> Context:
+    """A context with ``span`` as the active span (for explicit parenting)."""
+    return set_span_in_context(span, context=context)
+
+
+def resolve_parent_context(threaded: Span | None = None) -> Context:
+    """The context a child span should parent under.
+
+    Ambient-first: parent to the active OTel context (the server span, restored
+    by the logging worker or active in the request task), falling back to a span
+    passed explicitly (``threaded``) only when the ambient context has no
+    recordable span — e.g. a background service call with no request on the
+    stack. When neither is recordable the ambient context is returned unchanged,
+    so the span starts a new root trace.
+
+    Only service/DB spans pass ``threaded`` (the ``parent_otel_span`` handed to
+    the service hook). Request-level spans — the LLM call and guardrails — are
+    created where the server span is genuinely ambient, so they never need it.
+    """
+    ctx = get_current()
+    if is_recordable_span(threaded) and not is_recordable_span(get_current_span(ctx)):
+        ctx = context_from_span(threaded, context=ctx)  # type: ignore[arg-type]
+    return ctx
+
+
+def resolve_request_span_context() -> Context:
+    """The parent context for a request-level span (the LLM call, a guardrail).
+
+    These are direct children of the request's root server span — siblings of the
+    ``auth`` phase span and of each other, never nested under whatever span is
+    momentarily active. So prefer the explicitly anchored root span; fall back to
+    ambient context only when there is no anchor (the SDK / no-proxy path), where
+    the span legitimately starts its own root trace.
+
+    Unlike :func:`resolve_parent_context` (used by DB/service spans, which DO want
+    to nest under the active phase span, e.g. an auth DB lookup under ``auth``),
+    this never returns the active span when an anchor exists.
+    """
+    root = request_root_span()
+    if root is not None:
+        return context_from_span(root)
+    return get_current()
+
+
+def is_recordable_span(obj: object) -> bool:
+    """True if ``obj`` is a live span with a valid context (safe to parent under)."""
+    if not isinstance(obj, Span):
+        return False
+    try:
+        ctx = obj.get_span_context()
+    except Exception:
+        return False
+    return ctx is not None and ctx.is_valid
+
+
+def extract_traceparent(headers: Mapping[str, str]) -> Context | None:
+    """Extract a remote parent context from incoming HTTP headers, if present."""
+    if not any(key.lower() == "traceparent" for key in headers):
+        return None
+    carrier = {str(key).lower(): value for key, value in headers.items()}
+    return _PROPAGATOR.extract(carrier)

litellm/integrations/otel/plumbing/metrics.py

@@ -0,0 +1,28 @@
+"""GenAI client metrics (token usage + operation duration histograms)."""
+
+from dataclasses import dataclass
+
+from opentelemetry.metrics import Histogram, Meter
+
+from litellm.integrations.otel.model.semconv import Metric
+
+
+@dataclass(frozen=True)
+class GenAIMetrics:
+    token_usage: Histogram
+    operation_duration: Histogram
+
+
+def create_genai_metrics(meter: Meter) -> GenAIMetrics:
+    return GenAIMetrics(
+        token_usage=meter.create_histogram(
+            name=Metric.TOKEN_USAGE,
+            unit="{token}",
+            description="Number of tokens used per GenAI request.",
+        ),
+        operation_duration=meter.create_histogram(
+            name=Metric.OPERATION_DURATION,
+            unit="s",
+            description="GenAI operation duration.",
+        ),
+    )

litellm/integrations/otel/plumbing/providers.py

@@ -0,0 +1,220 @@
+"""Provider / exporter factory + the Baggage span processor."""
+
+from typing import Callable, Iterable
+
+from opentelemetry import baggage
+from opentelemetry.context import Context
+from opentelemetry.sdk.resources import Resource
+from opentelemetry.sdk.trace import ReadableSpan, SpanProcessor, TracerProvider
+from opentelemetry.sdk.trace.export import (
+    BatchSpanProcessor,
+    ConsoleSpanExporter,
+    SimpleSpanProcessor,
+    SpanExporter,
+)
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
+    InMemorySpanExporter,
+)
+from opentelemetry.trace import Span, SpanKind, Tracer
+
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.model.semconv import LiteLLM
+from litellm.integrations.otel.model.spans import LiteLLMSpanKind
+
+# Re-exported so ``providers.parse_headers`` remains a stable entry point.
+from litellm.integrations.otel.model.utils import parse_headers as parse_headers
+
+_SPAN_KIND_BY_ROLE_KIND: dict[LiteLLMSpanKind, SpanKind] = {
+    LiteLLMSpanKind.SERVER: SpanKind.SERVER,
+    LiteLLMSpanKind.CLIENT: SpanKind.CLIENT,
+    LiteLLMSpanKind.INTERNAL: SpanKind.INTERNAL,
+    LiteLLMSpanKind.PRODUCER: SpanKind.PRODUCER,
+    LiteLLMSpanKind.CONSUMER: SpanKind.CONSUMER,
+}
+
+
+def to_otel_span_kind(kind: LiteLLMSpanKind) -> SpanKind:
+    return _SPAN_KIND_BY_ROLE_KIND[kind]
+
+
+# Custom exporter factories keyed by ``ExporterSpec.kind``. A preset registers
+# one here when its destination needs construction logic the built-in kinds
+# can't express — e.g. an exporter that fetches an auth token lazily on its
+# first export (off the event loop) instead of blocking at config-build time.
+# Keeping the registry here lets this module stay vendor-agnostic: the factory
+# lives with the integration that needs it.
+_EXPORTER_FACTORIES: dict[str, Callable[[ExporterSpec], SpanExporter]] = {}
+
+
+def register_exporter_factory(
+    kind: str, factory: Callable[[ExporterSpec], SpanExporter]
+) -> None:
+    """Register a custom exporter ``factory`` for the exporter ``kind``."""
+    _EXPORTER_FACTORIES[kind.lower()] = factory
+
+
+class LiteLLMBaggageSpanProcessor(SpanProcessor):
+    """Stamps an allowlisted set of Baggage entries onto every span at start."""
+
+    def __init__(
+        self,
+        allowed_keys: Iterable[str],
+        allowed_prefixes: tuple[str, ...] = (LiteLLM.METADATA_PREFIX,),
+    ) -> None:
+        self._allowed_keys = frozenset(allowed_keys)
+        self._allowed_prefixes = tuple(allowed_prefixes)
+
+    def _is_allowed(self, key: str) -> bool:
+        return key in self._allowed_keys or any(
+            key.startswith(prefix) for prefix in self._allowed_prefixes
+        )
+
+    def on_start(self, span: Span, parent_context: Context | None = None) -> None:
+        for key, value in baggage.get_all(parent_context).items():
+            if self._is_allowed(key) and isinstance(value, (str, bool, int, float)):
+                span.set_attribute(key, value)
+
+    def on_end(self, span: ReadableSpan) -> None:  # noqa: D401 - no-op
+        return None
+
+    def shutdown(self) -> None:
+        return None
+
+    def force_flush(self, timeout_millis: int = 30000) -> bool:
+        return True
+
+
+def _otlp_traces_endpoint(endpoint: str | None) -> str | None:
+    """Point an OTLP/HTTP base endpoint at the ``/v1/traces`` signal path.
+
+    ``OTEL_EXPORTER_OTLP_ENDPOINT`` is a base URL (e.g. ``http://host:4318``).
+    The OTLP/HTTP exporter only appends the ``/v1/traces`` path when it reads
+    that env var itself; when an endpoint is passed explicitly it is used
+    verbatim, so a base URL would POST to the root and the collector returns
+    404. Append the signal path here (leaving an already-correct path intact).
+    """
+    if not endpoint:
+        return endpoint
+    endpoint = endpoint.rstrip("/")
+    # Splunk Observability uses ``/v2/trace/otlp``; never rewrite it.
+    if endpoint.endswith("/v1/traces") or "/v2/trace/otlp" in endpoint:
+        return endpoint
+    for other_signal in ("/v1/logs", "/v1/metrics"):
+        if endpoint.endswith(other_signal):
+            return endpoint[: -len(other_signal)] + "/v1/traces"
+    return endpoint + "/v1/traces"
+
+
+def _exporter_from_spec(spec: ExporterSpec) -> SpanExporter:
+    kind = (spec.kind or "console").lower()
+    factory = _EXPORTER_FACTORIES.get(kind)
+    if factory is not None:
+        return factory(spec)
+    if kind in ("in_memory", "inmemory", "memory"):
+        return InMemorySpanExporter()
+    if kind in ("otlp_http", "http", "http/protobuf", "http/json"):
+        from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
+            OTLPSpanExporter as HTTPExporter,
+        )
+
+        return HTTPExporter(
+            endpoint=_otlp_traces_endpoint(spec.endpoint),
+            headers=parse_headers(spec.headers),
+        )
+    if kind in ("otlp_grpc", "grpc"):
+        from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import (
+            OTLPSpanExporter as GRPCExporter,
+        )
+
+        return GRPCExporter(endpoint=spec.endpoint, headers=parse_headers(spec.headers))
+    return ConsoleSpanExporter()
+
+
+def _processor_for(exporter: SpanExporter, use_simple: bool | None) -> SpanProcessor:
+    """Pick a Simple or Batch span processor for ``exporter``.
+
+    When ``use_simple`` is unset, default to Simple for console and in-memory
+    exporters (spans export synchronously, which tests rely on) and Batch for
+    everything else (the right export semantics for production).
+    """
+    if use_simple is None:
+        use_simple = isinstance(exporter, (ConsoleSpanExporter, InMemorySpanExporter))
+    return SimpleSpanProcessor(exporter) if use_simple else BatchSpanProcessor(exporter)
+
+
+def build_span_exporter(config: OpenTelemetryV2Config) -> SpanExporter:
+    """Build a single exporter from the top-level config fields.
+
+    Convenience for the common single-exporter case (and for tests): reads the
+    ``exporter`` / ``endpoint`` / ``headers`` fields. To configure multiple
+    exporters, populate ``config.exporters`` directly.
+    """
+    return _exporter_from_spec(
+        ExporterSpec(
+            kind=config.exporter, endpoint=config.endpoint, headers=config.headers
+        )
+    )
+
+
+def build_resource(config: OpenTelemetryV2Config) -> Resource:
+    attributes: dict[str, str] = {"service.name": config.service_name}
+    if config.deployment_environment:
+        attributes["deployment.environment"] = config.deployment_environment
+    attributes.update(config.resource_attributes)
+    return Resource.create(attributes)
+
+
+def build_tracer_provider(
+    config: OpenTelemetryV2Config,
+    exporter: SpanExporter | None = None,
+    baggage_processor: SpanProcessor | None = None,
+    use_simple_processor: bool | None = None,
+) -> TracerProvider:
+    """Build the shared :class:`TracerProvider`.
+
+    Attach the Baggage processor first (so identity attributes land on each
+    span before any export decision), then add one ``SpanProcessor`` per
+    ``config.exporters`` entry — this is what fans spans out to multiple
+    backends. ``exporter`` and ``use_simple_processor`` are explicit overrides:
+    pass a single exporter to attach exactly that one (used by tests).
+    """
+    provider = TracerProvider(resource=build_resource(config))
+    if baggage_processor is None:
+        baggage_processor = LiteLLMBaggageSpanProcessor(
+            allowed_keys=config.baggage_promoted_keys
+        )
+    provider.add_span_processor(baggage_processor)
+
+    if exporter is not None:
+        provider.add_span_processor(_processor_for(exporter, use_simple_processor))
+        return provider
+
+    # ``config._normalize`` guarantees at least one spec (it folds the top-level
+    # ``exporter``/``endpoint``/``headers`` fields in when ``exporters`` is empty).
+    for spec in config.exporters:
+        exp = _exporter_from_spec(spec)
+        provider.add_span_processor(
+            _processor_for(
+                exp,
+                (
+                    spec.use_simple_processor
+                    if spec.use_simple_processor is not None
+                    else use_simple_processor
+                ),
+            )
+        )
+    return provider
+
+
+def get_tracer(provider: TracerProvider, name: str = "litellm") -> Tracer:
+    return provider.get_tracer(name)
+
+
+def in_memory_provider(
+    config: OpenTelemetryV2Config | None = None,
+) -> tuple[TracerProvider, InMemorySpanExporter]:
+    """Convenience for tests: a provider exporting to an in-memory buffer."""
+    cfg = config or OpenTelemetryV2Config(exporter="in_memory")
+    exporter = InMemorySpanExporter()
+    provider = build_tracer_provider(cfg, exporter=exporter)
+    return provider, exporter

litellm/integrations/otel/plumbing/routing.py

@@ -0,0 +1,101 @@
+"""Per-request multi-tenant tracer routing.
+
+When a request carries team/key vendor credentials in
+``standard_callback_dynamic_params``, its spans must export through a
+``TracerProvider`` whose OTLP headers carry those credentials.
+``TenantTracerCache`` builds and caches one provider per distinct credential
+set, and otherwise hands back the logger's default tracer. This lets a single
+logger fan requests out to many tenants without needing a logger per tenant.
+"""
+
+from collections import OrderedDict
+from typing import Any, Mapping
+
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.trace import Tracer
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.presets import dynamic_otlp_headers
+from litellm.integrations.otel.plumbing.providers import (
+    build_tracer_provider,
+    get_tracer,
+)
+
+# Exporter kinds that ignore headers — never rewritten with dynamic credentials.
+_NON_OTLP_KINDS = ("console", "in_memory", "inmemory", "memory")
+
+# Cap on distinct credential-scoped providers held at once. ``dynamic_params``
+# can be populated from request metadata, so an unbounded cache lets a caller
+# spawn one ``TracerProvider`` (plus its ``BatchSpanProcessor`` background
+# thread) per unique credential set and exhaust the proxy. The LRU bound keeps
+# the working set of active tenants resident while flushing and shutting down
+# evicted providers so their threads are reclaimed.
+_MAX_CACHED_PROVIDERS = 256
+
+
+def _shutdown_provider(provider: TracerProvider) -> None:
+    """Flush + stop an evicted provider's processors (reclaims their threads).
+
+    ``TracerProvider.shutdown`` force-flushes each ``SpanProcessor`` before
+    stopping it, so any spans already handed to a ``BatchSpanProcessor`` are
+    exported rather than dropped. Best-effort: a shutdown failure must not break
+    the request that triggered the eviction.
+    """
+    try:
+        provider.shutdown()
+    except Exception as e:  # pragma: no cover - defensive
+        verbose_logger.debug("OTel V2: error shutting down evicted provider: %s", e)
+
+
+class TenantTracerCache:
+    """Credential-scoped ``TracerProvider`` cache keyed by the dynamic headers."""
+
+    def __init__(
+        self,
+        config: OpenTelemetryV2Config,
+        callback_name: str | None,
+        tracer_name: str,
+    ) -> None:
+        self._config = config
+        self._callback_name = callback_name
+        self._tracer_name = tracer_name
+        self._providers: "OrderedDict[tuple[tuple[str, str], ...], TracerProvider]" = (
+            OrderedDict()
+        )
+
+    def tracer_for(self, default: Tracer, dynamic_params: Any) -> Tracer:
+        """Return the tracer for this request.
+
+        Use ``default`` unless the request's dynamic credentials require a
+        credential-scoped tracer, in which case build (or reuse) one. The cache
+        is a bounded LRU: the least-recently-used provider is flushed and shut
+        down on overflow so its exporter threads don't accumulate.
+        """
+        headers = dynamic_otlp_headers(self._callback_name, dynamic_params)
+        if not headers:
+            return default
+        cache_key = tuple(sorted(headers.items()))
+        provider = self._providers.get(cache_key)
+        if provider is not None:
+            self._providers.move_to_end(cache_key)
+        else:
+            provider = build_tracer_provider(self._config_with_headers(headers))
+            self._providers[cache_key] = provider
+            if len(self._providers) > _MAX_CACHED_PROVIDERS:
+                _, evicted = self._providers.popitem(last=False)
+                _shutdown_provider(evicted)
+        return get_tracer(provider, self._tracer_name)
+
+    def _config_with_headers(self, headers: Mapping[str, str]) -> OpenTelemetryV2Config:
+        """Clone the config, replacing OTLP exporter headers with ``headers``."""
+        header_str = ",".join(f"{key}={value}" for key, value in headers.items())
+        exporters = [
+            (
+                spec
+                if spec.kind.lower() in _NON_OTLP_KINDS
+                else spec.model_copy(update={"headers": header_str})
+            )
+            for spec in self._config.exporters
+        ]
+        return self._config.model_copy(update={"exporters": exporters})

