Emulated file_search now warns when the model returns a vector_store_id that is
not one of the request's stores, naming the dropped id and the stores that were
searched instead. H16 asserts the warning is emitted exactly once.
A live cost map older than this release, or a proxy whose map fetch lags, could
strip `reasoning` from a model this release knows accepts it. The bundled map is
now the floor: any OpenAI entry it flags as reasoning keeps the param whatever
the live map says. Fine-tuned ids with an empty suffix (`ft:gpt-4o-2024-08-06:org::id`)
now resolve to their base entry instead of failing open, `chat-latest` carries
the flag, and the schema test keeps every codex, deep-research, and chat-latest
entry flagged. The none-effort check goes through a public wrapper so the
responses config stops importing a private helper.
_average_latency skipped integer samples in the sum while counting them in the denominator, which contradicted its own
Sequence[float | int] signature; it now averages every sample. Both success loggers in cost-based routing computed
response_ms / completion_tokens and threw the result away, so a chat response with zero completion tokens raised
ZeroDivisionError inside the handler. The proxy swallows and logs it, but the handler then skips that request's tpm and
rpm update, so cost-based routing undercounts the deployment's usage. The QA run for the latency fix hit it on real
gpt-5.5 traffic through /v1/chat/completions and /v1/messages
The regression test's recording logger overrode async_post_call_failure_hook
with untyped parameters. It now mirrors the base signature, and the
UserAPIKeyAuth import moves to module level so the annotation resolves.
The emulated file_search handler searched whatever vector_store_id the
model returned, so a model steered to an id outside the request's
file_search tool reached a store the per-key vector store permission
check never saw. An id outside the request's stores now falls back to
those stores; an id that is one of them still narrows the search to it.
The changed-tests workflow overrode the suite's `--reruns 1` with `--reruns 0`, so a
transport blip failed a pass that pytest.ini already scopes to network errors and
5xx responses. Pass 2 of run 33692484803 also went red 15s after a model write with
"no healthy deployments": the barrier only polled /v1/models through nginx, which
proves one gateway converged, and the next request rolled the other. The stack now
exports LITELLM_PROXY_REPLICA_URLS, the barrier polls every replica with the full
budget before settling, and up.sh refuses to boot without DD_API_KEY, since the
gateway config enables the datadog callback on every run
Latency-based routing averaged a deployment's cached samples with total / len(samples) and raised ZeroDivisionError
once an entry held none, which the proxy answered as a 500 for every later request on that model group. Cost-based
routing writes the same {model_group}_map entry with minute counters only, so a group used by both strategies hit this
on every latency-routed request. A deployment with no samples now counts as 0 latency, the same as one the router has
never seen
Resolves LIT-7053
A failed pass-through call logged the httpx traceback, whose message
quotes the upstream URL with the provider API key in its query string,
into the spend log's error information and into every failure callback.
The error information built for logging now redacts its traceback and
error message, and the traceback is redacted once before the failure
callbacks receive it.
OCI Cohere restates the whole assistant text on the chunk that carries
the tool calls and again on the terminal chunk that carries chatHistory.
Only the terminal restatement was dropped, so a tool-calling turn streamed
the text twice. Treat a toolCalls-bearing chunk as a restatement too and
drop its text once deltas were already emitted.
Resolves LIT-6819
The test-quality gate counts every patch of a litellm internal against a ceiling, and the three patches this test needs pushed it over. The callback imports increment_spend_counters, update_cache and proxy_logging_obj from proxy_server inside its own body, so there is no seam to inject fakes through; every other test in this file uses the same three patches for the same reason
Claude-Session: https://claude.ai/code/session_01EX13mWex6RaBo9PYnkAtFW
The previous cleanup only ran inside a caller's own except handler, so a
build that failed after its only caller had already been cancelled left
the failed task cached with nothing left to clear it. Move the cleanup
onto the task itself as a done-callback, which fires whether or not
anyone is still awaiting it, so the next request always gets a fresh
attempt instead of replaying the stale failure.
