The /vllm and /azure router-model passthrough branches called
llm_router.allm_passthrough_route directly with no request metadata,
so the cost callback saw no user_api_key and no
user_api_key_budget_reservation. Spend for a budgeted virtual key hit
neither the key's spend nor the spend logs, and the reservation minted
at auth into the shared Redis counter was never released, drifting the
counter up until the key falsely tripped BudgetExceededError.
Thread the authenticated key's attribution metadata into both calls via
the same builder add_litellm_data_to_request uses, so the cost callback
attributes spend and reconciles the reservation. Regression tests cover
both branches.
`exception_type` decides the class and status a caller sees for every provider
failure, across 190 raise sites, and the tests for it were written one incident
at a time. Nothing said what a plain 401 from any given provider should be, so
mutating a raise site went unnoticed: swapping the class at each of the 190 in
turn, the mapped test file caught 15.
Adds two tables asserted end to end through `exception_type`: 25 providers by
the 9 upstream statuses, and the three error shapes the router branches on
(a full context window, a content policy block, a timeout). The same 190
mutants now fail 97 of them.
The tables record today's behavior, uneven where it is uneven. cloudflare,
ollama and vllm map no status at all, so every failure reaches the caller as a
500. A full context window is recognised by 15 of the 25, and a content policy
block by 11, which bounds where `context_window_fallbacks` and the content
policy retry policy can fire.
* fix(ci): make the migration DDL guard run, and stop it reading comments as SQL
TestMigrationSQLIdempotency requires guarded DDL across litellm-proxy-extras
and has never run in any job, so the convention eroded quietly. Four of its
assertions fail today, and it was allowlisted rather than wired up because
fixing the migrations is not an option: Prisma checksums an applied migration,
so editing one breaks `migrate deploy` for every existing install.
Two things were wrong with the guard itself. It scanned raw lines, so Prisma's
own `-- CREATE INDEX CONCURRENTLY ...` explanations counted as the statements
they describe, which is two of the reported migrations. And it had no way to
say "these predate the rule", so the only options were editing immutable files
or leaving the whole file unrun.
Comments are now stripped before matching, on the drop-column rule too, and the
migrations that already violate are named once in _PRE_GUARD_MIGRATIONS. The
rules bind everything after them, so a new migration with bare CREATE TABLE,
ADD COLUMN, CREATE INDEX or an unguarded ADD CONSTRAINT now fails a check
instead of landing unnoticed.
That set is 14 migrations, not the 13 previously recorded, measured after
comment-stripping. It can only shrink: a test fails if an entry names no
migration on disk, and another fails if an entry no longer violates anything.
The file now runs as a proxy-extras shard and comes off the coverage allowlist.
* fix(ci): strip block comments in the migration guard too
Prisma opens a destructive migration with a /* Warnings: You are about to
drop the column ... */ header. Nothing in the tree trips a rule on that text
today, but it is prose about a statement rather than the statement, and the
line-comment fix left the class open. Bodies are blanked rather than removed
so the reported line number still points at the real statement.
The allowlist recorded eight files in tests/local_testing, 118 tests, that
every job globbing that directory then deselects: local_testing_part1 and
part2 carry `-k "... and not caching and not cache"`, and the other three keep
one unrelated keyword each. They counted as covered while running nowhere.
Five of the eight need nothing. Measured with no provider credentials and no
Redis: test_cache_preset_key, test_caching_handler, test_prompt_caching,
test_responses_stream_cache_keys and test_unit_test_caching pass, 45 tests
together, and they now run as a caching-local shard. The other three stay
allowlisted with what they actually need recorded rather than a question:
test_caching wants Redis and a provider key for 37 of its 65, disk-cache wants
OPENAI_API_KEY for 2 of 4, gcs-cache wants GCS credentials for all 4.
Taking them off the allowlist exposed a gap in the slice guard itself: it
reasoned only about CircleCI `-k` expressions, so a file every slice drops read
as unrun even when a workflow names it outright. It now credits workflow
test-paths the way the census already does, and only workflows, so a tree only
CircleCI globs is still reported.