litellm/integrations/otel/presets/__init__.py

@@ -0,0 +1,78 @@
+"""Integration presets — each one returns an :class:`OpenTelemetryV2Config`.
+
+A preset is a callable that reads an integration's env vars and returns an
+``OpenTelemetryV2Config`` describing the exporter destination, the mapper
+vocabularies to apply, and any resource attributes. ``PRESET_BY_CALLBACK``
+maps a callback name (``"arize"``, ``"langfuse_otel"``, ...) to its preset so
+the factory in ``litellm_logging`` can resolve a name and build a single
+``OpenTelemetryV2`` instance from the result.
+"""
+
+from typing import Callable
+
+from litellm.integrations.otel.presets.agentops import agentops_preset
+from litellm.integrations.otel.presets.arize import arize_dynamic_headers, arize_preset
+from litellm.integrations.otel.presets.base import Preset
+from litellm.integrations.otel.presets.langfuse import (
+    langfuse_dynamic_headers,
+    langfuse_preset,
+)
+from litellm.integrations.otel.presets.langtrace import langtrace_preset
+from litellm.integrations.otel.presets.levo import levo_preset
+from litellm.integrations.otel.presets.phoenix import phoenix_preset
+from litellm.integrations.otel.presets.weave import weave_dynamic_headers, weave_preset
+from litellm.types.utils import StandardCallbackDynamicParams
+
+#: Callback name → preset. The ``Preset`` annotation makes mypy verify every
+#: registered value matches the preset interface.
+PRESET_BY_CALLBACK: dict[str, Preset] = {
+    "agentops": agentops_preset,
+    "arize": arize_preset,
+    "arize_phoenix": phoenix_preset,
+    "langfuse_otel": langfuse_preset,
+    "langtrace": langtrace_preset,
+    "levo": levo_preset,
+    "weave_otel": weave_preset,
+}
+
+#: Callback name → per-request OTLP header builder (team/key multi-tenant
+#: routing). Only integrations that support dynamic credentials appear here —
+#: Arize-Phoenix/Langtrace/Levo/AgentOps don't, so they use the logger's
+#: default tracer.
+DYNAMIC_HEADERS_BY_CALLBACK: dict[
+    str, Callable[[StandardCallbackDynamicParams], dict[str, str]]
+] = {
+    "arize": arize_dynamic_headers,
+    "langfuse_otel": langfuse_dynamic_headers,
+    "weave_otel": weave_dynamic_headers,
+}
+
+
+def dynamic_otlp_headers(
+    callback_name: str | None,
+    dynamic_params: StandardCallbackDynamicParams | None,
+) -> dict[str, str] | None:
+    """Per-request OTLP headers for ``callback_name``, or ``None`` if N/A.
+
+    ``None`` means "no per-request routing" — the caller uses its default tracer.
+    """
+    builder = DYNAMIC_HEADERS_BY_CALLBACK.get(callback_name or "")
+    if builder is None or not dynamic_params:
+        return None
+    headers = builder(dynamic_params)
+    return headers or None
+
+
+__all__ = [
+    "PRESET_BY_CALLBACK",
+    "DYNAMIC_HEADERS_BY_CALLBACK",
+    "Preset",
+    "dynamic_otlp_headers",
+    "agentops_preset",
+    "arize_preset",
+    "langfuse_preset",
+    "langtrace_preset",
+    "levo_preset",
+    "phoenix_preset",
+    "weave_preset",
+]

litellm/integrations/otel/presets/agentops.py

@@ -0,0 +1,139 @@
+"""AgentOps preset — OTLP/HTTP to AgentOps' endpoint with a lazily-fetched JWT.
+
+AgentOps authenticates with a short-lived JWT minted from the API key. Fetching
+it is blocking network I/O, so it must never run on the event loop: callback
+construction (where presets are built) can run inside the proxy's async startup
+or, in the SDK, on the first request. Instead of fetching at config-build time,
+this preset registers a custom exporter (``kind="agentops"``) that mints the JWT
+**on its first export** — which the ``BatchSpanProcessor`` runs in its own
+worker thread, off any event loop — and caches it for the process lifetime.
+"""
+
+from typing import Any
+
+import httpx
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm._logging import verbose_logger
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.plumbing.providers import register_exporter_factory
+
+_AGENTOPS_ENDPOINT = "https://otlp.agentops.cloud/v1/traces"
+_AGENTOPS_AUTH_ENDPOINT = "https://api.agentops.ai/v3/auth/token"
+_AGENTOPS_EXPORTER_KIND = "agentops"
+
+
+class _AgentOpsSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    api_key: str | None = Field(default=None, validation_alias="AGENTOPS_API_KEY")
+    service_name: str = Field(
+        default="agentops", validation_alias="AGENTOPS_SERVICE_NAME"
+    )
+    environment: str | None = Field(
+        default=None, validation_alias="AGENTOPS_ENVIRONMENT"
+    )
+
+
+def agentops_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    """Build the AgentOps config without any network I/O.
+
+    The ``agentops`` exporter mints (and caches) the JWT lazily on its first
+    export, so this stays non-blocking. ``project.id`` is therefore not a
+    resource attribute — it is encoded in the JWT, which AgentOps uses to route
+    the trace to the right project.
+    """
+    settings = _AgentOpsSettings()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=_AGENTOPS_EXPORTER_KIND,
+                    endpoint=_AGENTOPS_ENDPOINT,
+                    options=(
+                        {"api_key": settings.api_key} if settings.api_key else None
+                    ),
+                ),
+            ],
+            "resource_attributes": {
+                **base.resource_attributes,
+                "service.name": settings.service_name,
+                "telemetry.sdk.name": "agentops",
+                **(
+                    {"deployment.environment": settings.environment}
+                    if settings.environment
+                    else {}
+                ),
+            },
+        }
+    )
+
+
+def _build_agentops_exporter(spec: ExporterSpec) -> Any:
+    """Factory for the ``agentops`` exporter kind: a lazy-auth OTLP/HTTP exporter."""
+    from opentelemetry.exporter.otlp.proto.http.trace_exporter import (
+        OTLPSpanExporter,
+    )
+
+    class _LazyAuthAgentOpsExporter(OTLPSpanExporter):
+        """OTLP/HTTP exporter that mints the AgentOps JWT on its first export.
+
+        ``export`` runs in the ``BatchSpanProcessor`` worker thread, so the
+        blocking token fetch never touches an event loop. The result is cached
+        after the first attempt (success or failure) so it runs at most once.
+        """
+
+        def __init__(self, *, endpoint: str | None, api_key: str | None) -> None:
+            super().__init__(endpoint=endpoint)
+            self._agentops_api_key = api_key
+            self._auth_resolved = False
+
+        def _ensure_authenticated(self) -> None:
+            if self._auth_resolved:
+                return
+            self._auth_resolved = True
+            if not self._agentops_api_key:
+                return
+            try:
+                token = _fetch_agentops_jwt(self._agentops_api_key).get("token")
+                if token:
+                    # ``_session`` is the requests.Session the base exporter
+                    # POSTs through; updating its Authorization header is how the
+                    # minted JWT reaches every subsequent export.
+                    self._session.headers["Authorization"] = f"Bearer {token}"
+            except Exception as e:
+                verbose_logger.debug("AgentOps JWT fetch failed: %s", e)
+
+        def export(self, spans: Any) -> Any:
+            self._ensure_authenticated()
+            return super().export(spans)
+
+    options = spec.options or {}
+    return _LazyAuthAgentOpsExporter(
+        endpoint=spec.endpoint, api_key=options.get("api_key")
+    )
+
+
+def _fetch_agentops_jwt(api_key: str) -> dict[str, Any]:
+    # Own a short-lived client rather than ``_get_httpx_client()``: that returns
+    # a process-wide cached ``HTTPHandler`` whose connection pool is shared by
+    # every caller, so closing it here would break concurrent/subsequent
+    # requests. This one-shot auth call gets its own client to close.
+    with httpx.Client(timeout=10) as client:
+        response = client.post(
+            url=_AGENTOPS_AUTH_ENDPOINT,
+            headers={"Content-Type": "application/json", "Connection": "keep-alive"},
+            json={"api_key": api_key},
+        )
+        if response.status_code != 200:
+            raise RuntimeError(f"Failed to fetch AgentOps token: {response.text}")
+        return response.json()
+
+
+register_exporter_factory(_AGENTOPS_EXPORTER_KIND, _build_agentops_exporter)

litellm/integrations/otel/presets/arize.py

@@ -0,0 +1,75 @@
+"""Arize preset — OTLP exporter to Arize + OpenInference vocabulary."""
+
+from pydantic import Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm.integrations.arize.arize import ArizeLogger as _V1ArizeLogger
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+class _ArizeSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    # Standard OTLP headers env var, used as the fallback when no Arize
+    # credentials are configured.
+    otlp_traces_headers: str | None = Field(
+        default=None, validation_alias="OTEL_EXPORTER_OTLP_TRACES_HEADERS"
+    )
+
+
+def arize_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    arize_cfg = _V1ArizeLogger.get_arize_config()
+    headers = _arize_headers(arize_cfg)
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=arize_cfg.protocol or "otlp_grpc",
+                    endpoint=arize_cfg.endpoint or "https://otlp.arize.com/v1",
+                    headers=headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference"),
+            "resource_attributes": {
+                **base.resource_attributes,
+                **(
+                    {"model_id": arize_cfg.project_name}
+                    if arize_cfg.project_name
+                    else {}
+                ),
+            },
+        }
+    )
+
+
+def _arize_headers(arize_cfg) -> str | None:
+    pieces = []
+    if arize_cfg.space_id or arize_cfg.space_key:
+        pieces.append(f"space_id={arize_cfg.space_id or arize_cfg.space_key}")
+    if arize_cfg.api_key:
+        pieces.append(f"api_key={arize_cfg.api_key}")
+    if not pieces:
+        # Fall back to the standard OTLP headers env var when no Arize
+        # credentials are configured.
+        return _ArizeSettings().otlp_traces_headers
+    return ",".join(pieces)
+
+
+def arize_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Arize OTLP headers from team/key dynamic params."""
+    headers: dict[str, str] = {}
+    # ``arize_space_key`` is the suggested param and wins over ``arize_space_id``.
+    space = params.get("arize_space_key") or params.get("arize_space_id")
+    if space:
+        headers["arize-space-id"] = space
+    api_key = params.get("arize_api_key")
+    if api_key:
+        headers["api_key"] = api_key
+    return headers

litellm/integrations/otel/presets/base.py

@@ -0,0 +1,25 @@
+"""Preset interface.
+
+A preset is a callable that reads its integration's env vars and produces an
+:class:`OpenTelemetryV2Config` (exporter list + mapper-name list + resource
+attributes). This ``Protocol`` pins that contract so ``PRESET_BY_CALLBACK`` and
+the factory in ``litellm_logging`` are type-checked structurally against it,
+matching the ``AttributeMapper`` protocol the mappers use.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+
+
+@runtime_checkable
+class Preset(Protocol):
+    """Reads an integration's env config and returns an ``OpenTelemetryV2Config``.
+
+    ``config_overrides`` lets one preset layer onto another's config (or onto
+    test-supplied defaults); the factory calls presets with no arguments.
+    """
+
+    def __call__(
+        self, *, config_overrides: OpenTelemetryV2Config | None = None
+    ) -> OpenTelemetryV2Config: ...

litellm/integrations/otel/presets/langfuse.py

@@ -0,0 +1,43 @@
+"""Langfuse-OTEL preset."""
+
+from litellm.integrations.langfuse.langfuse_otel import (
+    LangfuseOtelLogger as _V1Langfuse,
+)
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+def langfuse_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Langfuse.get_langfuse_otel_config()
+    kind = cfg.exporter if isinstance(cfg.exporter, str) else "otlp_http"
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=kind,
+                    endpoint=cfg.endpoint,
+                    headers=cfg.headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "langfuse"),
+        }
+    )
+
+
+def langfuse_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Langfuse OTLP headers from team/key dynamic params."""
+    public_key = params.get("langfuse_public_key")
+    secret_key = params.get("langfuse_secret_key")
+    if public_key and secret_key:
+        return {
+            "Authorization": _V1Langfuse._get_langfuse_authorization_header(
+                public_key=public_key, secret_key=secret_key
+            )
+        }
+    return {}

litellm/integrations/otel/presets/langtrace.py

@@ -0,0 +1,22 @@
+"""Langtrace preset — Langtrace consumes generic OTLP + a vendor mapper."""
+
+from litellm.integrations.otel.model.config import OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+
+
+def langtrace_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    """Compose the Langtrace mapper on top of the customer's OTLP destination.
+
+    Unlike Arize / Phoenix / Langfuse, Langtrace doesn't ship its own endpoint
+    — users point their existing OTLP collector at Langtrace and just
+    need the vendor attribute schema applied to outgoing spans.
+    """
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "mapper_names": ensure_mappers(base.mapper_names, "langtrace"),
+        }
+    )

litellm/integrations/otel/presets/levo.py

@@ -0,0 +1,24 @@
+"""Levo preset — OTLP/HTTP to a Levo collector with org+workspace headers."""
+
+from litellm.integrations.levo.levo import LevoLogger as _V1Levo
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+
+
+def levo_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Levo.get_levo_config()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind="otlp_http",
+                    endpoint=cfg.endpoint,
+                    headers=cfg.otlp_auth_headers,
+                ),
+            ],
+        }
+    )

litellm/integrations/otel/presets/phoenix.py

@@ -0,0 +1,48 @@
+"""Arize-Phoenix preset."""
+
+from pydantic import AliasChoices, Field
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+from litellm.integrations.arize.arize_phoenix import (
+    ArizePhoenixLogger as _V1Phoenix,
+)
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+
+
+class _PhoenixSettings(BaseSettings):
+    model_config = SettingsConfigDict(case_sensitive=False, extra="ignore")
+
+    project_name: str = Field(
+        default="default",
+        validation_alias=AliasChoices(
+            "PHOENIX_PROJECT_NAME", "PHOENIX_COLLECTOR_PROJECT_NAME"
+        ),
+    )
+
+
+def phoenix_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    cfg = _V1Phoenix.get_arize_phoenix_config()
+    headers = cfg.otlp_auth_headers if hasattr(cfg, "otlp_auth_headers") else None
+    project_name = _PhoenixSettings().project_name
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=cfg.protocol if hasattr(cfg, "protocol") else "otlp_http",
+                    endpoint=cfg.endpoint,
+                    headers=headers,
+                ),
+            ],
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference"),
+            "resource_attributes": {
+                **base.resource_attributes,
+                "openinference.project.name": project_name,
+            },
+        }
+    )

litellm/integrations/otel/presets/utils.py

@@ -0,0 +1,16 @@
+"""Shared helpers for the integration presets."""
+
+from typing import Iterable
+
+
+def ensure_mappers(mapper_names: Iterable[str], *names: str) -> list[str]:
+    """Return ``mapper_names`` with each of ``names`` appended if not already present.
+
+    Order is preserved and duplicates are skipped, so composing several presets
+    (or re-applying one) never double-adds a vocabulary.
+    """
+    result = list(mapper_names)
+    for name in names:
+        if name not in result:
+            result.append(name)
+    return result

litellm/integrations/otel/presets/weave.py

@@ -0,0 +1,43 @@
+"""Weave (W&B) preset."""
+
+from litellm.integrations.otel.model.config import ExporterSpec, OpenTelemetryV2Config
+from litellm.integrations.otel.presets.utils import ensure_mappers
+from litellm.integrations.weave.weave_otel import (
+    _get_weave_authorization_header,
+    get_weave_otel_config,
+)
+from litellm.types.utils import StandardCallbackDynamicParams
+
+
+def weave_preset(
+    *,
+    config_overrides: OpenTelemetryV2Config | None = None,
+) -> OpenTelemetryV2Config:
+    weave_cfg = get_weave_otel_config()
+    base = config_overrides or OpenTelemetryV2Config()
+    return base.model_copy(
+        update={
+            "exporters": [
+                *base.exporters,
+                ExporterSpec(
+                    kind=weave_cfg.protocol or "otlp_http",
+                    endpoint=weave_cfg.endpoint,
+                    headers=weave_cfg.otlp_auth_headers,
+                ),
+            ],
+            # Weave consumes OpenInference + a small Weave-specific overlay.
+            "mapper_names": ensure_mappers(base.mapper_names, "openinference", "weave"),
+        }
+    )
+
+
+def weave_dynamic_headers(params: StandardCallbackDynamicParams) -> dict[str, str]:
+    """Per-request Weave OTLP headers from team/key dynamic params."""
+    headers: dict[str, str] = {}
+    api_key = params.get("wandb_api_key")
+    if api_key:
+        headers["Authorization"] = _get_weave_authorization_header(api_key=api_key)
+    project_id = params.get("weave_project_id")
+    if project_id:
+        headers["project_id"] = project_id
+    return headers

litellm/integrations/otel/runtime.py

@@ -0,0 +1,38 @@
+"""SDK-free entrypoints for proxy-core call sites (auth, …).
+
+Proxy code may run without the OpenTelemetry SDK installed, so it must not import
+``litellm.integrations.otel.logger`` (which imports the SDK at module scope) at
+module load. These wrappers import it lazily and no-op when the SDK is absent or
+V2 is not the active logger — so a call site can wrap a request phase or seed
+identity unconditionally.
+"""
+
+from contextlib import contextmanager
+from typing import Any, Iterator
+
+
+@contextmanager
+def phase_span(name: str) -> "Iterator[Any]":
+    """Run a request phase inside a live active span so its DB/service calls nest.
+
+    Yields ``None`` (a plain no-op) when the OTel SDK is unavailable or V2 is not
+    the active logger.
+    """
+    try:
+        from litellm.integrations.otel.logger import phase_span as _phase_span
+    except Exception:
+        yield None
+        return
+    with _phase_span(name) as span:
+        yield span
+
+
+def seed_request_identity(user_api_key_dict: Any, model: Any = None) -> None:
+    """Seed request-identity Baggage at the auth boundary (no-op without V2)."""
+    try:
+        from litellm.integrations.otel.logger import (
+            seed_request_identity as _seed_request_identity,
+        )
+    except Exception:
+        return
+    _seed_request_identity(user_api_key_dict, model=model)

litellm/integrations/prometheus.py

@@ -511,6 +511,23 @@ class PrometheusLogger(CustomLogger):
                 labelnames=self.get_labels_for_metric("litellm_cached_tokens_metric"),
             )
 
