* Refactor Bedrock response stream shape handling
- Introduced a module-level constant `BEDROCK_RESPONSE_STREAM_SHAPE` to cache the response stream shape, eliminating the need for per-instance caching in `BedrockEventStreamDecoderBase`.
- Updated relevant methods to utilize the new constant, improving performance by avoiding redundant loading of the shape.
- Added tests to ensure the shape is loaded correctly at import time and is consistent across different modules.
- Added a new mock server script for testing Bedrock pass-through functionality.
* Refactor response parsing for Bedrock and SageMaker
- Improved code readability by formatting the parsing method calls in `AWSEventStreamDecoder` for both Bedrock and SageMaker response stream shapes.
- Added blank lines for better separation of code blocks in `invoke_handler.py` and `common_utils.py` to enhance maintainability.
* Enhance error handling for Bedrock and SageMaker response stream shape loading
- Wrapped the loading logic in `_load_bedrock_response_stream_shape` and `_load_sagemaker_response_stream_shape` with try-except blocks to gracefully handle exceptions.
- Added logging to warn when the response stream shape cannot be pre-loaded, ensuring the module imports cleanly.
- Updated tests to verify that loading failures return `None` instead of propagating exceptions.
* Implement error handling for missing response stream shapes in Bedrock and SageMaker
- Added checks in `_parse_message_from_event` methods to raise appropriate errors when `BEDROCK_RESPONSE_STREAM_SHAPE` or `SAGEMAKER_RESPONSE_STREAM_SHAPE` is None, ensuring clearer error reporting.
- Updated logging messages to reflect the unavailability of event-stream decoding for both Bedrock and SageMaker.
- Enhanced unit tests to verify that the correct exceptions are raised when the response stream shapes are not loaded.
* fix(hosted_vllm): normalize custom tools for chat completions
Convert custom tool definitions into OpenAI function tools before forwarding hosted_vllm chat requests to avoid provider-side validation failures. Add a regression test and include a local curl verification screenshot.
Made-with: Cursor
* Fix black issue
* Fix hosted vllm custom tool schema fallback
* fix black
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
* feat(audio_transcription): add NVIDIA Riva STT provider
Adds nvidia_riva as a new audio transcription provider, supporting both
NVCF-hosted and self-hosted Riva ASR deployments via gRPC streaming.
- Auto-resamples input audio to 16 kHz mono LINEAR_PCM (soundfile + numpy,
audioread fallback) so callers can send any common format.
- Maps OpenAI params: language (en -> en-US), response_format (text/json/
verbose_json), timestamp_granularities=["word"] -> enable_word_time_offsets,
word offsets converted ms -> s for verbose_json.
- Auth: NVCF when nvcf_function_id is set (SSL on by default), self-hosted
otherwise (SSL off by default), with explicit use_ssl override.
- gRPC errors wrapped via NvidiaRivaException -> litellm exception classes.
- Optional deps gated behind [stt-nvidia-riva] extra (nvidia-riva-client,
soundfile, audioread, numpy).
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(nvidia_riva): address PR review feedback
- handler: forward call-level `timeout` to streaming_response_generator
(kwarg-detected via inspect for older riva-client compat) so a stalled
Riva server cannot block the caller indefinitely.
- audio_utils: spill bytes to a tempfile before audioread.audio_open;
most audioread backends (FFmpeg, GStreamer) require a real filesystem
path and previously raised TypeError on BytesIO, breaking the mp3/m4a
fallback path.
- audio_utils: prefer soxr / scipy.signal.resample_poly for resampling
(anti-aliased polyphase) when installed, falling back to linear only
as a last resort. Avoids aliasing on 44.1/48 kHz -> 16 kHz downsamples.
- transformation: bare `es` now maps to es-ES (Castilian) instead of
es-US, matching BCP-47 conventions.
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore: trigger CI re-run [stabilize loop 1/3]
* Update litellm/llms/nvidia_riva/audio_transcription/transformation.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* chore: trigger CI re-run [stabilize loop 1/3]
* fix code qa
* fix lint
* fix mypy
* fix mypy
* Fix NVIDIA Riva ASR service lookup
* Fix NVIDIA Riva transcription payload logging
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: oss-pr-review-agent-shin[bot] <281797381+oss-pr-review-agent-shin[bot]@users.noreply.github.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(anthropic, mcp): sanitize tool names to match Anthropic's `^[a-zA-Z0-9_-]{1,128}$`
Tool names with characters like `/` or `.` (commonly produced by the
OpenAPI -> MCP generator from `operationId`s such as
`actions/download-job-logs-for-workflow-run`) caused Anthropic to reject
requests with `tools.N.custom.name: String should match pattern
'^[a-zA-Z0-9_-]{1,128}$'`.
Two layers of fix:
1. Anthropic transformation: build a per-request forward map (original ->
sanitized, disambiguated by suffix on collisions) and a reverse map
(only for names actually rewritten). Forward map is applied to tool
defs, `tool_choice`, and historical assistant tool_calls in messages.
Reverse map is threaded through both the non-streaming and streaming
response paths so callers continue to see their original tool names
in `tool_use` blocks.
