MCP server CRUD endpoints (/v1/mcp/server*) were bundled with MCP
tool-call / passthrough endpoints under llm_api_routes, so setting
DISABLE_LLM_API_ENDPOINTS=true on admin-only nodes also blocked the
Admin UI from listing, adding, or attaching MCP servers.
Separate mcp_inference_routes (data-plane, gated by
DISABLE_LLM_API_ENDPOINTS) from mcp_management_routes (control-plane,
gated by DISABLE_ADMIN_ENDPOINTS). Keep mcp_routes as a union for
backward compat with allowed_routes=["mcp_routes"] virtual key configs.
Upgrade is_management_route to pattern-aware matching so
/v1/mcp/server/{path:path} resolves for concrete IDs.
Temporary MCP OAuth sessions were kept in process-local memory, so on
multi-instance/LB proxy deployments a session created on instance A could
not be found when the follow-up /server/oauth/{server_id}/... request
landed on instance B.
Persist temporary session records to Redis (encrypted with the existing
proxy encryption helpers) as a best-effort L2 cache alongside the current
in-memory L1. Convert get_cached_temporary_mcp_server to async and await
it from the authorize/token/register OAuth endpoints.
Made-with: Cursor
Vertex multi-region endpoints (e.g. us, eu) use the rep host pattern, not
{geo}-aiplatform.googleapis.com. Regional IDs still contain a hyphen.
common_utils.get_vertex_base_url centralizes the rule for SDK/API URL building.
Proxy pass-through duplicates the same branching in a local get_vertex_base_url
(with trailing slashes) to avoid importing from common_utils there; live
WebSocket passthrough uses the same multi-region host logic for wss://.
Tests cover us/eu for the common_utils helper.
Made-with: Cursor
Two fixes to proxy-db CI:
1. test_realtime_webrtc_endpoints.py's `proxy_app` fixture mutated the
module-global `proxy_server.master_key` without restoring it, leaking
state into any test that shared the same xdist worker. Under
--dist=loadscope with 2 workers (GHA proxy-endpoints), this caused the
google_endpoints tests to fail with "No api key passed in." because
user_api_key_auth saw a set master_key and a missing API key on the
test request. The fixture now saves and restores the original value.
2. Address the Greptile note that the semantic shard design has no
catch-all, so a new test file added to tests/proxy_unit_tests/ without
a matrix entry would silently skip CI. Adds an assert-shard-coverage
job that enumerates test_*.py files and fails the workflow if any are
not referenced by a matrix entry, with a clear message telling the
author which semantic shard to place it in. All proxy-db shards now
depend on this guard.
The mocked async_increment_cache_pipeline is invoked from Router's
deployment_callback_on_success, registered as an async success callback.
Those callbacks are enqueued to GLOBAL_LOGGING_WORKER and run on a
background task, so the mock may not have been called yet when the test
asserts on it. Flush the worker before asserting.
Two independent deflakes:
1. test_ui_view_spend_logs_unauthorized (unit) was returning 400 instead
of 401/403 when earlier tests in the file left proxy-auth globals
(prisma_client, master_key, user_custom_auth, general_settings,
user_api_key_cache) in a state that let invalid tokens pass auth and
fall through to the endpoint's own start_date/end_date validation.
Add an autouse fixture that pins those globals to their import-time
defaults for every test in the file. Harden the assertion to include
response body so future flakes are diagnosable.
2. test_basic_spend_accuracy (CI job proxy_spend_accuracy_tests) depends
on the Redis transaction buffer flushing spend to Postgres. The buffer
uses a single global pod-lock key (cronjob_lock:db_spend_update_job)
and a single global buffer list key. Pointing the proxy at the shared
remote Redis means concurrent CI pipelines contend for the same lock
and can drain each other's buffer into the wrong database. Add a
start_redis reusable command that boots a per-job redis:7-alpine
container (digest-pinned), and switch proxy_spend_accuracy_tests to
REDIS_HOST=host.docker.internal:6379 so lock and buffer state are
isolated per CI run.
