* 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.
* fix(router): coordinate async and sync failure handlers at remaining router call sites
Five router failure paths still scheduled logging_obj.async_failure_handler
as a task while starting logging_obj.failure_handler on a raw thread, so
both handlers mutated the same logging object concurrently. Route them
through dispatch_failure_handlers like the streaming paths already do.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(router): wait on the real logging executor and justify the callbacks global patch
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(logging): submit sync failure handler even when the dispatch task is cancelled
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(logging): justify the executor submit patch
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>
* fix(proxy): make the invalid-model 403 path cheap under a burst of rejections
Keep the wildcard pattern registry in specificity order at registration time
so route() no longer re-sorts every pattern per lookup, and reuse the
standardized failure payload across the async and threaded sync failure
handlers regardless of what a callback did to log_event_type. Rejections
are still logged and observable.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(router): wrap the filtered pattern tuple the way ruff format wants
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(router,logging): assert registry order and callback awaits instead of patching a class
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(router): inject the pattern sorter so the lookup test observes that route() never sorts
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>
* perf(logging): scan large base64 payloads for log truncation off the event loop
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(logging): make base64 offload threshold a plain constant
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>
* fix(hide-secrets): stop redacting benign identifiers and make redaction deterministic
The OpenAI key detector matched `sk-` anywhere inside a word, so `<task-notification>`
became `<ta[REDACTED]>`, and the Base64 entropy limit of 3.0 flagged ordinary quoted
identifiers such as `"application/json"` and model ids. Redaction also iterated a
hash-seeded set, so the same request produced different bytes on different workers and
broke prompt caching.
- require a standalone `sk-`/`sk_` token with a digit (still catches sk-proj-/sk-ant-)
- raise Base64HighEntropyString limit from 3.0 to the detect-secrets default 4.5
- redact overlapping matches longest-first in a stable order
Resolves LIT-7049
* fix(hide-secrets): treat separators as key boundaries and defer sk_live_ to the stripe detector
The standalone-token boundary also rejected keys glued to a preceding `_`, `-`
or percent-encoded delimiter (`openai_sk-…`, `key-sk-…`, `Bearer%20sk-…`), which
the old pattern redacted, and `sk_live_…` was counted by both the OpenAI and the
Stripe detector.
* fix(hide-secrets): keep the openai key scan linear on repeated sk separators
The digit requirement was a lookahead, so every `sk` inside a long
`[a-zA-Z0-9_-]` run re-scanned the rest of that run looking for a digit.
100 KB of `-sk-` took over 5s in the worker's event loop and the proxy
closed the connection without a response. The check now runs once per
match in `analyze_string` instead.
* chore(hide-secrets): remove redundant performance test comment
* fix(hide-secrets): consume complete openai key tokens
* chore(hide-secrets): remove redundant fixture comment
* chore(hide-secrets): remove redundant test docstrings
* fix(hide-secrets): redact whole stripe live keys
* style(hide-secrets): wrap secret sorting key
* fix(proxy): retry deadlocks and requeue spend logs on any DB write error
update_spend_logs dequeued the batch and only retried/requeued on transport
errors. A 40P01 deadlock surfaced as a plain prisma DataError and went through
poison-row isolation, which dropped every row it hit; every other DB error was
re-raised with the batch already gone from the queue.
Treat deadlocks as transient (retry, then requeue), keep them out of poison-row
isolation, and requeue the batch at the head of the queue on any other prisma
error so it lands once the DB is healthy.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(proxy): drop redundant docstrings and tighten test typing for spend-log requeue
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): assert deadlock retries from mock call history instead of mutable lists
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>
* fix(router): hold max_parallel_requests slot until streaming response is exhausted or closed
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(router): normalize deployment_slot once to keep stream_with_fallbacks under the C901 ceiling
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(router): close upstream stream before releasing max_parallel_requests slot
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>
* feat(responses): honor supported_endpoints /v1/responses opt-in for OpenAI-compatible deployments
custom_openai and other generic OpenAI-compatible deployments have no native
Responses API config, so every /v1/responses call is bridged through
/v1/chat/completions. When model_info.supported_endpoints lists /v1/responses,
resolve OpenAILikeResponsesConfig instead so the request is forwarded to
{api_base}/responses, for streaming, non-streaming and mode: responses
deployments alike. Providers with their own Responses config are unchanged.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): drop deployment supported_endpoints opt-in after cross-provider prompt swap
A prompt manager that moves the request to another provider leaves kwargs['model_info']
describing the original deployment; without this the swapped provider was sent an
OpenAI-like /responses request it does not serve.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(responses): carry prompt-swap deployment metadata as a return value instead of a kwargs marker
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>
Adds TestAnthropicChatCompletions to the chat completions regression suite,
registering a claude-haiku-4-5 deployment via /model/new and asserting the
streamed call delivers real content deltas and a tool-forced call returns a
well-formed get_weather tool_call on both the non-streamed and streamed paths.
