* test(responses): pin bridged commentary plus final answer as one chat choice
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): pin reasoning, two-message and two-tool-call bridge shapes
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): number bridged chat tool calls 0, 1 instead of all 0
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): fold bridged Responses output into one chat choice
A Chat Completions request bridged to /responses got one choice per
Responses message item, so a reply shaped [message, message,
function_call] put the tool call on choices[1] and clients reading
only choices[0] never saw it. Fold every message text, annotation,
reasoning item and tool call into choices[0]
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): drop mutable annotations and the extra cast in the choice fold
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): remove a cast-ok marker that suppressed nothing
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): join folded message texts with a newline
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* revert(responses): join folded message texts with no separator so stream and non-stream match
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(sagemaker): inference components and OpenAI-compatible embeddings
Send model_id on a SageMaker embedding deployment as the invoke call's
InferenceComponentName, sync and async, and keep it out of the request body.
Add the sagemaker/openai/<endpoint> route, which sends the OpenAI embeddings
request shape ({"input": [...]} plus extra_body fields) and reads
data[].embedding and usage from the response for containers that serve the
OpenAI embeddings API. Plain sagemaker/<endpoint> deployments keep the
Hugging Face shape
* refactor(sagemaker): type validate_environment messages and cover base64 embeddings
* fix(sagemaker): name model_id in inference component errors and reject a non-string model_id on embeddings
* fix(sagemaker): fall back to the error text when SageMaker sends a null message
A ClientError whose Error.Message is null (SageMaker's InternalFailure shape) reached the
inference component hint as None and crashed with a TypeError; the message now falls back
to the error's own text on the chat and embedding paths
* test(sagemaker): audit matrix for inference-component embeddings
* fix(sagemaker): read the error status and text through one shape-checked helper
botocore's ConnectionClosedError carries response=None, so the except blocks that
read e.response.get(...) crashed with an AttributeError and the proxy answered 400
"Invalid request format" for a dropped connection. client_error_details validates
the botocore error shape with pydantic and falls back to 500 plus the error's own
text for any other shape (a missing response, an httpx Response, a plain exception).
* fix(sagemaker): drop the inference component hint from chat and derive the SDK wire expectation from the request
The chat except blocks never see SageMaker's text: HTTPHandler.post raises from
raise_for_status before the handler's own SagemakerError line, so the hint check ran
on httpx's generic "Client error '400 Bad Request' for url" line and could not match on
any chat path. The hint now names litellm.embedding only and the chat blocks keep the
status and text they had, without the dead check. The two OpenAI SDK wire tests pinned
the SDK's default encoding_format; they now expect whatever the SDK sent.
* refactor(sagemaker): meet main's per-rule lint and type ceilings
Main started gating every lint rule on its merge-base count and added LIT015
(frozen pydantic models) after this branch was cut, so the merge with main
pushed LIT001, LIT011, LIT015, BLE001 and reportArgumentType past their
ceilings. The response models are frozen with tuple fields, the OpenAI
embedding response is built with model_copy instead of writing into the
caller's object, the invoke_endpoint arguments are a TypedDict, and
optional_params and extra_body are validated into Mapping[str, object] before
they are read
* refactor(sagemaker): pick the embedding config in one place and stop pinning upstream error text
The openai/, Voyage and Cohere checks lived in both get_model_config and
get_supported_openai_params, so they now share one lookup. Tests compare error
messages against the error's own text, or against botocore's and json's own
wording, instead of copying that wording, and assert whole messages where the
text is ours
* fix(caching): key direct calls on model_id so inference components don't share cached embeddings
* test(sagemaker): name the fake-boto3 SDK test class for what it covers
* test(sagemaker): kill and stop chaos workers through the owned-proxy harness
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
A caller hitting /v1/messages with a cache_control mark and an OpenAI