2. OpenAPI -> MCP generator: sanitize `operationId` (and the
method+path fallback) at registration time so generated MCP tools are
valid for any strict-name provider, not just Anthropic. The dashboard
preview endpoint applies the same sanitization for parity.
Includes unit tests covering: collision disambiguation between
`foo_bar` and `foo/bar` in the same request, reverse-map only firing
for actually-rewritten names, message rewrite for historical tool_calls,
streaming chunk_parser reverse-mapping, and sanitization of OpenAPI
operationIds plus the preview endpoint output.
Made-with: Cursor
* fix(anthropic): build tool-name maps in transform_request, not optional_params
The previous patch stashed the per-request forward and reverse tool-name
maps under ``optional_params["_anthropic_tool_name_forward_map"]`` and
``optional_params["_anthropic_tool_name_map"]``. ``optional_params`` is
the dict that becomes the JSON body via ``data = {**optional_params}``,
so those internal keys leaked over the wire and Anthropic 400'd with:
_anthropic_tool_name_forward_map: Extra inputs are not permitted
Worse, this meant *every* request whose tool list contained any name with
an invalid character (the exact case the patch was meant to fix) regressed
into a confusing meta-error pointing at LiteLLM's internal map instead of
the offending tool.
Fix: move all tool-name sanitization into ``transform_request``, which is
the single chokepoint already shared by ``AnthropicConfig``,
``AmazonAnthropicConfig`` (Bedrock invoke), ``VertexAIAnthropicConfig``,
and ``AzureAnthropicConfig`` (all call ``super().transform_request`` /
``AnthropicConfig.transform_request(self, ...)``). New static helper
``_sanitize_tool_names_in_request`` walks the already-Anthropic-shaped
``optional_params["tools"]`` (only ``type=="custom"`` entries -- hosted
tool names are reserved by Anthropic and must not be touched), builds
the per-request forward/reverse maps, and applies the forward map in
place to ``tools[*].name`` and ``tool_choice.name``. The reverse map is
stashed exclusively on ``litellm_params`` (which is never serialized to
a provider) under ``_anthropic_tool_name_map`` for the response paths
to consume.
Side effect of this restructure: ``map_openai_params`` is now a pure
OpenAI->Anthropic param translator with no side-channel state, which
matches its contract everywhere else in the codebase.
Tests: replaced the now-incorrect "stashes maps in optional_params"
tests with regressions that assert no underscore-prefixed keys appear
in either ``optional_params`` after ``map_openai_params`` or in the
final ``transform_request`` body. Added end-to-end coverage for:
sanitization in ``transform_request``, ``tool_choice`` rewriting,
historical ``tool_calls`` rewriting in messages, and hosted-tool
passthrough.
Made-with: Cursor
* fix(anthropic): always sanitize empty text content blocks
Anthropic 400s on `{"role": "user", "content": ""}` with:
"messages: text content blocks must be non-empty"
LiteLLM already had `_sanitize_empty_text_content` to rewrite empty text
to a placeholder, but it was gated behind `litellm.modify_params=True`.
With that flag off (default), empty content from upstream agent
frameworks (e.g. pydantic-ai) flowed straight through and tripped the
Anthropic validator.
Fix:
- Always run `_sanitize_empty_text_content` at the top of
`anthropic_messages_pt`, independent of `modify_params`. There is no
way to "pass through" an empty text block, so this is non-optional.
The richer tool-call sanitizations (Cases A/B/D, which actually
mutate conversation structure) remain gated on `modify_params`.
- Extend `_sanitize_empty_text_content` to also handle list-of-blocks
content (`[{"type": "text", "text": ""}]`), not just string content.
Adds 3 regression tests covering string content, list-of-blocks
content, and the no-op case (non-empty messages with modify_params off).
Made-with: Cursor
* fix(anthropic): drop dead tool-name forward-map params, fix mypy + caller-mutation
- remove unused `name_forward_map` param from `_map_tool_choice`,
`_map_tool_helper`, `_map_tools` and the `_apply_anthropic_tool_name_forward`
helper. Production sanitization runs in `_sanitize_tool_names_in_request`
at `transform_request`; these params were never threaded through.
- handler.py: use `ANTHROPIC_TOOL_NAME_REVERSE_MAP_KEY` constant instead of
the hardcoded `"_anthropic_tool_name_map"` string.
- fix mypy `"object" has no attribute "__iter__"` in
`_rewrite_tool_names_in_messages` by guarding `tool_calls` with
`isinstance(..., list)`.
- `_sanitize_tool_names_in_request`: build a new tools list with copy-on-
change entries (and copy `tool_choice` on rewrite) so a caller reusing
the same tool list/dicts across requests doesn't see its inputs
permanently rewritten.
- doc-comment `_build_request_tool_name_maps` clarifying it operates on
OpenAI-format tools (vs `_sanitize_tool_names_in_request` which runs
on Anthropic-format tools post-`_map_tools`).