* feat: add gpt-5.5 to model cost map
Add gpt-5.5 entry with pricing from OpenAI flagship page:
input $5/1M, cached input $0.50/1M, output $30/1M, 272K context.
* test: add gpt-5.5 coverage for model cost map and gpt-5 routing
- Add gpt-5.5 to GPT5_MODELS parametrized list so both OpenAIGPT5Config
and AzureOpenAIGPT5Config routing tests cover the new model.
- Add test_generic_cost_per_token_gpt55 verifying the new entry's
cost-map values ($5/$0.50/$30 per 1M) and that generic_cost_per_token
returns the expected prompt/completion costs.
The periodic budget-window reset job filtered keys/teams with
`where={"budget_limits": {"not": None}}`. The prisma-client-python
library does not support null-filtering on `Json?` columns (no
DbNull/JsonNull sentinel — upstream issue #714). The client drops the
`None` value during serialization and the engine rejects the query with
`MissingRequiredValueError: where.budget_limits.not: A value is
required but not set`, so neither the key nor team reset path runs.
Switch those two `find_many` calls to `query_raw` with
`WHERE budget_limits IS NOT NULL`, selecting only the PK and the
`budget_limits` column. Writes still go through the ORM. Add unit tests
covering the expired/unexpired paths for keys and teams, string-encoded
JSON payloads, empty payloads, error isolation between the two paths,
and a regression guard asserting the query still uses `IS NOT NULL`.
Align the Ruby, Node.js, and npm install path with the rest of the
config. Three separate upstream installers were being invoked via
\`curl ... | bash\` or unlocked \`npm install\`:
- RVM's \`get.rvm.io/stable\` installer (mutable upstream script).
Replace with a shallow git clone of the rvm/rvm repo at tag 1.29.12
and verify HEAD matches the published commit SHA before running the
local \`./install\` script. Same pattern already used for the
helm-unittest plugin in .github/workflows/helm_unit_test.yml.
- NodeSource's \`deb.nodesource.com/setup_18.x\` piped into sudo bash.
Replace with a direct download of the Node.js 18.20.8 linux-x64
tarball from nodejs.org, verified against the published
SHASUMS256.txt digest before extraction.
- \`npm install @google-cloud/vertexai @google/generative-ai\` and
\`--save-dev jest\` resolved fresh from the npm registry on every
run. Add \`tests/pass_through_tests/package.json\` with pinned
direct-dep versions and commit the generated package-lock.json, then
switch CI to \`npm ci\` (exact lockfile install, fails on drift).
Also scopes the Ruby+JS test runners to \`tests/pass_through_tests/\`
so they pick up the committed package.json rather than writing
node_modules at repo root.
The original check `"gpt-5-chat" not in model` already correctly
classifies all current gpt-5 variants (including gpt-5.3-chat and
gpt-5.1-chat, which do NOT contain the substring "gpt-5-chat"). This
change replaces it with an explicit `startswith("gpt-5-chat")` prefix
test on the provider-prefix-stripped model name.
The new check is functionally equivalent for all existing model names
but makes the classification boundary unambiguous and forward-safe:
future model names that might contain "gpt-5-chat" as an interior
substring won't accidentally be excluded from the GPT-5 reasoning path.
Also moves the new regression test from tests/ root to
tests/test_litellm/llms/openai/ so it is included in `make test-unit`.
* fix(anthropic): handle tool_choice type 'none' in messages API
* test(anthropic): add regression test for tool_choice type 'none'
---------
Co-authored-by: BillionClaw <267901332+BillionClaw@users.noreply.github.com>
Co-authored-by: Krrish Dholakia <krrish+github@berri.ai>
When reasoning_auto_summary is enabled (via litellm_settings or env var),
automatically set thinking.display="summarized" on native /v1/messages
requests. This ensures thinking content is returned in the response
instead of being omitted (the default on Claude 4.7+).
Only applies when thinking is enabled (type != "disabled").
The existing reasoning_auto_summary flag already handles the
/v1/responses path (summary="detailed") and the chat/completions
adapter path — this extends coverage to the native messages handler.
