GPT-5 / GPT-5-mini / GPT-5-nano / GPT-5.5 on OCI reject "maxTokens"
with HTTP 400:
Invalid 'maxTokens': Unsupported parameter: 'maxTokens' is not
supported with this model. Use 'maxCompletionTokens' instead.
(Same convention as OpenAI's reasoning-API contract.)
Add a model-aware rename in OCIChatConfig._get_optional_params so the
request payload uses maxCompletionTokens when the model id starts with
openai.gpt-5. Regular Llama / Cohere / Gemini / GPT-4.x continue to use
maxTokens unchanged.
Also widen OCIChatRequestPayload to carry the new optional field so it
survives Pydantic serialization.
Verified live against OCI us-chicago-1:
- openai.gpt-5, gpt-5-mini, gpt-5-nano, gpt-5.5 all return 200
- Full feature sweep on gpt-5.5 (basic, system, multi-turn, streaming,
tools, usage) all green
- meta.llama-3.3-70b-instruct still uses maxTokens (no regression)
4 new unit tests cover the helper, the routing in both pre- and
post-translation states, and Pydantic serialization.
- types/llms/oci.py: add "TOOL_CALL" to CohereChatResponse.finishReason
Literal so Pydantic does not raise ValidationError on non-streaming
Cohere tool-use calls (Greptile P1)
- test_oci_cohere_tool_calls.py: add test covering TOOL_CALL finish reason
- model_prices_and_context_window.json: remove 6 duplicate oci/cohere.embed-*
keys that were silently overridden by the more complete entries already
present in the file (Greptile P1)
- common_utils.py: move OCI_API_VERSION here from chat/transformation.py
so embed/transformation.py does not need to import chat/transformation;
change Protocol stub body from ... to pass (CodeQL "statement no effect");
add comment to sha256_base64 clarifying it implements OCI HTTP signing
spec, not password hashing (CodeQL false positive)
- chat/transformation.py: import CustomStreamWrapper from
litellm_core_utils.streaming_handler instead of litellm.utils to reduce
import cycle depth (CodeQL cyclic import)
- chat/cohere.py, chat/generic.py: import Usage and
ChatCompletionMessageToolCall from litellm.types.utils instead of
litellm.utils for the same reason
- embed/transformation.py: import OCI_API_VERSION from common_utils
instead of chat/transformation (removes the embed→chat import edge)
- Remove _extract_text_content duplication — single definition in cohere.py,
imported where needed; instance method on OCIChatConfig eliminated
- Move cryptography imports to module level with _CRYPTOGRAPHY_AVAILABLE flag
and _require_cryptography() guard; no more re-import on every signing call
- Move litellm version import to module level via litellm._version; remove
inline import inside validate_oci_environment
- sign_with_manual_credentials now returns Tuple[dict, bytes] matching
sign_with_oci_signer — asymmetry eliminated, Optional[bytes] guards removed
throughout stream wrappers (signed_json_body: bytes = b"")
- Rename _openai_to_oci_cohere_param_map → openai_to_oci_cohere_param_map
for consistency with openai_to_oci_generic_param_map
- Remove double-key bug in map_openai_params where responseFormat was stored
under both OCI and OpenAI key names simultaneously
- Remove delegating shims (adapt_messages_to_cohere_standard,
adapt_tool_definitions_to_cohere_standard, _handle_generic_stream_chunk)
from OCIChatConfig/OCIStreamWrapper; tests now import directly from
cohere.py and generic.py where symbols live
- Trim __all__ to 7 genuine public symbols; remove the 13-symbol list that
existed only to support test imports
- Collapse per-model integration test classes into pytest.mark.parametrize;
CHAT_MODELS list is the single source of truth for model-specific config
- Black + Ruff clean across all OCI files
- Add resolve_oci_schema_refs: inline $ref/$defs — OCI rejects JSON Schema refs
- Add resolve_oci_schema_anyof: flatten Optional[T] anyOf (Pydantic v2 emits these)
- Add sanitize_oci_schema: strip title, normalise null types, ensure array items
- Add OCI_JSON_TO_PYTHON_TYPES: Cohere expects Python type names (str/int/float),
not JSON Schema names (string/integer/number)
- Add enrich_cohere_param_description: embed enum/format/range/pattern constraints
into description since CohereParameterDefinition has no dedicated fields
- Apply all of the above in adapt_tool_definitions_to_cohere_standard and
adapt_tool_definition_to_oci_standard
- Fix toolChoice conversion: map OpenAI string ('auto','none','required') to OCI
dict form ({"type":"AUTO"} etc.) — the API rejects plain strings
- Update unit test expectations to match correct Python type names and enriched
descriptions
Addresses greptile P1: tests/test_litellm/ is for mock-only unit tests
(make test-unit target). Real-network OCI tests now live in the correct
location alongside other provider integration tests.
- Bug 1: sync and async streaming paths now use signed_json_body when provided
instead of re-serializing data with json.dumps() — the OCI RSA-SHA256 signature
covers the exact request body bytes, so re-serializing produces an invalid sig
- Bug 3: Cohere stop sequences now map to 'stopSequences' (was incorrectly 'stop')
- Bug 4: removed hardcoded Cohere defaults (maxTokens=600, temperature=1, topK=0,
topP=0.75, frequencyPenalty=0) that silently overrode user intent on every call
- Added 6 unit tests covering all three fixes
- Add OCIEmbedConfig with full Cohere embed support (7 models, batch up to 96)
- Fix sync streaming: split SSE events on \n\n before JSON parsing
- Fix reasoning models (Gemini 2.5, xAI Grok): make completionTokens and message
optional in OCIResponseChoice to handle max_tokens exhausted on reasoning
- Fix compartment_id resolution in chat transform to use resolve_oci_credentials
- Fix tool call id: make OCIToolCall.id optional, generate UUID fallback for
providers (Google via OCI) that omit it
- Add OCI_KEY env var support for inline PEM keys
- Fix datetime.utcnow() deprecation in request signing
- Expand model catalog: 29 OCI models including Llama 4, Gemini 2.5, xAI Grok,
Cohere Command A, and all Cohere embed variants
- Add 37 live integration tests: sync/async completions for Meta/Google/xAI/Cohere,
sync/async embeddings, tool use across all vendors, streaming, env var auth
- Add 23 embed unit tests covering all transform and validation paths
- router: raise ServiceUnavailableError (503) instead of RouterRateLimitErrorBasic (429)
when a specifically-addressed deployment is administratively blocked; 429 misleads
retry-enabled clients into spinning forever against a paused model
- proxy_server: compute get_fully_blocked_model_names() once before both branches in
model_list() instead of duplicating the call in each branch
- deepseek: upgrade silent debug log to warning when injecting placeholder
reasoning_content so callers are clearly notified of degraded multi-turn quality
- tests: update two blocked-deployment assertions to expect ServiceUnavailableError
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: reset org and tag budgets (#27326)
* reset org budgets
* reset tag budgets
---------
Co-authored-by: Michael Riad Zaky <michaelr@Mac.localdomain>
* fix(ui): omit allowed_routes from key edit save when unchanged (#27553)
* fix(ui): omit allowed_routes from key edit save when unchanged
When a team admin opens Edit Settings on a key with key_type=AI APIs and
saves without changing anything, the UI re-sends the existing allowed_routes
value, which the backend's _check_allowed_routes_caller_permission gate
rejects for non-proxy-admins (LIT-2681).