+            # Provider prompt-caching metrics
+            self.litellm_provider_cache_read_input_tokens_metric = self._counter_factory(
+                name="litellm_provider_cache_read_input_tokens_metric",
+                documentation="Total prompt/input tokens read from provider prompt cache (e.g. OpenAI/Anthropic/Gemini/Bedrock)",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_provider_cache_read_input_tokens_metric"
+                ),
+            )
+
+            self.litellm_provider_cache_creation_input_tokens_metric = self._counter_factory(
+                name="litellm_provider_cache_creation_input_tokens_metric",
+                documentation="Total prompt/input tokens written to provider prompt cache (e.g. Anthropic/Bedrock)",
+                labelnames=self.get_labels_for_metric(
+                    "litellm_provider_cache_creation_input_tokens_metric"
+                ),
+            )
+
             # User and Team count metrics
             self.litellm_total_users_metric = self._gauge_factory(
                 "litellm_total_users",
@@ -1458,11 +1475,11 @@ class PrometheusLogger(CustomLogger):
         """
         cache_hit = standard_logging_payload.get("cache_hit")
 
-        # Only track if cache_hit has a definite value (True or False)
         if cache_hit is None:
-            return
-
-        if cache_hit is True:
+            # Historically these metrics only tracked LiteLLM caching.
+            # Provider prompt-caching metrics are still emitted below.
+            pass
+        elif cache_hit is True:
             # Increment cache hits counter
             PrometheusLogger._inc_labeled_counter(
                 self,
@@ -1493,6 +1510,51 @@ class PrometheusLogger(CustomLogger):
                 label_context=label_context,
             )
 
+        # Provider prompt caching metrics are independent of LiteLLM cache_hit.
+        provider_cache_read_tokens = 0
+        provider_cache_creation_tokens = 0
+        usage_obj = (standard_logging_payload.get("metadata", {}) or {}).get(
+            "usage_object"
+        )
+        if isinstance(usage_obj, dict):
+            # Prefer explicit provider cache fields when available.
+            _read = usage_obj.get("cache_read_input_tokens")
+            _write = usage_obj.get("cache_creation_input_tokens")
+
+            if isinstance(_read, int):
+                provider_cache_read_tokens = _read
+            if isinstance(_write, int):
+                provider_cache_creation_tokens = _write
+
+            # Fallback to prompt_tokens_details.cached_tokens (common normalization point).
+            # Only fallback when the explicit field is genuinely absent (None).
+            if _read is None:
+                prompt_details = usage_obj.get("prompt_tokens_details")
+                if isinstance(prompt_details, dict):
+                    cached_tokens = prompt_details.get("cached_tokens")
+                    if isinstance(cached_tokens, int):
+                        provider_cache_read_tokens = cached_tokens
+
+            if provider_cache_read_tokens > 0:
+                PrometheusLogger._inc_labeled_counter(
+                    self,
+                    self.litellm_provider_cache_read_input_tokens_metric,
+                    "litellm_provider_cache_read_input_tokens_metric",
+                    enum_values,
+                    label_context=label_context,
+                    amount=float(provider_cache_read_tokens),
+                )
+
+            if provider_cache_creation_tokens > 0:
+                PrometheusLogger._inc_labeled_counter(
+                    self,
+                    self.litellm_provider_cache_creation_input_tokens_metric,
+                    "litellm_provider_cache_creation_input_tokens_metric",
+                    enum_values,
+                    label_context=label_context,
+                    amount=float(provider_cache_creation_tokens),
+                )
+
     async def _increment_remaining_budget_metrics(
         self,
         user_api_team: Optional[str],

litellm/litellm_core_utils/exception_mapping_utils.py

@@ -87,6 +87,7 @@ class ExceptionCheckers:
             "is longer than the model's context length",
             "input tokens exceed the configured limit",
             "`inputs` tokens + `max_new_tokens` must be",
+            "exceeds the available context size",  # llama.cpp/Lemonade
             "exceeds the maximum number of tokens allowed",  # Gemini
         ]
         for substring in known_exception_substrings:
@@ -891,12 +892,14 @@ def exception_type(  # type: ignore  # noqa: PLR0915
                         response=getattr(original_exception, "response", None),
                         litellm_debug_info=extra_information,
                     )
-                elif "model's maximum context limit" in error_str:
+                elif ExceptionCheckers.is_error_str_context_window_exceeded(error_str):
                     exception_mapping_worked = True
                     raise ContextWindowExceededError(
                         message=f"{custom_llm_provider.capitalize()}Exception: Context Window Error - {error_str}",
                         model=model,
                         llm_provider=custom_llm_provider,
+                        response=getattr(original_exception, "response", None),
+                        litellm_debug_info=extra_information,
                     )
                 elif "token_quota_reached" in error_str:
                     exception_mapping_worked = True

litellm/litellm_core_utils/fallback_utils.py

@@ -47,8 +47,9 @@ async def async_completion_with_fallbacks(**kwargs):
             completion_kwargs = safe_deep_copy(base_kwargs)
             # Handle dictionary fallback configurations
             if isinstance(fallback, dict):
-                model = fallback.pop("model", original_model)
-                completion_kwargs.update(fallback)
+                fallback_config = safe_deep_copy(dict(fallback))
+                model = fallback_config.pop("model", original_model)
+                completion_kwargs.update(fallback_config)
             else:
                 model = fallback
 

litellm/litellm_core_utils/get_llm_provider_logic.py

@@ -1,3 +1,4 @@
+import re
 from typing import Optional, Tuple
 from urllib.parse import urlparse
 
@@ -71,6 +72,25 @@ def _is_azure_claude_model(model: str) -> bool:
         return False
 
 
+_CLAUDE_PATTERN = re.compile(r"^claude-[a-z]+-\d+-\d+(?:-\d{8})?$", re.IGNORECASE)
+
+
+def _matches_claude_model_pattern(model: str) -> bool:
+    """
+    Check if a model string matches the Claude model naming pattern.
+
+    Matches patterns like:
+    - claude-opus-4-7
+    - claude-sonnet-4-6
+    - claude-haiku-4-5
+    - claude-opus-5-1-20270101 (with optional date suffix)
+
+    This allows future Claude models to be routed to the Anthropic provider
+    without requiring updates to model_prices_and_context_window.json.
+    """
+    return _CLAUDE_PATTERN.match(model) is not None
+
+
 def handle_cohere_chat_model_custom_llm_provider(
     model: str, custom_llm_provider: Optional[str] = None
 ) -> Tuple[str, Optional[str]]:
@@ -353,6 +373,9 @@ def get_llm_provider(  # noqa: PLR0915
                     elif endpoint == "https://api.lambda.ai/v1":
                         custom_llm_provider = "lambda_ai"
                         dynamic_api_key = get_secret_str("LAMBDA_API_KEY")
+                    elif endpoint == "https://api.inceptionlabs.ai/v1":
+                        custom_llm_provider = "inception"
+                        dynamic_api_key = get_secret_str("INCEPTION_API_KEY")
                     elif endpoint == "https://api.hyperbolic.xyz/v1":
                         custom_llm_provider = "hyperbolic"
                         dynamic_api_key = get_secret_str("HYPERBOLIC_API_KEY")
@@ -398,6 +421,9 @@ def get_llm_provider(  # noqa: PLR0915
                 custom_llm_provider = "anthropic_text"
             else:
                 custom_llm_provider = "anthropic"
+        ## anthropic - pattern-based matching for future Claude models
+        elif _matches_claude_model_pattern(model):
+            custom_llm_provider = "anthropic"
         ## cohere
         elif model in litellm.cohere_models or model in litellm.cohere_embedding_models:
             custom_llm_provider = "cohere"
@@ -931,6 +957,13 @@ def _get_openai_compatible_provider_info(  # noqa: PLR0915
         ) = litellm.LambdaAIChatConfig()._get_openai_compatible_provider_info(
             api_base, api_key
         )
+    elif custom_llm_provider == "inception":
+        (
+            api_base,
+            dynamic_api_key,
+        ) = litellm.InceptionChatConfig()._get_openai_compatible_provider_info(
+            api_base, api_key
+        )
     elif custom_llm_provider == "hyperbolic":
         (
             api_base,

litellm/litellm_core_utils/litellm_logging.py

@@ -1612,6 +1612,90 @@ class Logging(LiteLLMLoggingBaseClass):
     ) -> Optional[float]:
         return self._response_cost_calculator(result=result, cache_hit=cache_hit)
 
+    @staticmethod
+    def _is_sync_litellm_request(litellm_params: dict) -> bool:
+        """True for sync SDK entrypoints (``completion``), false for async (``acompletion``, etc.)."""
+        return (
+            litellm_params.get(CallTypes.acompletion.value, False) is not True
+            and litellm_params.get(CallTypes.aresponses.value, False) is not True
+            and litellm_params.get(CallTypes.aembedding.value, False) is not True
+            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
+            and litellm_params.get(CallTypes.atranscription.value, False) is not True
+        )
+
+    def _is_assembled_stream_success(self, result=None) -> bool:
+        """Final assembled stream export (not a per-chunk success call).
+
+        Per-chunk callers pass a ``ModelResponseStream`` (or ``None``); the
+        final assembled response is any other non-``None`` value (typically a
+        ``ModelResponse``). Treating a chunk as the assembled response would
+        prematurely set the ``has_dispatched_final_stream_success`` dedup
+        guard and silently suppress the real final stream log.
+        """
+        if self.stream is not True:
+            return False
+        if result is not None and not isinstance(result, ModelResponseStream):
+            return True
+        return (
+            "async_complete_streaming_response" in self.model_call_details
+            or self.model_call_details.get("complete_streaming_response") is not None
+        )
+
+    async def dispatch_success_handlers(
+        self,
+        result=None,
+        start_time=None,
+        end_time=None,
+        cache_hit=None,
+        prefer_async_handlers: bool = False,
+        **kwargs,
+    ) -> None:
+        """Route success logging to async and/or sync handlers for this request.
+
+        ``prefer_async_handlers`` only bypasses the sync-SDK-only shortcut (e.g.
+        ``async for`` on a stream from ``completion()``). Legacy string callbacks
+        still run via ``executor.submit(success_handler)`` when configured.
+        """
+        from litellm.litellm_core_utils.thread_pool_executor import executor
+
+        if self._is_assembled_stream_success(result):
+            if self.model_call_details.get("has_dispatched_final_stream_success"):
+                return
+            self.model_call_details["has_dispatched_final_stream_success"] = True
+
+        litellm_params = self.model_call_details.get("litellm_params", {}) or {}
+        sync_sdk = self._is_sync_litellm_request(litellm_params)
+        passthrough = self.call_type == CallTypes.pass_through.value
+        if sync_sdk and not prefer_async_handlers and not passthrough:
+            self.success_handler(
+                result,
+                start_time=start_time,
+                end_time=end_time,
+                cache_hit=cache_hit,
+                **kwargs,
+            )
+            return
+
+        await self.async_success_handler(
+            result,
+            start_time=start_time,
+            end_time=end_time,
+            cache_hit=cache_hit,
+            **kwargs,
+        )
+
+        if not self._should_run_sync_callbacks_for_async_calls():
+            return
+
+        executor.submit(
+            self.success_handler,
+            result,
+            start_time=start_time,
+            end_time=end_time,
+            cache_hit=cache_hit,
+            **kwargs,
+        )
+
     def should_run_logging(
         self,
         event_type: Literal[
@@ -2034,13 +2118,7 @@ class Logging(LiteLLMLoggingBaseClass):
             standard_logging_object=kwargs.get("standard_logging_object", None),
         )
         litellm_params = self.model_call_details.get("litellm_params", {})
-        is_sync_request = (
-            litellm_params.get(CallTypes.acompletion.value, False) is not True
-            and litellm_params.get(CallTypes.aresponses.value, False) is not True
-            and litellm_params.get(CallTypes.aembedding.value, False) is not True
-            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
-            and litellm_params.get(CallTypes.atranscription.value, False) is not True
-        )
+        is_sync_request = self._is_sync_litellm_request(litellm_params)
         try:
             ## BUILD COMPLETE STREAMED RESPONSE
             complete_streaming_response: Optional[
@@ -2496,9 +2574,11 @@ class Logging(LiteLLMLoggingBaseClass):
         print_verbose(
             "Logging Details LiteLLM-Async Success Call, cache_hit={}".format(cache_hit)
         )
-        if not self.should_run_logging(
+        if not self._is_assembled_stream_success(
+            result
+        ) and not self.should_run_logging(
             event_type="async_success"
-        ):  # prevent double logging
+        ):  # prevent double logging (non-streaming)
             return
 
         ## CALCULATE COST FOR BATCH JOBS
@@ -2948,13 +3028,7 @@ class Logging(LiteLLMLoggingBaseClass):
         ):  # prevent double logging
             return
         litellm_params = self.model_call_details.get("litellm_params", {})
-        is_sync_request = (
-            litellm_params.get(CallTypes.acompletion.value, False) is not True
-            and litellm_params.get(CallTypes.aresponses.value, False) is not True
-            and litellm_params.get(CallTypes.aembedding.value, False) is not True
-            and litellm_params.get(CallTypes.aimage_generation.value, False) is not True
-            and litellm_params.get(CallTypes.atranscription.value, False) is not True
-        )
+        is_sync_request = self._is_sync_litellm_request(litellm_params)
 
         try:
             start_time, end_time = self._failure_handler_helper_fn(
@@ -3718,6 +3792,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
     try:
         custom_logger_init_args = custom_logger_init_args or {}
         if logging_integration == "agentops":  # Add AgentOps initialization
+            _v2 = _maybe_construct_otel_v2("agentops", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             for callback in _in_memory_loggers:
                 if isinstance(callback, AgentOps):
                     return callback  # type: ignore
@@ -3870,6 +3947,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_opik_logger)
             return _opik_logger  # type: ignore
         elif logging_integration == "arize":
+            _v2 = _maybe_construct_otel_v2("arize", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
                 OpenTelemetryConfig,
@@ -3899,6 +3979,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_arize_otel_logger)
             return _arize_otel_logger  # type: ignore
         elif logging_integration == "arize_phoenix":
+            _v2 = _maybe_construct_otel_v2("arize_phoenix", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
                 OpenTelemetryConfig,
@@ -3929,6 +4012,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_arize_phoenix_otel_logger)
             return _arize_phoenix_otel_logger  # type: ignore
         elif logging_integration == "levo":
+            _v2 = _maybe_construct_otel_v2("levo", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.levo.levo import LevoLogger
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
@@ -3954,6 +4040,28 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_levo_otel_logger)
             return _levo_otel_logger  # type: ignore
         elif logging_integration == "otel":
+            # Gate the new typed V2 adapter behind LITELLM_OTEL_V2. When off,
+            # the legacy 3,227-line god-class is used unchanged. The two are
+            # never registered simultaneously — the dedup loop below treats
+            # any module under ``litellm.integrations.otel`` or
+            # ``litellm.integrations.opentelemetry`` as "the OTel callback".
+            from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+            if is_otel_v2_enabled():
+                from litellm.integrations.otel.logger import OpenTelemetryV2
+
+                for callback in _in_memory_loggers:
+                    if type(callback) is OpenTelemetryV2:
+                        return callback  # type: ignore
+                otel_logger_v2 = OpenTelemetryV2(
+                    **_get_custom_logger_settings_from_proxy_server(
+                        callback_name=logging_integration
+                    )
+                )
+                _in_memory_loggers.append(otel_logger_v2)
+                _maybe_auto_initialize_arize_phoenix(_in_memory_loggers)
+                return otel_logger_v2  # type: ignore
+
             from litellm.integrations.opentelemetry import OpenTelemetry
 
             for callback in _in_memory_loggers:
@@ -4092,6 +4200,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
         elif logging_integration == "langtrace":
             if "LANGTRACE_API_KEY" not in os.environ:
                 raise ValueError("LANGTRACE_API_KEY not found in environment variables")
+            _v2 = _maybe_construct_otel_v2("langtrace", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
 
             from litellm.integrations.opentelemetry import (
                 OpenTelemetry,
@@ -4132,6 +4243,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(langfuse_logger)
             return langfuse_logger  # type: ignore
         elif logging_integration == "langfuse_otel":
+            _v2 = _maybe_construct_otel_v2("langfuse_otel", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.langfuse.langfuse_otel import LangfuseOtelLogger
 
             for callback in _in_memory_loggers:
@@ -4148,6 +4262,9 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
             _in_memory_loggers.append(_otel_logger)
             return _otel_logger  # type: ignore
         elif logging_integration == "weave_otel":
+            _v2 = _maybe_construct_otel_v2("weave_otel", _in_memory_loggers)
+            if _v2 is not None:
+                return _v2  # type: ignore
             from litellm.integrations.opentelemetry import OpenTelemetryConfig
             from litellm.integrations.weave.weave_otel import (
                 WeaveOtelLogger,
@@ -4296,6 +4413,42 @@ def _init_custom_logger_compatible_class(  # noqa: PLR0915
     return None
 