gpt-5.6 or newer deployment behind it lost the mark when the request was
rebuilt as a Responses request, so OpenAI never cached at the boundary the
caller chose. The bridge now rewrites each marked system, user and
tool_result block to prompt_cache_breakpoint on the same block when the
deployment reads breakpoints, keeps the system prompt as a developer
message so its mark survives, and still strips the mark for deployments
that do not read it
The hook's gate becomes public so the bridge and the hook share one
answer for which deployments read breakpoints, and the one-block swap
lives next to the existing breakpoint helper in prompt_templates
* feat(providers): add CoralBricks as a native provider
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): harden cost map consistency tests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): type cost helpers, read cost maps from disk, cover streaming spend
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): drop Final on loop-rebound names
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): add the provider audit cells, the Add Model browser cell and the README row
* test(coralbricks): derive UI spend from model info, type browser rows, always release held upstream
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): release held upstream before the wire server joins its threads
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): number forwarded chats with itertools.count instead of a mutable list
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(coralbricks): drop retired glm-5.3-flash-fp4 and test against deepseek-v4.1-flash-fast-fp4
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(coralbricks): pin retries, await the respawned worker, read the UI rate from the catalog
The retry cells now set num_retries on the deployment, since the CI rig posts
router_settings.num_retries 0 after readiness and the router default of 2 no
longer applies there. The SIGKILL chaos cell waits for the respawned worker's
startup before the owned proxy is stopped and derives its timeout from the
harness budget. The Add Model spec waits for the catalog the form loads, reads
the expected rate from it, and finds the deployment through /v1/model/info,
which does not paginate
* fix(coralbricks): price the current model names and the new DeepSeek input rate
CoralBricks renamed glm-5.3-fp4 to glm-5.3-fast and deepseek-v4.1-flash-fast-fp4
to deepseek-v4.1-flash-fast (the fp4 names stay as deprecated aliases), and cut
DeepSeek V4.1 Flash input from $0.30 to $0.01 per 1M on 2026-10-08. Rates read
from https://www.coralbricks.ai/api/public/models on 2026-10-10; a live cache
write billed 6014 tokens at 9e-08 each, so the cache-write rate stays input
plus cacheWritePerM.
The test config builders now build new dicts instead of editing the parsed
YAML, and the cost-map test locals carry Final.
* fix(coralbricks): keep pricing for the deprecated fp4 slugs CoralBricks still serves
* fix(coralbricks): list the catalog rows for wildcard deployments and derive the test's input limit from the row
* fix(coralbricks): price the deprecated glm-5.3 and deepseek-v4.1-flash slugs
---------
Co-authored-by: Mubashir Osmani <mubashir@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(dd_span_tagger): emit litellm_user_email span tag for JWT-authenticated requests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(dd_span_tagger): use dotted litellm.user_email tag for consistency
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(model_hub): surface model_info.description in model group info and Model Hub UI
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(router): aggregate model group description without in-place mutation
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(health): skip background health check DB writes when the latest-row read fails
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock_mantle): map the OpenAI-shaped context overflow envelope too
Mantle reports context overflow in two envelopes. #37862 covered the
structured `validation_error` one carrying token counts
("prompt tokens (N) exceed model maximum (M)"). Mantle also returns the
OpenAI-shaped body:
{"error":{"code":"context_length_exceeded","message":"Your input
exceeds the context window of this model. ...","param":"input",
"type":"invalid_request_error"}}
That passes the existing `invalid_request_error` gate but matches neither
the token-count pattern nor `is_error_str_context_window_exceeded`'s
substring list, so it fell through to a generic BadRequestError. Clients
such as Claude Code only reactive-compact on a recognised overflow
phrase, so an overflowing request surfaced as an ordinary error with no
recovery.
Return the same normalised "prompt is too long: ..." message for this
envelope when it carries no counts, keeping the numeric wording when the
provider supplies them.