- tests: drop 3 tests pinning the now-removed param paths; add coverage
for tool_calls + None function_call rewrite and caller-dict immutability.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(mcp): inherit stored credentials in test/tools/list for edit flow
When editing an existing MCP server, the Tool Configuration preview
calls POST /mcp-rest/test/tools/list with server_id but no credentials
(management API redacts them). The endpoint now calls
_inherit_credentials_from_existing_server() so stored bearer tokens
and OAuth2 M2M credentials are loaded from global_mcp_server_manager
automatically — tools load without re-entering credentials.
New servers (no server_id) and requests with explicit credentials are
unaffected (function is a no-op in both cases).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(mcp): show all tools in edit panel, not just allowed tools
Edit flow was passing externalTools (from GET /tools/list, filtered by
allowed_tools) to MCPToolConfiguration, disabling the internal hook.
Remove the external props so the internal hook fires via
POST /test/tools/list, which returns all tools unfiltered. Combined
with the credential inheritance fix, tools load automatically without
re-entering credentials and all tools are visible for re-configuration.
existingAllowedTools still pre-checks previously allowed tools.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix order-dependent collision in _build_anthropic_tool_name_maps
Use a two-pass approach: first pre-register all already-valid tool names
in the 'used' set, then sanitize/disambiguate names that need rewriting.
This ensures valid names always have priority regardless of input order,
preventing duplicate tool names on the wire when e.g. 'foo/bar' appears
before 'foo_bar' in the tool list.
Add regression test for the reversed ordering case.
* Fix OpenAPI tool name collision: disambiguate sanitized names with numeric suffixes
sanitize_openapi_tool_name replaces all invalid chars with '_', but when
two operationIds differ only by sanitized characters (e.g. 'foo/list' and
'foo.list' both become 'foo_list'), the second registration silently
overwrites the first in the tool registry.
Add collision disambiguation in register_tools_from_openapi that appends
_2, _3, ... suffixes when a sanitized name is already taken, mirroring
the existing logic in _build_anthropic_tool_name_maps.
* Fix preview endpoint missing collision disambiguation for tool names
Add used_names tracking and _2/_3 suffix disambiguation to
_preview_openapi_tools, matching the logic in register_tools_from_openapi.
Without this, two operationIds that sanitize to the same string (e.g.
'foo/list' and 'foo.list' both becoming 'foo_list') would show duplicate
names in the preview while registration would disambiguate them.
* Align preview HTTP method order with register_tools_from_openapi
The preview endpoint and register_tools_from_openapi both use
order-dependent collision disambiguation (_2, _3 suffixes). When the
iteration order differs, two operations on the same path with sanitized
names that collide get different suffixes in preview vs registration,
so the dashboard shows names that don't match what actually got
registered.
Also adds a regression test that fails on the swapped order.
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
* Skip duplicate originals in _build_anthropic_tool_name_maps
If the same invalid tool name appeared twice in original_names (e.g.
['foo/bar', 'foo/bar']), the second occurrence overwrote the forward
map entry with a freshly-suffixed name (foo_bar_2), leaving foo_bar
orphaned in 'used' with no reverse mapping. _sanitize_tool_names_in_request
then rewrote both tool entries to foo_bar_2, and Anthropic 400'd on
duplicate tool names.
Skip the rewrite if forward already has the original mapped.
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
* feat(realtime): OpenAI Realtime GA support and beta compatibility
- Normalize beta-style session.update to GA for upstream OpenAI; optional GA→beta
event translation when client sends OpenAI-Beta: realtime=v1
- Default upstream WebSocket without OpenAI-Beta; forward header when client opts in
- Extend OpenAI realtime types for GA event names and conversation item shapes
- Relax LiteLLMRealtimeStreamLoggingObject.results to List[Any] for GA events
- Update proxy client_secrets fallback to omit beta header; dashboard RealtimePlayground
- Add unit tests for remap, translation, and beta header helper
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix results
* fix greptile
* Fix mypy issues
* Remove unused class constants _GA_TEXT_DELTA_TYPES and _GA_AUDIO_DELTA_TYPES
These frozensets were defined as class-level constants in realtime_streaming.py
but never referenced anywhere in the codebase. Removing dead code.
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
* fix(realtime): use GA-shaped session.update in guardrail injections
The guardrail VAD injection code sent a beta-style session.update with a
flat turn_detection field:
{"session": {"turn_detection": {"create_response": false}}}
When the upstream OpenAI backend operates in GA mode (no OpenAI-Beta
header forwarded), it requires the nested GA shape:
{"session": {"type": "realtime", "audio": {"input": {"turn_detection": {"create_response": false}}}}}
The _remap_beta_session_to_ga helper was only applied to client-
originated session.update messages in client_ack_messages. Internally-
generated session.updates (sent via _send_to_backend) in two paths:
- _handle_raw_backend_message (raw/no provider_config path, line 518)
- backend_to_client_send_messages provider_config path (line 481)
bypassed the remap, so GA upstreams ignored or rejected them, breaking
audio transcription guardrails for all non-beta clients.
Fix: add _make_disable_auto_response_message() helper that always emits
the correct GA-shaped session.update, and replace both injection sites
with it.