Strip allowed_routes from the patch in handleSubmit when it deep-equals the
original keyData.allowed_routes. The backend treats absence as "leave alone,"
so no-op saves now succeed for non-admins. Admins explicitly editing the
field still send the new value.
* fix(ui): order-insensitive allowed_routes diff + cover null-original case
Address Greptile review:
- Switch the "is allowed_routes unchanged" check to a Set-based comparison so
a server-side reorder of the array doesn't register as a user edit and
re-trigger LIT-2681.
- Add two regression tests: (1) keyData.allowed_routes is null and the form
is untouched — patch should strip the field; (2) server returned routes in
a different order than the user originally entered — patch should still
recognize the value as unchanged.
* chore(ui): strip ticket refs and tighten comments in key edit fix
- Remove internal-tracker references from in-code comments
- Tighten the WHY comment in handleSubmit to two lines
- Drop redundant test-block comments — test names already describe the case
* fix(ui): annotate Set<string> generic in allowed_routes diff to fix tsc
* fix(guardrails): return HTTP 400 instead of 500 for guardrail-blocked requests
GuardrailRaisedException and BlockedPiiEntityError both lacked a
status_code attribute. When these exceptions reached the proxy
exception handler (getattr(e, 'status_code', 500)), the fallback
defaulted to HTTP 500 — making intentional guardrail blocks
indistinguishable from server errors and causing unnecessary client
retries.
Changes:
- Add status_code=400 (keyword-only) to GuardrailRaisedException
- Add status_code=400 (keyword-only) to BlockedPiiEntityError
- Update _is_guardrail_intervention() to recognize both exceptions
so downstream loggers record 'guardrail_intervened' instead of
'guardrail_failed_to_respond'
- Add 6 unit tests for default/custom status codes and getattr pattern
- Strengthen existing blocked-action test with status_code assertion
Fixes#24348
---------
Co-authored-by: Michael-RZ-Berri <michael@berri.ai>
Co-authored-by: Michael Riad Zaky <michaelr@Mac.localdomain>
Co-authored-by: ryan-crabbe-berri <ryan@berri.ai>
Co-authored-by: Krrish Dholakia <krrish+github@berri.ai>
* fix(mcp-oauth): add PROXY_BASE_URL escape hatch + diagnostic logging for invalid_request
Customers hitting "{"detail":"invalid_request"}" on the MCP /authorize
endpoint had no way to recover when their ingress mangles X-Forwarded-*
headers (the same-origin check in validate_trusted_redirect_uri compares
the browser-supplied redirect_uri against get_request_base_url, which is
reconstructed from those headers).
Two contained changes:
1. get_request_base_url now honours PROXY_BASE_URL as the canonical
public origin when set, bypassing the X-Forwarded-* trust gate
entirely. Operators who know their public URL can set it once
instead of debugging ingress header rewrites.
2. The rejection path in validate_trusted_redirect_uri emits a WARN
log carrying the redirect_uri, computed proxy base, and the
X-Forwarded-* / Host headers seen. A bare 400 was undiagnosable;
this turns it into a one-line root-cause.
* test(mcp-oauth): capture warnings from correct logger ("LiteLLM")
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* fix(mcp-oauth): reject malformed PROXY_BASE_URL with one-shot diagnostic
A scheme-less PROXY_BASE_URL (e.g. "litellm.example.com" instead of
"https://litellm.example.com") would sail through urlparse with empty
scheme + netloc, silently breaking every same-origin compare in
validate_trusted_redirect_uri and leaving the operator staring at the
same opaque 400 the env var was meant to fix.
Validate it once at read time: only honour values that parse as
http(s) URLs with a non-empty netloc; otherwise log a one-shot WARN
naming the bad value and fall through to the request-derived origin
so the proxy still serves traffic.
* fix(mcp/oauth): normalize PROXY_BASE_URL to strip query/fragment
Match the X-Forwarded-* path's normalization so a configured
PROXY_BASE_URL containing a query string or fragment does not break
downstream f-string concatenation like f"{base_url}/callback".
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* refactor(mcp-oauth): drop non-essential comments from PROXY_BASE_URL changes
Strip narrative comments and verbose docstrings added in this PR; the
code is intuitive enough on its own and the log messages already carry
their own diagnostic context. Pre-existing comments are left untouched.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* fix(proxy): sort BYOK models by team_public_model_name in /v2/model/info
Team BYOK rows persist an internal `model_name` like
`model_name_{team_id}_{uuid}` and expose the user-facing name via
`model_info.team_public_model_name`. The UI's `getDisplayModelName`
and the search filter already fall back to that field, but
`_sort_models` was keying off the raw `model_name` — so BYOK rows
ranked by their opaque IDs and clumped at the end of the alphabetized
list instead of interleaving with non-BYOK rows.
Match the UI/search behavior: prefer `team_public_model_name` when
present, fall back to `model_name` otherwise.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(proxy): case-insensitive DB-side search for BYOK models
`_apply_search_filter_to_models` used Prisma's JSON path
`string_contains` to match the BYOK `team_public_model_name` field, but
that operator is case-sensitive in Postgres (no `mode: insensitive`
flag like column-level string filters have). So a search for "claude"
missed a stored "Claude Sonnet" via the DB branch even though the
router-side path matched it case-insensitively.