* fix(bedrock_mantle): map the streamed context overflow and keep its 400 on /v1/messages streams
* fix(anthropic_adapter): keep only the context overflow's 400 on streamed /v1/messages so other provider 4xx still fall back
* test(anthropic_adapter): wrap a parametrize line over the 120 char limit
* test(integration): audit Bedrock Mantle context overflow mapping across routes, clients, and fallbacks
* test(integration): move the streamed /v1/messages overflow spend row into a BUG-skipped test
* test(integration): keep every chaos cell in the burst and skip only the missing spend row
* fix(bedrock_mantle): answer 400 on an in-stream context overflow
The Mantle overflow mapper built its ContextWindowExceededError on the
original exception's response, so an overflow that Mantle reports inside
a 200 stream (the chat-mode token-count envelope, which the OpenAI chunk
parser wraps as a 500) kept that 500 and the streamed /v1/messages route
answered HTTP 500 api_error instead of 400 invalid_request_error. The
mapper now builds the error on a 400 response that keeps the original
headers and request, and the mapping test covers an in-stream 500 and 503.
The integration suites pin the streamed /v1/messages overflow to the
pre-content 400 route under a retry budget and to the message_start plus
invalid_request_error event without one, drop the chaos carve-out that
expected no spend row, and give the fallbacks file a timeout mark sized to
the proxy's graceful stop.
* fix(bedrock_mantle): keep the provider's 400 body and match the overflow code only as error.code
---------
Co-authored-by: milan <milan@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: ryan-crabbe-berri <ryan@berri.ai>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(decisions): support Clef-omni audio and video
* fix(decisions): accept Clef-omni media types with parameters and cover the proxy routes
The clip patterns now mirror Cloudflare's input schema, so a data URL or
content_type carrying parameters like audio/webm;codecs=opus is accepted.
The four proxy routes get media tests of their own, the backup cost map
is synced with the new Clef-omni rows, and the docs page moves to the
docs repo as a follow-up.
* fix(decisions): refuse parameterized media-type data URLs and refresh the OpenAPI snapshot
Cloudflare answers 422 to a data URL whose media type carries parameters even though its schema only requires the data:audio/ prefix, so LiteLLM now refuses those with a 400 before the call. The object form keeps passing parameters through, which Cloudflare accepts. The lazy OpenAPI snapshot and the dashboard API types are regenerated for the new audio and videos fields
* fix(decisions): gate media by cost-map flags, treat null clips as none, drop unsupported clips before validation
* fix(decisions): keep the SDK's positional order and gate media flags on evaluation-mode models
* fix(decisions): gate media on both decisions model modes and move the audit cells onto the harness helpers
* test(decisions): cover the Fireworks System One router in the media refusal cells
---------
Co-authored-by: moyai-devin-berriai[bot] <336287033+moyai-devin-berriai[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(gemini): do not inject temperature for Gemini 3+ models when omitted
map_openai_params added temperature=1.0 to every Gemini 3+ request that did not set one, on both the gemini and vertex_ai routes. Google is deprecating temperature, top_p and top_k for Gemini 3+ and says upcoming models will reject them, so this default turns a request that never set temperature into one that will fail
An omitted temperature now falls back to Google's own default, which their docs give as 1.0 for all Gemini 3 models. Explicit temperature, top_p and top_k are still forwarded
Fixes#38663
* test(gemini): annotate sampling regression tests
* test(gemini): drop the injected temperature from Gemini translation cases
The nine Gemini base cases for chat completions, responses and messages send no temperature, so the provider request no longer carries the removed 1.0 default. Each new body was sent to the real Gemini API for gemini-3.5-flash, gemini-3.8-flash and gemini-3.1-pro-preview and returned 200, and the stored provider responses are unchanged
* test(gemini): drop the default-temperature wording from the thinking level test docstring
* test(gemini): cover omitted and explicit sampling on both Gemini routes
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(fireworks_ai): add Decisions support for accounts/fireworks/routers/rune on /v1/systemone and /v1/decisions
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(fireworks_ai): point the rune cost row at the public Fireworks docs
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(decisions): patch litellm.callbacks with monkeypatch in the Fireworks cached-cost test
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): give chaos and timing tests a harness API and gate raw signals
Chaos tests killed and suspended proxy workers with raw os.kill and psutil
calls and bounded elapsed time with fixed literals, so they raced the uvicorn
worker healthcheck and failed on slow CI boxes
OwnedProxy now offers kill_worker and replacing_worker, which return only once
a replacement answers readiness with its own pid, and suspended, which is
refused unless the proxy was launched with a lenient worker healthcheck.
