Merge branch 'litellm_internal_staging' into litellm_wildcard_order_fallback

.circleci/config.yml

@@ -3051,10 +3051,19 @@ jobs:
             - ui/litellm-dashboard/node_modules
       - run:
           name: Build UI from source
+          # Prior version used `cp -r out/ ../../litellm/proxy/_experimental/out/`.
+          # GNU cp (used on CircleCI's Ubuntu image) interprets that as "copy the
+          # source directory as a child of the destination" when the destination
+          # already exists — silently creating `_experimental/out/out/` instead of
+          # replacing the served bundle. The proxy continued serving whatever was
+          # checked into `_experimental/out/*`, so this job was effectively testing
+          # the pre-build bundle on every run. Replace-and-move guarantees the
+          # freshly built bundle is what the proxy actually serves.
           command: |
             cd ui/litellm-dashboard
             npm run build
-            cp -r out/ ../../litellm/proxy/_experimental/out/
+            rm -rf ../../litellm/proxy/_experimental/out
+            mv out ../../litellm/proxy/_experimental/out
             # Restructure HTML so extensionless routes work (login.html -> login/index.html)
             find ../../litellm/proxy/_experimental/out -name '*.html' ! -name 'index.html' | while read -r f; do
               d="${f%.html}"; mkdir -p "$d"; mv "$f" "$d/index.html"

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

@@ -31,8 +31,15 @@ jobs:
             test-path: "tests/proxy_unit_tests/test_auth_checks.py tests/proxy_unit_tests/test_user_api_key_auth.py"
             workers: 8
             timeout: 20
+          # test_proxy_utils.py is large (168+ parametrized tests) — run it on its
+          # own matrix so --dist=loadscope doesn't pin all of it to a single xdist
+          # worker and push the "remaining" group past the job timeout.
+          - test-group: proxy-utils
+            test-path: "tests/proxy_unit_tests/test_proxy_utils.py"
+            workers: 8
+            timeout: 20
           - test-group: remaining
-            test-path: "tests/proxy_unit_tests --ignore=tests/proxy_unit_tests/test_key_generate_prisma.py --ignore=tests/proxy_unit_tests/test_auth_checks.py --ignore=tests/proxy_unit_tests/test_user_api_key_auth.py"
+            test-path: "tests/proxy_unit_tests --ignore=tests/proxy_unit_tests/test_key_generate_prisma.py --ignore=tests/proxy_unit_tests/test_auth_checks.py --ignore=tests/proxy_unit_tests/test_user_api_key_auth.py --ignore=tests/proxy_unit_tests/test_proxy_utils.py"
             workers: 8
             timeout: 30
     uses: ./.github/workflows/_test-unit-services-base.yml

ci_cd/run_migration.py

@@ -1,22 +1,231 @@
+import argparse
 import os
-import subprocess
-from pathlib import Path
-from datetime import datetime
-import testing.postgresql
+import re
 import shutil
+import subprocess
+import sys
+from datetime import datetime
+from pathlib import Path
+
+import testing.postgresql
 