Widen the JSON branch to "row has a team_public_model_name set" and
filter case-insensitively in Python so DB-only BYOK rows match the
same terms users see in the UI. This also drops the now-unused
DB-level page-size optimization and `sort_by` knob — the in-Python
filter is the source of truth for `db_models_total_count` now.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(proxy): scope BYOK search results to caller's accessible teams
`_apply_search_filter_to_models` was widened to fetch every row with a
`team_public_model_name` set so case-insensitive search could match
mixed-case stored names. `/v2/model/info` is reachable by non-admin
keys though, and the helper ran before `include_team_models` / `teamId`
filtering — so a non-admin caller could search a common substring like
"claude" and see BYOK rows belonging to teams they're not a member of.
Resolve the caller's team membership once (admin → no scoping, else
their `user_row.teams`) and drop BYOK rows (those with
`model_info.team_id` set) outside that scope on both the router-side
matches and the over-broad DB query, before display-name matching.
Non-team rows are unaffected and remain gated by the existing
`include_team_models` / `direct_access` paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(proxy): search by team_public_model_name and scope teamId queries
- /v2/model/info search now matches both `model_name` and
`model_info.team_public_model_name`, so team BYOK rows (which persist
an internal `model_name_{team_id}_{uuid}`) are findable by the public
name shown in the UI. DB query OR-includes a JSON-path match on
`team_public_model_name` for rows that exist only in the DB.
- `_filter_models_by_team_id` no longer short-circuits on the viewer's
`direct_access` flag — that describes the admin viewer's own
permissions and would leak every public model into a team-scoped view.
Models are kept only when they belong to the team (own BYOK, in
access_via_team_ids, or reachable via team.models / access groups).
- Added `_authorize_team_id_query`: the untrusted `teamId` query
parameter now requires the caller to be a proxy admin or a member of
the requested team, otherwise returns 403. Without this, any
authenticated user could enumerate another team's BYOK metadata by
guessing the team id.
- `_get_caller_byok_team_scope` now treats `PROXY_ADMIN_VIEW_ONLY` the
same as `PROXY_ADMIN` (both are admin roles); previously VIEW_ONLY
admins fell through to a user-id team lookup and saw only their own
teams' BYOK rows.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(proxy): bound BYOK search DB fetch in /v2/model/info
Previously the DB-side search OR'd a JSON-path predicate
`{model_info: {path: [team_public_model_name], string_contains: ""}}`
to compensate for Prisma's case-sensitive JSON `string_contains` on
Postgres. That predicate matches every row that has any
`team_public_model_name` set, so any authenticated caller could force a
full BYOK-table read with `/v2/model/info?search=x` regardless of page
size.
Drop the JSON-path branch. The DB query now does a bounded
`model_name contains <search>` lookup. BYOK rows that are loaded into
the router are still searchable by their `team_public_model_name` via
the router-side filter; only the rare edge case of a BYOK row that
exists only in the DB (router sync failed) loses display-name search,
which is an acceptable trade-off given the DoS surface.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix(proxy): bound DB find_many in /v2/model/info search
The previous bounding patch dropped the page-aware `take=N` on
`find_many`, so a broad `?search=model` would load and decrypt every
matching DB row on each request even though the response only returns
one page.
Restore bounded fetches in `_apply_search_filter_to_models`:
* Unsorted searches use `take = max(0, page * size - router_count)`,
i.e. exactly one page worth of remaining DB rows.
* Sorted searches need ordering across the full match set, so they cap
at `_SORTED_SEARCH_DB_FETCH_CAP = 500` instead of fetching everything.
* Total count comes from a cheap `count(...)` query so pagination stays
accurate without materializing every row.
Wired `page`, `size`, and `sortBy` through from the endpoint and added
a regression test covering both `take` values.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor(proxy): extract DB-fetch helper to satisfy PLR0915
_apply_search_filter_to_models tripped Ruff's "too many statements"
(51 > 50) after the bounded-fetch fix. Move the DB-side block into
`_fetch_db_models_for_search`, which keeps the same behavior:
* Bounded `take` via page math (unsorted) or `_SORTED_SEARCH_DB_FETCH_CAP`
(sorted)
* Cheap `count(...)` for accurate pagination totals
* Caller-team scope applied to fetched rows before decrypt
Pure refactor; no behavior change. All 8 BYOK/team tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* style: apply black formatting to _fetch_db_models_for_search
CI's "Check Black formatting" step flagged one line in the helper added
in d55eecf6af. No behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The proxy SERVER span ("Received Proxy Server Request") only carried
http.response.status_code on failures (set in _record_exception_on_span),
so success traces had no 2xx bucket — error-ratio and status-breakdown
dashboards were missing their denominator and the span violated the HTTP
semconv (the attribute is required whenever a response is sent). Add a
set_response_status_code_attribute helper and call it from
async_post_call_success_hook with 200, symmetric with the failure path
and the existing route/preprocessing-duration SERVER-span attributes.
* feat(otel): expose http.response.status_code on failure spans
Set the OTel-standard http.response.status_code (integer) on failure
spans alongside the existing OpenInference error.code (kept for
back-compat). error.type is already emitted via ERROR_TYPE.
Crucially, also record structured error attributes on the proxy SERVER
span ('Received Proxy Server Request') from async_post_call_failure_hook
- the only place the SERVER span is in hand. _handle_failure records on
the litellm_request child span (the parent span is not propagated into
its kwargs), so prior to this change the SERVER span that dashboards
query carried only span status, never error.code/error.type. Reuses
_record_exception_on_span + StandardLoggingPayloadSetup.get_error_information
so values match the child span.
Tests: recorder unit coverage + a hook-driven test asserting the SERVER
span is stamped (the gap recorder-only tests missed). Full
test_opentelemetry.py suite: 197 passed.
* feat(otel): set http.route + url.path on the proxy SERVER span
Add the OTel-standard http.route (low-cardinality route template, e.g.
/v1/threads/{thread_id}/runs) and url.path (literal path) to the SERVER
span ('Received Proxy Server Request') so dashboards can group traffic
by endpoint instead of seeing every path param as a unique value.
Same architectural gap as the status-code commit: the success/failure
logging handlers write the litellm_request CHILD span, and
_handle_success explicitly refuses to copy to the SERVER span. Verified
with a console-exporter run that the SERVER span was bare on success.
