Azure AI is the only provider that retries a 422 inside the translation
layer: when the endpoint rejects a field, litellm drops that field and sends
the request again, up to twice. That is the difference between a customer's
tool call working and coming back as a hard 400, and none of it was covered.
The retry loop in llm_http_handler.py is 13,419 lines of source against a
0.20 test-to-source ratio, and nothing exercised this path at all.
Drives real litellm.completion and litellm.acompletion calls against a
recorded Azure AI endpoint, so the assertions read the bytes that actually
went over the wire rather than a mock's call list. Nothing internal is
patched: respx fakes the HTTP boundary and the provider config, retry loop
and serialization are all the real ones.
Pins:
- a tool field the endpoint rejects is dropped and the call retried, and the
caller gets a normal completion
- the retry changes only the field the provider named
- a provider that keeps rejecting stops after exactly two attempts
- a rejection the provider cannot fix is not retried at all
- an extra input outside a tool is retried only when drop_params was asked for
Mutating the source confirms these bite: raising the retry cap from 2 to 3,
and making the tool-level field check always return False, each turn the
suite red.
The async cases pin the transport to httpx, because the aiohttp default
carries its own transport that an httpx-level fake cannot intercept. Without
that the two async tests reached the real Azure endpoint and failed on a 401.
proxy/_types.py is 4,965 lines holding 202 request and auth models with 27
validators, and its mapped test file was 32 lines covering one of them. The
validators decide what a caller is allowed to send, so a silent change here
reaches customers as a request that should have been refused and wasn't, or
the reverse.
Pins the contracts that carry real consequence:
- the server-only MCP markers and via_virtual_key are stripped from any
caller-supplied input, so they cannot be forged through the constructor or
model_validate, while the server can still set them by assignment
- a virtual key is hashed out of the auth object, and Bearer-prefixed and
bare keys hash alike
- a JWT issuer must name an audience or opt out of one, never both and never
neither
- a boolean spend reset is refused rather than silently read as 1.0 or 0.0
- a key or user update must say which key or user it updates
- a key lookup naming nothing is refused rather than matching everything
- an organization member cannot be given a role that lives outside an
organization
- an audit log stores the key it recorded a change to only masked, and keeps
the non-secret fields intact
Every case asserts the observed value rather than that a call happened, and
nothing is patched. Verified by mutating the source: dropping the marker
strip, flipping the audience rule's and to or, letting booleans through the
spend reset, treating an empty key list as naming a key, and disabling the
role check each turn the suite red.
Moves the file to the path that mirrors litellm/proxy/_types.py, which the
old file's own first line already said it should have been at, and carries
its two tests over.
`_delete_cache_key_object` awaited the Redis delete unguarded, so any cache
backend error surfaced as a failure on an operation that had already been
committed. A Redis ACL that denies DEL on LiteLLM's unprefixed token-hash keys
turned a persisted `/key/update` into `400 Authentication Error, No permissions
to access a key`, and `/key/block` and `/key/regenerate` into 500s
Make the helper best-effort, the way `delete_cache_team_object` and
`delete_cache_key_objects` on either side of it already are: log the failure and
carry on. Nothing ends up staler for it, since the in-memory entry is dropped
before the Redis round trip and the write has already committed, so raising only
misreported a success
* feat(prometheus): configure deployment caller identity
* test(prometheus): satisfy strict caller identity lint
* fix(prometheus): align caller identity on latency metrics
* fix(prometheus): validate caller identity mode before collectors register
Fail config load on an invalid prometheus_deployment_and_latency_caller_identity
value (including null) and on include_labels entries the selected mode removes
from a target metric, instead of booting green with an empty /metrics.
Validate the mode at the top of PrometheusLogger.__init__ so an invalid value
raises before any collector lands in the process-global registry, keeping
retries free of duplicated-timeseries errors. Label-validation errors now name
the mode setting alongside the rejected label.
---------
Co-authored-by: Mark Philipp <mphilipp622@gmail.com>
Co-authored-by: Yucheng Zhu <yucheng@berri.ai>
- test_custom_callback_input: audio redaction assertion expects None content
(redaction leaves None untouched, gpt-audio-1.5 returns content=None)
- local_testing conftest: drain GLOBAL_LOGGING_WORKER in isolate_litellm_state
teardown so mocked-router tests stop leaking pending logging tasks into
test_gcs_pub_sub
- test_together_ai: tools is always a supported param now; only response_format
is gated by function-calling support
- test_keys: /team/new omits models instead of sending null (422), so the key's
team really exists and auth no longer raises TeamNotFoundError
- test_team_delete_member_add_race: per-test unique team and user ids so xdist
workers sharing one Postgres stop deleting each other's team mid-race
PR #38114 dropped whichever header user_api_key_auth would read the caller's
key from, by precedence. Under custom_auth, JWT auth, or no master key that
header is the caller's own Google token, so the bring-your-own-credentials
Vertex branch answered 401 to every valid request.