timing.py bounds durations against a measured control or the injected delay.
run.py measures the proxy import time once and derives readiness deadlines and
the per-test timeout from it, writing boot-calibration.json
check_test_quality.py adds TQ010 (raw process signals) and TQ011 (elapsed time
asserted under a numeric literal) for tests/integration outside _support, and
every call site is migrated so both report zero with no suppressions
* test(integration): steer the assistants wire burst onto both workers
The mixed-status burst assumed the OS would spread 48 connections across both workers and failed when one worker took them all. Each half is now sent while the other worker is suspended, and the test asserts an exact 24 and 24 split
* test(integration): restore product-enforced timing bounds lost in the migration
Some TQ011 migrations dropped or loosened a bound whose time is the behaviour
under test. The websocket session cap only kept its floor, so a cap closing at
300 s would have passed, and several timeouts were bounded by the injected stall
instead of the configured product value
assert_enforced_within bounds an elapsed time by the configured value plus
explicit slack, with an optional floor. The websocket cap, the MCP validation
timeout, the guardrail litellm_params timeout, the prisma command timeouts, the
baseline projection lock timeout and the key metadata statement timeout use it.
The device code refusal keeps its 15 s margin, and the langfuse retried prompt
fetch is now bounded by a cold fetch of a steady prompt as well as by the
ignored retry-after
* fix(tests): wait for the supervisor to reap killed workers and fix real-kill migrations
The lenient worker healthcheck was the whole case budget, so uvicorn's supervisor sat
in a ping on the killed worker for up to that long before respawning it. It now uses
the graceful stop budget, and a worker kill returns only after the victim is reaped,
so pooled connections no longer race the dead worker's socket.
Several tests on main picked the multiprocessing resource tracker as their victim and
never killed a worker. Now that they kill a real one, presidio counts streams cut by
the death, delete_model waits for the config reload that removes a deleted model from
a replacement worker, and the wedged coordination Redis test kills before the wedge
and waits for every parked publish to fail before resuming Redis.
The redis cluster test imports the harness through the tests package so it collects
in the unit job, and the per-model fallback rig exposes the owned proxy it now needs.
* test(observability): only require S3 deliveries the killed worker could not have buffered
A real SIGKILL drops the victim's unflushed batch, so requests it already answered can't be expected to land. The test now requires every request served after the replacement to land, some of the surviving worker's burst share to land, and nothing unsent to land
---------
Co-authored-by: Yuneng Jiang <yuneng@berri.ai>
GPT reasoning-series models (gpt-5.x, gpt-6) on Bedrock Mantle chat completions now send max_completion_tokens, carrying a deployment or client max_tokens over with its value. Every other Mantle chat model keeps sending max_tokens.