Unlike error info, route/path are known at request time, so set them
directly on the freshly-created SERVER span in user_api_key_auth (one
edit point, works for success and failure, no hook-ordering risk):
- http.route from the matched FastAPI route (scope['route'].path),
empirically confirmed populated at auth-dependency time.
- url.path from the existing literal-path variable.
New get_request_route_template helper + set_proxy_request_route_attributes
(no-op on None span, so the Langfuse override stays safe).
Tests: route-attribute setter + route-template helper edges. Full
test_opentelemetry.py and test_auth_utils.py green.
* feat(otel): set litellm.preprocessing.duration_ms on the proxy SERVER span
Expose the total time LiteLLM spends before the upstream provider
request begins (auth + parsing + pre-call hooks) as a single number on
the SERVER span ('Received Proxy Server Request'). Window:
proxy-receive -> FIRST provider handoff.
Retry semantics: first attempt only (pure preprocessing, excludes
retry loops + backoff). api_call_start_time is overwritten on every
attempt, so a set-once first_api_call_start_time pins the first handoff.
Same architectural gap as the prior two commits: the success/failure
logging handlers write the litellm_request CHILD span, not the SERVER
span. Set it instead from the post-call hooks on
user_api_key_dict.parent_otel_span.
Failure-path subtlety: request_data.pop('litellm_logging_obj') runs
before the failure-hook loop, so the failure hook can't read the
logging object. litellm_received_at is propagated via the existing
request->metadata channel, and first_api_call_start_time is mirrored
onto litellm_params.metadata, so both anchors survive into request_data
and the OTel helper reads them uniformly for success and failure.
Edits: user_api_key_auth (stash receive instant), litellm_pre_call_utils
(propagate it), litellm_logging (set-once first handoff + metadata
mirror), opentelemetry (constant + set_preprocessing_duration_attribute,
called from both post-call hooks).
Tests: duration helper (both container shapes, missing/negative/None
edges) + set-once invariant (retry doesn't overwrite, metadata mirror).
test_opentelemetry.py + test_auth_utils.py + test_litellm_logging.py:
447 passed. Verified live: SERVER span carries the attribute on success
and failure, coexisting with the status-code and route attributes.
* fix(otel): MyPy type-narrowing for status-code + preprocessing-duration
No behavior change. MyPy (CI lint) flagged:
- error_information["error_code"] is str|None: narrow via a None-checked
local before int().
- _to_timestamp returns Optional[float]: resolve both anchors and return
early if either is None instead of subtracting possibly-None floats.
* fix(otel): stop polluting user request metadata with first_api_call_start_time
The PR3 set-once preprocessing anchor was mirrored into
litellm_params["metadata"] from core litellm_logging.py. That dict is
the caller's request metadata, mutated in place and shared across every
call path including pure SDK (litellm.acreate_batch). It got echoed into
LiteLLMBatch(metadata=...), which the OpenAI batch schema types as
Dict[str, str] -> pydantic ValidationError on a datetime value.
- litellm_logging.py: set first_api_call_start_time only on
model_call_details (success path reads it there directly).
- proxy/utils.py: post_call_failure_hook lifts it off the logging object
into request_data (internal top-level key, same convention as the
other proxy-internal request_data keys) right before the existing
litellm_logging_obj pop. Never touches user metadata.
- opentelemetry.py: read the anchor from the container top level
(model_call_details on success, request_data on failure).
- Tests updated; add TestPostCallFailureHookLiftsFirstApiCallStartTime.
Fixes the batches_testing regression introduced on this branch.
* chore(otel): trim verbose comments to concise rationale
Collapse multi-line why-blocks to one or two lines and drop process/plan references (PR-numbering, "the plan") from test comments. No behavior change.
Split the monolithic LiteLLM proxy into independently scalable Kubernetes components to allow separate horizontal scaling of the LLM data plane and management API surfaces
- Add DatabaseURLSettings pydantic-settings model that assembles DATABASE_URL (and optional DATABASE_URL_READ_REPLICA) from discrete DATABASE_* env vars before Prisma initializes, supporting both IAM token auth (minting short-lived RDS tokens) and password auth; replaces the CLI-only path that componentized entrypoints bypass
- Add gateway component (port 4000) that trims the proxy route table to the LLM data-plane surface (chat, embeddings, completions, audio, realtime, provider passthroughs, health/metrics) via an allowlist applied inside the lifespan context so plugin-registered routes are captured
- Add backend component (port 4001) that exposes the management/admin surface (keys, users, teams, orgs, spend analytics, model management, SSO, audit logs) with a complementary allowlist
- Add ui component — Next.js static export served by nginx (port 3000) with RSC payload routing, asset prefix aliasing, and SPA fallback for dashboard routes
- Add migrations component with dedicated Dockerfile that runs prisma migrate deploy via a Helm pre-install/pre-upgrade Job, eliminating per-pod schema contention on the Prisma advisory lock
- Add Helm chart (helm/litellm) with separate Deployments, Services, HPAs, and ConfigMap for each component; shared _helpers.tpl emits DATABASE_*, IAM_TOKEN_DB_AUTH, REDIS_*, and DISABLE_SCHEMA_UPDATE env vars from chart values; ingress template routes traffic to the correct component by path prefix
- Add comprehensive tests for DatabaseURLSettings covering IAM auth, password auth, read replica fallbacks, operator-pinned URL preservation, and percent-encoding; add coverage test asserting gateway + backend allowlist union equals the full proxy route set
- Add pydantic-settings>=2.14.1 as a proxy extra dependency and update liccheck allowlist
Co-authored-by: Yassin Kortam <yassinkortam@g.ucla.edu>
Google restructured the live spec at ai.google.dev/static/api/
interactions.openapi.json: the output-only fields (notably the
`steps` array, formerly `outputs`) moved off the request schema
`CreateModelInteractionParams` onto a dedicated `Interaction`
response schema. The response-side tests still read from the
request schema, so `test_interaction_response_fields` failed with
"Output field 'steps' not in spec".
Point `test_interaction_response_fields` and `test_status_enum_values`
at the `Interaction` schema (the semantic response object; all output
fields incl. `steps` present there). Request-side tests keep using
`CreateModelInteractionParams` (all request fields verified still
present). 13/13 pass against the current live spec.