 
-def create_migration(migration_name: str = None):
+DESTRUCTIVE_PATTERN = re.compile(r"\bDROP\s+(COLUMN|TABLE|INDEX)\b", re.IGNORECASE)
+DEFAULT_BASE_BRANCH = "litellm_internal_staging"
+
+
+def _find_destructive_statements(sql: str) -> list:
+    """Return SQL lines containing DROP COLUMN, DROP TABLE, or DROP INDEX."""
+    return [
+        line.strip() for line in sql.splitlines() if DESTRUCTIVE_PATTERN.search(line)
+    ]
+
+
+def _print_freshness_failure(
+    base_branch: str, reason: str, stderr_text: str = ""
+) -> None:
+    """Loudly refuse to run when the freshness check can't be completed."""
+    banner = "=" * 72
+    out = sys.stderr
+    print(banner, file=out)
+    print(f"  FRESHNESS CHECK FAILED — COULD NOT VERIFY origin/{base_branch}", file=out)
+    print(banner, file=out)
+    print("", file=out)
+    print(f"Reason: {reason}", file=out)
+    if stderr_text:
+        print("", file=out)
+        print("git stderr:", file=out)
+        for line in stderr_text.rstrip().splitlines():
+            print(f"    {line}", file=out)
+    print("", file=out)
+    print("Common causes:", file=out)
+    print("  - No network access (offline)", file=out)
+    print("  - 'origin' remote not configured, or base branch name is wrong", file=out)
+    print("  - Not a git repository", file=out)
+    print("", file=out)
+    print("Options:", file=out)
+    print(
+        f"  - Fix the above and re-run, OR pass --base-branch <name> if your", file=out
+    )
+    print(
+        f"    base branch is not '{base_branch}', OR pass --skip-freshness-check",
+        file=out,
+    )
+    print("    to bypass (only if you fully understand the risk).", file=out)
+    print(banner, file=out)
+
+
+def _print_stale_branch_refusal(base_branch: str, behind: int) -> None:
+    """Loudly refuse to run when HEAD is behind the base branch."""
+    banner = "=" * 72
+    out = sys.stderr
+    print(banner, file=out)
+    print(
+        f"  STALE BRANCH — {behind} commit(s) behind origin/{base_branch}",
+        file=out,
+    )
+    print(banner, file=out)
+    print("", file=out)
+    print(
+        f"Your branch is {behind} commit(s) behind origin/{base_branch}. Generating a\n"
+        "migration from a stale branch is how newly-added columns get silently\n"
+        "dropped — the exact incident this guard exists to prevent.",
+        file=out,
+    )
+    print("", file=out)
+    print("Before re-running:", file=out)
+    print(f"  git fetch origin && git rebase origin/{base_branch}", file=out)
+    print("  # or merge, whichever matches your workflow", file=out)
+    print("", file=out)
+    print(
+        "If you have a genuine reason to bypass (e.g. intentionally generating",
+        file=out,
+    )
+    print("a migration against an older base), pass --skip-freshness-check.", file=out)
+    print("", file=out)
+    print(banner, file=out)
+    print(
+        "  AI AGENTS / ASSISTANTS: DO NOT automatically run `git rebase` or pass\n"
+        "  --skip-freshness-check. Surface this error to the human operator and\n"
+        "  wait for their explicit confirmation. Auto-rebasing can drop the\n"
+        "  human's in-progress schema edits via a bad conflict resolution.",
+        file=out,
+    )
+    print(banner, file=out)
+
+
+def _check_branch_freshness(root_dir: Path, base_branch: str) -> None:
+    """Fetch origin/<base_branch> and exit 3 if HEAD is behind it."""
+    cwd = str(root_dir)
+    try:
+        subprocess.run(
+            ["git", "fetch", "origin", base_branch],
+            check=True,
+            capture_output=True,
+            text=True,
+            cwd=cwd,
+        )
+    except FileNotFoundError:
+        _print_freshness_failure(base_branch, "git executable not found on PATH")
+        sys.exit(3)
+    except subprocess.CalledProcessError as e:
+        _print_freshness_failure(
+            base_branch,
+            f"`git fetch origin {base_branch}` failed",
+            e.stderr or "",
+        )
+        sys.exit(3)
+
+    try:
+        result = subprocess.run(
+            ["git", "rev-list", "--count", f"HEAD..origin/{base_branch}"],
+            check=True,
+            capture_output=True,
+            text=True,
+            cwd=cwd,
+        )
+        behind = int(result.stdout.strip())
+    except subprocess.CalledProcessError as e:
+        _print_freshness_failure(
+            base_branch,
+            f"`git rev-list HEAD..origin/{base_branch}` failed",
+            e.stderr or "",
+        )
+        sys.exit(3)
+    except ValueError:
+        _print_freshness_failure(
+            base_branch,
+            "could not parse commit count from `git rev-list`",
+        )
+        sys.exit(3)
+
+    if behind > 0:
+        _print_stale_branch_refusal(base_branch, behind)
+        sys.exit(3)
+
+    print(f"Branch freshness OK: up to date with origin/{base_branch}.")
+
+
+def _print_destructive_refusal(destructive_lines: list) -> None:
+    """Loudly refuse to write a destructive migration and explain how to proceed."""
+    banner = "=" * 72
+    out = sys.stderr
+    print(banner, file=out)
+    print(
+        "  DESTRUCTIVE MIGRATION DETECTED — REFUSING TO WRITE MIGRATION FILE", file=out
+    )
+    print(banner, file=out)
+    print("", file=out)
+    print(
+        "The generated migration contains the following destructive statements:",
+        file=out,
+    )
+    print("", file=out)
+    for line in destructive_lines:
+        print(f"    {line}", file=out)
+    print("", file=out)
+    print(
+        "This very often means your branch is OUT OF DATE, or your local\n"
+        "schema.prisma is inconsistent with the migrations/ directory. A\n"
+        "previous incident saw newly-added columns silently dropped this way\n"
+        "and merged to main. Stop and verify before proceeding.",
+        file=out,
+    )
+    print("", file=out)
+    print("Before re-running:", file=out)
+    print(
+        "  1. git fetch origin && git status   # confirm branch is up to date", file=out
+    )
+    print(
+        "  2. Re-check all schema.prisma files are in sync (runbook Step 0)", file=out
+    )
+    print("  3. Review EACH DROP above — is it actually intended?", file=out)
+    print("  4. If (and only if) the drops are intentional, re-run with:", file=out)
+    print("         --allow-destructive", file=out)
+    print("", file=out)
+    print(banner, file=out)
+    print(
+        "  AI AGENTS / ASSISTANTS: DO NOT automatically re-run this command\n"
+        "  with --allow-destructive. Surface this error to the human operator\n"
+        "  and wait for their explicit confirmation before passing the flag.\n"
+        "  Passing the flag without human review is the exact failure mode\n"
+        "  this guard exists to prevent.",
+        file=out,
+    )
+    print(banner, file=out)
+
+
+def create_migration(
+    migration_name: str = None,
+    allow_destructive: bool = False,
+    base_branch: str = DEFAULT_BASE_BRANCH,
+    skip_freshness_check: bool = False,
+):
     """
     Create a new migration SQL file in the migrations directory by comparing
-    current database state with schema
+    current database state with schema.
 
