feat(skills)!: add portable Agent Skills Hub API (#50159)

Co-authored-by: mateo <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

.agents/skills/database-access/SKILL.md

@@ -0,0 +1,73 @@
+---
+name: database-access
+description: Organize database repositories and proxy dependency wiring, manage connection ownership, or port Prisma operations to Rust and SQLx without expanding migration scope
+metadata:
+  internal: true
+---
+
+# Database access
+
+## Situations
+
+- Adding a query or an operation to a repository, read [adding a query](references/python-repository.md#adding-a-query)
+- Moving existing Prisma calls behind a repository, read [moving existing access](references/python-repository.md#moving-existing-access-into-a-repository)
+- Porting a repository from Prisma to Rust and SQLx, read [porting to SQLx](references/python-repository.md#porting-a-repository-to-sqlx)
+- Injecting a repository into a route, read [Python and FastAPI](references/python-fastapi.md) or [Rust and Axum](references/rust-axum.md)
+
+## Adding new database access
+
+- Put Python database access behind an operation in `litellm/repositories`
+- Give each repository its own FastAPI provider, so a route builds only the repositories it uses
+- A provider borrows existing handles and performs no database I/O during construction
+- Define the operations a route calls as a Protocol in `litellm/models/<feature>.py`, next to its domain types
+  - The Protocol takes and returns domain types, never Prisma models, raw rows or serialized payloads
+  - Parsing and serialization of stored values live in the repository
+  - Add a Protocol when a repository gets a provider, not up front for every repository
+- Keep query execution inside the repository, with one implementation per repository and no backend switch or fallback
+- A repository is either all Prisma or all native, never a mix of both
+- Prisma is fine for the implementation, adding a query does not require a Rust port
+- Native Rust callers use `litellm-rust/crates/db` directly
+
+## Framework wiring
+
+- Python routes receive a per-repository dependency such as `UserBannerRepositoryDep` and call its operations, such as `banners.get_banner()`
+- `DatabaseDep` lives in `litellm/proxy/deps/database.py`, repository providers in `litellm/proxy/deps/repositories.py`
+  - A provider returns the feature's Protocol, so routes never see which backend implements it
+  - Dependencies used by one feature only, like the banner audit binding, stay next to that feature's routes
+- Repositories never import FastAPI, so jobs and helpers can use them without a request
+- Use named FastAPI providers with `Annotated[..., Depends(...)]`, ordinary helpers and jobs receive explicit arguments
+- Preserve each route's no-database behavior, authorization precedence and background-task scheduling
+- Read [Python and FastAPI](references/python-fastapi.md) for provider examples, lifespan state and route verification
+- Read [Rust and Axum](references/rust-axum.md) when adding database access to Rust handlers
+
+## Connection ownership
+
+- Keep connection setup and shutdown separate from repositories, which own no pools
+- Reuse existing handles, never connect per request or disconnect from a repository operation
+- Give each pool one owner for setup and shutdown, borrowed handles never close it
+- Finish requests, jobs and queued database work before closing connections
+- Clean up resources opened before a startup failure, without a second disconnect fallback
+- Add a backend to the existing lifecycle owner and pass its handle only to repositories that need it
+- Preserve cancellation, timeout, error and transaction semantics across the native bridge
+
+## Changing ownership or reloading resources
+
+- Inventory initialization, reload, tests, background jobs and shutdown together
+- Move every writer when changing authoritative storage, copying a replaceable global into state is insufficient
+- A frozen repository does not make its clients immutable
+- Decide whether each request captures one resource generation or reads it again during processing
+- Build and validate replacements before publishing them, preserving the old generation on failure
+- Serialize reloads and coordinate publication with shutdown
+- Drain streams and background users before retiring old resources, retaining every generation grows resource use
+- Define multi-worker coordination separately from an in-process swap
+- Choose mechanisms such as `ArcSwap` only after defining publication, capture and retirement
+
+## Migration context and boundaries
+
+- Long-term direction is to centralize Python database access in repositories, then replace their Prisma implementations with Rust APIs one repository at a time
+- Prisma and native repositories coexist during the transition, unrelated existing calls can stay as they are
+- Prisma and SQLx keep separate client pools to PgBouncer, which pools connections to PostgreSQL
+  - Size both layers together and preserve reader/writer routing and pooling-mode compatibility
+- PgBouncer does not share transactions between Prisma and SQLx, keep each atomic unit of work on one backend and transaction
+- Never retry a failed Rust write through Prisma when it may already have committed
+- Prisma owns schema migrations today, eventual SQLx migration ownership and Prisma removal are separate planned steps, not part of a query port

.agents/skills/database-access/references/python-fastapi.md

@@ -0,0 +1,101 @@
+# Python and FastAPI
+
+## Existing pattern
+
+- `UserBannerStore` in `litellm/models/user_banner.py` is the Protocol the banner routes call
+- `get_database` in `litellm/proxy/deps/database.py` reads the lifespan-owned handle from request state
+- `get_user_banner_repository` in `litellm/proxy/deps/repositories.py` builds the banner repository from that handle
+- Choose each repository's replica routing once in its provider, never at call sites
+- A handler receives `UserBannerStore | None`, never a Prisma client or a concrete repository
+- `UserBannerRepository` implements `UserBannerStore` structurally, keeping the existing Prisma table, serialization and configuration-sync behavior
+- The banner audit provider binds the legacy audit callable at request time, without performing audit I/O
+- A native repository can later implement the same Protocol without changing handler signatures, see [porting a repository to SQLx](python-repository.md#porting-a-repository-to-sqlx)
+
+## Dependency injection
+
+- Add one provider and one `Annotated` alias per repository in `litellm/proxy/deps/repositories.py`
+- Annotate the provider and alias with the feature's Protocol, not the concrete repository
+- A route depends only on the repositories it calls, so a request never builds repositories it does not use
+- Do not build repositories once at lifespan, a reconnect can replace the Prisma client they wrap
+
+```python
+class UserBannerStore(Protocol):
+    async def get_banner(self) -> UserBanner: ...
+
+    async def save_banner(self, banner: UserBanner) -> None: ...
+```
+
+```python
+async def get_user_banner_repository(database: DatabaseDep) -> UserBannerStore | None:
+    return None if database is None else UserBannerRepository(database.prisma)
+
+
+UserBannerRepositoryDep: TypeAlias = Annotated[UserBannerStore | None, Depends(get_user_banner_repository)]
+```
+
+- FastAPI caches each provider within a request, so routes and sub-dependencies share one instance
+- Ordinary helpers and jobs receive explicit arguments, never unresolved `Depends` defaults
+- A route still reading the `prisma_client` global or building a repository inline switches to the repository's `...RepositoryDep`, keeping its no-database behavior and error precedence
+
+```python
+async def get_user_banner(banners: UserBannerRepositoryDep) -> UserBanner:
+    if banners is None:
+        return UserBanner()
+    return await banners.get_banner()
+```
+
+## Lifespan state
+
+- Prisma startup owns connection setup, authentication refresh, health checks and migrations
+- `Database` holds the initialized Prisma client and the `NativeDatabase` handle, without holding fields per feature
+  - The native handle opens no pool yet, its `connect` and `close` are noops
+- The proxy lifespan builds the handle with `Database.open`, registers `Database.aclose` on its exit stack and publishes it in request state
+  - Callers only open and close `Database`, which closes Prisma and the native handle together, and `open` closes Prisma itself when the native connect fails
+- Starlette shallow-copies lifespan state into request state, which is distinct from `app.state`
+- The database module never imports a repository, legacy globals borrow its handle without owning shutdown
+- Drain requests, jobs, logging, spend and audit work before database teardown
+- The exit stack closes the handle on exit, including startup failures after the client is created
+- When adding native resources, initialize them in the serving process, respecting runtime fork restrictions
+- Await native shutdown instead of depending on Python garbage collection
+- Apply the [connection ownership rules](../SKILL.md#connection-ownership) when changing the owner
+
+## Timing and errors
+
+- FastAPI resolves dependencies before entering the handler
+- Handler exception handling cannot catch failures raised during dependency resolution
+- A universal required-database dependency can override authorization and validation behavior
+- Banner GET without a database returns the default disabled banner
+- Banner PATCH checks the admin role before returning the existing missing-database error
+- Invalid stored banner data falls back to the default, but storage failures still propagate
+- Banner writes finish before the success audit is queued as a FastAPI background task
+- Queue post-response work with `BackgroundTasks`, not bare `asyncio.create_task`, so the in-flight drain waits for it before the database closes
+- Do not move background work into an awaited provider merely to remove an import
+
+## Verification
+
+- Exercise real FastAPI handlers using `dependency_overrides` for external boundaries
+- Keep production parsing and persistence code real
+- Restore overrides after each test
+- Verify complete response bodies, status codes and OpenAPI against the base
+- Include combined failure conditions to pin error precedence
+- Check saved values, fresh revisions, audit contents and failed-write behavior
+- Check lifespan state with a running lifespan, not only a TestClient that skips startup
+- Verify providers do not connect, query, disconnect or audit eagerly
+- Use real PostgreSQL and live proxy requests when changing persistence behavior
+
+## Scope
+
+- Preserve direct Python call sites or migrate them explicitly
+- Inventory imports, state writers, side effects and error ordering before migrating a feature
+- Preserve configuration sync and serialization, and state which legacy reads remain
+- Do not migrate every route in a large module just to establish the pattern
+- Do not invent providers for resources without consumers
+
+## Framework references
+
+- [Lifespan events](https://fastapi.tiangolo.com/advanced/events/) and [Starlette lifespan state](https://starlette.dev/lifespan/)
+- [`SessionDep`-style `Annotated` aliases](https://fastapi.tiangolo.com/tutorial/sql-databases/#create-a-session-dependency), which `DatabaseDep` and `UserBannerRepositoryDep` follow
+- [Dependency composition and caching](https://fastapi.tiangolo.com/tutorial/dependencies/sub-dependencies/)
+- [Background tasks and dependencies with `yield`](https://fastapi.tiangolo.com/advanced/advanced-dependencies/#background-tasks-and-dependencies-with-yield-technical-details)
+- [Dependency overrides](https://fastapi.tiangolo.com/advanced/testing-dependencies/)
+- [Testing lifespan](https://fastapi.tiangolo.com/advanced/testing-events/)

.agents/skills/database-access/references/python-repository.md

@@ -0,0 +1,119 @@
+# Python repositories
+
+## Adding a query
+
+Use this when a feature needs a read or write that no repository offers yet, including one more operation on an existing repository
+
+### Before starting
+
+- Search `litellm/repositories` for a repository that already owns the table, and extend it instead of adding a second one
+- Check `litellm/models/<feature>.py` for an existing Protocol and domain types
+- Decide the route's behavior when no database is configured, and where that check sits relative to authorization and validation
+
+### Steps
+
+1. Domain types and Protocol in `litellm/models/<feature>.py`
+   - The Protocol takes and returns domain types, never Prisma models, raw rows or serialized payloads
+   - Add the Protocol together with the provider, not for repositories nothing injects
+2. Repository in `litellm/repositories/<feature>_repository.py`
+   - Implement the operation with Prisma, a Rust port is not required
+   - Compose the Prisma client and type the tables it touches with `TableActions` from `prisma_protocols.py`, never inherit `PrismaTableRepository`
+   - Parse and serialize stored values here, so callers only see domain types
+   - Never import FastAPI, connect, disconnect or choose replicas per call
+3. Provider and `Annotated` alias in `litellm/proxy/deps/repositories.py`, see [dependency injection](python-fastapi.md#dependency-injection)
+4. The route takes `<Feature>RepositoryDep`, handles `None` with the behavior decided above, and calls the operation
+5. Jobs and helpers that need the operation receive the repository as an explicit argument, built by their caller from the handle it already holds
+
+### Done when
+
+- Handler tests drive the real route with `dependency_overrides` and the real repository code, see [verification](python-fastapi.md#verification)
+- Tests cover the no-database path, error precedence, and what was actually stored
+- Persistence changes are checked against real PostgreSQL through a live proxy
+
+### Out of scope
+
+- Porting the query to Rust, see [porting a repository to SQLx](#porting-a-repository-to-sqlx)
+- Moving unrelated direct Prisma calls in the same module
+
+## Moving existing access into a repository
+
+Use this when code calls Prisma directly, through `prisma_client.db.<table>` or a helper, and the calls should live behind a repository. The backend stays Prisma
+
+### Before starting
+
+- List every call being moved, its callers, and the transaction or batch it runs in
+- List the module's imports, global state writers, side effects such as audit and config sync, and error ordering
+- Record current return values, raised errors and response bodies, these are what the change must preserve
+
+### Steps
+
+1. Move the existing Prisma code into the relevant repository in `litellm/repositories` without rewriting the queries
+2. Add domain types and a Protocol in `litellm/models/<feature>.py` when a route will inject the repository
+3. Replace each direct call with the repository operation, through a provider for routes and explicit arguments for jobs and helpers, see [Python and FastAPI](python-fastapi.md)
+4. Calls sharing a transaction move into the same repository or unit of work in one change
+5. State which legacy reads remain and why, when the change does not move all of them
+
+### Done when
+
+- Focused regression tests pin the preserved return values, errors and transaction behavior
+- Response bodies, status codes and OpenAPI match the base
+- Configuration sync and serialization behave the same, including invalid stored data
+
+### Out of scope
+
+- Switching the backend, even when an equivalent Rust query exists
+- Converting other routes in the module just to establish the pattern
+- Cleanups unrelated to the moved calls
+
+## Porting a repository to SQLx
+
+Status: TBD. No repository has been ported yet. `database.native` is a `NativeDatabase` from `crates/python-bridge/src/db.rs` whose `connect` and `close` are noops that open no pool, and the PyO3 repository surface does not exist. A real pool takes typed writer/reader settings and a token provider, never the Prisma `DATABASE_URL`, which carries Prisma-only params and rotating IAM tokens. The constraints below are settled, the steps are a sketch to revise after the first port lands
+
+Use this only when a port is explicitly requested. Repository extraction does not require one
+
+### When not to port
+
+- An equivalent Rust or SQLx query existing is not a reason to switch a Prisma caller
+- Leave existing backend selection unchanged unless integration is explicitly requested
+- Never port part of a repository, it is either all Prisma or all native
+
+### Before starting
+
+- Identify the repository's inputs, outputs, errors and transaction boundary
+- List repositories that share a transaction with it, they move together
+- Check `litellm-rust/crates/db` for an equivalent query and reuse it when its behavior matches
+
+### Steps
+
+1. Rust queries, following `litellm-rust/crates/db/AGENTS.md`
+   - Static SQL in `queries/`, compile-checked with SQLx, typed inputs and row projections in `src/`
+   - Dynamic filters and sort keys become closed Rust enums that select static SQL, every value is bound
+   - Refresh SQLx query metadata after SQL changes
+2. PyO3 surface exposing typed operations, never SQL text, table names or column names from Python. TBD: where it lives and how the handle reaches it
+3. Native repository implementing the same Protocol in `litellm/repositories`
+   - Converts typed PyO3 results to domain types, preserving defaults, invalid-data handling and storage errors
+   - Python conversion stays in the bridge, not in `crates/db`
+4. Switch the provider in `litellm/proxy/deps/repositories.py`, the Protocol and routes stay unchanged
+
+```diff
+ async def get_user_banner_repository(database: DatabaseDep) -> UserBannerStore | None:
+-    return None if database is None else UserBannerRepository(database.prisma)
++    return None if database is None else NativeUserBannerRepository(database.native)
+```
+
+5. Delete the Prisma repository in the same change and switch every caller, including jobs and helpers
+
+### Done when
+
+- Behavior is verified against real PostgreSQL, covering null handling, conflicts, ordering, errors and transaction semantics
+- `DATABASE_URL=... cargo test -p litellm-db --locked -- --include-ignored` passes against a Prisma-migrated database
+- Live proxy requests show unchanged responses. TBD: a Python-side real-PostgreSQL test harness
+
+### Rollback
+
+- Roll back the release, never retry an uncertain native write through Prisma
+
+### Out of scope
+
+- SQLx owning schema migrations, and removing Prisma
+- Converting unrelated routes to dependency injection

.agents/skills/database-access/references/rust-axum.md

@@ -0,0 +1,4 @@
+# Rust and Axum
+
+- This should use same `crates/db` what is exposed to Python world.
+- Details are TBD.

.circleci/AGENTS.md

@@ -0,0 +1,7 @@
+# .circleci
+
+The `integration` workflow on `main` runs from a scheduled pipeline in the CircleCI project settings, cron `9 0-8,14-23 * * *` (UTC) on `main` with no pipeline parameters. It has no `triggers:` block in this file because a workflow with one runs only on its schedule, so the API pipelines that set `run_migration_tests` or `routing_parity_base` would stop running it. If main stops getting integration runs, check that the schedule still exists and recreate it with these values
+
+Jobs clone with the `checkout_over_https` command, never the built-in `checkout` step, because CircleCI falls back to the project's SSH checkout key without its GitHub App access, GitHub rejects that key, and the public repo needs no key. It fetches the job's commit with full history and no file contents (`--filter=blob:none`), like the built-in step did, so `scripts/path_filter.sh` can find a merge base
+
+A step that checks out files from another commit fetches that commit with `--no-filter`, as `routing_parity` does, because the clone has no file contents and git would otherwise download each missing file in its own request, which took 210 to 896 seconds for the routing parity base

.circleci/config.yml

@@ -21,19 +21,30 @@ orbs:
   win: circleci/windows@5.0 # Add Windows orb
 
 commands:
-  checkout_migration_source:
+  checkout_over_https:
+    parameters:
+      revision:
+        type: string
+        default: ""
+      shell:
+        type: string
+        default: /bin/bash -eo pipefail
     steps:
       - run:
-          name: Select the requested migration test revision
+          name: Check out the public repository over HTTPS
+          shell: << parameters.shell >>
           environment:
-            MIGRATION_SOURCE_SHA: << pipeline.parameters.migration_source_sha >>
+            REQUESTED_REVISION: << parameters.revision >>
           command: |
-            revision="${MIGRATION_SOURCE_SHA:-$CIRCLE_SHA1}"
+            revision="${REQUESTED_REVISION:-$CIRCLE_SHA1}"
             [[ "$revision" =~ ^[0-9a-f]{40}$ ]] || exit 1
             git init
             git remote add origin https://github.com/BerriAI/litellm.git
-            git fetch --depth 1 origin "$revision"
-            git checkout --detach FETCH_HEAD
+            for attempt in 1 2 3; do
+              git fetch --filter=blob:none origin "$revision" && git checkout --detach FETCH_HEAD && break
+              [[ $attempt -lt 3 ]] || exit 1
+              sleep 5
+            done
   skip_if_unrelated_changes:
     parameters:
       category:
@@ -347,7 +358,7 @@ commands:
       - when:
           condition: << parameters.checkout >>
           steps:
-            - checkout
+            - checkout_over_https
       - setup_google_dns
       - install_uv
       - install_rust
@@ -377,7 +388,8 @@ jobs:
       CARGO_HTTP_MULTIPLEXING: "false"
       CARGO_NET_RETRY: "5"
     steps:
-      - checkout
+      - checkout_over_https:
+          shell: bash.exe
       - run:
           name: Skip job when no windows-release-relevant files changed
           shell: bash.exe
@@ -451,7 +463,7 @@ jobs:
           password: ${DOCKERHUB_PASSWORD}
     working_directory: ~/project
     steps:
-      - checkout
+      - checkout_over_https
       - setup_google_dns
       - install_uv
       - install_rust
@@ -482,7 +494,7 @@ jobs:
     working_directory: ~/project
 
     steps:
-      - checkout
+      - checkout_over_https
       - skip_if_unrelated_changes
       - attach_workspace:
           at: ~/project
@@ -559,7 +571,7 @@ jobs:
     resource_class: medium
     working_directory: ~/project
     steps:
-      - checkout
+      - checkout_over_https
       - run:
           name: Generate LiteLLM master key
           command: |
@@ -659,7 +671,7 @@ jobs:
     resource_class: large
     working_directory: ~/project
     steps:
-      - checkout
+      - checkout_over_https
       - run:
           name: Generate LiteLLM master key
           command: |
@@ -740,7 +752,7 @@ jobs:
     resource_class: large
     working_directory: ~/project
     steps:
-      - checkout
+      - checkout_over_https
       - skip_if_unrelated_changes:
           category: client
       - setup_google_dns
@@ -794,7 +806,7 @@ jobs:
       # the rest of the suite: nothing else performs a logout.
       PROXY_LOGOUT_URL: "https://www.example.com"
     steps:
-      - checkout
+      - checkout_over_https
       - run:
           name: Generate LiteLLM master key
           command: |
@@ -940,7 +952,7 @@ jobs:
       # and by migration.serverRootPath.config.ts, which refuses to run without it.
       SERVER_ROOT_PATH: "/litellm"
     steps:
-      - checkout
+      - checkout_over_https
       - run:
           name: Generate LiteLLM master key
           command: |
@@ -1066,11 +1078,12 @@ jobs:
       - when:
           condition: << parameters.migration_qualification >>
           steps:
-            - checkout_migration_source
+            - checkout_over_https:
+                revision: << pipeline.parameters.migration_source_sha >>
       - unless:
           condition: << parameters.migration_qualification >>
           steps:
-            - checkout
+            - checkout_over_https
             - skip_if_unrelated_changes
 
       - run:
@@ -1122,7 +1135,8 @@ jobs:
       MIGRATION_TEST_OUTPUT: /tmp/migration-results
       PYTHONPATH: tests/e2e
     steps:
-      - checkout_migration_source
+      - checkout_over_https:
+          revision: << pipeline.parameters.migration_source_sha >>
       - run:
           name: Generate LiteLLM master key
           command: |
@@ -1197,7 +1211,7 @@ jobs:
     resource_class: medium
     working_directory: ~/project
     steps:
-      - checkout
+      - checkout_over_https
       - skip_if_unrelated_changes
       - attach_workspace:
           at: ~/project
@@ -1291,8 +1305,21 @@ jobs:
                 name: Install MCP SDK1 peer
                 command: |
                   uv venv --python 3.12 .venv-mcp-peer
-                  uv pip install --python .venv-mcp-peer 'mcp==1.28.1' 'langchain-mcp-adapters==0.2.1'
+                  uv pip install --no-config --exclude-newer "3 days" --python .venv-mcp-peer 'mcp==1.28.1' 'langchain-mcp-adapters==0.2.1'
                   echo "export MCP_TEST_PEER_PYTHON=$PWD/.venv-mcp-peer/bin/python" >> "$BASH_ENV"
+      - when:
+          condition:
+            equal: [sdk, << parameters.suite >>]
+          steps:
+            - run:
+                name: Install redis-server 7.2.16
+                command: |
+                  mkdir -p /tmp/redis-bin
+                  cid="$(docker create redis:7.2.16@sha256:0637954999d01b7c9ce9167db2da50656e2590d3b884f1c600c5f63bb6e6773c)"
+                  docker cp "${cid}:/usr/local/bin/redis-server" /tmp/redis-bin/redis-server
+                  docker rm "$cid"
+                  /tmp/redis-bin/redis-server --version | grep -q 'v=7.2.16'
+                  echo 'export PATH="/tmp/redis-bin:$PATH"' >> "$BASH_ENV"
       - start_postgres:
           image: postgres:16@sha256:e17e86066e5ef83e0952a9347f5c792b7ece00972e2aa787a6986f471b3dd3d5
           server_args: "-c max_connections=300 -c shared_preload_libraries=pg_stat_statements -c pg_stat_statements.track=all -c pg_stat_statements.max=20000"
@@ -1358,7 +1385,7 @@ jobs:
     environment:
       UV_PYTHON_DOWNLOADS: never
     steps:
-      - checkout
+      - checkout_over_https
       - install_uv
       - run:
           name: Install the locked shard verifier and coverage dependencies
@@ -1366,7 +1393,6 @@ jobs:
             uv export --frozen --all-groups --all-extras --no-emit-project --no-emit-workspace --no-header --no-annotate \
               | awk '/^[^ ]/ { keep = $0 ~ /^(annotated-types|coverage|pydantic|pydantic-core|typing-extensions|typing-inspection)==/ } keep' \
               > /tmp/fan-in-requirements.txt
-            test "$(grep -c '^[^ ]' /tmp/fan-in-requirements.txt)" -eq 6
             uv venv .venv --python 3.12
             uv pip install --no-config --python .venv --require-hashes -r /tmp/fan-in-requirements.txt
       - attach_workspace:
@@ -1407,7 +1433,7 @@ jobs:
             ROUTING_PARITY_BASE: << pipeline.parameters.routing_parity_base >>
           command: |
             [[ "$ROUTING_PARITY_BASE" =~ ^[0-9a-f]{40}$ ]] || exit 1
-            git fetch --depth 1 origin "$ROUTING_PARITY_BASE"
+            git fetch --depth 1 --no-filter origin "$ROUTING_PARITY_BASE"
             git rm -r -f --quiet litellm enterprise litellm-proxy-extras
             git checkout "$ROUTING_PARITY_BASE" -- litellm enterprise litellm-proxy-extras
             git reset --quiet
@@ -1482,7 +1508,7 @@ jobs:
     environment:
       DATABASE_URL: postgresql://postgres:postgres@localhost:5432/litellm_test
     steps:
-      - checkout
+      - checkout_over_https
       - skip_if_unrelated_changes
       - setup_litellm_test_deps:
           checkout: false

.circleci/scripts/classify_changes.sh

@@ -33,7 +33,7 @@ while IFS= read -r file || [ -n "$file" ]; do
     *) has_backend=true ;;
   esac
   case "$file" in
-    model_prices_and_context_window.json | litellm/model_prices_and_context_window_backup.json | model_prices_and_context_window.schema.json)
+    model_prices_and_context_window.json | litellm/model_prices_and_context_window_backup.json)
       has_cost_map=true ;;
     tests/unit/proxy/*) : ;;
     *) outside_cost_map_set=true ;;

.circleci/scripts/run_integration.sh

@@ -1,14 +1,14 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
-if [ "${GITHUB_ACTIONS:-}" = true ]; then
-  echo "Integration contracts are owned by CircleCI" >&2
-  exit 1
-fi
-
 suite="${1:?integration suite required}"
 mode="${2:-standard}"
 side="${3:-}"
+requested_workers="${INTEGRATION_WORKERS:-}"
+license_env=()
+if [ -n "${LITELLM_LICENSE:-}" ]; then
+  license_env=("LITELLM_LICENSE=$LITELLM_LICENSE")
+fi
 if [ "$mode" = replica ]; then
   results="test-results/integration-${suite}-replica"
 elif [ "$mode" = parity ]; then
@@ -84,7 +84,7 @@ if [ "$suite" = browser ]; then
   export LITELLM_UI_PATH="$PWD/ui/litellm-dashboard/out"
   test -f "$LITELLM_UI_PATH/index.html"
 fi
-export INTEGRATION_SEED="$(.venv/bin/python -c 'import hashlib,os; print(int(hashlib.sha256((os.environ.get("CIRCLE_SHA1", "local") + os.environ.get("CIRCLE_WORKFLOW_ID", "local")).encode()).hexdigest()[:8],16))')"
+export INTEGRATION_SEED="$(.venv/bin/python -c 'import hashlib,os; print(int(hashlib.sha256(((os.environ.get("CIRCLE_SHA1") or os.environ.get("GITHUB_SHA") or "local") + (os.environ.get("CIRCLE_WORKFLOW_ID") or os.environ.get("GITHUB_RUN_ID") or "local")).encode()).hexdigest()[:8],16))')"
 export INTEGRATION_ORDER_SEED="$INTEGRATION_SEED"
 
 uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma > "$results/prisma-generate.log" 2>&1
@@ -107,7 +107,7 @@ sudo iptables -N integration_only
 guard_created=true
 sudo iptables -A integration_only -o lo -j ACCEPT
 sudo iptables -A integration_only -m conntrack --ctstate ESTABLISHED,RELATED -j ACCEPT
-for service in postgres-db redis-cache; do
+for service in "${INTEGRATION_POSTGRES_CONTAINER:-postgres-db}" "${INTEGRATION_REDIS_CONTAINER:-redis-cache}"; do
   address="$(docker inspect --format '{{range .NetworkSettings.Networks}}{{.IPAddress}}{{end}}' "$service")"
   sudo iptables -A integration_only -d "$address" -j ACCEPT
 done
@@ -128,6 +128,8 @@ fi
 sudo iptables -L integration_only -n -v -x > "$results/egress-guard.txt"
 awk '$3 == "REJECT" && $1 > 0 { rejected=1 } END { exit !rejected }' "$results/egress-guard.txt"
 
+export UV_OFFLINE=true
+
 if [ "$mode" = parity ]; then
   export INTEGRATION_WORKERS=1
 else
@@ -145,6 +147,9 @@ fi
 if [ "$suite" = mcp ]; then
   export INTEGRATION_WORKERS=2 INTEGRATION_COVERAGE=1
 fi
+if [ -n "$requested_workers" ]; then
+  export INTEGRATION_WORKERS="$requested_workers"
+fi
 isolated_workers=false
 if [ "${INTEGRATION_WORKERS:-1}" -gt 1 ]; then
   isolated_workers=true
@@ -185,7 +190,7 @@ start_proxy() {
     "${database_env[@]}" REDIS_HOST="$REDIS_HOST" REDIS_PORT="$REDIS_PORT" \
     INTEGRATION_UPSTREAM_URL="$INTEGRATION_UPSTREAM_URL" \
     LITELLM_MASTER_KEY="$LITELLM_MASTER_KEY" LITELLM_SALT_KEY="$LITELLM_SALT_KEY" LITELLM_UI_PATH="$LITELLM_UI_PATH" PROXY_BASE_URL="http://127.0.0.1:$port" \
-    LITELLM_LICENSE="${LITELLM_LICENSE:-}" \
+    "${license_env[@]}" \
     LITELLM_MODE=PRODUCTION STORE_MODEL_IN_DB=True LITELLM_ENABLE_MCP_STDIO=true "${cost_map_env[@]}" \
     AWS_EC2_METADATA_DISABLED=true DO_NOT_TRACK=1 COVERAGE_FILE="$coverage_data" \
     "${proxy_command[@]}" --config tests/integration/proxy_config.yaml \
@@ -257,7 +262,7 @@ env -i PATH="$PATH" HOME="$HOME" PYTHONPATH="$PYTHONPATH" \
   "${shared_services[@]}" "${worker_coverage[@]}" \
   INTEGRATION_WORKERS="${INTEGRATION_WORKERS:-1}" \
   LITELLM_MODE=PRODUCTION \
-  LITELLM_LICENSE="${LITELLM_LICENSE:-}" \
+  "${license_env[@]}" \
   INTEGRATION_SEED="$INTEGRATION_SEED" \
   INTEGRATION_ORDER_SEED="$INTEGRATION_ORDER_SEED" \
   INTEGRATION_CASE_SHARDS="${CIRCLE_NODE_TOTAL:-1}" INTEGRATION_CASE_SHARD_INDEX="${CIRCLE_NODE_INDEX:-0}" \

.githooks/pre-push

@@ -1,90 +0,0 @@
-#!/usr/bin/env bash
-#
-# pre-push — enforce Conventional Branches
-#   https://conventional-branch.github.io/
-#
-# Branch format:  <type>/<description>
-#   <type> must be one of: feature, bugfix, hotfix, release, chore
-#
-# Protected branches (always allowed):
-#   - main
-#   - dependabot/*
-#   - gh-readonly-queue/*
-#
-# Tag pushes and branch deletions are skipped.
-# Bypass: git push --no-verify.
-
-set -eu
-
-ZERO_OID="0000000000000000000000000000000000000000"
-ZERO_OID_SHA256="0000000000000000000000000000000000000000000000000000000000000000"
-ALLOWED_TYPES="feature|bugfix|hotfix|release|chore"
-BRANCH_PATTERN="^(${ALLOWED_TYPES})/.+"
-
-PROTECTED_NAMES="main"
-PROTECTED_PREFIXES="dependabot/ gh-readonly-queue/"
-
-is_protected() {
-    branch="$1"
-    for name in $PROTECTED_NAMES; do
-        if [ "$branch" = "$name" ]; then
-            return 0
-        fi
-    done
-    for prefix in $PROTECTED_PREFIXES; do
-        case "$branch" in "$prefix"*) return 0 ;; esac
-    done
-    return 1
-}
-
-invalid=""
-
-while read -r local_ref local_oid remote_ref remote_oid; do
-    # Branch deletion (no local commit being pushed).
-    if [ "$local_oid" = "$ZERO_OID" ] || [ "$local_oid" = "$ZERO_OID_SHA256" ]; then
-        continue
-    fi
-
-    # Only validate branch pushes; ignore tags and other ref namespaces.
-    case "$remote_ref" in
-        refs/heads/*) ;;
-        *) continue ;;
-    esac
-
-    branch="${remote_ref#refs/heads/}"
-
-    if is_protected "$branch"; then
-        continue
-    fi
-
-    if ! printf '%s' "$branch" | grep -Eq "$BRANCH_PATTERN"; then
-        invalid="$invalid $branch"
-    fi
-done
-
-if [ -n "$invalid" ]; then
-    cat >&2 <<EOF
-✗ Branch name does not follow Conventional Branches.
-
-  Invalid:$invalid
-
-  Expected: <type>/<description>
-
-  Allowed types: feature, bugfix, hotfix, release, chore
-  Examples:
-    feature/weighted-round-robin
-    bugfix/streaming-empty-chunks
-    chore/bump-deps
-    hotfix/auth-bypass
-
-  Protected (always allowed): main, dependabot/*, gh-readonly-queue/*.
-
-See https://conventional-branch.github.io/
-
-Rename with: git branch -m <new-name>
-To bypass (use sparingly): git push --no-verify
-EOF
-    exit 1
-fi
-
-exit 0

.github/CODEOWNERS

@@ -1,2 +1,6 @@
-/model_prices_and_context_window.json @mateo-berri @ryan-crabbe-berri @kerry-berri
-/litellm/model_prices_and_context_window_backup.json @mateo-berri @ryan-crabbe-berri @kerry-berri
+/model_prices_and_context_window.json @mateo-berri @ryan-crabbe-berri @kerry-berri @joshua-berri
+/litellm/model_prices_and_context_window_backup.json @mateo-berri @ryan-crabbe-berri @kerry-berri @joshua-berri
+/ci_cd/cost_map_pins.json @mateo-berri @ryan-crabbe-berri @kerry-berri @joshua-berri
+/helm/ @yassin-berriai
+/terraform/ @yassin-berriai
+/.github/workflows/zizmor.yml @yuneng-berri

.github/PULL_REQUEST_TEMPLATE/general.md

@@ -4,6 +4,25 @@
      Drop every section you have nothing to put in, heading included: a bare "## Relevant issues" or
      "## Affected release" with nothing under it must not appear in the final description -->
 
+## Handwritten 1-3 sentence description
+
+<!-- The author writes a plain summary in their own words before requesting review.
+     Wrap newly AI-written summary text in the make-me-human-readable tags below, including backticks.
+     The author reviews or rewrites the draft and removes the tags.
+     When updating an existing PR, preserve the author's wording and removed tags.
+     Changes elsewhere in the body do not make an unchanged summary an AI draft again.
+     Add tags again only around newly AI-written replacement summary text, never merely to match this template.
+     This preserves author review across later updates (2026-11-02, #49779).
+     Say what the PR does for users and why, then anything that keeps it working over time (a sync job, a config someone owns)
+     No bullets, tables, or code names here. Everything else goes inside the Details block below -->
+
+`<make-me-human-readable>`
+<blah>
+`</make-me-human-readable>`
+
+<details>
+<summary>Details</summary>
+
 ## TLDR
 
 <!-- Fill in the bullets below and keep each one short and concrete: one line per bullet, roughly 10 words max
@@ -77,9 +96,10 @@ If you're seeing a delay in your PR being merged, ping the LiteLLM Team on [Slac
 ## Screenshots / Proof of Fix
 
 <!-- Include screenshots, screen recordings, or command (e.g., curl) + output demonstrating that your changes work as expected
-     The proof must be completely e2e with no mocks, using actual LLM calls costing real $$$ if applicable. `pytest` commands are not enough
-     Show ONLY the latest run: capture Before at the merge base and After at the PR's current tip, and when new commits change behavior, replace this whole section with the fresh run instead of stacking it on top of older ones. The run must be up to date. As soon as a new commit is made and it makes this PR description's after sha stale (it's no longer tip of PR), you must re-run the QA
-     Structure the section exactly as below: Before and After one heading level below this section, each naming the commit hash it was captured at, one lower-level heading per case inside each, the same case names in the same order on both sides, and numbered steps (command, observed output) under every case, never loose prose; shared setup (config, payloads) goes above Before, and with a single case, drop the case headings and number the steps directly
+     Simple PR exception (2026-10-28, #49030). A mechanical change fully proven by focused checks may skip before/after e2e. Follow AGENTS.md's eligibility rule, trace callers, and verify bug-fix regression sensitivity. Record E2E exception at <tip SHA>, scope, rationale, commands and results here. Tests/static checks suffice only for this exception. Keep CI and bot gates, reassess changed diffs, and honor explicit e2e requests. Omit the Before/After scaffold for an exception
+     Otherwise proof must be fully e2e, no mocks, with real provider calls where applicable. Pytest alone is insufficient. The following format applies to live proof
+     Show only the latest proof. Before uses the merge base and After the tested tip. Rerun affected QA after behavior changes. Verified neutral diffs may carry proof with the tested SHA, current tip, equivalence reason and focused checks, following AGENTS.md and /qa
+     Use matching Before/After case headings with numbered commands and observed output. Put shared setup above them. For one case, omit case headings
 
 ### Before (<hash>)
 
@@ -164,3 +184,5 @@ Example checklists:
 ## Final Attestation
 
 - [ ] The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR
+
+</details>

.github/ci-coverage-allowlist.yml

@@ -5,7 +5,7 @@ description: >-
 
 test_paths:
   - reason: >-
-      litellm.agent() end-to-end suite. It drives the real claude, codex and opencode CLIs and
+      litellm.agent() end-to-end suite. It drives the real claude, codex, opencode and pi CLIs and
       deepagents against a live LiteLLM AI Gateway, so it needs those binaries on PATH plus
       LITELLM_PROXY_API_BASE / LITELLM_PROXY_API_KEY, and skips without them. Run manually
       before changing litellm/harness; the mocked coverage runs in tests/unit/harness and

.github/e2e-stack/assert_tests_ran.py

@@ -1,69 +0,0 @@
-import os
-import re
-import sys
-import xml.etree.ElementTree as ET
-from pathlib import Path
-from typing import Final
-
-sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "tests/e2e"))
-from coverage_registry.management_cases import MANAGEMENT_CASES
-
-
-def main() -> int:
-    selected: Final = tuple(sys.argv[2:])
-    try:
-        report: Final = ET.parse(Path(sys.argv[1])).getroot()
-    except (ET.ParseError, OSError):
-        _ = sys.stdout.write("::error::could not read the test execution report\n")
-        return 1
-    cases: Final = tuple(report.iter("testcase"))
-    expected_count: Final = os.environ.get("E2E_REQUIRED_TEST_COUNT")
-    passed: Final = frozenset(
-        case.get("file") for case in cases if all(case.find(tag) is None for tag in ("skipped", "failure", "error"))
-    )
-    missing: Final = tuple(path for path in selected if path not in passed)
-    required_nodes: Final = frozenset(case.node for case in MANAGEMENT_CASES if case.node.split("::", 1)[0] in selected)
-    passed_nodes: Final = frozenset(
-        prop.get("value")
-        for case in cases
-        if all(case.find(tag) is None for tag in ("skipped", "failure", "error"))
-        for prop in case.findall("./properties/property")
-        if prop.get("name") == "management_node"
-    )
-    missing_nodes: Final = required_nodes - passed_nodes
-    for node in sorted(missing_nodes):
-        _ = sys.stdout.write(f"::error::required management case did not pass: {node}\n")
-    for path in selected:
-        collected: Final = sum(case.get("file") == path for case in cases)
-        skipped: Final = sum(case.get("file") == path and case.find("skipped") is not None for case in cases)
-        _ = sys.stdout.write(f"{path}: {collected} collected, {skipped} skipped\n")
-        for case in cases:
-            if case.get("file") != path or all(case.find(tag) is None for tag in ("failure", "error", "skipped")):
-                continue
-            outcome = "skipped" if case.find("skipped") is not None else "failed"
-            _ = sys.stdout.write(f"  {outcome}: {case.get('classname', '')}::{case.get('name', '')}\n")
-            for prop in case.findall("./properties/property"):
-                name = prop.get("name", "")
-                value = prop.get("value", "")
-                if name in ("oauth_failure_phase", "oauth_exception_type", "oauth_frame") and re.fullmatch(
-                    r"[A-Za-z0-9_.:<>-]{1,240}", value
-                ):
-                    _ = sys.stdout.write(f"    {name}: {value}\n")
-    if expected_count is not None and (
-        len(cases) != int(expected_count) or any(case.find("skipped") is not None for case in cases)
-    ):
-        _ = sys.stdout.write("::error::required test count was not met or a required case was skipped\n")
-        return 1
-    if (
-        selected
-        and not missing
-        and not missing_nodes
-        and not any(case.find(tag) is not None for case in cases for tag in ("failure", "error"))
-    ):
-        return 0
-    _ = sys.stdout.write("::error::every selected file must execute a passing test, with no failures or errors\n")
-    return 1
-
-
-if __name__ == "__main__":
-    sys.exit(main())

.github/scripts/assert_ci_coverage.py

@@ -32,6 +32,7 @@ IGNORE_ARG_RE: Final = re.compile(r"--ignore(?:-glob)?[= ](\S+)")
 DOCKERFILE_TOKEN_RE = re.compile(r"[A-Za-z0-9_./-]*Dockerfile[A-Za-z0-9_.-]*")
 COMMENT_RE = re.compile(r"^\s*#.*$", re.MULTILINE)
 GLOB_CHARS = frozenset("*?")
+DEDICATED_INTEGRATION_WORKFLOWS: Final = MappingProxyType({"image-scan.yml": "tests/integration/database/image/"})
 
 # Trees whose jobs are sharded with no catch-all bucket, so every child that holds
 # tests has to be named by some shard or it runs nowhere. A child listed here is
@@ -560,6 +561,17 @@ def _integration_github_files(runner: pathlib.Path) -> frozenset[str]:
     return frozenset(ast.literal_eval(values))
 
 
+def _owns(workflow: str, relative_path: str) -> bool:
+    owned_root: Final = DEDICATED_INTEGRATION_WORKFLOWS.get(workflow)
+    return owned_root is not None and relative_path.startswith(owned_root)
+
+
+def _tokens_outside_owner(relative_path: str, workflow_tokens: Mapping[str, frozenset[str]]) -> frozenset[str]:
+    return frozenset[str]().union(
+        *(tokens for workflow, tokens in workflow_tokens.items() if not _owns(workflow, relative_path))
+    )
+
+
 def _integration_ownership(repo_root: pathlib.Path = REPO_ROOT) -> tuple[frozenset[str], tuple[Finding, ...]]:
     runner: Final = repo_root / "tests/integration/run.py"
     if not runner.exists():
@@ -611,15 +623,20 @@ def _integration_ownership(repo_root: pathlib.Path = REPO_ROOT) -> tuple[frozens
         )
         != 1
     )
-    gha_tokens: Final = _invoked_test_tokens(
-        scalar
-        for path in (repo_root / ".github/workflows").glob("*.y*ml")
-        for scalar in _scalars(yaml.safe_load(path.read_text()), path.name)
+    workflow_tokens: Final = MappingProxyType(
+        {
+            path.name: _invoked_test_tokens(_scalars(yaml.safe_load(path.read_text()), path.name))
+            for path in (repo_root / ".github/workflows").glob("*.y*ml")
+        }
+    )
+    gha_tokens: Final = frozenset[str]().union(*workflow_tokens.values())
+    circle_owned: Final = (
+        frozenset(str(path.relative_to(repo_root)) for path in integration_root.rglob("test_*.py")) - github_files
     )
     findings: Final = (
         tuple(
-            Finding(path, "integration contract is also selected by GitHub Actions")
-            for path in paths
+            Finding(path, "integration file outside GITHUB_FILES is selected by GitHub Actions")
+            for path in sorted(circle_owned)
             if any(_token_covers(token, path) for token in gha_tokens)
         )
         + tuple(
@@ -632,6 +649,11 @@ def _integration_ownership(repo_root: pathlib.Path = REPO_ROOT) -> tuple[frozens
             for path in sorted(github_files)
             if not (repo_root / path).is_file()
         )
+        + tuple(
+            Finding(path, "GitHub-owned integration contract runs outside a dedicated integration workflow")
+            for path in sorted(github_files)
+            if any(_token_covers(token, path) for token in _tokens_outside_owner(path, workflow_tokens))
+        )
     )
     browser_commands: Final = tuple(
         scalar.value
@@ -688,13 +710,9 @@ def main() -> int:
     scalars = _all_scalars()
 
     integration_paths, ownership_findings = _integration_ownership()
-    test_findings = (
-        _uncovered_tests(
-            allowlist,
-            _invoked_selections(scalars)
-            + (Selection(included=integration_paths, ignored=frozenset()),),
-        )
-        + ownership_findings
+    test_findings = _uncovered_tests(
+        allowlist,
+        _invoked_selections(scalars) + (Selection(included=integration_paths, ignored=frozenset()),),
     )
     dockerfile_findings = _uncovered_dockerfiles(allowlist, _built_dockerfile_tokens(scalars))
     stale_findings = _stale_allowlist_paths(allowlist, test_files=_test_files(), dockerfiles=_dockerfiles())
@@ -711,13 +729,25 @@ def main() -> int:
             test_findings,
             "Add each to a job's test path, or list it in .github/ci-coverage-allowlist.yml with a reason.",
         )
+    if ownership_findings:
+        browser_remedy = (
+            " CircleCI's browser group owns the integrationCritical Playwright contracts, so no GitHub workflow may "
+            "select them and the default Playwright config must exclude them."
+            if any("browser" in finding.detail or "Playwright" in finding.detail for finding in ownership_findings)
+            else ""
+        )
+        _report(
+            "integration contracts with the wrong CI owner",
+            ownership_findings,
+            "Follow tests/integration/AGENTS.md, which says which CI job owns each integration file." + browser_remedy,
+        )
     if dockerfile_findings:
         _report(
             "Dockerfiles that no CI job builds",
             dockerfile_findings,
             "Build each in a workflow, or list it in .github/ci-coverage-allowlist.yml with a reason.",
         )
-    if stale_findings or test_findings or dockerfile_findings:
+    if stale_findings or test_findings or ownership_findings or dockerfile_findings:
         return 1
 
     _write(

.github/scripts/assert_secret_absent.py

@@ -0,0 +1,119 @@
+import base64
+import io
+import os
+import sys
+import urllib.parse
+import zipfile
+from collections.abc import Callable, Iterator, Mapping
+from dataclasses import dataclass
+from functools import reduce
+from itertools import chain
+from pathlib import Path
+from types import MappingProxyType
+from typing import Final, TextIO
+
+SECRET_VARIABLE: Final = "SCANNED_SECRET"
+MIN_ENCODED_LENGTH: Final = 12
+MAX_ARCHIVE_DEPTH: Final = 3
+MASK: Final = "***"
+
+Forms = tuple[tuple[str, str], ...]
+
+
+@dataclass(frozen=True, slots=True)
+class Hit:
+    location: str
+    form: str
+
+
+def _aligned_encodings(raw: bytes, name: str, encode: Callable[[bytes], bytes]) -> Iterator[tuple[str, str]]:
+    for offset in range(3):
+        chunk: Final = raw[offset:]
+        yield f"{name} offset {offset}", encode(chunk[: len(chunk) - len(chunk) % 3]).decode()
+
+
+def secret_forms(secret: str) -> Forms:
+    encoded: Final = (
+        ("url-encoded", urllib.parse.quote(secret, safe="")),
+        *_aligned_encodings(secret.encode(), "base64", base64.b64encode),
+        *_aligned_encodings(secret.encode(), "base64url", base64.urlsafe_b64encode),
+    )
+    by_value: Final = MappingProxyType(
+        {value: name for name, value in encoded if value != secret and len(value) >= MIN_ENCODED_LENGTH}
+    )
+    return (("raw", secret), *((name, value) for value, name in by_value.items()))
+
+
+def mask(text: str, forms: Forms) -> str:
+    return reduce(lambda masked, form: masked.replace(form[1], MASK), forms, text)
+
+
+def _name_hits(location: str, name: str, forms: Forms) -> Iterator[Hit]:
+    for form, value in forms:
+        if value in name:
+            yield Hit(mask(location, forms), f"{form} in name")
+
+
+def _content_hits(location: str, data: bytes, forms: Forms, depth: int) -> Iterator[Hit]:
+    for form, value in forms:
+        if value.encode() in data:
+            yield Hit(mask(location, forms), form)
+    if depth >= MAX_ARCHIVE_DEPTH or not zipfile.is_zipfile(io.BytesIO(data)):
+        return
+    with zipfile.ZipFile(io.BytesIO(data)) as archive:
+        for member in archive.infolist():
+            if member.is_dir():
+                continue
+            yield from _content_hits(f"{location}!{member.filename}", archive.read(member), forms, depth + 1)
+
+
+def _walk(root: Path) -> Iterator[tuple[Path, bool]]:
+    for directory, subdirectories, files in os.walk(root):
+        yield from ((Path(directory, name), True) for name in subdirectories if Path(directory, name).is_symlink())
+        yield from ((Path(directory, name), False) for name in files)
+
+
+def _path_hits(root: Path, path: Path, linked_directory: bool, forms: Forms) -> Iterator[Hit]:
+    location: Final = str(path.relative_to(root))
+    if linked_directory:
+        yield Hit(mask(location, forms), "directory symlink, not scanned")
+        return
+    yield from _name_hits(location, location, forms)
+    yield from _content_hits(location, path.read_bytes(), forms, 0)
+
+
+def scan(root: Path, secret: str) -> tuple[int, tuple[Hit, ...]]:
+    forms: Final = secret_forms(secret)
+    entries: Final = tuple(_walk(root))
+    hits: Final = tuple(chain.from_iterable(_path_hits(root, path, linked, forms) for path, linked in entries))
+    return len(entries), hits
+
+
+def main(argv: tuple[str, ...], environ: Mapping[str, str], out: TextIO) -> int:
+    if len(argv) != 1:
+        print(f"usage: assert_secret_absent.py <directory>, with the secret in ${SECRET_VARIABLE}", file=out)
+        return 2
+    secret: Final = environ.get(SECRET_VARIABLE, "")
+    root: Final = Path(argv[0])
+    if not secret:
+        print(f"{SECRET_VARIABLE} is empty, nothing to scan for", file=out)
+        return 0
+    if not root.is_dir():
+        print(f"{root} is not a directory, nothing to scan", file=out)
+        return 0
+    try:
+        scanned, hits = scan(root, secret)
+    except Exception as error:
+        print(f"scan failed with {type(error).__name__}, treating the results as unsafe", file=out)
+        return 1
+    for hit in hits:
+        print(f"secret found: {hit.location} ({hit.form})", file=out)
+    if hits:
+        print(f"{len(hits)} copies of the secret under {root}, refusing to upload", file=out)
+        return 1
+    print(f"no copy of the secret in {scanned} files under {root}", file=out)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(tuple(sys.argv[1:]), os.environ, sys.stdout))