* feat(ci): gate patching of SDK internals in tests as TQ008
TQ002 catches the narrowest symptom of the suite's dominant mocking idiom,
patch X then assert only that X was called. The idiom itself is wider: tests
reach for litellm's own functions instead of faking the wire, so they pin how
the code is wired rather than what it does, and a test that patches internals
but makes weak real assertions trips nothing today.
TQ008 counts patch targets rooted at `litellm`, both the dotted string form and
the attribute chain handed to patch.object, and ratchets like every other rule.
Mocking anything outside the SDK is untouched: respx, httpx transports and
third-party clients do not trip it, which is the point, since those are the
patterns this is meant to move the suite toward.
Seeded at 9,643, in line with the ~9.4k patch sites an independent grep found
in the mirror. The burn-down horizon is long; the value here is stopping the
flow rather than clearing the stock.
Five existing rule tests patched `litellm.completion` incidentally and now
report TQ008 alongside what they were pinning. Their expected values are
updated to the accurate pair rather than loosened, so they keep failing on a
regression in either rule.
* test: add TQ008 to the shipped-budget rule canary
* fix(ci): resolve imported SDK names in TQ008
patch.object(handler.OpenAIChatCompletion, ...) after a from-import reaches the
same internal as the dotted string form, but the rule only saw the bare local
name and let it through. Import bindings are now resolved to the path they
stand for, so the aliased, renamed and from-imported forms all read alike and
the reported target is the real one.
That is 1,496 patches the ratchet could not see, so the TQ008 limit moves from
9,643 to 11,139. Third-party names and locals with no SDK import behind them
stay unflagged.
* perf(ci): fan the budget checkers out across cores
check_type_discipline.py and check_test_quality.py each walk a few thousand
files and parse every one, single-threaded. In the lint job those two steps
measure 2.3 and 1.6 minutes, second and third behind dependency install, and
lint is the slowest required check on 9 of the last 10 merged staging PRs.
check_file is already pure per-file work, so the walk fans out over a process
pool with no change to what either rule reports. Callers sort, which is what
keeps output order stable when results land out of order. Runs below
PARALLEL_MIN_PATHS stay serial rather than pay for process startup, and the
worker count is capped so a large runner does not oversubscribe.
Measured locally over the same trees, output byte-identical both times:
type-discipline 17.8s -> 3.0s over litellm/ (78,768 report lines), test-quality
14.4s -> 2.3s over tests/ (6,321 report lines), per-rule counts unchanged.
* test(ci): type the fan-out helpers and skip the comparison on one core
The 20% promotion that runs to 2027-01-31 covers every gemini model, not
just the 2.5 pair, so a constant naming two of them implied the other four
were exempt. Six covered entries live in the registry: two store the
discounted rate and four store list, which is a pre-existing overcharge
this branch does not touch, since it only adds cache fields and derives
them from each entry's own input rate. Name both groups for what they
store, pin the expiry, and tighten the tolerance to 2e-4.
The two gemini 2.5 entries price a factor of 1.25 under the published DBU
table because the published figures exclude a 20% promotion that runs to
2027-01-31. The previous constant name and test called them an older vintage
awaiting a refresh, which would have led a future reader to scale them up and
overcharge. Pin the discount and the cache relationship instead.
databricks-claude-fable-5 was the only fable-5 entry in the registry
declaring supports_vision false, and the only one of the five new
entries to do so.
Only the five new models were pinned against the published DBU table,
so the 26 cache literals added to pre-existing entries were checked by
nothing independent. Extend the table to all 33 entries carrying cache
rates and assert both cache fields against it for the 31 that take the
published rates, leaving the two older-vintage gemini-2-5 entries to
their existing guard.
Also widen the cache-declaration guard to both cache fields, and
replace the single-model equals-input assertion with one that covers
all 14 entries publishing no cache rates.
Cache rates were derived as ratios of the dollar input rate (1.25x write,
0.1x read) while input and output derive from the published DBU table
times $0.070. Databricks publishes cache write and cache read DBU per
model, and those are not exact multiples of the input DBU, so the two
rules disagreed by up to 0.1 percent.