     Args:
         migration_name (str): Name for the migration
+        allow_destructive (bool): Required to write a migration that contains
+            DROP COLUMN, DROP TABLE, or DROP INDEX statements. Without this
+            flag, the script exits non-zero and prints guidance.
+        base_branch (str): Branch to check freshness against
+            (default: "litellm_internal_staging").
+        skip_freshness_check (bool): Skip the "branch is up to date" check.
+            Only for intentional migrations against an older base.
     """
+    root_dir = Path(__file__).parent.parent
+
+    if skip_freshness_check:
+        print(
+            "WARNING: freshness check skipped (--skip-freshness-check). "
+            "Generating a migration from a stale branch can silently drop columns."
+        )
+    else:
+        _check_branch_freshness(root_dir, base_branch)
+
     try:
-        # Get paths
-        root_dir = Path(__file__).parent.parent
         migrations_dir = (
             root_dir / "litellm-proxy-extras" / "litellm_proxy_extras" / "migrations"
         )
@@ -59,7 +268,27 @@ def create_migration(migration_name: str = None):
                     check=True,
                 )
 
-                if result.stdout.strip():
+                # Prisma emits the literal "-- This is an empty migration." when
+                # there's no real drift. Treat that as "no changes".
+                diff_sql = result.stdout
+                stripped = diff_sql.strip()
+                is_empty_diff = (
+                    not stripped or stripped == "-- This is an empty migration."
+                )
+
+                if not is_empty_diff:
+                    destructive_lines = _find_destructive_statements(diff_sql)
+                    if destructive_lines and not allow_destructive:
+                        _print_destructive_refusal(destructive_lines)
+                        sys.exit(2)
+                    if destructive_lines and allow_destructive:
+                        print(
+                            "WARNING: writing destructive migration "
+                            "(--allow-destructive passed). Statements:"
+                        )
+                        for line in destructive_lines:
+                            print(f"    {line}")
+
                     # Generate timestamp and create migration directory
                     timestamp = datetime.now().strftime("%Y%m%d%H%M%S")
                     migration_name = migration_name or "unnamed_migration"
@@ -68,7 +297,7 @@ def create_migration(migration_name: str = None):
 
                     # Write the SQL to migration.sql
                     migration_file = migration_dir / "migration.sql"
-                    migration_file.write_text(result.stdout)
+                    migration_file.write_text(diff_sql)
 
                     print(f"Created migration in {migration_dir}")
                     return True
@@ -90,8 +319,48 @@ def create_migration(migration_name: str = None):
 
 
 if __name__ == "__main__":
-    # If running directly, can optionally pass migration name as argument
-    import sys
-
-    migration_name = sys.argv[1] if len(sys.argv) > 1 else None
-    create_migration(migration_name)
+    parser = argparse.ArgumentParser(
+        description=(
+            "Generate a Prisma migration by diffing the temp DB "
+            "(existing migrations applied) against schema.prisma."
+        )
+    )
+    parser.add_argument(
+        "migration_name",
+        nargs="?",
+        default=None,
+        help="Name for the migration (used in the generated directory name).",
+    )
+    parser.add_argument(
+        "--allow-destructive",
+        action="store_true",
+        help=(
+            "Required to write a migration that contains DROP COLUMN, "
+            "DROP TABLE, or DROP INDEX. Without this flag, destructive "
+            "diffs are refused."
+        ),
+    )
+    parser.add_argument(
+        "--base-branch",
+        default=DEFAULT_BASE_BRANCH,
+        help=(
+            f"Branch to check freshness against (default: {DEFAULT_BASE_BRANCH}). "
+            "The script fetches origin/<base-branch> and refuses to run if HEAD "
+            "is behind it."
+        ),
+    )
+    parser.add_argument(
+        "--skip-freshness-check",
+        action="store_true",
+        help=(
+            "Bypass the 'branch is up to date' check. Only for intentional "
+            "migrations against an older base. Pairs poorly with automation."
+        ),
+    )
+    args = parser.parse_args()
+    create_migration(
+        args.migration_name,
+        allow_destructive=args.allow_destructive,
+        base_branch=args.base_branch,
+        skip_freshness_check=args.skip_freshness_check,
+    )