- Introduce `OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental` opt-in that switches OTEL traces to conform with the OpenTelemetry GenAI semantic conventions specification
- Extract all semconv behavior into a new `OTELGenAISemconvMixin` class in `gen_ai_semconv.py`, mixed into `OpenTelemetry` to keep concerns separated
- In semconv mode, span name follows `{operation} {model}` pattern (e.g. `chat gpt-4`) and span kind is set to `CLIENT` instead of legacy `litellm_request`
- Replace `gen_ai.system` with `gen_ai.provider.name` and drop `llm.is_streaming` in semconv mode; add `gen_ai.request.{frequency_penalty,presence_penalty,top_k,seed,stop_sequences,stream,choice.count}` and `gen_ai.usage.cache_{creation,read}.input_tokens` attributes
- Replace per-message `gen_ai.content.prompt` / per-choice `gen_ai.content.completion` log events with a single consolidated `gen_ai.client.inference.operation.details` event; omit `gen_ai.input/output.messages` when content capture is disabled
- Suppress the non-standard `raw_gen_ai_request` child span entirely in semconv mode
- Support both programmatic (`OpenTelemetryConfig.semconv_stability_opt_in` field) and environment variable activation; the two sources are unioned so either or both can enable the opt-in
- Extract OTEL SDK `LogRecord` / `SeverityNumber` version-compatibility shim into a reusable `_otel_log_types()` static method to deduplicate the `< 1.39.0` / `>= 1.39.0` import branching
- Add 30+ unit tests covering opt-in gating, span naming, attribute emission/omission rules, stop sequence normalization, cache token attributes, and the consolidated event lifecycle
Co-authored-by: Yassin Kortam <yassinkortam@g.ucla.edu>
test_keepalive_timeout_flag and test_timeout_worker_healthcheck_flag
were the only run_server tests in test_proxy_cli.py that neither
stripped DATABASE_URL/DIRECT_URL nor mocked the prisma DB path. When a
DATABASE_URL is present (CI/env leak), run_server --local enters the DB
block and blocks in the un-timeout'd subprocess.run(["prisma"]) at
proxy_cli.py:987 plus the ProxyExtrasDBManager migrate-deploy retry
loops, ~370s per test on the CI runner. --dist=loadscope pins both to
one xdist worker, so the proxy-infra job appears stuck at 99% and hits
the 20-min timeout.
Apply the same isolation every other run_server test in this file
already uses: mock PrismaManager.setup_database +
should_update_prisma_schema and strip DATABASE_URL/DIRECT_URL. Full
module drops from 31.7s to 2.9s locally; both tests fall off the slow
list.
Adds test_update_config_env_var_round_trip_not_double_encrypted, which
drives the real /config/update handler: first write plaintext, then
re-POST the stored ciphertext (the Admin UI round-trip) and assert the
value is not stacked with a second encryption layer and untouched keys
stay byte-identical. Verified to fail against the pre-fix handler and
pass after. Also tightens the unit test to exactly three ciphertext
re-feeds.
A single decrypt-then-encrypt chokepoint (_encrypt_env_variables_for_db)
now backs both update_config and save_config. Re-submitting a value the
Admin UI read back from /get/config/callbacks as ciphertext no longer
stacks a second encryption layer, which previously decrypted to garbage
and silently broke the callback. The chokepoint decrypts with the pure
_decrypt_db_variables (no os.environ mutation on the write path) and
encrypts exactly once; update_config merges only the sent keys so
untouched env vars keep their stored ciphertext byte-for-byte.
* fix(bedrock-mantle): use /anthropic/v1/messages path for Mantle endpoint (#27943)
* docs: add one-line docstring to _disable_debugging (#27894)
Squash-merged by litellm-agent from oss-agent-shin's PR.
* Add jp. Bedrock cross-region inference profile for claude-sonnet-4-6 (#27831)
Squash-merged by litellm-agent from Cyberfilo's PR.
* Sanitize empty text content blocks on /v1/messages (#27832)
Squash-merged by litellm-agent from Cyberfilo's PR.
* fix(bedrock-mantle): use /anthropic/v1/messages path for Mantle endpoint
The bedrock-mantle gateway (Claude Mythos Preview) serves the Anthropic
Messages API at /anthropic/v1/messages; /v1/messages returns 404 Not
Found. Both AmazonMantleConfig (chat/completions caller route) and
AmazonMantleMessagesConfig (anthropic-messages caller route) hardcoded
the wrong path, so every Mantle request 404'd before reaching the model.
Per the Anthropic docs: "[Claude in Amazon Bedrock] uses the Messages
API at /anthropic/v1/messages with SSE streaming."
https://platform.claude.com/docs/en/api/claude-on-amazon-bedrock
Confirmed independently against the live endpoint:
/v1/chat/completions -> 200 OK
/v1/messages -> 404 Not Found (what litellm used)
/anthropic/v1/messages -> 200 OK (Claude only)
Adds a regression test asserting both Mantle configs build the
/anthropic/v1/messages path, and updates the existing assertions that
encoded the wrong path.
---------
Co-authored-by: oss-agent-shin <ext-agent-shin@berri.ai>
Co-authored-by: Filippo Menghi <113345637+Cyberfilo@users.noreply.github.com>
* fix: sanitize empty text blocks in sync anthropic_messages_handler path
Co-authored-by: Yassin Kortam <yassin@berri.ai>
---------
Co-authored-by: João Costa <13508071+jpv-costa@users.noreply.github.com>
Co-authored-by: oss-agent-shin <ext-agent-shin@berri.ai>
Co-authored-by: Filippo Menghi <113345637+Cyberfilo@users.noreply.github.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* Feat: Add Weighted-Routing Failover
* test(router): cover weighted failover helper functions
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(router): align weighted failover deployment list type with mypy
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(router): address greptile review on weighted failover
- Narrow exception swallowing in `_maybe_run_weighted_failover` to
`openai.APIError` so model failures defer to the regular fallback
while programming bugs (AttributeError/KeyError/TypeError) surface.
- Note async-only limitation of `enable_weighted_failover` in the
Router constructor docstring.