Update existing tests to assert the GA nested shape instead of the old
flat beta shape, and add a new unit test for the helper itself.
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
* Log realtime session type
* Fix beta realtime session payloads
* Fix realtime audio format remapping edge case
* Fix Azure realtime beta session shape
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* refactor(BaseAWSLLM): implement shared IAM cache and static credential caching
- Introduced a process-wide shared IAM cache to optimize credential management across instances.
- Added a method to handle caching of static credentials, ensuring only long-lived credentials are cached.
- Updated the get_credentials method to utilize the new caching mechanism for static credential flows.
- Enhanced unit tests to verify the correct behavior of the shared cache and static credential usage.
* refactor(BaseAWSLLM): enhance IAM credential caching and update related tests
- Improved the process-wide IAM credential caching mechanism to better handle static and AssumeRole credentials.
- Renamed the caching method for clarity and updated comments to reflect the new caching behavior.
- Added a fixture to ensure the IAM cache is flushed between tests to prevent leakage of cached entries.
- Updated unit tests to verify the correct behavior of the shared IAM cache, particularly for static credentials and role assumptions.
* refactor(BaseAWSLLM): clarify IAM credential caching behavior and enhance tests
- Updated documentation to specify that only static and ambient environment credentials are cached, excluding AssumeRole and other credential types.
- Modified the caching logic to ensure that AssumeRole credentials are not stored in the IAM cache, requiring STS calls for each request.
- Enhanced unit tests to verify that AssumeRole credentials are not cached and to ensure proper behavior of the IAM cache across different scenarios.
* Code Readability improvement for aws auth path
* refactor(BaseAWSLLM): enhance IAM credential caching documentation and add tests
- Updated comments to clarify the behavior of the in-process IAM credential cache, specifying the TTL for static and ambient credentials.
- Added new unit tests to verify the caching behavior for ambient environment credentials across instances and ensure that static access key sessions are constructed only once when cached.
- Ensured that temporary session tokens and AWS profiles are not cached, validating the expected behavior through additional tests.
* refactor(BaseAWSLLM): improve IAM credential handling and add tests for role assumption
- Updated comments to clarify the behavior of IAM credential caching, particularly regarding the handling of ambient credentials and role assumptions.
- Enhanced unit tests to verify that the caching mechanism correctly distinguishes between already running roles and new role assumptions, ensuring that cached environment credentials are not reused incorrectly.
- Added a new test case to validate the behavior when switching roles, confirming that the system correctly uses AssumeRole when the role changes.
- Add image_generation/http_utils.azure_deployment_image_generation_json_body; call
from azure.py (keeps AzureChatCompletion focused on chat).
- Rename finalize_image_edit_multipart_data to finalize_image_edit_request_data with
docstring covering multipart and JSON POST payloads (review feedback).
Co-authored-by: Cursor <cursoragent@cursor.com>
Substantial reduction (~765 LOC) without changing the security
boundary:
* Drop ContainerOwnershipStore and LiteLLMSkillsStore — both were
one-method-per-Prisma-call wrappers. Inline the calls instead,
matching the established pattern in vector_store_endpoints,
agent_endpoints, and mcp_server/db.py.
* Drop the prisma_client is None in-memory fallback. Production
deploys always have Prisma; running ownership-critical paths on a
process-local dict is a security footgun in the dev-mode case it
was meant to support, and complicates every code path with a
branch. Fail-secure: skip recording if Prisma is unavailable, and
treat reads as "not found" (admin-only).
* Drop the hand-rolled module-level cache. Replace with the existing
litellm.caching.in_memory_cache.InMemoryCache, which already has
TTL + max-size + eviction tested in its own module. Sentinel string
for negative caching since InMemoryCache can't disambiguate "miss"
from "cached as None".
* Tests: drop coverage for removed code paths (in-memory fallback,
hand-rolled cache internals). Keep tests for actual behavior (cache
hit-rate, negative caching, owner check, list filtering,
identity-less reject, admin bypass).
Two cleanups:
* ``LiteLLMSkillsHandler.create_skill`` raised ``HTTPException`` for
identity-less callers, importing FastAPI from a ``litellm/llms/``
module — that violates the project rule that FastAPI lives only
under ``proxy/``. Switch to ``ValueError`` (the same shape the rest
of the handler uses for not-found/forbidden) and update the test.
* The proxy-auth body bouncer derived its observability ban list from
``_supported_callback_params`` only, missing
``_request_blocked_callback_params`` (where ``gcs_bucket_name`` and
``gcs_path_service_account`` live). Two recently-merged sibling PRs
(#27019 added the deny list, #27081 added the test asserting these
are rejected at the request body root) crossed without folding them
together. Union the GCS deny list into the bouncer's derivation so
the single source of truth covers both code paths.
UNSCOPED_RESOURCE_OWNER_SCOPE collapsed every caller without an
identity field (no user_id / team_id / org_id / api_key / token) into
a single shared owner — a cross-tenant access primitive: any two such
callers could see and delete each other's containers and skills.