litellm-proxy-extras/migration_runbook.md

@@ -2,6 +2,8 @@
 
 This is a runbook for creating and running database migrations for the LiteLLM proxy. For use for litellm engineers only.
 
+> **AI AGENTS / ASSISTANTS:** If the script refuses with either a "STALE BRANCH" or "DESTRUCTIVE MIGRATION DETECTED" error, **do NOT** bypass it on your own (no `git rebase`, no `--skip-freshness-check`, no `--allow-destructive`). Surface the error to the human operator and wait for their explicit confirmation. See the [Branch freshness](#branch-freshness-check) and [Destructive migrations](#destructive-migrations-drop-column--drop-table) sections below.
+
 ## Step 0: Sync All `schema.prisma` Files
 
 Before doing anything else, make sure all `schema.prisma` files in the repo are in sync. There are multiple copies that must match:
@@ -46,10 +48,50 @@ uv run --with testing.postgresql python ci_cd/run_migration.py "your_migration_n
 
 ## What It Does
 
-1. Creates temp PostgreSQL DB
-2. Applies existing migrations
-3. Compares with `schema.prisma`
-4. Generates new migration if changes found
+1. **Verifies the current branch is up to date with `origin/litellm_internal_staging`** (see [Branch freshness](#branch-freshness-check))
+2. Creates temp PostgreSQL DB
+3. Applies existing migrations
+4. Compares with `schema.prisma`
+5. Generates new migration if changes found
+6. **Refuses to write if the generated SQL contains `DROP COLUMN`, `DROP TABLE`, or `DROP INDEX`**, unless `--allow-destructive` is passed
+
+## Branch Freshness Check
+
+Before generating anything, `run_migration.py` runs `git fetch origin <base>` and refuses to proceed if `HEAD` is behind `origin/<base>`. Default base is `litellm_internal_staging` (the branch PRs target). A previous incident saw a stale branch silently drop production columns; freshness is the first-line defense.
+
+Flags:
+
+- `--base-branch <name>` — check against a different base (e.g. `main`). Default is `litellm_internal_staging`.
+- `--skip-freshness-check` — bypass entirely. Only for intentional migrations against an older base.
+
+When the guard fires:
+
+1. Update your branch:
+
+   ```bash
+   git fetch origin && git rebase origin/litellm_internal_staging
+   # or git merge origin/litellm_internal_staging — whichever matches your workflow
+   ```
+2. Re-run `run_migration.py`.
+
+> **AI AGENTS / ASSISTANTS:** Do **not** auto-rebase or auto-pass `--skip-freshness-check`. A bad conflict resolution during rebase can itself drop the human's in-progress schema edits. Surface the error and wait for explicit confirmation.
+
+## Destructive Migrations (DROP COLUMN / DROP TABLE / DROP INDEX)
+
+If the generated diff contains `DROP COLUMN`, `DROP TABLE`, or `DROP INDEX`, `run_migration.py` exits non-zero and refuses to write the migration file. A previous incident saw newly-added columns silently dropped by a stale branch and merged to main — this guard exists to prevent a repeat.
+
+When the guard fires:
+
+1. Run `git fetch origin && git status` — confirm your branch is up to date with the base branch.
+2. Re-check all `schema.prisma` files are in sync (Step 0).
+3. Review EACH `DROP` statement printed in the error — is it actually intended?
+4. Only if the drops are genuinely intentional, re-run with the flag:
+
+   ```bash
+   uv run --with testing.postgresql python ci_cd/run_migration.py "your_migration_name" --allow-destructive
+   ```
+
+> **AI AGENTS / ASSISTANTS:** Do **not** automatically re-run the command with `--allow-destructive`. If the guard fires while you are driving the runbook for a human, stop, show them the error, and wait for their explicit confirmation before passing the flag. Auto-passing `--allow-destructive` is the exact failure mode this guard exists to prevent.
 