- Make the weighted distribution test less flaky (1000 iterations,
looser bound) and make the non-simple-shuffle test deterministic by
failing both deployments instead of relying on the latency strategy's
first pick.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(router): ensure weighted failover metadata persists in kwargs
The previous `kwargs.setdefault(metadata_variable_name, {}) or {}` returned
a brand-new dict whenever the existing metadata was falsy (empty dict or
None), so writes to `_failover_excluded_ids` never made it back into
`kwargs`. Multi-hop weighted failover then re-selected previously failed
deployments and exhausted `max_fallbacks` prematurely.
Explicitly assign a fresh dict into kwargs when metadata is missing so
mutations are visible to subsequent failover hops.
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* test(router): regression for weighted failover metadata persistence
Asserts kwargs["metadata"]["_failover_excluded_ids"] is populated after
_maybe_run_weighted_failover, proving the metadata dict written by the
helper is the same object that lives in kwargs (no disconnected copy).
Pairs with the prior fix that replaced `setdefault(..., {}) or {}` with
an explicit get/assign so writes survive across hops.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(router): harden weighted failover error/state handling
- Catch RouterRateLimitError (ValueError) alongside openai.APIError in
_maybe_run_weighted_failover so an exhausted intra-group retry falls
through to the regular cross-group fallback path instead of bubbling
out and bypassing configured fallbacks.
- Stop mutating the shared input_kwargs dict; build a local copy with
the weighted-failover keys so the entry (with _excluded_deployment_ids)
cannot leak into later fallback paths reading the same dict.
- _get_excluded_filtered_deployments now returns an empty list when the
exclusion filter removes every healthy deployment, instead of falling
back to the original list. The original-list behavior risked re-picking
the just-failed deployment; callers already handle the empty case by
raising their no-deployments error, which weighted failover now catches
and converts into a normal cross-group fallback.
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* fix(router): fall through to rpm/tpm when total weight is zero
When the weight metric's total is zero (e.g. after weighted-failover
exclusion leaves only zero-weight backups), continue to the next metric
(rpm/tpm) instead of returning a uniform random pick immediately. This
lets rpm/tpm still drive routing when present, and only falls back to
the uniform random pick at the end if no metric provides a positive
total weight.
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* fix(router): skip weighted failover when remaining deployments are all in cooldown
_maybe_run_weighted_failover was computing 'remaining' from all_deployments
(every deployment in the model group, including those in cooldown). This meant
that when all non-excluded deployments were in cooldown the method still invoked
run_async_fallback unnecessarily, which propagated into async_get_healthy_deployments,
found no eligible deployments, and raised RouterRateLimitError — only safely
caught thanks to the earlier exception-broadening fix.
The fix: before computing 'remaining', fetch the current cooldown set via
_async_get_cooldown_deployments and subtract it from all_ids. This allows
_maybe_run_weighted_failover to return None immediately (skipping the
run_async_fallback call entirely) when every non-failed deployment is in cooldown,
letting the caller fall through to the correct cross-group fallback path without
the wasteful extra round-trip.
Tests added:
- unit: _maybe_run_weighted_failover returns None without calling run_async_fallback
when all remaining deployments are in cooldown
- unit: _maybe_run_weighted_failover still calls run_async_fallback when at least
one healthy (non-cooldown) deployment is available
- integration: end-to-end fallthrough to cross-group fallback when remaining
deployments are in cooldown
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Yassin Kortam <yassin@berri.ai>
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
* fix(vertex-ai): fix zero cost/usage on completed Vertex AI batch jobs
Vertex batch jobs recorded 0 spend and 0 tokens after PR #25627 added
automatic transformation of GCS predictions.jsonl to OpenAI format.
Two bugs fixed:
1. batch_utils.py: the Vertex-specific cost/usage reader
(calculate_vertex_ai_batch_cost_and_usage) was always invoked and
reads raw usageMetadata fields that no longer exist in the
OpenAI-shaped output. Now the reader is only used when
disable_vertex_batch_output_transformation=True; otherwise the
generic path handles the already-transformed OpenAI-shaped content.
2. cost_calculator.py: batch_cost_calculator skipped the global
litellm.get_model_info() lookup when a model_info dict was passed
in, even when that dict had no pricing fields (e.g. deployment
metadata with only id/db_model). It now falls back to the global
pricing table when the provided model_info has no pricing data.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Update litellm/cost_calculator.py
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
* fix(cost-calculator): use not-any guard for pricing fallback in batch_cost_calculator
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(cost-calculator): treat explicit zero batch pricing as set in model_info
The fallback to litellm.get_model_info() used truthy checks on pricing
fields, so 0.0 was treated as missing and replaced by global rates.
Use `is not None` like elsewhere in cost calculation. Add regression test.
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Sameer Kankute <Sameerlite@users.noreply.github.com>
* fix(managed_batches): convert raw output_file_id to managed ID in CheckBatchCost poller
CheckBatchCost bypasses async_post_call_success_hook, causing raw provider
output_file_ids to be persisted in LiteLLM_ManagedObjectTable. This fix converts
output_file_id and error_file_id to managed base64 IDs before the DB write.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(check_batch_cost): persist managed file before mutating response and propagate team_id
- Move setattr after store_unified_file_id so the response only receives the
managed ID once the DB record is successfully written. Avoids serializing
an orphaned managed ID into file_object when the store call fails.
- Populate team_id on the minimal UserAPIKeyAuth from job.team_id so the
managed file record is created with the correct team ownership, allowing
other team members to access the batch output file via /files/{id}/content.
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* test(managed_batches): extend test to cover error_file_id conversion
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix managed file test
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Yassin Kortam <yassin@berri.ai>
* feat(proxy): fix vector store retrieve/list/update/delete routing without model
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(proxy): remove unchecked query-param injection in vector store management endpoints
Co-authored-by: Cursor <cursoragent@cursor.com>
* test(proxy): use subset assertion for vector store route test to allow extra kwargs like shared_session
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
_build_mcp_server_table omitted delegate_auth_to_upstream, so GET /v1/mcp/server always returned the default false while the registry kept the DB value.
Co-authored-by: Cursor <cursoragent@cursor.com>
Log verbose_logger.warning when loading oauth2 interactive servers with
available_on_public_internet=false and delegate_auth_to_upstream=true
(config + DB). Dashboard Alert for the same combo. CLAUDE note for
operators. Tests for log and M2M skip.
Remove available_on_public_internet gating from delegate-auth-to-upstream
paths so oauth2 + delegate_auth_to_upstream interactive servers behave
the same when marked internal. Keeps M2M exclusion. Updates tests.