 
+def _maybe_construct_otel_v2(
+    callback_name: str, _in_memory_loggers: list
+) -> Optional[Any]:
+    """If ``LITELLM_OTEL_V2`` is on, build (or reuse) a single ``OpenTelemetryV2``
+    instance configured via the preset for ``callback_name``.
+
+    Returns ``None`` when V2 is off OR when there's no preset registered for
+    ``callback_name`` — callers should then fall through to the legacy path.
+    """
+    from litellm.integrations.otel.model.config import is_otel_v2_enabled
+
+    if not is_otel_v2_enabled():
+        return None
+    from litellm.integrations.otel.logger import OpenTelemetryV2
+    from litellm.integrations.otel.presets import PRESET_BY_CALLBACK
+
+    preset_fn = PRESET_BY_CALLBACK.get(callback_name)
+    if preset_fn is None:
+        return None
+    for callback in _in_memory_loggers:
+        if (
+            isinstance(callback, OpenTelemetryV2)
+            and getattr(callback, "callback_name", None) == callback_name
+        ):
+            return callback
+    try:
+        config = preset_fn()
+    except Exception:
+        # If env vars are missing or the preset raises, defer to the legacy path
+        # so customers get the same error story they had before V2 landed.
+        return None
+    v2_logger = OpenTelemetryV2(config=config, callback_name=callback_name)
+    _in_memory_loggers.append(v2_logger)
+    return v2_logger
+
+
 def _maybe_auto_initialize_arize_phoenix(_in_memory_loggers: list) -> None:
     """
     Auto-initialize ArizePhoenixLogger when Phoenix env vars are detected.
@@ -5147,8 +5300,12 @@ class StandardLoggingPayloadSetup:
                     tb_lines[:MAXIMUM_TRACEBACK_LINES_TO_LOG]
                 )  # Limit to first 100 lines
 
-        # Get additional error details
-        error_message = str(original_exception)
+        explicit_message = getattr(original_exception, "message", None)
+        error_message = (
+            explicit_message
+            if isinstance(explicit_message, str) and explicit_message
+            else str(original_exception)
+        )
 
         return StandardLoggingPayloadErrorInformation(
             error_code=error_status,

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -34,6 +34,14 @@ _IMAGE_RESPONSE_CALL_TYPES = frozenset(
 _VALID_DATA_RESIDENCIES = frozenset(r.value for r in DataResidency)
 
 
+def _get_token_detail_value(details: object, key: str) -> Optional[int]:
+    if isinstance(details, dict):
+        value = details.get(key)
+    else:
+        value = getattr(details, key, None)
+    return value if isinstance(value, int) else None
+
+
 def _is_above_128k(tokens: float) -> bool:
     if tokens > 128000:
         return True
@@ -870,17 +878,47 @@ def calculate_image_response_cost_from_usage(
             cached_tokens=0,
         )
 
+    output_tokens_details = getattr(usage, "completion_tokens_details", None)
+    if output_tokens_details is None:
+        output_tokens_details = getattr(usage, "output_tokens_details", None)
+
+    if output_tokens_details is None:
+        completion_tokens_details = CompletionTokensDetailsWrapper(
+            text_tokens=0,
+            image_tokens=completion_tokens,
+            reasoning_tokens=0,
+            audio_tokens=0,
+        )
+    else:
+        text_tokens = _get_token_detail_value(output_tokens_details, "text_tokens") or 0
+        image_tokens = (
+            _get_token_detail_value(output_tokens_details, "image_tokens") or 0
+        )
+        audio_tokens = (
+            _get_token_detail_value(output_tokens_details, "audio_tokens") or 0
+        )
+        reasoning_tokens = (
+            _get_token_detail_value(output_tokens_details, "reasoning_tokens") or 0
+        )
+        known_output_tokens = (
+            text_tokens + image_tokens + audio_tokens + reasoning_tokens
+        )
+        if completion_tokens > known_output_tokens:
+            text_tokens += completion_tokens - known_output_tokens
+
+        completion_tokens_details = CompletionTokensDetailsWrapper(
+            text_tokens=text_tokens,
+            image_tokens=image_tokens,
+            reasoning_tokens=reasoning_tokens,
+            audio_tokens=audio_tokens,
+        )
+
     normalized_usage = Usage(
         prompt_tokens=prompt_tokens,
         completion_tokens=completion_tokens,
         total_tokens=total_tokens,
         prompt_tokens_details=prompt_tokens_details,
-        completion_tokens_details=CompletionTokensDetailsWrapper(
-            text_tokens=0,
-            image_tokens=completion_tokens,
-            reasoning_tokens=0,
-            audio_tokens=0,
-        ),
+        completion_tokens_details=completion_tokens_details,
     )
 
     prompt_cost, completion_cost = generic_cost_per_token(

litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py

@@ -144,6 +144,19 @@ async def convert_to_streaming_response_async(response_object: Optional[dict] =
 
     choice_list: List[StreamingChoices] = []
 
+    if not response_object.get("choices"):
+        from litellm.exceptions import APIError
+
+        raise APIError(
+            status_code=500,
+            message=(
+                "LiteLLM: provider returned a response with no 'choices'. "
+                f"Raw keys: {list(response_object.keys())}"
+            ),
+            llm_provider="",
+            model="",
+        )
+
     for idx, choice in enumerate(response_object["choices"]):
         if (
             choice["message"].get("tool_calls", None) is not None
@@ -213,6 +226,20 @@ def convert_to_streaming_response(response_object: Optional[dict] = None):
 
     model_response_object = ModelResponseStream()
     choice_list: List[StreamingChoices] = []
+
+    if not response_object.get("choices"):
+        from litellm.exceptions import APIError
+
+        raise APIError(
+            status_code=500,
+            message=(
+                "LiteLLM: provider returned a response with no 'choices'. "
+                f"Raw keys: {list(response_object.keys())}"
+            ),
+            llm_provider="",
+            model="",
+        )
+
     for idx, choice in enumerate(response_object["choices"]):
         delta = Delta(**choice["message"])
         finish_reason = choice.get("finish_reason", None)
@@ -536,9 +563,20 @@ def convert_to_model_response_object(  # noqa: PLR0915
                 return convert_to_streaming_response(response_object=response_object)
             choice_list: List[Choices] = []
 
-            assert response_object["choices"] is not None and isinstance(
+            if not response_object.get("choices") or not isinstance(
                 response_object["choices"], Iterable
-            )
+            ):
+                from litellm.exceptions import APIError
+
+                raise APIError(
+                    status_code=500,
+                    message=(
+                        "LiteLLM: provider returned a response with no 'choices'. "
+                        f"Raw keys: {list(response_object.keys())}"
+                    ),
+                    llm_provider="",
+                    model="",
+                )
 
             for idx, choice in enumerate(response_object["choices"]):
                 ## HANDLE JSON MODE - anthropic returns single function call]
@@ -816,7 +854,12 @@ def convert_to_model_response_object(  # noqa: PLR0915
                 model_response_object.results = response_object["results"]
 
             return model_response_object
-    except Exception:
+    except Exception as e:
+        from litellm.exceptions import APIError
+
+        if isinstance(e, APIError):
+            raise
+
         received_args = dict(
             response_object=response_object,
             model_response_object=model_response_object,

litellm/litellm_core_utils/logging_worker.py

@@ -3,6 +3,7 @@
 
 import asyncio
 import contextvars
+import logging
 from typing import Coroutine, Optional
 import atexit
 from typing_extensions import TypedDict
@@ -494,31 +495,43 @@ class LoggingWorker:
             processed = 0
             start_time = loop.time()
 
-            while not self._queue.empty() and processed < MAX_ITERATIONS_TO_CLEAR_QUEUE:
-                if loop.time() - start_time >= MAX_TIME_TO_CLEAR_QUEUE:
-                    self._safe_log(
-                        "warning",
-                        f"[LoggingWorker] atexit: Reached time limit ({MAX_TIME_TO_CLEAR_QUEUE}s), stopping flush",
-                    )
-                    break
+            # logging.raiseExceptions is a process-wide global; scope the
+            # suppression to just the drain loop, where shutdown callbacks may
+            # log to already-closed handler streams, so other threads keep their
+            # logging error reporting for as little of the window as possible.
+            previous_raise_exceptions = logging.raiseExceptions
+            logging.raiseExceptions = False
+            try:
+                while (
+                    not self._queue.empty()
+                    and processed < MAX_ITERATIONS_TO_CLEAR_QUEUE
+                ):
+                    if loop.time() - start_time >= MAX_TIME_TO_CLEAR_QUEUE:
+                        self._safe_log(
+                            "warning",
+                            f"[LoggingWorker] atexit: Reached time limit ({MAX_TIME_TO_CLEAR_QUEUE}s), stopping flush",
+                        )
+                        break
 
-                try:
-                    task = self._queue.get_nowait()
-                except asyncio.QueueEmpty:
-                    break
+                    try:
+                        task = self._queue.get_nowait()
+                    except asyncio.QueueEmpty:
+                        break
 
-                # Run the coroutine synchronously in new loop
-                # Note: We run the coroutine directly, not via create_task,
-                # since we're in a new event loop context
-                try:
-                    loop.run_until_complete(task["coroutine"])
-                    processed += 1
-                except Exception:
-                    # Silent failure to not break user's program
-                    pass
-                finally:
-                    # Clear reference to prevent memory leaks
-                    task = None
+                    # Run the coroutine synchronously in new loop
+                    # Note: We run the coroutine directly, not via create_task,
+                    # since we're in a new event loop context
+                    try:
+                        loop.run_until_complete(task["coroutine"])
+                        processed += 1
+                    except Exception:
+                        # Silent failure to not break user's program
+                        pass
+                    finally:
+                        # Clear reference to prevent memory leaks
+                        task = None
+            finally:
+                logging.raiseExceptions = previous_raise_exceptions
 
             self._safe_log(
                 "info",

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -3,6 +3,7 @@ Common utility functions used for translating messages across providers
 """
 
 import io
+import json
 import mimetypes
 import re
 from os import PathLike
@@ -132,6 +133,39 @@ def strip_none_values_from_message(message: AllMessageValues) -> AllMessageValue
     return cast(AllMessageValues, {k: v for k, v in message.items() if v is not None})
 
 
+def extract_search_results_text(search_results: object) -> str:
+    """
+    Extract model-visible text from OpenAI tool-message ``search_results``.
+
+    Used by token estimators and TPM limiters so large search result payloads
+    cannot bypass preflight checks via a small ``content`` field.
+
+    Counts every string field forwarded on Bedrock ``SearchResultBlock``:
+    ``source``, ``title``, ``content[].text``, and ``citations``.
+    """
+    if not isinstance(search_results, list):
+        return ""
+    texts = ""
+    for result in search_results:
+        if not isinstance(result, dict):
+            continue
+        for key in ("source", "title"):
+            value = result.get(key)
+            if isinstance(value, str):
+                texts += value
+        content = result.get("content")
+        if isinstance(content, list):
+            for block in content:
+                if isinstance(block, dict):
+                    text = block.get("text")
+                    if isinstance(text, str):
+                        texts += text
+        citations = result.get("citations")
+        if citations is not None:
+            texts += json.dumps(citations, separators=(",", ":"))
+    return texts
+
+
 def convert_content_list_to_str(
     message: Union[AllMessageValues, ChatCompletionResponseMessage],
 ) -> str:
@@ -152,6 +186,7 @@ def convert_content_list_to_str(
         elif message_content is not None and isinstance(message_content, str):
             texts = message_content
 
+    texts += extract_search_results_text(message.get("search_results"))
     return texts
 
 
@@ -815,7 +850,47 @@ def extract_file_data(file_data: FileTypes) -> ExtractedFileData:
 # ---------------------------------------------------------------------------
 
 
-def unpack_defs(schema: dict, defs: dict) -> None:
+def _estimate_json_bytes(obj: Any) -> int:
+    """Estimate the JSON-serialised byte size of ``obj`` without materialising
+    JSON. Walks iteratively (no recursion stack risk).
+
+    String length is read via ``len()`` (O(1) on Python ``str``) so a target
+    containing a 100MB description costs ~one walk step, not a 100MB
+    serialisation. Escape sequences are not counted exactly, so this is an
+    approximation -- but always within a small constant factor of the real
+    serialised size, which is what a schema-bomb budget needs.
+    """
+    total = 0
+    stack: list = [obj]
+    while stack:
+        x = stack.pop()
+        if isinstance(x, dict):
+            total += 2  # `{}`
+            for k, v in x.items():
+                total += len(str(k)) + 4  # `"k":,`
+                stack.append(v)
+        elif isinstance(x, list):
+            total += 2  # `[]`
+            total += max(0, len(x) - 1)  # commas between items
+            stack.extend(x)
+        elif isinstance(x, str):
+            total += len(x) + 2
+        elif isinstance(x, bool):  # bool subclasses int -- check first
+            total += 4 if x else 5
+        elif x is None:
+            total += 4
+        elif isinstance(x, (int, float)):
+            total += 24  # generous upper bound for stringified numbers
+        else:
+            total += 24
+    return total
+
+
+def unpack_defs(
+    schema: dict,
+    defs: dict,
+    max_inlined_bytes: Optional[int] = None,
+) -> None:
     """Expand *all* ``$ref`` entries pointing into ``$defs`` / ``definitions``.
 
     This utility walks the entire schema tree (dicts and lists) so it naturally
@@ -825,6 +900,15 @@ def unpack_defs(schema: dict, defs: dict) -> None:
     It mutates *schema* in-place and does **not** return anything.  The helper
     keeps memory overhead low by resolving nodes as it encounters them rather
     than materialising a fully dereferenced copy first.
+
+    ``max_inlined_bytes`` caps the cumulative JSON-byte size of every target
+    that has been inlined and is checked *before* each ``copy.deepcopy``, so
+    an oversized expansion is rejected without first materialising it. A byte
+    bound is the universal measure of expansion -- it simultaneously caps
+    ref-count fan-out, node-count amplification, and scalar-byte amplification
+    (a target containing a large string, ``const``, or ``enum`` entry).
+    Defaults to ``None`` (unbounded) so existing callers are unaffected;
+    raises ``ValueError`` on overflow.
     """
 
     import copy
@@ -844,6 +928,7 @@ def unpack_defs(schema: dict, defs: dict) -> None:
     queue: deque[
         tuple[Any, Union[dict, list, None], Union[str, int, None], dict, set]
     ] = deque([(schema, None, None, root_defs, set())])
+    inlined_bytes = 0
 
     while queue:
         node, parent, key, active_defs, ref_chain = queue.popleft()
@@ -864,6 +949,16 @@ def unpack_defs(schema: dict, defs: dict) -> None:
                 if target_schema is None:
                     continue
 
+                if max_inlined_bytes is not None:
+                    inlined_bytes += _estimate_json_bytes(target_schema)
+                    if inlined_bytes > max_inlined_bytes:
+                        raise ValueError(
+                            f"unpack_defs: inlined schema exceeded the "
+                            f"{max_inlined_bytes:,}-byte budget. Refusing to "
+                            f"deep-copy further to prevent schema-bomb "
+                            f"resource exhaustion."
+                        )
+
                 # Merge defs from the target to capture nested definitions
                 child_defs = {
                     **active_defs,
@@ -911,6 +1006,61 @@ def unpack_defs(schema: dict, defs: dict) -> None:
                 queue.append((item, node, idx, active_defs, ref_chain))
 