.github/scripts/auto_update_price_and_context_window_file.py

@@ -1,319 +0,0 @@
-import asyncio
-import aiohttp
-import json
-import math
-from typing import Any
-
-# Asynchronously fetch data from a given URL
-async def fetch_data(url):
-    try:
-        # Create an asynchronous session
-        async with aiohttp.ClientSession() as session:
-            # Send a GET request to the URL
-            async with session.get(url) as resp:
-                # Raise an error if the response status is not OK
-                resp.raise_for_status()
-                # Parse the response JSON
-                resp_json = await resp.json()
-                print("Fetch the data from URL.")
-                # Return the 'data' field from the JSON response
-                return resp_json['data']
-    except Exception as e:
-        # Print an error message if fetching data fails
-        print("Error fetching data from URL:", e)
-        return None
-
-
-FRIENDLI_API_URL = "https://api.friendli.ai/serverless/v1/models"
-FRIENDLI_PROVIDER = "friendliai"
-
-INHERITABLE_BASE_KEYS = (
-    "supports_pdf_input",
-    "supports_assistant_prefill",
-    "supports_adaptive_thinking",
-    "supports_output_config",
-)
-
-REASONING_EFFORT_LEVEL_ORDER = ("none", "minimal", "low", "medium", "high", "xhigh", "max")
-
-
-def _find_base_model_entry(base_model: str, local_data: dict) -> str | None:
-    if not base_model:
-        return None
-    bm_tail = base_model.split("/")[-1].lower()
-    if base_model in local_data:
-        return base_model
-    for key in local_data:
-        if key.startswith("sample_spec") or key == "fallback_generalizations":
-            continue
-        if key.split("/")[-1].lower() == bm_tail:
-            return key
-    return None
-
-
-def _reasoning_effort_levels(reasoning_options: list) -> list:
-    offered = {
-        val
-        for opt in reasoning_options or []
-        if opt.get("type") == "effort"
-        for val in opt.get("values", [])
-    }
-    return [level for level in REASONING_EFFORT_LEVEL_ORDER if level in offered]
-
-
-def _valid_token_price(value: object) -> bool:
-    try:
-        price = float(value)  # pyright: ignore[reportArgumentType]  # non-numeric values are rejected via the except
-    except (TypeError, ValueError):
-        return False
-    return math.isfinite(price) and price >= 0
-
-
-def _has_valid_token_prices(pricing: dict | None) -> bool:
-    prices = pricing or {}
-    return _valid_token_price(prices.get("input")) and _valid_token_price(prices.get("output"))
-
-
-def _pricing(pricing: dict) -> dict:
-    out: dict[str, Any] = {}
-    if not pricing:
-        return out
-    if "input" in pricing:
-        out["input_cost_per_token"] = float(pricing["input"])
-    if "output" in pricing:
-        out["output_cost_per_token"] = float(pricing["output"])
-    if "input_cache_read" in pricing and pricing["input_cache_read"] is not None:
-        out["cache_read_input_token_cost"] = float(pricing["input_cache_read"])
-    return out
-
-
-def _modality_flags(input_mods: list) -> dict:
-    mods = input_mods or []
-    has_image = "image" in mods
-    return {
-        "supports_vision": has_image,
-        "supports_image_input": has_image,
-        "supports_video_input": "video" in mods,
-    }
-
-
-def transform_friendli_data(data: list, local_data: dict) -> dict:
-    transformed: dict[str, dict] = {}
-    if not data:
-        return transformed
-    for model in data:
-        # An unpriced row must never wholesale-replace an already priced local entry:
-        # missing prices cost-calculate as zero, silently zeroing tracked spend
-        if not _has_valid_token_prices(model.get("pricing")):
-            continue
-        model_id = model["id"]
-        base_model = model.get("base_model") or ""
-        entry: dict[str, Any] = {
-            "litellm_provider": FRIENDLI_PROVIDER,
-        }
-
-        base_key = _find_base_model_entry(base_model, local_data)
-        if base_key:
-            base_entry = local_data[base_key]
-            for k in INHERITABLE_BASE_KEYS:
-                if k in base_entry:
-                    entry[k] = base_entry[k]
-
-        ctx = model.get("context_length")
-        if ctx is not None:
-            entry["max_input_tokens"] = int(ctx)
-        max_out = model.get("max_completion_tokens")
-        if max_out is not None:
-            entry["max_output_tokens"] = int(max_out)
-            entry["max_tokens"] = int(max_out)
-
-        pricing = _pricing(model.get("pricing", {}))
-        entry.update(pricing)
-        entry["supports_prompt_caching"] = "cache_read_input_token_cost" in pricing
-
-        reasoning = model.get("reasoning") is True
-        entry["supports_reasoning"] = reasoning
-        if reasoning:
-            entry["reasoning_effort_levels"] = _reasoning_effort_levels(
-                model.get("reasoning_options", [])
-            )
-
-        func = model.get("functionality", {})
-        entry["supports_function_calling"] = func.get("tool_call") is True
-        entry["supports_parallel_function_calling"] = func.get("parallel_tool_call") is True
-        is_struct = func.get("structured_output") is True
-        entry["supports_response_schema"] = is_struct
-        entry["supports_native_structured_output"] = is_struct
-        entry["supports_system_messages"] = func.get("system_messages") is True
-        entry["supports_tool_choice"] = func.get("tool_choice") is True
-
-        entry.update(_modality_flags(model.get("input_modalities", [])))
-
-        entry["mode"] = model.get("mode", "chat")
-
-        desc = model.get("description")
-        if desc:
-            entry["comment"] = desc
-
-        dep = model.get("deprecation_date")
-        if dep:
-            entry["deprecation_date"] = dep.split("T")[0]
-
-        entry["source"] = FRIENDLI_API_URL
-
-        transformed[f"{FRIENDLI_PROVIDER}/{model_id}"] = entry
-    return transformed
-
-# Synchronize local data with remote data
-def sync_local_data_with_remote(local_data, remote_data, replace_keys=frozenset()):
-    # Update existing keys in local_data with values from remote_data
-    # (replace_keys entries are swapped wholesale so a field the remote catalog
-    #  dropped, e.g. cache pricing, cannot survive as a stale value)
-    for key in (set(local_data) & set(remote_data)):
-        if key in replace_keys:
-            local_data[key] = remote_data[key]
-        else:
-            local_data[key].update(remote_data[key])
-
-    # Add new keys from remote_data to local_data
-    for key in (set(remote_data) - set(local_data)):
-        local_data[key] = remote_data[key]
-
-# Write data to the json file
-def write_to_file(file_path, data):
-    try:
-        # Open the file in write mode
-        with open(file_path, "w") as file:
-            # Dump the data as JSON into the file
-            json.dump(data, file, indent=4)
-        print("Values updated successfully.")
-    except Exception as e:
-        # Print an error message if writing to file fails
-        print("Error updating JSON file:", e)
-
-# Update the existing models and add the missing models for OpenRouter
-def transform_openrouter_data(data):
-    transformed = {}
-    if not data:
-        return transformed
-    for row in data:
-        # Add the fields 'max_tokens' and 'input_cost_per_token'
-        obj = {
-            "max_tokens": row["context_length"],
-            "input_cost_per_token": float(row["pricing"]["prompt"]),
-        }
-
-        # Add 'max_output_tokens' as a field if it is not None
-        if "top_provider" in row and "max_completion_tokens" in row["top_provider"] and row["top_provider"]["max_completion_tokens"] is not None:
-            obj['max_output_tokens'] = int(row["top_provider"]["max_completion_tokens"])
-
-        # Add the field 'output_cost_per_token'
-        obj.update({
-            "output_cost_per_token": float(row["pricing"]["completion"]),
-        })
-
-        # Add field 'input_cost_per_image' if it exists and is non-zero
-        if "pricing" in row and "image" in row["pricing"] and float(row["pricing"]["image"]) != 0.0:
-            obj['input_cost_per_image'] = float(row["pricing"]["image"])
-
-        # Add the fields 'litellm_provider' and 'mode'
-        obj.update({
-            "litellm_provider": "openrouter",
-            "mode": "chat"
-        })
-
-        # Add the 'supports_vision' field if the modality is 'multimodal'
-        if row.get('architecture', {}).get('modality') == 'multimodal':
-            obj['supports_vision'] = True
-
-        # Use a composite key to store the transformed object
-        transformed[f'openrouter/{row["id"]}'] = obj
-
-    return transformed
-
-# Update the existing models and add the missing models for Vercel AI Gateway
-def transform_vercel_ai_gateway_data(data):
-    transformed = {}
-    if not data:
-        return transformed
-    for row in data:
-        # Rows without token pricing or token limits (video/embedding models) previously KeyError'd the whole sync
-        if any(row.get(k) is None for k in ("context_window", "max_tokens")) or any(
-            row.get("pricing", {}).get(k) is None for k in ("input", "output")
-        ):
-            continue
-        obj = {
-            "max_tokens": row["context_window"],
-            "input_cost_per_token": float(row["pricing"]["input"]),
-            "output_cost_per_token": float(row["pricing"]["output"]),
-            'max_output_tokens': row['max_tokens'],
-            'max_input_tokens': row["context_window"],
-        }
-
-        # Handle cache pricing if available
-        if "pricing" in row:
-            if "input_cache_read" in row["pricing"] and row["pricing"]["input_cache_read"] is not None:
-                obj['cache_read_input_token_cost'] = float(f"{float(row['pricing']['input_cache_read']):e}")
-            
-            if "input_cache_write" in row["pricing"] and row["pricing"]["input_cache_write"] is not None:
-                obj['cache_creation_input_token_cost'] = float(f"{float(row['pricing']['input_cache_write']):e}")
-
-        mode = "embedding" if "embedding" in row["id"].lower() else "chat"
-        
-        obj.update({"litellm_provider": "vercel_ai_gateway", "mode": mode})
-
-        transformed[f'vercel_ai_gateway/{row["id"]}'] = obj
-
-    return transformed
-
-
-# Load local data from a specified file
-def load_local_data(file_path):
-    try:
-        # Open the file in read mode
-        with open(file_path, "r") as file:
-            # Load and return the JSON data
-            return json.load(file)
-    except FileNotFoundError:
-        # Print an error message if the file is not found
-        print("File not found:", file_path)
-        return None
-    except json.JSONDecodeError as e:
-        # Print an error message if JSON decoding fails
-        print("Error decoding JSON:", e)
-        return None
-
-def main():
-    local_file_path = "model_prices_and_context_window.json"  # Path to the local data file
-    openrouter_url = "https://openrouter.ai/api/v1/models"  # URL to fetch OpenRouter data
-    vercel_ai_gateway_url = "https://ai-gateway.vercel.sh/v1/models"  # URL to fetch Vercel AI Gateway data
-
-    # Load local data from file
-    local_data = load_local_data(local_file_path)
-    
-    # Fetch OpenRouter data
-    openrouter_data = asyncio.run(fetch_data(openrouter_url))
-    # Transform the fetched OpenRouter data
-    openrouter_data = transform_openrouter_data(openrouter_data)
-    
-    # Fetch Vercel AI Gateway data
-    vercel_data = asyncio.run(fetch_data(vercel_ai_gateway_url))
-    # Transform the fetched Vercel AI Gateway data
-    vercel_data = transform_vercel_ai_gateway_data(vercel_data)
-
-    friendli_data = asyncio.run(fetch_data(FRIENDLI_API_URL))
-    friendli_data = transform_friendli_data(friendli_data, local_data)
-    
-    # Combine both datasets
-    all_remote_data = {**openrouter_data, **vercel_data, **friendli_data}
-
-    # If both local and openrouter data are available, synchronize and save
-    if local_data and all_remote_data:
-        sync_local_data_with_remote(local_data, all_remote_data, replace_keys=frozenset(friendli_data))
-        write_to_file(local_file_path, local_data)
-    else:
-        print("Failed to fetch model data from either local file or URL.")
-
-# Entry point of the script
-if __name__ == "__main__":
-    main()

.github/scripts/smoke_test_native_wheel.py

@@ -7,6 +7,8 @@ import zipfile
 from pathlib import Path
 from typing import Final
 
+COST_MAP: Final = Path(__file__).resolve().parents[2] / "model_prices_and_context_window.json"
+
 CHILD_SCRIPT: Final = """
 from importlib.util import module_from_spec, spec_from_file_location
 from pathlib import Path
@@ -34,6 +36,9 @@ else:
 after = module.gil_stats()
 if not isinstance(after.get("releases"), int):
     raise AssertionError(f"native module unusable after panic: {after!r}")
+
+if module.bundled_model_cost_map() != Path(sys.argv[2]).read_bytes():
+    raise AssertionError("the wheel embeds a different model_prices_and_context_window.json than this checkout")
 """
 
 
@@ -55,7 +60,9 @@ def main() -> int:
 
         native_path: Final = Path(temporary_directory) / Path(native_members[0].filename).name
         native_path.write_bytes(archive.read(native_members[0]))
-        result: Final = subprocess.run((sys.executable, "-c", CHILD_SCRIPT, str(native_path)), check=False)
+        result: Final = subprocess.run(
+            (sys.executable, "-c", CHILD_SCRIPT, str(native_path), str(COST_MAP)), check=False
+        )
 
     if result.returncode != 0:
         sys.stderr.write(f"native wheel smoke test exited with status {result.returncode}\n")

.github/workflows/auto_update_price_and_context_window.yml

@@ -1,39 +0,0 @@
-name: Updates model_prices_and_context_window.json and Create Pull Request
-
-on:
-  schedule:
-    - cron: "0 0 * * 0" # Run every Sundays at midnight
-    #- cron: "0 0 * * *" # Run daily at midnight
-
-permissions:
-  contents: write
-  pull-requests: write
-
-jobs:
-  auto_update_price_and_context_window:
-    if: github.repository == 'BerriAI/litellm'
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          persist-credentials: false
-      - name: Set up uv
-        uses: ./.github/actions/setup-uv-with-retries
-        with:
-          version: "0.10.9"
-      - name: Update JSON Data
-        run: |
-          uv run --frozen --with 'aiohttp==3.13.3' python ".github/scripts/auto_update_price_and_context_window_file.py"
-      - name: Regenerate JSON Schema
-        run: |
-          uv run --frozen python ci_cd/generate_model_prices_schema.py
-      - name: Create Pull Request
-        run: |
-          git add model_prices_and_context_window.json model_prices_and_context_window.schema.json
-          git commit -m "Update model_prices_and_context_window.json file: $(date +'%Y-%m-%d')" 
-          gh pr create --title "Update model_prices_and_context_window.json file" \
-            --body "Automated update for model_prices_and_context_window.json" \
-            --head auto-update-price-and-context-window-$(date +'%Y-%m-%d') \
-            --base main
-        env:
-          GH_TOKEN: ${{ secrets.GH_TOKEN }}

.github/workflows/compat-matrix-image.yml

@@ -4,7 +4,7 @@ on:
   pull_request:
     paths:
       - tests/e2e/claude_code/cron_vm/**
-      - tests/e2e/claude_code/pr_gate_version_resolver.py
+      - tests/e2e/cli_harness/npm_version_resolver.py
       - .github/workflows/compat-matrix-image.yml
   workflow_dispatch:
 
@@ -36,7 +36,7 @@ jobs:
             whoami
             gh --version
             uv --version
-            version="$(uv run --no-project --python 3.12 python /opt/litellm/tests/e2e/claude_code/pr_gate_version_resolver.py)"
+            version="$(uv run --no-project --python 3.12 python /opt/litellm/tests/e2e/cli_harness/npm_version_resolver.py @anthropic-ai/claude-code)"
             /opt/litellm/tests/e2e/claude_code/cron_vm/install_claude_code.sh "${version}" /tmp/claude-cli
             /tmp/claude-cli/claude --version
           '

.github/workflows/cost-map-guard.yml

@@ -1,13 +1,14 @@
 name: Cost map guard
 
-on: # zizmor: ignore[dangerous-triggers] runs the base branch's code only; the PR's cost map files are read as data and never executed
-  pull_request_target:
+on:
+  pull_request:
     branches:
       - main
       - "litellm_**"
 
 permissions:
   contents: read
+  pull-requests: read
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
@@ -19,17 +20,23 @@ jobs:
     steps:
       - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
         with:
+          ref: ${{ github.event.pull_request.base.sha }}
           persist-credentials: false
-      - name: Fetch the pull request head and its merge base
+      - name: Check out the base branch tip and fetch the pull request head and its merge base
         id: revisions
         env:
           GH_TOKEN: ${{ github.token }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          EVENT_BASE_REF: ${{ github.event.pull_request.base.ref }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
-          merge_base="$(gh api "repos/${GITHUB_REPOSITORY}/compare/${BASE_SHA}...${HEAD_SHA}" --jq '.merge_base_commit.sha')"
-          git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA"
+          base_ref="$(gh api "repos/${GITHUB_REPOSITORY}/branches/${EVENT_BASE_REF}" --jq .name)" || base_ref="$(gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" --jq .base.ref)"
+          target="$(gh api "repos/${GITHUB_REPOSITORY}/branches/${base_ref}" --jq '.commit.sha')"
+          merge_base="$(gh api "repos/${GITHUB_REPOSITORY}/compare/${target}...${HEAD_SHA}" --jq '.merge_base_commit.sha')"
+          git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA" "$target"
+          git checkout -q --detach "$target"
           echo "merge_base=$merge_base" >> "$GITHUB_OUTPUT"
+          echo "target=$target" >> "$GITHUB_OUTPUT"
       - name: Set up uv
         uses: ./.github/actions/setup-uv-with-retries
         with:
@@ -37,7 +44,30 @@ jobs:
       - name: Run the guard
         env:
           MERGE_BASE: ${{ steps.revisions.outputs.merge_base }}
+          COST_MAP_GUARD_TARGET: ${{ steps.revisions.outputs.target }}
+          COST_MAP_GUARD_AUTHOR_TYPE: ${{ github.event.pull_request.user.type }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
           HEAD_REF: ${{ github.event.pull_request.head.ref }}
         run: |
-          uv run --frozen python ci_cd/cost_map_guard.py --base "$MERGE_BASE" --head "$HEAD_SHA" --head-ref "$HEAD_REF"
+          uv run --no-project python ci_cd/cost_map_guard.py --base "$MERGE_BASE" --head "$HEAD_SHA" --head-ref "$HEAD_REF"
+      - name: Check whether the pull request changes the cost map
+        id: changed
+        env:
+          MERGE_BASE: ${{ steps.revisions.outputs.merge_base }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          if git diff --quiet "$MERGE_BASE" "$HEAD_SHA" -- model_prices_and_context_window.json; then
+            echo "cost_map=false" >> "$GITHUB_OUTPUT"
+          else
+            echo "cost_map=true" >> "$GITHUB_OUTPUT"
+          fi
+      - if: steps.changed.outputs.cost_map == 'true'
+        run: rustup toolchain install --no-self-update
+      - name: Load the cost map through model-catalog
+        if: steps.changed.outputs.cost_map == 'true'
+        env:
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          git show "${HEAD_SHA}:model_prices_and_context_window.json" > "$RUNNER_TEMP/model_prices_and_context_window.json"
+          cargo run --locked --quiet --manifest-path litellm-rust/Cargo.toml -p litellm-model-catalog --bin check-registry \
+            -- "$RUNNER_TEMP/model_prices_and_context_window.json"

.github/workflows/cost-map-pins-recheck.yml

@@ -0,0 +1,91 @@
+name: Cost map pins recheck
+
+on:
+  push:
+    branches:
+      - main
+      - "litellm_**"
+    paths:
+      - model_prices_and_context_window.json
+      - litellm/model_prices_and_context_window_backup.json
+      - ci_cd/cost_map_pins.json
+
+permissions: {}
+
+jobs:
+  rerun-cost-map-guard:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      actions: write
+      contents: read
+      pull-requests: read
+    steps:
+      - name: Re-run the cost map guard on open pull requests the push could break
+        env:
+          GH_TOKEN: ${{ github.token }}
+          BASE_REF: ${{ github.ref_name }}
+          BEFORE_SHA: ${{ github.event.before }}
+          AFTER_SHA: ${{ github.event.after }}
+          FORCED: ${{ github.event.forced }}
+          COST_MAP_PATHS: '["model_prices_and_context_window.json", "litellm/model_prices_and_context_window_backup.json", "ci_cd/cost_map_pins.json"]'
+          PINS_PATHS: '["ci_cd/cost_map_pins.json"]'
+        run: |
+          if [[ "$BEFORE_SHA" =~ ^0+$ || "$FORCED" == true ]] \
+            || ! gh api "repos/${GITHUB_REPOSITORY}/compare/${BEFORE_SHA}...${AFTER_SHA}?per_page=1" \
+              --jq '(env.PINS_PATHS | fromjson) as $pins | (.files | length) < 300 and all(.files[]; [.filename, .previous_filename] | any(IN($pins[])) | not)' \
+              | grep -qx true; then
+            export RECHECK_PATHS="$COST_MAP_PATHS"
+          else
+            export RECHECK_PATHS="$PINS_PATHS"
+          fi
+          rerun_cutoff="$(date -u -d '30 days ago' +%Y-%m-%dT%H:%M:%SZ)"
+          candidates="$(gh api graphql --paginate -F owner="${GITHUB_REPOSITORY_OWNER}" -F name="${GITHUB_REPOSITORY#*/}" -F base="$BASE_REF" \
+            -f query='query($owner: String!, $name: String!, $base: String!, $endCursor: String) {
+              repository(owner: $owner, name: $name) {
+                pullRequests(states: OPEN, baseRefName: $base, first: 25, after: $endCursor) {
+                  nodes { number headRefOid changedFiles files(first: 100) { nodes { path } } }
+                  pageInfo { hasNextPage endCursor }
+                }
+              }
+            }' \
+            --jq '(env.RECHECK_PATHS | fromjson) as $paths | .data.repository.pullRequests.nodes[]
+              | {number, headRefOid, listed: any(.files.nodes[].path; IN($paths[])), unlisted: (.changedFiles > 100)}
+              | select(.listed or .unlisted) | "\(.number) \(.headRefOid) \(.listed)"')"
+          failed=0
+          declare -A rerun_heads
+          while read -r number head_sha listed; do
+            [ -n "$number" ] || continue
+            if [ "$listed" = false ]; then
+              if ! changed="$(gh api --paginate "repos/${GITHUB_REPOSITORY}/pulls/${number}/files?per_page=100" \
+                --jq '(env.RECHECK_PATHS | fromjson) as $paths | .[].filename | select(IN($paths[]))')"; then
+                echo "::error::could not list the files of #$number"
+                failed=1
+                continue
+              fi
+              [ -n "$changed" ] || continue
+            fi
+            [ -z "${rerun_heads[$head_sha]:-}" ] || continue
+            rerun_heads[$head_sha]=1
+            if ! runs="$(gh run list --repo "$GITHUB_REPOSITORY" --workflow cost-map-guard.yml --commit "$head_sha" --limit 100 --json databaseId,createdAt --jq '.[] | "\(.databaseId) \(.createdAt)"')"; then
+              echo "::error::could not list the cost map guard runs of #$number"
+              failed=1
+              continue
+            fi
+            if [ -z "$runs" ]; then
+              echo "::warning::no cost map guard run found for #$number"
+              continue
+            fi
+            while read -r run_id created_at; do
+              if [[ "$created_at" < "$rerun_cutoff" ]]; then
+                echo "::warning::cost map guard run $run_id for #$number is over 30 days old, so GitHub cannot re-run it; #$number gets the pin checks on its next push"
+                continue
+              fi
+              gh run watch "$run_id" --repo "$GITHUB_REPOSITORY" > /dev/null || true
+              if ! gh run rerun "$run_id" --repo "$GITHUB_REPOSITORY"; then
+                echo "::error::could not re-run cost map guard run $run_id for #$number"
+                failed=1
+              fi
+            done <<< "$runs"
+          done <<< "$candidates"
+          exit "$failed"

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

@@ -2,7 +2,7 @@ name: Create RC Branch
 
 on:
   schedule:
-    - cron: "0 3 * * 5"
+    - cron: "0 3 * * 1"
       timezone: "America/Los_Angeles"
   workflow_dispatch:
 

.github/workflows/deploy-dev-on-main.yml

@@ -1,30 +0,0 @@
-name: Deploy main to gateway-dev
-
-on:
-  push:
-    branches: [main]
-
-permissions: {}
-
-concurrency:
-  group: deploy-dev-dispatch
-  cancel-in-progress: true
-
-jobs:
-  dispatch:
-    if: github.repository == 'BerriAI/litellm'
-    runs-on: ubuntu-latest
-    timeout-minutes: 5
-    steps:
-      - uses: actions/create-github-app-token@fee1f7d63c2ff003460e3d139729b119787bc349 # v2.2.2
-        id: token
-        with:
-          app-id: ${{ secrets.LITELLM_OPS_DISPATCH_APP_ID }}
-          private-key: ${{ secrets.LITELLM_OPS_DISPATCH_APP_PRIVATE_KEY }}
-          owner: BerriAI
-          repositories: litellm-ops
-          permission-actions: write
-      - name: Dispatch the litellm-ops dev deploy
-        env:
-          GH_TOKEN: ${{ steps.token.outputs.token }}
-        run: gh api --method POST repos/BerriAI/litellm-ops/actions/workflows/deploy-dev.yml/dispatches -f ref=main

.github/workflows/image-scan.yml

@@ -13,8 +13,10 @@ on:
       - migrations/Dockerfile
       - migrations/run.py
       - gateway/Dockerfile
+      - gateway/launch.py
       - gateway/main.py
       - gateway/routes/allowlist.py
+      - litellm_cli/**
       - backend/Dockerfile
       - backend/main.py
       - litellm-rust/**

.github/workflows/interactions-spec-canary.yml

@@ -0,0 +1,94 @@
+name: Interactions spec canary
+
+on:
+  schedule:
+    - cron: "17 7 * * *"
+  pull_request:
+    paths:
+      - .github/workflows/interactions-spec-canary.yml
+      - tests/unit/interactions/**
+      - litellm/types/interactions/**
+  workflow_dispatch:
+
+permissions: {}
+
+concurrency: interactions-spec-canary
+
+jobs:
+  interactions-spec-canary:
+    name: interactions-spec-canary
+    if: github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      contents: read
+    outputs:
+      compliance: ${{ steps.compliance.outcome }}
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - name: Set up Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: "3.12"
+
+      - name: Set up uv
+        uses: ./.github/actions/setup-uv-with-retries
+        with:
+          version: "0.10.9"
+
+      - name: Set up the Rust bridge
+        uses: ./.github/actions/rust-bridge
+
+      - name: Install dependencies
+        run: .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra proxy
+
+      - name: Run the compliance tests against Google's live spec
+        id: compliance
+        run: |
+          uv run --no-sync pytest tests/unit/interactions/test_openapi_compliance.py \
+            --interactions-spec-url=https://ai.google.dev/static/api/interactions.openapi.json \
+            --tb=short -vv -p no:rerunfailures -p no:pytest-retry \
+            --allow-hosts=127.0.0.1,::1,localhost,ai.google.dev
+
+  tracking-issue:
+    name: tracking-issue
+    needs: interactions-spec-canary
+    if: always() && github.event_name != 'pull_request' && github.ref_name == github.event.repository.default_branch && github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    permissions:
+      issues: write
+    steps:
+      - name: File the tracking issue unless the compliance step passed, close it when it did
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
+        env:
+          COMPLIANCE_OUTCOME: ${{ needs.interactions-spec-canary.outputs.compliance }}
+          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        with:
+          script: |
+            const title = "Interactions spec canary is red";
+            const outcome = process.env.COMPLIANCE_OUTCOME;
+            const runUrl = process.env.RUN_URL;
+            const query = `repo:${context.repo.owner}/${context.repo.repo} is:issue is:open in:title "${title}"`;
+            const { data: found } = await github.rest.search.issuesAndPullRequests({ q: query, per_page: 20 });
+            const open = found.items.find((issue) => issue.title === title);
+            if (outcome === "success") {
+              if (open === undefined) return;
+              await github.rest.issues.createComment({ ...context.repo, issue_number: open.number, body: `Green again: ${runUrl}` });
+              await github.rest.issues.update({ ...context.repo, issue_number: open.number, state: "closed" });
+              return;
+            }
+            const state = outcome === "failure" ? "red" : `not run (compliance step outcome: ${outcome || "none"})`;
+            if (open !== undefined) {
+              await github.rest.issues.createComment({ ...context.repo, issue_number: open.number, body: `Still ${state}: ${runUrl}` });
+              return;
+            }
+            await github.rest.issues.create({
+              ...context.repo,
+              title,
+              assignees: ["mateo-berri"],
+              body: `The scheduled run of tests/unit/interactions/test_openapi_compliance.py against Google's live Interactions spec is ${state}: ${runUrl}\n\nlitellm/types/interactions/README.md says how to refresh the snapshot and regenerate the types.`,
+            });

.github/workflows/lens-install-smoke.yml

@@ -1,152 +0,0 @@
-name: Lens installation smoke
-
-on:
-  workflow_dispatch:
-    inputs:
-      lens_image:
-        description: Verified baseline Lens image at ghcr.io/berriai/lens@sha256
-        required: true
-        type: string
-      lens_upgrade_image:
-        description: Compatible Lens upgrade image at ghcr.io/berriai/lens@sha256
-        required: true
-        type: string
-      gateway_baseline_ref:
-        description: Previously qualified gateway source commit, full SHA
-        required: true
-        type: string
-
-permissions:
-  contents: read
-
-concurrency:
-  group: lens-install-${{ github.event.pull_request.number || github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  build-images:
-    runs-on: ubuntu-latest-16-cores
-    timeout-minutes: 45
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - component: gateway
-            dockerfile: gateway/Dockerfile
-          - component: backend
-            dockerfile: backend/Dockerfile
-          - component: ui
-            dockerfile: ui/Dockerfile
-          - component: migrations
-            dockerfile: migrations/Dockerfile
-          - component: monolith
-            dockerfile: Dockerfile
-          - component: gateway
-            dockerfile: gateway/Dockerfile
-            baseline: true
-          - component: backend
-            dockerfile: backend/Dockerfile
-            baseline: true
-          - component: monolith
-            dockerfile: Dockerfile
-            baseline: true
-    env:
-      COMPONENT: ${{ matrix.component }}
-      DOCKERFILE: ${{ matrix.dockerfile }}
-      SOURCE_REF: ${{ matrix.baseline && inputs.gateway_baseline_ref || github.sha }}
-      IMAGE_TAG: ${{ matrix.baseline && 'v0.0.0-lens-ci-baseline' || 'v0.0.0-lens-ci' }}
-      CURRENT_REF: ${{ github.sha }}
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          persist-credentials: false
-          ref: ${{ env.SOURCE_REF }}
-      - name: Verify the selected source
-        shell: bash -euo pipefail {0}
-        run: |
-          [[ "$SOURCE_REF" =~ ^[0-9a-f]{40}$ ]]
-          test "$(git rev-parse HEAD)" = "$SOURCE_REF"
-          if [[ "$IMAGE_TAG" == v0.0.0-lens-ci-baseline ]]; then
-            test "$SOURCE_REF" != "$CURRENT_REF"
-          fi
-      - name: Build the matching release image
-        run: |
-          docker build --build-arg "LITELLM_RELEASE_TAG=$IMAGE_TAG" \
-            -f "$DOCKERFILE" -t "lens-ci-$COMPONENT:$IMAGE_TAG" .
-      - name: Save the matching release image
-        run: |
-          docker save "lens-ci-$COMPONENT:$IMAGE_TAG" \
-            | gzip -1 > "$RUNNER_TEMP/lens-install-$COMPONENT-$IMAGE_TAG.tar.gz"
-      - uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1
-        with:
-          name: lens-install-${{ matrix.component }}-${{ env.IMAGE_TAG }}-${{ github.sha }}
-          path: ${{ runner.temp }}/lens-install-${{ matrix.component }}-${{ env.IMAGE_TAG }}.tar.gz
-          compression-level: 0
-          retention-days: 3
-          if-no-files-found: error
-          overwrite: true
-
-  helm-install:
-    needs: build-images
-    runs-on: ubuntu-latest-16-cores
-    timeout-minutes: 25
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          persist-credentials: false
-      - uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
-        with:
-          pattern: lens-install-*-${{ github.sha }}
-          merge-multiple: true
-          path: ${{ runner.temp }}/lens-install-images
-      - name: Load the matching release images
-        run: |
-          for archive in "$RUNNER_TEMP"/lens-install-images/lens-install-*.tar.gz; do
-            gzip -dc "$archive" | docker load
-            rm "$archive"
-          done
-      - uses: sigstore/cosign-installer@3454372f43399081ed03b604cb2d021dabca52bb # v3.8.2
-        with:
-          cosign-release: v2.4.3
-      - uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
-        with:
-          registry: ghcr.io
-          username: ${{ vars.LENS_READ_USER || github.actor }}
-          password: ${{ secrets.LENS_READ_TOKEN }}
-      - name: Verify independent Lens images before loading them
-        env:
-          LENS_IMAGE: ${{ inputs.lens_image }}
-          LENS_UPGRADE_IMAGE: ${{ inputs.lens_upgrade_image }}
-          LENS_PUBLIC_KEY: ${{ vars.LENS_COSIGN_PUBLIC_KEY }}
-          LENS_PUBLIC_KEY_SHA256: ${{ vars.LENS_COSIGN_PUBLIC_KEY_SHA256 }}
-        shell: bash -euo pipefail {0}
-        run: |
-          [[ "$LENS_PUBLIC_KEY_SHA256" =~ ^[0-9a-f]{64}$ ]]
-          printf '%s' "$LENS_PUBLIC_KEY" > "$RUNNER_TEMP/lens.pub"
-          echo "$LENS_PUBLIC_KEY_SHA256  $RUNNER_TEMP/lens.pub" | sha256sum --check
-          for image in "$LENS_IMAGE" "$LENS_UPGRADE_IMAGE"; do
-            [[ "$image" =~ ^ghcr.io/berriai/lens@sha256:[0-9a-f]{64}$ ]]
-            cosign verify --key "$RUNNER_TEMP/lens.pub" "$image" > /dev/null
-            cosign verify-attestation --key "$RUNNER_TEMP/lens.pub" --type spdxjson "$image" > /dev/null
-            docker pull "$image"
-          done
-          docker tag "$LENS_IMAGE" lens-ci-worker:baseline
-          docker tag "$LENS_UPGRADE_IMAGE" lens-ci-worker:upgrade
-      - name: Install pinned Kubernetes test tools
-        run: |
-          curl --fail --location --output "$RUNNER_TEMP/kind" \
-            https://kind.sigs.k8s.io/dl/v0.27.0/kind-linux-amd64
-          echo "a6875aaea358acf0ac07786b1a6755d08fd640f4c79b7a2e46681cc13f49a04b  $RUNNER_TEMP/kind" | sha256sum --check
-          chmod +x "$RUNNER_TEMP/kind"
-          curl --fail --location --output "$RUNNER_TEMP/kubectl" \
-            https://dl.k8s.io/release/v1.32.2/bin/linux/amd64/kubectl
-          echo "4f6a959dcc5b702135f8354cc7109b542a2933c46b808b248a214c1f69f817ea  $RUNNER_TEMP/kubectl" | sha256sum --check
-          chmod +x "$RUNNER_TEMP/kubectl"
-          curl --fail --location --output "$RUNNER_TEMP/helm.tar.gz" \
-            https://get.helm.sh/helm-v3.19.0-linux-amd64.tar.gz
-          echo "a7f81ce08007091b86d8bd696eb4d86b8d0f2e1b9f6c714be62f82f96a594496  $RUNNER_TEMP/helm.tar.gz" | sha256sum --check
-          tar -xzf "$RUNNER_TEMP/helm.tar.gz" -C "$RUNNER_TEMP"
-          echo "$RUNNER_TEMP" >> "$GITHUB_PATH"
-          echo "$RUNNER_TEMP/linux-amd64" >> "$GITHUB_PATH"
-      - name: Install, ingest, upgrade, and restart both charts
-        run: bash tests/e2e/migrations/lens_helm_smoke.sh

.github/workflows/osv-scan.yml

@@ -5,6 +5,9 @@ on:
     branches:
       - main
       - "litellm_**"
+      - "rc/**"
+      - "stable/**"
+      - "lts/**"
   schedule:
     - cron: "23 6 * * *"
   workflow_dispatch:

.github/workflows/spend-logs-guard.yml

@@ -0,0 +1,45 @@
+name: Spend logs guard
+
+on:
+  pull_request:
+    branches:
+      - main
+
+permissions:
+  contents: read
+  pull-requests: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  spend-logs-guard:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          ref: ${{ github.event.pull_request.base.ref }}
+          persist-credentials: false
+      - name: Fetch the pull request head, its merge base and its reviews
+        id: revisions
+        env:
+          GH_TOKEN: ${{ github.token }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          BASE_REF: ${{ github.event.pull_request.base.ref }}
+        run: |
+          merge_base="$(gh api "repos/${GITHUB_REPOSITORY}/compare/${BASE_REF}...${HEAD_SHA}" --jq '.merge_base_commit.sha')"
+          git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA"
+          gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}/reviews" --paginate --slurp | jq 'add // []' > "$RUNNER_TEMP/reviews.json"
+          echo "merge_base=$merge_base" >> "$GITHUB_OUTPUT"
+      - name: Set up uv
+        uses: ./.github/actions/setup-uv-with-retries
+        with:
+          version: "0.10.9"
+      - name: Run the guard
+        env:
+          MERGE_BASE: ${{ steps.revisions.outputs.merge_base }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          uv run --no-project --with pydantic==2.14.1 python ci_cd/spend_logs_guard.py --base "$MERGE_BASE" --head "$HEAD_SHA" --approver kerry-berri --reviews "$RUNNER_TEMP/reviews.json"

.github/workflows/sync-together-ai-models.yml

@@ -1,70 +0,0 @@
-name: Sync Together AI model registry
-
-on:
-  schedule:
-    - cron: "30 6 * * *"
-  workflow_dispatch:
-
-permissions:
-  contents: write
-  pull-requests: write
-
-jobs:
-  sync_together_ai_models:
-    if: github.repository == 'BerriAI/litellm'
-    runs-on: ubuntu-latest
-    env:
-      BASE_BRANCH: ${{ github.event.repository.default_branch }}
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          ref: ${{ env.BASE_BRANCH }}
-          persist-credentials: false
-      - name: Set up uv
-        uses: ./.github/actions/setup-uv-with-retries
-        with:
-          version: "0.10.9"
-      - name: Look for an already-open sync PR
-        id: existing
-        run: |
-          open_pr="$(gh pr list --repo "$GITHUB_REPOSITORY" --state open --limit 1000 --json headRefName \
-            --jq '[.[].headRefName | select(startswith("litellm_together_registry_sync_"))] | first // empty')"
-          echo "open_pr=$open_pr" >> "$GITHUB_OUTPUT"
-          if [ -n "$open_pr" ]; then
-            echo "An open sync PR already exists on branch $open_pr; skipping this run."
-          fi
-        env:
-          GH_TOKEN: ${{ secrets.GH_TOKEN || github.token }}
-      - name: Run the sync
-        if: steps.existing.outputs.open_pr == ''
-        run: |
-          uv run --frozen python scripts/sync_together_ai_models.py --write --pr-body-file "$RUNNER_TEMP/pr_body.md"
-        env:
-          TOGETHER_API_KEY: ${{ secrets.TOGETHER_API_KEY }}
-      - name: Regenerate the JSON schema
-        if: steps.existing.outputs.open_pr == ''
-        run: |
-          uv run --frozen python ci_cd/generate_model_prices_schema.py
-      - name: Create a pull request when the registry changed
-        if: steps.existing.outputs.open_pr == ''
-        run: |
-          if git diff --quiet; then
-            echo "Registry already in sync; no PR needed."
-            exit 0
-          fi
-          branch="litellm_together_registry_sync_$(date +'%Y-%m-%d')"
-          git config user.name "github-actions[bot]"
-          git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
-          git checkout -b "$branch"
-          git add model_prices_and_context_window.json \
-            litellm/model_prices_and_context_window_backup.json \
-            model_prices_and_context_window.schema.json
-          git commit -m "feat(models): sync together_ai model registry $(date +'%Y-%m-%d')"
-          gh auth setup-git
-          git push origin "$branch"
-          gh pr create --title "feat(models): sync together_ai model registry" \
-            --body-file "$RUNNER_TEMP/pr_body.md" \
-            --head "$branch" \
-            --base "$BASE_BRANCH"
-        env:
-          GH_TOKEN: ${{ secrets.GH_TOKEN || github.token }}

.github/workflows/test-integration.yml

@@ -0,0 +1,199 @@
+name: integration
+
+on:
+  pull_request:
+    types: [labeled]
+  schedule:
+    - cron: "5 8-20/3 * * 1-6"
+      timezone: America/Los_Angeles
+  workflow_dispatch:
+    inputs:
+      suite:
+        description: Integration group to run
+        type: choice
+        options:
+          - all
+          - providers
+          - extensions
+          - accounting
+          - management
+          - management-replica
+          - database
+          - database-replica
+          - mcp
+          - security
+          - cost
+          - sdk
+          - browser
+        default: all
+      workers:
+        description: Workers per group, blank keeps each group's default
+        type: string
+        default: ""
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}-${{ github.event.label.name }}-${{ inputs.suite }}-${{ inputs.workers }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  integration:
+    name: integration ${{ matrix.suite }}
+    if: github.event_name != 'pull_request' || github.event.label.name == 'run-ci'
+    environment: integration-license
+    runs-on: ${{ fromJSON('{"providers":"ubuntu-latest-64-cores","extensions":"ubuntu-latest-64-cores","accounting":"ubuntu-latest-32-cores","management":"ubuntu-latest-32-cores","management-replica":"ubuntu-latest-32-cores"}')[matrix.suite] || 'ubuntu-latest-16-cores' }}
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        suite: ${{ fromJSON((inputs.suite || 'all') == 'all' && '["providers","extensions","accounting","management","management-replica","database","database-replica","mcp","security","cost","sdk","browser"]' || format('["{0}"]', inputs.suite)) }}
+    env:
+      UV_PYTHON: "3.12"
+      INTEGRATION_WORKERS: ${{ inputs.workers || fromJSON('{"providers":"32","extensions":"48","accounting":"16","management":"16","management-replica":"16","cost":"16","mcp":"8","security":"8","database":"4","database-replica":"4"}')[matrix.suite] || '' }}
+    services:
+      postgres:
+        image: postgres:16@sha256:e17e86066e5ef83e0952a9347f5c792b7ece00972e2aa787a6986f471b3dd3d5
+        command: -c max_connections=1000 -c shared_preload_libraries=pg_stat_statements -c pg_stat_statements.track=all -c pg_stat_statements.max=20000
+        env:
+          POSTGRES_USER: postgres
+          POSTGRES_PASSWORD: postgres
+          POSTGRES_DB: circle_test
+        ports:
+          - 5432:5432
+        options: >-
+          --shm-size 1g
+          --health-cmd "pg_isready -U postgres"
+          --health-interval 5s
+          --health-timeout 5s
+          --health-retries 20
+      redis:
+        image: redis:7-alpine@sha256:7aec734b2bb298a1d769fd8729f13b8514a41bf90fcdd1f38ec52267fbaa8ee6
+        ports:
+          - 6379:6379
+        options: >-
+          --health-cmd "redis-cli ping"
+          --health-interval 5s
+          --health-timeout 5s
+          --health-retries 20
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - name: Set up Python
+        timeout-minutes: 3
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: ${{ env.UV_PYTHON }}
+
+      - name: Set up uv
+        timeout-minutes: 3
+        uses: ./.github/actions/setup-uv-with-retries
+        with:
+          version: "0.10.9"
+
+      - name: Cache the Rust build
+        timeout-minutes: 5
+        uses: ./.github/actions/cache-cargo-build
+
+      - name: Cache Prisma binaries
+        timeout-minutes: 5
+        uses: ./.github/actions/cache-prisma-binaries
+
+      - name: Install dependencies
+        timeout-minutes: 20
+        run: .github/scripts/uv_sync_with_retries.sh --frozen --all-groups --all-extras
+
+      - name: Install redis-server 7.2.16
+        timeout-minutes: 3
+        run: |
+          mkdir -p "$RUNNER_TEMP/redis-bin"
+          cid="$(docker create redis:7.2.16@sha256:0637954999d01b7c9ce9167db2da50656e2590d3b884f1c600c5f63bb6e6773c)"
+          docker cp "${cid}:/usr/local/bin/redis-server" "$RUNNER_TEMP/redis-bin/redis-server"
+          docker rm "$cid"
+          "$RUNNER_TEMP/redis-bin/redis-server" --version | grep -q 'v=7.2.16'
+          echo "$RUNNER_TEMP/redis-bin" >> "$GITHUB_PATH"
+
+      - name: Install MCP SDK1 peer
+        if: matrix.suite == 'mcp'
+        timeout-minutes: 5
+        run: |
+          uv venv --python 3.12 .venv-mcp-peer
+          uv pip install --no-config --exclude-newer "3 days" --python .venv-mcp-peer 'mcp==1.28.1' 'langchain-mcp-adapters==0.2.1'
+          echo "MCP_TEST_PEER_PYTHON=$PWD/.venv-mcp-peer/bin/python" >> "$GITHUB_ENV"
+
+      - name: Set up Node.js
+        if: matrix.suite == 'browser'
+        timeout-minutes: 5
+        uses: actions/setup-node@a0853c24544627f65ddf259abe73b1d18a591444 # v5.0.0
+        with:
+          node-version-file: ui/litellm-dashboard/.nvmrc
+          cache: "npm"
+          cache-dependency-path: |
+            ui/litellm-dashboard/package-lock.json
+            tests/e2e/ui/package-lock.json
+
+      - name: Install browser dependencies and build the dashboard
+        if: matrix.suite == 'browser'
+        timeout-minutes: 20
+        run: |
+          (cd ui/litellm-dashboard && npm ci)
+          (cd tests/e2e/ui && npm ci)
+          sudo env PATH="$PATH" DEBIAN_FRONTEND=noninteractive \
+            node tests/e2e/ui/node_modules/@playwright/test/cli.js install-deps chromium
+          node tests/e2e/ui/node_modules/@playwright/test/cli.js install chromium
+          (cd ui/litellm-dashboard && NEXT_TELEMETRY_DISABLED=1 npm run build)
+
+      - name: Run integration contracts
+        timeout-minutes: 45
+        env:
+          SUITE: ${{ matrix.suite }}
+          INTEGRATION_POSTGRES_CONTAINER: ${{ job.services.postgres.id }}
+          INTEGRATION_REDIS_CONTAINER: ${{ job.services.redis.id }}
+          LITELLM_LICENSE: ${{ secrets.LITELLM_LICENSE }}
+        run: |
+          case "$SUITE" in
+            *-replica) bash .circleci/scripts/run_integration.sh "${SUITE%-replica}" replica ;;
+            *) bash .circleci/scripts/run_integration.sh "$SUITE" ;;
+          esac
+
+      - name: Verify every case ran exactly once
+        if: ${{ !cancelled() && matrix.suite != 'browser' }}
+        timeout-minutes: 2
+        run: |
+          record="$(find test-results -mindepth 2 -maxdepth 2 -name execution.json -print -quit)"
+          test -n "$record"
+          mkdir -p "$RUNNER_TEMP/shards/0"
+          cp "$record" "$RUNNER_TEMP/shards/0/execution.json"
+          .venv/bin/python -m tests.integration._support.sharding "$RUNNER_TEMP/shards" --shards 1
+
+      - name: Report MCP coverage
+        if: ${{ !cancelled() && matrix.suite == 'mcp' }}
+        timeout-minutes: 5
+        run: |
+          coverage_dir=test-results/integration-mcp/coverage
+          test -f "$coverage_dir/coverage.txt" && exit 0
+          export COVERAGE_FILE="$PWD/$coverage_dir/data"
+          .venv/bin/python -m coverage combine --rcfile=tests/integration/mcp_coverage.toml
+          .venv/bin/python -m coverage report --rcfile=tests/integration/mcp_coverage.toml > "$coverage_dir/coverage.txt"
+          .venv/bin/python -m coverage html --rcfile=tests/integration/mcp_coverage.toml -d "$coverage_dir/html"
+          tail -n 1 "$coverage_dir/coverage.txt"
+
+      - name: Check results for the license
+        id: secret-scan
+        if: ${{ !cancelled() }}
+        timeout-minutes: 5
+        env:
+          SCANNED_SECRET: ${{ secrets.LITELLM_LICENSE }}
+        run: python .github/scripts/assert_secret_absent.py test-results
+
+      - name: Upload results
+        if: ${{ !cancelled() && steps.secret-scan.outcome == 'success' }}
+        timeout-minutes: 10
+        uses: actions/upload-artifact@4cec3d8aa04e39d1a68397de0c4cd6fb9dce8ec1 # v4.6.1
+        with:
+          name: integration-${{ matrix.suite }}-${{ github.run_id }}-${{ github.run_attempt }}
+          path: test-results/
+          retention-days: 14

.github/workflows/test-linting.yml

@@ -21,7 +21,7 @@ jobs:
       pull-requests: read
       actions: read
     runs-on: ubuntu-latest
-    timeout-minutes: 15
+    timeout-minutes: 20
     services:
       redis:
         image: redis@sha256:e2debfb7956fa12c7ddc79d7e645c8cf26b30c99a6e9161ea9bf4171e1668a5f
@@ -48,11 +48,13 @@ jobs:
         if: steps.changes.outputs.decision != 'skip'
         env:
           GH_TOKEN: ${{ github.token }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          EVENT_BASE_REF: ${{ github.event.pull_request.base.ref }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
           retry() { "$@" || { sleep 15; "$@"; } || { sleep 30; "$@"; }; }
-          MERGE_BASE=$(retry gh api "repos/${{ github.repository }}/compare/${BASE_SHA}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
+          BASE_REF=$(gh api "repos/${GITHUB_REPOSITORY}/branches/${EVENT_BASE_REF}" --jq .name) || BASE_REF=$(retry gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" --jq .base.ref)
+          MERGE_BASE=$(retry gh api "repos/${GITHUB_REPOSITORY}/compare/${BASE_REF}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
           test -n "$MERGE_BASE"
           retry git fetch --no-tags --depth=1 origin "$MERGE_BASE"
           echo "GATE_BASE_SHA=$MERGE_BASE" >> "$GITHUB_ENV"
@@ -206,7 +208,7 @@ jobs:
             exit 0
           fi
           retry() { "$@" || { sleep 15; "$@"; } || { sleep 30; "$@"; }; }
-          CLAUDE_VERSION="$(retry uv run --no-sync python tests/e2e/claude_code/pr_gate_version_resolver.py)"
+          CLAUDE_VERSION="$(retry uv run --no-sync python tests/e2e/cli_harness/npm_version_resolver.py @anthropic-ai/claude-code)"
           tests/e2e/claude_code/cron_vm/install_claude_code.sh "$CLAUDE_VERSION" "$RUNNER_TEMP/claude-cli"
           PATH="$RUNNER_TEMP/claude-cli:$PATH" uv run --no-sync pytest -q tests/e2e_harness
           uv run --no-sync pytest --tb=short -q --noconftest -o addopts= -o "pythonpath=tests/e2e tests/e2e_harness" -p no:rerunfailures \
@@ -264,6 +266,7 @@ jobs:
     name: ui
     permissions:
       contents: read
+      pull-requests: read
     runs-on: ubuntu-latest
     timeout-minutes: 8
     defaults:
@@ -281,16 +284,12 @@ jobs:
         id: changed
         env:
           GH_TOKEN: ${{ github.token }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          EVENT_BASE_REF: ${{ github.event.pull_request.base.ref }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
-          # base.sha is the base branch tip from when the PR was opened, while
-          # actions/checkout leaves HEAD on a merge of the PR into the *current*
-          # base tip. "$BASE_SHA"...HEAD therefore spans every base-branch commit
-          # landed since, so a PR that touches no UI file still gets linted
-          # against hundreds of other people's files. Diff the PR head against its
-          # own merge base instead, which is exactly what this PR changed.
-          merge_base=$(gh api "repos/${{ github.repository }}/compare/${BASE_SHA}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
+          base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/branches/${EVENT_BASE_REF}" --jq .name) || base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" --jq .base.ref)
+          merge_base=$(gh api "repos/${GITHUB_REPOSITORY}/compare/${base_ref}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
           test -n "$merge_base"
           git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA"
           : > "$RUNNER_TEMP/prettier_files.txt"
@@ -376,10 +375,10 @@ jobs:
             exit 0
           fi
           files="$(git diff --name-only "$base" HEAD)"
-          if grep -Eq '^(litellm/(proxy|types)/|ui/litellm-dashboard/(src/lib/http/schema\.d\.ts|scripts/gen-api-types\.mjs|package(-lock)?\.json)$|\.github/workflows/test-linting\.yml$)' <<< "$files"; then
+          if grep -Eq '^(litellm/|enterprise/|(pyproject\.toml|uv\.lock)$|ui/litellm-dashboard/(src/lib/http/schema\.d\.ts|scripts/gen-api-types\.mjs|package(-lock)?\.json)$|\.github/workflows/test-linting\.yml$)' <<< "$files"; then
             echo "relevant=true" >> "$GITHUB_OUTPUT"
           else
-            echo "No proxy, types or generator changes in this pull request, nothing to verify."
+            echo "No litellm, enterprise, dependency or generator changes in this pull request, nothing to verify."
             echo "relevant=false" >> "$GITHUB_OUTPUT"
           fi
 
@@ -524,6 +523,12 @@ jobs:
       - name: check_provider_folders_documented
         run: uv run --no-sync python ./tests/code_coverage_tests/check_provider_folders_documented.py
 
+      - name: check_provider_endpoints_support_generated
+        run: uv run --no-sync python scripts/generate_provider_endpoints_support.py --check
+
+      - name: test_generate_provider_endpoints_support
+        run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_generate_provider_endpoints_support.py
+
       - name: check_prisma_binary_cache
         run: uv run --no-sync python ./tests/code_coverage_tests/check_prisma_binary_cache.py
 
@@ -539,8 +544,8 @@ jobs:
       - name: test_unit_passed_gate
         run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_unit_passed_gate.py
 
-      - name: test_e2e_changed_gate
-        run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_e2e_changed_gate.py tests/code_coverage_tests/test_e2e_idp_stack.py
+      - name: test_e2e_idp_stack
+        run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_e2e_idp_stack.py
 
       - name: test_e2e_metadata
         env:
@@ -555,6 +560,9 @@ jobs:
       - name: router_code_coverage
         run: uv run --no-sync python ./tests/code_coverage_tests/router_code_coverage.py
 
+      - name: router_surface_check
+        run: uv run --no-sync python ./tests/code_coverage_tests/router_surface_check.py
+
       - name: test_chat_completion_imports
         run: uv run --no-sync python ./tests/code_coverage_tests/test_chat_completion_imports.py
 
@@ -621,6 +629,12 @@ jobs:
       - name: check_migrations_lock_timeout
         run: uv run --no-sync python ./tests/code_coverage_tests/check_migrations_lock_timeout.py
 
+      - name: check_no_none_string_error_fields
+        run: uv run --no-sync python ./tests/code_coverage_tests/check_no_none_string_error_fields.py
+
+      - name: test_check_no_none_string_error_fields
+        run: uv run --no-sync pytest -q --noconftest -p no:cacheprovider -c /dev/null tests/code_coverage_tests/test_check_no_none_string_error_fields.py
+
       - name: check_no_publicly_known_master_key
         run: uv run --no-sync python ./tests/code_coverage_tests/check_no_publicly_known_master_key.py
 

.github/workflows/test-mcp-dependency-resolution.yml

@@ -45,7 +45,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14", "3.15"]
 
     steps:
       - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
@@ -132,7 +132,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14", "3.15"]
     steps:
       - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
         with:

.github/workflows/test-mcp-oauth-e2e.yml

@@ -1,180 +0,0 @@
-name: MCP OAuth happy path
-
-on:
-  pull_request:
-    paths:
-      - '.github/workflows/test-mcp-oauth-e2e.yml'
-      - '.github/e2e-stack/**'
-      - 'tests/e2e/*.py'
-      - 'tests/e2e/pytest.ini'
-      - 'tests/e2e/idp_realm.json'
-      - 'tests/e2e/mcp/**'
-      - 'litellm/experimental_mcp_client/**'
-      - 'litellm/proxy/_experimental/mcp_server/**'
-      - 'litellm/proxy/auth/**'
-      - 'litellm/proxy/management_endpoints/mcp_management_endpoints.py'
-      - 'litellm/proxy/_types.py'
-      - 'litellm/types/mcp_server/mcp_server_manager.py'
-      - 'litellm/proxy/management_endpoints/*sso*.py'
-      - 'litellm/proxy/management_endpoints/sso/**'
-      - 'litellm/proxy/common_utils/encrypt_decrypt_utils.py'
-      - 'litellm/proxy/proxy_server.py'
-      - 'litellm/proxy/schema.prisma'
-      - 'ui/litellm-dashboard/src/app/connect/**'
-      - 'ui/litellm-dashboard/src/app/mcp/oauth/**'
-      - 'pyproject.toml'
-      - 'uv.lock'
-  workflow_dispatch:
-
-permissions: {}
-
-concurrency:
-  group: mcp-oauth-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  oauth:
-    if: github.event_name == 'workflow_dispatch' || github.event.pull_request.head.repo.full_name == github.repository
-    runs-on: ubuntu-latest
-    environment: e2e-changed
-    timeout-minutes: 45
-    permissions:
-      contents: read
-      id-token: write
-    services:
-      postgres:
-        image: postgres:16.6
-        env:
-          POSTGRES_USER: litellm
-          POSTGRES_PASSWORD: dbpassword9090
-          POSTGRES_DB: litellm
-        ports:
-          - 5432:5432
-        options: >-
-          --health-cmd "pg_isready -U litellm"
-          --health-interval 5s
-          --health-timeout 5s
-          --health-retries 10
-    env:
-      DATABASE_HOST: 127.0.0.1
-      DATABASE_PORT: '5432'
-      DATABASE_USER: litellm
-      DATABASE_PASSWORD: dbpassword9090
-      DATABASE_NAME: litellm
-      DATABASE_URL: postgresql://litellm:dbpassword9090@127.0.0.1:5432/litellm
-      E2E_KEYCLOAK_URL: http://127.0.0.1:8081
-      E2E_KEYCLOAK_ADMIN_USER: admin
-      E2E_KEYCLOAK_ADMIN_PASSWORD: e2e-ephemeral-idp-not-a-secret
-      E2E_FIXTURE_MODE: live
-      E2E_PROVIDER_CACHE: '0'
-      E2E_MCP_OAUTH_LIVE: '1'
-      E2E_REQUIRED_TEST_COUNT: '4'
-    steps:
-      - name: Checkout the tested source
-        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          ref: ${{ github.event.pull_request.head.sha || github.sha }}
-          persist-credentials: false
-
-      - name: Require and materialize the upstream login
-        env:
-          STORAGE_STATE: ${{ secrets.E2E_LINEAR_STORAGE_STATE_B64 }}
-        run: |
-          umask 077
-          python3 - <<'PY'
-          import base64
-          import json
-          import os
-          import secrets
-          from pathlib import Path
-          encoded = os.environ.get("STORAGE_STATE", "")
-          if not encoded:
-              raise SystemExit("E2E_LINEAR_STORAGE_STATE_B64 is required; capture and provision a test-account login")
-          state = json.loads(base64.b64decode(encoded, validate=True))
-          if not isinstance(state, dict) or not state.get("cookies"):
-              raise SystemExit("The captured login must contain browser cookies")
-          directory = Path(os.environ["RUNNER_TEMP"]) / "mcp-oauth-private"
-          directory.mkdir(mode=0o700)
-          path = directory / "linear-state.json"
-          path.write_text(json.dumps(state))
-          with open(os.environ["GITHUB_ENV"], "a") as output:
-              output.write(f"E2E_LINEAR_STORAGE_STATE={path}\n")
-              for name in ("LITELLM_MASTER_KEY", "LITELLM_SALT_KEY"):
-                  value = "sk-e2e-" + secrets.token_hex(24)
-                  print(f"::add-mask::{value}")
-                  output.write(f"{name}={value}\n")
-          PY
-
-      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-        with:
-          python-version: '3.13'
-      - uses: ./.github/actions/setup-uv-with-retries
-        with:
-          version: '0.10.9'
-      - uses: ./.github/actions/cache-cargo-build
-      - name: Install the frozen E2E environment
-        run: |
-          .github/scripts/uv_sync_with_retries.sh --frozen --extra proxy --extra proxy-runtime --extra extra_proxy --group ci --group proxy-dev --group e2e-dev
-          uv run --no-sync python scripts/prisma_generate_if_needed.py
-          uv run --no-sync playwright install --with-deps chromium
-
-      - name: Configure license access
-        id: aws
-        uses: aws-actions/configure-aws-credentials@e7f100cf4c008499ea8adda475de1042d6975c7b # v6.2.0
-        with:
-          role-to-assume: ${{ vars.E2E_AWS_ROLE_TO_ASSUME }}
-          aws-region: us-east-1
-          role-session-name: mcp-oauth-${{ github.run_id }}
-          role-duration-seconds: 900
-          output-env-credentials: false
-          output-credentials: true
-      - name: Load the E2E license
-        env:
-          AWS_ACCESS_KEY_ID: ${{ steps.aws.outputs.aws-access-key-id }}
-          AWS_SECRET_ACCESS_KEY: ${{ steps.aws.outputs.aws-secret-access-key }}
-          AWS_SESSION_TOKEN: ${{ steps.aws.outputs.aws-session-token }}
-          AWS_DEFAULT_REGION: us-east-1
-        run: |
-          license="$(aws secretsmanager get-secret-value --secret-id litellm-e2e-changed-license --query SecretString --output text)"
-          test -n "${license}"
-          echo "::add-mask::${license}"
-          echo "LITELLM_LICENSE=${license}" >> "${GITHUB_ENV}"
-
-      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
-        with:
-          node-version-file: ui/litellm-dashboard/.nvmrc
-      - name: Build the gateway consent UI at the tested commit
-        run: |
-          cd ui/litellm-dashboard
-          ../../scripts/with_dashboard_node.sh npm ci
-          ../../scripts/with_dashboard_node.sh npm run build
-          mkdir -p ../../litellm/proxy/_experimental/out
-          cp -r out/. ../../litellm/proxy/_experimental/out/
-          find ../../litellm/proxy/_experimental/out -name '*.html' ! -name index.html | while read -r page; do
-            mkdir -p "${page%.html}"
-            mv "${page}" "${page%.html}/index.html"
-          done
-
-      - name: Prepare the isolated database and IdP
-        run: |
-          umask 077
-          bash .github/e2e-stack/start-idp.sh
-          uv run --no-sync python migrations/run.py > "${RUNNER_TEMP}/mcp-oauth-private/migrations.log" 2>&1
-
-      - name: Run every required OAuth variant without retries
-        run: |
-          umask 077
-          uv run --no-sync pytest -c tests/e2e/pytest.ini tests/e2e/mcp/test_mcp_oauth_happy_path_e2e.py \
-            --rootdir=. --reruns 0 --tb=short -o junit_family=xunit1 \
-            --junitxml="${RUNNER_TEMP}/mcp-oauth-private/results.xml" \
-            > "${RUNNER_TEMP}/mcp-oauth-private/pytest.log" 2>&1
-      - name: Report JUnit results and reject skipped or missing cases
-        if: always()
-        run: |
-          uv run --no-sync python .github/e2e-stack/assert_tests_ran.py \
-            "${RUNNER_TEMP}/mcp-oauth-private/results.xml" tests/e2e/mcp/test_mcp_oauth_happy_path_e2e.py
-      - name: Remove private login and logs
-        if: always()
-        run: |
-          docker rm -f e2e-keycloak >/dev/null 2>&1 || true
-          rm -rf "${RUNNER_TEMP}/mcp-oauth-private"