Drop the sentinel. ``get_primary_resource_owner_scope`` returns
``None`` and ``get_resource_owner_scopes`` returns ``[]`` for
identity-less callers. ``record_container_owner`` and
``LiteLLMSkillsHandler.create_skill`` now reject creates from
identity-less callers with a 403 instead of stamping the placeholder.
Read paths already deny ``owner is None`` correctly so legacy rows
(if any) are admin-only.
LITELLM_ALLOW_UNTRACKED_CONTAINER_ACCESS and
LITELLM_ALLOW_UNOWNED_SKILL_ACCESS were operator-toggleable opt-outs
for the cross-tenant access primitive this PR closes — flipping either
on re-enabled exactly the VERIA-20 read path. Default-secure with no
escape hatch matches sibling fixes (vector-store cred isolation, semantic
cache key isolation, user_config strip): all rejected the
opt-out-of-security pattern.
Untracked containers and unowned skills (rows that pre-date this
enforcement) are admin-only. Non-admin owners need to either re-create
via the now-tracked flow or have an admin assign ``created_by`` on the
existing row. Update tests to assert the strict-only behaviour.
Two cleanups from the /simplify pass:
* ``_CONTAINER_OWNER_CACHE`` and ``_SKILL_CACHE`` now LRU-evict via
``OrderedDict.popitem(last=False)`` instead of full ``clear()`` at
capacity. Full clears converted a steady-state cached workload into a
periodic full-DB-load oscillation as the cache repopulated from zero
and cleared again. Reads now ``move_to_end`` so the just-touched
entry survives the next eviction. Mirrors the pre-existing LRU
pattern in ``_remember_container_owner``.
* ``LiteLLM_ManagedObjectTable.file_purpose`` Literal now includes
``"container"`` so Pydantic validation accepts rows written by the
ownership store.
Strip out the explanatory and historical comments that don't carry
business-logic justification. Comments that simply narrate what code
does — or that explain prior behavior, what was changed, or which PR
introduced a fix — are removed. Docstrings are reduced to a one-line
summary where the long form repeated information already evident from
the code or test data.
No code-behavior changes. All 643 affected unit tests still pass.
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
xAI's chat completions API accounts reasoning_tokens separately from
completion_tokens, but rolls them into total_tokens. This breaks the
OpenAI invariant total_tokens == prompt_tokens + completion_tokens
that downstream consumers (including litellm's own _usage_format_tests
in tests/llm_translation/base_llm_unit_tests.py:58) rely on.
Live capture (grok-3-mini-beta, 2026-05-04):
prompt=14, completion=10, total=336, reasoning=312
14 + 10 = 24, NOT 336.
OpenAI's o1/o3 reasoning models include reasoning_tokens in
completion_tokens, leaving the prompt+completion=total invariant
intact. xAI deviates. This patch aligns xAI to OpenAI semantics by
folding reasoning_tokens into completion_tokens after the parent
OpenAI parser runs.
The fold is idempotent and defensive:
- Only fires when total_tokens == prompt_tokens + completion_tokens
+ reasoning_tokens (the documented xAI shape). Refuses to fold if
the gap doesn't match, guarding against silent corruption when xAI
changes accounting.
- Skips if completion_tokens already covers the gap (already
normalised — e.g. cost calc replays a previously-folded Usage).
xai.cost_calculator.cost_per_token already added reasoning_tokens to
the visible completion count for billing. Post-fold the Usage block
now satisfies that invariant directly, so the cost calc would
double-bill. Updated cost_per_token to detect the OpenAI-normalised
shape (total == prompt + completion) and skip the reasoning add-on
in that case, falling through to the legacy raw-shape behaviour for
callers that bypass the transformation (e.g. proxy log replay).
Tests:
- Adds TestXAIReasoningTokenFolding covering: gap-explained-fold,
idempotent-no-double-fold, no-reasoning-skip, gap-mismatch-skip.
- Adds test_already_normalised_usage_does_not_double_count_reasoning
to lock the cost-calc idempotency.
- Updates 7 pre-existing cost-calc tests whose total_tokens was
internally inconsistent (used the OpenAI-normalised total but kept
reasoning_tokens external) to use the documented xAI raw shape
total = prompt + visible completion + reasoning. Pre-existing
values masked the missing-fold by accident.
Verified end-to-end against the live xAI API:
LITELLM_LOCAL_MODEL_COST_MAP=False (CI default) +
XAI_API_KEY set +
pytest tests/llm_translation/test_xai.py::TestXAIChat::test_prompt_caching
-> PASSED in 18.81s (was: AssertionError on
usage.total_tokens == usage.prompt_tokens + usage.completion_tokens)
20/20 tests in tests/test_litellm/llms/xai/test_xai_cost_calculator.py
and 8/8 in tests/test_litellm/llms/xai/test_xai_chat_transformation.py
pass.
