- The Redis stall cell now reads the burst key's live counter and requires at least the full failure spend
- Unit tests drop the charging writer double, bare dicts, list mutation and wall-clock timestamps
- Partial-stream test models are frozen
* fix(http_handler): name the timeout phase and the limit that fired instead of 'after None seconds'
Every HTTP verb on both handlers raised litellm.Timeout with the per-call
timeout argument interpolated into the message, which is None whenever the
caller relied on the client default, and always said "Connection" even for a
read timeout. One helper now reads the limit httpx actually enforced from the
failed request and names the phase that fired
Fixes#14635
* fix(rust_bridge): name the timeout phase in the vault timeout message to match the httpx handler
* fix(http_handler): drop the unreachable timeout fallbacks httpx already rules out
* refactor(rust_bridge): drop a doc comment that restated the timeout helper
OAuthScopeResponse holds a tools/call response's http.response.start so a managed OAuth scope challenge can replace it with a 403. Once the call was allowed, the held start still waited for the next body message, which for a slow tool is the 15s SSE keep-alive ping. Send it as soon as the call is allowed instead.
Fixes#46659
* fix(router): stop concurrent deployment deletes from crashing deployment lookups
delete_deployment and upsert_deployment shrink model_list before they repair
the index maps, so a lookup on another thread in that window read a stale
index: IndexError, or the neighbouring deployment and its credentials. An add
racing a delete could corrupt the index for good.
One RLock now covers model_list and the three index maps. Delete, upsert, add
and the set_model_list reset change them under it, and every lookup that turns
an index into a deployment reads under it.
Fixes#45242
* test(router): release the held delete when the read reaches the index lock
The race tests released the held write after a fixed 0.2 s, so a reader the runner
stalled past it read the repaired list and passed on unfixed code. The read now runs
on its own thread and the write stays held until the read returns or reaches the
router's index lock; the 10 s bound only fails a broken synchronization.
The integration test gains the audit cells: chat, streaming, Responses and
Anthropic messages calls (sync and async, by model name and team public name)
during a held delete, the same for upsert, add and set_model_list, a request for
the deployment being deleted, a provider error, a double delete, and thread and
async bursts during upsert churn.
* test(router): cover the early return of a delete and the team filter of get_model_ids
delete_deployment now returns None from inside the index lock, and get_model_ids by
name reads its deployments through _deployments_named; pin both: deleting an unknown
id returns None and leaves the list alone, and exclude_team_models drops only the
team deployment.
* fix(router): read deployments by name under the index lock outside the Router
The response polling provider check, the complexity router's adaptive pool and the
savings baseline turned the name index into list positions themselves, so a concurrent
delete could still hand them a stale slot. They now read through Router.deployments_named,
which takes the deployment index lock.
* fix(router): type the locked deployment lookups as DeploymentTypedDict
deployments_named and _team_deployments_named returned tuple[Any, ...]. They now return
DeploymentTypedDict rows, which is what model_list holds. The adaptive pool setup in the
Router and the complexity router drop their branches for pydantic Deployment rows, which
these lookups never return, and should_include_deployment accepts the typed row.
---------
Co-authored-by: yucheng-berri <293926326+yucheng-berri@users.noreply.github.com>
* feat(ui): simplify auto router setup with deployed model defaults
* fix(ui): recognize catalog Opus deployment formats
* fix(ui): address auto router setup review gaps
* fix(ui): preserve Gemini auto router suggestions
* fix(ui): prefer current aliases and wait for team selection
* fix(ui): default auto routers to classify each user turn
* fix(ui): parse Grok snapshot dates separately from releases
* refactor(ui): use non-mutating auto router sorting
* feat(ui): allow custom models in simple auto router setup
* fix(ui): keep Aria accessible in the dashboard header
* test(ui): wait for every pagination fixture log
* test(ui): update assistant setup for Aria branding
* fix(guardrails): exclude litellm_metadata from pass-through guardrail scan text
The pass-through guardrail translation json-dumps the request body as the
scan text when no request_fields are configured. The unified guardrail hook
injects litellm_metadata into that body first, so the dump carried the
caller's key metadata. Exclude it alongside metadata and litellm_logging_obj;
vendors still receive it through request_data.
* test(guardrails): type the pass-through scan-text tests per AGENTS.md
* fix(otel): decide per-team Arize sampling from the trace id so every pod agrees
* test(otel): cover per-team sampling across pods, workers, endpoints and restarts
* fix(otel): remember only a failed tree's sampling verdict so a later failed call answers to the error rate
A successful tree's verdict follows from the trace id alone, so remembering it only made a later failed call of the same trace follow the success rate on the pod that served both, while other pods judged it at the error rate
* test(otel): flush every worker before the restart cell kills it and cover a failed call after a dropped one
* fix(otel): decide team sampling per request, not per trace
Calls sharing a caller's trace id are separate requests, keyed by their root span.
Each is decided by the trace hash at the rate for its own outcome, and only its own
late spans follow its verdict, so no pod applies an earlier call's verdict to a later
call of the trace
* fix(otel): forget a span's request root when it ends, not when newer spans start
The root of each open span was kept in a first-in-first-out table of 16384 entries,
so a span two levels below the root that stayed open while the worker started more
spans than that lost its entry and ended under its parent's id. Its request then
never closed and was held from every sampled destination until shutdown.
