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.
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
Four registry cells that had no e2e test. The cache_control suite gains a direct
Anthropic case (the same cache_control prefix the Bedrock and Vertex rows send)
and an OpenAI case, where caching is automatic so the prefix goes out as a plain
system string with a prompt_cache_key; both assert the second identical call
reports cache-read tokens. The shared second-call helper now takes the send
callable so the OpenAI shape fits without a second copy of the retry loop.
The embeddings suite gains a cohere/embed-v4.0 deployment that must return a
non-zero vector, and the passthrough suite gains an OpenAI-format chat through
the raw /openai/v1/chat/completions prefix that must relay a real completion and
log a costed pass_through_endpoint row whose token counts match the usage the
caller was served.
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.
/v1/batches now rejects a missing input_file_id with a 400 through
raise_if_required_body_param_missing, so the contract negative that was
skipped for "500s instead of 400" passes as written. Verified against a
live proxy.
The three Datadog MCP tests were skipped because each one sent a
`telemetry` argument that search_datadog_logs rejects with "unexpected
additional properties". That argument was never a documented Datadog
parameter and no assertion reads it, so it is dropped and the tests run
again unchanged otherwise.
The two vision tests pointed at a Wikipedia-hosted cat photo, so every run
depended on upload.wikimedia.org staying up and unthrottled. It throttled,
and the 429 surfaced as a bedrock APIConnectionError, which reads as a
gateway failure rather than what it was.
The image is now a fixture in the repo, passed as a data URL. That also puts
the two providers on the same bytes: litellm downloads the image itself for
bedrock, while openai is handed the link and fetches it from its own servers,
so the hosted URL quietly meant the two tests were not testing the same thing.
The image was generated for this repo rather than borrowed, so nothing here
carries a third-party license. Also drops a stale comment about openai prompt
caching that sat above the vision helper; no caching test uses it.
The /v1/messages validator checked a tool_use block's name and id but not its
input, so a block whose location came back empty or wrong still passed, while
the chat side rejected the same damage. That gap predates this branch; it is
worth closing here because the point of the change is that every parallel call
is checked rather than counted.
AnthropicContentBlock now declares input as a typed field. It already survived
on extra="allow", but reaching it from a test needs a real field to keep the
e2e basedpyright gate at zero. Serialization is unchanged: bodies are dumped
with exclude_none, so a block without an input still replays exactly as before.
The together backend is picked as the cheapest chat row that supports both
tools and reasoning, which currently resolves to together_ai/openai/gpt-oss-120b.
That row is marked supports_parallel_function_calling, so one weather prompt
can legitimately come back as several get_weather calls. Both tool tests
asserted exactly one call, so a parallel answer failed them even though the
gateway handled it correctly.
They now check every returned call instead of counting them: each one has to
be a get_weather naming Paris, with an id a tool result can answer. Dropping,
misnaming, or mangling a call is still red; only the count is the model's
business. The round trips answer every call rather than just the first, which
is also what the Anthropic Messages spec asks for.
`test_gemini_chat_returns_content_and_logs_cost` asks gemini-2.5-flash to
"reply with the single word pong" under `max_tokens=32`, and has been seen
returning no content at all:
completion_tokens=29, reasoning_tokens=29, content=None
gemini-2.5-flash defaults to dynamic thinking, and `max_tokens` maps to
`maxOutputTokens`, which on the 2.5 family counts thinking tokens as well as
visible output. So the model is free to spend the entire budget on thoughts and
emit nothing, which is exactly what the usage above shows.
Raising the limit alone does not fix this. Dynamic thinking on 2.5 Flash is
documented up to 24576 tokens, so no budget small enough to be reasonable for a
one-word smoke test is safe. The fix is to take thinking out of the picture:
`reasoning_effort="none"` maps to `thinkingConfig.thinkingBudget=0` for the 2.5
family, so the whole limit is available to visible output. Verified against this
checkout:
get_optional_params(model="gemini-2.5-flash", custom_llm_provider="gemini",
max_tokens=32)
-> {'max_output_tokens': 32} # no thinkingConfig at all
get_optional_params(model="gemini-2.5-flash", custom_llm_provider="gemini",
max_tokens=64, reasoning_effort="none")
-> {'max_output_tokens': 64,
'thinkingConfig': {'thinkingBudget': 0, 'includeThoughts': False}}
This mirrors what the OpenAI tool tests in this same file already do with
gpt-5.6 for the same failure mode. `max_tokens` goes to 64 for headroom; with
thinking disabled that is ample for a one-word answer.
Neither `covers` claim changes: the call still exercises the gemini chat
translation path and still produces a costed SpendLogs row.
`_cacheable_system_block` embedded the per-run marker in all 300 paragraphs, so
the block's token count moved with the marker's own tokenization. Measured over
40 random markers the size ranged 3611-5408 tokens (median 4509): 15% of runs
landed under the 4096-token minimum cacheable prefix of Haiku 4.5, despite the
docstring claiming the prompt was comfortably above it.