.github/workflows/test-merge-smoke.yml

@@ -33,7 +33,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14"]
+        python-version: ["3.10", "3.11", "3.12", "3.13", "3.14", "3.15"]
     env:
       LITELLM_LOCAL_MODEL_COST_MAP: "True"
     steps:
@@ -76,6 +76,12 @@ jobs:
       - name: Import and CLI checks
         run: $IN_NETNS .venv/bin/python .github/scripts/run_merge_smoke.py cli
 
+      - name: Routing and UUID compatibility
+        run: >-
+          $IN_NETNS .venv/bin/python -m pytest
+          tests/unit/router_strategy/test_complexity_router.py::TestCustomDimensions
+          tests/unit/test_uuid_helper.py -q
+
       - name: Proxy startup check
         run: $IN_NETNS .venv/bin/python .github/scripts/run_merge_smoke.py proxy-startup --diagnostics-dir "${RUNNER_TEMP}/smoke-diagnostics"
 
@@ -126,7 +132,7 @@ jobs:
   windows-sdk:
     name: windows-sdk
     runs-on: windows-2025
-    timeout-minutes: 10
+    timeout-minutes: 30
     env:
       LITELLM_LOCAL_MODEL_COST_MAP: "True"
     defaults:
@@ -148,8 +154,11 @@ jobs:
         with:
           version: "0.10.9"
 
+      - name: Set up the Rust bridge
+        uses: ./.github/actions/rust-bridge
+
       - name: Install dependencies
-        run: .github/scripts/uv_sync_with_retries.sh --frozen --no-install-workspace --group dev --python 3.11
+        run: .github/scripts/uv_sync_with_retries.sh --frozen --group dev --python 3.11
 
       - name: Run the Windows tests
         run: uv run --no-sync python -m pytest --tb=short -v tests/windows_tests

.github/workflows/test-rust-db.yml

@@ -0,0 +1,116 @@
+name: LiteLLM Rust DB
+
+on:
+  push:
+    paths:
+      - "litellm-rust/crates/db/**"
+      - "litellm-rust/Cargo.toml"
+      - "litellm-rust/Cargo.lock"
+      - "litellm-rust/clippy.toml"
+      - "litellm-proxy-extras/litellm_proxy_extras/migrations/**"
+      - "**/schema.prisma"
+      - ".cargo/**"
+      - "rust-toolchain.toml"
+      - "pyproject.toml"
+      - "uv.lock"
+      - ".github/actions/setup-uv-with-retries/**"
+      - ".github/workflows/test-rust-db.yml"
+  pull_request:
+    branches:
+      - main
+      - "litellm_**"
+    paths:
+      - "litellm-rust/crates/db/**"
+      - "litellm-rust/Cargo.toml"
+      - "litellm-rust/Cargo.lock"
+      - "litellm-rust/clippy.toml"
+      - "litellm-proxy-extras/litellm_proxy_extras/migrations/**"
+      - "**/schema.prisma"
+      - ".cargo/**"
+      - "rust-toolchain.toml"
+      - "pyproject.toml"
+      - "uv.lock"
+      - ".github/actions/setup-uv-with-retries/**"
+      - ".github/workflows/test-rust-db.yml"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+env:
+  CARGO_TERM_COLOR: always
+  RUSTC_WRAPPER: sccache
+  SCCACHE_GHA_ENABLED: "true"
+  SCCACHE_GHA_VERSION: litellm-ci-v1
+  SCCACHE_GHA_RW_MODE: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' && 'READ_WRITE' || 'READ_ONLY' }}
+  CARGO_INCREMENTAL: "0"
+
+jobs:
+  db-queries:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    services:
+      postgres:
+        image: postgres:16.6@sha256:557fea37a744d5f4c8faab304b0a90858b53ab119735a88c131fd19dab802f36
+        env:
+          POSTGRES_DB: litellm
+          POSTGRES_PASSWORD: postgres
+          POSTGRES_USER: postgres
+        ports:
+          - 5432:5432
+        options: >-
+          --health-cmd "pg_isready -U postgres -d litellm"
+          --health-interval 10s
+          --health-timeout 5s
+          --health-retries 5
+    env:
+      DATABASE_URL: postgresql://postgres:postgres@localhost:5432/litellm
+      SQLX_OFFLINE: "false"
+    defaults:
+      run:
+        working-directory: litellm-rust
+    steps:
+      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
+        with:
+          python-version: "3.12"
+
+      - uses: ./.github/actions/setup-uv-with-retries
+        with:
+          version: "0.10.9"
+
+      - run: rustup toolchain install --no-self-update
+
+      - uses: Swatinem/rust-cache@6323deb102c322ba6fcbdcafc7e3dddab59af2b6 # v2.9.2
+        with:
+          workspaces: litellm-rust
+          cache-on-failure: true
+
+      - name: Set up sccache
+        uses: mozilla-actions/sccache-action@d651010b8da762cde178750d8eda7b5febfe147a
+        with:
+          version: "v0.18.0"
+
+      - name: Install Python dependencies
+        working-directory: .
+        run: uv sync --frozen --extra extra-proxy
+
+      - name: Deploy Prisma migrations
+        working-directory: .
+        run: uv run --no-sync prisma migrate deploy --schema litellm-proxy-extras/litellm_proxy_extras/schema.prisma
+
+      - name: Install SQLx CLI
+        run: cargo install sqlx-cli --version 0.9.0 --locked --no-default-features -F postgres
+
+      - name: Check SQLx query metadata
+        working-directory: litellm-rust/crates/db
+        run: cargo sqlx prepare --check
+
+      - name: Test database queries
+        run: cargo test -p litellm-db --locked -- --include-ignored

.github/workflows/test-rust.yml

@@ -4,6 +4,7 @@ on:
   push:
     paths:
       - "litellm-rust/**"
+      - "model_prices_and_context_window.json"
       - "litellm/rust_bridge/**"
       - "scripts/generate_trace_types.py"
       - "scripts/trace_codegen/**"
@@ -16,6 +17,10 @@ on:
       - "litellm/proxy/guardrails/**"
       - "litellm/utils.py"
       - "litellm/ocr/**"
+      - "litellm/chat_completions/dispatch.py"
+      - "litellm/embeddings/dispatch.py"
+      - "litellm/messages/dispatch.py"
+      - "litellm/responses/dispatch.py"
       - "litellm/llms/base_llm/ocr/**"
       - "litellm/llms/custom_httpx/llm_http_handler.py"
       - "Makefile"
@@ -34,6 +39,7 @@ on:
       - "litellm_**"
     paths:
       - "litellm-rust/**"
+      - "model_prices_and_context_window.json"
       - "litellm/rust_bridge/**"
       - "scripts/generate_trace_types.py"
       - "scripts/trace_codegen/**"
@@ -46,6 +52,10 @@ on:
       - "litellm/proxy/guardrails/**"
       - "litellm/utils.py"
       - "litellm/ocr/**"
+      - "litellm/chat_completions/dispatch.py"
+      - "litellm/embeddings/dispatch.py"
+      - "litellm/messages/dispatch.py"
+      - "litellm/responses/dispatch.py"
       - "litellm/llms/base_llm/ocr/**"
       - "litellm/llms/custom_httpx/llm_http_handler.py"
       - "Makefile"
@@ -68,6 +78,11 @@ concurrency:
 
 env:
   CARGO_TERM_COLOR: always
+  RUSTC_WRAPPER: sccache
+  SCCACHE_GHA_ENABLED: "true"
+  SCCACHE_GHA_VERSION: litellm-ci-v1
+  SCCACHE_GHA_RW_MODE: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' && 'READ_WRITE' || 'READ_ONLY' }}
+  CARGO_INCREMENTAL: "0"
 
 jobs:
   rust-lint:
@@ -91,6 +106,11 @@ jobs:
           cache-on-failure: true
           save-if: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Set up sccache
+        uses: mozilla-actions/sccache-action@d651010b8da762cde178750d8eda7b5febfe147a
+        with:
+          version: "v0.18.0"
+
       - run: cargo clippy --workspace --all-targets --locked -- -D warnings
 
   rust-test:
@@ -133,6 +153,11 @@ jobs:
           cache-on-failure: true
           save-if: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Set up sccache
+        uses: mozilla-actions/sccache-action@d651010b8da762cde178750d8eda7b5febfe147a
+        with:
+          version: "v0.18.0"
+
       - name: Check generated trace contracts
         working-directory: .
         run: uv run scripts/generate_trace_types.py --check
@@ -179,6 +204,11 @@ jobs:
           cache-on-failure: true
           save-if: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Set up sccache
+        uses: mozilla-actions/sccache-action@d651010b8da762cde178750d8eda7b5febfe147a
+        with:
+          version: "v0.18.0"
+
       - run: uv build --wheel --out-dir dist
 
       - run: python .github/scripts/verify_linux_native_wheel.py dist/*.whl

.github/workflows/test-terraform-provider.yml

@@ -12,6 +12,7 @@ on:
     paths:
       - "terraform/provider/**"
       - "litellm/proxy/**"
+      - "enterprise/**"
       - ".github/workflows/test-terraform-provider.yml"
 
 permissions:

.github/workflows/test-unit.yml

@@ -5,6 +5,9 @@ on:
     branches:
       - main
       - "litellm_**"
+      - "rc/**"
+      - "stable/**"
+      - "lts/**"
   push:
     branches:
       - main
@@ -42,7 +45,7 @@ jobs:
         id: rust-bridge-key
         if: steps.changes.outputs.decision != 'skip'
         env:
-          RUST_BRIDGE_CACHE_KEY: ${{ runner.os }}-rust-bridge-editable-${{ hashFiles('litellm-rust/Cargo.lock', 'litellm-rust/Cargo.toml', 'litellm-rust/crates/**', 'pyproject.toml') }}
+          RUST_BRIDGE_CACHE_KEY: ${{ runner.os }}-rust-bridge-editable-${{ hashFiles('litellm-rust/Cargo.lock', 'litellm-rust/Cargo.toml', 'litellm-rust/crates/**', 'model_prices_and_context_window.json', 'pyproject.toml') }}
         run: echo "key=$RUST_BRIDGE_CACHE_KEY" >> "$GITHUB_OUTPUT"
 
       - name: Set up Python
@@ -73,9 +76,21 @@ jobs:
         timeout-minutes: 5
         uses: ./.github/actions/cache-cargo-build
 
+      - name: Set up sccache
+        if: steps.changes.outputs.decision != 'skip' && steps.rust-bridge-cache.outputs.cache-hit != 'true'
+        uses: mozilla-actions/sccache-action@d651010b8da762cde178750d8eda7b5febfe147a
+        with:
+          version: "v0.18.0"
+
       - name: Build editable Rust bridge
         if: steps.changes.outputs.decision != 'skip' && steps.rust-bridge-cache.outputs.cache-hit != 'true'
         timeout-minutes: 15
+        env:
+          RUSTC_WRAPPER: sccache
+          SCCACHE_GHA_ENABLED: "true"
+          SCCACHE_GHA_VERSION: litellm-ci-v1
+          SCCACHE_GHA_RW_MODE: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' && 'READ_WRITE' || 'READ_ONLY' }}
+          CARGO_INCREMENTAL: "0"
         run: |
           mkdir -p rust-bridge-dist
           uv run --no-project --with maturin==1.15.0 python -c 'import maturin; maturin.build_editable("rust-bridge-dist")'
@@ -137,8 +152,10 @@ jobs:
         include:
           - shard: core-utils
             test-path: |-
+              tests/unit/accounting
               tests/unit/batch_completion
               tests/unit/decisions
+              tests/unit/extraction
               tests/unit/litellm_core_utils
               tests/unit/systemone
             python-version: "3.12"
@@ -215,14 +232,17 @@ jobs:
 
           - shard: llms-providers
             test-path: |-
+              tests/unit/llms/test_api_base_trailing_slash_routes.py
               tests/unit/llms/test_cache_control_and_reasoning.py
               tests/unit/llms/test_custom_llm.py
               tests/unit/llms/test_file_content_block.py
               tests/unit/llms/test_file_search_responses.py
+              tests/unit/llms/test_get_complete_url_trailing_slash.py
               tests/unit/llms/test_lifecycle_fix.py
               tests/unit/llms/test_oss_decision.py
               tests/unit/llms/test_polling_url_origin_match.py
               tests/unit/llms/test_predibase_transformation.py
+              tests/unit/llms/test_subscription_accounts.py
               tests/unit/llms/a2a
               tests/unit/llms/aiml
               tests/unit/llms/aiohttp_openai
@@ -232,6 +252,7 @@ jobs:
               tests/unit/llms/azure_ai
               tests/unit/llms/base_llm
               tests/unit/llms/baseten
+              tests/unit/llms/bigmodel
               tests/unit/llms/black_forest_labs
               tests/unit/llms/brave
               tests/unit/llms/bytez
@@ -240,7 +261,9 @@ jobs:
               tests/unit/llms/chatgpt
               tests/unit/llms/clarifai
               tests/unit/llms/claude_code
+              tests/unit/llms/clf_ai_gateway
               tests/unit/llms/cloudflare
+              tests/unit/llms/codestral
               tests/unit/llms/codex
               tests/unit/llms/cohere
               tests/unit/llms/cometapi
@@ -260,12 +283,14 @@ jobs:
               tests/unit/llms/e2b
               tests/unit/llms/edenai
               tests/unit/llms/elevenlabs
+              tests/unit/llms/empiriolabs
               tests/unit/llms/exa_ai
               tests/unit/llms/fal_ai
               tests/unit/llms/fastcrw
               tests/unit/llms/featherless_ai
               tests/unit/llms/firecrawl
               tests/unit/llms/fireworks_ai
+              tests/unit/llms/friendliai
               tests/unit/llms/gdc
               tests/unit/llms/gemini
               tests/unit/llms/gigachat
@@ -298,6 +323,7 @@ jobs:
               tests/unit/llms/modelscope
               tests/unit/llms/mongodb
               tests/unit/llms/moonshot
+              tests/unit/llms/morph
               tests/unit/llms/nadir
               tests/unit/llms/nebius
               tests/unit/llms/neosantara
@@ -314,6 +340,7 @@ jobs:
               tests/unit/llms/openai_like
               tests/unit/llms/opencode
               tests/unit/llms/opensandbox
+              tests/unit/llms/opensearch
               tests/unit/llms/openrouter
               tests/unit/llms/ovhcloud
               tests/unit/llms/parallel_ai
@@ -321,6 +348,7 @@ jobs:
               tests/unit/llms/pass_through
               tests/unit/llms/perplexity
               tests/unit/llms/pg_vector
+              tests/unit/llms/pi
               tests/unit/llms/publicai
               tests/unit/llms/ragflow
               tests/unit/llms/recraft
@@ -346,6 +374,7 @@ jobs:
               tests/unit/llms/tinyfish
               tests/unit/llms/together_ai
               tests/unit/llms/tool_loop
+              tests/unit/llms/topaz
               tests/unit/llms/triton
               tests/unit/llms/v0
               tests/unit/llms/valkey
@@ -390,6 +419,7 @@ jobs:
 
           - shard: router
             test-path: >-
+              tests/unit/router_rust
               tests/unit/test_router
             python-version: "3.12"
             workers: 4
@@ -409,15 +439,18 @@ jobs:
               tests/unit/embeddings
               tests/unit/endpoints
               tests/unit/files
+              tests/unit/fine_tuning
               tests/unit/images
               tests/unit/interactions
               tests/unit/messages
               tests/unit/ocr
               tests/unit/passthrough
               tests/unit/rag
+              tests/unit/cookbook
               tests/unit/realtime_api
               tests/unit/rerank_api
               tests/unit/search
+              tests/unit/text_completions
               tests/unit/vector_stores
               tests/unit/videos
             python-version: "3.12"
@@ -429,6 +462,13 @@ jobs:
 
           - shard: rust-bridge-harness
             test-path: >-
+              tests/test_litellm_rust/test_dispatch.py
+              tests/test_litellm_rust/test_public_entries.py
+              tests/test_litellm_rust/chat_completions/test_dispatch.py
+              tests/test_litellm_rust/embeddings/test_dispatch.py
+              tests/test_litellm_rust/messages/test_dispatch.py
+              tests/test_litellm_rust/ocr/test_dispatch.py
+              tests/test_litellm_rust/responses/test_dispatch.py
               tests/unit/compression
               tests/unit/harness
               tests/unit/rust_bridge
@@ -455,11 +495,11 @@ jobs:
               tests/unit/telemetry tests/unit/proxy/telemetry
               tests/unit/tracing
             python-version: "3.12"
-            workers: 2
+            workers: 4
             reruns: 0
             timeout-minutes: 20
             job-timeout-minutes: 60
-            dist: loadscope
+            dist: load
 
           - shard: proxy-auth
             test-path: |-
@@ -543,6 +583,7 @@ jobs:
               tests/unit/proxy/rerank_endpoints
               tests/unit/proxy/realtime_endpoints
               tests/unit/proxy/ui_crud_endpoints
+              tests/unit/proxy/deps
               tests/unit/proxy/config_resolvers
               tests/unit/proxy/utils
             python-version: "3.12"
@@ -564,17 +605,15 @@ jobs:
 
           - shard: mcp-elicitation
             test-path: |-
-              --override-ini=pythonpath=tests
-              tests/unit/experimental_mcp_client/test_mcp_client.py
-              tests/unit/proxy/_experimental/mcp_server/test_capabilities.py
-              tests/unit/proxy/_experimental/mcp_server/test_interactions.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_elicitation_handler.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_server_manager.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_server_tool_calls_and_headers.py
-              tests/unit/proxy/_experimental/mcp_server/test_operations.py
-              tests/integration/mcp/test_interactions.py
+              tests/unit/mcp_client/test_mcp_client.py
+              tests/unit/proxy/mcp_server/test_capabilities.py
+              tests/unit/proxy/mcp_server/test_interactions.py
+              tests/unit/proxy/mcp_server/test_mcp_elicitation_handler.py
+              tests/unit/proxy/mcp_server/test_mcp_server_manager.py
+              tests/unit/proxy/mcp_server/test_mcp_server_tool_calls_and_headers.py
+              tests/unit/proxy/mcp_server/test_operations.py
             python-version: "3.12"
-            workers: 2
+            workers: 4
             reruns: 0
             timeout-minutes: 20
             job-timeout-minutes: 60
@@ -590,7 +629,7 @@ jobs:
               tests/unit/proxy/spend_tracking
               --ignore=tests/unit/proxy/spend_tracking/test_search_api_logging.py
               tests/unit/proxy/pass_through_endpoints
-              tests/unit/proxy/experimental
+              tests/unit/proxy/mcp_server/test_tool_registry.py
               tests/unit/proxy/common_utils
               --ignore=tests/unit/proxy/common_utils/test_cache_aware_routing.py
               --ignore=tests/unit/proxy/common_utils/test_check_batch_cost.py
@@ -659,6 +698,7 @@ jobs:
           - shard: proxy-extras
             test-path: >-
               tests/unit/litellm_proxy_extras
+              tests/unit/db_scripts
             python-version: "3.12"
             workers: 2
             reruns: 0
@@ -697,7 +737,6 @@ jobs:
               tests/unit/responses/litellm_completion_transformation
               tests/unit/responses/test_additional_tools.py
               tests/unit/responses/test_custom_tool_call.py
-              tests/unit/responses/test_dispatch.py
               tests/unit/responses/test_metadata_codex_callback.py
               tests/unit/responses/test_no_duplicate_spend_logs.py
               tests/unit/responses/test_null_test_fix.py
@@ -798,14 +837,14 @@ jobs:
 
           - shard: mcp-oauth
             test-path: |-
-              tests/unit/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_server_manager.py
-              tests/unit/proxy/_experimental/mcp_server/test_db_credentials.py
-              tests/unit/proxy/_experimental/mcp_server/test_gateway_dcr_flow.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_oauth_passthrough_tools.py
-              tests/unit/proxy/_experimental/mcp_server/test_mcp_max_concurrent_requests.py
-              tests/unit/proxy/_experimental/mcp_server/faults/test_list_outcomes.py
-              tests/unit/proxy/_experimental/mcp_server/outbound_credentials
+              tests/unit/proxy/mcp_server/test_discoverable_endpoints.py
+              tests/unit/proxy/mcp_server/test_mcp_server_manager.py
+              tests/unit/proxy/mcp_server/test_db_credentials.py
+              tests/unit/proxy/mcp_server/test_gateway_dcr_flow.py
+              tests/unit/proxy/mcp_server/test_mcp_oauth_passthrough_tools.py
+              tests/unit/proxy/mcp_server/test_mcp_max_concurrent_requests.py
+              tests/unit/proxy/mcp_server/faults/test_list_outcomes.py
+              tests/unit/proxy/mcp_server/outbound_credentials
             python-version: "3.12"
             workers: 4
             reruns: 2
@@ -815,19 +854,21 @@ jobs:
 
           - shard: mcp-server
             test-path: |-
+              tests/unit/mcp_client
               tests/unit/experimental_mcp_client
+              tests/unit/proxy/mcp_server
               tests/unit/proxy/_experimental/mcp_server
               tests/unit/responses/mcp
-              --ignore=tests/unit/experimental_mcp_client/test_mcp_client.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_capabilities.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_interactions.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_mcp_elicitation_handler.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_mcp_server_manager.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_mcp_server_tool_calls_and_headers.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_operations.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_discoverable_endpoints.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/test_db_credentials.py
-              --ignore=tests/unit/proxy/_experimental/mcp_server/outbound_credentials
+              --ignore=tests/unit/mcp_client/test_mcp_client.py
+              --ignore=tests/unit/proxy/mcp_server/test_capabilities.py
+              --ignore=tests/unit/proxy/mcp_server/test_interactions.py
+              --ignore=tests/unit/proxy/mcp_server/test_mcp_elicitation_handler.py
+              --ignore=tests/unit/proxy/mcp_server/test_mcp_server_manager.py
+              --ignore=tests/unit/proxy/mcp_server/test_mcp_server_tool_calls_and_headers.py
+              --ignore=tests/unit/proxy/mcp_server/test_operations.py
+              --ignore=tests/unit/proxy/mcp_server/test_discoverable_endpoints.py
+              --ignore=tests/unit/proxy/mcp_server/test_db_credentials.py
+              --ignore=tests/unit/proxy/mcp_server/outbound_credentials
             python-version: "3.12"
             workers: 4
             reruns: 0
@@ -940,7 +981,7 @@ jobs:
         timeout-minutes: 3
         with:
           persist-credentials: false
-          fetch-depth: ${{ matrix.shard == 'db-and-spend' && '0' || '1' }}
+          fetch-depth: 1
 
       - name: Detect relevant changes
         id: changes
@@ -973,27 +1014,37 @@ jobs:
         with:
           artifact: ${{ needs.rust-bridge.outputs.artifact }}
 
-      - name: Install dependencies and check schema
+      - name: Install dependencies
         if: steps.changes.outputs.decision != 'skip'
-        timeout-minutes: 8
+        timeout-minutes: 5
         env:
           RUST_BRIDGE_ARTIFACT: ${{ needs.rust-bridge.outputs.artifact }}
-          UNIT_SHARD: ${{ matrix.shard }}
-          SCHEMA_BASE_REF: origin/${{ github.event.repository.default_branch }}
         run: |
           diff -u model_prices_and_context_window.json litellm/model_prices_and_context_window_backup.json
           if [ -z "$RUST_BRIDGE_ARTIFACT" ]; then
-            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra caching --extra extra_proxy --extra proxy-runtime --extra utils
+            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra extra_proxy --extra proxy-runtime --extra utils
           else
-            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra caching --extra extra_proxy --extra proxy-runtime --extra utils --no-install-project
+            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra extra_proxy --extra proxy-runtime --extra utils --no-install-project
             uv pip install --no-deps --python .venv/bin/python rust-bridge-dist/*.whl
             cp rust-bridge-dist/litellm/rust_bridge/_native.abi3.so litellm/rust_bridge/_native.abi3.so
             uv run --no-sync python -c "import importlib.metadata; import litellm.rust_bridge._native; print(importlib.metadata.version('litellm'))"
           fi
           uv run --no-sync python -c 'import os, sys; print(sys.version); assert f"{sys.version_info.major}.{sys.version_info.minor}" == os.environ["UV_PYTHON"]'
-          if [ "$UNIT_SHARD" = "db-and-spend" ]; then
-            uv run --no-sync python scripts/check_schema_compatibility.py --base-ref "$SCHEMA_BASE_REF"
-          fi
+
+      - name: Check schema.prisma against the merge base
+        if: steps.changes.outputs.decision != 'skip' && matrix.shard == 'db-and-spend' && github.event_name == 'pull_request'
+        timeout-minutes: 3
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          EVENT_BASE_REF: ${{ github.event.pull_request.base.ref }}
+          HEAD_SHA: ${{ github.event.pull_request.head.sha }}
+        run: |
+          base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/branches/${EVENT_BASE_REF}" --jq .name) || base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" --jq .base.ref)
+          merge_base=$(gh api "repos/${GITHUB_REPOSITORY}/compare/${base_ref}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
+          test -n "$merge_base"
+          git fetch --no-tags --depth=1 origin "$merge_base"
+          uv run --no-sync python scripts/check_schema_compatibility.py --base-commit "$merge_base"
 
       - name: Cache Prisma binaries
         if: steps.changes.outputs.decision != 'skip'
@@ -1130,18 +1181,20 @@ jobs:
         env:
           CI: "true"
           GH_TOKEN: ${{ github.token }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          EVENT_BASE_REF: ${{ github.event.pull_request.base.ref }}
           HEAD_SHA: ${{ github.event.pull_request.head.sha }}
         run: |
           full_suite() { npm run test -- --run --pool forks --maxWorkers=14; }
 
-          if [ -z "$BASE_SHA" ]; then
+          if [ -z "$PR_NUMBER" ]; then
             echo "Push to $GITHUB_REF_NAME: running the full suite"
             full_suite
             exit 0
           fi
 
-          merge_base=$(gh api "repos/${{ github.repository }}/compare/${BASE_SHA}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
+          base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/branches/${EVENT_BASE_REF}" --jq .name) || base_ref=$(gh api "repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}" --jq .base.ref)
+          merge_base=$(gh api "repos/${GITHUB_REPOSITORY}/compare/${base_ref}...${HEAD_SHA}?per_page=1" --jq '.merge_base_commit.sha')
           test -n "$merge_base"
           git fetch --no-tags --depth=1 origin "$merge_base" "$HEAD_SHA"
           changed_files=()
@@ -1290,6 +1343,8 @@ jobs:
           done
   openai-sdk-compat:
     name: OpenAI ${{ matrix.openai-version }} compatibility
+    needs: rust-bridge
+    if: ${{ !cancelled() }}
     permissions:
       contents: read
       pull-requests: read
@@ -1326,10 +1381,18 @@ jobs:
         with:
           version: "0.10.9"
 
+      - name: Set up the Rust bridge
+        if: steps.changes.outputs.decision != 'skip'
+        timeout-minutes: 5
+        uses: ./.github/actions/rust-bridge
+        with:
+          artifact: ${{ needs.rust-bridge.outputs.artifact }}
+
       - name: Install dependencies
         if: steps.changes.outputs.decision != 'skip'
         run: |
           .github/scripts/uv_sync_with_retries.sh --frozen --group dev --group proxy-dev --extra proxy --no-install-workspace
+          cp rust-bridge-dist/litellm/rust_bridge/_native.* litellm/rust_bridge/
           uv pip install "openai==${OPENAI_VERSION:?}" --exclude-newer-package openai=2026-10-07T00:00:00Z
           uv run --no-sync python -c 'import os, openai; assert openai.__version__ == os.environ["OPENAI_VERSION"], openai.__version__'
 
@@ -1421,9 +1484,9 @@ jobs:
           RUST_BRIDGE_ARTIFACT: ${{ needs.rust-bridge.outputs.artifact }}
         run: |
           if [ -z "$RUST_BRIDGE_ARTIFACT" ]; then
-            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra caching --extra extra_proxy --extra proxy-runtime --extra utils
+            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra extra_proxy --extra proxy-runtime --extra utils
           else
-            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra caching --extra extra_proxy --extra proxy-runtime --extra utils --no-install-project
+            .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra google --extra proxy --extra semantic-router --extra saml --extra extra_proxy --extra proxy-runtime --extra utils --no-install-project
             uv pip install --no-deps --python .venv/bin/python rust-bridge-dist/*.whl
             cp rust-bridge-dist/litellm/rust_bridge/_native.abi3.so litellm/rust_bridge/_native.abi3.so
           fi
@@ -1452,7 +1515,6 @@ jobs:
             tests/unit/caching/test_redis_connection_pool.py \
             tests/unit/caching/test_redis_cluster_cache.py \
             tests/unit/caching/test_evicted_client_closer.py \
-            tests/integration/sdk/test_redis_cluster_iam_auth.py \
             --tb=short -vv -rs -p no:rerunfailures -p no:pytest-retry --timeout=120 --durations=20 \
             --cov=./litellm --cov-report=xml:coverage.xml
 

.github/workflows/weekly_load_anomaly.yml

@@ -1,85 +0,0 @@
-name: "Weekly Load Anomaly Check"
-
-on:
-  schedule:
-    - cron: "0 12 * * 6"
-  workflow_dispatch:
-
-permissions:
-  contents: read
-
-jobs:
-  weekly-load-anomaly:
-    if: github.event_name != 'schedule' || github.repository == 'BerriAI/litellm'
-    runs-on: ubuntu-latest
-    timeout-minutes: 45
-    services:
-      postgres:
-        image: postgres:16.6
-        env:
-          POSTGRES_USER: llmproxy
-          POSTGRES_PASSWORD: dbpassword9090
-          POSTGRES_DB: litellm
-        ports:
-          - 5432:5432
-        options: >-
-          --health-cmd "pg_isready -U llmproxy"
-          --health-interval 5s
-          --health-timeout 5s
-          --health-retries 10
-    env:
-      DATABASE_URL: postgresql://llmproxy:dbpassword9090@localhost:5432/litellm
-      LITELLM_MASTER_KEY: sk-weekly-anomaly-check
-      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
-      AWS_BEARER_TOKEN_BEDROCK: ${{ secrets.AWS_BEARER_TOKEN_BEDROCK }}
-    steps:
-      - uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          persist-credentials: false
-
-      - name: Set up Python
-        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-        with:
-          python-version: "3.12"
-
-      - name: Set up uv
-        uses: ./.github/actions/setup-uv-with-retries
-        with:
-          version: "0.10.9"
-
-      - name: Cache the Rust build
-        uses: ./.github/actions/cache-cargo-build
-
-      - name: Install dependencies
-        run: |
-          .github/scripts/uv_sync_with_retries.sh --frozen --group ci --group proxy-dev --extra proxy
-
-      - name: Cache Prisma binaries
-        uses: ./.github/actions/cache-prisma-binaries
-
-      - name: Generate Prisma client
-        run: |
-          uv run --no-sync prisma generate --schema litellm/proxy/schema.prisma
-
-      - name: Start the proxy
-        run: |
-          nohup uv run --no-sync litellm --config tests/e2e/perf/weekly_anomaly_config.yml --port 4000 > proxy.log 2>&1 &
-          for _ in $(seq 1 90); do
-            if curl -fs http://localhost:4000/health/liveliness > /dev/null; then
-              exit 0
-            fi
-            sleep 2
-          done
-          echo "proxy never became live"
-          tail -n 100 proxy.log
-          exit 1
-
-      - name: Run the weekly session anomaly test
-        env:
-          E2E_WEEKLY_ANOMALY: "1"
-        run: |
-          uv run --no-sync pytest tests/e2e/perf/test_weekly_session_anomaly_e2e.py -v --tb=short -rA
-
-      - name: Show proxy log on failure
-        if: failure()
-        run: tail -n 300 proxy.log

.github/workflows/zizmor.yml

@@ -27,10 +27,15 @@ jobs:
         with:
           persist-credentials: false
 
+      - name: Install zizmor
+        run: |
+          curl -fsSL --retry 3 -o "$RUNNER_TEMP/zizmor.tar.gz" \
+            https://github.com/zizmorcore/zizmor/releases/download/v1.31.1/zizmor-x86_64-unknown-linux-gnu.tar.gz
+          echo "d185f726bdaf92df78273217dfdd24ebcd6b0ef48cf3ce776ff5762b6e138c99  $RUNNER_TEMP/zizmor.tar.gz" | sha256sum -c -
+          tar xzf "$RUNNER_TEMP/zizmor.tar.gz" -C "$RUNNER_TEMP" zizmor
+
       - name: Run zizmor
-        uses: zizmorcore/zizmor-action@5f14fd08f7cf1cb1609c1e344975f152c7ee938d # v0.5.6
-        with:
-          version: "1.24.1"
-          min-severity: medium
-          advanced-security: false
-          annotations: true
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          "$RUNNER_TEMP/zizmor" --no-ignores --no-config --min-severity=medium --format=github .

.gitignore

@@ -105,7 +105,7 @@ litellm_config.yaml
 .cursor
 litellm/proxy/to_delete_loadtest_work/*
 update_model_cost_map.py
-tests/unit/proxy/_experimental/mcp_server/test_mcp_server_manager.py
+tests/unit/proxy/mcp_server/test_mcp_server_manager.py
 scripts/test_vertex_ai_search.py
 LAZY_LOADING_IMPROVEMENTS.md
 STABILIZATION_TODO.md

.grype.yaml

@@ -11,3 +11,12 @@ ignore:
     package:
       name: zlib
       type: apk
+  - vulnerability: GHSA-v4x8-3wr5-3wq8
+    reason: >-
+      filelock ReadWriteLock advisory, and no package in the image takes ReadWriteLock;
+      the fix (4.2.1) is blocked by litellm's own filelock<4.0 cap until #48635 or #48802 lifts it,
+      and the image then moves off 3.32.7, which retires this version-scoped rule
+    package:
+      name: filelock
+      version: 3.32.7
+      type: python

AGENTS.md

@@ -29,7 +29,9 @@ A test must only fail when litellm code changes. Never pin facts we don't own (a
 
 `tests/unit/` mirrors `litellm/` in a parallel path (see `tests/unit/AGENTS.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
 
-End-to-end tests belong in `tests/e2e/` and must follow the harness conventions documented in that directory's `AGENTS.md`
+End-to-end tests belong in `tests/e2e/`. A rule about one directory's files, like its harness or which CI runs it, goes in that directory's `AGENTS.md`, never only in a README or here. Read a directory's `AGENTS.md` before editing files in it
+
+Before you build a new way to move data between the endpoints, the provider transformations, and cost tracking or logging, read the architecture decision records in `docs/adrs/`, which say why those layers are shaped the way they are
 
 When creating PRs, target the repository's current default branch for both internal and external / OSS contributions. Check it with `python3 scripts/default_branch.py --branch` instead of assuming a branch name or relying on cached `origin/HEAD`
 
@@ -39,7 +41,11 @@ Same applies for filing bug reports and feature requests, with .github/ISSUE_TEM
 
 If you're resolving a linear ticket, in the "## Linear ticket" section of the PR, say "Resolves LIT-1234", replacing "LIT-1234" with the actual ticket id that you're resolving. If you don't have the ticket id, don't make one up or search for it. Just drop the section
 
-Never use `pytest` commands or the like as "Screenshots / Proof of Fix". We prefer curl'ing a live proxy instance running on localhost:4000 (I like to run it with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`. The Admin UI dev server is `npm run dev` in `ui/litellm-dashboard`, served on port 3000) and showing both the command run and the output. Also, it should hit real LLM provider APIs, not mocks, and cost real $$$ because that is the most realistic test. The proof of fix should be exactly what the end user / customer would see / do. The run logs in PR #27703 is a prime example of how to do it (not a huge fan of using a python test script that future me and the team will have no visibility into. I prefer just curl commands or a short list of bash commands (e.g., using `for`)). If it's a UI thing, or the main use case runs through a headful agentic coding tool like Claude Code or Codex, drive that surface yourself and embed your own before and after screenshots of it in the PR (the Admin UI page, or what the coding tool shows), next to an ordered list of the URLs to go to (e.g., http://localhost:4000/ui/?page=logs), where to click, and what fields to fill out so a reviewer can reproduce it
+Run QA proxies with `python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --num_workers 2 --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log`, and swap its `--config` for a copy of `litellm/proxy/dev_config.yaml` that points both `router_settings` (`redis_host`, `redis_port`) and `general_settings.coordination_redis` (`host`, `port`) at one dedicated Redis instance, with credentials when required, so the workers share router state and spend, budget, and rate-limit counters. Check that two worker PIDs are serving, and use the same worker count and Redis setup for the before and after proof. When the PR is about `--run_gunicorn`, `--run_hypercorn`, or `--run_granian`, which `--reload` does not support, drop `--reload`. When it is about behavior without Redis, drop Redis. When the proxy prints `"workers" flag is ignored when reloading is enabled`, that checkout or its uvicorn reloads only one worker, so drop the two-worker check for it. Say in the proof which part you dropped and why
+
+Simple PRs may skip before/after e2e when the full diff is mechanical and focused checks completely prove it, such as #49030 restoring an existing helper name. Trace affected callers and run relevant tests/static checks at the tip. For a bug fix, establish that the regression catches the original defect. Size alone never qualifies. New behavior, API/provider semantics, auth, billing, persistence, concurrency or uncertain integration risk need live e2e. In Screenshots / Proof of Fix, record `E2E exception at <tip SHA>`, scope, rationale, commands and results. Tests count as proof only for this exception. Reassess changed diffs, retain CI/bot gates, and honor explicit e2e requests (2026-10-28, unnecessary live QA for #49030)
+
+For other PRs, proof must show the end user's real flow before and after, with commands and output against a live proxy and real provider calls where applicable. Pytest alone is insufficient. For UI or interactive coding-tool flows, drive that surface and embed before/after screenshots, with URLs, clicks and input fields a reviewer can reproduce
 
 If you ever write any human-facing text (pull requests, issues, commit messages, discussion posts, github comments, release notes, docs, etc.), always follow these guidelines to sound less AI-y:
 - don't use emojis
@@ -72,7 +78,7 @@ When referencing or running models (coding, QA'ing, writing docs, writing tests,
 
 Always pull before starting any work. The checkout or worktree may be sitting on a stale branch
 
-If you're an internal contributor, when creating a new PR, the typical flow is to branch off the repository's current default branch and create a branch prefixed with litellm_. Do not create a branch prefixed with claude/ and generally do not have / in your branch names
+Internal contributors branch a new PR off the repository's current default branch, with a branch name prefixed with litellm_
 
 Do not add `Co-Authored-By: Claude` or any Claude attribution to commit messages. Never use a `claude/` prefix or put a `/` in a branch name. Do not add "Generated with Claude Code" (or any similar attribution) to PR descriptions or comments. Do not create a new PR/branch off the existing PR to fix/add something that is related and could've just been committed directly to the existing PR's branch
 
@@ -86,6 +92,8 @@ Do not put names of customers or customer company names in code, PR descriptions
 
 CI supply-chain safety: Never pipe a remote script into a shell (`curl ... | bash`, `wget ... | sh`). Download the artifact to a file, verify its SHA-256 checksum, then install. Pin every external tool to a specific version with a full URL (not `latest` or `stable`). Verify checksums for all downloaded binaries, using the provider's official `.sha256` / `.sha256sum` sidecar when available. These rules apply to every download in CI
 
+For database operations, connection ownership and proxy dependency wiring, follow [database-access](.agents/skills/database-access/SKILL.md)
+
 Prisma migrations apply synchronously at proxy boot, before it serves traffic, so a migration must only change schema, never rewrite rows. No `UPDATE`, `DELETE` or `MERGE`, and no `INSERT ... SELECT`: on a spend-log-sized table any of those is minutes of downtime plus a doubled heap that plain autovacuum won't give back. `tests/code_coverage_tests/check_migrations_no_data_rewrites.py` enforces this. When a rewrite is genuinely bounded and has to ship inside the migration, mark the statement `-- data-migration-ok: <what bounds it>`
 
 Follow these coding conventions for new/updated code (a three-line fix in a legacy file shouldn't trigger huge drive-by refactors):
@@ -105,6 +113,7 @@ Follow these coding conventions for new/updated code (a three-line fix in a lega
 - No file sprawl: deliberate file and folder structure
 - Standard over hand-rolled: use the official SDK or a library where one exists. Where none does, follow industry standards instead of inventing local conventions
 - API-fragmentation-aware: when logic must branch on which API surface produced or consumes data (e.g. chat completions vs Anthropic Messages vs Responses API shapes), proactively look for an existing shared helper (e.g. `litellm_core_utils/prompt_templates/factory.py`) before writing per-surface parsing in the new module. If none exists, add one there instead of duplicating the same format-detection logic in every new guardrail/integration
+- Current default for new model names: logic that picks behavior from a model-name table (tokenizer, message framing, supported params) gives a provider's model the table does not list yet that provider's current behavior, never its oldest, since providers ship names faster than any table is updated
 
 Follow conventional commits for commit names and PR titles
 

ARCHITECTURE.md

@@ -2,6 +2,8 @@
 
 This document helps contributors understand where to make changes in LiteLLM.
 
+For why a layer is shaped the way it is, read the architecture decision records in [`docs/adrs/`](docs/adrs/AGENTS.md) (Python SDK and proxy) and [`litellm-rust/docs/adrs/`](litellm-rust/docs/adrs/AGENTS.md) (Rust core)
+
 ---
 
 ## How It Works
@@ -239,6 +241,8 @@ graph LR
 7. `DBSpendUpdateWriter.update_database()` queues spend increments to Redis
 8. Background job `update_spend` flushes queued spend to PostgreSQL every 60s
 
+Cost lookup falls back to the response model when the provider-reported name cannot be priced. Explicitly registered zero rates remain zero, and custom deployment pricing or an explicit `base_model` retain precedence
+
 ### Data Access Layer (Models & Repositories)
 
 Database entities and the operations on them live in two packages at the root of `litellm/` so both the gateway (`proxy/`) and the SDK can use them without importing proxy internals:
@@ -332,13 +336,13 @@ Each translation is isolated in its own file, making it easy to test and modify
 |--------------|----------|------------------|
 | `/v1/chat/completions` | Anthropic | `llms/anthropic/chat/transformation.py` |
 | `/v1/chat/completions` | Bedrock Converse | `llms/bedrock/chat/converse_transformation.py` |
-| `/v1/chat/completions` | Bedrock Invoke | `llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py` |
+| `/v1/chat/completions` | Bedrock Invoke | `llms/bedrock/chat/invoke_transformations/anthropic_claude_transformation.py` |
 | `/v1/chat/completions` | Gemini | `llms/gemini/chat/transformation.py` |
 | `/v1/chat/completions` | Vertex AI | `llms/vertex_ai/gemini/transformation.py` |
 | `/v1/chat/completions` | OpenAI | `llms/openai/chat/gpt_transformation.py` |
-| `/v1/messages` (passthrough) | Anthropic | `llms/anthropic/pass_through/messages/transformation.py` |
-| `/v1/messages` (passthrough) | Bedrock | `llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py` |
-| `/v1/messages` (passthrough) | Vertex AI | `llms/vertex_ai/vertex_ai_partner_models/anthropic/experimental_pass_through/transformation.py` |
+| `/v1/messages` (passthrough) | Anthropic | `llms/anthropic/messages/native/transformation.py` |
+| `/v1/messages` (passthrough) | Bedrock | `llms/bedrock/messages/invoke_transformations/anthropic_claude_transformation.py` |
+| `/v1/messages` (passthrough) | Vertex AI | `llms/vertex_ai/vertex_ai_partner_models/anthropic/messages/transformation.py` |
 | Passthrough endpoints | All | `proxy/pass_through_endpoints/llm_provider_handlers/` |
 
 ### Example: Debugging prompt caching
@@ -346,7 +350,7 @@ Each translation is isolated in its own file, making it easy to test and modify
 If `/v1/messages` → Bedrock Converse prompt caching isn't working but Bedrock Invoke works:
 
 1. **Bedrock Converse translation**: `llms/bedrock/chat/converse_transformation.py`
-2. **Bedrock Invoke translation**: `llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py`
+2. **Bedrock Invoke translation**: `llms/bedrock/chat/invoke_transformations/anthropic_claude_transformation.py`
 3. Compare how each handles `cache_control` in `transform_request()`
 
 ### How translations work

CONTRIBUTING.md

@@ -8,6 +8,7 @@ Here are the core requirements for any PR submitted to LiteLLM:
 
 - [ ] **Sign the Contributor License Agreement (CLA)** - [see details](#contributor-license-agreement-cla)
 - [ ] **Keep scope isolated** - Your changes should address 1 specific problem at a time
+- [ ] **Prove it works end to end** - If your change affects what the proxy or SDK does, paste the commands you ran against a live proxy and their output in the PR - [see details](#proving-your-change-works-end-to-end)
 
 #### Proxy (Backend) PRs
 
@@ -39,12 +40,12 @@ git clone https://github.com/YOUR_USERNAME/litellm.git
 cd litellm
 
 # Create a new branch for your feature (see "Commit and Branch Conventions" below)
-git checkout -b feature/your-feature
+git checkout -b litellm_your_feature
 
 # Install development dependencies
 make install-dev
 
-# Install git hooks that enforce commit + branch conventions (one-time, opt-in)
+# Install git hooks that enforce commit conventions (one-time, opt-in)
 make install-hooks
 
 # Verify your setup works
@@ -55,7 +56,7 @@ That's it! Your local development environment is ready.
 
 ## Commit and Branch Conventions
 
-Commits follow [Conventional Commits](https://www.conventionalcommits.org/en/v1.0.0/) and branches follow [Conventional Branches](https://conventional-branch.github.io/). Run `make install-hooks` once per clone to enable the local git hooks that enforce these — see the [contributor docs](https://docs.litellm.ai/docs/extras/contributing_code#commit-and-branch-conventions) for the full type list, examples, the protected-branch bypass list, and how to opt out.
+Commits follow [Conventional Commits](https://www.conventionalcommits.org/en/v1.0.0/). Internal contributors name branches with the `litellm_` prefix and no `/`, and fork branches can use any name. Run `make install-hooks` once per clone to enable the local hook that enforces the commit convention. See the [contributor docs](https://docs.litellm.ai/docs/extras/contributing_code#commit-and-branch-conventions) for details
 
 ### 2. Development Workflow
 
@@ -78,8 +79,8 @@ uv run pytest tests/unit/<your_test_file>.py -v
 git add .
 git commit -m "feat(scope): your descriptive commit message"
 