A header value is now dropped only when it is the master key or when its
hash is the api_key that authenticated the request, so a Google token that
auth never consumed keeps flowing while a LiteLLM key still never reaches
Google.
test_passthrough_post_call_guardrails.py no longer plants a MagicMock
proxy_server module in sys.modules at import, which poisoned sibling tests
that read module globals at call time.
The auth handler stamps the requested model's provider onto BudgetExceededError before logging it, which made a key-over-budget 429 look provider-originated and regain its traceback (and an OTel stack_trace) after the provider 4xx carve-out. Any exception whose unified rate-limit category names litellm's own limiter is now a proxy rejection, matching the HTTPException rule.
ProxyRateLimitError derives from HTTPException but carries an llm_provider,
so it read as provider-originated and regained its traceback. Any
HTTPException is a proxy rejection regardless of llm_provider.
The /v1/messages route logs the provider's raw BaseLLMException, which carries
no llm_provider, so its 4xx still counted as an expected client error and lost
its traceback. Treat BaseLLMException as provider-originated as well.
is_expected_client_error treated every HTTP 4xx as a rejection the proxy
issued itself, so a 401 or 429 the provider returned lost its traceback in
the standard logging payload and the OTel error span dropped
litellm.provider.error.stack_trace. An exception carrying llm_provider is
an upstream or deployment problem and keeps its traceback; the proxy's own
pre-call rejections still skip it
Document the intentional internal seams used by the DCR bridge admission tests and normalize import ordering.\n\nGenerated with AI\n\nCo-Authored-By: Codex
Preserve standard Authorization key validation while preventing client MCP credentials from receiving anonymous bridge admission.
Generated with AI
Co-Authored-By: Codex
* feat(logging): add async_post_call_failure_deployment_hook
CustomLogger already has async_pre_call_deployment_hook and
async_post_call_success_deployment_hook, both firing once per real
deployment attempt from wrapper_async since the router re-enters that
wrapper fresh on every retry and fallback step. There was no failure-side
counterpart; the only failure signal, async_log_failure_event, fires once
per logical client request behind a dedup gate, so fallback chain attempts
2+ were invisible to callbacks needing per-deployment-attempt granularity.
Adds async_post_call_failure_deployment_hook(request_data, exception,
call_type) to CustomLogger and a matching dispatcher in utils.py, called
from wrapper_async's except block. It needs no dedup coordination since
each real attempt naturally re-enters the wrapper once. Unlike its two
siblings, the dispatcher wraps each callback call in its own try/except
since it runs on the wrapper's own exception path and a broken callback
must never mask the exception about to be re-raised to the caller.
* feat(logging): pass fallback_depth through to async_post_call_failure_deployment_hook
Router already tracks fallback_depth internally on each fallback hop
(litellm/router_utils/fallback_event_handlers.py), incrementing it once per
target tried, but nothing surfaced it to CustomLogger callbacks. Reads it
off request_data in the dispatcher and passes it through as a best-effort
int | None keyword: None on the first, pre-fallback attempt or a bare SDK
call with no router, 1 on the first fallback hop, 2 on the second, and so
on. Verified live against a real multi-hop Router fallback chain before
adding the regression tests.
* fix(logging): fire async_post_call_failure_deployment_hook on internal calls too
The failure hook was gated behind the same not _is_litellm_internal_call
check as the request-level dedup-gated failure logging, so a failed
internal sub-call (e.g. an emulated file-search step) never reached it,
even though its async_pre_call_deployment_hook and
async_post_call_success_deployment_hook siblings already fire
unconditionally for such calls.
* chore: retrigger CI (lint job hit a transient GitHub Actions infra outage on the prior push)
* chore: retrigger CI (lint job hit the same GitHub Actions infra outage again)
* fix(logging): scope async_post_call_failure_deployment_hook to the actual model call
The hook was dispatched from the wrapper's broad outer except, which also
catches BudgetExceededError (raised before any deployment attempt),
errors from async_pre_call_deployment_hook, and errors raised after a
successful model call (post_call_processing, async_post_call_success_deployment_hook,
caching). None of those are a deployment attempt failing, so the hook
misreported them as one.
Scoped the hook to a try/except around the model call itself, so it only
fires when that specific call raises, matching its own documented contract.
* test: assert the callback actually ran in the failure-hook error-isolation test
An upstream test-quality gate (TQ001) flagged this test for asserting
nothing, so it could only fail by raising. Track whether the exploding
callback actually ran and assert on it, so the test would catch a
dispatcher that silently skipped every callback instead of isolating a
raising one.