When the system block is under the minimum, no cache entry is written at the
system breakpoint. The entry at the second breakpoint still gets written,
because system + first user turn clears the minimum -- which is why the failures
report a large cache_creation with cache_read stuck at 0
(`cache_creation_input_tokens=5610 cache_read_input_tokens=0`, and 5610 is the
whole prefix, not the user turn's share). `_prime_prompt_cache` rotates the user
turn on every attempt, so that second entry never prefix-matches the next
attempt either. Every attempt re-creates the full prefix, cache_read never rises
above 0, and the loop burns its 60s deadline:
prompt cache never became readable in full within 60.0s
That is the single most frequent flake in the e2e suite, 9 of 38 runs, and it
hits all three provider classes identically because they share this helper.
Move the marker out of the repeated paragraph so it appears once, and size the
block at 1500 paragraphs. The prefix is now 8056-8060 tokens across markers --
spread 4 tokens instead of 1797, and 1.97x the minimum in the worst case. The
same marker-per-repetition pattern in `_first_turn_user_text` is fixed the same
way. Both copies of the helpers stay byte-identical.
Google withdrew gemini-live-2.5-flash-preview-native-audio-09-2025 from the
Vertex Live API. Every session dies at setup:
received 1007 (invalid frame payload data)
gemini-live-2.5-flash-preview-native-audio-09-2025 is not supported in the live api.
The client sees session.created (the proxy synthesizes it on connect) and then
nothing, so both vertex_ai realtime tests time out waiting for session.updated.
Confirmed by probing the Vertex Live endpoint directly with the e2e stack's own
credentials:
gemini-live-2.5-flash-preview-native-audio-09-2025 -> 1007, not supported
gemini-live-2.5-flash-native-audio -> setupComplete
so this swaps to the non-preview sibling, which is the same native-audio class
and is what the cost map already carries for vertex_ai.
Not a litellm regression. The suspicion fell on #38395 because it removed the
native-audio speechConfig strip, but the setup payload this suite sends is
byte-identical either side of that change: the strip only fires when a client
sends a voice, and the e2e SessionConfig has no voice field. Google's rejection
names the model, not a field.
The gemini (Google AI Studio) provider keeps the -09-2025 id, which still works
there; only the Vertex endpoint dropped it.
Adds live e2e coverage for the Bedrock combinations behind recent customer
incidents: llm_provider-* response-header forwarding on /chat/completions
(nonstream and stream), regional us.anthropic.* inference-profile ids over
the invoke route, and the Admin UI Test Connection probe for a
responses-mode Bedrock Mantle deployment. Registers the matching cells in
the coverage registry and publishes the provider x feature matrix table in
its README.
Adds a scheduled GitHub Actions lane on top of the merged record/replay
transport. A Saturday cron records the `replayable` e2e tests against the
real providers and publishes the fixture bundle as a private
`e2e-fixtures-bundle` artifact with a SHA-256 sidecar. Weekday crons pull
that artifact by its pinned digest, verify the checksum before extracting,
and replay it with provider credentials set to bogus values, so a run that
ever reached a real provider fails instead of passing.
An egress sentinel pins the provider hostnames to a local sink for the whole
replay job and counts every connection that reaches them; the job asserts
that count is zero, so hermeticity is proven by measurement. A red Saturday
publishes no bundle, so the next weekday finds nothing fresh and fails loudly
rather than replaying a week-old recording, and the transport's seven-day
freshness gate hard-fails any bundle that has drifted too far. The lane also
runs on demand from the Actions tab with a record/replay `mode` input.
Tests join the lane with `@pytest.mark.replayable`. The streaming Anthropic
test now counts to twenty so its recorded response banks several content
deltas, matching the assertion that the stream arrives incrementally.
The record/replay harness stored a streamed provider response as one
buffered body, so a replayed stream arrived coalesced and the
/v1/messages streaming test could not be edge-wired. Keep each SSE
transfer chunk in the bundle in the order the provider sent it (a new
streamed response shape at BUNDLE_FORMAT_VERSION 4) so replay reproduces
the provider's split points, the recorded usage chunk keeps its
position, and a mid-stream upstream error replays as the same
mid-stream error rather than a clean body.
Resolves LIT-5742
* 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.
The mid-conversation system tests prime the prompt cache by re-sending an
identical /v1/messages body until its usage shows the full prefix read back
three times in a row. The e2e stack runs with the litellm response cache on,
so every resend after the first is served from redis with the first call's
usage and the streak can never form; the three unflagged-model tests have
failed on every litellm-e2e build since the consecutive-read check landed.
Send cache: {"no-cache": true} on RichMessagesRequest, as test_cache_control
already does, so each resend reaches the provider.
The two fallback tests sent the same "say hi" / max_tokens=16 body to the
gpt-5.5 fallback, so one empty (finish_reason=length) completion served the
second test from the response cache and failed both. Give each test a unique
prompt and leave gpt-5.5 enough tokens to emit text.
Chat completions, embeddings, the non-streaming /v1/messages tests, and the
OpenAI batch deployment now register through the provider edge, so
E2E_FIXTURE_MODE=record captures their provider calls and replay serves them
back offline. None of them was wired before, so record was a silent no-op over
these suites and replay quietly went live instead of using the bundle
Multipart uploads now key on their parsed parts: every ordinary form field,
plus the field name, filename, content digest, and length of each file part.