Covers three P0 registry cells that had no e2e test.
per_user_usage.tsx conflicted with the server pagination that already landed on
staging (default 50 rows, stale-response guard, tag and page-size resets). Took
the staging version and dropped this PR's now-redundant 25-row test for it
Claude-Session: https://claude.ai/code/session_01HkaXiD6gssHnx3kqu1rR8C
Generalizes the heuristic_v2 ceiling from #39468 into a capability table whose
records own their in-process predicate, SQL spelling and refusal wording. The
existing heuristic_v2 capability keeps its own one-router ceiling. A single
customization capability combines operator-defined tier definitions with every
operator-written part of the classifier prompt. The prompt half only applies to
classifier types that call an LLM. The shipped default prompt, classification
rubric presets, tier-label renames and tier model choices remain ungated.
Scope every enforcement point to actual complexity routers. A model-less PATCH
or legacy update now decrypts the stored model before accepting strategy-router
settings, so a regular model cannot acquire a router config or spend a license
slot. Under the existing advisory lock, the cross-pod candidate query returns
only model scalars and the count decrypts and classifies them in process; old
non-router rows carrying a capability-shaped config no longer block a real
complexity router. The signed auto_router license feature makes both ceilings
unlimited.
A slow or unreachable counter store made a pre-relay rejection wait behind
the reservation release before the client saw the error event and the close.
Close first and release in finally, mirroring the relay's own failure path,
so a client that already hung up still gets its reservation released.
The three pre-relay exits of realtime_websocket_endpoint (missing model,
key/model access denied, pre-call rejection such as a rate limit or a
guardrail) returned before the finally that releases the auth-time budget
reservation, so a rejected session pinned the key at the reserved amount
until the counter TTL expired and its next requests got budget_exceeded
while /key/info showed spend 0. A single _reject_realtime_session helper
now releases the reservation before sending the error event and closing,
and release_or_invalidate_budget_reservation shields the release from a
second cancellation and logs, rather than raises, a failing invalidate
fallback so it can never mask the session's own outcome.
The blanket finally release from the previous commit also zeroed the reservation
of successful sessions. Success settlement is enqueued on the logging worker, not
awaited, so the endpoint's finally ran first and released the reservation the cost
callback still had to reconcile, dropping the real spend from the key/team/user
counters.
The relay now stamps a synchronous marker (REALTIME_SESSION_SUCCESS_LOGGED_KEY) on
the shared logging object at the single success-dispatch site, and the endpoint
releases the reservation only when that marker is absent. Refused or failed
sessions, which never log success, still release; successful sessions leave the
reservation for the cost callback to settle to actual spend. Exactly one settler
touches each reservation, so the idempotent reconcile never double-adjusts.
A refused or failed /v1/realtime session never ran the success cost callback
or a failure hook, so its pre-call budget reservation stayed open and kept the
key/team/user spend counters pinned above real spend, 429ing later requests on
the same key until the counter's TTL expired. The endpoint now reconciles the
reservation in a finally, reusing a shared release_or_invalidate_budget_reservation
helper that mirrors the success/failure paths (release to zero, else invalidate
the reserved counters and finalize).
The relayed upstream close message and reason also go through the proxy's
client-facing redaction, so a credential, internal hostname, private IP, or
server path echoed by the upstream never reaches the client verbatim.
* fix(batches): register ownership for every batch create path
Since the team isolation change, the managed files hook decided whether a
response came from a create by looking for the managed input file id on it,
which only the unified input path sets. Batches created from a model-encoded
input file id, a model param, or a raw provider id with ?provider= never got
an ownership row, so they vanished from GET /v1/batches for the key that
created them.