 ## Common Fixes
 

litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py

@@ -34,6 +34,7 @@ from litellm.llms.bedrock.common_utils import (
     remove_custom_field_from_tools,
 )
 from litellm.types.llms.anthropic import ANTHROPIC_TOOL_SEARCH_BETA_HEADER
+from litellm.types.llms.bedrock import BedrockInvokeAnthropicMessagesRequest
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import GenericStreamingChunk
@@ -59,6 +60,10 @@ class AmazonAnthropicClaudeMessagesConfig(
 
     DEFAULT_BEDROCK_ANTHROPIC_API_VERSION = "bedrock-2023-05-31"
 
+    BEDROCK_INVOKE_ALLOWED_TOP_LEVEL_FIELDS = frozenset(
+        BedrockInvokeAnthropicMessagesRequest.__annotations__.keys()
+    )
+
     def __init__(self, **kwargs):
         BaseAnthropicMessagesConfig.__init__(self, **kwargs)
         AmazonInvokeConfig.__init__(self, **kwargs)
@@ -500,10 +505,6 @@ class AmazonAnthropicClaudeMessagesConfig(
                 anthropic_messages_request=anthropic_messages_request,
             )
 
-        # 5b. Strip `output_config` — Bedrock Invoke doesn't support it
-        # Fixes: https://github.com/BerriAI/litellm/issues/22797
-        anthropic_messages_request.pop("output_config", None)
-
         # 5a. Remove `custom` field from tools (Bedrock doesn't support it)
         # Claude Code sends `custom: {defer_loading: true}` on tool definitions,
         # which causes Bedrock to reject the request with "Extra inputs are not permitted"
@@ -550,14 +551,43 @@ class AmazonAnthropicClaudeMessagesConfig(
         if "tool-search-tool-2025-10-19" in beta_set:
             beta_set.add("tool-examples-2025-10-29")
 
-        filtered_auto_betas = filter_and_transform_beta_headers(
-            beta_headers=list(beta_set - user_beta_set),
-            provider="bedrock",
+        filtered_betas = sorted(
+            filter_and_transform_beta_headers(
+                beta_headers=list(beta_set),
+                provider="bedrock",
+            )
         )
-        filtered_betas = sorted(user_beta_set.union(set(filtered_auto_betas)))
+
+        dropped_user_betas = sorted(
+            b
+            for b in user_beta_set
+            if not filter_and_transform_beta_headers([b], provider="bedrock")
+        )
+        if dropped_user_betas:
+            verbose_logger.warning(
+                "Bedrock Invoke: dropping unsupported anthropic-beta values "
+                "from client headers: %s. Bedrock has no mapping entry for "
+                "these; forwarding them would cause a 400.",
+                dropped_user_betas,
+            )
+
         if filtered_betas:
             anthropic_messages_request["anthropic_beta"] = filtered_betas
 
+        # 7. Final safety net: filter top-level fields to the Bedrock Invoke allowlist.
+        # Catches Anthropic-only extensions (context_management, output_config, speed,
+        # mcp_servers, ...) and any future additions Claude Code may start sending.
+        allowed = self.BEDROCK_INVOKE_ALLOWED_TOP_LEVEL_FIELDS
+        stripped = sorted(k for k in anthropic_messages_request if k not in allowed)
+        if stripped:
+            verbose_logger.debug(
+                "Bedrock Invoke: stripping unsupported top-level request fields: %s",
+                stripped,
+            )
+        anthropic_messages_request = {
+            k: v for k, v in anthropic_messages_request.items() if k in allowed
+        }
+
         return anthropic_messages_request
 