-# Push and create a PR (branch must follow Conventional Branches — see above)
-git push origin feature/your-feature
+# Push and create a PR
+git push origin litellm_your_feature
 ```
 
 ## Adding Testing
@@ -94,6 +95,8 @@ Add your tests to the [`tests/unit/` directory](https://github.com/BerriAI/litel
 - **Only add mocked tests** - no real LLM API calls in this directory
 - For integration tests with real APIs, use the appropriate test directories
 
+Tests that call `litellm.register_model` or change `litellm.model_cost` should use the `local_model_cost_map` fixture from `tests/unit/conftest.py`. It provides a fresh catalog and restores the previous map, so pricing and capability overrides cannot leak into later tests
+
 ### File Naming Convention
 
 The `tests/unit/` directory follows the same structure as `litellm/`:
@@ -246,6 +249,43 @@ If `make test-unit` fails:
 - **Test edge cases**: Don't just test the happy path
 - **Update documentation**: If you change APIs, update docs
 
+## Proving Your Change Works End to End
+
+Tests are required, but they do not prove that your change works for a user. A PR that changes what the proxy or SDK does shows a real run in its "Screenshots / Proof of Fix" section, meaning the exact commands you sent to a live proxy and the output that came back, against real provider APIs and not mocks. `pytest` output does not count as proof, because a passing test only shows that the code does what its mocks expect. A PR that changes only docs or CI shows the rendered page or the CI run instead
+
+A good proof looks like this:
+
+1. Install the proxy dependencies with `make install-proxy-dev`, then start the proxy from your branch with a config that has the model you changed, `uv run litellm --config your_config.yaml --detailed_debug`
+2. Send the request a user would send with `curl`, and paste both the command and the response. When cost or routing is involved, include the response headers, since `x-litellm-response-cost` is what the user sees
+3. For a bug fix, do this twice, once at the merge base with BerriAI/litellm's main (`git merge-base origin/main HEAD`, with your remote for BerriAI/litellm in place of `origin` in a fork clone) to show the bug and once at your branch's latest commit to show the fix. Name both commit hashes
+4. If your change is reachable from more than one endpoint (`/v1/chat/completions`, `/v1/responses`, `/v1/messages`), show each one
+5. If the request can stream, show a `"stream": true` run too, because streaming builds usage and cost on a different path. A stream sends its headers before its usage exists, so show the `usage.cost` on the last chunk, which `/v1/chat/completions` sends when the request has `"stream_options": {"include_usage": true}` and other endpoints send with `include_cost_in_streaming_usage: true` under `litellm_settings`, or, when the proxy has a database, the request's row on the Admin UI's Logs page or `GET /spend/logs`
+6. For UI changes, add before and after screenshots and name the page
+7. If users reach your change through a coding tool such as Claude Code or Codex, drive that tool the way a user does and add before and after screenshots of it, because a `curl` or a headless run alone does not show what the user sees
+
+```bash
+curl -sD - http://localhost:4000/v1/chat/completions \
+  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "your-model", "messages": [{"role": "user", "content": "hi"}]}'
+```
+
+When the provider reports its own usage or spend numbers, compare the proxy's numbers against them in the proof
+
+## Contributing an LLM Provider Integration
+
+Provider work (a new provider, a new parameter, a new usage or cost field) is the most common kind of contribution, and it is where PRs get reworked most often. Follow these rules
+
+**Preserve zero prices.** In [`model_prices_and_context_window.json`](model_prices_and_context_window.json), an explicit `0.0` means that the priced operation is free. A missing field or `null` means no rate is specified, so the relevant calculator may use a fallback. For example, when no batch input rate is configured, batch cache writes use `cache_creation_input_token_cost` if set, otherwise the regular input rate, then apply the batch discount. A zero ordinary input rate does not make separately priced cache reads or writes free. Test `rate is None` before substituting a fallback, since `row.get("cache_creation_input_token_cost") or fallback` would discard a valid zero. Speech keeps character pricing ahead of token pricing even when the character rate is zero. If neither rate is configured, speech cost calculation raises an error. Its model lookup preserves unset rates instead of treating compatibility defaults as explicit zero prices. Perplexity keeps a configured legacy search rate ahead of the current search rate, including zero. When registering custom prices, `None` leaves an existing rate unchanged and `0.0` overrides it
+
+**Find the existing machinery before you add plumbing.** LiteLLM already has generic paths that move data between the endpoints, the provider transformations, and cost tracking. A second, hand-rolled path for the same data is a common reason a provider PR gets rewritten before merge. [`ARCHITECTURE.md`](ARCHITECTURE.md) shows where the layers live, and the decision records in [`docs/adrs/`](docs/adrs/AGENTS.md) explain why they are shaped that way. For example, [the ADR on provider usage fields](docs/adrs/adr_001_provider_usage_fields_ride_as_usage_extras.md) covers how provider-specific usage fields reach cost tracking. If you cannot find a mechanism for what you need, say so in the PR or ask in [#pr-review on Slack](https://join.slack.com/t/litellmossslack/shared_invite/zt-3o7nkuyfr-p_kbNJj8taRfXGgQI1~YyA) before you build one
+
+**Bill a new provider usage field from the usage object.** Leave the field on the usage object as an extra and read it in that provider's `cost_calculator.py`. A provider without one gets a new `cost_calculator.py` and an arm for it in `cost_per_token` (`litellm/cost_calculator.py`), which picks the calculator by provider. If the provider's transform builds `Usage` from named fields, as the Anthropic, Gemini, and Bedrock Converse ones do, pass the new field there too. Streamed requests rebuild usage in `ChunkProcessor.calculate_usage` (`litellm/litellm_core_utils/streaming_chunk_builder_utils.py`), which does not copy extras yet the way the Responses-to-chat bridge does (minus a short exclusion list), so until it does, bill the field on streams by naming it there, as Anthropic's `inference_geo` is. A provider that reports its own charge for the request, as OpenRouter does in `usage.cost`, is billed at that charge from `_hidden_params` before any calculator runs, unless the deployment sets its own prices. When the field counts a built-in tool such as web search, add the provider to `get_cost_for_web_search_request` in `litellm/llms/__init__.py`, and add a gate to `litellm/litellm_core_utils/llm_cost_calc/tool_call_cost_tracking.py` only when no existing gate fires. Then show the field's cost in your proof on every endpoint it reaches, streaming included
+
+**Do not change what a caller sees to make internals easier.** Each endpoint keeps the schema of the API it emulates. A `/v1/responses` caller gets `input_tokens` and a `/v1/chat/completions` caller gets `prompt_tokens`, whatever the provider sent and whatever cost tracking needs. A provider usage field that shows up as an extra key is fine, since it only adds a key, but renaming, removing, or reshaping a key the endpoint defines is not. Reshaping a public response to feed an internal consumer breaks every user of that endpoint
+
+If your change makes an architecture decision, such as a new cross-cutting mechanism or a deliberate difference from a provider's API shape, add an ADR to [`docs/adrs/`](docs/adrs/AGENTS.md) in the same PR
+
 ## Building and Running Locally
 
 ### LiteLLM Proxy Server

Dockerfile

@@ -1,10 +1,10 @@
 # syntax=docker/dockerfile:1.7
 
 # Base image for building
-ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 
 # Runtime image
-ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
 # Pinned by digest like the other base images; bump explicitly on Node upgrades.
 ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@@ -61,6 +61,7 @@ RUN apk add --no-cache \
     python-3.13-dev \
     rust \
     openssl \
+    libssl3 \
     openssl-dev \
     nodejs \
     npm \
@@ -120,8 +121,8 @@ RUN sed -i 's/\r$//' docker/entrypoint.sh && chmod +x docker/entrypoint.sh && \
 FROM $LITELLM_BUILD_IMAGE AS liteadmin-builder
 COPY --from=uvbin /uv /usr/local/bin/uv
 RUN apk add --no-cache python-3.13
-ADD --checksum=sha256:2f7ae5cdd9d91731c0990e74a58239dc3e3fd2bf28dab23b55eafcdc47aaf87e \
-    https://github.com/BerriAI/litellm-admin-agent/archive/ef501e94bc9fbacb9233b922abf71427f030408c.tar.gz /tmp/liteadmin.tar.gz
+ADD --checksum=sha256:4fb1847e28e2be30ba76dac94f02c4903291326516a0f4ba772f6ffb8b25d05c \
+    https://github.com/BerriAI/litellm-admin-agent/archive/5661c805ce479a0cf11b842ea133cf8d22cd7e34.tar.gz /tmp/liteadmin.tar.gz
 RUN mkdir /tmp/liteadmin && tar xzf /tmp/liteadmin.tar.gz --strip-components=1 -C /tmp/liteadmin && \
     uv venv /opt/liteadmin --python python3.13 && \
     uv pip install --python /opt/liteadmin/bin/python --require-hashes -r /tmp/liteadmin/requirements.txt && \
@@ -141,7 +142,7 @@ USER root
 RUN echo "https://packages.wolfi.dev/os" >> /etc/apk/repositories
 
 # node (without npm) is required by the prisma CLI at runtime
-RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile libevent
+RUN apk add --no-cache bash openssl libssl3 tzdata nodejs python-3.13 libsndfile libevent
 COPY --from=pgbouncer-builder /usr/local/bin/pgbouncer /usr/local/bin/pgbouncer
 
 WORKDIR /app

Makefile

@@ -22,7 +22,7 @@ help:
 	@echo "  make install-proxy-dev-ci - Install proxy dev dependencies (CI-compatible)"
 	@echo "  make install-test-deps  - Install the full local test environment"
 	@echo "  make install-helm-unittest - Install helm unittest plugin"
-	@echo "  make install-hooks      - Install git hooks (Conventional Commits + Branches)"
+	@echo "  make install-hooks      - Install git hooks (Conventional Commits)"
 	@echo "  make check              - Run CI-equivalent lint on staged files, or on the diff vs the base branch when nothing is staged"
 	@echo "  make pre-commit         - Legacy alias for make check"
 	@echo "  make format             - Apply ruff format code formatting"
@@ -120,7 +120,7 @@ install-helm-unittest:
 		helm plugin install https://github.com/helm-unittest/helm-unittest --version v0.8.2; \
 	}
 
-# Install git hooks that enforce Conventional Commits and Conventional Branches.
+# Install git hooks that enforce Conventional Commits.
 # Opt-in: not chained into install-dev.
 install-hooks:
 	./scripts/install_git_hooks.sh
@@ -153,7 +153,7 @@ lint-install:
 # the same set CI's ':(glob)litellm/**/*.py' selects.
 lint-format-check-changed: $(LINT_DEP_INSTALL) $(LINT_DEP_BASE)
 	@base_ref=$$($(RESOLVE_BASE)) && \
-	changed=$$(git diff --name-only --diff-filter=ACMR "$$base_ref...HEAD" -- 'litellm/*.py') && \
+	changed=$$(git diff --name-only --diff-filter=ACMR --merge-base "$$base_ref" -- 'litellm/*.py') && \
 	files=$$(printf '%s\n' "$$changed" | grep -v '^litellm/enterprise/' || true) || exit $$?; \
 	if [ -z "$$files" ]; then \
 		echo "No changed litellm Python files to format-check."; \
@@ -290,7 +290,7 @@ test-rust-extension:
 	"$$temporary/venv/bin/python" -I -m mypy.stubtest \
 		--mypy-config-file tests/unit/rust_bridge/stubtest.ini \
 		litellm.rust_bridge._native && \
-	LITELLM_RUST=1 LITELLM_LOCAL_MODEL_COST_MAP=True \
+	LITELLM_LOCAL_MODEL_COST_MAP=True \
 	"$$temporary/venv/bin/python" -I -m pytest --import-mode=importlib -m requires_rust_extension tests/test_litellm_rust
 
 rust-sqlx-prepare:
@@ -313,7 +313,7 @@ test-unit-proxy-core: install-test-deps
 	$(UV_RUN) pytest tests/unit/proxy/auth tests/unit/proxy/client tests/unit/proxy/db tests/unit/proxy/hooks tests/unit/proxy/policy_engine --ignore=tests/unit/proxy/db/db_transaction_queue/test_e2e_pod_lock_manager.py --ignore=tests/unit/proxy/db/test_update_daily_tag_spend.py --tb=short -vv -n 4 --durations=20
 
 test-unit-proxy-misc: install-test-deps
-	$(UV_RUN) pytest tests/unit/proxy/agent_endpoints tests/unit/proxy/anthropic_endpoints tests/unit/proxy/common_utils --ignore=tests/unit/proxy/common_utils/test_cache_aware_routing.py --ignore=tests/unit/proxy/common_utils/test_check_batch_cost.py --ignore=tests/unit/proxy/common_utils/test_check_responses_cost.py --ignore=tests/unit/proxy/common_utils/test_proxy_encrypt_decrypt.py --ignore=tests/unit/proxy/common_utils/test_realtime_cache.py tests/unit/proxy/discovery_endpoints tests/unit/proxy/experimental tests/unit/proxy/google_endpoints tests/unit/proxy/health_endpoints tests/unit/proxy/image_endpoints tests/unit/proxy/middleware --ignore=tests/unit/proxy/middleware/test_request_size_limit_middleware.py tests/unit/proxy/openai_files_endpoint tests/unit/proxy/pass_through_endpoints tests/unit/proxy/prompts tests/unit/proxy/public_endpoints tests/unit/proxy/response_api_endpoints tests/unit/proxy/shutdown tests/unit/proxy/spend_tracking --ignore=tests/unit/proxy/spend_tracking/test_search_api_logging.py tests/unit/proxy/ui_crud_endpoints tests/unit/proxy/vector_store_endpoints tests/unit/proxy/_experimental/mcp_server/test_mcp_server_tool_calls_and_headers.py --ignore=tests/unit/proxy/google_endpoints/test_gemini_agents_endpoints.py --ignore=tests/unit/proxy/google_endpoints/test_google_endpoint_routing.py --ignore=tests/unit/proxy/google_endpoints/test_google_gemini_proxy_request.py --ignore=tests/unit/proxy/public_endpoints/test_blog_posts_endpoint.py --tb=short -vv -n 4 --durations=20
+	$(UV_RUN) pytest tests/unit/proxy/agent_endpoints tests/unit/proxy/anthropic_endpoints tests/unit/proxy/common_utils --ignore=tests/unit/proxy/common_utils/test_cache_aware_routing.py --ignore=tests/unit/proxy/common_utils/test_check_batch_cost.py --ignore=tests/unit/proxy/common_utils/test_check_responses_cost.py --ignore=tests/unit/proxy/common_utils/test_proxy_encrypt_decrypt.py --ignore=tests/unit/proxy/common_utils/test_realtime_cache.py tests/unit/proxy/discovery_endpoints tests/unit/proxy/mcp_server/test_tool_registry.py tests/unit/proxy/google_endpoints tests/unit/proxy/health_endpoints tests/unit/proxy/image_endpoints tests/unit/proxy/middleware --ignore=tests/unit/proxy/middleware/test_request_size_limit_middleware.py tests/unit/proxy/openai_files_endpoint tests/unit/proxy/pass_through_endpoints tests/unit/proxy/prompts tests/unit/proxy/public_endpoints tests/unit/proxy/response_api_endpoints tests/unit/proxy/shutdown tests/unit/proxy/spend_tracking --ignore=tests/unit/proxy/spend_tracking/test_search_api_logging.py tests/unit/proxy/ui_crud_endpoints tests/unit/proxy/vector_store_endpoints tests/unit/proxy/mcp_server/test_mcp_server_tool_calls_and_headers.py --ignore=tests/unit/proxy/google_endpoints/test_gemini_agents_endpoints.py --ignore=tests/unit/proxy/google_endpoints/test_google_endpoint_routing.py --ignore=tests/unit/proxy/google_endpoints/test_google_gemini_proxy_request.py --ignore=tests/unit/proxy/public_endpoints/test_blog_posts_endpoint.py --tb=short -vv -n 4 --durations=20
 
 test-unit-proxy-root: install-test-deps
 	$(UV_RUN) pytest tests/unit/proxy/test_*.py --ignore=tests/unit/proxy/test_aproxy_startup.py --ignore=tests/unit/proxy/test_credential_slot_registry.py --ignore=tests/unit/proxy/test_custom_callback_input.py --ignore=tests/unit/proxy/test_custom_logger_s3_gcs.py --ignore=tests/unit/proxy/test_custom_tokenizer_bug.py --ignore=tests/unit/proxy/test_db_schema_changes.py --ignore=tests/unit/proxy/test_deprecated_key_grace_period.py --ignore=tests/unit/proxy/test_get_favicon.py --ignore=tests/unit/proxy/test_get_image.py --ignore=tests/unit/proxy/test_prisma_client_backoff_retry.py --ignore=tests/unit/proxy/test_prompt_test_endpoint.py --ignore=tests/unit/proxy/test_proxy_config_unit_test.py --ignore=tests/unit/proxy/test_proxy_custom_auth.py --ignore=tests/unit/proxy/test_proxy_reject_logging.py --ignore=tests/unit/proxy/test_proxy_server.py --ignore=tests/unit/proxy/test_proxy_setting_guardrails.py --ignore=tests/unit/proxy/test_proxy_token_counter.py --ignore=tests/unit/proxy/test_proxy_utils.py --ignore=tests/unit/proxy/test_reducto_ocr_route.py --ignore=tests/unit/proxy/test_response_polling_pre_call_checks.py --ignore=tests/unit/proxy/test_server_root_path.py --ignore=tests/unit/proxy/test_ui_path_detection.py --ignore=tests/unit/proxy/test_unit_test_proxy_hooks.py --ignore=tests/unit/proxy/test_update_spend.py --ignore=tests/unit/proxy/test_zero_cost_model_budget_bypass.py --tb=short -vv -n 4 --durations=20
@@ -325,7 +325,7 @@ test-unit-core-utils: install-test-deps
 	$(UV_RUN) pytest tests/unit/litellm_core_utils --tb=short -vv -n 2 --durations=20
 
 test-unit-other: install-test-deps
-	$(UV_RUN) pytest tests/unit/caching tests/unit/responses tests/unit/secret_managers tests/unit/vector_stores tests/unit/a2a_protocol tests/unit/completion_extras tests/unit/containers tests/unit/enterprise tests/unit/experimental_mcp_client tests/unit/google_genai tests/unit/images tests/unit/interactions tests/unit/router_strategy tests/unit/router_utils tests/unit/types --tb=short -vv -n 4 --durations=20
+	$(UV_RUN) pytest tests/unit/caching tests/unit/responses tests/unit/secret_managers tests/unit/vector_stores tests/unit/a2a_protocol tests/unit/completion_extras tests/unit/containers tests/unit/enterprise tests/unit/mcp_client tests/unit/google_genai tests/unit/images tests/unit/interactions tests/unit/router_strategy tests/unit/router_utils tests/unit/types --tb=short -vv -n 4 --durations=20
 
 test-unit-root: install-test-deps
 	$(UV_RUN) pytest tests/unit/test_*.py --tb=short -vv -n 4 --durations=20

README.md

@@ -141,12 +141,56 @@ response = client.chat.completions.create(
 )
 ```
 
-[**Docs: LLM Providers**](https://docs.litellm.ai/docs/providers)
+</details>
+
+<details>
+<summary><b>Auto Routing</b> - Choose a model for each prompt (Python SDK + AI Gateway)</summary>
+
+### Python SDK
+
+[**Docs: Autorouter SDK**](https://docs.litellm.ai/docs/autorouter_sdk)
+
+Install with `uv add litellm` and set `OPENAI_API_KEY` for the models below.
+
+```python
+import asyncio
+from litellm import Router
+
+router = Router(model_list=[
+    {"model_name": "fast", "litellm_params": {"model": "openai/gpt-6-luna"}},
+    {"model_name": "capable", "litellm_params": {"model": "openai/gpt-6.1-sol"}},
+    {"model_name": "auto", "litellm_params": {
+        "model": "auto_router/complexity_router",
+        "complexity_router_config": {
+            "classifier_type": "heuristic",
+            "tiers": {"SIMPLE": "fast", "MEDIUM": "fast",
+                      "COMPLEX": "capable", "REASONING": "capable"},
+            "return_raw_model_name": True,
+        },
+        "complexity_router_default_model": "capable",
+    }},
+])
+response = asyncio.run(router.acompletion(
+    model="auto", messages=[{"role": "user", "content": "What is 2+2?"}],
+))
+print(response.model, response.choices[0].message.content)
+```
+
+### AI Gateway (Proxy Server)
+
+[**Create an Auto Router**](https://docs.litellm.ai/docs/auto_router/setup), then call its model name. Replace `auto` and the URL below with yours, and set `LITELLM_API_KEY` to your virtual key.
+
+```shell
+curl http://localhost:4000/v1/chat/completions \
+  -H "Authorization: Bearer $LITELLM_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model":"auto","messages":[{"role":"user","content":"What is 2+2?"}]}'
+```
 
 </details>
 
 <details>
-<summary><b>Agents</b> - Invoke A2A Agents (Python SDK + AI Gateway)</summary>
+<summary><b>A2A</b> - Invoke A2A Agents (Python SDK + AI Gateway)</summary>
 
 [**Supported Providers**](https://docs.litellm.ai/docs/a2a#add-a2a-agents) - LangGraph, Vertex AI Agent Engine, Azure AI Foundry, Bedrock AgentCore, Pydantic AI
 
@@ -223,7 +267,7 @@ async with httpx.AsyncClient(headers=headers, timeout=60.0) as http_client:
 ```python
 from mcp import ClientSession, StdioServerParameters
 from mcp.client.stdio import stdio_client
-from litellm import experimental_mcp_client
+from litellm import mcp_client
 import litellm
 
 server_params = StdioServerParameters(command="python", args=["mcp_server.py"])
@@ -233,7 +277,7 @@ async with stdio_client(server_params) as (read, write):
         await session.initialize()
 
         # Load MCP tools in OpenAI format
-        tools = await experimental_mcp_client.load_mcp_tools(session=session, format="openai")
+        tools = await mcp_client.load_mcp_tools(session=session, format="openai")
 
         # Use with any LiteLLM model
         response = await litellm.acompletion(
@@ -287,7 +331,7 @@ For MCP OAuth, an upstream may advertise dynamic client registration but refuse
 </details>
 
 <details>
-<summary><b>Agents</b> - Run Claude Code, Codex, OpenCode or Deep Agents on any model (Python SDK)</summary>
+<summary><b>Agents</b> - Run Claude Code, Codex, OpenCode, pi or Deep Agents on any model (Python SDK)</summary>
 
 ### Python SDK - Agents
 
@@ -296,7 +340,7 @@ import litellm
 from litellm import Harness, sandbox
 
 result = litellm.agent(
-    Harness.CLAUDE_CODE,  # or Harness.CODEX, Harness.OPENCODE, Harness.DEEPAGENTS
+    Harness.CLAUDE_CODE,  # or Harness.CODEX, Harness.OPENCODE, Harness.PI, Harness.DEEPAGENTS
     "Find why tests/test_router.py is flaky and fix it.",
     sandbox=sandbox.local("./repo"),
     model="litellm_proxy/claude-sonnet-4-5",  # a model group on your AI Gateway
@@ -305,7 +349,7 @@ result = litellm.agent(
 print(result.text, result.cost, [f.path for f in result.files])
 ```
 
-Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call the agent makes goes through your AI Gateway, tagged `harness,claude_code`. Drop the `litellm_proxy/` prefix to call a provider directly. Install `starlette uvicorn` plus the agent's CLI (`claude`, `codex` or `opencode`), or `deepagents langchain-litellm` for Deep Agents.
+Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call the agent makes goes through your AI Gateway, tagged `harness,claude_code`. Drop the `litellm_proxy/` prefix to call a provider directly. Install `starlette uvicorn` plus the agent's CLI (`claude`, `codex`, `opencode` or `pi`), or `deepagents langchain-litellm` for Deep Agents.
 
 [**Docs: Agent Harnesses**](https://docs.litellm.ai/docs/harness)
 
@@ -335,6 +379,7 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th
 | [Bytez (`bytez`)](https://docs.litellm.ai/docs/providers/bytez) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Cerebras (`cerebras`)](https://docs.litellm.ai/docs/providers/cerebras) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Clarifai (`clarifai`)](https://docs.litellm.ai/docs/providers/clarifai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [CLF AI Gateway (`clf_ai_gateway`)](https://docs.litellm.ai/docs/providers/clf_ai_gateway) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Cloudflare AI Workers (`cloudflare`)](https://docs.litellm.ai/docs/providers/cloudflare_workers) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Codestral (`codestral`)](https://docs.litellm.ai/docs/providers/codestral) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Cognition (`cognition`)](https://docs.litellm.ai/docs/providers/cognition) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
@@ -353,7 +398,9 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th
 | [Deepseek (`deepseek`)](https://docs.litellm.ai/docs/providers/deepseek) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Eden AI (`edenai`)](https://docs.litellm.ai/docs/providers/edenai) | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |  |  |  |
 | [ElevenLabs (`elevenlabs`)](https://docs.litellm.ai/docs/providers/elevenlabs) | ✅ | ✅ | ✅ |  |  | ✅ | ✅ |  |  |  |
+| [EmpirioLabs AI (`empiriolabs`)](https://docs.litellm.ai/docs/providers/empiriolabs) | ✅ |  | ✅ |  |  |  |  |  |  |  |
 | [Empower (`empower`)](https://docs.litellm.ai/docs/providers/empower) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [eu/jev Decisions API (`eu_jev`)](https://docs.litellm.ai/docs/pass_through/eu_jev) |  |  |  |  |  |  |  |  |  |  |
 | [Fal AI (`fal_ai`)](https://docs.litellm.ai/docs/providers/fal_ai) | ✅ | ✅ | ✅ |  | ✅ |  |  |  |  |  |
 | [Featherless AI (`featherless_ai`)](https://docs.litellm.ai/docs/providers/featherless_ai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Fireworks AI (`fireworks_ai`)](https://docs.litellm.ai/docs/providers/fireworks_ai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
@@ -384,6 +431,7 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th
 | [ModelScope (`modelscope`)](https://docs.litellm.ai/docs/providers/modelscope) | ✅ | ✅ | ✅ |  | ✅ |  |  |  |  |  |
 | [Moonshot (`moonshot`)](https://docs.litellm.ai/docs/providers/moonshot) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Morph (`morph`)](https://docs.litellm.ai/docs/providers/morph) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [Nace (`nace`)](https://console.nace.ai/docs/drex) |  |  |  |  |  |  |  |  |  |  |
 | [Nebius AI Studio (`nebius`)](https://docs.litellm.ai/docs/providers/nebius) | ✅ | ✅ | ✅ | ✅ |  |  |  |  |  |  |
 | [NLP Cloud (`nlp_cloud`)](https://docs.litellm.ai/docs/providers/nlp_cloud) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Novita AI (`novita`)](https://novita.ai/models/llm?utm_source=github_litellm&utm_medium=github_readme&utm_campaign=github_link) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
@@ -397,6 +445,7 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th
 | [OpenAI-like (`openai_like`)](https://docs.litellm.ai/docs/providers/openai_compatible) |  |  |  | ✅ |  |  |  |  |  |  |
 | [OpenRouter (`openrouter`)](https://docs.litellm.ai/docs/providers/openrouter) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [OVHCloud AI Endpoints (`ovhcloud`)](https://docs.litellm.ai/docs/providers/ovhcloud) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
+| [Parallel AI (`parallel_ai`)](https://docs.litellm.ai/docs/search/parallel_ai) |  |  |  |  |  |  |  |  |  |  |
 | [Perplexity AI (`perplexity`)](https://docs.litellm.ai/docs/providers/perplexity) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Petals (`petals`)](https://docs.litellm.ai/docs/providers/petals) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Pinstripes (`pinstripes`)](https://docs.litellm.ai/docs/providers/pinstripes) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
@@ -426,6 +475,7 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th
 | [Watsonx Text (`watsonx_text`)](https://docs.litellm.ai/docs/providers/watsonx) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [xAI (`xai`)](https://docs.litellm.ai/docs/providers/xai) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 | [Xinference (`xinference`)](https://docs.litellm.ai/docs/providers/xinference) |  |  |  | ✅ |  |  |  |  |  |  |
+| [ZeroGPU (`zerogpu`)](https://docs.litellm.ai/docs/providers/zerogpu) | ✅ | ✅ | ✅ |  |  |  |  |  |  |  |
 
 [**Read the Docs**](https://docs.litellm.ai/docs/)
 
@@ -440,7 +490,7 @@ You can use LiteLLM through either the Proxy Server or Python SDK. Both give you
 <tr>
 <th style={{width: '14%'}}></th>
 <th style={{width: '43%'}}><strong><a href="https://docs.litellm.ai/docs/simple_proxy">LiteLLM AI Gateway</a></strong></th>
-<th style={{width: '43%'}}><strong><a href="https://docs.litellm.ai/docs/">LiteLLM Python SDK</a></strong></th>
+<th style={{width: '43%'}}><strong><a href="https://docs.litellm.ai/docs/python_sdk">LiteLLM Python SDK</a></strong></th>
 </tr>
 </thead>
 <tbody>
@@ -615,15 +665,15 @@ The Terraform modules live at [`terraform/litellm/aws/`](./terraform/litellm/aws
 
 ### Verify Docker Image Signatures
 
-All LiteLLM Docker images published to GHCR are signed with [cosign](https://docs.sigstore.dev/cosign/overview/). Every release is signed with the same key introduced in [commit `0112e53`](https://github.com/BerriAI/litellm/commit/0112e53046018d726492c814b3644b7d376029d0).
+All LiteLLM Docker images published to GHCR are signed with [cosign](https://docs.sigstore.dev/cosign/overview/). Releases are signed with the key introduced in [commit `e82a3c6`](https://github.com/BerriAI/litellm/commit/e82a3c646908ba0a6ee0beddc4310d7a6cac4065).
 
 **Verify using the pinned commit hash (recommended):**
 
-A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the original signing key:
+A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:
 
 ```bash
 cosign verify \
-  --key https://raw.githubusercontent.com/BerriAI/litellm/0112e53046018d726492c814b3644b7d376029d0/cosign.pub \
+  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
   ghcr.io/berriai/litellm:<release-tag>
 ```
 
@@ -648,7 +698,7 @@ For companies that need better security, user management and professional suppor
 [Talk to founders](https://enterprise.litellm.ai/demo)
 
 This covers:
-- ✅ **Features under the [LiteLLM Commercial License](https://docs.litellm.ai/docs/proxy/enterprise):**
+- ✅ **Features under the [LiteLLM Commercial License](https://docs.litellm.ai/docs/enterprise):**
 - ✅ **Feature Prioritization**
 - ✅ **Custom Integrations**
 - ✅ **Professional Support - Dedicated discord + slack**
@@ -659,6 +709,8 @@ This covers:
 
 We welcome contributions to LiteLLM! Whether you're fixing bugs, adding features, or improving documentation, we appreciate your help.
 
+If you want to work closely with our team, check out [LiteLLM OSS Champions](https://www.litellm.ai/champions), our program for trusted OSS contributors with their own Slack channel and a voice in the roadmap
+
 ## Quick Start for Contributors
 
 This requires uv to be installed.

SECURITY.md

@@ -79,7 +79,7 @@ To qualify for a bounty, reports must include clear reproduction steps, a reprod
 - Audit Logs with retention policy
 - Control Allowed IP Addresses that can access your Cloud LiteLLM Instance
 
-For security inquiries, please contact us at support@berri.ai
+For security inquiries, please contact us at security@berri.ai
 
 #### Supported data regions for LiteLLM Cloud
 

backend/Dockerfile

@@ -1,9 +1,37 @@
-ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
-ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
+ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
 
 FROM $UV_IMAGE AS uvbin
 
+# ---------- Rust extension ----------
+# Copies only the native build's inputs, so the fat-LTO compile stays cached across Python-only changes
+FROM $LITELLM_BUILD_IMAGE AS rust-native
+
+WORKDIR /app
+USER root
+
+COPY --from=uvbin /uv /uvx /usr/local/bin/
+
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
+
+ENV UV_PYTHON_DOWNLOADS=0
+
+COPY pyproject.toml README.md rust-toolchain.toml ./
+COPY .cargo ./.cargo
+COPY litellm-rust ./litellm-rust
+COPY model_prices_and_context_window.json ./
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    mkdir -p litellm/rust_bridge /wheels && \
+    touch litellm/__init__.py litellm/rust_bridge/__init__.py && \
+    uvx --python python3.13 --from maturin==1.15.0 --with puccinialin==0.1.14 \
+        python -c "import maturin; maturin.build_wheel('/wheels')"
+
 # ---------- Builder ----------
 FROM $LITELLM_BUILD_IMAGE AS builder
 
@@ -16,7 +44,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
 # instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
 # (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
 RUN for i in 1 2 3; do \
-      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
+      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl libssl3 openssl-dev libsndfile nodejs npm && break; \
       [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
       sleep 5; \
     done
@@ -54,7 +82,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 COPY . .
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-default-groups --no-editable \
+    uv sync --frozen --no-install-project --no-default-groups --no-editable \
         --extra proxy \
         --extra proxy-runtime \
         --group admin-mcp \
@@ -63,6 +91,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra saml \
         --python python3.13
 
+RUN --mount=type=bind,from=rust-native,source=/wheels,target=/wheels \
+    uv pip install --python /app/.venv/bin/python --no-deps /wheels/litellm-*.whl
+
 RUN cp "$(python -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])')"/litellm/rust_bridge/_native*.so litellm/rust_bridge/
 
 RUN HOME=/opt/prisma XDG_CACHE_HOME=/opt/prisma/.cache PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries \
@@ -79,7 +110,7 @@ ENV LITELLM_RELEASE_TAG=${LITELLM_RELEASE_TAG}
 USER root
 
 RUN for i in 1 2 3; do \
-      apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic && break; \
+      apk add --no-cache bash openssl libssl3 tzdata python-3.13 libsndfile libatomic && break; \
       [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
       sleep 5; \
     done
@@ -94,7 +125,12 @@ ENV HOME=/home/nonroot \
     PYTHONUNBUFFERED=1 \
     PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries
 
-COPY --from=builder --chown=nonroot:nonroot /app /app
+COPY --from=builder --chown=nonroot:nonroot /app/.venv /app/.venv
+COPY --from=builder --chown=nonroot:nonroot /app/litellm /app/litellm
+COPY --from=builder --chown=nonroot:nonroot /app/enterprise /app/enterprise
+COPY --from=builder --chown=nonroot:nonroot /app/backend /app/backend
+COPY --from=builder --chown=nonroot:nonroot /app/docker/component_entrypoint.sh /app/docker/component_entrypoint.sh
+COPY --from=builder --chown=nonroot:nonroot /app/schema.prisma /app/schema.prisma
 COPY --from=builder /opt/prisma /opt/prisma
 
 RUN find /app/.venv -type f -path "*/tornado/test/*" -delete && \

backend/routes/allowlist.py

@@ -55,6 +55,7 @@ BACKEND_PATH_PREFIXES: tuple[str, ...] = (
     "/cache_settings",
     "/coordination_redis/",
     "/cost_tracking",
+    "/subscription_accounts/",
     "/cost_optimization/",
     "/cost/",
     "/credentials",
@@ -165,6 +166,7 @@ BACKEND_EXACT_PATHS: frozenset[str] = frozenset(
 BACKEND_MOUNT_PATHS: frozenset[str] = frozenset(
     {
         "/admin",
+        "/metrics",
         "/swagger",  # API documentation static assets belong to the backend
         "/mcp",  # lazily-mounted MCP sub-app serves on the backend component
     }

ci_cd/cost_map_guard.py

@@ -1,41 +1,67 @@
 """Guard the cost map on pull requests.
 
-Every pull request whose diff against its merge base touches one of the three cost map files gets the file
-checks: the files parse, the backup copy matches the root file, and the JSON schema is in sync and validates the
-map. A pull request that leaves all three untouched skips them, since merging it keeps the base branch's copies
-and its head tree only carries whatever state the branch was cut from. Pull requests from the cost map sync bot
-(branches named litellm_cost_map_sync_*) always get the file checks and additionally may only touch those three
-files and may only add or update models.
+Every pull request whose diff against its merge base touches one of the two cost map files or the pins file gets the
+file checks: the files parse, the backup copy matches the root file, and every capability pinned in
+ci_cd/cost_map_pins.json (a value verified against a live provider call where the provider's model listing says
+otherwise) still holds in the tree that merging the pull request into its base branch produces, so a branch cut before
+a pin was added cannot lower it, and a pull request that deletes the pins file fails too. A pull request that leaves
+all three untouched skips them, since merging it keeps the base branch's copies and its head tree only carries
+whatever state the branch was cut from. Pull requests from the cost map sync bot (branches named
+litellm_cost_map_sync_*) always get the file checks and additionally may only touch the two cost map files and may
+only add or update models, so a sync can never lower a pinned capability or edit the pins. Price sync pull requests
+(branches named litellm-providers/price-sync*) and pull requests a bot account opened (COST_MAP_GUARD_AUTHOR_TYPE is
+Bot) get the human checks and may not touch the pins file either.
 """
 
 from __future__ import annotations
 
 import argparse
 import json
+import os
 import subprocess
 import sys
-from collections.abc import Mapping, Sequence
+from collections.abc import Iterator, Mapping, Sequence
 from dataclasses import dataclass
 from types import MappingProxyType
 from typing import Final
 
-from generate_model_prices_schema import SPECIAL_ROOT_KEYS, build_schema, render, validation_errors
-
 COST_MAP_PATH: Final = "model_prices_and_context_window.json"
 BACKUP_PATH: Final = "litellm/model_prices_and_context_window_backup.json"
-SCHEMA_PATH: Final = "model_prices_and_context_window.schema.json"
-GUARDED_PATHS: Final = (COST_MAP_PATH, BACKUP_PATH, SCHEMA_PATH)
+PINS_PATH: Final = "ci_cd/cost_map_pins.json"
+GUARDED_PATHS: Final = (COST_MAP_PATH, BACKUP_PATH)
+CHECKED_PATHS: Final = (*GUARDED_PATHS, PINS_PATH)
 BOT_BRANCH_PREFIX: Final = "litellm_cost_map_sync_"
+PRICE_SYNC_BRANCH_PREFIX: Final = "litellm-providers/price-sync"
+TARGET_ENV: Final = "COST_MAP_GUARD_TARGET"
+AUTHOR_TYPE_ENV: Final = "COST_MAP_GUARD_AUTHOR_TYPE"
+BOT_AUTHOR_TYPE: Final = "Bot"
+MERGE_CONFLICT_EXIT_CODE: Final = 1
+SPECIAL_ROOT_KEYS: Final = frozenset({"sample_spec", "fallback_generalizations"})
 EVALUATION_OUTPUT_PRICE_SOURCES: Final[Mapping[str, str]] = MappingProxyType({})
 
 CostMap = dict[str, object]
+_MISSING: Final = object()
 
 
 @dataclass(frozen=True, slots=True)
 class Snapshot:
     cost_map: str
     backup: str
-    schema: str
+    pins: str
+
+
+@dataclass(frozen=True, slots=True)
+class MergeFailed:
+    stderr: str
+
+
+@dataclass(frozen=True, slots=True)
+class Pin:
+    key: str
+    field: str
+    value: object
+    checked: str
+    source: str
 
 
 def _parse_object(text: str, path: str) -> CostMap | str:
@@ -46,13 +72,6 @@ def _parse_object(text: str, path: str) -> CostMap | str:
     return parsed if isinstance(parsed, dict) else f"{path} must be a JSON object at the root"
 
 
-def _rendered_schema(cost_map: CostMap) -> str:
-    try:
-        return render(build_schema(cost_map))
-    except SystemExit as error:
-        return str(error)
-
-
 def _evaluation_output_price_failure(
     key: str, entry: object, evaluation_output_price_sources: Mapping[str, str]
 ) -> str | None:
@@ -78,29 +97,13 @@ def _file_failures(
     *,
     evaluation_output_price_sources: Mapping[str, str] = EVALUATION_OUTPUT_PRICE_SOURCES,
 ) -> tuple[str, ...]:
-    schema_text: Final = _rendered_schema(head_map)
-    if not schema_text.startswith("{"):
-        return (schema_text,)
     backup_failure: Final = (
         ()
         if head.backup == head.cost_map
         else (f"{BACKUP_PATH} differs from {COST_MAP_PATH}; copy the root file over it",)
     )
-    schema_failure: Final = (
-        ()
-        if head.schema == schema_text
-        else (
-            f"{SCHEMA_PATH} is out of sync with {COST_MAP_PATH}; "
-            "run `python ci_cd/generate_model_prices_schema.py` and commit the result",
-        )
-    )
     return (
         *backup_failure,
-        *schema_failure,
-        *(
-            f"{COST_MAP_PATH} does not validate against its schema: {error}"
-            for error in validation_errors(head_map, json.loads(schema_text))[:20]
-        ),
         *(
             failure
             for key, entry in head_map.items()
@@ -118,6 +121,48 @@ def _entries(cost_map: CostMap) -> dict[str, dict[str, object]]:
     return {key: entry for key, entry in cost_map.items() if isinstance(entry, dict)}
 
 
+def _pin(key: str, field: str, spec: object) -> Pin | str:
+    malformed: Final = f"{PINS_PATH}: {key}.{field} needs value, checked, and source"
+    if not isinstance(spec, dict) or "value" not in spec:
+        return malformed
+    checked: Final = spec.get("checked")
+    source: Final = spec.get("source")
+    if not isinstance(checked, str) or not isinstance(source, str) or not checked or not source:
+        return malformed
+    return Pin(key, field, spec["value"], checked, source)
+
+
+def _pins(parsed: CostMap) -> Iterator[Pin | str]:
+    for key, fields in parsed.items():
+        if not isinstance(fields, dict):
+            yield f"{PINS_PATH}: {key} must map field names to pins"
+            continue
+        for field, spec in fields.items():
+            yield _pin(key, field, spec)
+
+
+def _pin_violations(parsed: CostMap, entries: Mapping[str, Mapping[str, object]]) -> Iterator[str]:
+    for pin in _pins(parsed):
+        if isinstance(pin, str):
+            yield pin
+        elif pin.key not in entries:
+            yield f"{PINS_PATH} pins {pin.key}.{pin.field} but {COST_MAP_PATH} has no {pin.key} entry"
+        elif entries[pin.key].get(pin.field, _MISSING) != pin.value:
+            yield (
+                f"{COST_MAP_PATH}: {pin.key}.{pin.field} must stay {json.dumps(pin.value)} "
+                f"(verified {pin.checked}: {pin.source}); change {PINS_PATH} with new evidence first"
+            )
+
+
+def pin_failures(pins_text: str, head_map: CostMap) -> tuple[str, ...]:
+    if not pins_text:
+        return (f"{PINS_PATH} is missing; restore it, its pins were verified against live provider calls",)
+    parsed: Final = _parse_object(pins_text, PINS_PATH)
+    if isinstance(parsed, str):
+        return (parsed,)
+    return tuple(_pin_violations(parsed, _entries(head_map)))
+
+
 def _bot_failures(base: Snapshot, head_map: CostMap, changed_files: Sequence[str]) -> tuple[str, ...]:
     base_map: Final = _parse_object(base.cost_map, COST_MAP_PATH)
     if isinstance(base_map, str):
@@ -147,8 +192,18 @@ def _bot_failures(base: Snapshot, head_map: CostMap, changed_files: Sequence[str
     )
 
 
+def pin_author_failures(head_ref: str, author_type: str, changed_files: Sequence[str]) -> tuple[str, ...]:
+    if PINS_PATH not in changed_files:
+        return ()
+    if head_ref.startswith(PRICE_SYNC_BRANCH_PREFIX):
+        return (f"price sync PRs may not change {PINS_PATH}; a pin moves only in a human PR with new evidence",)
+    if author_type == BOT_AUTHOR_TYPE:
+        return (f"PRs opened by a bot may not change {PINS_PATH}; a pin moves only in a human PR with new evidence",)
+    return ()
+
+
 def touches_cost_map(changed_files: Sequence[str]) -> bool:
-    return any(path in GUARDED_PATHS for path in changed_files)
+    return any(path in CHECKED_PATHS for path in changed_files)
 
 
 def contract_for(bot: bool, changed_files: Sequence[str]) -> str:
@@ -157,9 +212,21 @@ def contract_for(bot: bool, changed_files: Sequence[str]) -> str:
     return "human PR, file checks only" if touches_cost_map(changed_files) else "human PR, cost map untouched"
 
 
+def _merged_pin_failures(base: Snapshot, merged: Snapshot | None) -> tuple[str, ...]:
+    if merged is None:
+        return ("the pull request does not merge cleanly into its base branch; merge the base branch in first",)
+    if not merged.pins and not base.pins:
+        return ()
+    merged_map: Final = _parse_object(merged.cost_map, COST_MAP_PATH)
+    if isinstance(merged_map, str):
+        return (f"after merging into the base branch, {merged_map}",)
+    return pin_failures(merged.pins, merged_map)
+
+
 def guard_failures(
     base: Snapshot,
     head: Snapshot,
+    merged: Snapshot | None,
     changed_files: Sequence[str],
     bot: bool,
     *,
@@ -174,6 +241,7 @@ def guard_failures(
         *_file_failures(
             head, head_map, evaluation_output_price_sources=evaluation_output_price_sources
         ),
+        *_merged_pin_failures(base, merged),
         *(_bot_failures(base, head_map, changed_files) if bot else ()),
     )
 
@@ -184,7 +252,22 @@ def _git(*args: str) -> str | None:
 
 
 def snapshot(revision: str) -> Snapshot:
-    return Snapshot(*(_git("show", f"{revision}:{path}") or "" for path in GUARDED_PATHS))
+    return Snapshot(*(_git("show", f"{revision}:{path}") or "" for path in CHECKED_PATHS))
+
+
+def merged_tree(merge_base: str, target: str, head: str) -> str | MergeFailed | None:
+    result: Final = subprocess.run(
+        ("git", "merge-tree", "--write-tree", f"--merge-base={merge_base}", target, head),
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+    tree: Final = result.stdout.split("\n", 1)[0]
+    if result.returncode == 0:
+        return tree
+    if result.returncode == MERGE_CONFLICT_EXIT_CODE and tree:
+        return None
+    return MergeFailed(result.stderr.strip())
 
 
 def changed_files(base: str, head: str) -> tuple[str, ...] | None:
@@ -196,6 +279,13 @@ def main(argv: Sequence[str]) -> int:
     parser: Final = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--base", required=True, help="merge base of the pull request")
     parser.add_argument("--head", required=True, help="head commit of the pull request")
+    env_target: Final = os.environ.get(TARGET_ENV)
+    parser.add_argument(
+        "--target",
+        default=env_target,
+        required=env_target is None,
+        help=f"tip of the base branch the pull request merges into, defaulting to ${TARGET_ENV}",
+    )
     parser.add_argument("--head-ref", required=True, help="head branch name of the pull request")
     args: Final = parser.parse_args(argv)
     bot: Final = args.head_ref.startswith(BOT_BRANCH_PREFIX)
@@ -203,7 +293,19 @@ def main(argv: Sequence[str]) -> int:
     if changed is None:
         print(f"cost map guard failed: git diff {args.base} {args.head} failed, so the changed files are unknown")
         return 1
-    failures: Final = guard_failures(snapshot(args.base), snapshot(args.head), changed, bot)
+    if not bot and not touches_cost_map(changed):
+        print(f"cost map guard passed ({contract_for(bot, changed)})")
+        return 0
+    merged: Final = merged_tree(args.base, args.target, args.head)
+    if isinstance(merged, MergeFailed):
+        print(f"cost map guard failed: git merge-tree into {args.target} failed: {merged.stderr}")
+        return 1
+    failures: Final = (
+        *guard_failures(
+            snapshot(args.base), snapshot(args.head), None if merged is None else snapshot(merged), changed, bot
+        ),
+        *pin_author_failures(args.head_ref, os.environ.get(AUTHOR_TYPE_ENV, ""), changed),
+    )
     contract: Final = contract_for(bot, changed)
     if failures:
         print(f"cost map guard failed ({contract}):")

ci_cd/cost_map_pins.json

@@ -0,0 +1,16 @@
+{
+    "fireworks_ai/accounts/fireworks/models/minimax-m3": {
+        "supports_vision": {
+            "value": true,
+            "checked": "2026-09-26",
+            "source": "POST https://api.fireworks.ai/inference/v1/chat/completions read a 512x512 PNG back as digit=7, shape=circle, color=red while GET https://api.fireworks.ai/inference/v1/models still said supports_image_input false and GET https://api.fireworks.ai/v1/serverless/models?format=nested carried no supportsImageInput on the row"
+        }
+    },
+    "fireworks_ai/minimax-m3": {
+        "supports_vision": {
+            "value": true,
+            "checked": "2026-09-26",
+            "source": "POST https://api.fireworks.ai/inference/v1/chat/completions read a 512x512 PNG back as digit=7, shape=circle, color=red while GET https://api.fireworks.ai/inference/v1/models still said supports_image_input false and GET https://api.fireworks.ai/v1/serverless/models?format=nested carried no supportsImageInput on the row"
+        }
+    }
+}

ci_cd/generate_model_prices_schema.py

@@ -1,409 +0,0 @@
-from __future__ import annotations
-
-import json
-import sys
-from pathlib import Path
-from typing import Final, Optional
-
-import jsonschema
-
-REPO_ROOT = Path(__file__).parent.parent
-PRICES_PATH = REPO_ROOT / "model_prices_and_context_window.json"
-SCHEMA_PATH = REPO_ROOT / "model_prices_and_context_window.schema.json"
-
-SPECIAL_ROOT_KEYS = frozenset({"sample_spec", "fallback_generalizations"})
-
-JsonSchema = dict
-
-NONNEG_NUMBER: JsonSchema = {"type": "number", "minimum": 0}
-NONNEG_INTEGER: JsonSchema = {"type": "integer", "minimum": 0}
-BOOLEAN: JsonSchema = {"type": "boolean"}
-STRING: JsonSchema = {"type": "string"}
-TIME_WINDOW: Final[JsonSchema] = {"type": "string", "pattern": r"^([01]\d|2[0-3]):[0-5]\d-([01]\d|2[0-3]):[0-5]\d$"}
-WEEKDAY_PATTERN: Final = (
-    r"(?i)^(mon|monday|tue|tues|tuesday|wed|wednesday|thu|thur|thurs|thursday|fri|friday|sat|saturday|sun|sunday)$"
-)
-
-EXTRA_BOOLEAN_KEYS = frozenset(
-    {
-        "gemini_native_audio",
-        "gemini_audio_only_live",
-        "uses_embed_content",
-        "use_openai_responses_path",
-        "bedrock_converse_supports_strict_tools",
-        "thinking_always_on",
-    }
-)
-
-HOURS_UTC: Final[JsonSchema] = {
-    "description": 'UTC "HH:MM-HH:MM" window, or a list of them; a window may wrap past midnight.',
-    "oneOf": [TIME_WINDOW, {"type": "array", "items": TIME_WINDOW, "minItems": 1}],
-}
-
-OFF_PEAK_WINDOW: Final[JsonSchema] = {
-    "type": "object",
-    "properties": {
-        "hours_utc": HOURS_UTC,
-        "weekdays": {
-            "type": "array",
-            "description": "ISO-8601 weekday numbers (1 = Monday .. 7 = Sunday) or English day names the window applies on.",
-            "items": {
-                "oneOf": [
-                    {"type": "integer", "minimum": 1, "maximum": 7},
-                    {"type": "string", "pattern": WEEKDAY_PATTERN},
-                ]
-            },
-            "minItems": 1,
-        },
-    },
-    "required": ["hours_utc"],
-    "additionalProperties": False,
-}
-
-OBJECT_KEYS: dict[str, JsonSchema] = {
-    "off_peak_pricing": {
-        "type": "object",
-        "description": "Rates that replace the same-named base fields while the request falls inside the stated UTC windows.",
-        "properties": {
-            "hours_utc": HOURS_UTC,
-            "windows": {"type": "array", "items": OFF_PEAK_WINDOW, "minItems": 1},
-            "weekday_timezone": {
-                "type": "string",
-                "description": "IANA zone the weekdays of each window are read on; defaults to UTC.",
-            },
-            "input_cost_per_token": NONNEG_NUMBER,
-            "output_cost_per_token": NONNEG_NUMBER,
-            "output_cost_per_reasoning_token": NONNEG_NUMBER,
-            "cache_read_input_token_cost": NONNEG_NUMBER,
-            "cache_creation_input_token_cost": NONNEG_NUMBER,
-        },
-        "anyOf": [{"required": ["hours_utc"]}, {"required": ["windows"]}],
-        "additionalProperties": False,
-    },
-    "search_context_cost_per_query": {
-        "type": "object",
-        "description": "USD cost per web search query, keyed by search context size.",
-        "properties": {
-            "search_context_size_low": NONNEG_NUMBER,
-            "search_context_size_medium": NONNEG_NUMBER,
-            "search_context_size_high": NONNEG_NUMBER,
-        },
-        "additionalProperties": False,
-    },
-    "guardrail_cost_per_unit": {
-        "type": "object",
-        "description": "USD cost per billable guardrail unit, keyed by the provider's usage counter name (e.g. Bedrock's contentPolicyUnits).",
-        "additionalProperties": NONNEG_NUMBER,
-    },
-    "metadata": {
-        "type": "object",
-        "description": "Free-form notes about the entry (e.g. pricing derivation).",
-    },
-    "provider_specific_entry": {
-        "type": "object",
-        "description": "Provider-internal routing hints (e.g. bedrock_invocation_schema).",
-    },
-}
-
-ARRAY_KEYS: dict[str, JsonSchema] = {
-    "supported_audio_formats": {
-        "type": "array",
-        "description": "Audio container formats the model can return.",
-        "items": {"type": "string", "enum": ["mp3", "wav"]},
-    },
-    "supported_endpoints": {
-        "type": "array",
-        "description": "OpenAI-style API routes this model can be called through, e.g. /v1/chat/completions.",
-        "items": STRING,
-    },
-    "supported_modalities": {
-        "type": "array",
-        "description": "Input modalities the model accepts.",
-        "items": {"type": "string", "enum": ["text", "image", "audio", "video"]},
-    },
-    "supported_output_modalities": {
-        "type": "array",
-        "description": "Output modalities the model can produce.",
-        "items": {"type": "string", "enum": ["text", "image", "audio", "video", "code"]},
-    },
-    "reasoning_effort_levels": {
-        "type": "array",
-        "description": "Exact reasoning_effort levels this deployment accepts; wins over supports_* flags.",
-        "items": {"type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh", "max"]},
-    },
-    "supported_regions": {
-        "type": "array",
-        "description": "Cloud regions the model is available in ('global' or region ids).",
-        "items": STRING,
-    },
-    "tiered_pricing": {
-        "type": "array",
-        "description": "Context-length or result-count tiered rates; each tier's costs apply within its range.",
-        "items": {
-            "type": "object",
-            "properties": {
-                "range": {
-                    "type": "array",
-                    "description": "[min, max] prompt-token span this tier applies to.",
-                    "items": NONNEG_NUMBER,
-                    "minItems": 2,
-                    "maxItems": 2,
-                },
-                "max_results_range": {
-                    "type": "array",
-                    "description": "[min, max] result-count span this tier applies to (search models).",
-                    "items": NONNEG_NUMBER,
-                    "minItems": 2,
-                    "maxItems": 2,
-                },
-                "input_cost_per_token": NONNEG_NUMBER,
-                "output_cost_per_token": NONNEG_NUMBER,
-                "output_cost_per_reasoning_token": NONNEG_NUMBER,
-                "cache_read_input_token_cost": NONNEG_NUMBER,
-                "cache_creation_input_token_cost": NONNEG_NUMBER,
-                "input_cost_per_query": NONNEG_NUMBER,
-            },
-            "additionalProperties": False,
-        },
-    },
-}
-
-INTEGER_KEYS: dict[str, JsonSchema] = {
-    "max_tokens": {
-        **NONNEG_INTEGER,
-        "description": "Legacy field: max output tokens if the provider specifies it, else max input tokens.",
-    },
-    "max_input_tokens": {
-        **NONNEG_INTEGER,
-        "description": "Maximum prompt/context tokens the model accepts.",
-    },
-    "max_output_tokens": {
-        **NONNEG_INTEGER,
-        "description": "Maximum tokens the model can generate in one response.",
-    },
-    "output_vector_size": {
-        **NONNEG_INTEGER,
-        "description": "Embedding dimension for embedding models.",
-    },
-    "prompt_cache_min_tokens": {
-        **NONNEG_INTEGER,
-        "description": "Smallest prefix the provider will actually cache; absent means the provider default applies.",
-    },
-    "tpm": {**NONNEG_INTEGER, "description": "Provider default tokens-per-minute limit."},
-    "rpm": {**NONNEG_INTEGER, "description": "Provider default requests-per-minute limit."},
-}
-
-NUMBER_KEYS: dict[str, JsonSchema] = {
-    "regional_processing_uplift_multiplier_eu": {
-        "type": "number",
-        "minimum": 1,
-        "description": "Multiplier applied to all token costs for EU data residency (e.g. 1.10 = +10%).",
-    },
-    "regional_processing_uplift_multiplier_us": {
-        "type": "number",
-        "minimum": 1,
-        "description": "Multiplier applied to all token costs for US data residency (e.g. 1.10 = +10%).",
-    },
-    "regional_endpoint_uplift_multiplier": {
-        "type": "number",
-        "minimum": 1,
-        "description": "Multiplier applied to all token costs when served from a non-global Vertex AI endpoint (e.g. 1.10 = +10%).",
-    },
-}
-
-COST_DESCRIPTIONS: dict[str, str] = {
-    "input_cost_per_token": "USD per prompt token.",
-    "output_cost_per_token": "USD per generated token.",
-    "output_cost_per_reasoning_token": "USD per reasoning/thinking token, when billed separately.",
-    "google_maps_grounding_cost_per_query": (
-        "USD per Grounding with Google Maps request; billed per query or per prompt per web_search_billing_unit."
-    ),
-    "cache_creation_input_token_cost": "USD per token written to the provider's prompt cache.",
-    "cache_read_input_token_cost": "USD per prompt token served from the provider's prompt cache.",
-    "input_cost_per_token_batches": "USD per prompt token via the provider's batch API.",
-    "output_cost_per_token_batches": "USD per generated token via the provider's batch API.",
-}
-
-
-def cost_description(key: str) -> Optional[str]:
-    if key in COST_DESCRIPTIONS:
-        return COST_DESCRIPTIONS[key]
-    if key.endswith("_flex"):
-        return "Flex service-tier rate for the same-named base field."
-    if key.endswith("_priority"):
-        return "Priority service-tier rate for the same-named base field."
-    if "_above_" in key:
-        return "Rate applied once the prompt exceeds the token threshold in the field name."
-    return None
-
-
-def cost_schema(key: str) -> JsonSchema:
-    description = cost_description(key)
-    return {**NONNEG_NUMBER, "description": description} if description else dict(NONNEG_NUMBER)
-
-
-def string_key_schemas(modes: tuple) -> dict[str, JsonSchema]:
-    return {
-        "litellm_provider": {
-            "type": "string",
-            "description": "LiteLLM provider slug; one of https://docs.litellm.ai/docs/providers.",
-        },
-        "mode": {
-            "type": "string",
-            "description": "Primary API surface / task type of the model.",
-            "enum": list(modes),
-        },
-        "source": {
-            "type": "string",
-            "description": "URL of the provider pricing/model page this entry was taken from.",
-        },
-        "deprecation_date": {
-            "type": "string",
-            "description": "Date the provider deprecates the model, YYYY-MM-DD.",
-            "format": "date",
-            "pattern": "^\\d{4}-(0[1-9]|1[0-2])-(0[1-9]|[12]\\d|3[01])$",
-        },
-        "web_search_billing_unit": {
-            "type": "string",
-            "description": "Whether web search is billed per query or per prompt.",
-            "enum": ["per_query", "per_prompt"],
-        },
-        "bedrock_output_config_effort_ceiling": {
-            "type": "string",
-            "description": "Highest reasoning effort the Bedrock output_config accepts for this model.",
-            "enum": ["low", "medium", "high", "max", "xhigh"],
-        },
-        "default_reasoning_effort": {
-            "type": "string",
-            "description": (
-                "Reasoning effort the provider applies when the request omits reasoning_effort. "
-                "Gates whether a non-default temperature or the top_p/logprobs sampling params are "
-                "accepted, which hold only when the effort resolves to 'none'."
-            ),
-            "enum": ["none", "minimal", "low", "medium", "high", "xhigh"],
-        },
-        "comment": STRING,
-        "audio_transcription_config": STRING,
-        "vertex_ai_audio_api": {
-            "type": "string",
-            "enum": ["lyria_predict", "lyria_interactions"],
-        },
-    }
-
-
-def classify(key: str, modes: tuple) -> Optional[JsonSchema]:
-    curated = {**OBJECT_KEYS, **ARRAY_KEYS, **string_key_schemas(modes), **INTEGER_KEYS, **NUMBER_KEYS}
-    if key in curated:
-        return curated[key]
-    if key.startswith("supports_") or key in EXTRA_BOOLEAN_KEYS:
-        return BOOLEAN
-    if "cost" in key:
-        return cost_schema(key)
-    return None
-
-
-def build_schema(prices: dict) -> JsonSchema:
-    entries = {name: entry for name, entry in prices.items() if name not in SPECIAL_ROOT_KEYS}
-    all_keys = tuple(sorted({key for entry in entries.values() for key in entry}))
-    modes = tuple(sorted({entry["mode"] for entry in entries.values() if "mode" in entry}))
-    unclassified = tuple(key for key in all_keys if classify(key, modes) is None)
-    if unclassified:
-        raise SystemExit(
-            f"Unclassified keys in {PRICES_PATH.name}: {', '.join(unclassified)}. "
-            f"Add them to the key tables in {Path(__file__).name} and rerun it."
-        )
-    entry_properties = {key: classify(key, modes) for key in all_keys}
-    return {
-        "$schema": "https://json-schema.org/draft/2020-12/schema",
-        "title": "LiteLLM model_prices_and_context_window.json",
-        "description": (
-            "Schema for LiteLLM's model price and context window registry "
-            "(https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json). "
-            "Every top-level key except 'sample_spec' and 'fallback_generalizations' is a model id, "
-            "optionally prefixed with its provider (e.g. 'azure/gpt-5.4'), mapping to a model entry. "
-            "All costs are USD per unit. New optional fields are added regularly, so consumers should "
-            "ignore unknown fields rather than reject them."
-        ),
-        "type": "object",
-        "properties": {
-            "sample_spec": {
-                "type": "object",
-                "description": (
-                    "Documentation placeholder illustrating the entry shape; not a real model and not "
-                    "schema-conformant (several values are prose)."
-                ),
-            },
-            "fallback_generalizations": {
-                "type": "object",
-                "description": "Regex rules that generalize unknown model ids to known families; not a model entry.",
-                "properties": {
-                    "rules": {
-                        "type": "array",
-                        "items": {
-                            "type": "object",
-                            "properties": {
-                                "name": STRING,
-                                "pattern": STRING,
-                                "description": STRING,
-                            },
-                            "required": ["name", "pattern"],
-                            "additionalProperties": True,
-                        },
-                    }
-                },
-                "additionalProperties": False,
-            },
-        },
-        "additionalProperties": {"$ref": "#/$defs/modelEntry"},
-        "$defs": {
-            "modelEntry": {
-                "type": "object",
-                "description": (
-                    "Pricing, limits, and capability flags for one model. Fields other than litellm_provider "
-                    "are optional; boolean capability flags are simply omitted when unknown or false."
-                ),
-                "required": ["litellm_provider"],
-                "properties": entry_properties,
-                "additionalProperties": True,
-            }
-        },
-    }
-
-
-def render(schema: JsonSchema) -> str:
-    return json.dumps(schema, indent=2) + "\n"
-
-
-def validation_errors(prices: dict, schema: JsonSchema) -> tuple:
-    validator = jsonschema.Draft202012Validator(schema, format_checker=jsonschema.Draft202012Validator.FORMAT_CHECKER)
-    return tuple(
-        f"{'.'.join(str(part) for part in error.absolute_path)}: {error.message}"
-        for error in validator.iter_errors(prices)
-    )
-
-
-def main() -> int:
-    check = "--check" in sys.argv[1:]
-    prices = json.loads(PRICES_PATH.read_text())
-    rendered = render(build_schema(prices))
-    errors = validation_errors(prices, json.loads(rendered))
-    if errors:
-        print(f"{PRICES_PATH.name} does not validate against the generated schema:")
-        print("\n".join(errors[:20]))
-        return 1
-    if not check:
-        SCHEMA_PATH.write_text(rendered)
-        print(f"wrote {SCHEMA_PATH}")
-        return 0
-    if not SCHEMA_PATH.exists() or SCHEMA_PATH.read_text() != rendered:
-        print(
-            f"{SCHEMA_PATH.name} is out of sync with {PRICES_PATH.name}. "
-            f"Run `python {Path(__file__).relative_to(REPO_ROOT)}` and commit the result."
-        )
-        return 1
-    print(f"{SCHEMA_PATH.name} is in sync and {PRICES_PATH.name} validates against it")
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())