* fix(logging): harden async_post_call_failure_deployment_hook against 5 maintainer-verified issues
A maintainer's live-proxy A/B review against base found five real
problems with the failure hook, all reproduced and fixed:
- The dispatcher called overrides with fallback_depth as a required
keyword, so an override matching this PR's own earlier 3-arg
proof-of-fix example raised TypeError, swallowed at debug level, on
every call. Now checks the override's signature once per class and
omits the keyword when unsupported.
- A callback mutating the exception it receives (e.g. status_code)
changed what the real caller got back, since it was the same live
object about to be re-raised. Callbacks now receive a same-class
snapshot instead.
- request_data exposed attempted_targets, the router's own live
fallback-walk bookkeeping shared by reference across every hop, so a
callback calling .record() on it could make the router skip a
deployment it never actually tried. Now excluded from what the hook
receives.
- The hook's own await sat directly in the model-call except block, so
a caller-side cancellation landing mid-await (e.g. asyncio.wait_for)
replaced the real deployment exception with CancelledError/
TimeoutError. Now isolated so hook dispatch can never mask the real
failure.
- The timestamp used for the reported failure duration was captured
after the hook ran, so a slow callback inflated
async_log_failure_event's duration. Now captured before the hook
dispatches.
* fix(logging): preserve traceback/cause/context on the failure-hook exception snapshot
Bugbot found a real gap in the previous round's exception-mutation fix:
_snapshot_exception_for_hook only copied __dict__ and args, so a
callback formatting or inspecting the failure chain saw an empty
traceback and lost chained-exception context, even though the live
exception still has them. __traceback__/__cause__/__context__ aren't
stored in __dict__, so they need copying explicitly.
* fix(logging): preserve __suppress_context__ on the failure-hook exception snapshot
Setting __cause__ has a documented CPython side effect of implicitly
forcing __suppress_context__ to True. Since the previous round's
traceback fix set __cause__ before __suppress_context__, a normal
implicit-chaining exception (no `raise ... from`, __suppress_context__
naturally False) got its context wrongly suppressed on the snapshot.
Now __suppress_context__ is set explicitly, after __cause__, so it
always reflects the real exception.
* fix(logging): use MappingProxyType for the failure-hook's sanitized request_data
A LIT002 budget check (surfaced by rebasing onto a moved base) flagged
the dict comprehension building safe_request_data as mutable
construction. MappingProxyType is also a strictly better fit here: a
genuinely read-only view, not just an immutable-looking dict, matching
the intent that callbacks should never be able to mutate what they're
handed.
---------
Co-authored-by: Deepanshu <deepanshu.lulla@alpha-sense.com>
Adds streaming, async, /v1/responses, and /v1/messages coverage for the
Together AI overhaul (#38233, #38248, #38230, #38265, #38275), plus the
legacy api.together.xyz host and TOGETHER_AI_API_BASE through
litellm.completion. Each new test fails under a one-line mutation of the
merged code.
Type the GPT-5.x reasoning payload with a ReadOnly TypedDict so the dict
literal satisfies the type-discipline budget, and drop the now-redundant
thinking pop (the thinking mapping is already skipped for these models).
Update the cross-region capability test to expect reasoning_effort offered
and thinking/output_config withheld for GPT-5.x on Converse.
Keys minted by /key/generate get no LiteLLM_UserTable row, so /v2/model/info?user_models_only=true for such a user hit the new None guard and returned 400 where the merge base returned the user's own models. Skip the team-model merge for a missing row instead of raising, since a user with no row belongs to no team
The router's pre-content ping filter dropped AgenticAnthropicStreamingIterator's
hold-back keepalive, so a held-back turn sent the client nothing until the buffer
settled. A ping that no lifecycle frame precedes is now forwarded live, since a
fallback's message_start can still follow it without overlapping lifecycles
The proxy's cancel-refund guard checked isinstance against the iterator, but the
proxy only ever sees it behind FallbackAwareAnthropicMessagesStream and
AnthropicMessagesStreamingResponse, so a disconnect during hold-back refunded the
budget reservation anyway. Both wrappers now forward a duck-typed
has_buffered_provider_output flag, and the router wrapper follows a fallback
source so the flag tracks the stream actually being consumed
Stop advertising thinking/output_config as supported for OpenAI GPT-5.x and
skip the thinking mapping for these models, so a request combining thinking
with reasoning_effort can no longer leak a thinking block into
additionalModelRequestFields regardless of parameter order, which Bedrock
rejects with unknown_parameter.
prisma-client-py delete() returns None when the row is already gone, so the
handler reported deleted: true for a row this caller never removed. Surface
the same 404 the read path uses instead
OpenAI GPT-5.x models on Bedrock Converse expect reasoning effort under
additionalModelRequestFields as {"reasoning": {"effort": ...}}. They were
falling into the Anthropic branch and emitting a `thinking` block, which
Converse rejects with unknown_parameter.