`get_file_ids_from_messages` and `update_messages_with_model_file_ids`
assume every content block with `type: "file"` has a nested `file` dict in
the OpenAI Chat Completions shape. That assumption is too strong: `type:
"file"` is a public content-block discriminator and several real producers
emit blocks that use it without the OpenAI `file` sub-dict. For example,
LangChain v1's `_normalize_messages` rewrites OpenAI file blocks into
`{"type":"file","id":"...","base64":"...","mime_type":"...","extras":{}}`
before they reach LiteLLM.
`AnthropicConfig.validate_environment` calls both helpers unconditionally
on every Anthropic (and Anthropic-via-Vertex) request, so any such block
raises `KeyError: 'file'` which the Vertex partner layer then wraps as a
`500 InternalServerError` before the LLM is even contacted.
This patch switches both helpers from `c["file"]` to a defensive
`c.get("file")` + dict check. When the block does not match the OpenAI
shape there is no file_id to extract or remap, so we skip it and leave
the block untouched for the downstream provider transformer to handle.
Adds 5 regression tests covering the LangChain v1 shape, the OpenAI
happy path, mixed shapes in one message, `file` set to a non-dict value,
and the remap path for non-OpenAI blocks.
Related to #24503, which proposed raising `BadRequestError` in the same
spots. For these two discovery functions specifically, the skip semantics
is strictly more permissive: well-formed OpenAI blocks still yield their
file_id, and legitimate non-OpenAI blocks stop crashing the request.
* fix(model-info): include reasoning effort support fields in get_model_info
_get_model_info_helper constructs ModelInfoBase explicitly but never
reads supports_xhigh/minimal/none_reasoning_effort from the cost map
JSON. Add the three fields so get_model_info() returns them correctly.
Also add supports_minimal_reasoning_effort to the ModelInfo TypedDict
(xhigh and none were already declared, minimal was missing).
* fix(model-registry): add missing reasoning effort fields for claude 4.6/4.7
Claude Opus 4.7 supports max reasoning effort (above xhigh).
The field was present for Opus 4.6 but missing for all Opus 4.7
entries (base, dated, Bedrock, Vertex AI, Azure AI).
All Claude 4.6/4.7 models (Opus 4.6, Sonnet 4.6, Opus 4.7) support
minimal reasoning effort via adaptive thinking. Add the field to all
provider variants.
* fix(adapter): map output_config.effort to reasoning_effort (#25079)
Anthropic's adaptive thinking (thinking.type="adaptive") and
output_config.effort were silently dropped when translating to
OpenAI format, resulting in no reasoning_effort on the outgoing
request.
Adapter changes (format translation):
- adapters/transformation.py: add "adaptive" branch to
translate_anthropic_thinking_to_reasoning_effort(); pass through
output_config.effort as-is in _translate_thinking_to_openai();
add "output_config" to translatable_anthropic_params
- adapters/handler.py: extract output_config from extra_kwargs into
request_data so it reaches the translation layer
- responses_adapters/transformation.py: add "adaptive" branch and
output_config param to translate_thinking_to_reasoning()
Handler changes (model-aware normalization):
- utils.py: add normalize_reasoning_effort_value() that uses
get_model_info() to map "max" → "xhigh"/"high" and
"minimal" → "minimal"/"low" based on model capabilities
- adapters/handler.py: call normalization before responses routing
- responses_adapters/handler.py: call normalization after translation
Relates to BerriAI/litellm#25079
* test(reasoning-effort): add tests for effort capability fields and normalize logic
Test coverage for:
- get_model_info returning supports_minimal/max_reasoning_effort fields
- JSON registry entries for claude 4.6/4.7 across all providers
- normalize_reasoning_effort_value degradation chains and exception fallback
- Adapter translation of adaptive thinking + output_config.effort
* fix: forward custom_llm_provider to normalize_reasoning_effort_value in responses adapter
* fix(mcp_semantic_tool_filter): match canonical tools that arrive with
a client-side namespace prefix.