 
+def _has_legacy_defs(schema: object) -> bool:
+    if not isinstance(schema, dict):
+        return False
+    components = schema.get("components")
+    return "definitions" in schema or (
+        isinstance(components, dict) and isinstance(components.get("schemas"), dict)
+    )
+
+
+# Schema-bomb budget for ``unpack_legacy_defs``: cap the cumulative JSON-byte
+# size of every inlined target. A byte cap is the universal measure of
+# expansion -- it simultaneously bounds ref-count fan-out, node-count
+# amplification, and scalar-byte amplification (large ``description`` /
+# ``const`` / ``enum`` values). Real-world MCP / OpenAPI-derived tool schemas
+# inline well under 1MB; 10MB sits two orders of magnitude above that, well
+# below memory-pressure territory, and rejects request-supplied bombs before
+# the proxy materialises them.
+_LEGACY_DEFS_MAX_INLINED_BYTES = 10_000_000
+
+
+def unpack_legacy_defs(
+    schema: dict,
+    *,
+    copy: bool = False,
+    max_inlined_bytes: int = _LEGACY_DEFS_MAX_INLINED_BYTES,
+) -> dict:
+    """Inline ``$ref``s backed by draft-04 ``definitions`` / OpenAPI
+    ``components.schemas``. ``$defs`` is left untouched.
+
+    Anthropic and Fireworks tool-schema resolvers only recognise ``$defs``;
+    legacy / OpenAPI def blocks are otherwise silently dropped and leave
+    dangling pointers. See https://github.com/BerriAI/litellm/issues/26692.
+
+    Mutates ``schema`` in place and returns it. Pass ``copy=True`` to deep-copy
+    first (only when there is actually work to do). ``max_inlined_bytes``
+    bounds the cumulative JSON-byte size of inlined targets so request-supplied
+    schemas cannot expand into a schema-bomb before reaching the upstream
+    provider -- raises ``ValueError`` on overflow.
+    """
+    if not _has_legacy_defs(schema):
+        return schema
+    if copy:
+        import copy as _copy
+
+        schema = _copy.deepcopy(schema)
+    # On key collision, ``definitions`` wins over ``components.schemas`` --
+    # ``unpack_defs`` keys refs by last path segment so a single name can only
+    # resolve to one body, and ``definitions`` is the JSON-Schema-native
+    # namespace.
+    defs = schema.pop("components", {}).get("schemas") or {}
+    defs.update(schema.pop("definitions", None) or {})
+    unpack_defs(schema, defs, max_inlined_bytes=max_inlined_bytes)
+    return schema
+
+
 def _get_image_mime_type_from_url(url: str) -> Optional[str]:
     """
     Get mime type for common image URLs

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -1670,15 +1670,15 @@ def convert_to_gemini_tool_call_result(  # noqa: PLR0915
     if gemini_call_id:
         _function_response["id"] = gemini_call_id
 
-    # Create part with function_response, and optionally inline_data for images (Computer Use)
     _part: VertexPartType = {"function_response": _function_response}
 
-    # For Computer Use, if we have images/files, we need separate parts:
-    # - One part with function_response
-    # - One part per inline_data item
-    # Gemini's PartType is a oneof, so we can't have both in the same part
+    # For multimodal function responses, Gemini expects media parts nested
+    # inside functionResponse.parts instead of sibling content parts.
     if inline_data_list:
-        return [_part] + [{"inline_data": d} for d in inline_data_list]
+        _function_response["parts"] = [
+            {"inline_data": inline_data} for inline_data in inline_data_list
+        ]
+        return [_part]
 
     return _part
 
@@ -3658,6 +3658,7 @@ from litellm.types.llms.bedrock import (
     ToolInputSchemaBlock as BedrockToolInputSchemaBlock,
 )
 from litellm.types.llms.bedrock import ToolJsonSchemaBlock as BedrockToolJsonSchemaBlock
+from litellm.types.llms.bedrock import SearchResultBlock
 from litellm.types.llms.bedrock import ToolResultBlock as BedrockToolResultBlock
 from litellm.types.llms.bedrock import (
     ToolResultContentBlock as BedrockToolResultContentBlock,
@@ -4063,6 +4064,122 @@ def _convert_to_bedrock_tool_call_invoke(
         )
 
 
+def _append_bedrock_tool_result_media_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    processed_block: BedrockContentBlock,
+    content: dict,
+    content_type: str,
+) -> None:
+    if "image" in processed_block:
+        tool_result_content_blocks.append(
+            BedrockToolResultContentBlock(image=processed_block["image"])
+        )
+    elif "document" in processed_block:
+        tool_result_content_blocks.append(
+            BedrockToolResultContentBlock(document=processed_block["document"])
+        )
+    else:
+        verbose_logger.warning(
+            "Bedrock Converse: unrecognized BedrockContentBlock keys "
+            "%s for %s tool-result block %s; dropping.",
+            list(processed_block.keys()),
+            content_type,
+            content,
+        )
+
+
+def _append_bedrock_tool_result_image_url_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    content: dict,
+) -> None:
+    format: Optional[str] = None
+    if isinstance(content["image_url"], dict):
+        image_url = content["image_url"]["url"]
+        format = content["image_url"].get("format")
+    else:
+        image_url = content["image_url"]
+    processed_block = BedrockImageProcessor.process_image_sync(
+        image_url=image_url,
+        format=format,
+    )
+    _append_bedrock_tool_result_media_block(
+        tool_result_content_blocks, processed_block, content, "image_url"
+    )
+
+
+def _append_bedrock_tool_result_file_block(
+    tool_result_content_blocks: List[BedrockToolResultContentBlock],
+    content: dict,
+) -> None:
+    # Match the user-message path (_process_file_message): accept either
+    # file_data (base64 data URI) or file_id (server-side reference / URL).
+    file_obj = content.get("file") or {}
+    file_data = file_obj.get("file_data")
+    file_id = file_obj.get("file_id")
+    if file_data is None and file_id is None:
+        raise litellm.BadRequestError(
+            message="file_data and file_id cannot both be None. Got={}".format(content),
+            model="",
+            llm_provider="bedrock",
+        )
+    processed_block = BedrockImageProcessor.process_image_sync(
+        image_url=cast(str, file_id or file_data),
+        format=file_obj.get("format"),
+    )
+    _append_bedrock_tool_result_media_block(
+        tool_result_content_blocks, processed_block, content, "file"
+    )
+
+
+def _parse_bedrock_tool_result_content_list(
+    content_list: List,
+) -> List[BedrockToolResultContentBlock]:
+    tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
+    for content in content_list:
+        if content["type"] == "text":
+            tool_result_content_blocks.append(
+                BedrockToolResultContentBlock(text=content["text"])
+            )
+        elif content["type"] == "image_url":
+            _append_bedrock_tool_result_image_url_block(
+                tool_result_content_blocks, content
+            )
+        elif content["type"] == "file":
+            _append_bedrock_tool_result_file_block(tool_result_content_blocks, content)
+    return tool_result_content_blocks
+
+
+def _build_bedrock_tool_result_content_blocks(
+    message: Union[ChatCompletionToolMessage, ChatCompletionFunctionMessage],
+) -> tuple[List[BedrockToolResultContentBlock], bool]:
+    # Optional OpenAI tool-message extension:
+    # allow structured Bedrock search results on tool messages and map them
+    # directly to toolResult.content[].searchResult for Converse API.
+    #
+    # If `search_results` is present, we intentionally prefer it over `content`
+    # to avoid generating mixed text + searchResult blocks.
+    search_results = message.get("search_results")
+    if isinstance(search_results, list):
+        tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
+        for result in search_results:
+            if not isinstance(result, dict):
+                continue
+            tool_result_content_blocks.append(
+                BedrockToolResultContentBlock(
+                    searchResult=cast(SearchResultBlock, result)
+                )
+            )
+        if tool_result_content_blocks:
+            return tool_result_content_blocks, True
+
+    message_content = message["content"]
+    if isinstance(message_content, str):
+        return [BedrockToolResultContentBlock(text=message_content)], False
+    if isinstance(message_content, List):
+        return _parse_bedrock_tool_result_content_list(message_content), False
+    return [], False
+
+
 def _convert_to_bedrock_tool_call_result(
     message: Union[ChatCompletionToolMessage, ChatCompletionFunctionMessage],
 ) -> BedrockContentBlock:
@@ -4106,90 +4223,18 @@ def _convert_to_bedrock_tool_call_result(
     """
     - 
     """
-    tool_result_content_blocks: List[BedrockToolResultContentBlock] = []
-    if isinstance(message["content"], str):
-        tool_result_content_blocks.append(
-            BedrockToolResultContentBlock(text=message["content"])
-        )
-    elif isinstance(message["content"], List):
-        content_list = message["content"]
-        for content in content_list:
-            if content["type"] == "text":
-                tool_result_content_blocks.append(
-                    BedrockToolResultContentBlock(text=content["text"])
-                )
-            elif content["type"] == "image_url":
-                format: Optional[str] = None
-                if isinstance(content["image_url"], dict):
-                    image_url = content["image_url"]["url"]
-                    format = content["image_url"].get("format")
-                else:
-                    image_url = content["image_url"]
-                _block: BedrockContentBlock = BedrockImageProcessor.process_image_sync(
-                    image_url=image_url,
-                    format=format,
-                )
-                if "image" in _block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(image=_block["image"])
-                    )
-                elif "document" in _block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(document=_block["document"])
-                    )
-                else:
-                    verbose_logger.warning(
-                        "Bedrock Converse: unrecognized BedrockContentBlock keys "
-                        "%s for image_url tool-result block %s; dropping.",
-                        list(_block.keys()),
-                        content,
-                    )
-            elif content["type"] == "file":
-                # Match the user-message path (_process_file_message): accept
-                # either file_data (base64 data URI) or file_id (server-side
-                # reference / URL) and hand off to BedrockImageProcessor. Raise
-                # BadRequestError on both-None rather than silently dropping.
-                file_obj = content.get("file") or {}
-                file_data = file_obj.get("file_data")
-                file_id = file_obj.get("file_id")
-                if file_data is None and file_id is None:
-                    raise litellm.BadRequestError(
-                        message="file_data and file_id cannot both be None. Got={}".format(
-                            content
-                        ),
-                        model="",
-                        llm_provider="bedrock",
-                    )
-                file_format = file_obj.get("format")
-                _file_block: BedrockContentBlock = (
-                    BedrockImageProcessor.process_image_sync(
-                        image_url=cast(str, file_id or file_data),
-                        format=file_format,
-                    )
-                )
-                if "document" in _file_block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(document=_file_block["document"])
-                    )
-                elif "image" in _file_block:
-                    tool_result_content_blocks.append(
-                        BedrockToolResultContentBlock(image=_file_block["image"])
-                    )
-                else:
-                    verbose_logger.warning(
-                        "Bedrock Converse: unrecognized BedrockContentBlock keys "
-                        "%s for file tool-result block %s; dropping.",
-                        list(_file_block.keys()),
-                        content,
-                    )
+    tool_result_content_blocks, used_search_results = (
+        _build_bedrock_tool_result_content_blocks(message)
+    )
 
     message.get("name", "")
     id = str(message.get("tool_call_id", str(uuid.uuid4())))
 
     tool_result = BedrockToolResultBlock(
-        content=tool_result_content_blocks,
-        toolUseId=id,
+        content=tool_result_content_blocks, toolUseId=id
     )
+    if used_search_results:
+        tool_result["status"] = cast(Literal["success"], "success")
 
     content_block = BedrockContentBlock(toolResult=tool_result)
 

litellm/litellm_core_utils/safe_json_dumps.py

@@ -6,10 +6,16 @@ from pydantic import BaseModel
 from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH
 
 
+def strip_null_bytes(value: str) -> str:
+    """Strip NUL bytes, which PostgreSQL text/jsonb columns reject (error 22P05)."""
+    return value.replace("\x00", "")
+
+
 def safe_dumps(data: Any, max_depth: int = DEFAULT_MAX_RECURSE_DEPTH) -> str:
     """
     Recursively serialize data while detecting circular references.
     If a circular reference is detected then a marker string is returned.
+    NUL bytes are stripped from strings to prevent PostgreSQL 22P05 errors.
     """
 
     def _serialize(obj: Any, seen: set, depth: int) -> Any:
@@ -17,7 +23,9 @@ def safe_dumps(data: Any, max_depth: int = DEFAULT_MAX_RECURSE_DEPTH) -> str:
         if depth > max_depth:
             return "MaxDepthExceeded"
         # Base-case: if it is a primitive, simply return it.
-        if isinstance(obj, (str, int, float, bool, type(None))):
+        if isinstance(obj, str):
+            return strip_null_bytes(obj)
+        if isinstance(obj, (int, float, bool, type(None))):
             return obj
         # Check for circular reference.
         if id(obj) in seen:
@@ -28,7 +36,7 @@ def safe_dumps(data: Any, max_depth: int = DEFAULT_MAX_RECURSE_DEPTH) -> str:
             result = {}
             for k, v in obj.items():
                 if isinstance(k, (str)):
-                    result[k] = _serialize(v, seen, depth + 1)
+                    result[strip_null_bytes(k)] = _serialize(v, seen, depth + 1)
             seen.remove(id(obj))
             return result
         elif isinstance(obj, list):
@@ -51,7 +59,7 @@ def safe_dumps(data: Any, max_depth: int = DEFAULT_MAX_RECURSE_DEPTH) -> str:
         else:
             # Fall back to string conversion for non-serializable objects.
             try:
-                return str(obj)
+                return strip_null_bytes(str(obj))
             except Exception:
                 return "Unserializable Object"
 

litellm/litellm_core_utils/secret_redaction.py

@@ -50,13 +50,15 @@ def _build_secret_patterns() -> "re.Pattern[str]":
         r"(?<=://)[^\s'\"]*:[^\s'\"@]+(?=@)",
         # Databricks personal access tokens
         r"dapi[0-9a-f]{32}",
+        # Module-level provider keys logged as litellm.<provider>_key=<value>
+        r"litellm\.[A-Za-z0-9_]*_key['\"]?\s*[:=]\s*['\"]?[^\s,'\"})\]{}>]+",
         # ── Key-name-based redaction ──
         # Catches secrets inside dicts/config dumps by matching on the KEY name
         # regardless of what the value looks like.
         # e.g. 'master_key': 'any-value-here', "database_url": "postgres://..."
         # private_key with PEM-aware value capture
         r"""private_key['\"]?\s*[:=]\s*['\"]?(?:-----BEGIN[A-Z \-]*PRIVATE KEY-----[\s\S]*?-----END[A-Z \-]*PRIVATE KEY-----|[^\s,'\"})\]{}>]+)""",
-        r"(?:master_key|database_url|db_url|connection_string|"
+        r"(?:master_key|xai_key|database_url|db_url|connection_string|"
         r"signing_key|encryption_key|"
         r"auth_token|access_token|refresh_token|"
         r"slack_webhook_url|webhook_url|"

litellm/litellm_core_utils/streaming_handler.py

@@ -1835,8 +1835,10 @@ class CustomStreamWrapper:
                     processed_chunk, None, None, cache_hit
                 )
             )
-        ## SYNC LOGGING
-        self.logging_obj.success_handler(processed_chunk, None, None, cache_hit)
+        ## SYNC LOGGING — only for sync SDK entrypoints; async proxy paths export via async_success_handler
+        litellm_params = self.logging_obj.model_call_details.get("litellm_params", {})
+        if self.logging_obj._is_sync_litellm_request(litellm_params):
+            self.logging_obj.success_handler(processed_chunk, None, None, cache_hit)
 
     def finish_reason_handler(self):
         model_response = self.model_response_creator()
@@ -2231,23 +2233,19 @@ class CustomStreamWrapper:
                         cache_hit,
                     )
                 else:
+                    # prefer_async_handlers routes CustomLogger to async_success_handler
+                    # when consumers use ``async for`` on sync-SDK streams. Legacy string
+                    # callbacks still run via executor.submit inside dispatch_success_handlers.
                     asyncio.create_task(
-                        self.logging_obj.async_success_handler(
+                        self.logging_obj.dispatch_success_handlers(
                             complete_streaming_response,
                             cache_hit=cache_hit,
                             start_time=None,
                             end_time=None,
+                            prefer_async_handlers=True,
                         )
                     )
 
-                    executor.submit(
-                        self.logging_obj.success_handler,
-                        complete_streaming_response,
-                        cache_hit=cache_hit,
-                        start_time=None,
-                        end_time=None,
-                    )
-
                 raise StopAsyncIteration  # Re-raise StopIteration
             else:
                 self.sent_last_chunk = True

litellm/litellm_core_utils/token_counter.py

@@ -486,6 +486,14 @@ def _count_messages(
                     use_default_image_token_count,
                     default_token_count,
                 )
+            elif key == "search_results" and isinstance(value, list):
+                from litellm.litellm_core_utils.prompt_templates.common_utils import (
+                    extract_search_results_text,
+                )
+
+                search_results_text = extract_search_results_text(value)
+                if search_results_text:
+                    num_tokens += params.count_function(search_results_text)
             else:
                 # Skip unsupported keys instead of raising an error
                 continue
@@ -764,11 +772,29 @@ def _format_function_definitions(tools):
     lines.append("namespace functions {")
     lines.append("")
     for tool in tools:
+        if not isinstance(tool, dict):
+            continue
         function = tool.get("function")
+        if not isinstance(function, dict):
+            # Anthropic tool shape → OpenAI function dict for token counting.
+            params = tool.get("input_schema") or tool.get("parameters") or {}
+            if not isinstance(params, dict):
+                params = {}
+            function = {
+                "name": tool.get("name"),
+                "description": tool.get("description"),
+                "parameters": params,
+            }
+        function_name = function.get("name")
+        if not function_name:
+            # Skip malformed tools missing a name to avoid emitting
+            # ``type None = ...`` which would produce inaccurate token counts.
+            continue
         if function_description := function.get("description"):
             lines.append(f"// {function_description}")
-        function_name = function.get("name")
-        parameters = function.get("parameters", {})
+        parameters = function.get("parameters") or {}
+        if not isinstance(parameters, dict):
+            parameters = {}
         properties = parameters.get("properties")
         if properties and properties.keys():
             lines.append(f"type {function_name} = (_: {{")

litellm/llms/anthropic/chat/transformation.py

@@ -29,6 +29,7 @@ from litellm.constants import (
     RESPONSE_FORMAT_TOOL_NAME,
 )
 from litellm.litellm_core_utils.core_helpers import map_finish_reason
+from litellm.litellm_core_utils.prompt_templates.common_utils import unpack_legacy_defs
 from litellm.llms.base_llm.base_utils import type_to_response_format_param
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
 from litellm.types.llms.anthropic import (
@@ -680,6 +681,10 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                 if "properties" not in _input_schema:
                     _input_schema["properties"] = {}
 