The boundary is envelope rather than content, so it stays out of the digest
instead of changing the key on every run. A body that does not parse as its
declared envelope still has the boundary normalized away before hashing, so
the fallback is at least stable, and it records a name that says why
Binary uploads hash byte for byte. Canonicalizing them first meant decoding
with errors="replace", which collapsed every invalid byte to one U+FFFD and
gave two different PDFs of the same length the same key
Bundles stay out of the repo: they hold verbatim provider response bodies and
expire seven days after recording. Publishing them for CI is LIT-5748, and
streaming fidelity is LIT-5742
The websocket routes under /openai_passthrough and /openai had no e2e
coverage, so nothing catches the regression from issue #36088, where both
prefixes carried HTTP routes only and refused every upgrade with a 403
before a socket ever existed.
Two tests cover it. The realtime one opens /openai_passthrough/v1/realtime
and asserts OpenAI's own session.created frame comes back, which proves the
route is registered and relayed upstream. The responses one asserts
/openai/v1/responses accepts the upgrade, since a responses.connect socket
waits for the client to speak first and has no opening frame to check.
A refused upgrade is an HTTP response rather than a close frame, so both
assert on the handshake. ws_base_url moves into e2e_config now that a
second suite needs it
The passthrough tests and their coverage registry rows pointed at the internal
ticket id, which does not resolve for anyone following a link from
status.litellm.ai. Each test docstring and registry rationale now names the
GitHub issue it pins: #36086 for the two prefix routing cases, #36087 for the
file list cursors, #36523 for streamed Responses cost, and #36646 for
embeddings spend.
Five e2e tests over routes a customer drives through the gateway, each one
pinning a fix that currently has no live coverage.
The dedicated /openai_passthrough prefix used to be swallowed by the
provider-scoped /{provider}/v1/files and /{provider}/v1/batches routes, which
bound "openai_passthrough" as a provider name and failed inside the gateway
before ever reaching OpenAI. Two tests now upload a file and list batches
through that prefix and assert OpenAI's own objects come back.
Streamed /openai_passthrough/v1/responses and /openai_passthrough/v1/embeddings
are relayed to OpenAI but still have to be costed, since the customer budgets
against this traffic. Both used to land a row the gateway could not use: the
streamed responses call logged a zero-cost row under a random id, and
embeddings wrote no row at all. Each test now reconciles the logged spend and
token counts against the response the caller was actually served.
GET /v1/files narrowed its data to the caller's own rows but left first_id and
last_id addressing the shared provider account's page, handing any caller raw
provider file ids belonging to other tenants. The new test asserts both cursors
address rows in the page the caller can see.
ResourceManager.defer now accepts any callable rather than one returning None,
so a delete that answers with a response model can be deferred as-is.
On models without supports_mid_conversation_system, a system entry between
an assistant tool_use turn and the user tool_result turn became a user turn
in that position and the provider rejected the request ("tool_use ids were
found without tool_result blocks immediately after"). That run of entries
now goes right after the tool_result turn, where consecutive user turns
merge upstream. The converted turn also carries only role and content, as
the hoist did, so an entry with extra keys no longer 400s with "Extra
inputs are not permitted".
The e2e cache priming re-sends the identical first turn until its own cache
entry reads back before the reminder turn goes out, since Vertex can take a
few seconds to serve a freshly written entry.
/audio/transcriptions answers a model-less request with one of two 400s
depending on whether any wildcard deployment is registered at the time, and
every suite shares one proxy, so run order decided which message came back.
The assertion pinned only the no-wildcard wording, so it went red whenever
the model-access-group suite had registered its wildcards first. It now
accepts either message and still holds the error to naming the model
Verified against a live proxy in both states: with a wildcard registered
(the message CI was seeing) and with none (the message the assertion
expected), the suite passes 3/3 either way
This cell needs the websearch_interception callback and a declared search
backend, both listed in its own module docstring. The ephemeral e2e stack
ships neither, so the request falls through to the bedrock transformation
and takes the by-design 400 that tells you to enable interception.
The cell has never been green here: the error path merged about an hour and
a half before the cell did, and the last full suite to pass predates the
cell entirely. Skip it with the reason recorded so the run reports honestly
instead of carrying a permanent red, and unskip once the stack ships the
config the docstring already spells out.
Both providers reworded the error strings these two cells pinned, so the
suite went red without any behavior changing. Anthropic's auth error is now
"API key is invalid." rather than "invalid x-api-key", and OpenAI rejects an
empty upload with "This model does not support the format you provided.",
which names neither "file" nor "audio".
Assert the durable shape instead. The otel cell pins the machine-readable
authentication_error type plus a non-empty message, and the embedded JSON
still has to parse, which is what proves the attribute survived untruncated.
The transcription cell pins that the 400 relays the provider's own rejection
and is typed as a client input error, so a regression that swallows the
provider reason or returns a 500 still fails.