`SemanticMCPToolFilter._get_tools_by_names` matched by exact equality
between the canonical name stored in the router
(`<server><MCP_TOOL_PREFIX_SEPARATOR><tool>`) and the name in the
incoming `tools[]` list. MCP clients such as opencode wrap every tool
name with their own additive alias prefix
(`<client_alias>_<canonical>`), so the two never matched, the filter
dropped every tool to zero, and the proxy forwarded `tools: []` with
`tool_choice: auto` — which strict upstream providers reject with a 400.
The fix adds anchored suffix matching with a separator check: the
canonical must form the complete tail of the incoming name and be
preceded by `_` or `-`. Exact matches still win over suffix matches,
incoming tools are returned at most once, and the original tool object
is passed through unchanged so the client-facing name survives for
tool-call round-trips.
Seven unit tests in a new TestGetToolsByNames class cover exact
match, underscore- and dash-prefixed variants, non-separator-anchored
suffixes (which must not match), exact-wins-over-prefixed precedence,
deduplication when two canonicals suffix-match the same incoming tool,
and ordering-follows-router-output.
Fixes#26078
* review: strengthen the suffix-fallback tie-breaker and the
deduplication regression test (Greptile comments on #26117)
- test_same_tool_not_returned_twice now passes two distinct canonicals
("read_file" and "file") that both suffix-match the same incoming
tool, rather than the same canonical twice, so the assertion
actually exercises the used_ids dedup path instead of the
duplicate-input-list path.
- The suffix fallback in _get_tools_by_names now prefers the shortest
incoming name that still qualifies under the separator-anchored
match. In the one-prefix-per-client opencode scenario this is a
no-op, but in multi-namespace configurations the shortest qualifying
name is the least-wrapped one and is the most defensible deterministic
choice, replacing the dict-insertion-order fallback.
- Adds test_suffix_fallback_prefers_shortest_candidate covering the
new tie-breaker directly.
Still 15 tests passing locally (was 14).
* review(#26117): gate suffix-matching on canonical containing MCP_TOOL_PREFIX_SEPARATOR
@krrish-berri-2 flagged a possible collision in the suffix fallback:
a local user function whose name happens to end in a bare canonical
substring (e.g. my_firecrawl_scrape vs canonical firecrawl_scrape)
would be spuriously selected.
Server-registered MCP tools are always emitted as
<server_name><MCP_TOOL_PREFIX_SEPARATOR><tool_name> via
add_server_prefix_to_name, so a canonical without the separator is
not a namespaced MCP tool and does not warrant suffix matching.
Added that guard to _name_matches_canonical with a regression test
(test_does_not_collide_with_local_function_on_unprefixed_canonical)
that reproduces the collision before the fix and is pinned after.
Pre-existing TestGetToolsByNames fixtures that relied on bare
canonicals (get_weather, search, read_file, write/delete/read) were
switched to realistic server-prefixed ones so they continue to
exercise the suffix-fallback path under the new guard. The opencode
scenario (client prefix on already-server-prefixed canonical) is
unchanged.
---------
Co-authored-by: sakenuGOD <sakenuGOD@users.noreply.github.com>
Co-authored-by: Krrish Dholakia <krrish+github@berri.ai>
Two changes, both test-only:
- Configure the aiohttp session with TCPConnector(force_close=True) and an
explicit ClientTimeout(total=30, connect=10). Prevents reuse of idle TCP
connections that the proxy/kernel may have closed during the long window
between setup POSTs and the later poll loop, and surfaces a blocked proxy
event loop quickly instead of hanging on aiohttp's 5-minute default.
- In poll_key_spend_until, catch aiohttp.ClientError and asyncio.TimeoutError
around the single /key/info call. A transient transport hiccup now logs and
retries on the next tick instead of failing the entire polling loop.
Addresses the ConnectionTimeoutError observed on the first /key/info call
after the 20 chat completions.
store_in_memory_spend_updates_in_redis drained the in-memory queues
into local variables before the rpush pipeline. If rpush raised (cloud
Redis hiccup, timeout, connection blip), those already-drained
transactions were garbage-collected with the scheduler job, silently
losing all spend aggregated during that tick.