+            # Inline legacy / OpenAPI $refs before the allow-list filter strips
+            # their backing def blocks (https://github.com/BerriAI/litellm/issues/26692).
+            _input_schema = unpack_legacy_defs(_input_schema, copy=True)
+
             _allowed_properties = set(AnthropicInputSchema.__annotations__.keys())
             input_schema_filtered = {
                 k: v for k, v in _input_schema.items() if k in _allowed_properties

litellm/llms/anthropic/experimental_pass_through/adapters/handler.py

@@ -4,6 +4,7 @@ from typing import (
     AsyncIterator,
     Coroutine,
     Dict,
+    Iterator,
     List,
     Optional,
     Tuple,
@@ -12,9 +13,16 @@ from typing import (
 )
 
 import litellm
+from litellm._logging import verbose_logger
+from litellm.litellm_core_utils.asyncify import run_async_function
 from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import (
     AnthropicAdapter,
 )
+from litellm.llms.anthropic.experimental_pass_through.context_management import (
+    AnthropicContextManagementError,
+    PolyfillResult,
+    apply_context_management,
+)
 from litellm.llms.anthropic.experimental_pass_through.utils import (
     is_reasoning_auto_summary_enabled,
 )
@@ -28,15 +36,266 @@ if TYPE_CHECKING:
     pass
 
 
-# Anthropic-only fields that the translator above already maps into the
-# OpenAI-format completion_kwargs (output_config → reasoning_effort /
-# response_format, etc.). They must be filtered out of the raw
-# extra_kwargs re-merge below or non-Anthropic backends reject the call
-# with 400 "Extra inputs are not permitted". Add new entries here when
-# extending AnthropicMessagesRequestOptionalParams with another Anthropic-
-# specific key.
+# Anthropic-only keys already mapped by the translator; strip on extra_kwargs re-merge.
 ANTHROPIC_ONLY_REQUEST_KEYS: frozenset[str] = frozenset({"output_config"})
 
+
+def _messages_have_compaction_block(messages: List[Dict]) -> bool:
+    """Return True when any message carries a ``compaction`` content block."""
+    for msg in messages:
+        content = msg.get("content")
+        if not isinstance(content, list):
+            continue
+        for block in content:
+            if isinstance(block, dict) and block.get("type") == "compaction":
+                return True
+    return False
+
+
+def _extract_proxy_litellm_metadata(kwargs: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    """Return ``kwargs["litellm_metadata"]`` when it's a dict; ``None`` otherwise.
+
+    The proxy attaches its auth/spend-attribution fields (``user_api_key``,
+    ``user_api_key_team_id``, ``litellm_call_id``, the full ``UserAPIKeyAuth``
+    object under ``user_api_key_auth``, ...) to ``data["litellm_metadata"]``
+    for ``/v1/messages`` (see
+    ``LiteLLMProxyRequestSetup.add_user_api_key_auth_to_request_metadata`` and
+    ``LITELLM_METADATA_ROUTES``). The Anthropic-shape ``metadata`` arg only
+    carries ``user_id`` and must not be conflated. Returns ``None`` for SDK
+    callers that bypass the proxy entirely.
+    """
+    litellm_metadata = kwargs.get("litellm_metadata")
+    if not isinstance(litellm_metadata, dict):
+        return None
+    return litellm_metadata
+
+
+async def _prepare_context_managed_request(
+    *,
+    model: str,
+    messages: List[Dict],
+    tools: Optional[List[Dict]],
+    system: Optional[Any],
+    context_management_spec: Any,
+    litellm_metadata: Optional[Dict],
+    drop_params: Optional[bool],
+    llm_router: Any,
+    user_api_key_auth: Any = None,
+) -> Optional[PolyfillResult]:
+    """Apply client compaction history, then optional context_management polyfill."""
+    from litellm.llms.anthropic.experimental_pass_through.context_management.editors.compact import (
+        apply_client_compaction_block_history,
+    )
+
+    # Skip the client-history pre-processing when a ``compact_20260112``
+    # polyfill spec will run: that editor already slices around any client-sent
+    # compaction block in its Phase A (and uses the full post-compaction tail
+    # for its token-threshold check). Pre-collapsing to just the latest user
+    # question here would starve the polyfill of conversation context and
+    # silently drop intermediate turns.
+    polyfill_will_run = _polyfill_will_run(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+
+    if polyfill_will_run:
+        history_result: Optional[PolyfillResult] = None
+        working_messages: List[Dict] = messages
+        working_system: Optional[Any] = system
+    else:
+        history_result = apply_client_compaction_block_history(
+            messages=cast(List[Dict[str, Any]], messages),
+            system=system,
+        )
+        working_messages = (
+            history_result.messages if history_result is not None else messages
+        )
+        working_system = history_result.system if history_result is not None else system
+
+    polyfill_result = await _run_polyfill_if_enabled(
+        model=model,
+        messages=working_messages,
+        tools=tools,
+        system=working_system,
+        context_management_spec=context_management_spec,
+        litellm_metadata=litellm_metadata,
+        drop_params=drop_params,
+        llm_router=llm_router,
+        user_api_key_auth=user_api_key_auth,
+    )
+
+    if polyfill_result is not None:
+        return polyfill_result
+
+    # Safety net: if we skipped client-history pre-processing because a
+    # ``compact_20260112`` polyfill was expected to handle the compaction
+    # block itself but the polyfill ultimately did not produce a result
+    # (e.g. it crashed and was best-effort swallowed in
+    # ``_run_polyfill_if_enabled``), apply the slice-only fallback now so
+    # Anthropic-specific ``compaction`` content blocks don't leak through
+    # to non-Anthropic backends that would reject them.
+    if polyfill_will_run and history_result is None:
+        history_result = apply_client_compaction_block_history(
+            messages=cast(List[Dict[str, Any]], messages),
+            system=system,
+        )
+    return history_result
+
+
+def _polyfill_will_run(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> bool:
+    """Return True when ``compact_20260112`` will run via the polyfill dispatcher.
+
+    Mirrors the gating in ``_run_polyfill_if_enabled``: an empty spec or
+    effective ``drop_params`` short-circuits the polyfill. The pre-processing
+    skip only applies when the dispatcher will actually invoke
+    ``apply_compact_20260112`` (which has its own compaction-block slicing).
+    """
+    edits = _normalize_spec_edits(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+    if edits is None:
+        return False
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.constants import (
+        COMPACT_EDIT_TYPE,
+    )
+
+    return any(
+        isinstance(edit, dict) and edit.get("type") == COMPACT_EDIT_TYPE
+        for edit in edits
+    )
+
+
+def _spec_has_non_compact_edits(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> bool:
+    """Return True when the spec includes edits other than ``compact_20260112``.
+
+    Used to decide whether a polyfill failure can be silently swallowed
+    (compact-only specs have a safe compaction-block slicing fallback) or
+    must be surfaced (other editors like ``clear_tool_uses_20250919`` have
+    no slice-only fallback and would otherwise be dropped without notice).
+    """
+    edits = _normalize_spec_edits(
+        context_management_spec=context_management_spec,
+        drop_params=drop_params,
+    )
+    if edits is None:
+        return False
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.constants import (
+        COMPACT_EDIT_TYPE,
+    )
+
+    return any(
+        isinstance(edit, dict)
+        and isinstance(edit.get("type"), str)
+        and edit.get("type") != COMPACT_EDIT_TYPE
+        for edit in edits
+    )
+
+
+def _normalize_spec_edits(
+    *,
+    context_management_spec: Any,
+    drop_params: Optional[bool],
+) -> Optional[List[Dict[str, Any]]]:
+    """Return the normalized ``edits`` list, or ``None`` if the polyfill won't run.
+
+    Delegates spec-shape normalization to the dispatcher's ``_normalize_spec``
+    so the prediction here can't drift from what the dispatcher actually does.
+    """
+    if not context_management_spec:
+        return None
+
+    effective_drop_params = (
+        drop_params if drop_params is not None else litellm.drop_params
+    )
+    if effective_drop_params:
+        return None
+
+    from litellm.llms.anthropic.experimental_pass_through.context_management.dispatcher import (
+        _normalize_spec,
+    )
+
+    try:
+        return _normalize_spec(context_management_spec)
+    except Exception:
+        return None
+
+
+async def _run_polyfill_if_enabled(
+    *,
+    model: str,
+    messages: List[Dict],
+    tools: Optional[List[Dict]],
+    system: Optional[Any],
+    context_management_spec: Any,
+    litellm_metadata: Optional[Dict],
+    drop_params: Optional[bool],
+    llm_router: Any,
+    user_api_key_auth: Any = None,
+) -> Optional[PolyfillResult]:
+    """Run the async context_management polyfill if a spec is present.
+
+    Returns ``None`` when the spec is empty or drop_params is on. Raises
+    ``AnthropicContextManagementError`` so the /v1/messages endpoint can
+    emit an Anthropic-format 400. All other exceptions are best-effort
+    swallowed (matches v0 behavior).
+    """
+    if not context_management_spec:
+        return None
+
+    effective_drop_params = (
+        drop_params if drop_params is not None else litellm.drop_params
+    )
+    if effective_drop_params:
+        return None
+
+    try:
+        return await apply_context_management(
+            model=model,
+            messages=messages,
+            tools=tools,
+            system=system,
+            context_management_spec=context_management_spec,
+            litellm_metadata=litellm_metadata,
+            llm_router=llm_router,
+            user_api_key_auth=user_api_key_auth,
+        )
+    except AnthropicContextManagementError:
+        # Surface validation errors so the endpoint can emit an Anthropic-format
+        # 400. Other exception types fall into the best-effort branch below.
+        raise
+    except Exception as e:
+        verbose_logger.exception(
+            "context_management polyfill: skipping edits due to error: %s", e
+        )
+        # Best-effort swallow is only safe for compact-only specs, where the
+        # caller's compaction-block-slicing safety net produces a correct
+        # (if degraded) result. When the spec also requested non-compact
+        # edits (e.g. ``clear_tool_uses_20250919``), the safety net does
+        # NOT re-run those editors, so silently returning ``None`` would
+        # drop them with no error surface. Raise instead so the endpoint
+        # emits an Anthropic-format error.
+        if _spec_has_non_compact_edits(
+            context_management_spec=context_management_spec,
+            drop_params=drop_params,
+        ):
+            raise AnthropicContextManagementError(
+                status_code=500,
+                message=f"context_management polyfill failed: {e}",
+            ) from e
+        return None
+
+
 ########################################################
 # init adapter
 ANTHROPIC_ADAPTER = AnthropicAdapter()
@@ -163,7 +422,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         metadata: Optional[Dict] = None,
         stop_sequences: Optional[List[str]] = None,
         stream: Optional[bool] = False,
-        system: Optional[str] = None,
+        system: Optional[Union[str, List[Dict[str, Any]]]] = None,
         temperature: Optional[float] = None,
         thinking: Optional[Dict] = None,
         tool_choice: Optional[Dict] = None,
@@ -307,19 +566,56 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         top_p: Optional[float] = None,
         output_format: Optional[Dict] = None,
         **kwargs,
-    ) -> Union[AnthropicMessagesResponse, AsyncIterator]:
+    ) -> Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]]:
         """Handle non-Anthropic models asynchronously using the adapter"""
+        context_management = kwargs.pop("context_management", None)
+        drop_params: Optional[bool] = kwargs.get("drop_params", None)
+        litellm_router = kwargs.pop("litellm_router", None)
+        if litellm_router is None:
+            try:
+                from litellm.proxy.proxy_server import llm_router as _proxy_router
+
+                litellm_router = _proxy_router
+            except Exception:
+                pass
+
+        proxy_litellm_metadata = _extract_proxy_litellm_metadata(kwargs)
+        user_api_key_auth = (
+            proxy_litellm_metadata.get("user_api_key_auth")
+            if proxy_litellm_metadata is not None
+            else None
+        )
+
+        polyfill_result = await _prepare_context_managed_request(
+            model=model,
+            messages=messages,
+            tools=tools,
+            system=system,
+            context_management_spec=context_management,
+            litellm_metadata=proxy_litellm_metadata,
+            drop_params=drop_params,
+            llm_router=litellm_router,
+            user_api_key_auth=user_api_key_auth,
+        )
+
+        effective_messages = (
+            polyfill_result.messages if polyfill_result is not None else messages
+        )
+        effective_system = (
+            polyfill_result.system if polyfill_result is not None else system
+        )
+
         (
             completion_kwargs,
             tool_name_mapping,
         ) = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
             max_tokens=max_tokens,
-            messages=messages,
+            messages=effective_messages,
             model=model,
             metadata=metadata,
             stop_sequences=stop_sequences,
             stream=stream,
-            system=system,
+            system=effective_system,
             temperature=temperature,
             thinking=thinking,
             tool_choice=tool_choice,
@@ -338,6 +634,8 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                     completion_response,
                     model=model,
                     tool_name_mapping=tool_name_mapping,
+                    polyfill_result=polyfill_result,
+                    is_async=True,
                 )
             )
             if transformed_stream is not None:
@@ -347,6 +645,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             anthropic_response = ANTHROPIC_ADAPTER.translate_completion_output_params(
                 cast(ModelResponse, completion_response),
                 tool_name_mapping=tool_name_mapping,
+                polyfill_result=polyfill_result,
             )
             if anthropic_response is not None:
                 return anthropic_response
@@ -372,8 +671,13 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         **kwargs,
     ) -> Union[
         AnthropicMessagesResponse,
+        Iterator[bytes],
         AsyncIterator[Any],
-        Coroutine[Any, Any, Union[AnthropicMessagesResponse, AsyncIterator[Any]]],
+        Coroutine[
+            Any,
+            Any,
+            Union[AnthropicMessagesResponse, AsyncIterator[Any], Iterator[bytes]],
+        ],
     ]:
         """Handle non-Anthropic models using the adapter."""
         if _is_async is True:
@@ -395,17 +699,72 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                 **kwargs,
             )
 
+        # Run the context_management polyfill on the sync path too so that
+        # ``litellm.messages.create()`` callers don't silently lose edits like
+        # ``clear_tool_uses_20250919``. The dispatcher is async (so the
+        # ``compact_20260112`` editor can ``await`` the summarization model);
+        # bridge to it via ``run_async_function``.
+        context_management = kwargs.pop("context_management", None)
+        drop_params: Optional[bool] = kwargs.get("drop_params", None)
+        # Deliberately do NOT auto-attach the proxy ``llm_router`` here:
+        # ``run_async_function`` spawns a new event loop in a worker thread
+        # to bridge to the async dispatcher, but the proxy router's httpx
+        # ``AsyncClient`` instances are bound to the proxy's main event loop.
+        # Reusing them from the new thread's loop violates httpx's single-loop
+        # invariant and can raise ``RuntimeError: Event loop is closed`` or
+        # produce stalled connections. The summary editor falls back to
+        # ``litellm.acompletion`` (which creates a fresh client per call) when
+        # ``llm_router`` is ``None``, which is safe to call from the bridged
+        # loop. The async ``async_anthropic_messages_handler`` path is
+        # unaffected because it ``await``s within the original event loop.
+        litellm_router = kwargs.pop("litellm_router", None)
+
+        # Skip the async bridge entirely when there is nothing for either the
+        # polyfill or the client-history slice-only fallback to do. The vast
+        # majority of sync ``litellm.messages.create()`` requests carry no
+        # ``context_management`` spec and no client-sent ``compaction`` block,
+        # and bridging through a worker-thread event loop just to discover
+        # there is no work is pure overhead.
+        if context_management is None and not _messages_have_compaction_block(messages):
+            polyfill_result: Optional[PolyfillResult] = None
+        else:
+            proxy_litellm_metadata = _extract_proxy_litellm_metadata(kwargs)
+            user_api_key_auth = (
+                proxy_litellm_metadata.get("user_api_key_auth")
+                if proxy_litellm_metadata is not None
+                else None
+            )
+            polyfill_result = run_async_function(
+                _prepare_context_managed_request,
+                model=model,
+                messages=messages,
+                tools=tools,
+                system=system,
+                context_management_spec=context_management,
+                litellm_metadata=proxy_litellm_metadata,
+                drop_params=drop_params,
+                llm_router=litellm_router,
+                user_api_key_auth=user_api_key_auth,
+            )
+
+        effective_messages = (
+            polyfill_result.messages if polyfill_result is not None else messages
+        )
+        effective_system = (
+            polyfill_result.system if polyfill_result is not None else system
+        )
+
         (
             completion_kwargs,
             tool_name_mapping,
         ) = LiteLLMMessagesToCompletionTransformationHandler._prepare_completion_kwargs(
             max_tokens=max_tokens,
-            messages=messages,
+            messages=effective_messages,
             model=model,
             metadata=metadata,
             stop_sequences=stop_sequences,
             stream=stream,
-            system=system,
+            system=effective_system,
             temperature=temperature,
             thinking=thinking,
             tool_choice=tool_choice,
@@ -424,6 +783,8 @@ class LiteLLMMessagesToCompletionTransformationHandler:
                     completion_response,
                     model=model,
                     tool_name_mapping=tool_name_mapping,
+                    polyfill_result=polyfill_result,
+                    is_async=False,
                 )
             )
             if transformed_stream is not None:
@@ -433,6 +794,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             anthropic_response = ANTHROPIC_ADAPTER.translate_completion_output_params(
                 cast(ModelResponse, completion_response),
                 tool_name_mapping=tool_name_mapping,
+                polyfill_result=polyfill_result,
             )
             if anthropic_response is not None:
                 return anthropic_response