ci_cd/spend_logs_guard.py

@@ -0,0 +1,226 @@
+"""Required check: a pull request that touches LiteLLM_SpendLogs in schema.prisma or a migration
+needs an approving review from the spend-log owner before it can merge.
+
+CODEOWNERS can only match file paths, and every schema change edits the same schema.prisma copies,
+so a path rule would put every migration on the owner. This gate reads the diff instead: a schema.prisma
+hunk counts when it adds or removes a line inside the `model LiteLLM_SpendLogs { ... }` block, a migration
+hunk counts when an added line names the table. Without such a hunk it passes at once, otherwise it passes
+only once the owner's latest review on the pull request is an approval that still covers every spend-log line
+in the head (approvals survive pushes, so a later commit may not add spend-log lines the owner never saw).
+Reviews do not trigger the workflow, so the check is re-run by hand after approving.
+"""
+
+from __future__ import annotations
+
+import argparse
+import re
+import subprocess
+import sys
+from collections.abc import Callable, Mapping, Sequence
+from dataclasses import dataclass
+from pathlib import Path
+from types import MappingProxyType
+from typing import Final
+
+from pydantic import BaseModel, ConfigDict, TypeAdapter
+
+WATCHED: Final = (
+    "schema.prisma",
+    "litellm/proxy/schema.prisma",
+    "litellm-proxy-extras/litellm_proxy_extras/schema.prisma",
+)
+MIGRATIONS_PREFIX: Final = "litellm-proxy-extras/litellm_proxy_extras/migrations/"
+TABLE: Final = re.compile(r"(?<![A-Za-z0-9])LiteLLM_SpendLogs(?![A-Za-z0-9])", re.IGNORECASE)
+DECIDING_STATES: Final = frozenset({"APPROVED", "CHANGES_REQUESTED", "DISMISSED"})
+MODEL_START: Final = re.compile(r"^model\s+LiteLLM_SpendLogs\s*\{")
+HUNK: Final = re.compile(r"^@@ -(\d+)(?:,\d+)? \+(\d+)(?:,\d+)? @@[^\n]*\n", re.MULTILINE)
+
+
+class Reviewer(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+    login: str
+
+
+class Review(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+    state: str
+    user: Reviewer
+    commit_id: str | None = None
+
+
+_REVIEWS: Final = TypeAdapter(list[Review])
+
+
+@dataclass(frozen=True, slots=True)
+class Block:
+    """Lines of `model LiteLLM_SpendLogs { ... }` in one schema.prisma copy at the merge base and at the head."""
+
+    base: range
+    head: range
+
+
+@dataclass(frozen=True, slots=True)
+class Hunk:
+    removed: tuple[tuple[int, str], ...]
+    added: tuple[tuple[int, str], ...]
+
+
+def watched(path: str) -> bool:
+    return path in WATCHED or (path.startswith(MIGRATIONS_PREFIX) and path.endswith("/migration.sql"))
+
+
+def model_block(schema: str) -> range:
+    """1-based line range of `model LiteLLM_SpendLogs { ... }` including both braces, empty when the schema has none."""
+    lines: Final = schema.splitlines()
+    starts: Final = tuple(number for number, line in enumerate(lines, 1) if MODEL_START.match(line))
+    if not starts:
+        return range(0)
+    ends: Final = tuple(number for number, line in enumerate(lines, 1) if number > starts[0] and line.startswith("}"))
+    return range(starts[0], (ends[0] if ends else len(lines)) + 1)
+
+
+def _hunks(section: str) -> tuple[Hunk, ...]:
+    headers: Final = tuple(HUNK.finditer(section))
+    ends: Final = tuple(header.start() for header in headers[1:]) + (len(section),)
+    return tuple(_hunk(header, section[header.end() : end]) for header, end in zip(headers, ends))
+
+
+def _hunk(header: re.Match[str], body: str) -> Hunk:
+    lines: Final = body.splitlines()
+    return Hunk(
+        removed=tuple(enumerate((line[1:] for line in lines if line.startswith("-")), int(header.group(1)))),
+        added=tuple(enumerate((line[1:] for line in lines if line.startswith("+")), int(header.group(2)))),
+    )
+
+
+def spend_log_hits(diff: str, blocks: Mapping[str, Block]) -> tuple[str, ...]:
+    """`path: +line` or `path: -line` for every changed line that touches the table.
+
+    A schema.prisma line counts when its number falls inside that copy's `blocks` entry (head for added lines,
+    merge base for removed ones), a migration line when it is added and names the table."""
+    sections: Final = re.split(r"^diff --git [^\n]*\n", diff, flags=re.MULTILINE)
+    paths: Final = tuple((_path_of(section), section) for section in sections)
+    return tuple(
+        hit for path, section in paths if watched(path) for hit in _section_hits(path, _hunks(section), blocks)
+    )  # comprehension-ok: path then its hits
+
+
+def _section_hits(path: str, hunks: Sequence[Hunk], blocks: Mapping[str, Block]) -> tuple[str, ...]:
+    if path not in blocks:
+        return tuple(f"{path}: +{line.strip()}" for hunk in hunks for _, line in hunk.added if TABLE.search(line))
+    block: Final = blocks[path]
+    removed: Final = tuple(
+        f"{path}: -{line.strip()}" for hunk in hunks for number, line in hunk.removed if number in block.base
+    )  # comprehension-ok: hunk then its lines
+    added: Final = tuple(
+        f"{path}: +{line.strip()}" for hunk in hunks for number, line in hunk.added if number in block.head
+    )  # comprehension-ok: hunk then its lines
+    return removed + added
+
+
+def _path_of(section: str) -> str:
+    match: Final = re.search(r"^\+\+\+ b/(.+)$", section, flags=re.MULTILINE)
+    return match.group(1) if match else ""
+
+
+def approved_by(reviews: Sequence[Review], login: str) -> bool:
+    """True when `login`'s most recent review that changes GitHub's approval state is an approval.
+
+    COMMENTED and PENDING reviews leave an earlier approval in force on GitHub, so they are skipped here too.
+    """
+    own: Final = tuple(review.state for review in reviews if review.user.login == login)
+    states: Final = tuple(state for state in own if state in DECIDING_STATES)
+    return bool(states) and states[-1] == "APPROVED"
+
+
+def latest_approval(reviews: Sequence[Review], login: str) -> Review | None:
+    """`login`'s most recent APPROVED review on the pull request, the one GitHub keeps in force after pushes."""
+    own: Final = tuple(review for review in reviews if review.user.login == login and review.state == "APPROVED")
+    return own[-1] if own else None
+
+
+def approval_covers(
+    approval: Review, head: str, hits: Sequence[str], hits_at: Callable[[str], Sequence[str] | None]
+) -> bool:
+    """True when the approval is on the head, or the head adds no spend-log line the approved commit did not."""
+    if approval.commit_id is None:
+        return False
+    if approval.commit_id == head:
+        return True
+    reviewed: Final = hits_at(approval.commit_id)
+    return reviewed is not None and set(hits) <= set(reviewed)
+
+
+def added_diff(base: str, head: str) -> str:
+    result: Final = subprocess.run(
+        ("git", "diff", "--unified=0", "--no-color", f"{base}..{head}", "--", *WATCHED, MIGRATIONS_PREFIX),
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    return result.stdout
+
+
+def file_at(revision: str, path: str) -> str:
+    """Contents of `path` at `revision`, empty when the file does not exist there."""
+    result: Final = subprocess.run(("git", "show", f"{revision}:{path}"), check=False, capture_output=True, text=True)
+    return result.stdout if result.returncode == 0 else ""
+
+
+def schema_blocks(base: str, head: str) -> Mapping[str, Block]:
+    return MappingProxyType(
+        {path: Block(base=model_block(file_at(base, path)), head=model_block(file_at(head, path))) for path in WATCHED}
+    )
+
+
+def fetched(commit: str) -> bool:
+    result: Final = subprocess.run(("git", "fetch", "--no-tags", "--depth=1", "origin", commit), check=False)
+    return result.returncode == 0
+
+
+def reviewed_hits(base: str, commit: str) -> tuple[str, ...] | None:
+    return spend_log_hits(added_diff(base, commit), schema_blocks(base, commit)) if fetched(commit) else None
+
+
+class Arguments(BaseModel):
+    model_config = ConfigDict(frozen=True)
+    base: str
+    head: str
+    approver: str
+    reviews: Path
+
+
+def main(argv: Sequence[str]) -> int:
+    parser: Final = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--base", required=True, help="merge base of the pull request")
+    parser.add_argument("--head", required=True, help="head sha of the pull request")
+    parser.add_argument("--approver", required=True, help="GitHub login whose approval unlocks the merge")
+    parser.add_argument("--reviews", required=True, help="JSON file with the pull request's reviews")
+    args: Final = Arguments.model_validate(vars(parser.parse_args(argv)))
+    hits: Final = spend_log_hits(added_diff(args.base, args.head), schema_blocks(args.base, args.head))
+    if not hits:
+        sys.stdout.write(
+            "No change inside model LiteLLM_SpendLogs in schema.prisma and no added migration line names it; "
+            "nothing to review.\n"
+        )
+        return 0
+    reviews: Final = _REVIEWS.validate_json(args.reviews.read_bytes())
+    approval: Final = latest_approval(reviews, args.approver) if approved_by(reviews, args.approver) else None
+    if approval is not None and approval_covers(
+        approval, args.head, hits, lambda commit: reviewed_hits(args.base, commit)
+    ):
+        sys.stdout.write(f"LiteLLM_SpendLogs change approved by {args.approver}: {', '.join(hits)}\n")
+        return 0
+    listed: Final = "".join(f"  {hit}\n" for hit in hits)
+    reason: Final = (
+        f"This pull request changes LiteLLM_SpendLogs and needs an approving review from @{args.approver}:\n"
+        if approval is None
+        else f"LiteLLM_SpendLogs lines changed after @{args.approver} approved {approval.commit_id}, "
+        f"a new approving review is needed:\n"
+    )
+    sys.stdout.write(f"{reason}{listed}After approving, re-run this job.\n")
+    return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1:]))

codecov.yaml

@@ -15,9 +15,10 @@ ignore:
 flag_management:
   default_rules:
     carryforward: true
-  # Dead flags no CI job uploads anymore: their carried-forward sessions were
-  # measured against old revisions, and the stale line maps mark comment lines
-  # of since-edited files as missed, sinking patch coverage on unrelated PRs.
+  # Dead flags no CI job uploads anymore, and flags whose job is path-gated on
+  # PRs: their carried-forward sessions were measured against old revisions, and
+  # the stale line maps mark comment lines of since-edited files as missed,
+  # sinking patch coverage on unrelated PRs.
   individual_flags:
     - name: proxy-mgmt-behavior
       carryforward: false
@@ -99,6 +100,10 @@ flag_management:
       carryforward: false
     - name: proxy-db-endpoints-and-responses
       carryforward: false
+    - name: roi-postgres
+      carryforward: false
+    - name: redis-compat
+      carryforward: false
 
 component_management:
   individual_components:

cookbook/ai_coding_tool_guides/chatgpt_subscription_passthrough.md

@@ -0,0 +1,78 @@
+# Use your ChatGPT subscription through LiteLLM
+
+Sign into ChatGPT in Codex or Pi and send model requests through your existing LiteLLM proxy. The client keeps your login and refreshes it. Each request uses that client's ChatGPT account, while a separate LiteLLM virtual key controls gateway access and usage attribution.
+
+An administrator adds a `chatgpt/` model using the usual model configuration. No ChatGPT login needs to be added to the proxy. This uses the same separate-key convention as provider passthrough, including Claude Code subscription forwarding.
+
+## Add a model to the proxy
+
+Reuse your existing proxy. If you need one, follow the [proxy quickstart](https://docs.litellm.ai/docs/proxy/docker_quick_start). Merge this entry into its model list and restart the proxy:
+
+```yaml
+model_list:
+  - model_name: gpt-6-luna
+    litellm_params:
+      model: chatgpt/gpt-6-luna
+    model_info:
+      mode: responses
+```
+
+Use a model available to the caller's subscription. The virtual key must allow the exposed model name. Leave saved ChatGPT logins and API credentials unset for a client-only deployment. Explicit cross-provider fallbacks can incur separate API charges; do not configure them for subscription-only traffic.
+
+## Configure Codex
+
+Use your existing LiteLLM URL and a virtual key assigned to your gateway user. Export that key as `LITELLM_API_KEY` in the environment that launches Codex. Keep the value out of committed configuration.
+
+Add this provider to `~/.codex/config.toml`, preserving your other settings:
+
+```toml
+model = "gpt-6-luna"
+model_provider = "litellm_subscription"
+
+[model_providers.litellm_subscription]
+name = "LiteLLM subscription"
+base_url = "https://litellm.example.com/v1"
+wire_api = "responses"
+supports_websockets = false
+requires_openai_auth = true
+env_http_headers = { "x-litellm-api-key" = "LITELLM_API_KEY" }
+```
+
+Replace `https://litellm.example.com` with your proxy URL. A local proxy normally uses `http://localhost:4000/v1`. Launch Codex, sign into ChatGPT if needed, and send a prompt with a model available to your subscription. The selected model must match the name exposed by the gateway.
+
+The `lite` CLI can set up this provider each time it starts Codex, with no `config.toml` edit. Export `LITELLM_PROXY_URL` and `LITELLM_PROXY_API_KEY`, then run `lite codex --subscription`. Run `lite config set codex_subscription true` to make it the default.
+
+For a managed workspace rollout, use OpenAI's [gateway configuration guide](https://learn.chatgpt.com/docs/enterprise/sign-in-with-chatgpt-through-a-gateway).
+
+## Native compatibility and Pi
+
+For a native Codex model catalog and request format, change Codex's `base_url` to `https://litellm.example.com/backend-api/codex`. This route uses the same credentials but does not need a model deployment or resolve model aliases. Restricted keys must allow the upstream model name.
+
+### Configure Pi
+
+Sign into the built-in `openai-codex` provider in Pi. Merge this override into `~/.pi/agent/models.json` and export your LiteLLM virtual key as `LITELLM_API_KEY` before launching Pi:
+
+```json
+{
+  "providers": {
+    "openai-codex": {
+      "baseUrl": "https://litellm.example.com/backend-api/codex",
+      "headers": {
+        "x-litellm-api-key": "LITELLM_API_KEY"
+      }
+    }
+  }
+}
+```
+
+Keep Pi's subscription login as the provider credential. The header value names an environment variable, not a key to paste into the file.
+
+## Verify account usage
+
+Send a prompt and check the completed request in LiteLLM's Logs page. Its gateway user/key and token counts should match the caller. Give each person their own LiteLLM virtual key for separate attribution. The gateway credential and ChatGPT login are independently authenticated; LiteLLM does not establish that both identities belong to the same person.
+
+The native routes are `GET /backend-api/codex/models` and `POST /backend-api/codex/responses`. Model requests use the upstream model name, so a restricted LiteLLM key must allow that name. OpenAI's model catalog is returned in its native format; seeing a model there does not override gateway permissions.
+
+The proxy forwards the ChatGPT bearer and account header to the fixed OpenAI Codex endpoint. It removes gateway credentials and cookies from the upstream request and redacts credential headers in logging callbacks. The client remains responsible for refresh and reauthentication. Native passthrough returns authentication and quota errors without switching subscriptions. Standard model routes retain the caller login across ChatGPT retries; explicitly configured cross-provider fallbacks use their own credentials. An explicit workspace header is preserved for OpenAI to validate.
+
+This flow does not connect accounts through the dashboard or pool subscription allowances. Use a Codex build that supports ChatGPT authentication with a custom model provider. The integration forwards HTTP and SSE requests; it does not add native WebSocket or other Codex backend routes.

cookbook/gpt_realtime_translate.py

@@ -0,0 +1,245 @@
+#!/usr/bin/env python3
+
+import argparse
+import asyncio
+import base64
+import json
+import os
+import sys
+import wave
+from collections.abc import Iterator, Sequence
+from dataclasses import dataclass
+from pathlib import Path
+from urllib.parse import urlencode, urlsplit, urlunsplit
+
+import websockets
+from websockets.asyncio.client import ClientConnection
+
+SAMPLE_RATE = 24_000
+CHANNELS = 1
+SAMPLE_WIDTH = 2
+CHUNK_DURATION_SECONDS = 0.1
+CHUNK_BYTES = int(SAMPLE_RATE * CHANNELS * SAMPLE_WIDTH * CHUNK_DURATION_SECONDS)
+RESPONSE_TIMEOUT_SECONDS = 30.0
+AUDIO_EVENT_TYPES = frozenset(
+    {
+        "session.output_audio.delta",
+        "response.audio.delta",
+        "response.output_audio.delta",
+    }
+)
+TRANSCRIPT_EVENT_TYPES = frozenset(
+    {
+        "session.output_transcript.delta",
+        "response.text.delta",
+        "response.output_audio_transcript.delta",
+    }
+)
+
+
+@dataclass(frozen=True, slots=True)
+class Settings:
+    input_wav: Path
+    output_wav: Path
+    base_url: str
+    model: str
+    target_language: str
+    trailing_silence_seconds: float
+    api_key: str
+
+
+def write_stdout(message: str = "", *, end: str = "\n", flush: bool = False) -> None:
+    sys.stdout.write(f"{message}{end}")
+    if flush:
+        sys.stdout.flush()
+
+
+def write_stderr(message: str) -> None:
+    sys.stderr.write(f"{message}\n")
+
+
+def parse_args(argv: Sequence[str] | None = None) -> Settings | str:
+    parser = argparse.ArgumentParser(
+        description="Stream a 24 kHz PCM16 WAV through gpt-realtime-translate and save the translated audio",
+    )
+    parser.add_argument("input_wav", type=Path)
+    parser.add_argument("--output", type=Path, default=Path("translated.wav"))
+    parser.add_argument("--base-url", default=os.getenv("LITELLM_BASE_URL", "http://localhost:4000"))
+    parser.add_argument("--model", default=os.getenv("REALTIME_TRANSLATE_MODEL", "gpt-realtime-translate"))
+    parser.add_argument("--target-language", default="fr")
+    parser.add_argument("--trailing-silence", type=float, default=1.5)
+    parsed = parser.parse_args(argv)
+    api_key = os.getenv("LITELLM_API_KEY") or os.getenv("OPENAI_API_KEY")
+    if not api_key:
+        return "Set LITELLM_API_KEY or OPENAI_API_KEY before running the script"
+    if parsed.trailing_silence < 0:
+        return "--trailing-silence must be zero or greater"
+    return Settings(
+        input_wav=parsed.input_wav,
+        output_wav=parsed.output,
+        base_url=parsed.base_url,
+        model=parsed.model,
+        target_language=parsed.target_language,
+        trailing_silence_seconds=parsed.trailing_silence,
+        api_key=api_key,
+    )
+
+
+def translation_url(base_url: str, model: str) -> str | None:
+    parsed = urlsplit(base_url.rstrip("/"))
+    scheme = {"http": "ws", "https": "wss", "ws": "ws", "wss": "wss"}.get(parsed.scheme)
+    if not scheme or not parsed.netloc:
+        return None
+    base_path = parsed.path.rstrip("/")
+    realtime_path = (
+        f"{base_path}/realtime/translations" if base_path.endswith("/v1") else f"{base_path}/v1/realtime/translations"
+    )
+    return urlunsplit((scheme, parsed.netloc, realtime_path, urlencode({"model": model}), ""))
+
+
+def read_pcm16_wav(path: Path) -> bytes | str:
+    try:
+        with wave.open(str(path), "rb") as source:
+            actual_format = (
+                source.getnchannels(),
+                source.getsampwidth(),
+                source.getframerate(),
+                source.getcomptype(),
+            )
+            expected_format = (CHANNELS, SAMPLE_WIDTH, SAMPLE_RATE, "NONE")
+            if actual_format != expected_format:
+                return (
+                    f"{path} must be mono, 16-bit PCM, 24 kHz WAV; received "
+                    f"channels={actual_format[0]}, sample_width={actual_format[1]}, "
+                    f"sample_rate={actual_format[2]}, compression={actual_format[3]}"
+                )
+            return source.readframes(source.getnframes())
+    except (OSError, EOFError, wave.Error) as exc:
+        return f"Unable to read {path}: {exc}"
+
+
+def audio_chunks(audio: bytes) -> Iterator[bytes]:
+    return (audio[offset : offset + CHUNK_BYTES] for offset in range(0, len(audio), CHUNK_BYTES))
+
+
+def audio_message(audio: bytes) -> str:
+    return json.dumps(
+        {
+            "type": "session.input_audio_buffer.append",
+            "audio": base64.b64encode(audio).decode("ascii"),
+        }
+    )
+
+
+async def configure_session(connection: ClientConnection, target_language: str) -> str | None:
+    await connection.send(
+        json.dumps(
+            {
+                "type": "session.update",
+                "session": {"audio": {"output": {"language": target_language}}},
+            }
+        )
+    )
+    while True:
+        raw_event = await asyncio.wait_for(connection.recv(), timeout=20)
+        event = json.loads(raw_event)
+        event_type = event.get("type")
+        if event_type == "session.created":
+            write_stdout(f"Session: {event.get('session', {}).get('id', 'created')}")
+        if event_type == "session.updated":
+            return None
+        if event_type == "error":
+            return f"Realtime API error: {json.dumps(event.get('error', event), ensure_ascii=False)}"
+
+
+async def send_audio(
+    connection: ClientConnection, pcm: bytes, trailing_silence_seconds: float, finished: asyncio.Event
+) -> None:
+    silence = bytes(round(trailing_silence_seconds * SAMPLE_RATE * CHANNELS * SAMPLE_WIDTH))
+    try:
+        for chunk in audio_chunks(pcm + silence):
+            await connection.send(audio_message(chunk))
+            await asyncio.sleep(len(chunk) / (SAMPLE_RATE * CHANNELS * SAMPLE_WIDTH))
+        await connection.send(json.dumps({"type": "session.close"}))
+    finally:
+        finished.set()
+
+
+async def receive_translation(
+    connection: ClientConnection, output_path: Path, sender_finished: asyncio.Event
+) -> str | None:
+    audio_received = asyncio.Event()
+    try:
+        with wave.open(str(output_path), "wb") as output:
+            output.setnchannels(CHANNELS)
+            output.setsampwidth(SAMPLE_WIDTH)
+            output.setframerate(SAMPLE_RATE)
+            write_stdout("Translation: ", end="", flush=True)
+            while True:
+                try:
+                    raw_event = await asyncio.wait_for(connection.recv(), timeout=RESPONSE_TIMEOUT_SECONDS)
+                except TimeoutError:
+                    if not sender_finished.is_set():
+                        continue
+                    return (
+                        f"The translation stream went quiet for {RESPONSE_TIMEOUT_SECONDS:g} s without closing the session"
+                    )
+                event = json.loads(raw_event)
+                event_type = event.get("type")
+                if event_type in AUDIO_EVENT_TYPES:
+                    output.writeframes(base64.b64decode(event.get("delta", ""), validate=True))
+                    audio_received.set()
+                elif event_type in TRANSCRIPT_EVENT_TYPES:
+                    write_stdout(event.get("delta", event.get("text", "")), end="", flush=True)
+                elif event_type == "error":
+                    return f"Realtime API error: {json.dumps(event.get('error', event), ensure_ascii=False)}"
+                elif event_type == "session.closed":
+                    write_stdout()
+                    return None if audio_received.is_set() else "The translation stream ended without translated audio"
+    except (OSError, wave.Error) as exc:
+        return f"Unable to write {output_path}: {exc}"
+
+
+async def translate(settings: Settings, pcm: bytes) -> str | None:
+    url = translation_url(settings.base_url, settings.model)
+    if not url:
+        return f"Invalid --base-url: {settings.base_url}"
+    sender_finished = asyncio.Event()
+    try:
+        async with websockets.connect(
+            url,
+            additional_headers={"Authorization": f"Bearer {settings.api_key}"},
+            proxy=None,
+            open_timeout=20,
+            close_timeout=5,
+        ) as connection:
+            configuration_error = await configure_session(connection, settings.target_language)
+            if configuration_error:
+                return configuration_error
+            async with asyncio.TaskGroup() as tasks:
+                receiver = tasks.create_task(receive_translation(connection, settings.output_wav, sender_finished))
+                tasks.create_task(send_audio(connection, pcm, settings.trailing_silence_seconds, sender_finished))
+            return receiver.result()
+    except Exception as exc:
+        return f"Translation failed: {type(exc).__name__}: {exc}"
+
+
+def main(argv: Sequence[str] | None = None) -> int:
+    settings = parse_args(argv)
+    if isinstance(settings, str):
+        write_stderr(settings)
+        return 2
+    pcm = read_pcm16_wav(settings.input_wav)
+    if isinstance(pcm, str):
+        write_stderr(pcm)
+        return 2
+    error = asyncio.run(translate(settings, pcm))
+    if error:
+        write_stderr(error)
+        return 1
+    write_stdout(f"Translated audio: {settings.output_wav}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

cookbook/litellm_proxy_server/batch_api/bedrock/README.md

@@ -0,0 +1,41 @@
+# Bedrock batch endpoints
+
+Set `aws_bedrock_control_plane_endpoint` on a deployment when batch create, retrieve, and cancel requests must use a custom Bedrock control-plane endpoint. The proxy signs these requests for the configured URL. Without this setting, batch requests use the standard regional Bedrock endpoint
+
+```yaml
+model_list:
+  - model_name: bedrock-batch
+    litellm_params:
+      model: bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0
+      aws_region_name: us-east-1
+      aws_access_key_id: os.environ/AWS_ACCESS_KEY_ID
+      aws_secret_access_key: os.environ/AWS_SECRET_ACCESS_KEY
+      aws_batch_role_arn: os.environ/AWS_BATCH_ROLE_ARN
+      s3_output_bucket_name: os.environ/AWS_S3_OUTPUT_BUCKET_NAME
+      aws_bedrock_control_plane_endpoint: os.environ/BEDROCK_CONTROL_PLANE_ENDPOINT
+```
+
+Set `BEDROCK_CONTROL_PLANE_ENDPOINT` to your Bedrock control-plane VPC endpoint or compatible gateway base URL, without `/model-invocation-job`. The proxy adds the operation path. The endpoint must be reachable from the proxy, and the configured AWS region must match the endpoint's signing region
+
+`api_base` and `aws_bedrock_runtime_endpoint` configure runtime inference separately. They do not override the batch control-plane endpoint. `s3_endpoint_url` configures S3 file access separately
+
+The control-plane setting belongs in deployment configuration. Sending `aws_bedrock_control_plane_endpoint` in an API request body is rejected. There is no automatic `AWS_BEDROCK_CONTROL_PLANE_ENDPOINT` environment fallback, so reference the desired environment variable explicitly as shown above
+
+Start the proxy with this configuration and set `LITELLM_API_BASE` to its URL. Set `BEDROCK_BATCH_INPUT_URI` to a valid input JSONL file in S3 that the batch service role can read
+
+```bash
+curl -sS "$LITELLM_API_BASE/v1/batches" \
+  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
+  -H 'Content-Type: application/json' \
+  -d "{\"model\":\"bedrock-batch\",\"input_file_id\":\"$BEDROCK_BATCH_INPUT_URI\",\"endpoint\":\"/v1/chat/completions\",\"completion_window\":\"24h\"}"
+```
+
+Use the returned `id` as `BATCH_ID` to retrieve or cancel the same job through the deployment's configured endpoint
+
+```bash
+curl -sS "$LITELLM_API_BASE/v1/batches/$BATCH_ID?model=bedrock-batch" \
+  -H "Authorization: Bearer $LITELLM_MASTER_KEY"
+
+curl -sS -X POST "$LITELLM_API_BASE/v1/batches/$BATCH_ID/cancel?model=bedrock-batch" \
+  -H "Authorization: Bearer $LITELLM_MASTER_KEY"
+```

cookbook/litellm_proxy_server/grafana_dashboard/dashboard_all_metrics/grafana_dashboard.json

@@ -6563,6 +6563,87 @@
       ],
       "title": "litellm_project_model_rate_limit_used_metric",
       "type": "timeseries"
+    },
+    {
+      "collapsed": false,
+      "gridPos": {
+        "h": 1,
+        "w": 24,
+        "x": 0,
+        "y": 447
+      },
+      "id": 115,
+      "panels": [],
+      "title": "Agent tracing export (needs LITELLM_LENS_URL)",
+      "type": "row"
+    },
+    {
+      "datasource": {
+        "type": "prometheus",
+        "uid": "${DS_PROMETHEUS}"
+      },
+      "description": "Request records the Lens exporter delivered to the Lens service against the records it dropped before the service accepted them",
+      "fieldConfig": {
+        "defaults": {
+          "color": {
+            "mode": "palette-classic"
+          },
+          "custom": {
+            "drawStyle": "line",
+            "fillOpacity": 10,
+            "lineWidth": 1,
+            "showPoints": "never",
+            "spanNulls": false
+          },
+          "unit": "short"
+        },
+        "overrides": []
+      },
+      "gridPos": {
+        "h": 8,
+        "w": 24,
+        "x": 0,
+        "y": 448
+      },
+      "id": 116,
+      "options": {
+        "legend": {
+          "calcs": [],
+          "displayMode": "list",
+          "placement": "bottom",
+          "showLegend": true
+        },
+        "tooltip": {
+          "mode": "multi",
+          "sort": "desc"
+        }
+      },
+      "targets": [
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${DS_PROMETHEUS}"
+          },
+          "editorMode": "code",
+          "expr": "sum(rate(litellm_lens_export_rows_written_total[$__rate_interval]))",
+          "legendFormat": "written",
+          "range": true,
+          "refId": "A"
+        },
+        {
+          "datasource": {
+            "type": "prometheus",
+            "uid": "${DS_PROMETHEUS}"
+          },
+          "editorMode": "code",
+          "expr": "sum(rate(litellm_lens_export_rows_dropped_total[$__rate_interval]))",
+          "legendFormat": "dropped",
+          "range": true,
+          "refId": "B"
+        }
+      ],
+      "title": "litellm_lens_export_rows_written and litellm_lens_export_rows_dropped rate",
+      "type": "timeseries"
     }
   ],
   "preload": false,

cookbook/litellm_proxy_server/grafana_dashboard/dashboard_all_metrics/readme.md

@@ -1,10 +1,10 @@
 # LiteLLM All Prometheus Metrics dashboard
 
-Every `litellm_*` metric family the proxy can expose on `/metrics` (141 families across 97 panels), grouped into rows: proxy traffic, latency, spend and tokens, cache, LLM API deployments, key and team rate limits, budgets, guardrails, MCP, managed files and batches, users and teams, the Redis circuit breaker, the spend log cleanup job, and the `prometheus_system` service callback metrics (per-service latency, request and failure rates, spend update queue sizes). Panel titles are the metric names so you can grep the JSON for the metric you care about
+Every `litellm_*` metric family the proxy can expose on `/metrics` (143 families across 98 panels), grouped into rows: proxy traffic, latency, spend and tokens, cache, LLM API deployments, key and team rate limits, budgets, guardrails, MCP, managed files and batches, users and teams, the Redis circuit breaker, the spend log cleanup job, the `prometheus_system` service callback metrics (per-service latency, request and failure rates, spend update queue sizes), project model rate limits, and the agent tracing export to Lens. Panel titles are the metric names so you can grep the JSON for the metric you care about
 
 Import `grafana_dashboard.json` from **Dashboards > New > Import** and pick your Prometheus data source when prompted (the `DS_PROMETHEUS` variable). Counters are plotted as `rate()` over `$__rate_interval`, histograms as p50 / p95 / p99, gauges as the raw value grouped by the most useful label. Every query names the metric exactly as the proxy emits it (counters carry the `_total` suffix the Prometheus client adds), and `tests/unit/integrations/test_prometheus_metric_name_consistency.py` fails if a metric is renamed without updating this dashboard
 
-The first eleven rows need only `callbacks: ["prometheus"]`. The last three rows and the `litellm_admission_*` panels are emitted by other subsystems and stay empty until those are on: the service callback row needs `service_callback: ["prometheus_system"]` in `litellm_settings`, the circuit breaker row needs a Redis cache, the cleanup row needs spend log retention, and admission control needs its middleware enabled. Within the base rows, many panels only fill in once the matching feature is in use: budgets need keys, teams, users or orgs with `max_budget` set, cache panels need caching on, guardrail and MCP panels need those features configured, deployment health needs the router with more than one deployment or a failure to record, and `litellm_in_flight_requests` needs traffic at scrape time. An empty panel for a feature you do not use is expected
+The first eleven rows and the project model rate limits row need only `callbacks: ["prometheus"]`. The other rows and the `litellm_admission_*` panels are emitted by other subsystems and stay empty until those are on: the service callback row needs `service_callback: ["prometheus_system"]` in `litellm_settings`, the circuit breaker row needs a Redis cache, the cleanup row needs spend log retention, the agent tracing export row needs `LITELLM_LENS_URL`, and admission control needs its middleware enabled. Within the base rows, many panels only fill in once the matching feature is in use: budgets need keys, teams, users or orgs with `max_budget` set, cache panels need caching on, guardrail and MCP panels need those features configured, deployment health needs the router with more than one deployment or a failure to record, and `litellm_in_flight_requests` needs traffic at scrape time. An empty panel for a feature you do not use is expected
 
 ## Pre-requisites
 

cosign.pub

@@ -1,4 +1,4 @@
 -----BEGIN PUBLIC KEY-----
-MFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAEKi4ivqGpE231OGH50PKbqy1Y1Kkb
-POJC8+i2Wko82gBOUCe3M0Vw86H/4rhUhfoYEti4gdJ9wZbYmK0I2EE96g==
+MFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAEG06dpKM+vEDnmp2zcjyXgVCXxzRg
+JA5H2LnmwBKQafZGka6OOrWujCEWATQewiAgyD4WQdJ0pvU9RC4mvNAjvQ==
 -----END PUBLIC KEY-----

db_scripts/strip_vector_store_internal_metadata.py

@@ -0,0 +1,201 @@
+import argparse
+import os
+import sys
+from collections.abc import Iterator, Sequence
+from dataclasses import dataclass
+from itertools import accumulate, repeat, takewhile
+from typing import Final, Literal, TextIO
+
+import httpx
+from pydantic import BaseModel, ConfigDict, JsonValue, ValidationError
+
+from litellm.vector_stores.utils import is_litellm_written_metadata_key
+
+DEFAULT_API_BASE: Final = "https://api.openai.com/v1"
+LIST_PAGE_SIZE: Final = "100"
+
+
+class VectorStore(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+
+    id: str
+    metadata: dict[str, JsonValue] | None = None
+
+
+class VectorStorePage(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+
+    data: tuple[VectorStore, ...]
+    has_more: bool = False
+
+
+class OpenAIErrorBody(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+
+    message: str = ""
+
+
+class OpenAIErrorResponse(BaseModel):
+    model_config = ConfigDict(frozen=True, extra="ignore")
+
+    error: OpenAIErrorBody
+
+
+@dataclass(frozen=True, slots=True)
+class Options:
+    api_key: str
+    api_base: str
+    vector_store_ids: tuple[str, ...]
+    extra_keys: frozenset[str]
+    apply: bool
+
+
+@dataclass(frozen=True, slots=True)
+class StoreResult:
+    store_id: str
+    matched_keys: tuple[str, ...]
+    status: Literal["dry_run", "updated", "failed"]
+    error: str = ""
+
+
+def is_proxy_written_key(key: str, extra_keys: frozenset[str]) -> bool:
+    return is_litellm_written_metadata_key(key) or key in extra_keys
+
+
+def matching_keys(store: VectorStore, extra_keys: frozenset[str]) -> tuple[str, ...]:
+    return tuple(sorted(key for key in (store.metadata or {}) if is_proxy_written_key(key, extra_keys)))
+
+
+def fetch_page(client: httpx.Client, after: str | None) -> VectorStorePage:
+    params: Final = {"limit": LIST_PAGE_SIZE} if after is None else {"limit": LIST_PAGE_SIZE, "after": after}
+    return VectorStorePage.model_validate_json(client.get("/vector_stores", params=params).raise_for_status().content)
+
+
+def fetch_next_page(client: httpx.Client, page: VectorStorePage | None) -> VectorStorePage | None:
+    if page is None or not page.has_more or not page.data:
+        return None
+    return fetch_page(client, after=page.data[-1].id)
+
+
+def list_all_stores(client: httpx.Client) -> Iterator[VectorStore]:
+    pages: Final = accumulate(
+        repeat(None), lambda page, _: fetch_next_page(client, page), initial=fetch_page(client, after=None)
+    )
+    for page in takewhile(lambda page: page is not None, pages):
+        yield from page.data
+
+
+def retrieve_stores(client: httpx.Client, store_ids: tuple[str, ...]) -> Iterator[VectorStore]:
+    for store_id in store_ids:
+        yield VectorStore.model_validate_json(client.get(f"/vector_stores/{store_id}").raise_for_status().content)
+
+
+def describe_http_error(response: httpx.Response) -> str:
+    try:
+        message: Final = OpenAIErrorResponse.model_validate_json(response.content).error.message
+    except ValidationError:
+        return f"HTTP {response.status_code}"
+    return f"HTTP {response.status_code}: {message}"
+
+
+def delete_keys(client: httpx.Client, store_id: str, keys: tuple[str, ...]) -> StoreResult:
+    try:
+        response: Final = client.post(f"/vector_stores/{store_id}", json={"metadata": dict.fromkeys(keys)})
+    except httpx.TransportError as error:
+        return StoreResult(store_id=store_id, matched_keys=keys, status="failed", error=type(error).__name__)
+    if response.is_success:
+        return StoreResult(store_id=store_id, matched_keys=keys, status="updated")
+    return StoreResult(store_id=store_id, matched_keys=keys, status="failed", error=describe_http_error(response))
+
+
+def clean_store(client: httpx.Client, store: VectorStore, options: Options) -> StoreResult | None:
+    keys: Final = matching_keys(store, options.extra_keys)
+    if not keys:
+        return None
+    if not options.apply:
+        return StoreResult(store_id=store.id, matched_keys=keys, status="dry_run")
+    return delete_keys(client, store.id, keys)
+
+
+def format_result(result: StoreResult) -> str:
+    key_list: Final = ", ".join(result.matched_keys)
+    count: Final = len(result.matched_keys)
+    match result.status:
+        case "dry_run":
+            return f"{result.store_id}: would delete {count} keys: {key_list}"
+        case "updated":
+            return f"{result.store_id}: deleted {count} keys: {key_list}"
+        case "failed":
+            return f"{result.store_id}: failed to delete {count} keys ({result.error}): {key_list}"
+
+
+def clean_and_report(
+    client: httpx.Client, stores: Iterator[VectorStore], options: Options, out: TextIO
+) -> Iterator[StoreResult | None]:
+    for store in stores:
+        result: Final = clean_store(client, store, options)
+        if result is not None:
+            print(format_result(result), file=out)
+        yield result
+
+
+def run(options: Options, client: httpx.Client, out: TextIO) -> int:
+    stores: Final = (
+        retrieve_stores(client, options.vector_store_ids) if options.vector_store_ids else list_all_stores(client)
+    )
+    scanned: Final = tuple(clean_and_report(client, stores, options, out))
+    results: Final = tuple(result for result in scanned if result is not None)
+    updated: Final = sum(1 for result in results if result.status == "updated")
+    failed: Final = sum(1 for result in results if result.status == "failed")
+    mode: Final = "applied" if options.apply else "dry run, pass --apply to delete"
+    print(
+        f"scanned={len(scanned)} matched={len(results)} updated={updated} failed={failed} ({mode})",
+        file=out,
+    )
+    return 1 if failed else 0
+
+
+def parse_options(argv: Sequence[str] | None) -> Options:
+    parser: Final = argparse.ArgumentParser(
+        description="Remove LiteLLM proxy-written keys from OpenAI vector store metadata. Dry run by default."
+    )
+    parser.add_argument("--api-key", default=os.environ.get("OPENAI_API_KEY"), help="defaults to $OPENAI_API_KEY")
+    parser.add_argument("--api-base", default=DEFAULT_API_BASE)
+    parser.add_argument("--vector-store-id", action="append", default=[], help="repeatable, defaults to every store")
+    parser.add_argument("--extra-key", action="append", default=[], help="repeatable, another key name to delete")
+    parser.add_argument("--apply", action="store_true", help="delete the matching keys instead of listing them")
+    args: Final = parser.parse_args(argv)
+    if not args.api_key:
+        parser.error("--api-key or OPENAI_API_KEY is required")
+    return Options(
+        api_key=args.api_key,
+        api_base=args.api_base.rstrip("/"),
+        vector_store_ids=tuple(args.vector_store_id),
+        extra_keys=frozenset(args.extra_key),
+        apply=args.apply,
+    )
+
+
+def main(argv: Sequence[str] | None = None, transport: httpx.BaseTransport | None = None) -> int:
+    options: Final = parse_options(argv)
+    with httpx.Client(
+        base_url=options.api_base,
+        headers={"Authorization": f"Bearer {options.api_key}"},
+        timeout=60.0,
+        transport=transport,
+    ) as client:
+        try:
+            return run(options, client, sys.stdout)
+        except httpx.HTTPStatusError as error:
+            sys.stderr.write(f"could not read vector stores: {describe_http_error(error.response)}\n")
+            return 2
+        except httpx.TransportError as error:
+            sys.stderr.write(f"could not reach {options.api_base}: {type(error).__name__}\n")
+            return 2
+        except ValidationError:
+            sys.stderr.write(f"unexpected vector store response shape from {options.api_base}\n")
+            return 2
+
+
+if __name__ == "__main__":
+    sys.exit(main())

docker/Dockerfile.database

@@ -1,10 +1,10 @@
 # syntax=docker/dockerfile:1.7
 
 # Base image for building
-ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 
 # Runtime image
-ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
 # Pinned by digest like the other base images; bump explicitly on Node upgrades.
 ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@@ -59,6 +59,7 @@ RUN apk add --no-cache \
     python-3.13 \
     python-3.13-dev \
     openssl \
+    libssl3 \
     openssl-dev \
     nodejs \
     npm \
@@ -122,7 +123,7 @@ ENV LITELLM_RELEASE_TAG=${LITELLM_RELEASE_TAG}
 USER root
 
 # node (without npm) is required by the prisma CLI at runtime
-RUN apk add --no-cache bash openssl tzdata nodejs python-3.13 libsndfile libevent
+RUN apk add --no-cache bash openssl libssl3 tzdata nodejs python-3.13 libsndfile libevent
 COPY --from=pgbouncer-builder /usr/local/bin/pgbouncer /usr/local/bin/pgbouncer
 
 WORKDIR /app

docker/Dockerfile.non_root

@@ -1,8 +1,8 @@
 # syntax=docker/dockerfile:1.7
 
 # Base images
-ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
-ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
+ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
 # Pinned by digest like the other base images; bump explicitly on Node upgrades.
 ARG UI_BUILD_IMAGE=node:24.19-alpine3.24@sha256:d32cdf619f63fe0471182d08996dd516c6275bb5fd31ae06e55a570bd9e1ad43
@@ -60,6 +60,7 @@ RUN for i in 1 2 3; do \
         coreutils \
         curl \
         openssl \
+        libssl3 \
         libsndfile \
         nodejs \
         npm && break || sleep 5; \
@@ -134,7 +135,7 @@ RUN for i in 1 2 3; do \
       apk upgrade --no-cache && break || sleep 5; \
     done && \
     for i in 1 2 3; do \
-      apk add --no-cache python-3.13 bash openssl tzdata libsndfile nodejs libevent && break || sleep 5; \
+      apk add --no-cache python-3.13 bash openssl libssl3 tzdata libsndfile nodejs libevent && break || sleep 5; \
     done
 COPY --from=pgbouncer-builder /usr/local/bin/pgbouncer /usr/local/bin/pgbouncer
 

docker/build_from_pip/Dockerfile.build_from_pip

@@ -30,7 +30,7 @@ RUN uv venv --python python && \
       "google-cloud-aiplatform==1.133.0" \
       "google-genai==1.37.0" \
       "anthropic[vertex]==0.84.0" \
-      "grpcio==1.78.0" \
+      "grpcio==1.84.1" \
       "prometheus-client==0.20.0" \
       "langfuse==2.59.7" \
       "opentelemetry-api==1.28.0" \

docker/docker-compose.quickstart.yml

@@ -11,7 +11,7 @@
 # beyond local evaluation, pin the image to a specific release tag.
 services:
   litellm:
-    image: docker.litellm.ai/berriai/litellm:main-stable
+    image: docker.litellm.ai/berriai/litellm:latest
     ports:
       # LITELLM_BIND is empty by default, so this stays "4000:4000". The quickstart
       # script sets it to "127.0.0.1:" so new installs listen on this machine only.

docs/adrs/AGENTS.md

@@ -0,0 +1,5 @@
+# Architecture decision records for the Python SDK and proxy
+
+This folder records decisions about the Python SDK and proxy the way [`litellm-rust/docs/adrs/`](../../litellm-rust/docs/adrs/AGENTS.md) records them for the Rust core, and that file's process applies here unchanged: one short `adr_NNN_short_title.md` per decision that is hard to reverse or that a new contributor would reasonably question, with Context, Decision, Alternatives Considered, and Consequences, superseded by a new ADR instead of edited once accepted. Each folder numbers its own ADRs, so a reference to an ADR in the other folder names its path
+
+Read the ADRs here before you build a new way to move data between the endpoints, the provider transformations, and cost tracking or logging. A PR that makes a decision of that kind adds its ADR in the same PR

docs/adrs/adr_001_provider_usage_fields_ride_as_usage_extras.md

@@ -0,0 +1,39 @@
+# ADR 001: Provider usage fields ride as extras on the normalized usage object
+
+Status: Accepted
+
+## Context
+
+Providers report usage fields that no other provider has, such as xAI's `usage.server_side_tool_usage_details.web_search_calls` or Anthropic's `usage.inference_geo`, and cost tracking has to read them or that spend never reaches `LiteLLM_SpendLogs`. One provider response can be served on `/v1/chat/completions`, `/v1/responses`, and `/v1/messages`, streaming or not, so a field that only one of those paths carries is billed wrong on the others
+
+[#30817](https://github.com/BerriAI/litellm/pull/30817) fixed unbilled xAI web searches. One draft did the copy in the xAI Responses transform to keep provider code out of shared files, and that swapped `/v1/responses` usage to the chat shape, so callers of that endpoint lost the Responses schema. The merged version keeps the schema, copies unknown usage fields in the shared Responses-to-chat bridge for every provider, and gates on the reported field without naming the provider
+
+In October 2026, `Usage` and `ResponseAPIUsage` kept unknown usage keys as pydantic extras, and that bridge copied them. Paths that build usage from named fields kept only the fields they name: the Anthropic, Gemini, and Bedrock Converse transforms, the stream builder, and the bridges from chat usage to `/v1/responses` and `/v1/messages`
+
+## Decision
+
+Provider-only usage fields stay on the usage object as pydantic extras, and cost tracking reads them there, in that provider's cost calculator or in a shared gate that checks the field instead of the provider. An extra reaches callers as an added usage key. No transform renames, removes, or reshapes a usage field the endpoint's schema defines to feed cost tracking, though it may fill one LiteLLM's usage object already defines, such as `server_tool_use`
+
+A shared path that rebuilds usage, such as the stream builder or a bridge between endpoint shapes, copies the extras instead of naming provider fields. A provider transform that builds usage from named fields names the new field. Neither is bypassed with a side channel. Older code that differs, such as the Azure branches in the shared tracker and the Gemini pricing tier read from `_hidden_params`, predates this ADR
+
+## Alternatives Considered
+
+**Reshape usage in the provider transform**, as a draft of #30817 did. It changes the schema that callers of that endpoint rely on, and every provider and endpoint needs its own copy
+
+**A typed field on `ResponseAPIUsage` and `Usage` for each provider's quirk**. The types grow without bound and every bridge needs an update for each one. A field several providers report, like `server_tool_use`, does get one
+
+**A side channel such as `_hidden_params`**. Cost calculators receive `Usage`, so a side channel makes a second source of truth for the same number
+
+**Provider branches or provider field names in shared code**. They put one provider's logic in a file every provider shares, which is why #30817 replaced its xAI-only gate with one that checks the reported field
+
+## Consequences
+
+Benefits:
+
+- Billing a new provider usage field touches that provider's calculator (plus a `cost_per_token` arm when the calculator is new), its transform when that names fields, and for a built-in tool at most one gate and one `get_cost_for_web_search_request` arm
+
+Costs:
+
+- Extras are untyped, so a reader validates them before use, as `_usage_reports_server_side_web_search_calls` does
+- The stream builder and the bridges from chat usage did not copy extras when this was written, so until they do, a field billed on those paths is named in them, and tests and proof runs cover each endpoint and `"stream": true`
+- This covers usage fields LiteLLM prices, not request-side params or a provider's own charge for the request, which OpenRouter and opencode Go report and `response_cost_calculator` reads from `_hidden_params` first

enterprise/enterprise_hooks/aporia_ai.py

@@ -15,9 +15,10 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import json
 import sys
-from typing import Any, List, Literal, Optional
+from typing import Final, List, Literal, Optional
 
 from fastapi import HTTPException
+from pydantic import TypeAdapter
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -29,11 +30,13 @@ from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     httpxSpecialProvider,
 )
+from litellm.llms.openai.common_utils import decoded_json_object
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.guardrails.guardrail_helpers import should_proceed_based_on_metadata
 from litellm.types.guardrails import GuardrailEventHooks
 
 GUARDRAIL_NAME = "aporia"
+_DECODED_JSON: Final = TypeAdapter(object)
 
 
 class AporiaGuardrail(CustomGuardrail):
@@ -68,7 +71,7 @@ class AporiaGuardrail(CustomGuardrail):
     async def prepare_aporia_request(
         self, new_messages: List[dict], response_string: Optional[str] = None
     ) -> dict:
-        data: dict[str, Any] = {}
+        data: dict[str, object] = {}
         if new_messages is not None:
             data["messages"] = new_messages
         if response_string is not None:
@@ -122,8 +125,10 @@ class AporiaGuardrail(CustomGuardrail):
         verbose_proxy_logger.debug("Aporia AI response: %s", response.text)
         if response.status_code == 200:
             # check if the response was flagged
-            _json_response = response.json()
-            action: str = _json_response.get(
+            _json_response: Final = decoded_json_object(
+                _DECODED_JSON.validate_python(response.json())
+            )
+            action: Final = _json_response.get(
                 "action"
             )  # possible values are modify, passthrough, block, rephrase
             if action == "block":

enterprise/enterprise_hooks/blocked_user_list.py

@@ -7,9 +7,10 @@
 ## This accepts a list of user id's for whom calls will be rejected
 
 
-from typing import Literal, Optional
+from typing import Final, Literal, Optional
 
 from fastapi import HTTPException
+from pydantic import TypeAdapter
 
 import litellm
 from litellm._internal_context import with_service_target
@@ -19,6 +20,9 @@ from litellm.integrations.custom_logger import CustomLogger
 from litellm.proxy._types import LiteLLM_EndUserTable, UserAPIKeyAuth
 from litellm.proxy.common_utils.user_api_key_cache import AUTH_OBJECTS_TARGET
 from litellm.proxy.utils import PrismaClient
+from litellm.repositories.table_repositories import EndUserRepository
+
+_REQUEST_BODY_VALUE: Final = TypeAdapter(object)
 
 
 class ENTERPRISE_BlockedUserList(CustomLogger):
@@ -94,15 +98,14 @@ class ENTERPRISE_BlockedUserList(CustomLogger):
                 )
                 if end_user_cache_obj is None and self.prisma_client is not None:
                     # check db
-                    end_user_obj = (
-                        await self.prisma_client.db.litellm_endusertable.find_unique(
-                            where={"user_id": user}
-                        )
-                    )
+                    end_user_id: Final = _REQUEST_BODY_VALUE.validate_python(user)
+                    end_user_obj = await EndUserRepository(
+                        self.prisma_client
+                    ).table.find_unique(where={"user_id": end_user_id})
                     if end_user_obj is None:  # user not in db - assume not blocked
                         end_user_obj = LiteLLM_EndUserTable(user_id=user, blocked=False)
                     cache.set_cache(key=cache_key, value=end_user_obj, ttl=60)
-                    if end_user_obj is not None and end_user_obj.blocked is True:
+                    if end_user_obj.blocked is True:
                         raise HTTPException(
                             status_code=400,
                             detail={

enterprise/enterprise_hooks/google_text_moderation.py

@@ -6,15 +6,23 @@
 # +-----------------------------------------------+
 #  Thank you users! We ❤️ you! - Krrish & Ishaan
 