The bedrock_converse gpt-5.6 entries were also missing supports_reasoning,
so reasoning_effort was dropped before mapping. Setting the flag lets the
existing config-driven supported-params path accept it, rather than adding
another model-name branch.
* perf(streaming): add shared JSONFragmentAccumulator for Vertex and Anthropic
Vertex's handle_accumulated_json_chunk and Anthropic's
_handle_accumulated_json_chunk each independently accumulated SSE fragments
into a JSON envelope with self.accumulated_json += fragment. Because the
attribute holds a live reference, CPython copies the whole prior buffer on
every fragment, making buffer assembly O(n^2) in total payload size.
Anthropic additionally had no completeness heuristic at all and called
json.loads on the whole buffer after every fragment, and could wedge forever
on two concatenated envelopes.
Add JSONFragmentAccumulator in litellm_core_utils/: fragments append to a
list in O(1), a could_close_json heuristic lets callers skip the join+parse
entirely until a value could plausibly be complete, and pop_next_value peels
one JSON value off the front of the buffer at a time using
json.JSONDecoder().raw_decode, keeping any unconsumed remainder instead of
failing on concatenated values. Migrate both providers onto it; Anthropic's
__next__/__anext__ end-of-stream handlers now delegate to
_handle_accumulated_json_chunk(is_final=True) instead of duplicating the
parse-and-reset logic inline.
Fixes#31861.
* test(streaming): close diff-coverage gaps in JSONFragmentAccumulator migration
Codecov flagged 11 uncovered lines in the migration: the accumulated_json
setter, __next__/__anext__'s end-of-stream drain branches in both providers,
and the pop_next_value "not found" path when a buffer's newest fragment ends
in "}" but is genuinely incomplete (an inner object closed, the outer one
didn't). Add targeted tests for each.
* fix(streaming): make JSONFragmentAccumulator's completeness heuristic O(1)
could_close_json rescanned every preceding blank fragment on each call, so a
hostile upstream that sends malformed JSON (never closing) followed by many
blank keepalive fragments could drive that scan, and the join+parse it
gates, to O(n^2) total. Track the last non-blank fragment's trailing byte
incrementally in append/pop_next_value/set instead of rescanning the buffer.
Reported by automated review on PR #36610.
* fix(streaming): make JSONFragmentAccumulator.pop_next_value O(1) per call
pop_next_value previously rebuilt the full remaining string and sliced a
new remainder on every call, so draining N concatenated JSON values
already sitting in one buffer cost O(n^2) total. Replace the rebuild-and-
slice with a materialize-once cursor: pending fragments are joined into
the buffer only when new ones have arrived since the last pop, and
consumed values are dropped by advancing an offset instead of copying
the remaining string.
* test(streaming): make JSONFragmentAccumulator drain regression test CI-stable
The 80k-value drain test used an absolute ms budget that flaked on a
busier CI runner (233.5ms vs a 150ms budget calibrated on a quiet
machine). Replace it with a doubling-ratio check: draining twice as many
concatenated values should take roughly 2x as long for O(n), not
~4x for O(n^2), and that ratio holds regardless of machine speed.
* test(streaming): suppress TQ002 on the append-laziness spy test
A test-quality gate (TQ002: don't assert only that a mock was called)
landed upstream since this branch's last rebase and now flags
test_append_never_calls_raw_decode. The test verifies append() defers
all decoding to pop_next_value, which has no caller-observable proxy
other than spying on the stdlib call it must avoid making.
* test(vertex): spy on raw_decode instead of json.loads in accumulator regression tests
Post-migration to the shared JSONFragmentAccumulator, Vertex's decode
path goes through json.JSONDecoder.raw_decode, not json.loads. The two
O(n^2)/partial-fragment regression tests still patched json.loads, which
that path never calls, so both passed unconditionally regardless of
whether the underlying implementation regressed. Verified by simulating
an eager-reparse regression: both tests now fail against it and pass
against the correct implementation.
* fix(streaming): widen JSONFragmentAccumulator's whitespace skip to match str.strip()
pop_next_value's whitespace skip only matched json.decoder.WHITESPACE's
ASCII set, narrower than str.strip() (Unicode-aware) which the O(1)-cursor
rewrite replaced. A non-ASCII separator like U+00A0 between two
concatenated JSON values on one SSE line made raw_decode fail on it, and
the buffer never advanced past that byte again, permanently stranding
everything after it for the rest of the stream. Use str.isspace() to
match str.strip()'s tolerance.
---------
Co-authored-by: Deepanshu <deepanshu.lulla@alpha-sense.com>