* feat(custom_logger): add async_post_agentic_loop_response_hook
Lets a CustomLogger shape the response returned by the agentic-loop
follow-up call without bypassing the loop's safety / observability
machinery (depth tracking, fingerprinting, etc.). Default returns the
response unchanged.
Used by websearch_interception to inject Anthropic-native
web_search_tool_result blocks when the originating client requested a
native web_search_* tool.
* feat(llm_http_handler): call post-agentic-loop hook on the originating callback
In _execute_anthropic_agentic_plan, after anthropic_messages.acreate
returns, call the originating callback's
async_post_agentic_loop_response_hook so it can mutate the final
response (e.g. inject native tool_result blocks). Pass the callback
through from _call_agentic_completion_hooks.
Exceptions in the post-hook are caught and logged so a buggy callback
can't kill the request.
* feat(websearch_interception): add is_anthropic_native_web_search_tool
Identifies tools the Anthropic-native clients (Claude Desktop, the
Anthropic SDK, the Anthropic Console) use to request native search:
type starts with "web_search_" (e.g. web_search_20250305). Rejects the
LiteLLM standard tool, the OpenAI-function variant, the bare
"WebSearch" legacy name, and the bare "web_search" Claude Code shape.
This lets us decide per-request whether the client expects
web_search_tool_result content blocks in the response, without
renaming any existing constants or touching native-provider skip
logic.
* feat(websearch_interception): add build_web_search_tool_result_block
Produces the Anthropic-native web_search_tool_result content block
from a structured SearchResponse. Anthropic-native clients use this
block to populate citations / source links — the existing text-blob
flatten path only feeds readable evidence to the model and discards
the structure, so this builder gives us the missing piece.
Shape matches https://docs.anthropic.com/en/api/web-search-tool —
web_search_result items carry url, title, page_age, encrypted_content
(empty string when the search provider doesn't supply one).
* feat(websearch_interception): emit native web_search_tool_result blocks
When the originating client request carried a native Anthropic
web_search_* tool, the final response now also carries
web_search_tool_result content blocks alongside the model's text
answer — so Claude Desktop / Anthropic SDK clients can populate the
citations panel and replay conversation history with structured search
evidence.
Wiring:
- Pre-request hooks (both deployment + Anthropic path) set a flag on
kwargs when they see a native web_search_* tool, so the signal
survives the conversion-to-litellm_web_search step regardless of
which hook fires first.
- _execute_search now returns (text, SearchResponse) so the structured
results aren't lost when the text is flattened for the follow-up
model call.
- _build_anthropic_request_patch returns the parallel list of
SearchResponse objects.
- async_build_agentic_loop_plan pre-builds the web_search_tool_result
blocks (one per tool_use_id) and stashes them on plan.metadata when
the flag is set.
- async_post_agentic_loop_response_hook reads the metadata and
prepends the blocks to response.content.
- _execute_agentic_loop mirrors the injection for the legacy path so
both paths behave identically.
Clients that send the LiteLLM standard tool keep the existing
text-only behavior — no regression.
* test(websearch_interception): cover native web_search_tool_result emission
18 tests across:
- detector branches (native vs litellm-standard, OpenAI-function shape,
Claude Desktop builtin WebSearch, bare web_search, missing type)
- block-builder shape (results, none, empty)
- pre-request hook flag-setting (native sets, standard does not)
- async_build_agentic_loop_plan attaches blocks to plan.metadata when
the flag is present, leaves metadata untouched when absent
- post-hook injection into dict and object responses
- legacy _execute_agentic_loop mirrors the injection so both paths
return the same shape
* test(websearch_short_circuit): keep _execute_search mocks in sync with new tuple return
* test(websearch_thinking_constraint): keep _execute_search mocks in sync with new tuple return
* feat(websearch_interception): emit native blocks from try_short_circuit_search
The agentic-loop post-hook only fires when the model returns a tool_use
block. Cowork / Claude Desktop on Bedrock actually make TWO requests
per user turn: the main /v1/messages with their builtin tool, and a
separate standalone /v1/messages whose only tool is
web_search_20250305. That second request hits try_short_circuit_search
— no agentic loop, no post-hook — and was returning text-only, leaving
the citations panel empty.
When the short-circuit input carries a native web_search_* tool, build
a synthetic server_tool_use + web_search_tool_result pair (using the
structured SearchResponse already returned by _execute_search) so the
client gets the native shape it expects. The legacy text block is
preserved so non-native short-circuit callers (Claude Code,
github_copilot, etc.) see the same payload as before.
Failure path still emits the native block pair (with empty results)
plus the text-error block, so the client gets a well-formed response
rather than a malformed half-shape.
* test(websearch_native_blocks): cover short-circuit native-block emission
Three new cases on top of the existing 18:
- native web_search_20250305 short-circuit → [server_tool_use,
web_search_tool_result, text], ids paired, urls/titles carried.
- litellm_web_search short-circuit → text-only (no regression).
- native short-circuit on search failure → still emits the native
block pair (empty results) plus the text-error block, so the client
never sees a malformed half-shape.
* test(websearch_short_circuit): index assertions by block type, not by position
Native short-circuit responses now have [server_tool_use,
web_search_tool_result, text] when the input carries
web_search_20250305 — find the text block by type rather than relying
on content[0].
* fix(websearch_interception): gate legacy WebSearch name on schema absence
Clients like Cowork / Claude Desktop ship a client-side tool named
"WebSearch" with a full input_schema — they handle it themselves and
expect to make a separate native web_search_20250305 sub-request for
the actual search.
Today is_web_search_tool matches the bare name regardless of other
fields, which hijacks the client's tool server-side. The agentic loop
fires on the main request, the model never gets to emit the
client-side tool_use, and the separate native sub-request (where
citation data flows) is never made. Net: citations panel empty.
Real Anthropic client tools always carry input_schema (the API rejects
them otherwise), so a bare {name: "WebSearch"} with no schema is the
only thing that could be a legacy interception marker. Gate the match
on schema absence: legacy callers (if any) keep working, real
client-side WebSearch tools pass through untouched.
* fix(websearch_interception): drop "WebSearch" from response-detection lists
Post-conversion the model always sees ``litellm_web_search``, so the
"WebSearch" entry in the response-side tool_use detection lists was
dead at best. If a model ever did return ``tool_use(name="WebSearch")``
it would now (incorrectly) hijack the client's own ``WebSearch`` tool
again — same Cowork problem we just fixed on the input side. Drop it.