+from typing import Final
+
 from fastapi import HTTPException
+from pydantic import TypeAdapter
 
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.proxy._types import UserAPIKeyAuth
-from litellm.proxy.guardrails._content_utils import iter_message_text
+from litellm.proxy.guardrails._content_utils import (
+    is_non_conversational_request,
+    iter_message_text,
+)
 from litellm.types.utils import CallTypesLiteral
 
+_THRESHOLD_VALUE: Final = TypeAdapter(object)
+
 
 class ENTERPRISE_GoogleTextModeration(CustomLogger):
     user_api_key_cache = None
@@ -69,9 +77,11 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
                     f"{category}_confidence_threshold",
                     default_confidence_threshold,
                 )
-            set_confidence_value = getattr(
-                self,
-                f"{category}_confidence_threshold",
+            set_confidence_value = _THRESHOLD_VALUE.validate_python(
+                getattr(
+                    self,
+                    f"{category}_confidence_threshold",
+                )
             )
             verbose_proxy_logger.info(
                 f"Google Text Moderation: {category}_confidence_threshold: {set_confidence_value}"
@@ -87,7 +97,7 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
 
     async def async_moderation_hook(
         self,
-        data: dict,
+        data: dict[str, object],
         user_api_key_dict: UserAPIKeyAuth,
         call_type: CallTypesLiteral,
     ):
@@ -95,6 +105,9 @@ class ENTERPRISE_GoogleTextModeration(CustomLogger):
         - Calls Google's Text Moderation API
         - Rejects request if it fails safety check
         """
+        if is_non_conversational_request(call_type, data):
+            verbose_proxy_logger.debug("Google text moderation: skipping non-conversational call type %s", call_type)
+            return
         # Covers multimodal list content + Responses-API input.
         text = "".join(iter_message_text(data))
         if text:

enterprise/litellm_enterprise/enterprise_callbacks/callback_controls.py

@@ -33,9 +33,8 @@ class EnterpriseCallbackControls:
 
             try:
                 disabled_callbacks = EnterpriseCallbackControls.get_disabled_callbacks(litellm_params, standard_callback_dynamic_params)
-                verbose_logger.debug(f"Dynamically disabled callbacks from {X_LITELLM_DISABLE_CALLBACKS}: {disabled_callbacks}")
-                verbose_logger.debug(f"Checking if {callback} is disabled via headers. Disable callbacks from headers: {disabled_callbacks}")
                 if disabled_callbacks is not None:
+                    verbose_logger.debug(f"Checking if {callback} is in dynamically disabled callbacks: {disabled_callbacks}")
                     #########################################################
                     # premium user check
                     #########################################################

enterprise/litellm_enterprise/enterprise_callbacks/example_logging_api.py

@@ -1,15 +1,20 @@
 # this is an example endpoint to receive data from litellm
+from typing import Final
+
 from fastapi import FastAPI, HTTPException, Request
+from pydantic import TypeAdapter
 
 app = FastAPI()
 
+_LOGGED_EVENT: Final = TypeAdapter(object)
+
 
 @app.post("/log-event")
 async def log_event(request: Request):
     try:
         print("Received /log-event request")  # noqa
         # Assuming the incoming request has JSON data
-        data = await request.json()
+        data: Final = _LOGGED_EVENT.validate_python(await request.json())
         print("Received request data:")  # noqa
         print(data)  # noqa
 

enterprise/litellm_enterprise/enterprise_callbacks/llm_guard.py

@@ -8,25 +8,30 @@
 ## This provides an LLM Guard Integration for content moderation on the proxy
 
 import asyncio
-from typing import Final, Optional
+from collections.abc import Mapping
+from typing import Final
 
 import aiohttp
 from fastapi import HTTPException
+from pydantic import TypeAdapter
 
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.llms.openai.common_utils import decoded_json_object
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.utils import CallTypesLiteral
 
+_DECODED_JSON: Final = TypeAdapter(object)
+
 
 class _ENTERPRISE_LLMGuard(CustomLogger):
     # Class variables or attributes
     def __init__(
         self,
         mock_testing: bool = False,
-        mock_redacted_text: Optional[dict] = None,
+        mock_redacted_text: Mapping[str, object] | None = None,
     ):
         self.mock_redacted_text = mock_redacted_text
         self.llm_guard_mode = litellm.llm_guard_mode
@@ -66,7 +71,9 @@ class _ENTERPRISE_LLMGuard(CustomLogger):
                     async with session.post(
                         analyze_url, json={"prompt": text}
                     ) as response:
-                        redacted_text = await response.json()
+                        redacted_text = _DECODED_JSON.validate_python(
+                            await response.json()
+                        )
             verbose_proxy_logger.debug(
                 f"LLM Guard: Received response - {redacted_text}"
             )
@@ -77,12 +84,13 @@ class _ENTERPRISE_LLMGuard(CustomLogger):
                         "error": f"Invalid content moderation response: {redacted_text}"
                     },
                 )
-            if redacted_text.get("is_valid", None) is False:
+            redacted_response: Final = decoded_json_object(redacted_text)
+            if redacted_response.get("is_valid", None) is False:
                 raise HTTPException(
                     status_code=400,
                     detail={"error": "Violated content safety policy"},
                 )
-            sanitized_prompt = redacted_text.get("sanitized_prompt")
+            sanitized_prompt: Final = redacted_response.get("sanitized_prompt")
             return sanitized_prompt if isinstance(sanitized_prompt, str) else text
         except Exception as e:
             verbose_proxy_logger.exception(

enterprise/litellm_enterprise/enterprise_callbacks/secret_detection.py

@@ -13,21 +13,28 @@ sys.path.insert(
 )  # Adds the parent directory to the system path
 import configparser
 import contextlib
+import hashlib
 import itertools
+import json
 import re
 import tempfile
+from collections import OrderedDict
 from collections.abc import Generator, Iterator, Sequence
 from contextvars import ContextVar
+from functools import partial
 from typing import TYPE_CHECKING, ClassVar, Final, Literal, Optional
 
+from pydantic import JsonValue, TypeAdapter
+
 from litellm._logging import verbose_proxy_logger
 from litellm.caching.caching import DualCache
 from litellm.integrations.custom_guardrail import (
     CustomGuardrail,
     log_guardrail_information,
 )
+from litellm.llms.base_llm.guardrail_translation.utils import effective_skip_tool_message_for_guardrail
 from litellm.proxy._types import UserAPIKeyAuth
-from litellm.proxy.guardrails._content_utils import walk_user_text
+from litellm.proxy.guardrails._content_utils import iter_tool_json_fields, map_tool_json_text, walk_user_text
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.utils import GenericGuardrailAPIInputs
 
@@ -40,7 +47,7 @@ GUARDRAIL_PROVIDER = "hide-secrets"
 
 # Per-invocation tally of redacted secrets by detect-secrets plugin type; None
 # means the guardrail did not run, so _process_response records nothing.
-_masked_entity_count: ContextVar[Optional[dict]] = ContextVar(
+_masked_entity_count: ContextVar[Optional[dict[str, int]]] = ContextVar(
     "hide_secrets_masked_entity_count", default=None
 )
 
@@ -449,6 +456,19 @@ _default_detect_secrets_config = {
 
 _CONFIG_SECTION: Final = "litellm-prompt"
 
+_JSON_STRING: Final = TypeAdapter(str)
+
+_KEY_CONTEXT: Final = 64
+
+# Longest JSON line scanned for a multi-line tool string. Escaping puts the whole string on one line, and
+# detect-secrets regexes backtrack across it, so a longer multi-line string is cut into runs of whole lines,
+# each scanned as its own JSON line beside the key. Every run holds at least two lines, so the high-entropy
+# detectors, which match only a quoted string made wholly of their charset, never judge a lone line.
+_JSON_LINE_LIMIT: Final = 4096
+
+# A detector can stop a match inside an escape of a JSON-escaped line, as at the backslash of \".
+_CUT_ESCAPE: Final = re.compile(r"(?<=.)\\(?:u[0-9A-Fa-f]{0,3})?\Z", re.DOTALL)
+
 _ASSIGNMENT_LINE: Final = re.compile(r"[^\s\[#;:=][^:=]*[:=]")
 
 _SHELL_ASSIGNMENT: Final = re.compile(r"(?P<key>[^\s\[#;:=](?:[^:=]*[^\s:=])?)=(?P<value>\S+)")
@@ -459,6 +479,10 @@ _SHELL_TRAILER: Final = re.compile(r"\\|#.*|-*\w[\w.-]*=\S*")
 
 _SCAN_SUFFIX: Final = ".py"
 
+# Digests of scan inputs that held no secret, so the history an agent resends every turn is scanned once. 65536
+# digests take about 10 MB and hold a few hundred interleaved agent sessions per worker.
+_CLEAN_SCAN_CACHE_SIZE: Final = 65536
+
 
 @contextlib.contextmanager
 def _temp_file(text: str) -> Generator[str, None, None]:
@@ -532,6 +556,57 @@ def _quoted_assignments(text: str) -> tuple[str, ...]:
     )
 
 
+def _json_unescaped(value: str) -> str:
+    for candidate in (value, _CUT_ESCAPE.sub("", value)):
+        try:
+            return _JSON_STRING.validate_json(f'"{candidate}"')
+        except ValueError:
+            continue
+    return value
+
+
+def _key_context(key: str) -> str:
+    return json.dumps(" ".join(key.split()), ensure_ascii=False)[1:-1][-_KEY_CONTEXT:]
+
+
+def _json_line(context: str | None, text: str) -> str:
+    value: Final = json.dumps(text, ensure_ascii=False)
+    return value if context is None else f'"{context}": {value}'
+
+
+def _json_lines(context: str | None, text: str) -> Iterator[str]:
+    """Yield ``text`` as one JSON line, or a long multi-line ``text`` as JSON lines of two or more whole lines."""
+    line: Final = _json_line(context, text)
+    lines: Final = text.splitlines(keepends=True)
+    if len(line) <= _JSON_LINE_LIMIT or len(lines) < 2:
+        yield line
+        return
+    budget: Final = _JSON_LINE_LIMIT - len(_json_line(context, ""))
+    sizes: Final = (len(json.dumps(part, ensure_ascii=False)) - 2 for part in lines)
+    states: Final = itertools.accumulate(sizes, partial(_next_run_state, budget), initial=(0, 0, 0))
+    run_ids: Final = (state[0] for state in itertools.islice(states, 1, None))
+    runs: Final = tuple(
+        tuple(part for part, _ in run) for _, run in itertools.groupby(zip(lines, run_ids), key=lambda pair: pair[1])
+    )
+    whole_runs: Final = (*runs[:-2], runs[-2] + runs[-1]) if len(runs[-1]) < 2 else runs
+    for run in whole_runs:
+        yield _json_line(context, "".join(run[:-1]) + run[-1].splitlines()[0])
+
+
+def _next_run_state(budget: int, state: tuple[int, int, int], size: int) -> tuple[int, int, int]:
+    """Advance ``(run id, lines in run, escaped size of run)`` by one line of escaped ``size``."""
+    run_id, count, total = state
+    if count >= 2 and total + size > budget:
+        return run_id + 1, 1, size
+    return run_id, count + 1, total + size
+
+
+def _value_in_texts(value: str, texts: Sequence[str]) -> str | None:
+    """Return the detected value as it appears in ``texts``, decoding it when it came from a JSON-escaped key line."""
+    candidates: Final = (value, _json_unescaped(value))
+    return next((candidate for candidate in candidates if any(candidate in text for text in texts)), None)
+
+
 class _ENTERPRISE_SecretDetection(CustomGuardrail):
     # Keeps proxied traffic on async_pre_call_hook (the unified apply_guardrail
     # path skips should_run_check and never sees data["prompt"]).
@@ -539,11 +614,16 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail):
 
     def __init__(self, detect_secrets_config: Optional[dict] = None, **kwargs):
         self.user_defined_detect_secrets_config = detect_secrets_config
+        self._clean_scans: OrderedDict[bytes, None] = OrderedDict()
         super().__init__(**kwargs)
 
-    def scan_message_for_secrets(self, message_content: str):
+    def scan_message_for_secrets(self, message_content: str) -> list[dict[str, str]]:
         from detect_secrets.settings import transient_settings
 
+        digest: Final = hashlib.sha256(message_content.encode()).digest()
+        if digest in self._clean_scans:
+            self._clean_scans.move_to_end(digest)
+            return []
         detect_secrets_config = (
             self.user_defined_detect_secrets_config or _default_detect_secrets_config
         )
@@ -552,6 +632,10 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail):
                 (*message_content.splitlines(), *_quoted_assignments(message_content))
             )
 
+        if not found:
+            self._clean_scans[digest] = None
+            if len(self._clean_scans) > _CLEAN_SCAN_CACHE_SIZE:
+                self._clean_scans.popitem(last=False)
         return [
             {"type": secret_type, "value": value}
             for value, secret_type in sorted(
@@ -580,6 +664,35 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail):
         )
         return pattern.sub("[REDACTED]", text)
 
+    def redact_tool_payload(self, payload: JsonValue) -> JsonValue:
+        """Redact secrets in a tool payload's string values, scanning each value alone and as its JSON lines."""
+        fields: Final = tuple(iter_tool_json_fields(payload))
+        texts: Final = tuple(text for _, text in fields)
+        contexts: Final = {key: _key_context(key) for key in {key for key, _ in fields if key is not None}}
+        keyed: Final = tuple(
+            itertools.chain.from_iterable(
+                _json_lines(None if key is None else contexts[key], text) for key, text in fields
+            )
+        )
+        detected: Final = tuple(
+            {**secret, "value": value}
+            for secret in self.scan_message_for_secrets("\n".join((*texts, *keyed)))
+            if (value := _value_in_texts(secret["value"], texts)) is not None
+        )
+        if not detected:
+            return payload
+        counts: Final = _masked_entity_count.get()
+        if counts is not None:
+            for secret in detected:
+                counts[secret["type"]] = counts.get(secret["type"], 0) + 1
+        verbose_proxy_logger.warning(
+            "Detected and redacted secrets in tool payload: %s",
+            sorted(dict.fromkeys(secret["type"] for secret in detected)),
+        )
+        values: Final = sorted({secret["value"] for secret in detected}, key=len, reverse=True)
+        pattern: Final = re.compile("|".join(re.escape(value) for value in values))
+        return map_tool_json_text(payload, lambda text: pattern.sub("[REDACTED]", text))
+
     async def should_run_check(self, user_api_key_dict: UserAPIKeyAuth) -> bool:
         if user_api_key_dict.permissions is not None:
             if GUARDRAIL_NAME in user_api_key_dict.permissions:
@@ -641,7 +754,12 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail):
         _masked_entity_count.set({})
 
         # Covers multimodal list content + Responses-API input.
-        inspected = walk_user_text(data, self.redact_text) + self._redact_prompt(data)
+        inspected = walk_user_text(
+            data,
+            self.redact_text,
+            self.redact_tool_payload,
+            skip_tool_message=effective_skip_tool_message_for_guardrail(self),
+        ) + self._redact_prompt(data)
 
         if inspected == 0:
             # Image-only, empty-text, and unsupported payloads inspected

enterprise/litellm_enterprise/integrations/custom_guardrail.py

@@ -20,9 +20,15 @@ class EnterpriseCustomGuardrailHelper:
           the tag's value (the mode for that tag).
         - If no request tag matches, fall back to default mode(s).
         """
+        from litellm.litellm_core_utils.core_helpers import (
+            get_metadata_variable_name_from_kwargs,
+        )
         from litellm.litellm_core_utils.litellm_logging import (
             StandardLoggingPayloadSetup,
         )
+        from litellm.router_strategy.tag_based_routing import (
+            without_baggage_mapped_tags,
+        )
         from litellm.proxy._types import CommonProxyErrors
         from litellm.proxy.proxy_server import premium_user
 
@@ -36,9 +42,13 @@ class EnterpriseCustomGuardrailHelper:
 
         proxy_server_request = data.get("proxy_server_request", {})
 
-        request_tags = StandardLoggingPayloadSetup.get_request_tags(
-            litellm_params=data,
-            proxy_server_request=proxy_server_request,
+        request_tags = without_baggage_mapped_tags(
+            StandardLoggingPayloadSetup.get_request_tags(
+                litellm_params=data,
+                proxy_server_request=proxy_server_request,
+            ),
+            data.get(get_metadata_variable_name_from_kwargs(data)),
+            keep_serving_deployment_tags=True,
         )
 
         # Check if any request tag matches a Mode tag key

enterprise/litellm_enterprise/proxy/auth/route_checks.py

@@ -2,6 +2,8 @@ import os
 
 from fastapi import HTTPException, status
 
+from litellm.proxy._types import MODEL_LIST_ROUTES
+
 
 class EnterpriseRouteChecks:
     @staticmethod
@@ -43,7 +45,7 @@ class EnterpriseRouteChecks:
 
     # Routes that should remain accessible even when LLM API endpoints are disabled.
     # These are read-only model listing routes needed by the Admin UI.
-    LLM_API_EXEMPT_ROUTES = ["/models", "/v1/models"]
+    LLM_API_EXEMPT_ROUTES = MODEL_LIST_ROUTES
 
     @staticmethod
     def should_call_route(route: str):

enterprise/litellm_enterprise/proxy/auth/user_api_key_auth.py

@@ -1,14 +1,16 @@
-from typing import Any, Optional
+from collections.abc import Awaitable, Callable
 
 from fastapi import Request
 
 from litellm._logging import verbose_proxy_logger
-from litellm.proxy._types import ProxyException, UserAPIKeyAuth
+from litellm.proxy._types import ProxyException
 
 
 async def enterprise_custom_auth(
-    request: Request, api_key: str, user_custom_auth: Optional[Any]
-) -> Optional[UserAPIKeyAuth]:
+    request: Request,
+    api_key: str,
+    user_custom_auth: Callable[[Request, str], Awaitable[object]] | None,
+) -> object:
     from litellm_enterprise.proxy.proxy_server import custom_auth_settings
 
     if user_custom_auth is None:

enterprise/litellm_enterprise/proxy/common_utils/check_responses_cost.py

@@ -1,12 +1,43 @@
 """
 Polls LiteLLM_ManagedObjectTable to check if the response is complete.
-Cost tracking is handled by the get-responses call, which prices normally only because the
-poll stamps itself with BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN; user-facing reads of the
-same route are non-inference and free.
+A background response is billed exactly once: whichever client read or poll first sees it
+finished moves its row out of a pending status, then logs the body that read already fetched as
+one spend event stamped with BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN and the creating key's
+attribution, priced against the deployment that served it. The poll's status checks are stamped
+with BACKGROUND_RESPONSE_STATUS_POLL_CALL_ORIGIN and client reads are marked managed by the
+managed files hook, so both stay free. The spend event carries the id the client holds, the row's
+unified_object_id, so its spend log is found under that id with the _background_cost suffix.
+The spend event reaches the creating key's or team's own logging callbacks the way that creator's
+requests do. Its message content is redacted when the creator's key or team logging turns message
+logging off, else as the proxy is set, the way the batch poll decides it. The row is claimed before
+the creator is read, so reads that lose the race skip the lookups. A creator whose lookup fails, or
+whose logging settings the logger rejects, is still billed, through the proxy's own callbacks only
+and with its message content redacted.
+A read that finds no row, because the create could not store one, stores it under the reading key
+and bills it the same way, so a response the create left untracked is still billed once. A streamed
+background create is free on its own; its stream stores the row under the creating key and bills
+the finished response through it before the client sees response.completed.
+A read the caller retargets with client-side credentials, and a delete of a still-pending response,
+settle the bill from a fetch through the configured deployment instead: the client controls what a
+retargeted read returns, and a deleted response leaves the poll nothing to fetch.
+A row stored before this billing model keeps the encrypted create id as its model_object_id. The
+first lookup by that id rekeys it by the provider id, and the poll retires such a row without
+billing it once a read has already tracked the response under the provider id.
 """
 
+import uuid
+from collections.abc import Awaitable, Callable, Mapping
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
-from typing import TYPE_CHECKING, Dict, Final, Protocol, cast
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final
+
+from litellm_enterprise.proxy.common_utils.managed_object_attribution import (
+    ManagedObjectCreatorAttribution,
+    ManagedObjectRow,
+    ResponseCreator,
+)
+from pydantic import BaseModel, ConfigDict, TypeAdapter
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -14,36 +45,202 @@ from litellm.constants import (
     INTERNAL_CALL_ORIGIN_METADATA_KEY,
     MANAGED_OBJECT_STALENESS_CUTOFF_DAYS,
     MAX_OBJECTS_PER_POLL_CYCLE,
+    PENDING_RESPONSE_STATUSES,
     STALE_OBJECT_CLEANUP_BATCH_SIZE,
+    TERMINAL_RESPONSE_STATUSES,
 )
+from litellm.proxy.hooks.responses_id_security import decrypted_response_id
 from litellm.repositories.table_repositories import ManagedObjectRepository
 from litellm.responses.utils import ResponsesAPIRequestUtils
 from litellm.types.llms.openai import ResponsesAPIResponse
-from litellm.types.utils import BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
+from litellm.types.router import Deployment
+from litellm.types.utils import (
+    BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN,
+    BACKGROUND_RESPONSE_STATUS_POLL_CALL_ORIGIN,
+)
 
 if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+    from litellm.proxy.proxy_server import ProxyConfig
     from litellm.proxy.utils import PrismaClient, ProxyLogging
     from litellm.repositories.prisma_protocols import TableActions
     from litellm.router import Router
 
-TERMINAL_RESPONSE_STATUSES = frozenset({"completed", "failed", "cancelled", "incomplete"})
+_STORED_RESPONSE_ADAPTER: Final[TypeAdapter[dict[str, object]]] = TypeAdapter(dict[str, object])
+_CALLBACK_LIST_ADAPTER: Final[TypeAdapter[tuple[object, ...]]] = TypeAdapter(tuple[object, ...])
 
 
-class _ManagedObjectRow(Protocol):
-    @property
-    def id(self) -> str: ...
-
-    @property
-    def unified_object_id(self) -> str: ...
-
-    @property
-    def created_by(self) -> str | None: ...
-
-    @property
-    def file_object(self) -> object: ...
+class _RoutedModelInfo(BaseModel):
+    model_config = ConfigDict(frozen=True)
+    id: str | None = None
 
 
-def _managed_object_table(prisma_client: "PrismaClient") -> "TableActions[_ManagedObjectRow]":
+class _RoutedDeployment(BaseModel):
+    model_config = ConfigDict(frozen=True)
+    model_info: _RoutedModelInfo = _RoutedModelInfo()
+
+
+_ROUTED_DEPLOYMENTS: Final = TypeAdapter(tuple[_RoutedDeployment, ...])
+_RETRIEVE_RESPONSE_CALL_TYPE: Final = "aget_responses"
+_CALLBACK_LIST_FIELDS: Final = frozenset({"success_callback", "failure_callback", "litellm_disabled_callbacks"})
+_REJECTED_CREATOR_LOGGING: Final[Mapping[str, object]] = MappingProxyType({"turn_off_message_logging": True})
+
+
+@dataclass(frozen=True, slots=True)
+class _PollTarget:
+    response_id: str
+    model_metadata: Mapping[str, object]
+    deployment: Deployment | None
+
+
+@dataclass(frozen=True, slots=True)
+class BilledBackgroundResponse:
+    response: ResponsesAPIResponse
+    model: str
+    custom_llm_provider: str | None
+    litellm_params: Mapping[str, object]
+    metadata: Mapping[str, object]
+    creator_logging: Mapping[str, object]
+
+
+@dataclass(frozen=True, slots=True)
+class BilledCharge:
+    cost: float | None
+    logging_obj: "LiteLLMLogging"
+
+
+BilledResponseLogger = Callable[[BilledBackgroundResponse], Awaitable[BilledCharge]]
+UntrackedResponseTracker = Callable[[ResponsesAPIResponse], Awaitable[bool]]
+
+
+def _without_read_price(response: ResponsesAPIResponse) -> ResponsesAPIResponse:
+    unpriced: Final = response.model_copy(deep=True)
+    unpriced.hidden_params = {key: value for key, value in response.hidden_params.items() if key != "response_cost"}
+    return unpriced
+
+
+def _string_callbacks(callbacks: object) -> tuple[str, ...]:
+    if not isinstance(callbacks, list):
+        return ()
+    return tuple(
+        callback for callback in _CALLBACK_LIST_ADAPTER.validate_python(callbacks) if isinstance(callback, str)
+    )
+
+
+def _running_proxy_config() -> "ProxyConfig":
+    from litellm.proxy.proxy_server import proxy_config
+
+    return proxy_config
+
+
+def creator_logging_fields(creator: ResponseCreator, proxy_config: "ProxyConfig | None") -> Mapping[str, object]:
+    from litellm.proxy._types import UserAPIKeyAuth
+    from litellm.proxy.litellm_pre_call_utils import key_and_team_logging_fields
+
+    if not creator.settings_read:
+        return _REJECTED_CREATOR_LOGGING
+    try:
+        return key_and_team_logging_fields(
+            user_api_key_dict=UserAPIKeyAuth(
+                metadata=dict(creator.key_metadata),
+                team_id=creator.team_id,
+                team_metadata=dict(creator.team_metadata),
+            ),
+            proxy_config=proxy_config if proxy_config is not None else _running_proxy_config(),
+        )
+    except Exception as e:
+        verbose_proxy_logger.error(
+            f"CheckResponsesCost: could not read the creator's logging settings, so only the proxy's "
+            f"callbacks get the billed event: {e}"
+        )
+        return _REJECTED_CREATOR_LOGGING
+
+
+def _billed_logging_obj(billed: BilledBackgroundResponse, creator_logging: Mapping[str, object]) -> "LiteLLMLogging":
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
+    from litellm.utils import split_dynamic_success_callbacks
+
+    sync_callbacks, async_callbacks = split_dynamic_success_callbacks(
+        _string_callbacks(creator_logging.get("success_callback"))
+    )
+    return LiteLLMLogging(
+        model=billed.model,
+        messages=[{"role": "user", "content": "<retrieve_response>"}],
+        stream=False,
+        call_type=_RETRIEVE_RESPONSE_CALL_TYPE,
+        start_time=datetime.now(),
+        litellm_call_id=str(uuid.uuid4()),
+        function_id=str(uuid.uuid4()),
+        dynamic_success_callbacks=list(sync_callbacks),
+        dynamic_async_success_callbacks=list(async_callbacks) or None,
+        kwargs={key: value for key, value in creator_logging.items() if key not in _CALLBACK_LIST_FIELDS},
+    )
+
+
+def _creator_or_proxy_logging_obj(billed: BilledBackgroundResponse) -> "LiteLLMLogging":
+    try:
+        return _billed_logging_obj(billed, billed.creator_logging)
+    except Exception as e:
+        verbose_proxy_logger.error(
+            f"CheckResponsesCost: the creator's logging settings were rejected, so only the proxy's "
+            f"callbacks get the billed event: {e}"
+        )
+        return _billed_logging_obj(billed, _REJECTED_CREATOR_LOGGING)
+
+
+async def log_billed_background_response(billed: BilledBackgroundResponse) -> BilledCharge:
+    from litellm.litellm_core_utils.logging_worker import GLOBAL_LOGGING_WORKER
+
+    logging_obj: Final = _creator_or_proxy_logging_obj(billed)
+    logging_obj.update_environment_variables(
+        litellm_params={**billed.litellm_params, "metadata": dict(billed.metadata)},
+        optional_params={},
+        custom_llm_provider=billed.custom_llm_provider,
+    )
+    unpriced: Final = _without_read_price(billed.response)
+    cost: Final = logging_obj.response_cost_calculator(result=unpriced)
+    end_time: Final = datetime.now()
+    GLOBAL_LOGGING_WORKER.ensure_initialized_and_enqueue(
+        async_coroutine=logging_obj.async_success_handler(
+            result=unpriced, start_time=logging_obj.start_time, end_time=end_time
+        )
+    )
+    logging_obj.handle_sync_success_callbacks_for_async_calls(
+        result=unpriced, start_time=logging_obj.start_time, end_time=end_time
+    )
+    return BilledCharge(cost=cost, logging_obj=logging_obj)
+
+
+class _DeploymentBaseModel(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    base_model: str | None = None
+
+
+def _deployment_litellm_params(deployment: Deployment) -> Mapping[str, object]:
+    from litellm.litellm_core_utils.litellm_logging import mask_api_base_credentials
+
+    api_base: Final = deployment.litellm_params.api_base
+    base_model: Final = _DeploymentBaseModel.model_validate(deployment.litellm_params, from_attributes=True).base_model
+    return MappingProxyType(
+        {
+            **({"api_base": mask_api_base_credentials(api_base)} if api_base else {}),
+            **({"base_model": base_model} if base_model else {}),
+        }
+    )
+
+
+def _deployment_metadata(deployment: Deployment) -> Mapping[str, object]:
+    return MappingProxyType(
+        {
+            "model_info": deployment.model_info.model_dump(exclude_none=True),
+            "model_group": deployment.model_name,
+            "deployment": deployment.litellm_params.model,
+        }
+    )
+
+
+def _managed_object_table(prisma_client: "PrismaClient") -> "TableActions[ManagedObjectRow]":
     return ManagedObjectRepository(prisma_client).table
 
 
@@ -51,12 +248,18 @@ def _is_response_gone_at_provider(error: Exception, provider_response_id: str) -
     return getattr(error, "status_code", None) == 404 and provider_response_id in str(error)
 
 
+def _keyed_by_issued_id(job: ManagedObjectRow) -> bool:
+    return decrypted_response_id(job.model_object_id) != job.model_object_id
+
+
 class CheckResponsesCost:
     def __init__(
         self,
         proxy_logging_obj: "ProxyLogging",
         prisma_client: "PrismaClient",
         llm_router: "Router",
+        log_billed_response: BilledResponseLogger = log_billed_background_response,
+        proxy_config: "ProxyConfig | None" = None,
     ):
         from litellm.proxy.utils import PrismaClient, ProxyLogging
         from litellm.router import Router
@@ -64,15 +267,18 @@ class CheckResponsesCost:
         self.proxy_logging_obj: ProxyLogging = proxy_logging_obj
         self.prisma_client: PrismaClient = prisma_client
         self.llm_router: Router = llm_router
+        self._creator_attribution: Final = ManagedObjectCreatorAttribution(prisma_client, "CheckResponsesCost")
+        self._log_billed_response: Final = log_billed_response
+        self._proxy_config: Final = proxy_config
 
-    def _resolve_deployment(self, response_id: str) -> bool:
+    def _resolve_deployment(self, response_id: str) -> Deployment | None:
         model_id: str | None = ResponsesAPIRequestUtils.get_model_id_from_response_id(response_id)
-        return model_id is not None and self.llm_router.get_deployment(model_id=model_id) is not None
+        return self.llm_router.registry().deployment(model_id) if model_id is not None else None
 
     async def _get_response(
         self,
         response_id: str,
-        litellm_metadata: Dict[str, str],
+        litellm_metadata: dict[str, object],
         via_router: bool,
     ) -> ResponsesAPIResponse:
         """Fetch the upstream response, using deployment credentials when available.
@@ -84,16 +290,26 @@ class CheckResponsesCost:
         sees provider env vars, so it fails for every deployment whose credentials
         live in the config; the row then never leaves ``queued``.
         """
-        if not via_router:
-            return await litellm.aget_responses(response_id=response_id, litellm_metadata=litellm_metadata)
-        router_response = await self.llm_router.aget_responses(
-            response_id=response_id, litellm_metadata=litellm_metadata
+        retrieved: Final[object] = (
+            await self.llm_router.aget_responses(response_id=response_id, litellm_metadata=litellm_metadata)
+            if via_router
+            else await litellm.aget_responses(response_id=response_id, litellm_metadata=litellm_metadata)
         )
-        return cast(ResponsesAPIResponse, router_response)
+        if not isinstance(retrieved, ResponsesAPIResponse):
+            raise TypeError(f"Retrieving {response_id} did not return a response object")
+        return retrieved
 
-    async def _expire_stale_rows(
-        self, cutoff: datetime, batch_size: int
-    ) -> int:
+    async def _fetch_unbilled(self, target: _PollTarget) -> ResponsesAPIResponse:
+        return await self._get_response(
+            response_id=target.response_id,
+            litellm_metadata={
+                **target.model_metadata,
+                INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_STATUS_POLL_CALL_ORIGIN,
+            },
+            via_router=target.deployment is not None,
+        )
+
+    async def _expire_stale_rows(self, cutoff: datetime, batch_size: int) -> int:
         """Execute the bounded UPDATE that marks stale rows as 'stale_expired'.
 
         Isolated so it can be swapped / mocked in tests without touching the
@@ -140,14 +356,170 @@ class CheckResponsesCost:
                 f"(older than {MANAGED_OBJECT_STALENESS_CUTOFF_DAYS} days) as stale_expired"
             )
 
+    def _target(self, response_id: str, model_name: str | None) -> _PollTarget:
+        return _PollTarget(
+            response_id=response_id,
+            model_metadata=MappingProxyType({"model": model_name, "model_group": model_name} if model_name else {}),
+            deployment=self._resolve_deployment(response_id),
+        )
+
+    def _poll_target(self, job: ManagedObjectRow) -> _PollTarget:
+        stored_model: Final = _STORED_RESPONSE_ADAPTER.validate_python(job.file_object).get("model")
+        return self._target(
+            decrypted_response_id(job.unified_object_id), stored_model if isinstance(stored_model, str) else None
+        )
+
+    async def _claim_for_billing(self, job: ManagedObjectRow) -> bool:
+        claimed: Final = await _managed_object_table(self.prisma_client).update_many(
+            # bounded-ok: PENDING_RESPONSE_STATUSES is a two-value constant tuple
+            where={"id": job.id, "status": {"in": list(PENDING_RESPONSE_STATUSES)}},
+            data={"status": "completed"},
+        )
+        return claimed > 0
+
+    async def _release_billing_claim(self, job: ManagedObjectRow) -> None:
+        await _managed_object_table(self.prisma_client).update_many(
+            where={"id": job.id, "status": "completed"},
+            data={"status": job.status or PENDING_RESPONSE_STATUSES[0]},
+        )
+
+    def _billed_response(
+        self,
+        job: ManagedObjectRow,
+        target: _PollTarget,
+        response: ResponsesAPIResponse,
+        creator: ResponseCreator,
+    ) -> BilledBackgroundResponse:
+        deployment: Final = target.deployment
+        return BilledBackgroundResponse(
+            response=response.model_copy(update={"id": job.unified_object_id}),
+            model=deployment.litellm_params.model if deployment is not None else (response.model or ""),
+            custom_llm_provider=ResponsesAPIRequestUtils.decode_responses_api_response_id(target.response_id).get(
+                "custom_llm_provider"
+            ),
+            litellm_params=_deployment_litellm_params(deployment) if deployment is not None else MappingProxyType({}),
+            metadata=MappingProxyType(
+                {
+                    **creator.metadata,
+                    INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN,
+                    **target.model_metadata,
+                    **(_deployment_metadata(deployment) if deployment is not None else {}),
+                    **self._requested_model_group(target, job.team_id),
+                }
+            ),
+            creator_logging=creator_logging_fields(creator, self._proxy_config),
+        )
+
+    def _requested_model_group(self, target: _PollTarget, team_id: str | None) -> Mapping[str, object]:
+        requested: Final = target.model_metadata.get("model_group")
+        if target.deployment is None or not isinstance(requested, str):
+            return MappingProxyType({})
+        routed: Final = _ROUTED_DEPLOYMENTS.validate_python(
+            self.llm_router.registry().deployments(requested, team_id=team_id) or ()
+        )
+        serving_id: Final = target.deployment.model_info.id
+        return MappingProxyType(
+            {"model_group": requested} if any(row.model_info.id == serving_id for row in routed) else {}
+        )
+
+    async def _bill_once(
+        self,
+        job: ManagedObjectRow,
+        target: _PollTarget,
+        response: ResponsesAPIResponse,
+    ) -> BilledCharge | None:
+        if not await self._claim_for_billing(job):
+            verbose_proxy_logger.info(f"Response {job.unified_object_id} was already billed elsewhere, skipping")
+            return None
+        try:
+            creator: Final = await self._creator_attribution.response_creator(job, job.unified_object_id)
+            charge: Final = await self._log_billed_response(self._billed_response(job, target, response, creator))
+        except BaseException:
+            await self._release_billing_claim(job)
+            raise
+        verbose_proxy_logger.info(f"Response {job.unified_object_id} reached a terminal status and was billed")
+        return charge
+
+    async def _job_for_provider_id(self, provider_response_id: str) -> ManagedObjectRow | None:
+        return await _managed_object_table(self.prisma_client).find_first(
+            where={"model_object_id": provider_response_id, "file_purpose": "response"}
+        )
+
+    async def _rekey_by_provider_id(self, job: ManagedObjectRow) -> None:
+        await _managed_object_table(self.prisma_client).update_many(
+            where={"id": job.id}, data={"model_object_id": decrypted_response_id(job.model_object_id)}
+        )
+
+    async def _tracked_job(self, response_id: str) -> ManagedObjectRow | None:
+        if tracked := await self._job_for_provider_id(decrypted_response_id(response_id)):
+            return tracked
+        stored_by_issued_id: Final = await _managed_object_table(self.prisma_client).find_first(
+            where={"unified_object_id": response_id, "file_purpose": "response"}
+        )
+        if stored_by_issued_id is not None and _keyed_by_issued_id(stored_by_issued_id):
+            await self._rekey_by_provider_id(stored_by_issued_id)
+        return stored_by_issued_id
+
+    async def _superseded_by_a_provider_id_row(self, job: ManagedObjectRow) -> bool:
+        if not _keyed_by_issued_id(job):
+            return False
+        if await self._job_for_provider_id(decrypted_response_id(job.model_object_id)) is None:
+            await self._rekey_by_provider_id(job)
+            return False
+        await self._claim_for_billing(job)
+        return True
+
+    async def _track_from_deployment(
+        self, response_id: str, track_untracked: UntrackedResponseTracker | None
+    ) -> BilledCharge | None:
+        target: Final = self._target(decrypted_response_id(response_id), None)
+        if track_untracked is None or target.deployment is None:
+            return None
+        return await self.bill_if_finished(response_id, await self._fetch_unbilled(target), track_untracked)
+
+    async def _start_tracking(
+        self, response_id: str, response: ResponsesAPIResponse, track_untracked: UntrackedResponseTracker | None
+    ) -> ManagedObjectRow | None:
+        if track_untracked is None or not await track_untracked(response):
+            return None
+        return await self._tracked_job(response_id)
+
+    async def bill_if_finished(
+        self,
+        response_id: str,
+        response: ResponsesAPIResponse,
+        track_untracked: UntrackedResponseTracker | None = None,
+    ) -> BilledCharge | None:
+        if response.status not in TERMINAL_RESPONSE_STATUSES:
+            return None
+        job: Final = await self._tracked_job(response_id) or await self._start_tracking(
+            response_id, response, track_untracked
+        )
+        if job is None or job.status not in PENDING_RESPONSE_STATUSES:
+            return None
+        return await self._bill_once(job, self._poll_target(job), response)
+
+    async def bill_from_deployment(
+        self, response_id: str, track_untracked: UntrackedResponseTracker | None = None
+    ) -> BilledCharge | None:
+        job: Final = await self._tracked_job(response_id)
+        if job is None:
+            return await self._track_from_deployment(response_id, track_untracked)
+        if job.status not in PENDING_RESPONSE_STATUSES:
+            return None
+        target: Final = self._poll_target(job)
+        response: Final = await self._fetch_unbilled(target)
+        if response.status not in TERMINAL_RESPONSE_STATUSES:
+            return None
+        return await self._bill_once(job, target, response)
+
     async def check_responses_cost(self):
         """
         Check if background responses are complete and track their cost.
         - Get all status="queued" or "in_progress" and file_purpose="response" jobs
-        - Query the provider to check if response is complete
-        - Cost is tracked by the get-responses call, billed because the poll is stamped
-          with BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN
-        - Mark responses in a terminal state as complete in the database
+        - Query the provider, unbilled, to check if the response is complete
+        - Claim each terminal job by moving its row to completed, then log the body that check fetched
+          as its one billed event, so two pollers racing over the same job bill it once
         - Mark responses the provider no longer has (404 through a resolved
           deployment) as stale_expired
         """
@@ -160,99 +532,53 @@ class CheckResponsesCost:
 
         jobs = await _managed_object_table(self.prisma_client).find_many(
             where={
-                "status": {"in": ["queued", "in_progress"]},
+                # bounded-ok: PENDING_RESPONSE_STATUSES is a two-value constant tuple
+                "status": {"in": list(PENDING_RESPONSE_STATUSES)},
                 "file_purpose": "response",
             },
             take=MAX_OBJECTS_PER_POLL_CYCLE,
             order={"created_at": "asc"},
         )
-        
+
         verbose_proxy_logger.debug(f"Found {len(jobs)} response jobs to check")
-        completed_jobs: Final[list[_ManagedObjectRow]] = []
-        expired_jobs: Final[list[_ManagedObjectRow]] = []
+        expired_jobs: Final[list[ManagedObjectRow]] = []
 
         for job in jobs:
             unified_object_id = job.unified_object_id
 
             try:
-                from litellm.proxy.hooks.responses_id_security import (
-                    ResponsesIDSecurity,
-                )
-
-                # Get the stored response object to extract model information
-                stored_response = job.file_object
-                model_name = stored_response.get("model", None)
-
-                # Decrypt the response ID
-                responses_id_security, _, _ = ResponsesIDSecurity()._decrypt_response_id(unified_object_id)
-
-                # Prepare metadata with model information for cost tracking
-                litellm_metadata = {
-                    "user_api_key_user_id": job.created_by or "default-user-id",
-                    INTERNAL_CALL_ORIGIN_METADATA_KEY: BACKGROUND_RESPONSE_COST_POLL_CALL_ORIGIN,
-                }
-
-                # Add model information if available
-                if model_name:
-                    litellm_metadata["model"] = model_name
-                    litellm_metadata["model_group"] = model_name  # Use same value for model_group
-                via_router = self._resolve_deployment(responses_id_security)
+                if await self._superseded_by_a_provider_id_row(job):
+                    verbose_proxy_logger.info(f"Response {unified_object_id} is billed under its provider id row")
+                    continue
+                target = self._poll_target(job)
             except Exception as e:
-                verbose_proxy_logger.warning(
-                    f"Skipping job {unified_object_id} due to error: {e}"
-                )
+                verbose_proxy_logger.warning(f"Skipping job {unified_object_id} due to error: {e}")
                 continue
 
-            provider_response_id = ResponsesAPIRequestUtils.decode_responses_api_response_id(
-                responses_id_security
-            ).get("response_id", responses_id_security)
+            provider_response_id = ResponsesAPIRequestUtils.decode_responses_api_response_id(target.response_id).get(
+                "response_id", target.response_id
+            )
             try:
-                response = await self._get_response(
-                    response_id=responses_id_security,
-                    litellm_metadata=litellm_metadata,
-                    via_router=via_router,
-                )
+                response = await self._fetch_unbilled(target)
 
-                verbose_proxy_logger.debug(
-                    f"Response {unified_object_id} status: {response.status}, model: {model_name}"
-                )
+                verbose_proxy_logger.debug(f"Response {unified_object_id} status: {response.status}")
+                if response.status in TERMINAL_RESPONSE_STATUSES:
+                    await self._bill_once(job, target, response)
 
             except Exception as e:
-                if via_router and _is_response_gone_at_provider(e, provider_response_id):
+                if target.deployment is not None and _is_response_gone_at_provider(e, provider_response_id):
                     verbose_proxy_logger.info(
                         f"Response {unified_object_id} no longer available at provider (404), marking stale_expired: {e}"
                     )
                     expired_jobs.append(job)
                     continue
-                verbose_proxy_logger.warning(
-                    f"Skipping job {unified_object_id} due to error: {e}"
-                )
+                verbose_proxy_logger.warning(f"Skipping job {unified_object_id} due to error: {e}")
                 continue
 
-            if response.status in TERMINAL_RESPONSE_STATUSES:
-                verbose_proxy_logger.info(
-                    f"Response {unified_object_id} has terminal status {response.status}, marking as complete"
-                )
-                completed_jobs.append(job)
-
-        # Mark completed jobs in the database
-        if len(completed_jobs) > 0:
-            await _managed_object_table(self.prisma_client).update_many(
-                # bounded-ok: at most MAX_OBJECTS_PER_POLL_CYCLE rows per cycle, the find_many take above
-                where={"id": {"in": [job.id for job in completed_jobs]}},
-                data={"status": "completed"},
-            )
-            verbose_proxy_logger.info(
-                f"Marked {len(completed_jobs)} response jobs as completed"
-            )
-
         if len(expired_jobs) > 0:
             await _managed_object_table(self.prisma_client).update_many(
                 # bounded-ok: at most MAX_OBJECTS_PER_POLL_CYCLE rows per cycle, the find_many take above
                 where={"id": {"in": [job.id for job in expired_jobs]}},
                 data={"status": "stale_expired"},
             )
-            verbose_proxy_logger.info(
-                f"Marked {len(expired_jobs)} response jobs as stale_expired"
-            )
-
+            verbose_proxy_logger.info(f"Marked {len(expired_jobs)} response jobs as stale_expired")

enterprise/litellm_enterprise/proxy/common_utils/managed_object_attribution.py

@@ -0,0 +1,381 @@
+"""Creator attribution for the spend row a managed-object cost poll writes.
+
+A batch or a background response is billed by a poll job that runs long after the creating
+request returned, so the managed object row's creator columns are the only record of which
+key, team, organization, user, end user, and tags carry the spend.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass
+from datetime import datetime
+from typing import TYPE_CHECKING, Final, Protocol, TypeGuard, cast
+
+from pydantic import TypeAdapter
+
+from litellm._logging import verbose_proxy_logger
+from litellm.constants import CLI_SESSION_KEY_PREFIX
+from litellm.repositories.table_repositories import (
+    DeletedVerificationTokenRepository,
+    DeprecatedVerificationTokenRepository,
+)
+from litellm.repositories.team_repository import TeamRepository
+from litellm.repositories.user_repository import UserRepository
+from litellm.repositories.verification_token_repository import VerificationTokenRepository
+
+if TYPE_CHECKING:
+    from prisma import models as prisma_models
+
+    from litellm.proxy.utils import PrismaClient
+    from litellm.repositories.prisma_protocols import TableActions
+
+
+class ManagedObjectRow(Protocol):
+    @property
+    def created_at(self) -> datetime: ...
+
+    @property
+    def updated_at(self) -> datetime: ...
+
+    @property
+    def id(self) -> str: ...
+
+    @property
+    def unified_object_id(self) -> str: ...
+
+    @property
+    def model_object_id(self) -> str: ...
+
+    @property
+    def created_by(self) -> str | None: ...
+
+    @property
+    def status(self) -> str | None: ...
+
+    @property
+    def file_object(self) -> object: ...
+
+    @property
+    def org_id(self) -> str | None: ...
+
+    @property
+    def api_key(self) -> str | None: ...
+
+    @property
+    def team_id(self) -> str | None: ...
+
+    @property
+    def request_tags(self) -> object: ...
+
+    @property
+    def end_user_id(self) -> str | None: ...
+
+
+_TAG_LIST_ADAPTER: Final[TypeAdapter[list[object]]] = TypeAdapter(list[object])
+_KEY_OWNER_SCOPE_PREFIX: Final = "key:"
+_MAX_ROTATION_HOPS: Final = 16
+
+
+def _string_tags(request_tags: object) -> list[str] | None:
+    if not isinstance(request_tags, list):
+        return None
+    return [tag for tag in _TAG_LIST_ADAPTER.validate_python(request_tags) if isinstance(tag, str)] or None
+
+
+def _creator_user_id(created_by: str | None) -> str | None:
+    if created_by is None or created_by.startswith(_KEY_OWNER_SCOPE_PREFIX):
+        return None
+    return created_by
+
+
+def _user_table(prisma_client: PrismaClient) -> TableActions[prisma_models.LiteLLM_UserTable]:
+    return UserRepository(prisma_client).table
+
+
+def _token_table(prisma_client: PrismaClient) -> TableActions[prisma_models.LiteLLM_VerificationToken]:
+    return VerificationTokenRepository(prisma_client).table
+
+
+def _team_table(prisma_client: PrismaClient) -> TableActions[prisma_models.LiteLLM_TeamTable]:
+    return TeamRepository(prisma_client).table
+
+
+def _rotated_token_table(
+    prisma_client: PrismaClient,
+) -> TableActions[prisma_models.LiteLLM_DeprecatedVerificationToken]:
+    return DeprecatedVerificationTokenRepository(prisma_client).table
+
+
+def _deleted_token_table(prisma_client: PrismaClient) -> TableActions[prisma_models.LiteLLM_DeletedVerificationToken]:
+    return DeletedVerificationTokenRepository(prisma_client).table
+
+
+def _deleted_team_table(prisma_client: PrismaClient) -> TableActions[prisma_models.LiteLLM_DeletedTeamTable]:
+    return TeamRepository(prisma_client).deleted_table
+
+
+@dataclass(frozen=True, slots=True)
+class _CreatorRows:
+    key: prisma_models.LiteLLM_VerificationToken | None
+    team: prisma_models.LiteLLM_TeamTable | None
+    user: prisma_models.LiteLLM_UserTable | None
+
+
+@dataclass(frozen=True, slots=True)
+class ResponseCreator:
+    metadata: Mapping[str, object]
+    key_metadata: Mapping[str, object]
+    team_id: str | None
+    team_metadata: Mapping[str, object]
+    settings_read: bool = True
+
+
+class ManagedObjectCreatorAttribution:
+    """Rebuilds the spend-tracking metadata of whoever created a managed object, so the poll
+    job's spend row is attributed like the creating request: key hash, team, organization,
+    user, aliases, and tags.
+
+    Rows created before api_key and request_tags were persisted carry only created_by and
+    team_id, and fall back to those. A named creating key owns user_api_key_alias; when it
+    has no alias, or the key has since been rotated or deleted, the field keeps the creating
+    user's alias, because a resolvable name is more useful on the spend row than a null.
+    user_api_key_org_id is resolved too: the spend update writer reads it off this metadata
+    to increment organization spend, so leaving it out drops the cost from org accounting.
+    A background response also carries its end user, the creating key's project and its alias,
+    and the key, team, and user budgets and per-model budgets, the way the read that billed it
+    before did, and never names a key-scoped creator as its user. A key that sets no budget of
+    its own takes its budget tier's, the same rule auth applies. Its creator's key and team
+    metadata come along too, so the billed event reaches the key's and team's logging callbacks
+    like that read did. A rotated key's metadata is read from the live key its rotations lead
+    to, and a deleted key's or team's from its deleted row. A lookup that fails leaves the
+    creator's settings unread and the attribution to the rows that still load, so the
+    response still bills and still counts toward their budgets.
+    """
+
+    def __init__(self, prisma_client: PrismaClient, poller_name: str) -> None:
+        self._prisma_client: Final = prisma_client
+        self._poller_name: Final = poller_name
+
+    async def batch_metadata(self, job: ManagedObjectRow, object_id: str) -> Mapping[str, object]:
+        rows: Final = await self._rows(job, object_id)
+        return {"user_api_key_user_id": job.created_by, **_shared_metadata(job, rows)}
+
+    async def response_creator(self, job: ManagedObjectRow, object_id: str) -> ResponseCreator:
+        try:
+            return await self._read_response_creator(job, object_id)
+        except Exception as e:
+            verbose_proxy_logger.error(
+                f"{self._poller_name}: could not read the creator of {object_id}, so it bills without their "
+                f"settings: {e}"
+            )
+            return ResponseCreator(
+                metadata=_response_creator_metadata(job, await self._rows(job, object_id)),
+                key_metadata={},
+                team_id=getattr(job, "team_id", None),
+                team_metadata={},
+                settings_read=False,
+            )
+
+    async def _read_response_creator(self, job: ManagedObjectRow, object_id: str) -> ResponseCreator:
+        api_key: Final = getattr(job, "api_key", None)
+        team_id: Final = getattr(job, "team_id", None)
+        rows: Final = _CreatorRows(
+            key=await self._live_key_row(api_key, job.created_by),
+            team=await self._live_team_row(team_id),
+            user=await self._user_row(object_id, _creator_user_id(job.created_by)),
+        )
+        return ResponseCreator(
+            metadata=_response_creator_metadata(job, rows),
+            key_metadata=row_metadata(rows.key) if rows.key is not None else await self._retired_key_metadata(api_key),
+            team_id=team_id,
+            team_metadata=(
+                row_metadata(rows.team) if rows.team is not None else await self._deleted_team_metadata(team_id)
+            ),
+        )
+
+    async def _retired_key_metadata(self, api_key: str | None) -> Mapping[str, object]:
+        if not _names_a_key_row(api_key):
+            return {}
+        current_token: Final = await self._token_rotated_to(api_key, _MAX_ROTATION_HOPS)
+        current_key: Final = (
+            await _token_table(self._prisma_client).find_unique(where={"token": current_token})
+            if current_token != api_key
+            else None
+        )
+        if current_key is not None:
+            return row_metadata(current_key)
+        return row_metadata(
+            await _deleted_token_table(self._prisma_client).find_first(
+                where={"token": current_token}, order={"deleted_at": "desc"}
+            )
+        )
+
+    async def _token_rotated_to(self, token: str, hops_left: int) -> str:
+        rotation: Final = await _rotated_token_table(self._prisma_client).find_unique(where={"token": token})
+        if rotation is None or hops_left == 0:
+            return token
+        return await self._token_rotated_to(rotation.active_token_id, hops_left - 1)
+
+    async def _deleted_team_metadata(self, team_id: str | None) -> Mapping[str, object]:
+        if not team_id:
+            return {}
+        return row_metadata(
+            await _deleted_team_table(self._prisma_client).find_first(
+                where={"team_id": team_id}, order={"deleted_at": "desc"}
+            )
+        )
+
+    async def _rows(self, job: ManagedObjectRow, object_id: str) -> _CreatorRows:
+        return _CreatorRows(
+            key=await self._key_row(object_id, getattr(job, "api_key", None), job.created_by),
+            team=await self._team_row(object_id, getattr(job, "team_id", None)),
+            user=await self._user_row(object_id, _creator_user_id(job.created_by)),
+        )
+
+    async def _user_row(self, object_id: str, user_id: str | None) -> prisma_models.LiteLLM_UserTable | None:
+        """Objects created by a team or service account key carry no user id, and
+        find_unique(where={"user_id": None}) raises, so those resolve no row."""
+        if not user_id:
+            return None
+        try:
+            return await _user_table(self._prisma_client).find_unique(where={"user_id": user_id})
+        except Exception as e:
+            verbose_proxy_logger.error(f"{self._poller_name}: could not look up user {user_id} for {object_id}: {e}")
+            return None
+
+    async def _key_row(
+        self, object_id: str, api_key: str | None, created_by: str | None
+    ) -> prisma_models.LiteLLM_VerificationToken | None:
+        try:
+            return await self._live_key_row(api_key, created_by)
+        except Exception as e:
+            verbose_proxy_logger.error(f"{self._poller_name}: could not look up the creating key for {object_id}: {e}")
+            return None
+
+    async def _live_key_row(
+        self, api_key: str | None, created_by: str | None
+    ) -> prisma_models.LiteLLM_VerificationToken | None:
+        if not api_key or _is_cli_session_key(api_key, created_by):
+            return None
+        return await _token_table(self._prisma_client).find_unique(
+            where={"token": api_key}, include={"litellm_budget_table": True, "litellm_project_table": True}
+        )
+
+    async def _team_row(self, object_id: str, team_id: str | None) -> prisma_models.LiteLLM_TeamTable | None:
+        try:
+            return await self._live_team_row(team_id)
+        except Exception as e:
+            verbose_proxy_logger.error(f"{self._poller_name}: could not look up team {team_id} for {object_id}: {e}")
+            return None
+
+    async def _live_team_row(self, team_id: str | None) -> prisma_models.LiteLLM_TeamTable | None:
+        if not team_id:
+            return None
+        return await _team_table(self._prisma_client).find_unique(where={"team_id": team_id})
+
+
+def _response_creator_metadata(job: ManagedObjectRow, rows: _CreatorRows) -> Mapping[str, object]:
+    end_user_id: Final = getattr(job, "end_user_id", None)
+    return {
+        "user_api_key_user_id": _creator_user_id(job.created_by),
+        **_shared_metadata(job, rows),
+        **({"user_api_key_end_user_id": end_user_id} if end_user_id else {}),
+        "user_api_key_project_id": getattr(rows.key, "project_id", None),
+        "user_api_key_project_alias": getattr(getattr(rows.key, "litellm_project_table", None), "project_alias", None),
+        **_key_budget_fields(rows.key),
+        **_budget_fields(
+            rows.team,
+            "user_api_key_team_spend",
+            "user_api_key_team_max_budget",
+            "user_api_key_team_model_max_budget",
+        ),
+        **_budget_fields(
+            rows.user,
+            "user_api_key_user_spend",
+            "user_api_key_user_max_budget",
+            "user_api_key_user_model_max_budget",
+        ),
+    }
+
+
+def _shared_metadata(job: ManagedObjectRow, rows: _CreatorRows) -> Mapping[str, object]:
+    api_key: Final = getattr(job, "api_key", None)
+    team_id: Final = getattr(job, "team_id", None)
+    key_alias: Final = _key_alias(api_key, job.created_by, rows.key)
+    team_alias: Final = getattr(rows.team, "team_alias", None)
+    org_id: Final = getattr(job, "org_id", None) or _organization_id(rows.key) or _organization_id(rows.team)
+    tags: Final = _string_tags(getattr(job, "request_tags", None))
+    return {
+        "user_api_key": api_key,
+        "user_api_key_hash": api_key,
+        "user_api_key_team_id": team_id,
+        **(
+            {
+                "user_api_key_user_email": getattr(rows.user, "user_email", None),
+                "user_api_key_alias": getattr(rows.user, "user_alias", None),
+            }
+            if rows.user is not None
+            else {}
+        ),
+        **({"user_api_key_alias": key_alias} if key_alias is not None else {}),
+        **({"user_api_key_team_alias": team_alias} if team_alias is not None else {}),
+        **({"user_api_key_org_id": org_id} if org_id is not None else {}),
+        **({"tags": tags} if tags else {}),
+    }
+
+
+def row_metadata(row: object) -> Mapping[str, object]:
+    metadata: Final = getattr(row, "metadata", None)
+    if not isinstance(metadata, dict):
+        return {}
+    return cast(Mapping[str, object], metadata)  # cast-ok: prisma Json object column
+
+
+def _names_a_key_row(api_key: object) -> TypeGuard[str]:
+    from litellm.litellm_core_utils.litellm_logging import is_valid_sha256_hash
+
+    return isinstance(api_key, str) and is_valid_sha256_hash(api_key)
+
+
+def _is_cli_session_key(api_key: str, created_by: str | None) -> bool:
+    return bool(created_by) and api_key == f"{CLI_SESSION_KEY_PREFIX}-{created_by}"
+
+
+def _key_alias(
+    api_key: str | None, created_by: str | None, key_row: prisma_models.LiteLLM_VerificationToken | None
+) -> str | None:
+    if api_key and _is_cli_session_key(api_key, created_by):
+        return api_key
+    return getattr(key_row, "key_alias", None) if key_row is not None else None
+
+
+def _organization_id(row: object) -> str | None:
+    return cast("str | None", getattr(row, "organization_id", None)) if row is not None else None
+
+
+def _budget_fields(row: object, spend_key: str, max_budget_key: str, model_max_budget_key: str) -> Mapping[str, object]:
+    if row is None:
+        return {}
+    return {
+        spend_key: getattr(row, "spend", None),
+        max_budget_key: getattr(row, "max_budget", None),
+        model_max_budget_key: getattr(row, "model_max_budget", None),
+    }
+
+
+def _key_budget_fields(key: object) -> Mapping[str, object]:
+    if key is None:
+        return {}
+    tier: Final = getattr(key, "litellm_budget_table", None)
+    return {
+        "user_api_key_spend": getattr(key, "spend", None),
+        "user_api_key_max_budget": _own_or_tier_budget(key, tier, "max_budget"),
+        "user_api_key_model_max_budget": _own_or_tier_budget(key, tier, "model_max_budget"),
+    }
+
+
+def _own_or_tier_budget(key: object, tier: object, field: str) -> object:
+    own: Final = getattr(key, field, None)
+    tier_value: Final = getattr(tier, field, None)
+    return own if tier_value is None or own not in (None, {}) else tier_value