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -3,11 +3,26 @@
 import json
 import traceback
 from collections import deque
-from typing import TYPE_CHECKING, Any, AsyncIterator, Dict, Iterator, Literal, Optional
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    AsyncIterator,
+    Dict,
+    Iterator,
+    List,
+    Literal,
+    Optional,
+)
 
-from litellm import verbose_logger
+from litellm._logging import verbose_logger
 from litellm._uuid import uuid
-from litellm.types.llms.anthropic import UsageDelta
+from litellm.types.llms.anthropic import (
+    AppliedEdit,
+    CompactionBlock,
+    ContextManagementResponse,
+    UsageDelta,
+    UsageIteration,
+)
 from litellm.types.utils import AdapterCompletionStreamWrapper
 
 if TYPE_CHECKING:
@@ -37,22 +52,208 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
     holding_stop_reason_chunk: Optional[Any] = None
     queued_usage_chunk: bool = False
     current_content_block_index: int = 0
-    current_content_block_start: ContentBlockContentBlockDict = TextBlock(
-        type="text",
-        text="",
-    )
-    chunk_queue: deque = deque()  # Queue for buffering multiple chunks
 
     def __init__(
         self,
         completion_stream: Any,
         model: str,
         tool_name_mapping: Optional[Dict[str, str]] = None,
+        applied_edits: Optional[List[AppliedEdit]] = None,
+        compaction_block: Optional[CompactionBlock] = None,
+        iterations_usage: Optional[List[UsageIteration]] = None,
     ):
         super().__init__(completion_stream)
         self.model = model
         # Mapping of truncated tool names to original names (for OpenAI's 64-char limit)
         self.tool_name_mapping = tool_name_mapping or {}
+        # Polyfill applied_edits on final message_delta.
+        self.applied_edits: List[AppliedEdit] = list(applied_edits or [])
+        # Synthesized compaction block from compact_20260112 polyfill (streaming).
+        self.compaction_block = compaction_block
+        self.iterations_usage = iterations_usage
+        self.sent_compaction_block: bool = False
+        # Per-phase flags so the compaction block's start/delta/stop events
+        # are emitted (and the public state machine is advanced) in
+        # lock-step with the caller actually consuming each event. Pre-
+        # queuing all three would set ``sent_content_block_finish=True``
+        # before the client received ``content_block_stop``, leaving the
+        # observable state inconsistent during the drain window.
+        self.sent_compaction_block_start: bool = False
+        self.sent_compaction_block_delta: bool = False
+        # Per-instance queue for buffering multiple chunks. Must be initialized
+        # here (not at class level) so concurrent streams don't share the same
+        # deque and corrupt each other's SSE event order.
+        self.chunk_queue: deque = deque()
+        # Per-instance default content block. Must be initialized here (not at
+        # class level) so concurrent streams don't share the same mutable dict
+        # — `_should_start_new_content_block` mutates `tool_block["name"]` in
+        # place, which would otherwise leak across streams.
+        self.current_content_block_start: (
+            "AnthropicStreamWrapper.ContentBlockContentBlockDict"
+        ) = self.TextBlock(
+            type="text",
+            text="",
+        )
+
+    def _merge_usage_into_held_stop_reason_chunk(self, chunk: Any) -> Dict[str, Any]:
+        """Merge usage data from ``chunk`` into the held ``message_delta`` chunk.
+
+        Shared by both the sync ``__next__`` and async ``__anext__`` paths so
+        the subtle hold-and-merge logic (cache tokens, ``context_management``
+        attachment, ``UsageDelta`` shape) lives in exactly one place.
+
+        Caller is responsible for managing ``self.holding_stop_reason_chunk``
+        and ``self.queued_usage_chunk`` state and for queuing the returned
+        merged chunk.
+        """
+        assert self.holding_stop_reason_chunk is not None
+        merged_chunk = self.holding_stop_reason_chunk.copy()
+        if "delta" not in merged_chunk:
+            merged_chunk["delta"] = {}
+
+        uncached_input_tokens = chunk.usage.prompt_tokens or 0
+        if (
+            hasattr(chunk.usage, "prompt_tokens_details")
+            and chunk.usage.prompt_tokens_details
+        ):
+            cached_tokens = (
+                getattr(chunk.usage.prompt_tokens_details, "cached_tokens", 0) or 0
+            )
+            uncached_input_tokens -= cached_tokens
+
+        usage_dict: UsageDelta = {
+            "input_tokens": uncached_input_tokens,
+            "output_tokens": chunk.usage.completion_tokens or 0,
+        }
+        if (
+            hasattr(chunk.usage, "_cache_creation_input_tokens")
+            and chunk.usage._cache_creation_input_tokens > 0
+        ):
+            usage_dict["cache_creation_input_tokens"] = (
+                chunk.usage._cache_creation_input_tokens
+            )
+        if (
+            hasattr(chunk.usage, "_cache_read_input_tokens")
+            and chunk.usage._cache_read_input_tokens > 0
+        ):
+            usage_dict["cache_read_input_tokens"] = chunk.usage._cache_read_input_tokens
+        merged_chunk["usage"] = usage_dict
+        if self.applied_edits and "context_management" not in merged_chunk:
+            merged_chunk["context_management"] = ContextManagementResponse(
+                applied_edits=list(self.applied_edits)
+            )
+        return self._augment_message_delta_usage(merged_chunk)
+
+    def _ensure_context_management_attached(
+        self, message_delta_chunk: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """Attach ``context_management`` to a ``message_delta`` chunk if
+        ``self.applied_edits`` is non-empty and the chunk does not already
+        carry it. Returns the (possibly new) chunk dict.
+
+        Centralizing this guard ensures every ``message_delta`` emission
+        path (merge-with-usage and direct-flush-of-held) consistently
+        surfaces ``applied_edits`` to the client.
+        """
+        if not self.applied_edits or "context_management" in message_delta_chunk:
+            return message_delta_chunk
+        augmented = message_delta_chunk.copy()
+        augmented["context_management"] = ContextManagementResponse(
+            applied_edits=list(self.applied_edits)
+        )
+        return augmented
+
+    def _augment_message_delta_usage(
+        self, message_delta_chunk: Dict[str, Any]
+    ) -> Dict[str, Any]:
+        """Attach polyfill compaction iteration usage to the final message_delta.
+
+        Also defensively re-attaches ``context_management`` so the direct
+        held-chunk flush path stays in sync with the merge path's guarantee
+        when ``self.applied_edits`` is non-empty.
+        """
+        message_delta_chunk = self._ensure_context_management_attached(
+            message_delta_chunk
+        )
+        if self.iterations_usage is None:
+            return message_delta_chunk
+        usage = message_delta_chunk.get("usage")
+        if not isinstance(usage, dict) or "iterations" in usage:
+            return message_delta_chunk
+
+        input_tokens = usage.get("input_tokens", 0) or 0
+        output_tokens = usage.get("output_tokens", 0) or 0
+        augmented = message_delta_chunk.copy()
+        augmented_usage = dict(usage)
+        iterations: List[UsageIteration] = list(self.iterations_usage)
+        # Only emit a ``message`` iteration when we have real token data.
+        # Without a separate usage chunk (e.g. provider sent finish_reason
+        # alone), the held ``message_delta`` carries placeholder zeros from
+        # the translate step; reporting a zero-token iteration would be
+        # misleading and inconsistent with the non-streaming path.
+        if input_tokens > 0 or output_tokens > 0:
+            message_iteration: UsageIteration = {
+                "type": "message",
+                "input_tokens": input_tokens,
+                "output_tokens": output_tokens,
+            }
+            iterations.append(message_iteration)
+        augmented_usage["iterations"] = iterations  # type: ignore[typeddict-unknown-key]
+        augmented["usage"] = augmented_usage
+        return augmented
+
+    def _next_compaction_event(self) -> Optional[Dict[str, Any]]:
+        """Return the next compaction content-block SSE event, or ``None``.
+
+        Anthropic delivers compaction as a single delta (no token-by-token
+        streaming), but we still surface it as a proper
+        start → delta → stop trio. Each call returns exactly one event so
+        the state machine (``sent_content_block_finish``,
+        ``current_content_block_index``) is advanced *only* when the
+        terminal stop event is actually handed back to the caller. This
+        prevents an observable window where the flags claim the block is
+        finished while the stop event is still buffered.
+        """
+        if self.compaction_block is None or self.sent_compaction_block:
+            return None
+
+        compaction_index = self.current_content_block_index
+
+        if not self.sent_compaction_block_start:
+            self.sent_compaction_block_start = True
+            return {
+                "type": "content_block_start",
+                "index": compaction_index,
+                # Mirror the text-block shape ({"type": "text", "text": ""}):
+                # send an empty ``content`` field so clients that introspect
+                # ``content_block_start`` see the full block schema. The
+                # actual summary text arrives via the ``content_block_delta``
+                # below.
+                "content_block": {"type": "compaction", "content": ""},
+            }
+
+        if not self.sent_compaction_block_delta:
+            self.sent_compaction_block_delta = True
+            summary_content = self.compaction_block.get("content") or ""
+            return {
+                "type": "content_block_delta",
+                "index": compaction_index,
+                "delta": {"type": "compaction_delta", "content": summary_content},
+            }
+
+        stop_event = {
+            "type": "content_block_stop",
+            "index": compaction_index,
+        }
+        # Don't touch ``sent_content_block_finish`` here: that flag is the
+        # state machine for the regular text/tool_use/thinking block and is
+        # independent of the synthetic compaction block lifecycle. Conflating
+        # them would let outside observers (subclass overrides, introspection
+        # hooks, exception paths) see ``sent_content_block_finish=True``
+        # without any regular content block ever having started.
+        self._increment_content_block_index()
+        self.sent_compaction_block = True
+        return stop_event
 
     def _create_initial_usage_delta(self) -> UsageDelta:
         """
@@ -75,7 +276,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             cache_read_input_tokens=0,
         )
 
-    def __next__(self):
+    def __next__(self):  # noqa: PLR0915
         from .transformation import LiteLLMAnthropicMessagesAdapter
 
         try:
@@ -103,8 +304,17 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 )
                 return self.chunk_queue.popleft()
 
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
             if self.sent_content_block_start is False:
                 self.sent_content_block_start = True
+                self.sent_content_block_finish = False
                 self.chunk_queue.append(
                     {
                         "type": "content_block_start",
@@ -122,11 +332,45 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 if should_start_new_block:
                     self._increment_content_block_index()
 
+                # applied_edits only needs to flow to the final message_delta
+                # (when finish_reason is set); skip threading it through every
+                # intermediate chunk. For the hold-and-merge path below,
+                # context_management is attached directly to the merged chunk,
+                # so the translated ``processed_chunk`` would be discarded —
+                # skip the applied_edits attachment in that case to avoid
+                # allocating a throwaway ``MessageBlockDelta``.
+                will_merge_into_held = (
+                    self.holding_stop_reason_chunk is not None
+                    and getattr(chunk, "usage", None) is not None
+                )
+                is_final_chunk = chunk.choices[0].finish_reason is not None
                 processed_chunk = LiteLLMAnthropicMessagesAdapter().translate_streaming_openai_response_to_anthropic(
                     response=chunk,
                     current_content_block_index=self.current_content_block_index,
+                    applied_edits=(
+                        self.applied_edits
+                        if is_final_chunk and not will_merge_into_held
+                        else None
+                    ),
                 )
 
+                # Check if this is a usage chunk and we have a held stop_reason chunk
+                if will_merge_into_held:
+                    merged_chunk = self._merge_usage_into_held_stop_reason_chunk(chunk)
+                    self.chunk_queue.append(merged_chunk)
+                    self.queued_usage_chunk = True
+                    self.holding_stop_reason_chunk = None
+                    return self.chunk_queue.popleft()
+
+                if self.queued_usage_chunk:
+                    # Usage has already been merged + emitted. Any trailing
+                    # provider events would violate Anthropic SSE ordering
+                    # (no chunks may follow the final ``message_delta``), so
+                    # silently drop them — matches the async ``__anext__``
+                    # behavior where the block-handling logic is gated on
+                    # ``not self.queued_usage_chunk``.
+                    continue
+
                 if should_start_new_block and not self.sent_content_block_finish:
                     # Queue the sequence: content_block_stop -> content_block_start
                     # For text blocks the trigger chunk is not emitted as a separate
@@ -178,20 +422,64 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         }
                     )
                     self.sent_content_block_finish = True
-                    self.chunk_queue.append(processed_chunk)
+                    if processed_chunk.get("delta", {}).get("stop_reason") is not None:
+                        self.holding_stop_reason_chunk = processed_chunk
+                    else:
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
+                        self.chunk_queue.append(processed_chunk)
                     return self.chunk_queue.popleft()
                 elif self.holding_chunk is not None:
                     self.chunk_queue.append(self.holding_chunk)
+                    if processed_chunk.get("type") == "message_delta":
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
                     self.chunk_queue.append(processed_chunk)
                     self.holding_chunk = None
                     return self.chunk_queue.popleft()
                 else:
+                    if processed_chunk.get("type") == "message_delta":
+                        processed_chunk = self._augment_message_delta_usage(
+                            processed_chunk
+                        )
                     self.chunk_queue.append(processed_chunk)
                     return self.chunk_queue.popleft()
 
-            # Handle any remaining held chunks after stream ends
-            if self.holding_chunk is not None:
-                self.chunk_queue.append(self.holding_chunk)
+            # Handle any remaining held chunks after stream ends. The
+            # buffered ``holding_chunk`` (a ``content_block_delta``) must
+            # precede the final ``message_delta`` so Anthropic SSE event
+            # ordering is preserved. When ``queued_usage_chunk`` is True,
+            # the final ``message_delta`` has already been emitted; any
+            # buffered content delta is dropped rather than emitted after
+            # ``message_delta`` (which would violate SSE ordering and may
+            # confuse strict Anthropic SDK clients).
+            if not self.queued_usage_chunk:
+                if self.holding_chunk is not None:
+                    self.chunk_queue.append(self.holding_chunk)
+                    self.holding_chunk = None
+                if self.holding_stop_reason_chunk is not None:
+                    # A final ``message_delta`` must be preceded by
+                    # ``content_block_stop`` so the emitted SSE stays in
+                    # valid Anthropic order (... -> content_block_stop ->
+                    # message_delta). Emit ``content_block_stop`` here if
+                    # the active content block was not already closed.
+                    if not self.sent_content_block_finish:
+                        self.chunk_queue.append(
+                            {
+                                "type": "content_block_stop",
+                                "index": self.current_content_block_index,
+                            }
+                        )
+                        self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+            else:
                 self.holding_chunk = None
 
             if not self.sent_last_message:
@@ -205,6 +493,26 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         except StopIteration:
             if self.chunk_queue:
                 return self.chunk_queue.popleft()
+            # Handle any held stop_reason chunk. Emit ``content_block_stop``
+            # first if the active content block was not already closed, so
+            # Anthropic SSE ordering is preserved (content_block_stop ->
+            # message_delta).
+            if self.holding_stop_reason_chunk is not None:
+                if not self.sent_content_block_finish:
+                    self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+                    return {
+                        "type": "content_block_stop",
+                        "index": self.current_content_block_index,
+                    }
+                held = self._augment_message_delta_usage(self.holding_stop_reason_chunk)
+                self.holding_stop_reason_chunk = None
+                return held
             if self.sent_last_message is False:
                 self.sent_last_message = True
                 return {"type": "message_stop"}
@@ -213,7 +521,7 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             verbose_logger.error(
                 "Anthropic Adapter - {}\n{}".format(e, traceback.format_exc())
             )
-            raise StopAsyncIteration
+            raise StopIteration
 
     async def __anext__(self):  # noqa: PLR0915
         from .transformation import LiteLLMAnthropicMessagesAdapter
@@ -243,8 +551,17 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 )
                 return self.chunk_queue.popleft()
 