Adds a regression test for exactly that ordering (cancel the only
caller, let the build fail unobserved, then confirm the next request
builds successfully); it fails against the previous except-based
cleanup, which left the task cached.
test_load_state_from_db_handles_unknown_request_type compared the
WRITING cell after load against a cold-start value captured for
GENERAL. They happened to be equal for this fixture (the fast model's
empty strengths list makes every request type's prior identical), which
hid that the assertion was comparing the wrong baseline. Capture each
request type's own cold-start value instead.
The proxy cost callback zeroed response_cost whenever cache_hit was true. That rule dates from Jan 2024 when it was the only place cache hits were priced. The logging layer has priced the LLM share at 0 on a cache hit since Aug 2024, and since guardrail cost joined the standard logging payload the proxy-side zeroing has thrown away a real provider charge: a pre_call guardrail runs before the cache is consulted, so a cached response still cost whatever the guardrail billed. Drop the redundant zeroing so the payload's response_cost, which is already LLM 0 + guardrail cost, reaches spend logs, daily tables and budgets untouched
Claude-Session: https://claude.ai/code/session_01EX13mWex6RaBo9PYnkAtFW
_ensure_routelayer previously awaited asyncio.to_thread(...) directly
inside the lock. Under cancel_on_disconnect, cancelling that await
released the lock while the worker thread kept running, so a second
concurrent request would see no lock held and start a duplicate billed
build. Store the build as a task on self and have every caller await it
through asyncio.shield: cancelling one caller's wait no longer cancels
the build or lets another caller start a second one. A real build
failure (not merely a cancelled caller) clears the slot so the next
call retries fresh instead of replaying the same failure forever.
Also replaces the two tests that monkeypatched _build_routelayer (an
anti-pattern per this repo's conventions) with ones that instrument the
already-injected embedding router dependency instead, and adds a third
proving the cancellation race is actually closed.
- Shrink the fallback comments to one line each; the fuller rationale
was redundant per repo comment policy.
- Add test_pick_model_favors_the_cheaper_model_info_priced_deployment
and its hybrid-router counterpart, which exercise pick_model's actual
Thompson-sampling/scoring output instead of only asserting the
model_to_cost dict. Both are ordered so the expensive model wins
pick_best's insertion-order tie-break on the pre-fix code (proven by
reverting the production diff and rerunning), so they fail before the
fix and pass after it.
Both places that build an adaptive router's model_to_cost (the plain
auto_router/adaptive_router path in router.py, and the hybrid
adaptive-inside-complexity_router path in complexity_router.py) read
input_cost_per_token from litellm_params only. Custom pricing is
conventionally declared under model_info everywhere else in LiteLLM
(cost_calculator.py, add_deployment's litellm.model_cost registration),
so a deployment priced that way silently costs 0.0 in adaptive-router
scoring: every candidate ties on cost, the cost term contributes
nothing, and routing runs on quality alone with no warning.
Fall back to model_info at both call sites when litellm_params does not
declare a cost, matching how quality_router.py already sources cost.
litellm_params still wins when both are set.
Fixes#31481.
load_state_from_db assigned a DB row's (alpha, beta) straight into the
bandit cell, discarding the cold-start prior _init_cold_start_cells had
already put there. AdaptiveRouterUpdateQueue.flush_state_to_db only ever
persists accumulated deltas (its upsert creates a row with the raw delta
as the initial value, then increments it), never a full posterior, so a
cell whose first flush sees only one kind of signal persists a one-sided
row: e.g. alpha=1.0, beta=0.0. Loading that row as the whole cell hands
thompson_sample() a Beta(alpha, 0), and random.betavariate raises
'gammavariate: alpha and beta must be > 0.0' on every draw from that
cell from then on, surviving restarts since the bad row stays in place.
Fix: add the row on top of a freshly computed prior instead of replacing
the cell with it. Deltas are never negative, so both parameters stay
positive.
Fixes#35590. Fixes#29397.
AutoRouter's cold-start route layer construction (SemanticRouter with
auto_sync="local") ran directly on the event loop, doing at least one
synchronous embedding HTTP call inline behind a bare "if routelayer is
None" check with no lock, so it blocked the whole worker and let
concurrent cold-start requests each build a duplicate layer.