CI runs without LITELLM_LOCAL_MODEL_COST_MAP=True, so litellm.model_cost
is loaded from main-branch JSON (default model_cost_map_url) instead of
the PR's checked-out model_prices_and_context_window.json. Tests that
assert per-model flags added in this PR (supports_max_reasoning_effort,
supports_xhigh_reasoning_effort) therefore pass locally but fail in CI
with 'AssertionError: assert False is True' on 5 cases:
- test_anthropic_model_supports_effort_param_recognizes_supporting_models
[anthropic.claude-mythos-preview, bedrock/.../mythos-preview,
claude-opus-4-5-20251101]
- test_supports_effort_level_handles_provider_prefixes
[bedrock/invoke/us.anthropic.claude-sonnet-4-6-max-True,
claude-sonnet-4-6-max-True]
Add an autouse fixture at tests/test_litellm/llms/anthropic/chat/conftest.py
that monkey-patches litellm.model_cost to the PR-local JSON for every test
in this directory. The parent conftest already snapshots+restores
litellm.model_cost per-function, so the mutation is contained.
This is a scoped workaround. The proper fix is to set the env var
globally in the test workflow once the ~10 inline self-set test files
are audited; tracking that as a follow-up issue.
The chat completion path (`_apply_output_config`) and the /v1/messages
pass-through (`AnthropicMessagesConfig._translate_reasoning_effort_to_anthropic`)
both gate `max` / `xhigh` per model. The two sites had diverged from
near-identical copies into separately maintained blocks, creating a real
drift risk when a new model tier (e.g. Claude 4.8) lands -- a contributor
could update one site and miss the other.
Centralise the gating in `AnthropicConfig._validate_effort_for_model`,
which returns an error message string or `None`. Each call site keeps
its own provider-appropriate exception type (`BadRequestError` for the
chat path, `AnthropicError` for the /v1/messages pass-through) but the
gating decision now comes from one place. Net -11 LOC.
Adds a parametrised unit test exercising the helper directly across
4.5 / 4.6 / 4.7 model families and `max` / `xhigh` / lower-effort
inputs. Existing tests at both call sites continue to pass unchanged.
Addresses Greptile finding on PR #27074.
`azure_ai` is registered in `litellm.openai_compatible_providers`, so
`add_provider_specific_params_to_optional_params` (litellm/utils.py)
auto-stuffs any non-OpenAI kwarg (e.g. `output_config={"effort": "..."}`)
into `optional_params["extra_body"]`. `AzureAnthropicConfig.transform_request`
then strips `extra_body` entirely on the way out, silently dropping the
param — and `AnthropicConfig._apply_output_config` never sees it, so
`effort="invalid"` / `effort="xhigh"` on a non-supporting model
quietly reaches the model with default behavior instead of returning a
clean 400 (as the native `anthropic` provider does).
Promote the keys back to top-level `optional_params` (using `setdefault`
so explicit top-level values win) before delegating to the parent
`AnthropicConfig`. Apply in both `validate_environment` and
`transform_request` so flag detection (`is_mcp_server_used`, etc.) and
output-config validation both run.
Surfaced by the QA matrix expansion on PR #27074: 20 cells where Azure
returned 200 while `anthropic` returned 400 — all `output_config` mode
across haiku_4_5, sonnet_4_5, opus_4_5, sonnet_4_6, opus_4_6, opus_4_7
families with `effort` in {invalid, xhigh, max, low, medium, high}.
Tests:
* `test_output_config_promoted_from_extra_body`: valid effort reaches data
* `test_invalid_output_config_effort_raises_via_extra_body`: 400 on bad effort
* `test_unsupported_effort_xhigh_raises_via_extra_body`: 400 on xhigh-on-Sonnet-4.6
* `test_extra_body_promotion_does_not_clobber_top_level`: setdefault semantics
- Image generation tests patch HTTPHandler.post / get_async_httpx_client so
make_*_azure_httpx_request runs and wire json is asserted on call kwargs.
- Azure image edit: strip model in finalize_image_edit_multipart_data using the
same URL string the handler passes to POST (no second get_complete_url in
transform). BaseImageEditConfig default finalize is a no-op.
Co-authored-by: Cursor <cursoragent@cursor.com>
Azure OpenAI routes image gen/edit by deployment in the URL; sending the
deployment id in model breaks gpt-image-2 (invalid_value). Strip model from
JSON for deployments/.../images/generations and from multipart data for
.../images/edits. Non-deployment URLs (e.g. Azure AI FLUX) unchanged.
Fixes#26316.
Co-authored-by: Cursor <cursoragent@cursor.com>
- claude-sonnet-4-6 + reasoning_effort=max no longer 400s. Renamed
_is_opus_4_6_model to _is_claude_4_6_model at three sites and added
supports_max_reasoning_effort: true to 12 model entries in the JSON
cost map (10 sonnet 4.6 ids + OpenRouter opus 4.6/4.7).
- _map_reasoning_effort now raises BadRequestError(400) directly with
llm_provider, instead of letting Databricks (and similar callers)
surface its raw ValueError as a 500.
- output_config.effort on Opus 4.5 over Bedrock no longer 400s for
missing effort-2025-11-24 beta. Flipped JSON to "effort-2025-11-24"
for bedrock + bedrock_converse and added an auto-attach branch in
_process_tools_and_beta for non-adaptive Anthropic + output_config
on Converse.