* test: migrate legacy unit dirs and unrun folders into tests/unit, tests/integration and tests/e2e
* test: fix disk cache collection, ungate saturation check, drop RSS memory tests
* test: assert one delivered log per completion in Datadog and LangSmith checks
* test: drop SDK-direct e2e copies and align cost-map path filter cases
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test(integration): wait for customer spend rows and use owned OpenRouter cost
* ci: isolate integration workers and clean up resources after crashes
* docs: describe isolated extensions workers and shard census
* fix(ci): restore integration service and fixture guarantees
* fix(integration): preserve replay IDs and native route fixtures
* fix(tests): scope integration harness imports to each unit case
* fix(ci): preserve fixtures and wait for owned workers
* fix(ci): isolate nested runs and telemetry cases
* fix(ci): make worker readiness and fault tests deterministic
* chore(ci): drop MCP fixture repair already covered by main
* chore(ci): keep the local catalog fix from main
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test: move guardrails, image, audio, batches, search legacy tests to unit
* test: restore the custom guardrail and remaining chunk 1 unit rows
* test: move the ocr matrix to e2e
* test: drop comments from moved tests and fix vertex ocr credentials
* test: move litellm utils and unified google legacy tests to their tiers
* test: move logging callback and pass-through unit legacy tests to their tiers
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: move store_model_in_db legacy tests and start the admin ui migration
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: finish the proxy job rows and drop their CircleCI jobs
Moves the remaining multi-instance, openai endpoint, otel, pass-through and
top-level proxy tests into tests/integration with scripted upstreams, rewrites
the weak replacements to assert the outbound wire and exact responses, and
removes the build_and_test job group this PR owns from CircleCI
* test: tighten the key generation and error log assertions, rename migrated unit files to their mapped names
* test: shard the pagerduty unit tests and record the rust harness file on the coverage allowlist
* test: drive the SSO sign-in tests through a scripted IdP and isolate the cost center stand-in
* test: type the e2e beta header, prompt cache and secret manager values
* test: assert the missing model error message on the OpenAI routes
* test: drop the deleted otel base class import and restore mcp redaction and exact langfuse usage checks
* test: make the langfuse fixture directory a package for the layout check
* test: restore collected no-match cases, real langfuse and gcs payload checks, and gate the aws secret lane
* test: keep the ocr e2e pdf fixture next to the test
* test: put meta below covers on the streaming tool search e2e
* test: run the OCR matrix against a scripted upstream and fix two runner-only e2e failures
* test: count only OCR calls in the OCR matrix recorder
* test: pin OCR parse bodies and billed cost, group new e2e cases into classes
* ci: restore the python image anchor dropped in the main sync
---------
Co-authored-by: yuneng <yuneng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(databricks): call the ai_decide AI Function route for decisions and the auto-router
Databricks serves ai_decide at POST <host>/api/2.0/ai-functions/ai-decide, not
on a serving endpoint's /invocations route. The decisions provider and the
auto-router classifier now accept only the ai_decide model, post to that route
without a model field, unwrap the response envelope, and map a noul answer's
probability to the System One noul field.
* test(databricks): bound the megabyte model name test by scaling, not wall time
* test(databricks): assert the whole OpenAI-format decision and floor the liveliness bound
* test(decisions): move the Databricks OpenAI-format wire row to the ai_decide route
* fix(databricks): keep serving endpoint decisions beside the ai_decide route
* test(databricks): assert the whole classifier request body in the test routing case
* test(databricks): correct System One spend call type
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: mateo <mateo@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: migrate local_testing SDK and provider tests out of build_and_test
* test: cover sagemaker region, hf template and nova paths end to end
Region rows now go through litellm.completion on the sagemaker/ handler that
the legacy tests exercised, the prompt template test mocks the Hugging Face
tokenizer config instead of switching to custom_prompt_dict, streaming tests
assert role, finish reason and body, and Nova single turn and streaming run
through litellm.completion with a recorded request body
* test: tighten migrated integration tests for local OpenAI routes, timeouts and air-gapped boot
Split the catch-all providers migration file by behavior. The fake endpoint
route tests now hit the owned upstream directly, as the legacy harness tests
did. LM Studio asserts the outbound key, path and stripped model. The embedding
timeout test asserts the proxy maps a deployment timeout to a 408 and forwards
the read timeout to the provider client, instead of timing out the test client
The Azure streaming tool-call test also asserts the second identical stream is
served from the Redis cache with the same reassembled tool call
The Docker network test always skipped in CI because the security job has no
litellm/litellm image and no egress to pull one, and it tested a published image
rather than the checkout. It now boots the checkout proxy with every outbound
HTTP connection refused and asserts readiness, the bundled cost map fallback and
that the cost map host is the only egress attempted
* test: tighten the migrated e2e rows for local_testing SDK and provider tests
Drop the duplicated Anthropic covers marker and the Anthropic cell wrongly claimed by
the Bedrock and Vertex cache tests. Keep the file-wide Anthropic model as it is on main.