enterprise/litellm_enterprise/proxy/hooks/managed_files.py

@@ -3,9 +3,13 @@
 
 import asyncio
 import base64
+import copy
 import json
 import time
 from collections.abc import Awaitable, Callable, Iterator, Mapping, Sequence
+from dataclasses import dataclass
+from datetime import datetime, timedelta
+from itertools import chain
 from types import MappingProxyType
 from typing import (
     TYPE_CHECKING,
@@ -19,35 +23,40 @@ from typing import (
     TypedDict,
     Union,
     cast,
+    runtime_checkable,
 )
 from uuid import NAMESPACE_URL, uuid5
 
 import httpx
 from fastapi import HTTPException
-from pydantic import BaseModel, ValidationError
+from openai import APIConnectionError, AsyncOpenAI
+from openai.types.file_deleted import FileDeleted
+from pydantic import BaseModel, ConfigDict, JsonValue, TypeAdapter, ValidationError
 from typing_extensions import ReadOnly, Unpack
 
 import litellm
 from litellm import Router, verbose_logger
-from litellm._internal_context import with_service_target
+from litellm._internal_context import internal_sub_call, post_response_phase, with_service_target
 from litellm._uuid import uuid
 from litellm.caching.caching import DualCache
 from litellm.constants import (
     BATCH_OUTPUT_FILE_FALLBACK_RETRY_AFTER_SECONDS,
     BATCH_OUTPUT_FILE_LOOKUP_TIMEOUT_SECONDS,
+    MANAGED_BACKGROUND_RESPONSE_METADATA_KEY,
+    MANAGED_OBJECT_STALENESS_CUTOFF_DAYS,
     MAX_FILE_LIST_LIMIT,
+    PENDING_RESPONSE_STATUSES,
 )
 from litellm.files.types import FileRetrieveCallOptions, FileRetrieveProvider
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.litellm_core_utils.core_helpers import require_attributes
 from litellm.litellm_core_utils.hidden_params import HIDDEN_PARAMS_ATTR
+from litellm.litellm_core_utils.internal_call_metadata import requests_background
 from litellm.litellm_core_utils.prompt_templates.common_utils import (
     extract_file_metadata,
 )
-from openai import APIConnectionError, AsyncOpenAI
-from openai.types.file_deleted import FileDeleted
-
 from litellm.llms.base_llm.files.storage_backend_factory import get_storage_backend
-from litellm.llms.base_llm.files.transformation import BaseFileEndpoints
+from litellm.llms.base_llm.files.transformation import BaseFileEndpoints, FileListModelScope
 from litellm.llms.base_llm.managed_resources.isolation import (
     build_list_page,
     build_owner_filter,
@@ -88,6 +97,8 @@ from litellm.proxy.pass_through_endpoints.llm_provider_handlers.batch_attributio
     request_tags_from_metadata,
 )
 from litellm.repositories.managed_file_repository import ManagedFileRepository
+from litellm.responses.streaming_iterator import spawn_logging_task
+from litellm.responses.utils import ResponsesAPIRequestUtils
 from litellm.types.llms.openai import (  # pyright: ignore[reportAttributeAccessIssue]
     AllMessageValues,
     AsyncCursorPage,
@@ -100,6 +111,7 @@ from litellm.types.llms.openai import (  # pyright: ignore[reportAttributeAccess
     ResponsesAPIResponse,
 )
 from litellm.types.utils import (
+    ADDRESSED_RESPONSE_ID_FIELD,
     CallTypesLiteral,
     LiteLLMBatch,
     LiteLLMFineTuningJob,
@@ -107,6 +119,7 @@ from litellm.types.utils import (
     LLMResponseTypes,
     SpecialEnums,
 )
+from litellm.utils import get_utc_datetime
 
 
 class _ManagedFileRetrieve(Protocol):
@@ -128,7 +141,7 @@ if TYPE_CHECKING:
     from litellm.proxy.utils import InternalUsageCache as _InternalUsageCache
     from litellm.proxy.utils import PrismaClient as _PrismaClient
 
-    Span = Union[_Span, Any]
+    Span = _Span
     InternalUsageCache = _InternalUsageCache
     PrismaClient = _PrismaClient
 else:
@@ -236,7 +249,7 @@ def _provider_file_retrieve_credentials(
     if llm_router is None or model_id is None:
         return None
     try:
-        credentials: Final = llm_router.get_deployment_credentials_with_provider(model_id)
+        credentials: Final = llm_router.registry().credentials(model_id, form="with_provider")
     except Exception as error:
         verbose_logger.warning(
             "Failed to retrieve credentials for provider file "
@@ -300,6 +313,7 @@ def _is_xai_batch_output(
 class _ManagedFileRow(Protocol):
     unified_file_id: str
     file_object: OpenAIFileObject
+    model_mappings: Mapping[str, str]
     flat_model_file_ids: Sequence[str]
     storage_backend: Optional[str]
     storage_url: Optional[str]
@@ -346,11 +360,15 @@ class _ManagedObjectTableActions(Protocol):
 
     async def update_many(self, where: Mapping[str, object], data: Mapping[str, object]) -> int: ...
 
+    async def create(self, data: Mapping[str, object]) -> "PrismaManagedObjectRow": ...
+
 
 class _ManagedResourceDatabase(Protocol):
     @property
     def litellm_managedfiletable(self) -> _ManagedFileTableActions: ...
 
+    async def query_raw(self, query: str, *args: object) -> object: ...
+
     @property
     def litellm_managedobjecttable(self) -> _ManagedObjectTableActions: ...
 
@@ -360,6 +378,12 @@ class _ManagedResourcePrismaClient(Protocol):
     def db(self) -> _ManagedResourceDatabase: ...
 
 
+@runtime_checkable
+class _WriterRoutedDatabase(Protocol):
+    @property
+    def writer(self) -> _ManagedResourceDatabase: ...
+
+
 class _SchedulerWithJobLookup(Protocol):
     def get_job(self, job_id: str) -> object: ...
 
@@ -374,10 +398,211 @@ class _RouterFileCallKwargs(TypedDict, total=False):
     custom_llm_provider: ReadOnly[str | None]
 
 
+_MODEL_FILE_ID_MAPPING: Final = TypeAdapter(dict[str, str])
+_FILE_UPLOAD_ONLY_FIELDS: Final = frozenset({"file", "purpose", "expires_after", "extra_body"})
+_FILE_DELETE_RESERVED_FIELDS: Final = frozenset({"model", "file_id", "_litellm_internal_model_credentials"})
+
+
+@dataclass(frozen=True, slots=True)
+class _ProviderUpload:
+    model_id: str
+    provider_file_id: str
+    upload_call_id: str | None = None
+
+
+def _provider_uploads(file_object: OpenAIFileObject) -> tuple[_ProviderUpload, ...]:
+    file_hidden_params: Final = cast(  # cast-ok: preserve mapping operations on dynamic file metadata
+        dict[str, object], getattr(file_object, HIDDEN_PARAMS_ATTR)
+    )
+    model_file_id_mapping: Final = _MODEL_FILE_ID_MAPPING.validate_python(
+        file_hidden_params.get("model_file_id_mapping") or {}
+    )
+    upload_call_id: Final = file_hidden_params.get("litellm_call_id")
+    return tuple(
+        _ProviderUpload(
+            model_id=model_id,
+            provider_file_id=provider_file_id,
+            upload_call_id=upload_call_id if isinstance(upload_call_id, str) else None,
+        )
+        for model_id, provider_file_id in model_file_id_mapping.items()
+    )
+
+
+def _keeps_one_upload_per_target(llm_router: Router, model_id: str) -> bool:
+    credentials: Final = _provider_file_retrieve_credentials(llm_router=llm_router, model_id=model_id)
+    return (credentials or {}).get("custom_llm_provider") == LlmProviders.LITELLM_PROXY.value
+
+
+def _is_kept_upload(uploads: Sequence[_ProviderUpload], index: int, per_target_model_ids: frozenset[str]) -> bool:
+    upload: Final = uploads[index]
+    other_uploads: Final = uploads[index + 1 :] if upload.model_id in per_target_model_ids else uploads[:index]
+    return all(other.model_id != upload.model_id for other in other_uploads)
+
+
+def _upload_from_trailing_id_fields(decoded_unified_file_id: str) -> _ProviderUpload:
+    trailing_fields: Final = decoded_unified_file_id.rpartition(";llm_output_file_id,")[2]
+    provider_file_id, _, model_id = trailing_fields.partition(";llm_output_file_model_id,")
+    return _ProviderUpload(model_id=model_id, provider_file_id=provider_file_id)
+
+
+def _carried_upload_missing_from(unified_file_id: str, model_file_ids: Mapping[str, str]) -> _ProviderUpload | None:
+    decoded: Final = _is_base64_encoded_unified_file_id(unified_file_id)
+    carried: Final = _upload_from_trailing_id_fields(decoded) if decoded else None
+    if carried is None or model_file_ids.get(carried.model_id) in (None, carried.provider_file_id):
+        return None
+    return carried
+
+
+async def _delete_provider_file(
+    llm_router: Router,
+    model_id: str,
+    provider_file_id: str,
+    request_data: Mapping[str, object],
+) -> None:
+    credentials: Final = llm_router.registry().credentials(model_id, form="with_provider")
+    delete_data: Final = {
+        **{k: v for k, v in request_data.items() if k not in _FILE_DELETE_RESERVED_FIELDS},
+        **(
+            {"_litellm_internal_model_credentials": MappingProxyType(dict(credentials))}
+            if credentials is not None
+            else {}
+        ),
+    }
+    await llm_router.afile_delete(
+        model=model_id,
+        file_id=provider_file_id,
+        **cast(_RouterFileCallKwargs, delete_data),
+    )
+
+
+def _with_own_nested_dicts(request_data: Mapping[str, object]) -> MappingProxyType[str, object]:
+    return MappingProxyType({k: copy.copy(v) if isinstance(v, dict) else v for k, v in request_data.items()})
+
+
+async def _delete_redundant_upload(
+    llm_router: Router,
+    upload: _ProviderUpload,
+    delete_request: Mapping[str, object],
+    parent_call_id: str | None,
+) -> None:
+    try:
+        with internal_sub_call(parent_call_id):
+            await _delete_provider_file(
+                llm_router, upload.model_id, upload.provider_file_id, _with_own_nested_dicts(delete_request)
+            )
+    except Exception as error:
+        verbose_logger.warning(
+            f"Could not delete the redundant upload provider_file_id={sanitize_for_log(upload.provider_file_id)} "
+            f"at deployment model_id={sanitize_for_log(upload.model_id)}: "
+            f"{type(error).__name__} {sanitize_for_log(error)}"
+        )
+
+
+async def _delete_redundant_uploads(
+    llm_router: Router,
+    uploads: Sequence[_ProviderUpload],
+    delete_request: Mapping[str, object],
+) -> None:
+    await asyncio.gather(
+        *(_delete_redundant_upload(llm_router, upload, delete_request, upload.upload_call_id) for upload in uploads)
+    )
+
+
+def _start_deleting_redundant_uploads(
+    llm_router: Router,
+    uploads: Sequence[_ProviderUpload],
+    create_file_request: CreateFileRequest,
+) -> None:
+    if not uploads:
+        return
+    delete_request: Final = MappingProxyType(
+        {k: v for k, v in create_file_request.items() if k not in _FILE_UPLOAD_ONLY_FIELDS}
+    )
+    with post_response_phase():
+        spawn_logging_task(
+            asyncio.get_running_loop(),
+            _delete_redundant_uploads(llm_router, uploads, delete_request),
+            task_name="managed file repeat upload cleanup",
+        )
+
+
+def _managed_database(prisma_client: _ManagedResourcePrismaClient) -> _ManagedResourceDatabase:
+    return prisma_client.db
+
+
 def _managed_file_table(prisma_client: _ManagedResourcePrismaClient) -> _ManagedFileTableActions:
     return prisma_client.db.litellm_managedfiletable
 
 
+class _ListedFileRow(Protocol):
+    @property
+    def unified_file_id(self) -> str: ...
+
+    @property
+    def file_object(self) -> object: ...
+
+    @property
+    def model_mappings(self) -> Mapping[str, object]: ...
+
+    @property
+    def storage_backend(self) -> str | None: ...
+
+
+class _ScopedFileRow(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    unified_file_id: str
+    file_object: JsonValue = None
+    model_mappings: Mapping[str, JsonValue]
+    storage_backend: str | None = None
+
+
+_SCOPED_FILE_ROWS: Final = TypeAdapter(tuple[_ScopedFileRow, ...])
+
+
+def _owner_condition(owner_filter: Mapping[str, object]) -> tuple[str, tuple[str, ...]]:
+    match owner_filter:
+        case {"OR": [{"created_by": str() as user_id}, {"team_id": str() as team_id}]}:
+            return '("created_by" = $1 OR "team_id" = $2)', (user_id, team_id)
+        case {"created_by": str() as user_id}:
+            return '"created_by" = $1', (user_id,)
+        case {"team_id": str() as team_id}:
+            return '"team_id" = $1', (team_id,)
+        case _:
+            return ("FALSE", ()) if owner_filter else ("TRUE", ())
+
+
+async def _scoped_file_rows(
+    database: _ManagedResourceDatabase,
+    owner_filter: Mapping[str, object],
+    model_mapping_keys: frozenset[str],
+    cursor_id: str | None,
+    direction: Literal["asc", "desc"],
+    take: int,
+) -> tuple[_ScopedFileRow, ...]:
+    owner_condition, owner_params = _owner_condition(owner_filter)
+    keys_param: Final = len(owner_params) + 1
+    comparison: Final = "<" if direction == "desc" else ">"
+    cursor_condition: Final = (
+        f'AND ("created_at", "unified_file_id") {comparison} (SELECT "created_at", "unified_file_id" '
+        f'FROM "LiteLLM_ManagedFileTable" WHERE "unified_file_id" = ${keys_param + 1}) '
+        if cursor_id
+        else ""
+    )
+    order: Final = direction.upper()
+    rows: Final = await database.query_raw(
+        'SELECT "unified_file_id", "file_object", "model_mappings", "storage_backend" FROM "LiteLLM_ManagedFileTable" '
+        f'WHERE {owner_condition} AND ("model_mappings" ?| ${keys_param}::text[] OR "storage_backend" IS NOT NULL) '
+        f'{cursor_condition}ORDER BY "created_at" {order}, "unified_file_id" {order} '
+        f"LIMIT ${keys_param + (2 if cursor_id else 1)}",
+        *owner_params,
+        sorted(model_mapping_keys),
+        *((cursor_id,) if cursor_id else ()),
+        take,
+    )
+    return _SCOPED_FILE_ROWS.validate_python(rows)
+
+
 def _iter_provider_file_id_pairs(
     rows: Sequence[_ManagedFileRow],
     requested_provider_file_ids: frozenset[str],
@@ -392,6 +617,63 @@ def _managed_object_table(prisma_client: _ManagedResourcePrismaClient) -> _Manag
     return prisma_client.db.litellm_managedobjecttable
 
 
+def _writer_database(db: _ManagedResourceDatabase) -> _ManagedResourceDatabase:
+    """Reads that feed a write must see the create that just happened; a lagging read replica may not"""
+    return db.writer if isinstance(db, _WriterRoutedDatabase) else db
+
+
+_JSON_VALUE_ADAPTER: Final[TypeAdapter[JsonValue]] = TypeAdapter(JsonValue)
+_POSTGRES_UNSTORABLE_CHARACTER: Final = "\x00"
+
+
+def _storable_text(text: str) -> str:
+    return text.replace(_POSTGRES_UNSTORABLE_CHARACTER, "")
+
+
+def _storable_json_value(value: JsonValue) -> JsonValue:
+    if isinstance(value, str):
+        return _storable_text(value)
+    if isinstance(value, list):
+        return [_storable_json_value(item) for item in value]
+    if isinstance(value, dict):
+        return {_storable_text(key): _storable_json_value(item) for key, item in value.items()}
+    return value
+
+
+def postgres_storable_json(file_object: BaseModel) -> str:
+    dumped: Final = file_object.model_dump_json()
+    if "\\u0000" not in dumped:
+        return dumped
+    return _JSON_VALUE_ADAPTER.dump_json(_storable_json_value(_JSON_VALUE_ADAPTER.validate_json(dumped))).decode()
+
+
+def _without_model_output(
+    file_object: LiteLLMBatch | LiteLLMFineTuningJob | ResponsesAPIResponse,
+) -> LiteLLMBatch | LiteLLMFineTuningJob | ResponsesAPIResponse:
+    return (
+        file_object.model_copy(update={"output": []}) if isinstance(file_object, ResponsesAPIResponse) else file_object
+    )
+
+
+def _with_model_id(
+    file_object: LiteLLMBatch | LiteLLMFineTuningJob | ResponsesAPIResponse,
+    model_id: str | None,
+) -> LiteLLMBatch | LiteLLMFineTuningJob | ResponsesAPIResponse:
+    return file_object if model_id is None else file_object.model_copy(update={"model_id": model_id})
+
+
+async def _stored_model_id(managed_object_table: _ManagedObjectTableActions, unified_object_id: str) -> str | None:
+    row: Final = await managed_object_table.find_first(where={"unified_object_id": unified_object_id})
+    if row is None:
+        return None
+    try:
+        parsed: Final = _JSON_VALUE_ADAPTER.validate_python(_decode_json_blob(row.file_object))
+    except (json.JSONDecodeError, ValidationError):
+        return None
+    model_id: Final = parsed.get("model_id") if isinstance(parsed, dict) else None
+    return model_id if isinstance(model_id, str) and model_id else None
+
+
 def _storage_metadata_of(file_object: OpenAIFileObject | None) -> Mapping[str, str]:
     hidden_params: Final = cast(  # cast-ok: _hidden_params is an untyped attribute the upload path sets
         "Mapping[str, object]", getattr(file_object, "_hidden_params", None) or {}
@@ -405,10 +687,59 @@ def _storage_metadata_of(file_object: OpenAIFileObject | None) -> Mapping[str, s
     )
 
 
+_MANAGED_RESPONSE_READ_CALL_TYPES: Final = frozenset({"aget_responses", "acancel_responses"})
+_REQUEST_FIELDS: Final = TypeAdapter(dict[str, object])
+
+
+def _addresses_a_proxy_issued_response(data: Mapping[str, object]) -> bool:
+    from litellm.proxy.hooks.responses_id_security import decrypted_response_id
+
+    addressed_id: Final = data.get(ADDRESSED_RESPONSE_ID_FIELD) or data.get("response_id")
+    return (
+        isinstance(addressed_id, str)
+        and ResponsesAPIRequestUtils.get_model_id_from_response_id(decrypted_response_id(addressed_id)) is not None
+    )
+
+
+def _is_streamed_background_create(request: Mapping[str, object]) -> bool:
+    return (
+        request.get("stream") is True
+        and isinstance(request.get("litellm_metadata"), dict)
+        and requests_background(request)
+    )
+
+
+def _with_managed_background_marker(request: Mapping[str, object]) -> dict[str, object]:
+    metadata_field: Final = "litellm_metadata" if isinstance(request.get("litellm_metadata"), dict) else "metadata"
+    existing_metadata: Final = request.get(metadata_field)
+    return {
+        **request,
+        metadata_field: {
+            **(_REQUEST_FIELDS.validate_python(existing_metadata) if isinstance(existing_metadata, dict) else {}),
+            MANAGED_BACKGROUND_RESPONSE_METADATA_KEY: True,
+        },
+    }
+
+
 _MANAGED_FILES_TARGET: Final = "managed_files"
+_PROVIDER_FILE_ID_KWARGS: Final = MappingProxyType(
+    {
+        CallTypes.acreate_batch: ("input_file_id",),
+        CallTypes.acreate_fine_tuning_job: ("training_file", "validation_file"),
+    }
+)
+_PROVIDER_FILE_ID_KEYS: Final = tuple(dict.fromkeys(chain.from_iterable(_PROVIDER_FILE_ID_KWARGS.values())))
 _PROVIDER_FILE_RETRIEVE_RETRY_DELAYS_SECONDS: Final = (0.5, 1.0, 2.0)
 
 
+class _SupportsKeyedGet(Protocol):
+    def get(self, key: str, default: object, /) -> object: ...
+
+
+def _keyed_get(source: object, key: str, default: object = None) -> object:
+    return require_attributes(source, _SupportsKeyedGet, "get").get(key, default)
+
+
 class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
     # Class variables or attributes
     def __init__(
@@ -417,10 +748,12 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         prisma_client: PrismaClient,
         *,
         sleep: Callable[[float], Awaitable[None]] = asyncio.sleep,
+        clock: Callable[[], datetime] = get_utc_datetime,
     ):
         self.internal_usage_cache = internal_usage_cache
         self.prisma_client = prisma_client
         self._sleep = sleep
+        self._clock = clock
 
     @staticmethod
     def _get_prometheus_logger():
@@ -526,14 +859,25 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         persist_attribution: bool = False,
         create_if_missing: bool = True,
         batch_processed: bool = False,
+        creating_model_id: str | None = None,
     ) -> None:
         """Persist a managed object row, caching it and upserting it in the DB.
 
-        persist_attribution is set only by the batch create, which is the one caller
-        that can speak for the creator; it gates the api_key and request_tags columns
-        that CheckBatchCost bills against, so a later poll or retrieve of the same
-        batch cannot record itself as the paying key. Like created_by and team_id,
-        both are written only in the upsert create branch, never on update.
+        creating_model_id is the router deployment that created the object. It is kept
+        inside the stored file_object JSON as `model_id`, because the raw provider id an
+        unmanaged batch stores as unified_object_id says nothing about which deployment
+        made it, and the cost poller needs that deployment to read the result back. A
+        later observe of the row (create_if_missing=False) keeps the stored value rather
+        than overwriting it with whichever deployment served the retrieve.
+
+        persist_attribution is set only by the batch and background response creates,
+        the callers that can speak for the creator; it gates the api_key, request_tags, and,
+        for a response, end_user_id columns that the cost poll jobs bill against, so a later poll or
+        retrieve of the same object cannot record itself as the paying key. Like created_by and team_id,
+        they are written only in the upsert create branch, never on update.
+
+        A stored response keeps no model output: billing reads only its status and model from
+        the row, so the row never holds generated content.
 
         create_if_missing is cleared by callers that observe a batch they did not
         create, such as a poll. They still refresh status and file_object, but a
@@ -546,11 +890,19 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         It is written only in the upsert create branch.
         """
         verbose_logger.info(f"Storing LiteLLM Managed {file_purpose} object with id={unified_object_id} in cache")
+        managed_database: Final = _managed_database(self.prisma_client)
+        managed_object_table: Final = managed_database.litellm_managedobjecttable
+        preserved_model_id: Final = (
+            creating_model_id
+            if creating_model_id is not None or create_if_missing
+            else await _stored_model_id(_writer_database(managed_database).litellm_managedobjecttable, unified_object_id)
+        )
+        stored_object: Final = _with_model_id(_without_model_output(file_object), preserved_model_id)
         litellm_managed_object = LiteLLM_ManagedObjectTable(
             unified_object_id=unified_object_id,
             model_object_id=model_object_id,
             file_purpose=file_purpose,
-            file_object=file_object,
+            file_object=stored_object,
         )
         await self.internal_usage_cache.async_set_cache(
             key=unified_object_id,
@@ -561,32 +913,34 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         from prisma import Json
 
         api_key = LiteLLMProxyRequestSetup.get_logged_api_key(user_api_key_dict) or None
+        end_user_id: Final = user_api_key_dict.end_user_id if file_purpose == "response" else None
         attribution_columns = (
             {
                 **({"api_key": api_key} if api_key is not None else {}),
-                **({"request_tags": Json(list(request_tags))} if request_tags else {}),
+                **({"request_tags": Json([_storable_text(tag) for tag in request_tags])} if request_tags else {}),
+                **({"end_user_id": end_user_id} if end_user_id else {}),
             }
             if persist_attribution
             else {}
         )
         # FIX: Update status and file_object on every operation to keep state in sync
         update_columns: Final = {
-            "file_object": file_object.model_dump_json(),
+            "file_object": postgres_storable_json(stored_object),
             "status": file_object.status,
             "updated_by": user_api_key_dict.user_id,
         }
         if not create_if_missing:
-            await _managed_object_table(self.prisma_client).update_many(
+            await managed_object_table.update_many(
                 where={"unified_object_id": unified_object_id},
                 data=update_columns,
             )
             return
-        await _managed_object_table(self.prisma_client).upsert(
+        await managed_object_table.upsert(
             where={"unified_object_id": unified_object_id},
             data={
                 "create": {
                     "unified_object_id": unified_object_id,
-                    "file_object": file_object.model_dump_json(),
+                    "file_object": postgres_storable_json(stored_object),
                     "model_object_id": model_object_id,
                     "file_purpose": file_purpose,
                     "created_by": resolve_resource_owner_id(user_api_key_dict),
@@ -601,6 +955,42 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             },
         )
 
+    async def track_untracked_background_response(
+        self,
+        response_id: str,
+        response: ResponsesAPIResponse,
+        user_api_key_dict: UserAPIKeyAuth,
+        request_tags: Sequence[str] | None = None,
+    ) -> bool:
+        if not _addresses_a_proxy_issued_response({"response_id": response_id}):
+            return False
+        from prisma import Json
+        from prisma.errors import UniqueViolationError
+
+        from litellm.proxy.hooks.responses_id_security import decrypted_response_id
+
+        api_key: Final = LiteLLMProxyRequestSetup.get_logged_api_key(user_api_key_dict) or None
+        try:
+            await _managed_object_table(self.prisma_client).create(
+                data={
+                    "unified_object_id": response_id,
+                    "model_object_id": decrypted_response_id(response_id),
+                    "file_purpose": "response",
+                    "file_object": postgres_storable_json(_without_model_output(response)),
+                    "status": PENDING_RESPONSE_STATUSES[0],
+                    "created_by": resolve_resource_owner_id(user_api_key_dict),
+                    "team_id": user_api_key_dict.team_id,
+                    "org_id": await self._resolve_creator_org_id(user_api_key_dict),
+                    "updated_by": user_api_key_dict.user_id,
+                    **({"api_key": api_key} if api_key is not None else {}),
+                    **({"end_user_id": user_api_key_dict.end_user_id} if user_api_key_dict.end_user_id else {}),
+                    **({"request_tags": Json([_storable_text(tag) for tag in request_tags])} if request_tags else {}),
+                }
+            )
+        except UniqueViolationError:
+            return False
+        return True
+
     @with_service_target(_MANAGED_FILES_TARGET)
     async def get_unified_file_id(
         self, file_id: str, litellm_parent_otel_span: Optional[Span] = None
@@ -1375,6 +1765,9 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 unified_vs_ids = self.get_vector_store_ids_from_file_search_tools(tools)
                 if unified_vs_ids:
                     await self.check_vector_store_ids_access(unified_vs_ids, user_api_key_dict)
+            create_request: Final = _REQUEST_FIELDS.validate_python(data)
+            if _is_streamed_background_create(create_request):
+                return _with_managed_background_marker(create_request)
         elif call_type == CallTypes.afile_content.value:
             retrieve_file_id = cast(Optional[str], data.get("file_id"))
             potential_file_id = _is_base64_encoded_unified_file_id(retrieve_file_id) if retrieve_file_id else False
@@ -1432,6 +1825,10 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 data[accessor_key] = get_batch_id_from_unified_batch_id(potential_llm_object_id)
             elif retrieve_object_id and accessor_key == "batch_id":
                 await self.enforce_batch_object_access(retrieve_object_id, user_api_key_dict)
+        elif call_type in _MANAGED_RESPONSE_READ_CALL_TYPES:
+            request: Final = _REQUEST_FIELDS.validate_python(data)
+            if _addresses_a_proxy_issued_response(request):
+                return _with_managed_background_marker(request)
         elif call_type == CallTypes.acreate_fine_tuning_job.value:
             fine_tuning_file_ids: Final = tuple(
                 file_id
@@ -1439,6 +1836,10 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 if isinstance(file_id, str) and file_id
             )
             await self.check_file_ids_access(fine_tuning_file_ids, user_api_key_dict)
+            if fine_tuning_file_ids:
+                data["model_file_id_mapping"] = await self.get_model_file_id_mapping(
+                    fine_tuning_file_ids, user_api_key_dict.parent_otel_span
+                )
 
         return data
 
@@ -1453,50 +1854,49 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         if request_kwargs is None:
             return healthy_deployments
 
-        input_file_id = cast(Optional[str], request_kwargs.get("input_file_id"))
-        model_file_id_mapping = cast(
-            Optional[Dict[str, Dict[str, str]]],
-            request_kwargs.get("model_file_id_mapping"),
+        request: Final = cast(Mapping[str, object], request_kwargs)
+        model_file_id_mapping: Final = cast(Mapping[str, Mapping[str, str]], request.get("model_file_id_mapping") or {})
+        requested_file_ids: Final = (request.get(key) for key in _PROVIDER_FILE_ID_KEYS)
+        model_ids_per_file: Final = tuple(
+            frozenset(model_file_id_mapping[file_id])
+            for file_id in requested_file_ids
+            if isinstance(file_id, str) and file_id in model_file_id_mapping
+        )
+        allowed_model_ids: Final[frozenset[str]] = (
+            model_ids_per_file[0].intersection(*model_ids_per_file[1:]) if model_ids_per_file else frozenset()
         )
-        allowed_model_ids = []
-        if input_file_id and model_file_id_mapping:
-            model_id_dict = model_file_id_mapping.get(input_file_id, {})
-            allowed_model_ids = list(model_id_dict.keys())
 
         if len(allowed_model_ids) == 0:
             return healthy_deployments
 
-        return [
+        holders: Final[list[dict[str, object]]] = [
             deployment
             for deployment in healthy_deployments
             if deployment.get("model_info", {}).get("id") in allowed_model_ids
         ]
+        return holders if holders or "input_file_id" in request else healthy_deployments
 
     async def async_pre_call_deployment_hook(
-        self, kwargs: Dict[str, Any], call_type: Optional[CallTypes]
+        self, kwargs: dict[str, object], call_type: Optional[CallTypes]
     ) -> Optional[dict]:
         """
         Allow modifying the request just before it's sent to the deployment.
         """
-        accessor_key: Optional[str] = None
-        if call_type and call_type == CallTypes.acreate_batch:
-            accessor_key = "input_file_id"
-        elif call_type and call_type == CallTypes.acreate_fine_tuning_job:
-            accessor_key = "training_file"
-        else:
+        accessor_keys: Final = _PROVIDER_FILE_ID_KWARGS.get(call_type) if call_type else None
+        if accessor_keys is None:
             return kwargs
 
-        if accessor_key:
+        model_file_id_mapping = cast(Optional[Dict[str, Dict[str, str]]], kwargs.get("model_file_id_mapping"))
+        # model_info may be at top-level or nested under litellm_metadata
+        # (batch/file operations use litellm_metadata)
+        model_id = cast(Optional[str], _keyed_get(kwargs.get("model_info", {}), "id"))
+        if model_id is None:
+            model_id = cast(
+                Optional[str],
+                _keyed_get(_keyed_get(kwargs.get("litellm_metadata", {}), "model_info", {}), "id"),
+            )
+        for accessor_key in accessor_keys:
             input_file_id = cast(Optional[str], kwargs.get(accessor_key))
-            model_file_id_mapping = cast(Optional[Dict[str, Dict[str, str]]], kwargs.get("model_file_id_mapping"))
-            # model_info may be at top-level or nested under litellm_metadata
-            # (batch/file operations use litellm_metadata)
-            model_id = cast(Optional[str], kwargs.get("model_info", {}).get("id", None))
-            if model_id is None:
-                model_id = cast(
-                    Optional[str],
-                    kwargs.get("litellm_metadata", {}).get("model_info", {}).get("id", None),
-                )
             mapped_file_id: Optional[str] = None
             if input_file_id and model_file_id_mapping and model_id:
                 mapped_file_id = model_file_id_mapping.get(input_file_id, {}).get(model_id, None)
@@ -1690,7 +2090,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                     ),
                 )
 
-    async def get_model_file_id_mapping(self, file_ids: List[str], litellm_parent_otel_span: Span) -> dict:
+    async def get_model_file_id_mapping(self, file_ids: Sequence[str], litellm_parent_otel_span: Span | None) -> dict:
         """
         Get model-specific file IDs for a list of proxy file IDs.
         Returns a dictionary mapping litellm_proxy/ file_id -> model_id -> model_file_id
@@ -1732,7 +2132,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         llm_router: Optional[Router],
         _create_file_request: CreateFileRequest,
         target_model_names_list: List[str],
-        litellm_parent_otel_span: Span,
+        litellm_parent_otel_span: Span | None,
     ) -> List[OpenAIFileObject]:
         if llm_router is None:
             raise Exception("LLM Router not initialized. Ensure models added to proxy.")
@@ -1748,7 +2148,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         create_file_request: CreateFileRequest,
         llm_router: Router,
         target_model_names_list: List[str],
-        litellm_parent_otel_span: Span,
+        litellm_parent_otel_span: Span | None,
         user_api_key_dict: UserAPIKeyAuth,
     ) -> OpenAIFileObject:
         responses = await self.create_file_for_each_model(
@@ -1765,16 +2165,16 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             target_model_names_list=target_model_names_list,
         )
 
-        ## STORE MODEL MAPPINGS IN DB
-        model_mappings: Dict[str, str] = {}
-
-        for file_object in responses:
-            file_hidden_params = cast(  # cast-ok: preserve mapping operations on dynamic file metadata
-                dict[str, object], getattr(file_object, HIDDEN_PARAMS_ATTR)
-            )
-            model_file_id_mapping = file_hidden_params.get("model_file_id_mapping")
-            if model_file_id_mapping and isinstance(model_file_id_mapping, dict):
-                model_mappings.update(model_file_id_mapping)
+        uploads: Final = tuple(chain.from_iterable(map(_provider_uploads, responses)))
+        per_target_model_ids: Final = frozenset(
+            model_id
+            for model_id in {upload.model_id for upload in uploads}
+            if _keeps_one_upload_per_target(llm_router, model_id)
+        )
+        kept_uploads: Final = tuple(
+            upload for index, upload in enumerate(uploads) if _is_kept_upload(uploads, index, per_target_model_ids)
+        )
+        model_mappings: Final = {upload.model_id: upload.provider_file_id for upload in kept_uploads}
 
         await self.store_unified_file_id(
             file_id=response.id,
@@ -1783,6 +2183,15 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             model_mappings=model_mappings,
             user_api_key_dict=user_api_key_dict,
         )
+        _start_deleting_redundant_uploads(
+            llm_router,
+            tuple(
+                upload
+                for upload in uploads
+                if upload not in kept_uploads and upload.model_id not in per_target_model_ids
+            ),
+            create_file_request,
+        )
 
         # Emit Prometheus metrics for managed file creation
         prom_logger = self._get_prometheus_logger()
@@ -1815,7 +2224,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         file_objects: List[OpenAIFileObject],
         create_file_request: CreateFileRequest,
         internal_usage_cache: InternalUsageCache,
-        litellm_parent_otel_span: Span,
+        litellm_parent_otel_span: Span | None,
         target_model_names_list: List[str],
     ) -> OpenAIFileObject:
         ## GET THE FILE TYPE FROM THE CREATE FILE REQUEST
@@ -1946,6 +2355,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 request_tags=request_tags_from_metadata(request_metadata if isinstance(request_metadata, dict) else {}),
                 persist_attribution=is_batch_create,
                 create_if_missing=is_batch_create,
+                creating_model_id=model_id if is_batch_create else None,
             )
             if not is_batch_create:
                 await self.store_batch_output_file_ownership(
@@ -2117,14 +2527,23 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         user_api_key_dict: UserAPIKeyAuth,
         limit: Optional[int] = None,
         after: Optional[str] = None,
+        model_scope: Optional[FileListModelScope] = None,
+        before: Optional[str] = None,
         **data: Dict,
     ) -> FileListPage:
         """List the managed files the caller owns, newest first.
 
         Pagination is keyset based on ``unified_file_id`` so a key that owns
         every file on the proxy still reads one bounded page at a time.
-        ``purpose`` is applied after parsing, because the managed file table
-        keeps it inside the ``file_object`` blob instead of a column, and rows
+        ``after`` pages toward older files and ``before`` toward newer ones,
+        each page still ordered newest first.
+        ``model_scope`` is applied in the query, so only rows whose
+        ``model_mappings`` name one of its deployment ids or model names, and
+        storage-backed rows, leave the database; a storage-backed row then
+        stays only when its model name is one the router sends to the listed
+        provider. ``purpose`` is applied after parsing, because the managed
+        file table keeps it inside the ``file_object`` blob instead of a
+        column. Rows
         whose blob will not parse drop out there too, so a chunk of rows can
         yield fewer matches than the page holds. Successive chunks are read
         until the page is full or the caller's rows run out, which keeps
@@ -2133,9 +2552,8 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         scan has to continue past it, the chunk widens to
         ``FILE_LIST_CONTINUATION_CHUNK_SIZE``, so the walk costs one query per
         that many rows instead of one per page. That bound is per query, not
-        per request: the work is still linear in the rows the caller owns, and
-        a filter matching nothing reads every one of them, with no index
-        covering either the owner filter or the sort.
+        per request: a ``purpose`` filter matching nothing still reads every
+        row that passes the query.
         """
         validate_file_list_limit(limit)
         validate_file_list_purpose(purpose)
@@ -2144,36 +2562,60 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         if owner_filter is None:
             return FileListPage.model_validate(build_list_page([]))
 
-        if after:
-            cursor_row = await _managed_file_table(self.prisma_client).find_first(
-                where={**owner_filter, "unified_file_id": after}
+        if after and before:
+            raise ProxyException(
+                message="Pass either 'after' (or 'after_id') or 'before_id' when listing files, not both.",
+                type="invalid_request_error",
+                param="before_id",
+                code=400,
+                openai_code="invalid_value",
+            )
+        cursor_param: Final = "before_id" if before else "after"
+        cursor: Final = before or after
+        database: Final = _managed_database(self.prisma_client)
+        if cursor:
+            cursor_row = await database.litellm_managedfiletable.find_first(
+                where={**owner_filter, "unified_file_id": cursor}
             )
             if cursor_row is None:
                 raise ProxyException(
-                    message=f"Invalid 'after' cursor: no file found with id '{after}'.",
+                    message=f"Invalid '{cursor_param}' cursor: no file found with id '{cursor}'.",
                     type="invalid_request_error",
-                    param="after",
+                    param=cursor_param,
                     code=400,
                     openai_code="invalid_value",
                 )
 
+        if model_scope is not None and not model_scope.model_mapping_keys:
+            return FileListPage.model_validate(build_list_page([]))
+
         page_size: Final = min(limit or MAX_FILE_LIST_LIMIT, MAX_FILE_LIST_LIMIT)
         matches: Final[List[OpenAIFileObject]] = []
-        cursor_id = after
+        cursor_id = cursor
         chunk_size = page_size + 1
+        direction: Final[Literal["asc", "desc"]] = "asc" if before else "desc"
 
         while len(matches) <= page_size:
             cursor_args: _CursorPageArgs = {"cursor": {"unified_file_id": cursor_id}, "skip": 1} if cursor_id else {}
-            chunk = await _managed_file_table(self.prisma_client).find_many(
-                where=owner_filter,
-                take=chunk_size,
-                order=[{"created_at": "desc"}, {"unified_file_id": "desc"}],
-                **cursor_args,
+            chunk: Sequence[_ListedFileRow] = (
+                await _scoped_file_rows(
+                    database, owner_filter, model_scope.model_mapping_keys, cursor_id, direction, chunk_size
+                )
+                if model_scope is not None
+                else await database.litellm_managedfiletable.find_many(
+                    where=owner_filter,
+                    take=chunk_size,
+                    order=[{"created_at": direction}, {"unified_file_id": direction}],
+                    **cursor_args,
+                )
             )
             matches.extend(
                 _public_file_object(parsed_file_object, row.unified_file_id)
                 for row in chunk
-                if (parsed_file_object := _parse_managed_file_object(row.file_object, row.unified_file_id)) is not None
+                if (
+                    model_scope is None or model_scope.holds(row.model_mappings, stored=row.storage_backend is not None)
+                )
+                and (parsed_file_object := _parse_managed_file_object(row.file_object, row.unified_file_id)) is not None
                 and (purpose is None or parsed_file_object.purpose == purpose)
             )
             if len(chunk) < chunk_size:
@@ -2181,7 +2623,10 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             cursor_id = chunk[-1].unified_file_id
             chunk_size = max(chunk_size, FILE_LIST_CONTINUATION_CHUNK_SIZE)
 
-        return FileListPage.model_validate(build_list_page(matches[:page_size], has_more=len(matches) > page_size))
+        page: Final = matches[:page_size]
+        return FileListPage.model_validate(
+            build_list_page(page[::-1] if before else page, has_more=len(matches) > page_size)
+        )
 
     def _is_batch_polling_enabled(self) -> bool:
         """
@@ -2244,12 +2689,13 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
             where={
                 "file_purpose": "batch",
                 "batch_processed": False,
-                "status": {"not_in": ["failed", "expired", "cancelled"]},
+                "status": {"not_in": ["failed", "expired", "cancelled", "stale_expired"]},
             },
             take=MAX_MATCHES_TO_RETURN,
             order={"created_at": "desc"},
         )
 
+        input_file_cutoff: Final = self._clock() - timedelta(days=MANAGED_OBJECT_STALENESS_CUTOFF_DAYS)
         referencing_batches: Final[list[dict[str, object]]] = []
         for batch in batches:
             try:
@@ -2262,7 +2708,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 # Extract file IDs from batch
                 # Batches typically reference the unified file ID in input_file_id
                 # Output and error files are generated by the provider
-                input_file_id = batch_data.get("input_file_id")
+                input_file_id = batch_data.get("input_file_id") if batch.created_at >= input_file_cutoff else None
                 output_file_id = batch_data.get("output_file_id")
                 error_file_id = batch_data.get("error_file_id")
 
@@ -2386,23 +2832,13 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
         specific_model_file_id_mapping: Final = model_file_id_mapping.get(file_id)
         if not specific_model_file_id_mapping:
             return
-        filtered_data: Final = {
-            k: v for k, v in data.items() if k not in ("model", "file_id", "_litellm_internal_model_credentials")
-        }
         for model_id, model_file_id in specific_model_file_id_mapping.items():
-            credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_id)
-            delete_data = {
-                **filtered_data,
-                **(
-                    {"_litellm_internal_model_credentials": MappingProxyType(dict(credentials))}
-                    if credentials is not None
-                    else {}
-                ),
-            }
-            await llm_router.afile_delete(
-                model=model_id,
-                file_id=model_file_id,
-                **cast(_RouterFileCallKwargs, delete_data),
+            await _delete_provider_file(llm_router, model_id, model_file_id, data)
+        unrecorded_upload: Final = _carried_upload_missing_from(file_id, specific_model_file_id_mapping)
+        delete_call_id: Final = data.get("litellm_call_id")
+        if unrecorded_upload is not None:
+            await _delete_redundant_upload(
+                llm_router, unrecorded_upload, data, delete_call_id if isinstance(delete_call_id, str) else None
             )
 
     async def afile_content(
@@ -2434,7 +2870,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                     # against the deployment's configured bucket, which they only
                     # trust from this immutable server-side snapshot, never from
                     # request params.
-                    credentials = llm_router.get_deployment_credentials_with_provider(model_id=model_id)
+                    credentials = llm_router.registry().credentials(model_id, form="with_provider")
                     if credentials is not None:
                         data["_litellm_internal_model_credentials"] = cast(Dict, MappingProxyType(dict(credentials)))
                     else:
@@ -2518,7 +2954,7 @@ class _PROXY_LiteLLMManagedFiles(CustomLogger, BaseFileEndpoints):
                 # Continue with other files even if one fails
                 continue
 
-    def _get_content_type_from_file_object(self, file_object: Optional[Any]) -> str:
+    def _get_content_type_from_file_object(self, file_object: object) -> str:
         """
         Determine content type from file object.
 

enterprise/litellm_enterprise/proxy/hooks/managed_vector_stores.py

@@ -29,7 +29,7 @@ if TYPE_CHECKING:
     from litellm.proxy.utils import InternalUsageCache as _InternalUsageCache
     from litellm.proxy.utils import PrismaClient as _PrismaClient
 
-    Span = Union[_Span, Any]
+    Span = _Span
     InternalUsageCache = _InternalUsageCache
     PrismaClient = _PrismaClient
 else:
@@ -107,7 +107,7 @@ class _PROXY_LiteLLMManagedVectorStores(
         llm_router: Router,
         model: str,
         request_data: dict[str, object] | VectorStoreCreateOptionalRequestParams,
-        litellm_parent_otel_span: Span,
+        litellm_parent_otel_span: Span | None,
     ) -> VectorStoreCreateResponse:
         """
         Create a vector store for a specific model.
@@ -134,7 +134,7 @@ class _PROXY_LiteLLMManagedVectorStores(
         create_request: VectorStoreCreateOptionalRequestParams,
         llm_router: Router,
         target_model_names_list: List[str],
-        litellm_parent_otel_span: Span,
+        litellm_parent_otel_span: Span | None,
         user_api_key_dict: UserAPIKeyAuth,
     ) -> VectorStoreCreateResponse:
         """

enterprise/litellm_enterprise/proxy/liteadmin.py

@@ -18,8 +18,8 @@ from fastapi.responses import HTMLResponse, RedirectResponse, Response
 from pydantic import ConfigDict, Field, SecretStr, TypeAdapter, ValidationError
 
 from litellm.llms.custom_httpx.http_handler import get_async_httpx_client
-from litellm.proxy._experimental.mcp_server.oauth_utils import get_request_base_url
 from litellm.proxy._types import LiteLLM_UserTable, LitellmUserRoles, UserAPIKeyAuth
+from litellm.proxy.mcp_server.oauth_utils import get_request_base_url
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.proxy.auth.auth_checks import UserNotFoundError
 
@@ -122,7 +122,7 @@ def _cookie_name(token: str) -> str:
 
 
 async def _session_user(request: Request) -> str | None:
-    from litellm.proxy._experimental.mcp_server.byok_oauth_endpoints import (
+    from litellm.proxy.mcp_server.byok_oauth_endpoints import (
         get_authenticated_browser_user_id,
     )
 

enterprise/litellm_enterprise/proxy/management_endpoints/project_endpoints.py

@@ -15,7 +15,11 @@ from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Final, cast
 
 from fastapi import APIRouter, Depends, HTTPException, Request
-from pydantic import TypeAdapter
+from litellm_enterprise.proxy.management_endpoints.project_model_quota import (
+    describe_missing_quota,
+    missing_project_model_quota,
+)
+from pydantic import TypeAdapter, ValidationError
 
 from litellm._logging import verbose_proxy_logger
 from litellm._uuid import uuid
@@ -27,6 +31,15 @@ from litellm.proxy.management.teams.dependencies import get_team_access
 from litellm.proxy.management.users.service import org_admin_org_ids
 from litellm.proxy.management_endpoints.common_utils import _set_object_metadata_field
 from litellm.proxy.management_endpoints.team_admin_field_permissions import team_admin_may_manage_projects
+from litellm.proxy.management_helpers.model_quota_policy import (
+    ENFORCE_PROJECT_MODEL_QUOTA_SETTING,
+    REQUIRE_BOTH_LIMITS,
+    ModelQuotaPolicy,
+    RouteOf,
+    no_route,
+    parse_model_quota_setting,
+    router_model_routes,
+)
 from litellm.proxy.management_helpers.utils import (
     management_endpoint_wrapper,
 )
@@ -36,6 +49,7 @@ from litellm.repositories.budget_repository import BudgetRepository
 from litellm.repositories.object_permission_repository import ObjectPermissionRepository
 from litellm.repositories.prisma_protocols import TableActions
 from litellm.repositories.project_repository import ProjectRepository
+from litellm.repositories.table_repositories import PrismaTableRepository
 from litellm.repositories.team_repository import TeamRepository
 from litellm.repositories.user_repository import UserRepository
 from litellm.repositories.verification_token_repository import VerificationTokenRepository
@@ -73,6 +87,16 @@ def _verification_token_table(
     return VerificationTokenRepository(prisma_client).table
 
 
+class _WriterVerificationTokenRepository(PrismaTableRepository["prisma_models.LiteLLM_VerificationToken"]):
+    table_name = "litellm_verificationtoken"
+
+
+def _writer_verification_token_table(
+    prisma_client: PrismaClient,
+) -> TableActions["prisma_models.LiteLLM_VerificationToken"]:
+    return _WriterVerificationTokenRepository(prisma_client, use_writer=True).table
+
+
 def _budget_table(prisma_client: PrismaClient) -> TableActions["prisma_models.LiteLLM_BudgetTable"]:
     return BudgetRepository(prisma_client).table
 
@@ -265,28 +289,8 @@ def _check_team_project_limits(
         )
 
 
-def _project_models_missing_positive_quota(
-    models: list[str] | None,
-    rpm_limits: Mapping[str, object] | None,
-    tpm_limits: Mapping[str, object] | None,
-) -> list[str]:
-    """Return the models that lack a positive `rpm` AND `tpm` quota.
-
-    A valid quota is a positive integer; null, zero, and negative are rejected
-    because downstream rate limiters treat a non-positive limit as immediately
-    exhausted (every request blocked).
-    """
-
-    def _is_positive(value: object) -> bool:
-        return isinstance(value, int) and not isinstance(value, bool) and value > 0
-
-    rpm = rpm_limits or {}
-    tpm = tpm_limits or {}
-    return [model for model in (models or []) if not _is_positive(rpm.get(model)) or not _is_positive(tpm.get(model))]
-
-
 def _router_access_group_names(llm_router: "Router | None") -> frozenset[str]:
-    return frozenset(llm_router.get_model_access_groups()) if llm_router is not None else frozenset()
+    return frozenset(llm_router.registry().access_groups()) if llm_router is not None else frozenset()
 
 
 def _project_models_expanding_at_request_time(
@@ -311,64 +315,79 @@ def _raise_on_project_models_expanding_at_request_time(
     raise HTTPException(
         status_code=400,
         detail={
-            "error": f"models {list(expanding)} expand to multiple models at request time, so a per-model rpm/tpm quota cannot be enforced for them while 'enforce_project_model_quota' is enabled. List concrete model names instead."
+            "error": f"models {list(expanding)} expand to multiple models at request time, so a per-model rpm/tpm quota cannot be enforced for them while '{ENFORCE_PROJECT_MODEL_QUOTA_SETTING}' is enabled. List concrete model names instead."
         },
     )
 
 
+def _project_model_quota_policy(general_settings: Mapping[str, object]) -> ModelQuotaPolicy | None:
+    try:
+        return parse_model_quota_setting(general_settings.get(ENFORCE_PROJECT_MODEL_QUOTA_SETTING))
+    except ValidationError as e:
+        raise HTTPException(
+            status_code=500,
+            detail={"error": f"general_settings.{ENFORCE_PROJECT_MODEL_QUOTA_SETTING} is invalid: {e}"},
+        ) from e
+
+
 def _raise_on_missing_project_model_quota(
-    data: NewProjectRequest | UpdateProjectRequest, access_group_names: frozenset[str] = frozenset()
+    data: NewProjectRequest | UpdateProjectRequest,
+    access_group_names: frozenset[str] = frozenset(),
+    policy: ModelQuotaPolicy = REQUIRE_BOTH_LIMITS,
+    route_of: RouteOf = no_route,
 ) -> None:
-    """Require a positive `rpm`/`tpm` quota for every model on project CREATE.
+    """Require the positive `rpm`/`tpm` quotas the policy asks of each model on project CREATE.
 