- reasoning_effort=xhigh / =max on legacy budget-mode models
(Haiku 4.5, Sonnet 4.5, Opus 4.5) now map to thinking.budget_tokens
8192 / 16384 instead of returning 400. Added two constants in
litellm/constants.py.
Tests updated for all four flips. Validated end-to-end via 306-cell
live proxy matrix (6 model families x 3 routes x 17 effort cases),
all pass.
Replace hardcoded _EFFORT_SUPPORTING_MODEL_PATTERNS with a JSON-backed
check that uses supports_*_reasoning_effort flags from the model map.
Add supports_minimal_reasoning_effort: true to opus-4-5 and mythos-preview
entries (which previously only carried supports_reasoning) so the JSON
remains the single source of truth for effort capability.
When a proxy fronts Claude Code (which always sends `output_config.effort`)
at a pre-4.5 Anthropic model — haiku-3, sonnet-3.5, opus-3, sonnet-4 — the
forwarded knob causes a forced 400 the client can't fix. Gating a strip
behind the existing `drop_params` flag lets operators opt into silent
fixup once and stop worrying about per-model param hygiene.
Default (`drop_params=False`) still forwards and surfaces the provider's
error, preserving the strict, debuggable contract from #27074.
Per https://platform.claude.com/docs/en/build-with-claude/effort the
supporting set is Opus 4.5+, Sonnet 4.6+, and Mythos Preview; everything
else is dropped (with a verbose_logger warning so the strip is visible).
Recognition uses model-name patterns plus a fallback to any
`supports_*_reasoning_effort` flag in the model map for forward
compatibility with new entries.
https://claude.ai/code/session_01WjHq31rvXT6xYNdVmSJvRp
(cherry picked from commit 1233943e78)
The reasoning-effort mapping dict was a public class attribute on
AnthropicConfig, so BaseConfig.get_config returned it as a request
parameter and every Anthropic-backed call (Anthropic / Azure / Vertex /
Bedrock Invoke) hit a 400 'REASONING_EFFORT_TO_OUTPUT_CONFIG_EFFORT:
Extra inputs are not permitted' from the provider. Move the mapping
to a module-level constant.
_supports_effort_level only looked the model up under
custom_llm_provider='anthropic', so bedrock-prefixed model ids
(e.g. bedrock/invoke/us.anthropic.claude-opus-4-7) returned False
for both 'max' and 'xhigh' even when the underlying model entry has
the flag set. Strip known provider prefixes and retry the lookup
against litellm.model_cost directly so per-model gating works on
every route.
Mirror the per-model xhigh/max gate from
AnthropicConfig._apply_output_config in
AnthropicMessagesConfig._translate_reasoning_effort_to_anthropic so
the /v1/messages route also raises a clean 400 instead of forwarding
the unsupported tier.
Closes the remaining QA-sweep gap on PR #27074: Bedrock Invoke
/v1/messages was silently ignoring ``reasoning_effort`` because the
shared param filter only kept native Anthropic keys, so every effort
tier collapsed to the same behavior on the wire (27/231 cells failing
across opus-4-5 / opus-4-6 / sonnet-4-6).
Map ``reasoning_effort`` to native Anthropic ``thinking`` /
``output_config.effort`` at the ``AnthropicMessagesConfig`` layer so
all four /v1/messages routes (direct Anthropic, Azure AI, Vertex AI,
Bedrock Invoke) inherit the same translation:
- Add ``reasoning_effort`` to ``AnthropicMessagesRequestOptionalParams``
so the param filter in
``AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param``
no longer drops it before the transformation runs.
- Add ``_translate_reasoning_effort_to_anthropic`` and call it from
``transform_anthropic_messages_request``. Mirrors
``AnthropicConfig.map_openai_params`` on the chat completion path
(re-uses ``_map_reasoning_effort`` and
``REASONING_EFFORT_TO_OUTPUT_CONFIG_EFFORT``) so the two routes
cannot drift. Pops ``reasoning_effort`` so it never reaches the wire.
- Caller-supplied native ``thinking`` / ``output_config.effort`` always
win — same precedence as
``_translate_legacy_thinking_for_adaptive_model``.
- Garbage values (``""``, ``"disabled"``, ``"invalid"``) raise
``AnthropicError(status_code=400)`` instead of falling through and
surfacing as 500s from the provider.
- ``"none"`` clears thinking + output_config so callers can opt out
per request.
Also restores the non-adaptive-model test coverage on Bedrock Invoke
/v1/messages that the previous commit lost when
``test_bedrock_messages_strips_output_config`` was renamed to the
``forwards`` variant on Opus 4.7.
Adds a new test file
``test_reasoning_effort_translation.py`` covering the translation at
the shared config level (adaptive + non-adaptive models, none, garbage,
caller precedence) so all four /v1/messages routes are exercised by a
single suite.
Adds parametrized + behavioral tests on the Bedrock Invoke /v1/messages
suite covering: minimal/low/medium/high/xhigh/max mapping for adaptive
models, thinking-budget mapping for non-adaptive Opus 4.5, ``none``
clears both, garbage raises 400, explicit ``output_config`` wins.