Wrap the rpush in try/except. On failure, re-enqueue the aggregated
transactions into their respective in-memory queues so the next
scheduler tick retries.
Add a unit test that seeds real queues, simulates an rpush failure,
and asserts the transactions land back in-memory.
Replace the calibration step (one request + 10-minute poll) with an
independent ground truth computed from response usage via
litellm.cost_per_token. All N requests are made up front, so a single
dropped Redis write no longer kills the test.
Add /health/readiness checks at test start and on poll timeout so the
failure message surfaces proxy state (db, cache) instead of "calibration
timed out".
Set PROXY_BATCH_WRITE_AT=2 in the spend tracking CI job to shorten the
scheduler flush window.
Restore guardrail spend/UI event_type wiring, request_data on streaming
OUTPUT paths, and centralized match redaction after the upstream revert.
Made-with: Cursor
* add moonshot/kimi-k2.6 to model registry
* add moonshot/kimi-k2.6 to backup model registry
* add tests for moonshot/kimi-k2.6 model registry
* fix moonshot/kimi-k2.6 pricing and add reasoning support
* fix moonshot/kimi-k2.6 pricing and add reasoning support in backup
* update kimi-k2.6 tests: fix pricing, add tool_choice and reasoning checks
* fix: load kimi-k2.6 registry tests from local backup instead of remote cost map
P1 review: adaptive_router.py had a top-level import of
AdaptiveRouterUpdateQueue from litellm.proxy.db, which broke the
SDK/proxy boundary that every other router strategy respects. No
other router_strategy module imports from litellm.proxy at module
level.
The queue only depends on litellm._logging — it never needed to
live under litellm.proxy. Moved:
litellm/proxy/db/db_transaction_queue/adaptive_router_update_queue.py
→ litellm/router_strategy/adaptive_router/update_queue.py
tests/test_litellm/proxy/db/db_transaction_queue/
test_adaptive_router_update_queue.py
→ tests/test_litellm/router_strategy/adaptive_router/test_update_queue.py
Also switched the queue's logger from verbose_proxy_logger to
verbose_router_logger to match the new module's ownership.
P2 review: drop unused constant STAGNATION_JACCARD_EXACT from
config.py — it was defined but never referenced.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously, members added to a team without an explicit per-member budget were
all linked to the same `litellm_budgettable` row referenced by the team's
`metadata.team_member_budget_id`. Updating one member's budget via
`/team/member_update` mutated the shared row and silently changed every other
member's budget too.
Now both write paths produce a private, per-member budget:
- `add_new_member` clones the team's default budget into a fresh row when a
member is added without `max_budget_in_team`/`allowed_models`. If no team
default exists, the membership is created with no budget.
- `_upsert_budget_and_membership` detects when an existing membership still
points at the team's default budget id and clones-on-write, relinking the
membership to the new private budget before applying the update.
- `team_member_update` reads `team_member_budget_id` from team metadata and
passes it through so the helper can make this distinction.
Adds unit tests for clone-on-write, in-place update of a private budget, and
the no-default-no-budget add path.
Made-with: Cursor
P1: start the adaptive-router flusher loop unconditionally at proxy boot
instead of gating on 'adaptive_routers is non-empty'. Adaptive routers
added via /config/reload after boot now have their queues drained.
State is lazy-loaded per router on first flush tick (new _state_loaded
flag on AdaptiveRouter) so hot-reloaded routers still get their
persisted priors.
P2: _finalize_adaptive_router_if_configured now prunes stale
AdaptiveRouterPostCallHook callbacks from every litellm callback list
before registering new ones. Without this, every Router replacement
left the old hooks wired up in litellm.callbacks and double-fired
signal recording for every request. Uses
logging_callback_manager.remove_callbacks_by_type (same pattern as the
semantic tool filter).
CI fixes:
- black --check failure: reformatted litellm/router.py
- schema migration diff: aligned @@index with the explicit index name
('idx_adaptive_router_session_activity') from the original migration
by adding 'map:' to all three schema.prisma copies. No new migration
needed.
Tests: 1 new covering the prune-on-hot-reload path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