     def get_async_streaming_response_iterator(

litellm/types/llms/bedrock.py

@@ -997,3 +997,47 @@ class BedrockToolBlock(TypedDict, total=False):
     toolSpec: Optional[ToolSpecBlock]
     systemTool: Optional[SystemToolBlock]  # For Nova grounding
     cachePoint: Optional[CachePointBlock]
+
+
+class BedrockInvokeAnthropicMessagesRequest(TypedDict, total=False):
+    """
+    Top-level request body accepted by AWS Bedrock `InvokeModel` /
+    `InvokeModelWithResponseStream` when calling an Anthropic Claude model with
+    the Messages API format. The LiteLLM /v1/messages → Bedrock Invoke
+    transformation filters outgoing requests to the keys of this TypedDict; any
+    other field (Anthropic-only extension, internal metadata, future addition)
+    is dropped before signing so Bedrock doesn't 400 with
+    "Extra inputs are not permitted".
+
+    Reference:
+        https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages.html
+        https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages-request-response.html
+
+    Editing this type is the single source of truth — the runtime allowlist in
+    `AmazonAnthropicClaudeMessagesConfig.BEDROCK_INVOKE_ALLOWED_TOP_LEVEL_FIELDS`
+    is derived from `__annotations__`, and a test asserts the resolved set
+    exactly, so any edit forces a conscious review.
+
+    Value types are intentionally loose (`list`, `dict`) — this type exists to
+    pin the allowed field names, not to validate nested structure.
+    """
+
+    # Required by Bedrock
+    anthropic_version: str
+    max_tokens: int
+    messages: list
+
+    # Documented optional fields
+    anthropic_beta: List[str]
+    system: object  # str or list[TextBlock]
+    stop_sequences: List[str]
+    temperature: float
+    top_p: float
+    top_k: int
+    tools: list
+    tool_choice: dict
+
+    # `thinking` is required for Opus 4.5 / Sonnet 4 extended thinking,
+    # `metadata` is part of the common Anthropic Messages API shape.
+    thinking: dict
+    metadata: dict

pyproject.toml

@@ -208,7 +208,7 @@ build-backend = "uv_build"
 
 [tool.uv]
 default-groups = ["dev"]
-required-version = "==0.10.9"
+required-version = ">=0.10.9"
 exclude-newer = "3 days"
 
 [tool.uv.sources]

tests/proxy_unit_tests/conftest.py

@@ -1,48 +1,143 @@
 # conftest.py
 
-import importlib
+import asyncio
+import copy
+import inspect
 import os
 import sys
+import warnings
 
 import pytest
 
 sys.path.insert(
     0, os.path.abspath("../..")
 )  # Adds the parent directory to the system path
+
 import litellm
+import litellm.proxy.proxy_server
+
+
+# Top-level assignments of these types are the ones importlib.reload(litellm)
+# would have effectively reset. We snapshot them at conftest import time and
+# deep-copy the snapshot back before every test.
+_SNAPSHOT_TYPES = (list, dict, set, tuple, str, int, float, bool, bytes)
+
+
+def _snapshot_mutable_state(module):
+    """Capture a per-module snapshot of primitive and collection attributes."""
+    snapshot = {}
+    for attr in list(vars(module)):
+        if attr.startswith("_"):
+            continue
+        try:
+            value = getattr(module, attr)
+        except Exception as exc:
+            warnings.warn(
+                f"conftest: could not read {module.__name__}.{attr} during snapshot: {exc}",
+                stacklevel=2,
+            )
+            continue
+        if value is None or isinstance(value, _SNAPSHOT_TYPES):
+            try:
+                snapshot[attr] = copy.deepcopy(value)
+            except Exception as exc:
+                warnings.warn(
+                    f"conftest: could not snapshot {module.__name__}.{attr}: {exc}",
+                    stacklevel=2,
+                )
+    return snapshot
+
+
+def _restore_mutable_state(module, snapshot):
+    for attr, default in snapshot.items():
+        try:
+            setattr(module, attr, copy.deepcopy(default))
+        except Exception as exc:
+            warnings.warn(
+                f"conftest: could not restore {module.__name__}.{attr}: {exc}",
+                stacklevel=2,
+            )
+
+
+def _collect_flushable_caches():
+    """Return (module, attr) pairs whose values expose flush_cache()."""
+    targets = []
+    for module in (litellm, litellm.proxy.proxy_server):
+        for attr in list(vars(module)):
+            if attr.startswith("_"):
+                continue
+            try:
+                value = getattr(module, attr)
+            except Exception:
+                continue
+            # Only instances — a class reference has an unbound flush_cache
+            # that can't be called without a self argument.
+            if inspect.isclass(value) or inspect.ismodule(value):
+                continue
+            if callable(getattr(value, "flush_cache", None)):
+                targets.append((module, attr))
+    return targets
+
+
+def _flush_caches(targets):
+    for module, attr in targets:
+        try:
+            value = getattr(module, attr)
+        except Exception:
+            continue
+        flush = getattr(value, "flush_cache", None)
+        if callable(flush):
+            try:
+                flush()
+            except Exception as exc:
+                warnings.warn(
+                    f"conftest: flush_cache failed on {module.__name__}.{attr}: {exc}",
+                    stacklevel=2,
+                )
+
+
+# Snapshot once at conftest import — these are the "clean" module states.
+_LITELLM_STATE = _snapshot_mutable_state(litellm)
+_PROXY_SERVER_STATE = _snapshot_mutable_state(litellm.proxy.proxy_server)
+_FLUSHABLE_CACHES = _collect_flushable_caches()
 