Send the 2023-06-01 version header to Anthropic through the deployment's extra_headers
so the header actually reaches the provider, and mark the creation-token case 1h then
5m like the legacy test did
Reuse the Bedrock file's own registration for the pydantic response test, swap the
hand-built PNGs for the shared cat.jpg fixture, assert the SageMaker Nova image answer
names the cat and that top_logprobs=3 is honored, and register the SageMaker Nova opt-in
markers the same way as the other opt-in markers
* test: restore lost assertions in migrated test_main tests
Restore the role alternation check in test_completion_anthropic_hanging, the
embedding and streaming response-header checks, the logit_bias drop check and
the gemini system and safety settings checks. Add an azure 429 retry-after
test for test_exception_with_headers, trigger the router fallback with
mock_timeout like the legacy test, and stop the mock timeout tests sleeping 3s
* test: restore lost legacy assertions in migrated vertex ai unit tests
* test: tighten migrated cost, streaming and embedding tests and scope the hosted mock guard
* test: strengthen migrated anthropic, azure, bedrock and cohere provider tests
Assert full request bodies and usage instead of presence checks, restore the
dummy-tool and custom_llm_provider assertions the legacy tests made, add the
async and per-model variants the legacy parametrization covered, split
streamed tool arguments across deltas, and move the Azure chat streaming
tests to the azure.py mirror file
* test: tighten migrated llms provider tests and cover legacy intents lost in the move
Strengthen the migrated fireworks, gemini, mistral, together, openai text completion, openrouter, perplexity and openai error tests so they assert what the legacy tests proved (cost, provider tag, usage-only-on-last-chunk, function_call rebuild, reasoning param, sync/async/stream error fields). Replace the huggingface and jina tests that doubled litellm code with respx end-to-end tests. Add the tests the pre-existing mapped tests never covered: cohere 400 mapping, token-array text completion, hosted_vllm/llamafile/lm_studio env api base embeddings, watsonx embedding auth, ollama ssl_verify client and ollama vision prompt
* test: cover residual legacy paths for anthropic tool streams, watsonx async stream, ollama http images and caller async embedding clients
* test: restore single request assertions in the deep_infra and azure_ai embedding tests
* test: drop a no-op streaming cost flag from the openai streaming cost test
* test: cover residual legacy-test lines in caching, main, utils, replicate, vertex and bedrock invoke
* test: cover streamed audio assembly, choice membership, untranslated responses events and the stream log guard
* test: pin cost rates in unit tests, use sdk for cohere e2e, drop vertex llama e2e
* test: send the vertex maps e2e request through the openai sdk
* test: drop the local_testing jobs that select no tests after the migrations
* test: wait for the streamed cache entry and keep custom Mistral rates test-local
* test: keep untouched Mistral test lines as they were
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* feat(guardrails): add decision_model guardrail for decisions-API models
Adds a new guardrail provider that scores request and response text with a
decisions-API model (e.g. Jev) against named predicate checks. Each check
carries its own threshold and action (block or log); a flagged block check
rejects the request with a 400 naming the flagged checks, and call failures
honor unreachable_fallback (fail_open or fail_closed).