* test(websearch_native_blocks): cover the WebSearch legacy-name schema gate
Three new cases:
- {name: "WebSearch"} (bare interception marker) → still matched
- {name: "WebSearch", input_schema: {...}} (Cowork client tool) →
passes through untouched
- {name: "WebSearch", description: "..."} (no schema) → still matched
on the assumption it's a legacy marker rather than a malformed real
client tool.
---------
Co-authored-by: Ishaan Jaffer <ishaanjaffer0324@gmail.com>
* fix(rate-limit): stop v3 limiter from leaking internal stash to provider body
PR #27001 (atomic TPM rate limit) introduced a reservation flow that
writes four LiteLLM-internal keys onto the request data dict:
_litellm_rate_limit_descriptors
_litellm_tpm_reserved_tokens
_litellm_tpm_reserved_model
_litellm_tpm_reserved_scopes
_litellm_tpm_reservation_released
These keys are forwarded as request body params to the upstream provider,
which rejects them as unknown fields:
OpenAI -> 400 'Unknown parameter: _litellm_rate_limit_descriptors'
(mapped by litellm to RateLimitError / 429, hiding the bug
behind a misleading 'throttling_error' code)
Anthropic -> 400 '_litellm_rate_limit_descriptors: Extra inputs are
not permitted'
Net effect: every chat completion against any real provider fails the
moment a virtual key has any tpm_limit / rpm_limit set — i.e. v3-enforced
key-level TPM/RPM limits are broken end-to-end. The v3 RPM/TPM check
itself still runs (raises 429 on over-limit), but the success path
poisons the upstream body.
Reproduced on litellm_internal_staging HEAD (410ce761dc) against
gpt-4o-mini and claude-haiku-4-5 with a 1-RPM/1-TPM key — first request
fails with the provider's unknown-field error.
Fix: the stash is metadata only.
- Add RATE_LIMIT_DESCRIPTORS_KEY constant and a _LITELLM_STASH_KEYS
registry so we have a single source of truth for stash keys.
- New helper _stash_value_in_metadata_channels writes to
data['metadata'] / data['litellm_metadata'] without touching the
top level.
- _stash_reservation_in_data and the descriptor stash now route
through that helper. _mark_reservation_released stops writing
top-level.
- _lookup_stashed_value also checks kwargs['metadata'] /
kwargs['litellm_metadata'] (raw request_data shape) in addition to
kwargs['litellm_params']['metadata'] (completion kwargs shape).
- async_post_call_failure_hook now reads descriptors via the unified
metadata lookup instead of request_data.get(top-level).
- Defense in depth: async_pre_call_hook strips any stash key that
somehow surfaced at the top level (stale cache, future refactor,
test fixture) before returning.
Tests:
- New regression test asserts no _litellm_* stash key is present at
the top level of data after async_pre_call_hook, and that the
metadata channel still carries the reservation + descriptors so
success / failure reconciliation works.
- Existing test_tpm_concurrent.py tests that asserted top-level
presence are updated to read from data['metadata'] — the location
is an implementation detail; the spec is that post-call callbacks
can resolve the stash.
Verified end-to-end against OpenAI gpt-4o-mini and Anthropic
claude-haiku-4-5 via /v1/chat/completions on a low-rpm key:
- With limits not exceeded: HTTP 200, valid completion response,
no leaked fields in body.
- With RPM exceeded: HTTP 429 from v3 enforcement
('Rate limit exceeded ... Limit type: requests').
- With TPM exceeded: HTTP 429 from v3 enforcement
('Rate limit exceeded ... Limit type: tokens').
Full v3 hook test suite passes (171 tests).
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
* chore(rate-limit): use RATE_LIMIT_DESCRIPTORS_KEY constant in test, trim noisy comments
Address greptile P2: test fixture now uses the imported constant.
Drop comments that re-explain what well-named identifiers already convey.
* fix(rate-limit): reject caller-supplied stash values to prevent TPM-refund abuse
Strip _LITELLM_STASH_KEYS from data top-level and both metadata channels at
the start of async_pre_call_hook. Without this, an authenticated caller can
inject _litellm_rate_limit_descriptors plus _litellm_tpm_reserved_tokens in
body metadata, trigger a proxy-side rejection, and cause
async_post_call_failure_hook to refund TPM counters against attacker-named
scopes (e.g. another tenant's api_key).
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
- Introduce `_CallbackCapabilities` dataclass and `ProxyLogging._callback_capabilities()` static method that inspects `litellm.callbacks` once and caches capability flags keyed on (list length, member ids); invalidates automatically when the callback list mutates without per-request iteration overhead
- Replace O(n) `litellm.callbacks` walks in `async_pre_call_hook`, `during_call_hook`, `async_post_call_streaming_iterator_hook`, `async_post_call_streaming_hook`, and `post_call_response_headers_hook` with fast-path exits when no relevant callbacks are registered
- Add `needs_iterator_wrap()` and `needs_per_chunk_streaming_hook()` instance methods to decouple iterator-level wrapping from per-chunk hook execution; avoids `get_response_string` materialization per chunk when no guardrail or chunk-hook callback is active
- Introduce `_fast_serialize_simple_model_response_stream()` using `orjson` for common single-choice text streaming chunks, bypassing the full Pydantic serializer; falls back to `model_dump_json` for tool calls, logprobs, usage, and provider-specific fields
- Add early-return in `_restamp_streaming_chunk_model` when downstream model already matches the requested model, avoiding unnecessary string comparisons on every chunk
- Fix stale zero-cost cache bug in `_is_model_cost_zero`: move the per-router `_zero_cost_cache` dict onto the `Router` instance and clear it in `_invalidate_model_group_info_cache` so in-place pricing updates via `upsert_deployment` immediately resume budget enforcement
- Add `scripts/benchmark_chat_completions_perf.py`: standalone async benchmarking tool with a mock OpenAI provider, LiteLLM proxy process management, non-streaming RPS, streaming TTFT, and full-stream latency measurements with repeat/median run support
- Add comprehensive unit tests covering capability detection, cache invalidation, fast-path correctness, zero-cost cache regression, and the no-callback streaming fast path
Co-authored-by: Yassin Kortam <yassinkortam@g.ucla.edu>