The create endpoint now stamps a create marker on the response before the
hooks run, and the hook keys ownership registration and the batch-created
metric on that marker instead of on the input id format.
* test(batches): assert ownership registration through the managed files hook
The endpoint tests asserted the private create marker, which is wiring, not
behaviour. They now run the create through the real managed files hook and
assert the ownership row is written for the creating key on every create
path, with the unified path driven by a genuine encoded input file id
instead of patched decoders.
* fix(mcp): let config.yaml MCP servers pin server_id
A config-defined MCP server's id is a hash of server_name|url|transport|
auth_type|alias, recomputed on every config load, so editing any of those
fields mints a new id. Every key and team granted the old id via
object_permission.mcp_servers keeps pointing at an id that no longer exists,
and the server disappears from tools/list for them with nothing logged.
load_servers_from_config now uses an explicit server_id from the server's
config entry when present and falls back to the existing hash otherwise, so
grants survive url/name/alias edits. Rejected at config load: a blank or
non-string server_id, two entries claiming the same id, a pinned id already
held by a database-backed server, and a pinned id that is another entry's
server_name or alias (expand_permission_list matches ids before names, so
that one would capture the other server's grants). Because the database
registry loads after the config on startup, a database row that lands on a
pinned config id is reported as a warning from the database reload instead,
where it is decidable; the warning is latched on the shadowed set so the
config-reload timer does not reprint it every interval.
Deployments that do not set server_id keep the exact id they have today.
* fix(mcp): close two more pinned-id capture paths
A pinned server_id equal to an alias supplied through litellm_settings
mcp_aliases was accepted, because the collision index only held the entry's
own alias field. expand_permission_list matches ids before names, so grants
written for the aliased server resolved to the pinning one. mcp_aliases keys
whose target is a config server are now reserved the same way.
A pinned server_id equal to a database-backed server's name, server_name or
alias had the same effect against the database side, and could not be
rejected at config load because the database registry is not loaded yet. The
database reload now warns about it, latched like the existing shadow warning.
* fix(mcp): reserve only the aliases the loader actually assigns
Reserving every mcp_aliases key targeting a config server was too broad in
two ways: the mapping is ignored when the entry sets its own alias, and only
the first mapping for a server is ever applied. Both cases made a pinned
server_id that could never have collided abort proxy startup. Reserve only
the name load_servers_from_config will really assign.
The database capture warning also fired for a database server whose own id is
the config server_id. There the database row wins the id outright through
get_registry precedence, so the shadow warning above it is the accurate one
and the capture message contradicted it. Skip those rows.
Also mark the two litellm-internal patches in the reload test helper, which
the test-quality gate counts; the database reload has no other seam.
* fix(mcp): match the loader's alias check exactly, is None not falsiness
load_servers_from_config consults mcp_aliases only when the entry has no
alias key at all, so an entry setting alias: "" gets no mapped alias. The
collision index used falsiness and reserved the mapped name anyway, which
failed startup on a pinned server_id that could never have collided with it.
* fix(mcp): skip one identifier, not the whole database row
A database row can shadow one config server_id by id and capture another by
name at the same time. Skipping the entire row when its id shadowed a config
entry dropped the second warning, leaving the operator with half a diagnosis.
Skip only the identifier equal to the row's own id.
* fix(mcp): reject conflicting self-pinned server ids
* fix(mcp): validate config server names before building the identifier index
The collision check reads every entry's body up front, so a malformed
entry under an invalid name surfaced as an AttributeError instead of the
name validation error the loader gave before this change.
The handshake error path already runs client-facing error strings through
_redact_string; the relay's _close_client did not, so a secret echoed in an
upstream close reason could reach the client verbatim. Mirror the handshake
path and scrub the close message and reason before relaying them.
* fix(responses): decode JSON-string tool schemas before sending to the provider
A caller that hands a tool schema over already JSON-encoded reached the
Responses API with a string `parameters`, and the provider rejected the
request with a 400 naming the routed model instead of the offending tool.
Decode it at the one place every Responses request converges, and refuse
anything that is neither an object nor a string encoding one.