Ships four preset checks for the Admin UI (prompt injection, jailbreak,
system prompt extraction, data exfiltration) served from
/guardrails/ui/add_guardrail_settings, plus a Decision Model step in the
add-guardrail form filtered to evaluation-mode models.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(guardrails): fail closed on verdict errors and drop comments in decision_model
* feat(guardrails): cap decision_model input with head and tail elision
* fix(guardrails): list only block checks in decision_model 400s and tighten inputs
* feat(guardrails): honor timeout, cap checks at MAX_DECISION_QUESTIONS, and reject duplicate check names
* fix(guardrails): decorate decision_model apply_guardrail with log_guardrail_information
* fix(guardrails): suppress untyped-decorator pyright error on log_guardrail_information import
* feat(guardrails): screen each text in overlapping chunks with a shared decisions-call semaphore
Replace head/tail elision with per-text screening: each non-empty text is
sent to the decision model on its own, split into overlapping chunks when
it exceeds max_input_chars, and all calls run under a per-event-loop
semaphore shared across requests (max_concurrent_decision_calls,
default 8). Check verdicts aggregate the max probability across calls; a
missing predicate answer on a Block check blocks under fail_closed and
only warns under fail_open, while log-only checks are never flagged for
it. The 400 detail now lists flagged check names only. The provider
combobox also matches Decision Model on jev, typesafe, prompt injection,
jailbreak, and decision
* fix(ui): align decision model custom check rows with preset rows
* fix(ui): move decision model custom check remove button to the row end
* style(ui): run prettier on decision model check files
* fix(ui): inline decision model evaluation-mode filter after lens module removal
* fix(guardrails): keep decision model detections when a chunk call fails
* fix(ui): require an enabled decision model check before create
* fix(guardrails): screen decision model texts sequentially like presidio
* feat(guardrails): replace decision model presets with provider picker and per-question test
* fix(ui): render decision model question test as a full-width row
* feat(ui): add run-history test section to decision model form
* fix(ui): prevent threshold overlap and show test shortcut
* feat(guardrails): default decision model threshold to 0.7
* test(ui): expect decision model default threshold 0.7 in characterization payload
* feat(guardrails): screen tool call arguments with the decision model
* refactor(guardrails): freeze the decision model config model
* feat(guardrails): list only evaluation-mode deployments in the decision model picker
* feat(guardrails): hide the decision model questions table until a question exists
* style(ui): format DecisionModelFields
* style(ui): format decisionModelQuestion tests
* refactor(guardrails): type the tool call function extraction
* refactor(guardrails): build tool call texts without a mutable accumulator
* refactor(ui): share isDecisionMode for the decision model picker
* fix(guardrails): screen prior Responses function call arguments in the decision model guardrail
The Responses input handler leaves function_call arguments out of texts and tool_calls, so a prompt injection in a prior call's arguments passed unscreened. Assistant tool calls from structured messages are now screened too, deduplicated against the request's tool calls
* test(integration): send instructions with the decision model guardrail check
Checks require instructions, so creating the guardrail returned 422 and the test never sent a chat request
* fix(ui): score the decision model test preview the way the guardrail does
An unanswered block question now shows Block, disabled questions no longer count, and a run is hidden once the model or a question's text changes since it was scored
* fix(ui): clear the decision model error on pick and list questions under Add a question
Picking a model after a failed Create left "Select a decision model" on screen, because the form only validates on submit. The field now revalidates when it already shows an error. The Questions table and its heading appear only once a question exists, below the Add a question box
* test(ui): move the decision model form tests into an integration test file
They render the whole AddGuardrailForm, so they belong in a .integration.test.tsx file. Plain fields are now filled with fireEvent.change, and the two existing form test files go back to main apart from the modelHubCall mock the form now needs
* test(guardrails): build the decision model test params without mutating a local
* fix(guardrails): reject list checks outside decision_model and honor streaming scan settings
- LitellmParams now rejects a list `checks` for any guardrail other than
decision_model, so a Bedrock guardrail with list checks gets 422 again
instead of being stored and failing every request
- The decision model guardrail reads streaming_end_of_stream_only and
streaming_sampling_rate, so a streamed post_call response can be scored
once at end of stream instead of every 5 chunks
- max_concurrent_decision_calls is described as a per-worker limit
* fix(ui): match guardrail providers like the default picker filter
- The provider search uses the Base UI collator filter on the label and
the aliases, so searches such as presidio_pii or llm_as_a_judge match
again, as they did before the alias filter
- Model groups from /model_group/info are parsed defensively and after the
providers are populated, so a malformed response cannot empty the
provider list
* fix(ui): edit decision model questions in place and show passing test chips in green
Questions start as an Add question button. Each click adds a blank question
card with its name, text, action and threshold, editable until the guardrail
is created, mirroring the per-model budget editor. Create and Next reject a
blank or duplicate enabled question. Test run verdicts and chips use the
success tone for Pass, the error tone for Block and the warning tone for a
flagged Log only question
* fix(ui): drop the per-question enable checkbox from the decision model form
A question is either in the list or removed with its X, so every listed
question is sent and scored. The card now labels Name and Question the way
it labels Action and Threshold
* fix(ui): give the question name field the full card width
* fix(ui): show the provider logo on the selected decision provider
* fix(ui): show the full threshold slider track on decision model questions
* feat(ui): add decision model cards to the guardrail garden
Adds a Decision Model section to the guardrail garden with TypeSafe Jev,
Perplexity, Microsoft, OpenAI and Databricks cards. Each card opens the
create form scoped to that provider's decision models.