Rewrites 43 cache literals across 31 entries to published_cache_DBU x
$0.070. Skips databricks-gemini-2-5-pro and databricks-gemini-2-5-flash,
whose input and output rates predate the current table by a 1.25x
increase; their cache rates stay tied to their own input rate so each
entry remains internally consistent.
Replaces the ratio assertions with a test pinning the absolute published
DBU figures for the five new models, and adds a test pinning the
older-vintage exception. Corrects the metadata note on the five new
entries, which claimed the reference-only *_dbu_cost_per_token fields
drive cost calculation.
The new supports_prompt_caching assertion reads a capability this branch
adds to the registry, so it only holds against the bundled map. CI leaves
LITELLM_LOCAL_MODEL_COST_MAP unset and fetches main's copy, which lags the
branch until merge, so the test failed there while passing locally.
Use the local_model_cost_map fixture the repo already provides, matching
what the other two test files in this change do.
The shared cost calculator treats a missing cache rate as free, so routing
Databricks through it billed cached tokens at zero on the 14 entries that
publish no cache pricing. On a 10,000 token prompt with 8,000 cache reads
that is $0.0010000 against the correct $0.0050001, a fivefold undercharge.
Those entries now declare cache rates equal to their input rate, which is
what a model with no caching discount should charge, and a test pins every
priced Databricks entry to declaring cache rates so no future entry can
regress into it.
Also repoints the provider-neutral generalization test off an id the new
Opus 5 entry now shadows, adds backup-to-main parity tests for the five new
entries, pins that Databricks Claude is never auto-injected with cache
control despite reporting caching support, and trims the Sonnet 5 pricing
note, which is served on an unauthenticated route.
* test: add regression coverage for twelve closed issues
Adds targeted regression tests for behavior that was fixed but left ungated,
so the fixes cannot silently regress:
- #33772 openai cache_write_tokens cost
- #34309 Responses API cache cost_breakdown
- #35363 /v1/responses batch spend
- #36619 auto-router api_base/api_key leak on a shared model name
- #35359 batch fallbacks within the owning model group
- #36523 passthrough streamed Responses spend log
- #36646 passthrough embeddings spend log
- #37147 non-object metadata on create_batch is a 400
- #35362 unscoped list files reads the managed-file store
- #33221 gpt-5.6 bridges to Responses on function tools alone
- #34487 LLM complexity classifier runs for every caller metadata shape
- #35124 streamed /v1/messages emits success logging on both bridges
Cost assertions read rates from litellm.model_cost rather than hardcoding
dollar amounts, so they do not drift on repricing.
* fix: stop the new regression tests polluting and tripping over shared global state
Two shard failures, both from global state the new tests share with their
neighbours rather than from the behaviour under test.
test_main.py's local_cost_map pinned litellm.model_cost but left the
get_model_info lru_cache warm, so completion_cost billed at whatever prices
were cached earlier in the process while the assertions read the pinned map.
Clear the cache on both sides of the fixture, matching the local_model_cost_map
fixture in tests/test_litellm/conftest.py.
The anthropic messages streaming tests called GLOBAL_LOGGING_WORKER.flush()
on whatever queue happened to be around. A queue left non-empty by an earlier
test is still bound to that test's loop, so join() either hangs or raises
"bound to a different event loop". Rebind to the running loop before the call
and wait for the captured payload instead of a fixed sleep.
The introductory DBU rates run through 2026-08-31 and pricing carries no
expiry date, so a static introductory entry would undercharge by a third
from September 1 and let spend outrun enforced budgets. Ship the standard
rates, which match Sonnet 4.5 and 4.6, and keep the introductory numbers
in the entry notes.
Also give the new cost calculator tests full type annotations.
Databricks cost calculation multiplied every prompt token by the input rate, so
a cache read cost the same as an uncached token. Route it through
generic_cost_per_token, which already understands cache reads and cache writes,
and add the cache rates the registry was missing.
Adds Claude Opus 4.7, Opus 4.8, Opus 5, Sonnet 5 and Fable 5 on Databricks.