ComplexityRouter already solved this identically for its own semantic
keyword matching (_ensure_semantic_routelayer: a lock plus
asyncio.to_thread). Give AutoRouter the same treatment: extract the
build into _build_routelayer and gate it behind _ensure_routelayer's
double-checked async lock.
Fixes#33204.
The Anthropic and Together AI /v1/messages streaming tests required at
least two content_block_delta events. How many deltas a reply is split into
is the provider's choice, and Haiku answers a short count in one or two, so
the assertion failed on provider variance with no change in the proxy: four
of the day's full runs on the PR e2e gate went red on it on 2026-09-05.
The harness now stamps when each SSE event reached the client
(StreamingResponse.stream_event_arrivals, index-aligned with stream_events,
with the clock injectable so the reader has a unit test). Both tests ask for
a reply long enough to take seconds to generate and require the first
content delta to land at least STREAM_MIN_LEAD_SECONDS before message_stop.
A relayed stream shows a lead of about two seconds. A proxy that buffered
the response delivers every event in one burst and fails every time, which
a whole-response buffering relay in front of a live proxy confirmed. The
event-grammar assertions are unchanged.
Replay hands the proxy its recorded chunks back to back, so timing says
nothing there. The assertion is gated on provider_paces_stream() and replay
proves the grammar only, which tests/e2e/CLAUDE.md now says.
* fix(mcp): scan and mask MCP tool call arguments in unified guardrails
A guardrail configured with mode pre_mcp_call was handed only a synthetic
tool definition (name plus an empty parameters schema), so it never saw the
argument values it was configured to inspect, and any rewrite it returned was
discarded. Detection could not fire and masking could not take effect, while
the applied-guardrails metadata still reported the guardrail as having run.
Pass every string leaf of the tool call arguments as texts, and fold the
guardrail's rewritten leaves back into modified_arguments, which is the channel
the MCP call path reads to decide what to send upstream. The leaf walk reuses
the json_string_leaves / with_json_string_leaves helpers the tool result path
already uses, so both directions share one bounded traversal.
Two guardrails running concurrently under run_in_parallel scan the same payload
snapshot, so each returns a full replacement derived from the original leaf.
Rewrites of the same leaf to different values are rejected rather than silently
losing one redaction; a leaf that already holds this guardrail's own replacement
is convergent and still masks, which is what the bundled content filter does
when it rewrites the arguments itself as well as through texts.
* fix(mcp): annotate guardrail argument rewrites
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(tests): isolate MCP guardrail callback state
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore: ratchet LIT010 budget after merge
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(tests): remove duplicate Bedrock hook parameter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): fail closed when guardrail rewrites cannot be mapped to MCP arguments
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(tests): patch the guardrail translation mappings cache where staging now keeps it
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): serve Prometheus /metrics from a separate process via --prometheus_metrics_port
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(proxy): ruff format prometheus_metrics_server
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): fail fast when the separate metrics server cannot start and force the multiproc dir whenever it is enabled
- wait for the child's /health before starting uvicorn; raise a ClickException if it exits first (port in use)
- create PROMETHEUS_MULTIPROC_DIR whenever --prometheus_metrics_port is set, so DB-configured prometheus callbacks work
- honour lowercase prometheus_multiproc_dir; validate the port before spawning
- cover main() entry point, readiness, bind failure and wildcard-host probing in tests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): pin metrics-server readiness to the child pid so another service on the port cannot pass the health check
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): probe metrics-server readiness through the shared HTTPHandler instead of bare httpx.get
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): serve only /metrics on the prometheus metrics port
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): validate metrics server CLI args with pydantic instead of typing.cast
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): satisfy metrics server lint gate
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Anthropic now returns the 64-token 'count to 20' reply in one to three content_block_delta events, measured directly against api.anthropic.com and through proxies at 7672399 and 49a1145 alike, so the incrementality assertion (at least two deltas) failed in litellm-e2e builds 125, 130 and the 278 rerun with no proxy change behind it. A 'count to 100' reply at max_tokens 400 arrived in five to fifty deltas across every measured run
Custom code guardrails could only allow(), block(reason) or modify(). This adds flag(reason, metadata={}) which lets the request or response through unchanged and records a guardrail_flagged entry carrying the guardrail name, configured mode, evaluated input_type (request or response), reason and structured metadata. The new status is threaded through the request-level guardrail_status aggregation, the Guardrails Monitor rollup (flagged_count), Request Logs (action=flagged, most severe phase wins when a guardrail runs pre and post call) and the Request Logs detail view in the dashboard, which now renders FLAGGED with warning styling instead of falling into FAILED.