Each open span now keeps its entry until it ends, and a span starting under a parent
that has already ended reads the root from the parent span itself
* fix(otel): decide a request as its spans end once it falls out of the open span table
A request whose span was pushed out of the table of open spans, because the worker
held more than 16384 open at once, ended that span under another key and was held
from every sampled destination until shutdown. Pushing a span out now drops its
request's open count too, so the request is decided as each of its spans ends, the
way one pushed out of the open request table already is
* test(otel): annotate the team sampling test locals as Final
* test(otel): build the caller's remote parent in one helper instead of rebinding it per trace
* fix(otel): keep each open span's request until it ends, bounded by the open requests
The span table had its own bound of 16384, so a worker holding more spans open than
that pushed out the oldest, and a failure in such a span no longer reached its
request. Each request now holds the ids of its open spans, and a span's entry in the
lookup goes when the span ends or when its request is pushed out of the open request
table, the same bound main already applies. A failure anywhere in a tracked request
reaches its verdict however many spans are open
---------
Co-authored-by: yucheng-berri <293926326+yucheng-berri@users.noreply.github.com>
* fix(streaming): stop re-raising and re-logging a failed stream on every next()
A stream that raised keeps its provider iterator alive, so a consumer that iterates past the error re-detected, re-raised and re-logged the same failure on every call, flooding failure callbacks for one request. Latch the failure and re-raise it without logging again.
* fix(streaming): drop the retained traceback before re-raising a failed stream
* test(streaming): count failures through a real logging callback and require a clean stop on later reads
* test(streaming): require the stuck stream to close early and release its socket
The provider stream now records a GeneratorExit and caps its chunks, so draining it or detecting the loop late fails. Two wire cases check a sync and an async stuck stream release their TCP session.
* test(streaming): avoid failure recorder name collision
---------
Co-authored-by: Devin AI <devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: moe-berri <moe@berri.ai>
* build(docker): move to newer wolfi-base and nginx base images
Bump the pinned wolfi-base digest in every Wolfi Dockerfile to 360896ef, which
ships brotli 1.2.1-r1 and ca-certificates-bundle 20260909-r3. The previous
digest already carried libbrotli 1.2.0-r6 as an OpenSSL dependency, and
`apk add --no-cache` never upgrades a package the base already has
Move the UI image to nginx 1.31.8-alpine3.24, whose CA bundle is the 20261020
Mozilla set, and upgrade brotli-libs in its runtime stage since no nginx
digest ships 1.2.1 yet
* build(docker): copy only runtime files into the gateway, backend and migrations images
The component runtime stages copied the builder's whole /app, which is the
entire repository plus build output: six lockfiles, helm, terraform, docs,
.env.example and the Rust build tree. Measured at runtime with an audit hook
on a booted container, the gateway reads only gateway/, litellm/ and
enterprise/ from /app, the backend only backend/, litellm/ and enterprise/,
and the migrations job only litellm/ and run.py. The pyproject.toml read is
prisma-python's optional [tool.prisma] lookup, which ours does not define
Copy just those, plus .venv, the entrypoint script and schema.prisma for the
gateway and backend. PYTHONPATH=/app and the in-Dockerfile asserts are
unchanged
Preserve oversized JSON integers in proxy requests and model rewrites
Parse large bodies in a worker thread to avoid blocking request handling
Fixes#47947
Co-authored-by: Jeevan Babu Gotru <114808955+jeevan1098@users.noreply.github.com>
Co-authored-by: moe-berri <moe@berri.ai>
* fix(deepseek): pass reasoning_effort through to the API
DeepSeekChatConfig.map_openai_params mapped reasoning_effort to a binary
thinking {type: enabled/disabled} toggle and discarded the effort value, so
DeepSeek always ran default high effort. High-effort reasoning can run away
and exhaust max_tokens during reasoning, returning empty content. Preserve
the reasoning_effort value into the request body when thinking is enabled so
callers can bound effort (e.g. reasoning_effort="low"). Backward-compatible.
See https://api-docs.deepseek.com/guides/thinking_mode
* test(deepseek): reasoning_effort passthrough coverage
* fix(deepseek): exclude reasoning_effort=none sentinel from passthrough
"none" is the thinking-OFF sentinel, not a real effort value. Passing it
alongside explicit thinking:enabled would send contradictory instructions to
DeepSeek. Only forward real effort values (low/medium/high/xhigh/max).
Addresses review feedback on the reasoning_effort passthrough change.
* refactor(deepseek): name supported effort values and type mapped params
* fix(deepseek): keep mapping types local to the adapter
* fix(deepseek): remove cast and type mapped return
---------
Co-authored-by: moe-berri <moe@berri.ai>
* fix(proxy): reconcile team member total_spend with current cycle spend
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): read /team/list as a list and drain the scripted provider
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(repositories): compose the membership reset writer and drop SQL text asserts
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: type the cascade repair test and flatten memberships without a stacked comprehension
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: keep the reset job test file's existing typing untouched
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: joshua <joshua@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(cost): parse 12-hour cache write rates for batch and ultrafast tiers
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost): bill 12-hour batch cache writes once when only the 12hr batch key exists
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>
* fix(ui): show "Never resets" instead of "n/a" for an unset budget reset period
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(ui): label unset reset period as Keep current/No change where blank keeps the stored value
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(ui): show a single Never resets option in create team when no default is set
Without a default team budget_duration, the create team Reset Budget dropdown
offered both "Not set" and "Never resets", and both created a team that never
resets. The explicit Never resets item is now only shown when there is a
default to opt out of, otherwise the empty option itself reads "Never resets"
Also makes the Reset Budget create tests use chooseSelectOption so they no
longer flake on the select popup's pointer-events transition
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