An empty content list, or one holding only opaque blocks, still lets the
provider-bound branch replay the summary text. The inspection path treated
any non-None content as final, so that replayed text stayed invisible to
guardrails and token counting.
The field is declared optional on OpenAIFileObject and its own docstring says it
is absent on every upload guardrails did not touch, but the /v1/files routes have
no response_model, so FastAPI falls through to jsonable_encoder with exclude_none
off and serialises the unset default as an explicit null. Every create and
retrieve response on a proxy with no guardrails configured at all picked up a
litellm_batch_guardrail: null it never had before, and so did every row of a file
list, since those rows are the same object.
A wrap serializer drops the key only when nothing set it, so the populated report
still reaches the wire intact, including a record whose guardrail is null. The
managed-files list route spreads a stored file_object blob rather than the model,
so rows persisted before this lands keep their null until it is dropped there too.
* test(e2e): send no-cache on every cacheable request body, opt in only where a hit is the assertion
The e2e proxy runs with the response cache on, so any test that re-sends an
identical chat, messages, responses, completions, embeddings or rerank body
reads back a redis copy of an earlier call instead of reaching the provider.
Five tests in the last week failed that way. Default cache: {"no-cache": true}
on those request models and pass cache=None only in the two tests whose
assertion is the cache hit itself.
* test(e2e): give image edits and OCR a 180s client timeout
Both routes wait on providers that can legitimately take longer than the
60s transport-wide request timeout (gpt-image edits, Azure Document
Intelligence), and a client-side read timeout there fails a green request.
post/upload now accept a per-call timeout like get already does; only those
two call sites use it.
* test(e2e): rerun once on network errors and upstream 5xx only
Assertion failures still fail on the first attempt; only an outcome whose
error string carries the e2e_http network kind or a 5xx status gets one
more try. Test Engine records every attempt, so the flake rate stays
visible while a single provider blip no longer reds the rc run.
* test(e2e): let the reseed burst survive one upstream failure and print why
The burst is the precondition, not the property: one 5xx among six
concurrent calls still leaves five workers racing the cold counter, which
is what the reseed assertion measures. Two or more failures still abort,
and the failing bodies are now in the message instead of only the status
codes.
* test(e2e): keep polling Jaeger through a transient query failure
poll_traces_for_call already waits up to POLL_TIMEOUT for spans to land,
but a single refused connection to the query API failed the test on the
spot. Jaeger restarted twice during today's gate runs (19:05 and 19:41
UTC, each under a minute) and took ten and three otel tests with it while
the same tests passed on the rc build minutes later. A network failure
now counts as not-yet inside the same deadline; if Jaeger is still
unreachable when the deadline passes the test fails with that error, and
any non-network failure still fails immediately.
A reasoning input item that carries only summary text is replayed to the
provider as reasoning_content, so inspection-only callers must see that
text too. They used to fall through to the generic content branch, which
reads content and drops a summary-only item, leaving guardrails and token
counters blind to text the model still receives.
litellm_request_total_latency_metric's start_time is set inside
common_processing_pre_call_logic, which only runs after user_api_key_auth
has already succeeded, so the metric silently excluded authentication and
pre-call setup time despite being documented as total request latency. The
sibling litellm_request_queue_time_seconds metric had the same problem:
its arrival_time was captured after auth too, despite its own comment
claiming to track when the request arrived at the proxy.
request.state.litellm_received_at is now stamped unconditionally at the
very first line of user_api_key_auth (previously only when OTEL was
configured), giving a timestamp that precedes all auth work. Both metrics
now derive from it: queue_time_seconds genuinely spans arrival through the
start of pre-call processing, and the total-latency metric adds that
queue time on top of its existing start/end window so it becomes true
end-to-end latency.
queue_time_seconds ends exactly at start_time rather than a separately
captured timestamp, so its window and the total-latency window share a
boundary instead of overlapping and double-counting a few lines of setup
work on every request.