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): cover presidio post_call, tool_permission, and weave logging cells
Five registry cells in Logging & Guardrails had no covering test. Each one now
has a live scenario read back from the real destination:
- guardrail.presidio.post_call.masks: an output-scoped Presidio guardrail
anonymizes the PII the model repeats back. The prompt also asks for the
address's local part, which Presidio does not mask, so one response proves the
model saw the raw address (no pre-call masking) while the address itself comes
back as <EMAIL_ADDRESS>
- guardrail.tool_permission.pre_call.blocks / .allows: an allow-list of one tool.
A request declaring an unlisted tool is rejected 400 naming it; a request
declaring the permitted tool is served and carries
x-litellm-applied-guardrails, so the allow half cannot pass by the guardrail
never running
- logging.niche_integrations.success.logs_spend / .failure.logs_spend: a
key-scoped weave_otel callback delivers to the real Weave project, read back
through Weave's query API. Success asserts exactly one call whose
llm.response.cost equals the x-litellm-response-cost header; failure asserts
one ERROR-status call naming the provider exception and carrying no cost
Logging & Guardrails coverage goes 24/59 to 29/59. No registry rows are added.
* test(e2e): make the tool-permission allow case deterministic and scope the Weave read-back
Review follow-ups on the coverage PR.
- the allow scenario forced the outcome to depend on whether the model felt like
calling an optional tool, and checked for the tool name as a substring of the
whole body, which a prose mention would satisfy. It now sends
tool_choice="required" and asserts the parsed response carries exactly one tool
call, for the permitted tool
- the Weave read-back queried the newest 200 calls of a shared project and
filtered client-side, so busy traffic could push the target out of the window
and read as a delivery failure. The query now scopes server-side to the
litellm_request op and to calls started after the request, and pages through
the window with offset
- the reader builds its results as tuples instead of accumulating into lists
Also unblocks the lint gate: `basedpyright tests/e2e` runs only on PRs that touch
tests/e2e, and it has been failing on staging for three FakeItem arguments in
test_junit_properties.py. The stand-in now goes through one typed adapter that
says why, so the gate is green without touching junit_properties.py itself.
* test(e2e): scope the presidio post_call guardrail to email and phone
Running the suite three times in a row caught a real flake: Presidio's broader
recognizers sometimes claim the email's local part as an NRP entity, so the
answer came back as `<NRP>\n<EMAIL_ADDRESS>\n<PHONE_NUMBER>` and the assertion
that the raw local part survives failed. That token is what tells output masking
apart from input masking, so it has to survive.
The post_call guardrail now registers pii_entities_config for EMAIL_ADDRESS and
PHONE_NUMBER only, which is also the narrower thing the scenario means. Verified
against the exact marker that failed, plus two others.
* test(e2e): mark weave logging cells stage red
* test(e2e): use per-test stage red skips for the weave logging cells
* feat(mcp): warn when an oauth2_id_jag server outruns the SSO provider's assertion capture
Only the generic OIDC login path captures the IdP id_token that an oauth2_id_jag MCP
server spends as its RFC 8693 subject token. Under Google, Microsoft, SAML or no SSO at
all, registration succeeds and then every ID-JAG credential resolution fails for every
user, with nothing in the logs, the config or the API response to say why.
Report the mismatch from the two places it is knowable: when an oauth2_id_jag server is
created or updated through the management endpoint, and at SSO callback time when a login
hands the arm nothing while such a server is registered. Provider selection mirrors the
callback's precedence, so a generic client id sitting behind GOOGLE_CLIENT_ID does not
clear the warning.