     `model_rpm_limit`/`model_tpm_limit` are relocated into `metadata` by the request
     model's `set_model_info` validator, so they are read from there.
-
-    Only invoked when `general_settings.enforce_project_model_quota` is enabled
-    (default off), so it is opt-in and does not change behavior for existing users.
     """
     _raise_on_project_models_expanding_at_request_time(data.models, access_group_names)
     metadata = data.metadata or {}
-    missing = _project_models_missing_positive_quota(
-        data.models, metadata.get("model_rpm_limit"), metadata.get("model_tpm_limit")
+    missing: Final = missing_project_model_quota(
+        data.models or (), metadata.get("model_rpm_limit"), metadata.get("model_tpm_limit"), policy, route_of
     )
     if not missing:
         return
     raise HTTPException(
         status_code=400,
         detail={
-            "error": f"models {missing} added to project without a positive rpm/tpm quota. Set a positive model_rpm_limit and model_tpm_limit for each model."
+            "error": f"models added to project without a positive rpm/tpm quota: {describe_missing_quota(missing)}. Set a positive value for each listed limit."
         },
     )
 
 
 def _raise_on_missing_project_model_quota_on_update(
-    data: UpdateProjectRequest, existing_project: object, access_group_names: frozenset[str] = frozenset()
+    data: UpdateProjectRequest,
+    existing_project: object,
+    access_group_names: frozenset[str] = frozenset(),
+    policy: ModelQuotaPolicy = REQUIRE_BOTH_LIMITS,
+    route_of: RouteOf = no_route,
 ) -> None:
-    """Require a positive `rpm`/`tpm` quota over the RESULTING state on project UPDATE.
+    """Require the positive `rpm`/`tpm` quotas the policy asks of each model over the RESULTING state on project UPDATE.
 
     `/project/update` replaces `models` and `metadata` when they are provided, so the
     check runs on what the project WILL look like: a partial update that doesn't touch
     models/quota keeps the existing values, while one that adds a model or clears a
-    model's quota must leave every resulting model with a positive limit.
-
-    Only invoked when `general_settings.enforce_project_model_quota` is enabled
-    (default off), so it is opt-in and does not change behavior for existing users.
+    model's quota must leave every resulting model with the limits it needs.
     """
     resulting_models = data.models if data.models is not None else (getattr(existing_project, "models", None) or [])
     resulting_metadata = (
         data.metadata if data.metadata is not None else (getattr(existing_project, "metadata", None) or {})
     )
     _raise_on_project_models_expanding_at_request_time(resulting_models, access_group_names)
-    missing = _project_models_missing_positive_quota(
-        resulting_models, resulting_metadata.get("model_rpm_limit"), resulting_metadata.get("model_tpm_limit")
+    missing: Final = missing_project_model_quota(
+        resulting_models,
+        resulting_metadata.get("model_rpm_limit"),
+        resulting_metadata.get("model_tpm_limit"),
+        policy,
+        route_of,
     )
     if not missing:
         return
     raise HTTPException(
         status_code=400,
         detail={
-            "error": f"models {missing} would be left on the project without a positive rpm/tpm quota. Set a positive model_rpm_limit and model_tpm_limit for each model."
+            "error": f"models would be left on the project without a positive rpm/tpm quota: {describe_missing_quota(missing)}. Set a positive value for each listed limit."
         },
     )
 
@@ -519,6 +538,7 @@ async def new_project(
     """
     from litellm.proxy.proxy_server import (
         general_settings,
+        general_settings_view,
         litellm_proxy_admin_name,
         llm_router,
         premium_user,
@@ -567,10 +587,6 @@ async def new_project(
             data=data,
         )
 
-        # Opt-in (default off): require rpm/tpm for every model added to the project.
-        if general_settings.get("enforce_project_model_quota", False):
-            _raise_on_missing_project_model_quota(data, _router_access_group_names(llm_router))
-
         # Check if user has permission to create projects for this team
         # only team admins can create projects for their team
         has_permission = await _check_user_permission_for_project(
@@ -589,6 +605,15 @@ async def new_project(
                 },
             )
 
+        quota_policy: Final = _project_model_quota_policy(general_settings_view())
+        if quota_policy is not None:
+            _raise_on_missing_project_model_quota(
+                data,
+                _router_access_group_names(llm_router),
+                quota_policy,
+                router_model_routes(llm_router, data.team_id),
+            )
+
         # Generate project_id if not provided
         if data.project_id is None:
             data.project_id = str(uuid.uuid4())
@@ -710,6 +735,7 @@ async def update_project(
     """
     from litellm.proxy.proxy_server import (
         general_settings,
+        general_settings_view,
         litellm_proxy_admin_name,
         llm_router,
         premium_user,
@@ -817,10 +843,14 @@ async def update_project(
                 data=data,
             )
 
-        # Opt-in (default off): require rpm/tpm for every model the update would leave on the project.
-        if general_settings.get("enforce_project_model_quota", False):
+        quota_policy: Final = _project_model_quota_policy(general_settings_view())
+        if quota_policy is not None:
             _raise_on_missing_project_model_quota_on_update(
-                data, existing_project, _router_access_group_names(llm_router)
+                data,
+                existing_project,
+                _router_access_group_names(llm_router),
+                quota_policy,
+                router_model_routes(llm_router, target_team_id),
             )
 
         # Prepare update data
@@ -849,7 +879,7 @@ async def update_project(
                 else None
             )
             if current_project is not None and data.team_id != current_project.team_id:
-                mismatched_key_count: Final = await prisma_client.writer_db.litellm_verificationtoken.count(
+                mismatched_key_count: Final = await _writer_verification_token_table(prisma_client).count(
                     where={
                         "project_id": data.project_id,
                         "OR": [{"team_id": {"not": data.team_id}}, {"team_id": None}],

enterprise/litellm_enterprise/proxy/management_endpoints/project_model_quota.py

@@ -0,0 +1,49 @@
+from collections.abc import Mapping, Sequence
+from dataclasses import dataclass
+from types import MappingProxyType
+from typing import Final
+
+from litellm.proxy.management_helpers.model_quota_policy import (
+    ModelQuotaPolicy,
+    QuotaLimit,
+    RouteOf,
+    required_limits,
+    unmet_limits,
+)
+
+_NO_LIMITS_SET: Final[Mapping[str, object]] = MappingProxyType({})
+
+
+@dataclass(frozen=True, slots=True)
+class MissingModelQuota:
+    model: str
+    limits: tuple[QuotaLimit, ...]
+
+
+def _limits_set_for(
+    model: str, rpm_set: Mapping[str, object], tpm_set: Mapping[str, object]
+) -> Mapping[QuotaLimit, object]:
+    return MappingProxyType({"rpm": rpm_set.get(model), "tpm": tpm_set.get(model)})
+
+
+def missing_project_model_quota(
+    models: Sequence[str],
+    rpm_limits: Mapping[str, object] | None,
+    tpm_limits: Mapping[str, object] | None,
+    policy: ModelQuotaPolicy,
+    route_of: RouteOf,
+) -> tuple[MissingModelQuota, ...]:
+    rpm_set: Final = rpm_limits or _NO_LIMITS_SET
+    tpm_set: Final = tpm_limits or _NO_LIMITS_SET
+    gaps: Final = tuple(
+        MissingModelQuota(
+            model,
+            unmet_limits(required_limits(policy, model, route_of(model)), _limits_set_for(model, rpm_set, tpm_set)),
+        )
+        for model in models
+    )
+    return tuple(gap for gap in gaps if gap.limits)
+
+
+def describe_missing_quota(missing: Sequence[MissingModelQuota]) -> str:
+    return ", ".join(f"{gap.model} ({' and '.join(f'model_{limit}_limit' for limit in gap.limits)})" for gap in missing)

enterprise/litellm_enterprise/proxy/vector_stores/endpoints.py

@@ -14,6 +14,7 @@ from collections.abc import Mapping
 from typing import TYPE_CHECKING, Final, List, Optional, Protocol
 
 from fastapi import APIRouter, Depends, HTTPException
+from pydantic import TypeAdapter
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -37,6 +38,7 @@ if TYPE_CHECKING:
     from litellm.proxy.utils import PrismaClient
 
 router = APIRouter()
+_DECODED_JSON: Final = TypeAdapter(object)
 
 
 class ManagedVectorStoreRow(Protocol):
@@ -348,11 +350,11 @@ async def update_vector_store(
         raise HTTPException(status_code=500, detail="Database not connected")
 
     try:
-        update_data = data.model_dump(exclude_unset=True)
-        vector_store_id: Final[str] = update_data.pop("vector_store_id")
+        update_data = data.model_dump(exclude_unset=True, exclude={"vector_store_id"})
+        vector_store_id: Final = data.vector_store_id
         if update_data.get("vector_store_metadata") is not None:
             update_data["vector_store_metadata"] = safe_dumps(
-                update_data["vector_store_metadata"]
+                _DECODED_JSON.validate_python(update_data["vector_store_metadata"])
             )
 
         updated = await managed_vector_store_table(prisma_client).update(

enterprise/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "litellm-enterprise"
-version = "0.1.77"
+version = "0.1.81"
 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 = "0.1.77"
+version = "0.1.81"
 version_files = [
     "pyproject.toml:^version",
     "../pyproject.toml:litellm-enterprise==",

examples/bigmodel.md

@@ -0,0 +1,63 @@
+# BigModel (Zhipu AI)
+
+Use `bigmodel/` for the domestic BigModel platform and `zai/` for international Z.AI. Each platform uses its own credentials and default URL
+
+```yaml
+model_list:
+  - model_name: glm
+    litellm_params:
+      model: bigmodel/glm-5.3
+      api_key: os.environ/BIGMODEL_API_KEY
+      api_base: https://open.bigmodel.cn
+```
+
+One model configuration supports all three native APIs. The provider selects the upstream endpoint from the requested API
+
+| Proxy endpoint | BigModel endpoint |
+| --- | --- |
+| `/v1/chat/completions` | `/api/paas/v4/chat/completions` |
+| `/v1/responses` | `/api/v1/responses` |
+| `/v1/messages` | `/api/anthropic/v1/messages` |
+
+`api_base` is optional and defaults to `https://open.bigmodel.cn`. Set `BIGMODEL_API_BASE` to override the default, or pass an explicit `api_base` to override the environment value. Existing API-specific base URLs and complete endpoint URLs are normalized, preserving a custom proxy's path prefix
+
+A GLM Coding Plan key uses `https://open.bigmodel.cn/api/coding/paas/v4` as its `api_base`, since Coding Plan and standard keys do not work on each other's endpoints. Chat Completions then calls the Coding Plan endpoint, and Responses and Messages keep the endpoints in the table above
+
+The Admin UI lists BigModel (Zhipu AI), displays its official logo, and provides a required password API Key field and an optional API Base field with the default URL. Authentication uses an explicit `api_key` or `BIGMODEL_API_KEY`
+
+```python
+import litellm
+
+chat = litellm.completion(
+    model="bigmodel/glm-5.3",
+    messages=[{"role": "user", "content": "Hello"}],
+    thinking={"type": "enabled"},
+    reasoning_effort="low",
+)
+
+response = litellm.responses(
+    model="bigmodel/glm-5.3",
+    input="Hello",
+    reasoning={"effort": "low"},
+    store=True,
+)
+
+async def messages():
+    return await litellm.anthropic.messages.acreate(
+        model="bigmodel/glm-5.3",
+        max_tokens=1024,
+        messages=[{"role": "user", "content": "Hello"}],
+        thinking={"type": "adaptive"},
+        output_config={"effort": "low"},
+    )
+```
+
+Each API retains its native response format and streaming events. The native Messages SDK entry point is asynchronous, as is the proxy's `/v1/messages` route
+
+GLM-5.3, GLM-5.3-Flash, and GLM-5.3-FlashX use always-on thinking. GLM-5.3 supports `low`, `high`, and `max` reasoning effort. Messages preserves native `thinking`, `output_config`, and `reasoning_effort` parameters
+
+Set `store=True` when creating a Responses request if you need continuation, retrieval, or deletion. Use the response ID returned by LiteLLM. BigModel documents a seven-day retention period and does not support cancel, compact, native WebSocket, or file_search
+
+The model catalog and packaged backup include GLM-5.3, GLM-5.3-Flash, and GLM-5.3-FlashX. Prices are expressed in USD/token using a static 7.0 CNY/USD conversion, with the source and verification date in each entry's metadata. Use existing `litellm_params` cost fields for contract-specific deployment prices
+
+API references: [Chat Completions](https://docs.bigmodel.cn/cn/guide/develop/openai/introduction), [Responses](https://docs.bigmodel.cn/cn/guide/develop/responses/introduction), [Messages](https://docs.bigmodel.cn/cn/guide/develop/claude/introduction), and [pricing](https://docs.bigmodel.cn/cn/guide/start/pricing), checked on 2026-10-27

gateway/Dockerfile

@@ -1,5 +1,5 @@
-ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
-ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:1d95114038f76513a9ace6fca107d5582b08c65981f81f61cb56bf7fd2ef216d
+ARG LITELLM_BUILD_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
+ARG LITELLM_RUNTIME_IMAGE=cgr.dev/chainguard/wolfi-base@sha256:360896efb0753a5701c48f860e17130c205eae5a264797aa3e7ea126ee2923a7
 ARG UV_IMAGE=ghcr.io/astral-sh/uv:0.11.7@sha256:240fb85ab0f263ef12f492d8476aa3a2e4e1e333f7d67fbdd923d00a506a516a
 # Checksum from https://www.pgbouncer.org/downloads/ (the Wolfi repo only carries 1.24.x)
 ARG PGBOUNCER_VERSION=1.26.0
@@ -20,6 +20,34 @@ RUN curl -fsSL -o pgbouncer.tar.gz "https://www.pgbouncer.org/downloads/files/${
     make -j"$(nproc)" pgbouncer && \
     install -m 0755 pgbouncer /usr/local/bin/pgbouncer
 
+# ---------- Rust extension ----------
+# Copies only the native build's inputs, so the fat-LTO compile stays cached across Python-only changes
+FROM $LITELLM_BUILD_IMAGE AS rust-native
+
+WORKDIR /app
+USER root
+
+COPY --from=uvbin /uv /uvx /usr/local/bin/
+
+RUN for i in 1 2 3; do \
+      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev && break; \
+      [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
+      sleep 5; \
+    done
+
+ENV UV_PYTHON_DOWNLOADS=0
+
+COPY pyproject.toml README.md rust-toolchain.toml ./
+COPY .cargo ./.cargo
+COPY litellm-rust ./litellm-rust
+COPY model_prices_and_context_window.json ./
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    mkdir -p litellm/rust_bridge /wheels && \
+    touch litellm/__init__.py litellm/rust_bridge/__init__.py && \
+    uvx --python python3.13 --from maturin==1.15.0 --with puccinialin==0.1.14 \
+        python -c "import maturin; maturin.build_wheel('/wheels')"
+
 # ---------- Builder ----------
 FROM $LITELLM_BUILD_IMAGE AS builder
 
@@ -32,7 +60,7 @@ COPY --from=uvbin /uv /uvx /usr/local/bin/
 # instead of nodeenv downloading one whose dynamic deps may not be in Wolfi
 # (e.g. Node 26.2.0 needs libatomic). Retry for transient apk.cgr.dev flakes.
 RUN for i in 1 2 3; do \
-      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl openssl-dev libsndfile nodejs npm && break; \
+      apk add --no-cache bash gcc python-3.13 python-3.13-dev openssl libssl3 openssl-dev libsndfile nodejs npm && break; \
       [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
       sleep 5; \
     done
@@ -66,11 +94,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra bedrock-realtime \
         --python python3.13
 
-# Stage 2 — copy source and install the project + workspace members.
+# Stage 2 — copy source and install the workspace members.
 COPY . .
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --no-default-groups --no-editable \
+    uv sync --frozen --no-install-project --no-default-groups --no-editable \
         --extra proxy \
         --extra proxy-runtime \
         --group admin-mcp \
@@ -79,6 +107,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         --extra bedrock-realtime \
         --python python3.13
 
+RUN --mount=type=bind,from=rust-native,source=/wheels,target=/wheels \
+    uv pip install --python /app/.venv/bin/python --no-deps /wheels/litellm-*.whl
+
 # PYTHONPATH=/app makes the source tree shadow the installed package, so the
 # compiled Rust extension must live next to the source or it is never imported.
 RUN cp "$(python -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])')"/litellm/rust_bridge/_native*.so litellm/rust_bridge/
@@ -95,7 +126,7 @@ FROM $LITELLM_RUNTIME_IMAGE AS runtime
 USER root
 
 RUN for i in 1 2 3; do \
-      apk add --no-cache bash openssl tzdata python-3.13 libsndfile libatomic libevent && break; \
+      apk add --no-cache bash openssl libssl3 tzdata python-3.13 libsndfile libatomic libevent && break; \
       [ $i = 3 ] && { echo "apk add failed after 3 retries" >&2; exit 1; }; \
       sleep 5; \
     done
@@ -110,7 +141,13 @@ ENV HOME=/home/nonroot \
     PYTHONUNBUFFERED=1 \
     PRISMA_BINARY_CACHE_DIR=/opt/prisma/binaries
 
-COPY --from=builder --chown=nonroot:nonroot /app /app
+COPY --from=builder --chown=nonroot:nonroot /app/.venv /app/.venv
+COPY --from=builder --chown=nonroot:nonroot /app/litellm /app/litellm
+COPY --from=builder --chown=nonroot:nonroot /app/enterprise /app/enterprise
+COPY --from=builder --chown=nonroot:nonroot /app/gateway /app/gateway
+COPY --from=builder --chown=nonroot:nonroot /app/litellm_cli /app/litellm_cli
+COPY --from=builder --chown=nonroot:nonroot /app/docker/component_entrypoint.sh /app/docker/component_entrypoint.sh
+COPY --from=builder --chown=nonroot:nonroot /app/schema.prisma /app/schema.prisma
 COPY --from=builder /opt/prisma /opt/prisma
 COPY --from=pgbouncer-builder /usr/local/bin/pgbouncer /usr/local/bin/pgbouncer
 
@@ -118,7 +155,7 @@ RUN find /app/.venv -type f -path "*/tornado/test/*" -delete && \
     find /app/.venv -type d -path "*/tornado/test" -delete && \
     chmod -R a+rX /opt/prisma && \
     python -c "from prisma.client import BINARY_PATHS; paths = list(BINARY_PATHS.query_engine.values()); assert paths and all(p.startswith('/opt/prisma/') for p in paths), paths" && \
-    python -c "import litellm; from litellm.rust_bridge.loader import native_bridge_available; assert litellm.__file__ == '/app/litellm/__init__.py', litellm.__file__; assert native_bridge_available()"
+    python -c "import litellm, litellm.rust_bridge._native; assert litellm.__file__ == '/app/litellm/__init__.py', litellm.__file__"
 
 USER nonroot
 

gateway/launch.py

@@ -8,6 +8,11 @@ handed the loopback URL it listens on. A pre-existing ``DATABASE_URL`` wins in
 wins under token auth too, so exporting it here is enough for every worker to
 pick the pooled URL up unchanged.
 
+LiteLLM imports wait for ``main``: every uvicorn worker imports this module again
+before it can answer its first 5 s health ping, so a slow host would kill each one.
+Workers then load the gateway app through ``worker_app`` without garbage collection
+pauses, which hold the GIL long enough to miss a ping.
+
 Run with:
     python -m gateway.launch --workers 4 --host 0.0.0.0 --port 4000
 """
@@ -15,19 +20,19 @@ Run with:
 import os
 import sys
 from collections.abc import Callable, Mapping, Sequence
-from typing import Final
+from typing import TYPE_CHECKING, Final
 
 from uvicorn.main import main as uvicorn_main
 
-from litellm.proxy.db.db_url_settings import DatabaseURLSettings
-from litellm.proxy.db.pgbouncer import (
-    PgBouncerError,
-    PgBouncerSettings,
-    export_pooled_database_url,
-    start_in_container_pgbouncer,
-)
+from litellm_cli.main import without_gc_pauses
 
-GATEWAY_APP: Final = "gateway.main:app"
+if TYPE_CHECKING:
+    from fastapi import FastAPI
+
+    from litellm.proxy.db.db_url_settings import DatabaseURLSettings
+    from litellm.proxy.db.pgbouncer import PgBouncerError, PgBouncerSettings
+
+GATEWAY_APP: Final = "gateway.launch:worker_app"
 KEEPALIVE_FLAG: Final = "--timeout-keep-alive"
 
 
@@ -35,21 +40,23 @@ def uvicorn_argv(argv: Sequence[str], environ: Mapping[str, str]) -> tuple[str,
     """Honor ``KEEPALIVE_TIMEOUT`` like ``proxy_cli.py`` does, unless the flag was passed explicitly."""
     keepalive: Final = environ.get("KEEPALIVE_TIMEOUT")
     if keepalive is None or any(arg == KEEPALIVE_FLAG or arg.startswith(f"{KEEPALIVE_FLAG}=") for arg in argv):
-        return (GATEWAY_APP, *argv)
-    return (GATEWAY_APP, *argv, KEEPALIVE_FLAG, keepalive)
+        return (GATEWAY_APP, "--factory", *argv)
+    return (GATEWAY_APP, "--factory", *argv, KEEPALIVE_FLAG, keepalive)
 
 
 def pool_database_url(
-    settings: DatabaseURLSettings,
-    pgbouncer: PgBouncerSettings,
+    settings: "DatabaseURLSettings",
+    pgbouncer: "PgBouncerSettings",
     environ: Mapping[str, str],
-) -> str | PgBouncerError | None:
+) -> "str | PgBouncerError | None":
     """Start the in-container PgBouncer and return its loopback URL, or None when ``pgbouncer.enabled`` is off.
 
     The upstream URL is whatever ``apply_to_env`` assembled from the discrete
     ``DATABASE_*`` vars (or an operator-pinned ``DATABASE_URL``). Under token
     auth the pooler mints and renews the upstream token itself.
     """
+    from litellm.proxy.db.pgbouncer import PgBouncerError, start_in_container_pgbouncer
+
     if not pgbouncer.enabled:
         return None
     upstream_url: Final = environ.get("DATABASE_URL")
@@ -58,11 +65,20 @@ def pool_database_url(
     return start_in_container_pgbouncer(pgbouncer, upstream_url, token_auth=settings.token_auth())
 
 
+def worker_app() -> "FastAPI":
+    with without_gc_pauses():
+        from gateway.main import app
+    return app
+
+
 def _serve(argv: Sequence[str]) -> None:
     uvicorn_main(tuple(argv), prog_name="uvicorn")
 
 
 def main(argv: Sequence[str], serve: Callable[[Sequence[str]], None] = _serve) -> None:
+    from litellm.proxy.db.db_url_settings import DatabaseURLSettings
+    from litellm.proxy.db.pgbouncer import PgBouncerError, PgBouncerSettings, export_pooled_database_url
+
     settings: Final = DatabaseURLSettings.from_env()
     settings.apply_to_env()
     pooled_url: Final = pool_database_url(settings, PgBouncerSettings(), os.environ)

gateway/routes/allowlist.py

@@ -1,7 +1,7 @@
 """Path allowlist for the gateway component.
 
 The gateway exposes the LLM data-plane surface: chat/completions, embeddings,
-audio, batches, files, fine-tuning, rerank, decisions, ocr, rag, video, search, image,
+audio, batches, files, fine-tuning, rerank, decisions, extraction, ocr, rag, video, search, image,
 responses, vector stores, passthrough providers, realtime websockets, MCP
 tool-call endpoints, and operational endpoints (/health, /metrics, and the
 /debug/memory/summary read of the serving worker's RSS).
@@ -64,6 +64,8 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = (
     "/decisions",
     "/v1/systemone",
     "/systemone",
+    "/v1/extraction",
+    "/extraction",
     "/v1/ocr",
     "/ocr",
     "/v1/rag/",
@@ -140,6 +142,7 @@ GATEWAY_EXACT_PATHS: frozenset[str] = frozenset(
         "/test",
         "/debug/memory/summary",
         "/api/event_logging/batch",
+        "/eu_jev/v1/systemone",
     }
 )
 

helm/litellm-helm/Chart.lock

@@ -5,8 +5,5 @@ dependencies:
 - name: redis
   repository: oci://registry-1.docker.io/bitnamicharts
   version: 18.19.1
-- name: lens
-  repository: oci://ghcr.io/berriai/charts
-  version: 0.1.0-dev.0
-digest: sha256:16b92c3d7fc74632e8aa11e602c136d273c774ed39b2fe970b80273e0570898b
-generated: '2026-10-09T04:18:10.643228000Z'
+digest: sha256:38962e231f6596b93f82a8412bbe4cf5de696caecf5775dfbbd163383eb1c009
+generated: '2026-10-30T02:13:11.505726000Z'

helm/litellm-helm/Chart.yaml

@@ -18,7 +18,7 @@ type: application
 # This is the chart version. This version number should be incremented each time you make changes
 # to the chart and its templates, including the app version.
 # Versions are expected to follow Semantic Versioning (https://semver.org/)
-version: 1.1.3
+version: 1.1.4
 
 # This is the version number of the application being deployed. This version number should be
 # incremented each time you make changes to the application. Versions are not expected to
@@ -39,6 +39,3 @@ dependencies:
     version: "18.19.1"
     repository: oci://registry-1.docker.io/bitnamicharts
     condition: redis.enabled
-  - name: lens
-    version: 0.1.0-dev.0
-    repository: oci://ghcr.io/berriai/charts

helm/litellm-helm/README.md

@@ -56,8 +56,8 @@ If `db.useStackgresOperator` is used (not yet implemented):
 
 | `billingMetrics.enabled`    | Enable enterprise billable-request metering. Requires an enterprise license.                                                                                                                                                                  | `false`                   |
 | `billingMetrics.endpoint`   | Collector that the billable-request counter is pushed to.                                                                                                                                                                                    | `https://telemetry.litellm.ai` |
-| `billingMetrics.secretName` | Name of an existing Secret holding the mTLS client certificate, under the keys `tls.crt` and `tls.key`.                                                                                                                                       | `litellm-billing-metrics-mtls` |
-| `billingMetrics.caSecretName` | Name of an existing Secret holding a CA bundle under the key `ca.crt`. Only needed for a private or test collector whose server certificate is not on the public web PKI.                                                                    | `""`                      |
+| `billingMetrics.tokenUrl`   | Cognito OAuth2 client-credentials token endpoint.                                                                                                                                                                                               | `https://litellm-prod.auth.us-west-1.amazoncognito.com/oauth2/token` |
+| `billingMetrics.secretName` | Name of an existing Secret with `client-id` and `client-secret` keys.                                                                                                                                                                           | `litellm-billing-metrics` |
 | `billingMetrics.exportIntervalMs` | How often the counter is pushed, in milliseconds. The proxy defaults to `60000` when unset.                                                                                                                                             | `""`                      |
 
 #### Example `proxy_config` ConfigMap from values (default):
@@ -70,11 +70,18 @@ proxyConfigMap:
 proxy_config:
   general_settings:
     master_key: os.environ/PROXY_MASTER_KEY
+  model_list: []
+```
+
+The chart ships no models, so add them under `proxy_config.model_list` or from the Admin UI with `store_model_in_db: true`
+
+```
+proxy_config:
   model_list:
-    - model_name: gpt-3.5-turbo
+    - model_name: gpt-5.6
       litellm_params:
-        model: gpt-3.5-turbo
-        api_key: eXaMpLeOnLy
+        model: openai/gpt-5.6
+        api_key: os.environ/OPENAI_API_KEY
 ```
 
 #### Example using existing `proxyConfigMap` instead of creating it:
@@ -102,10 +109,10 @@ type: Opaque
 
 #### Enterprise billable-request metering
 
-Enterprise licenses meter billable requests by pushing a counter to LiteLLM's collector over mutual TLS. The chart does not create the client certificate; it mounts one you already hold, read-only, so the private key is never exposed through the environment. Create the Secret under the name the chart expects, then turn the block on:
+Enterprise licenses meter billable requests by pushing a counter to LiteLLM's collector with Cognito bearer authentication. Create a Secret with `client-id` and `client-secret` keys, then enable the chart values:
 
 ```
-kubectl create secret tls litellm-billing-metrics-mtls --cert=client.crt --key=client.key
+kubectl create secret generic litellm-billing-metrics --from-literal=client-id=... --from-literal=client-secret=...
 ```
 
 ```
@@ -113,7 +120,8 @@ billingMetrics:
   enabled: true
 ```
 
-Set `billingMetrics.caSecretName` only when the collector is a private or test one whose server certificate is not on the public web PKI; the production collector needs no CA override. The chart fails the render rather than deploying a proxy that silently never exports, so a missing `secretName` or an emptied `endpoint` surfaces at `helm install` time.
+The endpoint and token URL must use HTTPS. The chart requires both URLs and
+the Secret name when billing metrics are enabled
 
 ### Database Settings
 
@@ -126,6 +134,11 @@ Set `billingMetrics.caSecretName` only when the collector is a private or test o
 | `db.secret.name`          | If `db.useExisting` is `true`, the name of the Kubernetes Secret that contains credentials.                                                                                                                                                                                               | `postgres`                                                                                 |
 | `db.secret.usernameKey`   | If `db.useExisting` is `true`, the name of the key within the Kubernetes Secret that holds the username for authenticating with the Postgres instance.                                                                                                                                    | `username`                                                                                 |
 | `db.secret.passwordKey`   | If `db.useExisting` is `true`, the name of the key within the Kubernetes Secret that holds the password associates with the above user.                                                                                                                                                   | `password`                                                                                 |
+| `db.username`             | If `db.useExisting` is `true`, the database user as a plain value. When set, `db.secret.usernameKey` is not read.                                                                                                                                                                        | `""`                                                                                       |
+| `db.port`                 | If `db.useExisting` is `true` with token auth, the Postgres port used to build the connection URL. Password auth takes the port from `db.url`.                                                                                                                                           | `5432`                                                                                     |
+| `db.schema`               | With token auth, the Postgres schema holding the LiteLLM tables, rendered as `DATABASE_SCHEMA` when the tables are not in `public`. Ignored with password auth, where the schema stays in `db.url` as `?schema=`.                                                                        | `""`                                                                                       |
+| `db.useIAMAuth`           | If `db.useExisting` is `true` and `db.deployStandalone` is `false`, authenticate with an AWS RDS IAM token instead of a password. See [Token auth](#token-auth-to-an-existing-postgres-rds-iam-or-microsoft-entra-id).                                                                  | `false`                                                                                    |
+| `db.useAzureEntraAuth`    | If `db.useExisting` is `true` and `db.deployStandalone` is `false`, authenticate with a Microsoft Entra ID token instead of a password. Mutually exclusive with `db.useIAMAuth`.                                                                                                        | `false`                                                                                    |
 | `db.useStackgresOperator` | Not yet implemented.                                                                                                                                                                                                                                                                      | `false`                                                                                    |
 | `db.deployStandalone`     | Deploy a standalone, single instance deployment of Postgres, using the Bitnami postgresql chart. This is useful for getting started but doesn't provide HA or (by default) data backups.                                                                                                  | `true`                                                                                     |
 | `postgresql.*`            | If `db.deployStandalone` is `true`, configuration passed to the Bitnami postgresql chart. See the [Bitnami Documentation](https://github.com/bitnami/charts/tree/main/bitnami/postgresql) for full configuration details. See [values.yaml](./values.yaml) for the default configuration. | See [values.yaml](./values.yaml)                                                           |
@@ -156,6 +169,22 @@ data:
 type: Opaque
 ```
 
+#### Token auth to an existing Postgres (RDS IAM or Microsoft Entra ID)
+
+When the database accepts short-lived tokens instead of a password (AWS RDS or Aurora with IAM database authentication, or Azure Database for PostgreSQL with Microsoft Entra ID), there is no password to put in a Secret. Turn on `db.useIAMAuth` or `db.useAzureEntraAuth` and the chart stops referencing `db.secret.passwordKey` and `db.url`. It sets `IAM_TOKEN_DB_AUTH` (or `AZURE_POSTGRESQL_AUTH`) and emits `DATABASE_HOST`, `DATABASE_PORT`, `DATABASE_USERNAME` and `DATABASE_NAME` for the proxy, the migrations job and the collector sidecar. Each of them mints its own token at startup and builds `DATABASE_URL` from those parts, and the in-container pgbouncer (`db.connectionPool`) refreshes the token before it expires.
+
+```yaml
+db:
+  useExisting: true
+  deployStandalone: false
+  useIAMAuth: true
+  endpoint: litellm.cluster-abc123.us-west-1.rds.amazonaws.com
+  database: litellm
+  username: litellm_iam   # the role granted rds_iam; no db.secret needed
+```
+
+The pod needs cloud credentials for the token: on EKS use IRSA or Pod Identity on the chart's `serviceAccount` (and `migrationJob.serviceAccountName` when the Helm hook runs before the ServiceAccount exists), on AKS use workload identity. Set `AWS_RDS_REGION` under `envVars` when the RDS hostname is not a standard regional endpoint (the token must be signed for the region the database lives in). TLS options that would otherwise go in `db.url` are set as `DATABASE_SSLMODE` and `DATABASE_SSLROOTCERT` under `envVars`. Leaving `db.username` empty reads the user from `db.secret.usernameKey` as before.
+
 #### Examples for `environmentSecrets` and `environemntConfigMaps`
 
 ```yaml
@@ -228,6 +257,10 @@ itself.
 
 ## Connect Lens
 
-`lensWorker.mode` selects `disabled`, `bundled` or `external`. Bundled mode generates separate service and identity-signing credentials. External mode requires `lensWorker.gateway.secretName` and `lensWorker.serviceTokenSecret.name`; their keys default to `gateway-secret` and `service-token`. Provision distinct values matching the external Lens deployment
+This chart does not deploy Lens. Deploy Lens from [its own chart](https://github.com/BerriAI/lens/tree/main/helm/lens), then set `lens.enabled: true`, `lens.url`, `lens.publicUrl`, `lens.gatewaySecret.name` and `lens.serviceTokenSecret.name`. The secret keys default to `gateway-secret` and `service-token`. Set `lens.service.name` when that Lens runs in this release's namespace and the chart ingress should route `/lens-ingest` to it. The old `lensWorker` values are no longer read, and enabling them fails the render
 
-Follow the [Lens Helm connection guide](https://github.com/BerriAI/lens/blob/main/helm/lens/README.md#connect-a-gateway) for complete values, verification, GitOps credential requirements and existing-data migration. Source-chart installation requires a built Lens image until the first signed Lens release is published
+## Shutdown accounting
+
+`GRACEFUL_SHUTDOWN_TIMEOUT` defaults to 30 seconds. Pending success callbacks share the time remaining since SIGTERM or the first `/health/drain` call. Callback flushing reserves the existing 10-second teardown headroom for cancellation, tracked accounting tasks and persistence. Exhausting this budget can discard callbacks, which is logged
+
+Set `terminationGracePeriodSeconds` to at least `GRACEFUL_SHUTDOWN_TIMEOUT + 10`, allowing additional time for preStop hooks and database writes. This is not a hard shutdown bound or a guarantee of persistence. Database retries retain their own limits, and cancellation-resistant callbacks can remain alive. The existing preStop hook can also be followed by Uvicorn's separate connection drain, as described in `values.yaml`

helm/litellm-helm/templates/_helpers.tpl

@@ -50,53 +50,27 @@ app.kubernetes.io/name: {{ include "litellm.name" . }}
 app.kubernetes.io/instance: {{ .Release.Name }}
 {{- end }}
 
-{{/*
-Enterprise billable-request metering. The client certificate identifies the
-deployment to LiteLLM's collector, so it is mounted read-only from an existing
-Secret rather than passed through the environment.
-*/}}
-{{- define "litellm.billingMetrics.certDir" -}}/etc/litellm/billing-mtls{{- end -}}
-{{- define "litellm.billingMetrics.caDir" -}}/etc/litellm/billing-mtls-ca{{- end -}}
-
 {{- define "litellm.billingMetricsEnv" -}}
 - name: LITELLM_BILLING_METRICS_ENDPOINT
   value: {{ required "billingMetrics.endpoint is required when billingMetrics.enabled is true" .Values.billingMetrics.endpoint | quote }}
-- name: LITELLM_BILLING_METRICS_CLIENT_CERT
-  value: {{ printf "%s/tls.crt" (include "litellm.billingMetrics.certDir" .) | quote }}
-- name: LITELLM_BILLING_METRICS_CLIENT_KEY
-  value: {{ printf "%s/tls.key" (include "litellm.billingMetrics.certDir" .) | quote }}
-{{- if .Values.billingMetrics.caSecretName }}
-- name: LITELLM_BILLING_METRICS_CA_CERT
-  value: {{ printf "%s/ca.crt" (include "litellm.billingMetrics.caDir" .) | quote }}
-{{- end }}
+- name: LITELLM_BILLING_METRICS_TOKEN_URL
+  value: {{ required "billingMetrics.tokenUrl is required when billingMetrics.enabled is true" .Values.billingMetrics.tokenUrl | quote }}
+- name: LITELLM_BILLING_METRICS_CLIENT_ID
+  valueFrom:
+    secretKeyRef:
+      name: {{ required "billingMetrics.secretName is required when billingMetrics.enabled is true" .Values.billingMetrics.secretName }}
+      key: client-id
+- name: LITELLM_BILLING_METRICS_CLIENT_SECRET
+  valueFrom:
+    secretKeyRef:
+      name: {{ required "billingMetrics.secretName is required when billingMetrics.enabled is true" .Values.billingMetrics.secretName }}
+      key: client-secret
 {{- with .Values.billingMetrics.exportIntervalMs }}
 - name: LITELLM_BILLING_METRICS_EXPORT_INTERVAL_MS
   value: {{ . | quote }}
 {{- end }}
 {{- end -}}
 
-{{- define "litellm.billingMetricsVolumes" -}}
-- name: billing-metrics-mtls
-  secret:
-    secretName: {{ required "billingMetrics.secretName is required when billingMetrics.enabled is true (an existing Secret with tls.crt and tls.key)" .Values.billingMetrics.secretName }}
-{{- if .Values.billingMetrics.caSecretName }}
-- name: billing-metrics-mtls-ca
-  secret:
-    secretName: {{ .Values.billingMetrics.caSecretName }}
-{{- end }}
-{{- end -}}
-
-{{- define "litellm.billingMetricsVolumeMounts" -}}
-- name: billing-metrics-mtls
-  mountPath: {{ include "litellm.billingMetrics.certDir" . }}
-  readOnly: true
-{{- if .Values.billingMetrics.caSecretName }}
-- name: billing-metrics-mtls-ca
-  mountPath: {{ include "litellm.billingMetrics.caDir" . }}
-  readOnly: true
-{{- end }}
-{{- end -}}
-
 {{/*
 Create the name of the service account to use
 */}}
@@ -162,6 +136,77 @@ taken before the change, which by that point no longer exists.
 {{- end -}}
 {{- end -}}
 
+{{/*
+Whether the existing Postgres is reached with a short-lived token (AWS RDS IAM
+or Microsoft Entra ID) instead of a password.
+*/}}
+{{- define "litellm.db.tokenAuth" -}}
+{{- if or .Values.db.useIAMAuth .Values.db.useAzureEntraAuth -}}true{{- end -}}
+{{- end -}}
+
+{{- define "litellm.validateDbTokenAuth" -}}
+{{- if and .Values.db.useIAMAuth .Values.db.useAzureEntraAuth -}}
+{{- fail "db.useIAMAuth and db.useAzureEntraAuth are mutually exclusive: the database password can only come from one token source" -}}
+{{- end -}}
+{{- if and (include "litellm.db.tokenAuth" .) (or (not .Values.db.useExisting) .Values.db.deployStandalone) -}}
+{{- fail "db.useIAMAuth and db.useAzureEntraAuth need db.useExisting: true and db.deployStandalone: false (token auth targets an existing RDS or Azure Database for PostgreSQL; the bundled postgresql subchart only supports password auth)" -}}
+{{- end -}}
+{{- end -}}
+
+{{/*
+Database env for an existing Postgres, shared by the proxy container, the
+collector sidecar and the migrations Job. Under token auth every entrypoint
+mints its own token and assembles DATABASE_URL itself, so db.url is left out.
+*/}}
+{{- define "litellm.existingDbEnv" -}}
+{{- $tokenAuth := include "litellm.db.tokenAuth" . -}}
+- name: DATABASE_USERNAME
+  {{- if .Values.db.username }}
+  value: {{ .Values.db.username | quote }}
+  {{- else }}
+  valueFrom:
+    secretKeyRef:
+      name: {{ .Values.db.secret.name }}
+      key: {{ .Values.db.secret.usernameKey }}
+  {{- end }}
+{{- if .Values.db.useIAMAuth }}
+- name: IAM_TOKEN_DB_AUTH
+  value: "true"
+{{- else if .Values.db.useAzureEntraAuth }}
+- name: AZURE_POSTGRESQL_AUTH
+  value: "true"
+{{- else }}
+- name: DATABASE_PASSWORD
+  valueFrom:
+    secretKeyRef:
+      name: {{ .Values.db.secret.name }}
+      key: {{ .Values.db.secret.passwordKey }}
+{{- end }}
+- name: DATABASE_HOST
+  {{- if .Values.db.secret.endpointKey }}
+  valueFrom:
+    secretKeyRef:
+      name: {{ .Values.db.secret.name }}
+      key: {{ .Values.db.secret.endpointKey }}
+  {{- else }}
+  value: {{ .Values.db.endpoint }}
+  {{- end }}
+{{- if $tokenAuth }}
+- name: DATABASE_PORT
+  value: {{ .Values.db.port | quote }}
+{{- end }}
+- name: DATABASE_NAME
+  value: {{ .Values.db.database }}
+{{- if and $tokenAuth .Values.db.schema }}
+- name: DATABASE_SCHEMA
+  value: {{ .Values.db.schema | quote }}
+{{- end }}
+{{- if not $tokenAuth }}
+- name: DATABASE_URL
+  value: {{ .Values.db.url | quote }}
+{{- end }}
+{{- end -}}
+
 {{/*
 Environment shared by the proxy container and the opt-in collector sidecar:
 database, pgbouncer, master key, redis, user envVars. Both containers must see
@@ -169,6 +214,7 @@ the same DATABASE_URL and REDIS_* so the sidecar reaches the pod's pgbouncer
 and the same spend transaction buffer.
 */}}
 {{- define "litellm.proxyEnv" -}}
+{{- include "litellm.validateDbTokenAuth" . -}}
 - name: HOST
   value: "{{ .Values.listen | default "0.0.0.0" }}"
 - name: PORT
@@ -189,29 +235,7 @@ and the same spend transaction buffer.
 - name: DATABASE_NAME
   value: litellm
 {{- else if .Values.db.useExisting }}
-- name: DATABASE_USERNAME
-  valueFrom:
-    secretKeyRef:
-      name: {{ .Values.db.secret.name }}
-      key: {{ .Values.db.secret.usernameKey }}
-- name: DATABASE_PASSWORD
-  valueFrom:
-    secretKeyRef:
-      name: {{ .Values.db.secret.name }}
-      key: {{ .Values.db.secret.passwordKey }}
-- name: DATABASE_HOST
-  {{- if .Values.db.secret.endpointKey }}
-  valueFrom:
-    secretKeyRef:
-      name: {{ .Values.db.secret.name }}
-      key: {{ .Values.db.secret.endpointKey }}
-  {{- else }}
-  value: {{ .Values.db.endpoint }}
-  {{- end }}
-- name: DATABASE_NAME
-  value: {{ .Values.db.database }}
-- name: DATABASE_URL
-  value: {{ .Values.db.url | quote }}
+{{ include "litellm.existingDbEnv" . }}
 {{- end }}
 {{- if and .Values.db.useExisting .Values.db.readReplicaUrl .Values.db.secret.readReplicaEndpointKey (not .Values.db.secret.readReplicaUrlKey) }}
 - name: DATABASE_READER_HOST
@@ -322,10 +346,6 @@ through an emptyDir. Empty when the sidecar is off or uses 127.0.0.1 TCP.
   value: {{ .Values.collector.drainTimeoutSeconds | quote }}
 {{- end -}}
 
-{{- define "litellm.lensWorker.image" -}}
-{{- include "lens.image" (dict "Values" .Values.lensWorker "Chart" .Subcharts.lens.Chart) -}}
-{{- end -}}
-
 {{- define "litellm.gateway.collectorSocketDir" -}}
 {{- if and .Values.gateway.collector.enabled (hasPrefix "unix://" .Values.gateway.collector.address) -}}
 {{- dir (trimPrefix "unix://" .Values.gateway.collector.address) -}}
@@ -352,95 +372,28 @@ shutdown drain window.
 {{- end }}
 {{- end -}}
 
-{{- define "litellm.lensWorker.labels" -}}
-{{- $labels := include "litellm.labels" . | fromYaml -}}
-{{- $_ := set $labels "app.kubernetes.io/name" (printf "%s-lens-worker" (include "litellm.name" . | trunc 51 | trimSuffix "-")) -}}
-{{- toYaml $labels -}}
-{{- end -}}
-
-{{- define "litellm.lensWorker.serviceTokenSecretName" -}}
-{{- .Values.lensWorker.serviceTokenSecret.name | default (printf "%s-lens-service" (include "litellm.fullname" .)) -}}
-{{- end -}}
-
-{{- define "litellm.lensWorker.gatewaySecretName" -}}
-{{- .Values.lensWorker.gateway.secretName | default (printf "%s-lens-gateway" (include "litellm.fullname" . | trunc 50 | trimSuffix "-")) -}}
-{{- end -}}
-
-{{- define "litellm.lensWorker.bundledClickhouse" -}}
-{{- if and (eq (include "litellm.lens.mode" .) "bundled") .Values.lensWorker.clickhouse.enabled (not .Values.lensWorker.clickhouseSecret.name) -}}true{{- end -}}
-{{- end -}}
-
-{{- define "litellm.lensWorker.publicUrl" -}}
-{{- if .Values.lensWorker.publicUrl -}}
-{{- .Values.lensWorker.publicUrl -}}
-{{- else if eq (include "litellm.lens.mode" .) "external" -}}
-{{- fail "lensWorker.publicUrl is required for external Lens" -}}
-{{- else if .Values.lensWorker.ingress.enabled -}}
-{{- $tls := or (not (empty .Values.lensWorker.ingress.tls)) (hasKey .Values.lensWorker.ingress.annotations "alb.ingress.kubernetes.io/certificate-arn") -}}
-{{- printf "%s://%s" (ternary "https" "http" $tls) (required "lensWorker.ingress.host is required" .Values.lensWorker.ingress.host) -}}
-{{- else if and .Values.ingress.enabled (eq (len .Values.ingress.hosts) 1) -}}
-{{- $host := required "ingress.hosts[0].host is required" (first .Values.ingress.hosts).host -}}
-{{- $tls := or (not (empty .Values.ingress.tls)) (hasKey .Values.ingress.annotations "alb.ingress.kubernetes.io/certificate-arn") -}}
-{{- printf "%s://%s/lens-ingest" (ternary "https" "http" $tls) $host -}}
-{{- else -}}
-{{- fail "lensWorker.publicUrl is required when there is no single ingress hostname" -}}
-{{- end -}}
-{{- end -}}
-
-{{- define "litellm.lensWorker.clickhouseName" -}}
-{{- printf "%s-lens-clickhouse" (include "litellm.fullname" . | trunc 47 | trimSuffix "-") -}}
-{{- end -}}
-
 {{- define "litellm.lensConnectionEnv" -}}
-{{- $mode := include "litellm.lens.mode" . -}}
-{{- if ne $mode "disabled" }}
-{{- if and (eq $mode "external") (not .Values.lensWorker.serviceTokenSecret.name) -}}
-{{- fail "lensWorker.serviceTokenSecret.name is required for external Lens" -}}
+{{- with .Values.lensWorker }}
+{{- if or .enabled (and .mode (ne .mode "disabled")) }}
+{{- fail "lensWorker was removed; deploy Lens from its own chart and set lens.* instead, see the chart README" }}
 {{- end }}
-{{- if and (eq $mode "external") (not .Values.lensWorker.gateway.secretName) -}}
-{{- fail "lensWorker.gateway.secretName is required for external Lens" -}}
-{{- end -}}
+{{- end }}
+{{- with .Values.lens }}
+{{- if .enabled }}
 - name: LITELLM_LENS_URL
-  value: {{ if eq $mode "external" }}{{ required "lensWorker.externalUrl is required for external Lens" .Values.lensWorker.externalUrl | quote }}{{ else }}{{ printf "http://%s-lens-worker:%v" (include "litellm.fullname" .) .Values.lensWorker.service.port | quote }}{{ end }}
+  value: {{ required "lens.url is required when lens.enabled is true" .url | quote }}
 - name: LITELLM_LENS_PUBLIC_URL
-  value: {{ include "litellm.lensWorker.publicUrl" . | quote }}
+  value: {{ required "lens.publicUrl is required when lens.enabled is true" .publicUrl | quote }}
 - name: LENS_GATEWAY_SECRET
   valueFrom:
     secretKeyRef:
-      name: {{ include "litellm.lensWorker.gatewaySecretName" . | quote }}
-      key: {{ .Values.lensWorker.gateway.secretKey | quote }}
+      name: {{ required "lens.gatewaySecret.name is required when lens.enabled is true" .gatewaySecret.name | quote }}
+      key: {{ .gatewaySecret.key | quote }}
 - name: LITELLM_LENS_SERVICE_TOKEN
   valueFrom:
     secretKeyRef:
-      name: {{ include "litellm.lensWorker.serviceTokenSecretName" . | quote }}
-      key: {{ .Values.lensWorker.serviceTokenSecret.key | quote }}
+      name: {{ required "lens.serviceTokenSecret.name is required when lens.enabled is true" .serviceTokenSecret.name | quote }}
+      key: {{ .serviceTokenSecret.key | quote }}
+{{- end }}
 {{- end }}
 {{- end -}}
-
-{{- define "litellm.lens.mode" -}}
-{{- $mode := .Values.lensWorker.mode | default (ternary "bundled" "disabled" .Values.lensWorker.enabled) -}}
-{{- if not (has $mode (list "bundled" "external" "disabled")) -}}
-{{- fail "lensWorker.mode must be bundled, external or disabled" -}}
-{{- end -}}
-{{- $mode -}}
-{{- end -}}
-
-{{- define "litellm.lens.render" -}}
-{{- $root := .root -}}
-{{- $values := mergeOverwrite (deepCopy $root.Subcharts.lens.Values) (deepCopy $root.Values.lensWorker) -}}
-{{- $_ := set $values "fullnameOverride" (printf "%s-lens-worker" (include "litellm.fullname" $root)) -}}
-{{- $_ := set $values "component" "lens-worker" -}}
-{{- $_ := set $values "nameOverride" (printf "%s-lens-worker" (include "litellm.name" $root | trunc 51 | trimSuffix "-")) -}}
-{{- $_ := set $values "imagePullSecrets" $root.Values.imagePullSecrets -}}
-{{- $_ := set $values.gateway "enabled" true -}}
-{{- $_ := set $values.gateway "generatedName" (include "litellm.lensWorker.gatewaySecretName" $root) -}}
-{{- $_ := set $values.clickhouse "nameOverride" (include "litellm.lensWorker.clickhouseName" $root) -}}
-{{- $_ := set $values.serviceTokenSecret "generatedName" (include "litellm.lensWorker.serviceTokenSecretName" $root) -}}
-{{- $_ := set $values "publicUrl" ($root.Values.lensWorker.standaloneUrl | default $root.Subcharts.lens.Values.publicUrl) -}}
-{{- if and (eq .resource "deployment") (or $root.Values.lensWorker.publicUrl $root.Values.lensWorker.ingress.enabled $root.Values.ingress.enabled) -}}
-{{- $ingestion := include "litellm.lensWorker.publicUrl" $root -}}
-{{- $_ := set $values "ingestionUrl" $ingestion -}}
-{{- $_ := set $values "publicUrl" ($root.Values.lensWorker.standaloneUrl | default (trimSuffix "/lens-ingest" $ingestion)) -}}
-{{- end -}}
-{{- include (printf "lens.%s" .resource) (dict "Values" $values "Release" $root.Release "Chart" $root.Subcharts.lens.Chart "Capabilities" $root.Capabilities) -}}
-{{- end -}}

helm/litellm-helm/templates/deployment.yaml

@@ -148,9 +148,6 @@ spec:
             - name: npm
               mountPath: /.npm
           {{- end }}
-          {{- if .Values.billingMetrics.enabled }}
-            {{- include "litellm.billingMetricsVolumeMounts" . | nindent 12 }}
-          {{- end }}
           {{- if include "litellm.collector.socketDir" . }}
             - name: collector-socket
               mountPath: {{ include "litellm.collector.socketDir" . }}
@@ -234,9 +231,6 @@ spec:
             items:
               - key: {{ .Values.proxyConfigMap.key | default "config.yaml" }}
                 path: "config.yaml"
-        {{- if .Values.billingMetrics.enabled }}
-        {{- include "litellm.billingMetricsVolumes" . | nindent 8 }}
-        {{- end }}
         {{- if include "litellm.collector.socketDir" . }}
         - name: collector-socket
           emptyDir:

helm/litellm-helm/templates/ingress.yaml

@@ -1,8 +1,6 @@
 {{- if .Values.ingress.enabled -}}
 {{- $fullName := include "litellm.fullname" . -}}
 {{- $svcPort := .Values.service.port -}}
-{{- $lensMode := include "litellm.lens.mode" . -}}
-{{- $lensService := .Values.lensWorker.externalServiceName | default (printf "%s-lens-worker" $fullName) -}}
 {{- if and .Values.ingress.className (not (semverCompare ">=1.18-0" .Capabilities.KubeVersion.GitVersion)) }}
   {{- if not (hasKey .Values.ingress.annotations "kubernetes.io/ingress.class") }}
   {{- $_ := set .Values.ingress.annotations "kubernetes.io/ingress.class" .Values.ingress.className}}
@@ -46,18 +44,18 @@ spec:
     - host: {{ .host | quote }}
       http:
         paths:
-          {{- if or (eq $lensMode "bundled") (and (eq $lensMode "external") $.Values.lensWorker.externalServiceName) }}
+          {{- if and $.Values.lens.enabled $.Values.lens.service.name }}
           - path: /lens-ingest
             pathType: Prefix
             backend:
               {{- if semverCompare ">=1.19-0" $.Capabilities.KubeVersion.GitVersion }}
               service:
-                name: {{ if eq $lensMode "external" }}{{ $lensService }}{{ else }}{{ $fullName }}-lens-worker{{ end }}
+                name: {{ $.Values.lens.service.name }}
                 port:
-                  number: {{ $.Values.lensWorker.service.port }}
+                  number: {{ $.Values.lens.service.port }}
               {{- else }}
-              serviceName: {{ if eq $lensMode "external" }}{{ $lensService }}{{ else }}{{ $fullName }}-lens-worker{{ end }}
-              servicePort: {{ $.Values.lensWorker.service.port }}
+              serviceName: {{ $.Values.lens.service.name }}
+              servicePort: {{ $.Values.lens.service.port }}
               {{- end }}
           {{- end }}
           {{- range .paths }}