 
 @pytest.fixture(scope="function", autouse=True)
 def setup_and_teardown():
+    """Reset mutable module state on litellm and proxy_server before each test.
+
+    Replaces a previous importlib.reload(litellm) approach that cost ~17s
+    per test (re-executing the full litellm __init__ import chain).
+
+    What IS reset:
+      - Top-level module attributes of type list / dict / set / tuple
+        / str / int / float / bool / bytes, and None-valued attributes.
+        These cover callback lists, general_settings, master_key,
+        premium_user, prisma_client, etc. — anything the old reload() reset
+        by re-executing the module body.
+      - Any module-level object instance that exposes flush_cache() (the
+        DualCache and LLMClientCache family), which handles cache state
+        that can't round-trip through deepcopy because of internal locks.
+
+    What is NOT reset:
+      - Class instances without flush_cache() (e.g. ProxyLogging,
+        JWTHandler, FastAPI routers, loggers). If a test mutates such an
+        instance in-place (setattr on the instance, appending to one of
+        its internal lists, etc.), the mutation will leak into later tests.
+        Use pytest's monkeypatch.setattr() or a local fixture for those
+        cases — don't rely on this autouse fixture to undo them.
     """
-    This fixture reloads litellm before every function. To speed up testing by removing callbacks being chained.
-    """
-    curr_dir = os.getcwd()  # Get the current working directory
-    sys.path.insert(
-        0, os.path.abspath("../..")
-    )  # Adds the project directory to the system path
-
-    import litellm
-    from litellm import Router
-
-    importlib.reload(litellm)
-    try:
-        if hasattr(litellm, "proxy") and hasattr(litellm.proxy, "proxy_server"):
-            importlib.reload(litellm.proxy.proxy_server)
-    except Exception as e:
-        print(f"Error reloading litellm.proxy.proxy_server: {e}")
-
-    import asyncio
+    _restore_mutable_state(litellm, _LITELLM_STATE)
+    _restore_mutable_state(litellm.proxy.proxy_server, _PROXY_SERVER_STATE)
+    _flush_caches(_FLUSHABLE_CACHES)
 
     loop = asyncio.get_event_loop_policy().new_event_loop()
     asyncio.set_event_loop(loop)
-    print(litellm)
-    # from litellm import Router, completion, aembedding, acompletion, embedding
-    yield
-
-    # Teardown code (executes after the yield point)
-    loop.close()  # Close the loop created earlier
-    asyncio.set_event_loop(None)  # Remove the reference to the loop
+    try:
+        yield
+    finally:
+        loop.close()
+        asyncio.set_event_loop(None)
 
 
 def pytest_collection_modifyitems(config, items):

tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py

@@ -579,6 +579,155 @@ def test_bedrock_messages_strips_output_config_with_output_format():
     assert "output_format" not in result
 