Collapses the duplicated input/tool sanitization block shared by the
request and compact-request builders into a single owner, so the decode
cannot be wired into one path and not the other.
* test(responses): pin null tool schemas as accepted, and type the parametrized cases
The Responses API serves `parameters: null` and an omitted schema alike, so
neither may raise. Pin both against a future tightening, annotate the
parametrized inputs, and trim the docstrings back to what the code does not
already say.
The refusal predicate also required the session log to be empty, but that
log is not limited to upstream frames. With gemini_live_defer_setup the
handler stores a synthetic session.created before the relay starts, and
the transcription usage flush appends a usage event before the check
runs, so an upstream policy close with no received frames was still
logged as a $0 success. Key the check off the received-frames flag only
The relay's failure dispatch runs the async handler and then the legacy sync
failure_handler for the proxy's callable callbacks. The realtime logging object
carried no async marker, so failure_handler treated the session as a sync SDK
call and fired every CustomLogger's sync failure hook on top of the async one:
Langfuse recorded two ERROR observations per refused session, and OpenTelemetry,
MLflow, Braintrust, Literal AI, DeepEval and New Relic implement the same sync
hook. Plant the _arealtime marker in litellm_params the way aanthropic_messages
and agenerate_content already do, so both dispatchers classify the session async.
A shadow eval job could only be scoped by identity, so "this user's traffic on model X
across every key they own" was not expressible and a models field on the start body was
silently dropped. The job now carries a models list that every target is narrowed to,
matched on the requested model group with model_group_alias resolved on both sides. An
unresolvable name is a 400 at start. Empty means every model, which is what every existing
row reads as. The dashboard start form gains an "Only on models" picker and the job
headline shows the scope.
* feat(otel): stamp litellm.request.route on the LLM call span
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(otel): drop redundant comment on REQUEST_ROUTE
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(otel): Final-annotate route test locals, drop field comment
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): read litellm.request.route off the server span
The LLM call span took the auth-normalized literal path from logging
metadata, which disagrees with the SERVER span wherever FastAPI matched a
template: on /engines/{model:path}/chat/completions the LLM span spelled the
model name while http.route carried the template, so the two spans grouped
into different buckets and the PR's premise did not hold.
Read the value off the span that already holds it. The request's root SERVER
span is anchored per request for parenting, and its attributes stay readable
after it ends, so request_root_http_route() answers from the async close
callback with the same http.route the SERVER span exports: the route template
on a normal route, the literal path where the passthrough hook rewrote it, and
the mount point on an MCP call. Nothing has to re-derive any of that, so the
two spans cannot drift apart.
The route the proxy recorded at auth stays as the backstop for a deployment
whose FastAPI instrumentation never mounted, where there is no server span to
disagree with. Off the proxy the attribute is omitted rather than empty.
---------
Co-authored-by: shivam <shivam@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Yucheng He <yucheng@berri.ai>
client_ack_messages classified a websockets ConnectionClosed raised by
the client socket as the backend closing, so bidirectional_forward kept
waiting on the upstream instead of ending the session. Starlette clients
raise WebSocketDisconnect, but the realtime test client in
tests/llm_translation/realtime raises websockets.exceptions.ConnectionClosed,
which hung test_openai_realtime_simple.py until the run was killed.
Only the receive_text call now maps every exception to
CLIENT_DISCONNECTED; the loop body keeps ConnectionClosed as
BACKEND_CLOSED, since the backend socket is the only websockets socket
touched there.
When the upstream closes while the proxy is forwarding a client message,
the client loop ends before the backend relay sees the close, and the
relay skipped closing the client because it read the client loop's exit
as the client hanging up. The client loop now reports why it stopped, so
a close observed on the backend send still reaches the client with the
error event and the upstream close code
When the provider closes the realtime websocket (for example Vertex Live
refusing the session with 1008 "Publisher model ... was not found"), the
proxy swallowed the close and kept waiting on the client, so the client
sat on an open socket with nothing coming back and the session was logged
as a $0 success
The backend relay now returns the upstream close, and bidirectional_forward
sends the client an OpenAI-style error event naming the upstream code and
reason, then closes the client socket with the same code (or 1011 when the
upstream code is one a server may not send). A session the upstream refused
before sending any frame is logged through the failure handlers instead of
as a success