When get_team_object fails, the centralized auth gate rebuilds the team
from the token's own fields. A token whose team row was missing when the
key was read carries team_models=[] and team_blocked=False, and the
model-access check reads an empty model list as every model, so the
rebuilt team grants more than the real team ever did.
get_team_object reported a deleted team and a database that would not
answer as the same 404, so the fallback could not tell a definitive
answer from a degraded read. Raise a TeamNotFoundError subclass, still a
404 with the same detail so every other caller is unaffected, only when
the database answers and the row is absent.
A team that is provably gone now refuses, and no setting overrides that.
Otherwise the grant is merely unknown: a token carrying one may vouch,
since replaying a recorded grant cannot widen it, and a token carrying
none may not. allow_requests_on_db_unavailable still opts back out there,
and is only consulted once the failure is known to be a degraded read.
The Admin UI mints every session key against the UI_TEAM_ID sentinel,
which by design never has a team row, so every UI request hit the new
refusal with no override. Exempt UI_TEAM_ID explicitly so it keeps
reconstructing from the token unconditionally, matching how the MCP
handler and agent_permission_handler already special-case it.
Resolves LIT-5522
* fix(auth): resolve team object_permission independently in the unresolvable-team fallback
When get_team_object fails for a token's team_id, _user_api_key_auth_builder
reconstructs a LiteLLM_TeamTableCachedObj from the token's own cached fields,
carrying team_object_permission_id but leaving object_permission unset. That
silently dropped any vector-store or MCP restriction the team carried,
granting more access than the token's own object_permission_id vouches for.
Resolve the object permission by its id directly via get_object_permission,
independent of the unreadable team row, matching how every other consumer of
a team's object_permission (vector store access checks, MCP tool/server
resolvers) already treats an unresolvable team as "no restriction at this
level" and re-resolves on its own.
* fix(auth): trim ticket references and narrative docstrings per Greptile review
Drop the LIT-5539 ticket id from test names and fixture strings, and shorten
both the new helper's docstring and the regression test docstrings to their
contracts rather than restating the fix's history.
The team roster update, the user.teams update, the team membership
delete, and the team-scoped verification token delete ran as four
sequential writes with no transaction around them, so a failure
between any two left the removal half applied. Thread a single
prisma transaction through all four writes, following the same
tx.<table> pattern /team/member_add and /team/member_update already
use, so either all four land or none do.
The Presidio guardrail masks messages in place inside pre_call_hook, but three
paths independently persisted or emitted the raw pre-guardrail data: the
SpendLogs proxy_server_request body snapshot (taken before the hook runs),
a verbose_proxy_logger.debug dump of the raw request, and logging_only mode's
async_logging_hook, which never masked the model's response before it reached
external logging callbacks.
Resolves LIT-6015
The session lookup reads spend logs straight out of the database, so a
follow-up sent seconds after the turn it chains off found nothing while the
row was still queued in the worker that served it, and the conversation was
dropped without an error. Responses calls now ask the spend-log writer to
flush on its next pass instead of waiting out its poll interval, and the
lookup gives a just-finished turn a short second chance.
Replaying a session also accepted `input` only as a string or a single dict,
so the standard list shape dropped every user turn and left the model with
assistant messages alone.
The standalone migration entrypoint re-runs `prisma generate` after the
migration completes. That refresh writes into the installed prisma package in
site-packages, which an arbitrary non-root uid cannot do, and which no uid can
do under a read-only root filesystem. Both are supported configurations of the
migrations Job: helm/litellm-helm/tests/migrations-job_tests.yaml asserts
runAsNonRoot, runAsUser and readOnlyRootFilesystem all render.
The write has always failed there, but the failure used to be swallowed. Making
migration failures fatal turned it into a hard exit 1, so a Job that applied
every migration correctly now reports Failed and blocks the rollout it was
supposed to gate.
The refresh is redundant in the shipped images: every Dockerfile generates the
client at build time from the same baked schema, copies it into the runtime
stage, and asserts it resolves there. It stays load-bearing only for a source
checkout, where CircleCI runs the entrypoint under `set +e` and ignores the exit
code anyway. So the call stays and only its exit code stops propagating;
migration failures are still fatal.
image-scan never ran on the change that introduced this, because its path filter
did not list the entrypoint it exercises. Add prisma_migration.py and
entrypoint.sh so the non-root offline migration test gates them from now on.