* test(sso): update merged CLI diagnostic patch target
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(mcp): warn about the ID-JAG capture gap for config-declared servers and on the SSO debug page (#39350)
* feat(sso): surface the ID-JAG capture gap on the SSO debug page
/sso/debug/callback is where an operator lands when they are already trying to work out
why ID-JAG is failing, so the reason belongs on it. The annotation appears only when the
active SSO provider captures no identity assertion AND an oauth2_id_jag server is
registered for that gap to break; a deployment without both renders the page it rendered
before, byte for byte. Only the provider name and the remedy are rendered, never a
configured value, and an unreachable MCP table costs the page its annotation rather than
the page itself.
The payload carries the one mutable-ok in this work. Conditionally including a member of a
JSON document has to construct a mapping, and the rejected alternatives are recorded on the
helper so the next reader does not rediscover them.
Held out of the diagnosability PR deliberately: that PR is already reviewed and green, and
this surface ships with the remaining config-load warning as one follow-up.
* feat(mcp): warn at config load when an oauth2_id_jag server outruns the SSO provider's assertion capture
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(sso): trim comments on the ID-JAG debug page diagnostic
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(sso): clean up merged imports
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(sso): satisfy type discipline for diagnostic payload
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(sso): keep the optional ID-JAG payload member on one line for ruff format
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(sso): use Python 3.10-compatible assert_never
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Yassin Kortam <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(sso): keep the ID-JAG capture-gap diagnostic out of the unauthenticated debug page
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(sso): inject the retention check and log via caplog so the ID-JAG tests pass the test-quality gate
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(sso): keep the debug-page outage test on the capture-gap path
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(sso): annotate the retention check type alias
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(mcp): renew the stored SSO identity assertion behind ID-JAG
The oauth2_id_jag arm asserts the id_token captured at the user's last
interactive SSO login, and nothing ever renewed it, so an agent holding a
brokered LiteLLM key could act for that user only until that token's exp.
The assertion already carried the IdP refresh token beside it; this
redeems it.
RefreshingSSOAssertionStore wraps the database reader and satisfies the
same protocol, so the egress arm is unchanged. Renewal is lazy and
single-flighted per user through the same RefreshCoordinator the
authorization_code arm uses, since an IdP that rotates refresh tokens
treats two concurrent redemptions as replay. A refusal leaves the expired
assertion in place so the reader still challenges the user; an
unreachable IdP surfaces as a store outage instead.
* fix(mcp): let a cross-replica loser settle the SSO assertion renewal itself
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): satisfy type discipline lint budget
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(ci): rerun checks after docs main added the missing router setting row
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): answer a cross-replica loser retryable instead of re-electing it
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(mcp): bypass stale assertion cache during renewal
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore: ratchet type-discipline budget after merge
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Yassin Kortam <yassin.kortam@gmail.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cloudzero): infer daily batch schema from every row
pl.DataFrame defaults to inferring column types from the first 100 rows,
so a day whose batch starts with more than 100 rows missing team_alias,
api_key_alias or user_email typed that column as Null and then raised a
ComputeError on the first row that had a value, failing the whole export
with a 500 and sending nothing.
Pass infer_schema_length=None when rebuilding each day's DataFrame, the
same guard the usage query already uses.
* test(cloudzero): cover late tag schema inference
Exercise the CloudZero resource tag field after a long run of missing values so a finite inference window fails the regression test.
* fix(cloudzero): preserve late resource tags
* style(cloudzero): remove redundant test comment
* fix(cloudzero): infer daily batch schema from every row
pl.DataFrame defaults to inferring column types from the first 100 rows,
so a day whose batch starts with more than 100 rows missing team_alias,
api_key_alias or user_email typed that column as Null and then raised a
ComputeError on the first row that had a value, failing the whole export
with a 500 and sending nothing.
Pass infer_schema_length=None when rebuilding each day's DataFrame, the
same guard the usage query already uses.
* test(cloudzero): cover late tag schema inference
Exercise the CloudZero resource tag field after a long run of missing values so a finite inference window fails the regression test.