Refs: https://github.com/BerriAI/litellm/pull/27074
Follow-up bugs surfaced by the QA sweep on PR #27039
(https://github.com/BerriAI/litellm/pull/27039#issuecomment-4363363610).
1. Stop stripping output_config.effort on Bedrock + Vertex adaptive routes.
- Vertex AI Claude 4.6/4.7 accepts output_config.effort on rawPredict
(verified end-to-end against us-east5 / global). The strip helper now
no-ops for effort.
- Bedrock Converse routes output_config into additionalModelRequestFields
for anthropic base models so the requested adaptive tier (low/medium/
high/xhigh/max) actually reaches the wire instead of all collapsing to
identical thinking.
- Bedrock Invoke chat transformation (AmazonAnthropicClaudeConfig) stops
popping output_config from the post-AnthropicConfig request body.
- Bedrock Invoke /v1/messages allowlist (BedrockInvokeAnthropicMessagesRequest)
now lists output_config so the runtime allowlist filter forwards it.
2. Validate effort across Bedrock Converse so 'disabled' / 'invalid' / '' /
unsupported tiers (xhigh/max on Sonnet 4.6 or budget-mode 4.5 models)
surface as a clean 400 BadRequestError instead of 500.
3. ValueError -> BadRequestError throughout (AnthropicConfig.map_openai_params,
_apply_output_config, AmazonConverseConfig._handle_reasoning_effort_parameter).
Empty-string effort is now rejected (was silently passing the
'if effort and ...' short-circuit).
4. Floor reasoning_effort='minimal' at the Anthropic provider minimum
(1024 budget_tokens) via new ANTHROPIC_MIN_THINKING_BUDGET_TOKENS so it's
a usable tier on direct Anthropic / Azure AI Anthropic / Vertex AI Anthropic /
Bedrock Invoke (all of which 400 below 1024).
5. model_prices: dedupe duplicate supports_max_reasoning_effort key on
claude-opus-4-7 / claude-opus-4-7-20260416.
Adds regression tests across all five affected paths; existing tests asserting
the silent-strip behavior were updated to reflect the new pass-through and
clean 400 surfaces.
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
The Vertex batch output transformer was emitting both a populated 'response' and 'error' for failed batch entries. The OpenAI Batch output spec defines them as mutually exclusive: on error 'response' MUST be null. This broke any consumer using 'result["response"] is None' to detect failures.
Setting reasoning_effort="none" on Anthropic chat models (direct, Bedrock
Invoke, Bedrock Converse, Vertex AI Anthropic, Azure AI Anthropic) crashed
LiteLLM with:
litellm.APIConnectionError: 'NoneType' object has no attribute 'get'
Both the Anthropic chat transformation and Bedrock Converse called
``AnthropicConfig._map_reasoning_effort`` and assigned the ``None`` it returns
for ``"none"`` directly to ``optional_params["thinking"]``. Downstream
``is_thinking_enabled`` then did ``optional_params["thinking"].get("type")``
and crashed.
Pop ``thinking`` (and on Claude 4.6/4.7, ``output_config``) instead of
assigning ``None``, restoring the documented contract that
``reasoning_effort="none"`` means "do not enable thinking". This also
prevents downstream Anthropic 400s ("thinking: Input should be an object",
"output_config.effort: Input should be ...") if the bug were ever masked.
Verified end-to-end against the live Anthropic API and Bedrock Converse
on claude-opus-4-{5,6,7} and claude-sonnet-4-6, plus Bedrock Invoke for
Claude 4.5/4.6. Vertex AI Anthropic and Azure AI Anthropic inherit the
fixed ``map_openai_params`` from ``AnthropicConfig`` and need no further
changes.
Adds litellm.disable_vertex_batch_output_transformation (default False).
When True, afile_content returns raw Vertex predictions.jsonl untouched
so users that parse candidates/modelVersion directly are not broken.
The method was overwriting logging_obj.optional_params, logging_obj.model,
and logging_obj.start_time on the caller's Logging instance. When invoked
from llm_http_handler.py's generic framework path, the framework's own
logging_obj (which already went through pre_call) had its properties
clobbered, causing model and start_time to reflect the last batch line's
values rather than the original call context.
Fix: create a fresh local Logging instance for the per-line transformation
instead of mutating the incoming logging_obj. The caller's object is now
left entirely untouched regardless of whether a logging_obj was passed in
or not.
Regression tests added to verify model, start_time, and optional_params
are not mutated on the caller's logging_obj.
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
Container ownership and skill rows are looked up on every retrieve /
delete / list / file-content / chat-completion-with-skill call. The new
stores wrapped raw Prisma queries with no cache, putting one DB
round-trip on each request. Add an in-process TTL'd cache mirroring the
_byok_cred_cache pattern in mcp_server/server.py: per-key (value,
monotonic_timestamp), 60s TTL, 10000-entry cap with full-clear on
overflow, invalidated by every write. Negative results (`None`) are
cached too so untracked-resource checks also skip the DB.
Tests cover: cache-after-first-hit, negative caching, write
invalidation, no-caching-on-DB-error, TTL expiry, capacity eviction.
56 tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