Write the fallback reasoning item id back to the cache so the
reasoning-done path and the completed snapshot cannot drift apart, and
cover the shared delta id and the snapshot alignment with tests.
Guardrails, token counting and rate limiting share the input transform with
the provider path, so moving reasoning onto reasoning_content hid it from
them. Provider-bound callers opt in with replay_reasoning.
The ceiling used to go through `int(... or 3)`, so anything `int()` accepted
worked. Tightening the new shared validator to `isinstance(int)` turned a
config that boots today into a proxy that refuses to start, because
`max_agentic_loops: os.environ/MAX_AGENTIC_LOOPS` is resolved to a string
before it reaches either check, and a YAML-quoted "5" is a string too.
Accept ints, integral floats, and strings that parse to a whole number. Keep
refusing bools, fractional floats, words, and anything below 1.
Align the response.completed item IDs by copying each output item rather than
writing to it in place, and move the regression cases into the existing
completion-response and image-generation test modules.
Streaming /v1/responses over the completion bridge minted a fresh resp_{uuid4}
for every response, while spend tracking stored the inner chat completion id as
request_id. The session lookup queries on request_id, so a follow-up sent with
that response id matched no rows and the prior conversation was silently
dropped. The iterator now pulls the first upstream chunk before emitting
response.created, so created, in_progress and completed all carry the same
encoded chat completion id.
Two more ways the same history went missing:
- The session lookup only read spend logs already written to the DB, so a
follow-up sent inside the batch writer's window found nothing. It now also
reads the rows still queued in memory.
- Input was only accepted as a string or a single dict, so the list shape the
Responses API actually sends dropped every user turn from the reconstructed
history.
* fix(otel): emit LLM Call spans for speech, image, moderation, ocr and transcription
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): log the image request before caller headers are merged in
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): map non-chat routes to standard genai operations
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): stop caller image headers aliasing the logged request body
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): keep resolved api_base in async moderation pre_call
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): log resolved client endpoint for speech pre_call
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(otel): justify mutable request payloads in speech and image pre_call
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(otel): keep caller headers out of the logged speech request body
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>
The ceiling was only checked at the feature level, on
litellm_settings.websearch_interception_params. The per-deployment
litellm_params.max_agentic_loops, which wins over it, went straight into
int(kwargs.get("max_agentic_loops", 3) or 3), so a 0 was swallowed by the
falsy fallback and read as the default 3. Asking for the tightest ceiling
handed you the loosest one. A non-integer booted the proxy and then failed
every request to that model with "invalid literal for int() with base 10".
Both settings now share one validator, which names the field it rejected,
and the per-deployment value is checked while the model list is read at
startup so a bad value stops the proxy rather than surfacing per request.
The check sits in load_config rather than on LiteLLM_Params because the
proxy builds its router with ignore_invalid_deployments=True, where a
validation error drops the deployment silently instead of refusing to
start. This is the same placement the complexity_router_config plugin
check already uses.
Chat completions read the same key through a separate path that turned 0
into 1 and true into a ceiling of 1, so it now shares the validator too
and the key means one thing on both surfaces.
encrypted_content on a reasoning input item is written by LiteLLM's own
_encode_thinking_blocks as a JSON array of Anthropic/Bedrock thinking
blocks, so decode it back and replay the signed blocks on the assistant
message instead of dropping them. Providers without a native
ResponsesAPIConfig now keep the verifiable chain-of-thought across turns,
and prior-turn reasoning stops reaching the provider as visible
assistant text.
The Chat Completions -> Responses bridge stamped the upstream chatcmpl-*
ID onto message output items, so replaying bridged history into native
OpenAI Responses failed with "Expected an ID that begins with 'msg'".
Image generation calls were minted as chatcmpl-*_img_N instead of ig_*,
and reasoning items used a salted hash() that is not stable across
processes.
Streaming minted msg_* for its incremental events but rebuilt the
response.completed snapshot through the same broken transform, so the
snapshot contradicted the events it had just sent and streaming clients
hit the same 400. The snapshot now reuses the IDs already streamed.
Fixes#27333