+            if (
+                self.sent_compaction_block is False
+                and self.compaction_block is not None
+            ):
+                compaction_event = self._next_compaction_event()
+                if compaction_event is not None:
+                    return compaction_event
+
             if self.sent_content_block_start is False:
                 self.sent_content_block_start = True
+                self.sent_content_block_finish = False
                 self.chunk_queue.append(
                     {
                         "type": "content_block_start",
@@ -263,57 +580,31 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                 if should_start_new_block:
                     self._increment_content_block_index()
 
+                # applied_edits only needs to flow to the final message_delta
+                # (when finish_reason is set); skip threading it through every
+                # intermediate chunk. For the hold-and-merge path below,
+                # context_management is attached directly to the merged chunk,
+                # so the translated ``processed_chunk`` would be discarded —
+                # skip the applied_edits attachment in that case to avoid
+                # allocating a throwaway ``MessageBlockDelta``.
+                will_merge_into_held = (
+                    self.holding_stop_reason_chunk is not None
+                    and getattr(chunk, "usage", None) is not None
+                )
+                is_final_chunk = chunk.choices[0].finish_reason is not None
                 processed_chunk = LiteLLMAnthropicMessagesAdapter().translate_streaming_openai_response_to_anthropic(
                     response=chunk,
                     current_content_block_index=self.current_content_block_index,
+                    applied_edits=(
+                        self.applied_edits
+                        if is_final_chunk and not will_merge_into_held
+                        else None
+                    ),
                 )
 
                 # Check if this is a usage chunk and we have a held stop_reason chunk
-                if (
-                    self.holding_stop_reason_chunk is not None
-                    and getattr(chunk, "usage", None) is not None
-                ):
-                    # Merge usage into the held stop_reason chunk
-                    merged_chunk = self.holding_stop_reason_chunk.copy()
-                    if "delta" not in merged_chunk:
-                        merged_chunk["delta"] = {}
-
-                    # Add usage to the held chunk
-                    uncached_input_tokens = chunk.usage.prompt_tokens or 0
-                    if (
-                        hasattr(chunk.usage, "prompt_tokens_details")
-                        and chunk.usage.prompt_tokens_details
-                    ):
-                        cached_tokens = (
-                            getattr(
-                                chunk.usage.prompt_tokens_details, "cached_tokens", 0
-                            )
-                            or 0
-                        )
-                        uncached_input_tokens -= cached_tokens
-
-                    usage_dict: UsageDelta = {
-                        "input_tokens": uncached_input_tokens,
-                        "output_tokens": chunk.usage.completion_tokens or 0,
-                    }
-                    # Add cache tokens if available (for prompt caching support)
-                    if (
-                        hasattr(chunk.usage, "_cache_creation_input_tokens")
-                        and chunk.usage._cache_creation_input_tokens > 0
-                    ):
-                        usage_dict["cache_creation_input_tokens"] = (
-                            chunk.usage._cache_creation_input_tokens
-                        )
-                    if (
-                        hasattr(chunk.usage, "_cache_read_input_tokens")
-                        and chunk.usage._cache_read_input_tokens > 0
-                    ):
-                        usage_dict["cache_read_input_tokens"] = (
-                            chunk.usage._cache_read_input_tokens
-                        )
-                    merged_chunk["usage"] = usage_dict
-
-                    # Queue the merged chunk and reset
+                if will_merge_into_held:
+                    merged_chunk = self._merge_usage_into_held_stop_reason_chunk(chunk)
                     self.chunk_queue.append(merged_chunk)
                     self.queued_usage_chunk = True
                     self.holding_stop_reason_chunk = None
@@ -379,28 +670,63 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                         ):
                             self.holding_stop_reason_chunk = processed_chunk
                         else:
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                             self.chunk_queue.append(processed_chunk)
                         return self.chunk_queue.popleft()
                     elif self.holding_chunk is not None:
                         # Queue both chunks
                         self.chunk_queue.append(self.holding_chunk)
+                        if processed_chunk.get("type") == "message_delta":
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                         self.chunk_queue.append(processed_chunk)
                         self.holding_chunk = None
                         return self.chunk_queue.popleft()
                     else:
-                        # Queue the current chunk
+                        if processed_chunk.get("type") == "message_delta":
+                            processed_chunk = self._augment_message_delta_usage(
+                                processed_chunk
+                            )
                         self.chunk_queue.append(processed_chunk)
                         return self.chunk_queue.popleft()
 
-            # Handle any remaining held chunks after stream ends
+            # Handle any remaining held chunks after stream ends. The
+            # buffered ``holding_chunk`` (a ``content_block_delta``) must
+            # precede the final ``message_delta`` so Anthropic SSE event
+            # ordering is preserved. When ``queued_usage_chunk`` is True,
+            # the final ``message_delta`` has already been emitted; any
+            # buffered content delta is dropped rather than emitted after
+            # ``message_delta`` (which would violate SSE ordering and may
+            # confuse strict Anthropic SDK clients).
             if not self.queued_usage_chunk:
-                if self.holding_stop_reason_chunk is not None:
-                    self.chunk_queue.append(self.holding_stop_reason_chunk)
-                    self.holding_stop_reason_chunk = None
-
                 if self.holding_chunk is not None:
                     self.chunk_queue.append(self.holding_chunk)
                     self.holding_chunk = None
+                if self.holding_stop_reason_chunk is not None:
+                    # A final ``message_delta`` must be preceded by
+                    # ``content_block_stop`` so the emitted SSE stays in
+                    # valid Anthropic order (... -> content_block_stop ->
+                    # message_delta). Emit ``content_block_stop`` here if
+                    # the active content block was not already closed.
+                    if not self.sent_content_block_finish:
+                        self.chunk_queue.append(
+                            {
+                                "type": "content_block_stop",
+                                "index": self.current_content_block_index,
+                            }
+                        )
+                        self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+            else:
+                self.holding_chunk = None
 
             if not self.sent_last_message:
                 self.sent_last_message = True
@@ -416,9 +742,28 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             # Handle any remaining queued chunks before stopping
             if self.chunk_queue:
                 return self.chunk_queue.popleft()
-            # Handle any held stop_reason chunk
+            # Handle any held stop_reason chunk — clear after capturing so a
+            # subsequent ``__anext__`` call doesn't re-emit the same chunk
+            # (matches the sync ``__next__`` path). Emit ``content_block_stop``
+            # first if the active content block was not already closed, so
+            # Anthropic SSE ordering is preserved (content_block_stop ->
+            # message_delta).
             if self.holding_stop_reason_chunk is not None:
-                return self.holding_stop_reason_chunk
+                if not self.sent_content_block_finish:
+                    self.sent_content_block_finish = True
+                    self.chunk_queue.append(
+                        self._augment_message_delta_usage(
+                            self.holding_stop_reason_chunk
+                        )
+                    )
+                    self.holding_stop_reason_chunk = None
+                    return {
+                        "type": "content_block_stop",
+                        "index": self.current_content_block_index,
+                    }
+                held = self._augment_message_delta_usage(self.holding_stop_reason_chunk)
+                self.holding_stop_reason_chunk = None
+                return held
             if not self.sent_last_message:
                 self.sent_last_message = True
                 return {"type": "message_stop"}

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -6,6 +6,7 @@ from typing import (
     Any,
     AsyncIterator,
     Dict,
+    Iterator,
     List,
     Literal,
     Optional,
@@ -75,6 +76,9 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
 from litellm.litellm_core_utils.prompt_templates.factory import (
     THOUGHT_SIGNATURE_SEPARATOR,
 )
+from litellm.llms.anthropic.experimental_pass_through.context_management import (
+    PolyfillResult,
+)
 from litellm.types.llms.anthropic import (
     ANTHROPIC_HOSTED_TOOLS,
     AllAnthropicToolsValues,
@@ -87,14 +91,17 @@ from litellm.types.llms.anthropic import (
     AnthropicResponseContentBlockText,
     AnthropicResponseContentBlockThinking,
     AnthropicResponseContentBlockToolUse,
+    AppliedEdit,
     ContentBlockDelta,
     ContentJsonBlockDelta,
     ContentTextBlockDelta,
     ContentThinkingBlockDelta,
     ContentThinkingSignatureBlockDelta,
+    ContextManagementResponse,
     MessageBlockDelta,
     MessageDelta,
     UsageDelta,
+    UsageIteration,
 )
 from litellm.types.llms.anthropic_messages.anthropic_response import (
     AnthropicMessagesResponse,
@@ -195,6 +202,7 @@ class AnthropicAdapter:
         self,
         response: ModelResponse,
         tool_name_mapping: Optional[Dict[str, str]] = None,
+        polyfill_result: Optional[PolyfillResult] = None,
     ) -> Optional[AnthropicMessagesResponse]:
         """
         Translate OpenAI response to Anthropic format.
@@ -204,10 +212,12 @@ class AnthropicAdapter:
             tool_name_mapping: Optional mapping of truncated tool names to original names.
                               Used to restore original names for tools that exceeded
                               OpenAI's 64-char limit.
+            polyfill_result: PolyfillResult from context_management polyfill.
         """
         return LiteLLMAnthropicMessagesAdapter().translate_openai_response_to_anthropic(
             response=response,
             tool_name_mapping=tool_name_mapping,
+            polyfill_result=polyfill_result,
         )
 
     def translate_completion_output_params_streaming(
@@ -215,7 +225,9 @@ class AnthropicAdapter:
         completion_stream: Any,
         model: str,
         tool_name_mapping: Optional[Dict[str, str]] = None,
-    ) -> Union[AsyncIterator[bytes], None]:
+        polyfill_result: Optional[PolyfillResult] = None,
+        is_async: bool = True,
+    ) -> Union[AsyncIterator[bytes], Iterator[bytes], None]:
         """
         Translate OpenAI streaming response to Anthropic format.
 
@@ -223,14 +235,35 @@ class AnthropicAdapter:
             completion_stream: The OpenAI streaming response
             model: The model name
             tool_name_mapping: Optional mapping of truncated tool names to original names.
+            polyfill_result: PolyfillResult from context_management polyfill.
+            is_async: When ``True`` (default, for back-compat with existing
+                async callers) returns an ``AsyncIterator[bytes]``. When
+                ``False`` returns a sync ``Iterator[bytes]`` so sync callers
+                (e.g. ``litellm.anthropic.messages.create(stream=True)`` via
+                the sync handler) don't get back an async iterator they
+                can't iterate without an event loop.
         """
+        applied_edits = (
+            polyfill_result.applied_edits_for_response() if polyfill_result else None
+        )
+        compaction_block = (
+            polyfill_result.compaction_block if polyfill_result is not None else None
+        )
+        iterations_usage = (
+            polyfill_result.iterations_usage if polyfill_result is not None else None
+        )
         anthropic_wrapper = AnthropicStreamWrapper(
             completion_stream=completion_stream,
             model=model,
             tool_name_mapping=tool_name_mapping,
+            applied_edits=applied_edits,
+            compaction_block=compaction_block,
+            iterations_usage=iterations_usage,
         )
-        # Return the SSE-wrapped version for proper event formatting
-        return anthropic_wrapper.async_anthropic_sse_wrapper()
+        # Return the SSE-wrapped version for proper event formatting.
+        if is_async:
+            return anthropic_wrapper.async_anthropic_sse_wrapper()
+        return anthropic_wrapper.anthropic_sse_wrapper()
 
 
 class LiteLLMAnthropicMessagesAdapter:
@@ -1342,6 +1375,7 @@ class LiteLLMAnthropicMessagesAdapter:
         self,
         response: ModelResponse,
         tool_name_mapping: Optional[Dict[str, str]] = None,
+        polyfill_result: Optional[PolyfillResult] = None,
     ) -> AnthropicMessagesResponse:
         """
         Translate OpenAI response to Anthropic format.
@@ -1351,12 +1385,17 @@ class LiteLLMAnthropicMessagesAdapter:
             tool_name_mapping: Optional mapping of truncated tool names to original names.
                               Used to restore original names for tools that exceeded
                               OpenAI's 64-char limit.
+            polyfill_result: PolyfillResult from context_management polyfill.
         """
         ## translate content block
         anthropic_content = self._translate_openai_content_to_anthropic(
             choices=response.choices,  # type: ignore
             tool_name_mapping=tool_name_mapping,
         )
+
+        if polyfill_result is not None and polyfill_result.compaction_block is not None:
+            anthropic_content.insert(0, polyfill_result.compaction_block)  # type: ignore[arg-type]
+
         ## extract finish reason
         anthropic_finish_reason = self._translate_openai_finish_reason_to_anthropic(
             openai_finish_reason=response.choices[0].finish_reason  # type: ignore
@@ -1385,6 +1424,14 @@ class LiteLLMAnthropicMessagesAdapter:
         if cached_tokens > 0:
             anthropic_usage["cache_read_input_tokens"] = cached_tokens
 
+        if polyfill_result is not None and polyfill_result.iterations_usage is not None:
+            message_iteration: UsageIteration = {
+                "type": "message",
+                "input_tokens": uncached_input_tokens,
+                "output_tokens": usage.completion_tokens or 0,
+            }
+            anthropic_usage["iterations"] = list(polyfill_result.iterations_usage) + [message_iteration]  # type: ignore[typeddict-unknown-key]
+
         translated_obj = AnthropicMessagesResponse(
             id=response.id,
             type="message",
@@ -1396,6 +1443,14 @@ class LiteLLMAnthropicMessagesAdapter:
             stop_reason=anthropic_finish_reason,
         )
 
+        applied_edits = (
+            polyfill_result.applied_edits_for_response() if polyfill_result else None
+        )
+        if applied_edits:
+            translated_obj["context_management"] = ContextManagementResponse(
+                applied_edits=list(applied_edits)
+            )
+
         return translated_obj
 
     def _translate_streaming_openai_chunk_to_anthropic_content_block(
@@ -1455,6 +1510,17 @@ class LiteLLMAnthropicMessagesAdapter:
                         return "thinking", ChatCompletionThinkingBlock(
                             type="thinking", thinking=thinking, signature=signature
                         )
+            # OpenAI-compatible reasoning backends (e.g. vLLM/SGLang reasoning
+            # parsers) populate ``reasoning_content`` without ``thinking_blocks``.
+            # ``Delta`` deletes the ``thinking_blocks`` attribute when unset, so the
+            # branch above is skipped entirely; open a ``thinking`` block here so the
+            # matching ``thinking_delta`` stream is not emitted into a text block.
+            elif isinstance(choice, StreamingChoices) and getattr(
+                choice.delta, "reasoning_content", None
+            ):
+                return "thinking", ChatCompletionThinkingBlock(
+                    type="thinking", thinking="", signature=""
+                )
 
         return "text", TextBlock(type="text", text="")
 
@@ -1528,7 +1594,10 @@ class LiteLLMAnthropicMessagesAdapter:
             return "text_delta", ContentTextBlockDelta(type="text_delta", text=text)
 
     def translate_streaming_openai_response_to_anthropic(
-        self, response: ModelResponse, current_content_block_index: int
+        self,
+        response: ModelResponse,
+        current_content_block_index: int,
+        applied_edits: Optional[List[AppliedEdit]] = None,
     ) -> Union[ContentBlockDelta, MessageBlockDelta]:
         ## base case - final chunk w/ finish reason
         if response.choices[0].finish_reason is not None:
@@ -1578,9 +1647,14 @@ class LiteLLMAnthropicMessagesAdapter:
                     usage_delta["cache_read_input_tokens"] = cached_tokens
             else:
                 usage_delta = UsageDelta(input_tokens=0, output_tokens=0)
-            return MessageBlockDelta(
+            message_block = MessageBlockDelta(
                 type="message_delta", delta=delta, usage=usage_delta  # type: ignore
             )
+            if applied_edits:
+                message_block["context_management"] = ContextManagementResponse(
+                    applied_edits=list(applied_edits)
+                )
+            return message_block
         (
             type_of_content,
             content_block_delta,

litellm/llms/anthropic/experimental_pass_through/context_management/__init__.py

@@ -0,0 +1,11 @@
+from .constants import CLEARED_TOOL_RESULT_PLACEHOLDER
+from .dispatcher import apply_context_management
+from .errors import AnthropicContextManagementError
+from .result import PolyfillResult
+
+__all__ = [
+    "apply_context_management",
+    "AnthropicContextManagementError",
+    "CLEARED_TOOL_RESULT_PLACEHOLDER",
+    "PolyfillResult",
+]