 
+def test_bedrock_messages_strips_context_management():
+    """
+    Ensure context_management is stripped from the request before sending to
+    Bedrock Invoke, which doesn't support this Anthropic-specific parameter.
+
+    Claude Code sends context_management on every request; leaving it in the body
+    causes a 400 "context_management: Extra inputs are not permitted" from Bedrock.
+    """
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    messages = [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}]
+    optional_params = {
+        "max_tokens": 4096,
+        "context_management": {
+            "edits": [{"type": "clear_thinking_20251015", "keep": "all"}]
+        },
+    }
+
+    result = cfg.transform_anthropic_messages_request(
+        model="anthropic.claude-3-haiku-20240307-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params=optional_params,
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    assert (
+        "context_management" not in result
+    ), "context_management should be stripped — Bedrock Invoke rejects it"
+    assert result.get("max_tokens") == 4096
+
+
+def test_bedrock_messages_allowlist_filters_anthropic_only_fields():
+    """
+    Bedrock Invoke rejects any top-level body field it doesn't recognize with
+    "Extra inputs are not permitted". Defend against that by filtering the
+    outgoing body to a Bedrock-supported allowlist — catches Anthropic-only
+    extensions (speed, mcp_servers, container, ...) and any future additions
+    Claude Code starts sending before we learn about them.
+    """
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    messages = [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}]
+    optional_params = {
+        "max_tokens": 4096,
+        "temperature": 0.5,
+        "speed": "fast",
+        "mcp_servers": [{"type": "url", "url": "https://example.com"}],
+        "container": {"skills": []},
+        "inference_geo": "us",
+        "output_config": {"effort": "low"},
+        "context_management": {"edits": []},
+    }
+
+    result = cfg.transform_anthropic_messages_request(
+        model="anthropic.claude-3-haiku-20240307-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params=optional_params,
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+
+    for bad in (
+        "speed",
+        "mcp_servers",
+        "container",
+        "inference_geo",
+        "output_config",
+        "context_management",
+        "model",
+        "stream",
+    ):
+        assert bad not in result, f"{bad} should be stripped by the allowlist"
+
+    # Supported fields pass through.
+    assert result["max_tokens"] == 4096
+    assert result["temperature"] == 0.5
+    assert result["anthropic_version"] == cfg.DEFAULT_BEDROCK_ANTHROPIC_API_VERSION
+    # Every surviving key is in the allowlist.
+    assert set(result).issubset(cfg.BEDROCK_INVOKE_ALLOWED_TOP_LEVEL_FIELDS)
+
+
+def test_bedrock_messages_filters_user_provided_unsupported_beta_header():
+    """
+    In proxy deployments the client (e.g. Claude Code) doesn't know the backend
+    is Bedrock and may send Anthropic-direct beta headers Bedrock can't handle.
+    All betas must go through the provider mapping, not just auto-injected ones
+    — otherwise Bedrock 400s on the unsupported value.
+    """
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    messages = [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}]
+    optional_params = {"max_tokens": 128}
+    # `advisor-tool-2026-03-01` has no bedrock mapping entry → must be dropped.
+    # `context-1m-2025-08-07` does → must pass through.
+    headers = {
+        "anthropic-beta": "advisor-tool-2026-03-01,context-1m-2025-08-07",
+    }
+
+    result = cfg.transform_anthropic_messages_request(
+        model="anthropic.claude-3-haiku-20240307-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params=optional_params,
+        litellm_params=GenericLiteLLMParams(),
+        headers=headers,
+    )
+
+    betas = result.get("anthropic_beta") or []
+    assert (
+        "advisor-tool-2026-03-01" not in betas
+    ), "user-provided beta not in the Bedrock mapping must be dropped"
+    assert (
+        "context-1m-2025-08-07" in betas
+    ), "user-provided beta that IS in the Bedrock mapping should survive"
+
+
+def test_bedrock_messages_renames_user_provided_aliased_beta_header():
+    """
+    Bedrock's config maps `advanced-tool-use-2025-11-20` to
+    `tool-search-tool-2025-10-19`. User-provided betas must go through the
+    rename too, not be forwarded under their Anthropic-direct spelling.
+    """
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    messages = [{"role": "user", "content": [{"type": "text", "text": "Hello"}]}]
+    optional_params = {"max_tokens": 128}
+    headers = {"anthropic-beta": "advanced-tool-use-2025-11-20"}
+
+    result = cfg.transform_anthropic_messages_request(
+        model="anthropic.claude-3-haiku-20240307-v1:0",
+        messages=messages,
+        anthropic_messages_optional_request_params=optional_params,
+        litellm_params=GenericLiteLLMParams(),
+        headers=headers,
+    )
+
+    betas = result.get("anthropic_beta") or []
+    assert (
+        "advanced-tool-use-2025-11-20" not in betas
+    ), "Anthropic-direct spelling should be rewritten, not forwarded verbatim"
+    assert (
+        "tool-search-tool-2025-10-19" in betas
+    ), "user-provided beta should be renamed to the Bedrock-side spelling"
+
+
 @pytest.mark.asyncio
 async def test_promote_message_stop_usage_preserves_message_delta_output_tokens():
     """

tests/test_litellm/llms/bedrock/test_anthropic_beta_support.py

@@ -95,7 +95,7 @@ class TestAnthropicBetaHeaderSupport:
     def test_messages_transformation_anthropic_beta(self):
         """Test that Messages API transformation includes anthropic_beta in request."""
         config = AmazonAnthropicClaudeMessagesConfig()
-        headers = {"anthropic-beta": "output-128k-2025-02-19"}
+        headers = {"anthropic-beta": "context-1m-2025-08-07"}
 
         result = config.transform_anthropic_messages_request(
             model="anthropic.claude-haiku-4-5-20251001-v1:0",
@@ -107,7 +107,7 @@ class TestAnthropicBetaHeaderSupport:
 
         assert "anthropic_beta" in result
         # Sort both arrays before comparing to avoid flakiness from ordering differences
-        assert sorted(result["anthropic_beta"]) == sorted(["output-128k-2025-02-19"])
+        assert sorted(result["anthropic_beta"]) == sorted(["context-1m-2025-08-07"])
 
     def test_converse_computer_use_compatibility(self):
         """Test that user anthropic_beta headers work with computer use tools."""