The create form drops the Decision Provider picker and lists every
decision model on the proxy with its provider logo, prompts to add a
model when none is available, and says that test runs are billed and
logged.
* fix(ui): shorten the decision model test hint to one line
* fix(ui): prefill the decision model from the garden card's provider
* chore(ui): regenerate the dashboard API types after merging main
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: yucheng-berri <293926326+yucheng-berri@users.noreply.github.com>
* fix(sagemaker): keep AWS credentials out of the sagemaker_chat request body
sagemaker_chat and sagemaker_nova built the container body from every optional
param, so aws_access_key_id, aws_secret_access_key, aws_session_token and the
other aws_* auth params, plus model_id and sagemaker_base_url, were sent in the
JSON body to the endpoint container. The request is still SigV4-signed with the
same credentials, which stay in optional_params for signing
* fix(sagemaker): limit the sagemaker_chat body filter to aws_ params
model_id and sagemaker_base_url are not credentials, so they stay in the
body as before and the change sticks to what the fix is about
* test(sagemaker): assert the whole sagemaker_chat request body
* fix(sagemaker): filter the sagemaker_chat body with the shared AWS auth param list
* test(sagemaker): cover the sagemaker_chat body on every endpoint, credential source, and outage
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(vertex_ai): parse Vertex timestamps as UTC so created_at is not shifted by the host timezone
convert_vertex_datetime_to_openai_datetime parsed the trailing-Z Vertex and GCS
timestamps into a naive datetime, so .timestamp() read them as host local time
and every created_at it produced was off by the host UTC offset (about 7h ahead
on a PDT host). Attach UTC before converting, and accept the whole-second and
nanosecond forms Google's RFC 3339 timestamps also use, which used to raise.
Refs LIT-9383
* test(vertex_ai): type the UTC epoch test, guard tzset, build the rig config without mutation
* feat(sdk): add litellm.systemone and trim litellm.decisions to the OpenAI shape
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* ci: run tests/unit/systemone in the core-utils unit shard
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(sdk): keep litellm.decisions(state=...) working with a DeprecationWarning
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: make tests/unit/systemone a package
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(sdk): surface the litellm.decisions(state=...) deprecation as a UserWarning
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
PR #45667 removed openrouter/typesafe/jev-1.13 from the cost map after #45640 moved OpenRouter decisions billing to usage.cost. The OpenAI-format wire test still read the removed row and raised KeyError on CircleCI integration-providers
Co-authored-by: mateo <mateo@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* perf(ci): distribute unit cases and reuse security sweep connections
* revert(ci): reject first speed trial after new failures
* perf(ci): avoid repeated test discovery and parsing
* perf(ci): isolate parallel security workers
* revert(ci): restore measured security execution isolation
* perf(tests): reuse source scans and deduplicate cache cleanup
* perf(ci): restore parallel security with safe log batching
* test(ci): isolate fixture state and schema comparisons
* test(ci): make schema and scheduler fixtures deterministic
* test(ci): drain queued logging before fixture teardown
* test(ci): preserve logging delivery without idle worker restarts
* test(ci): keep guardrail sinks available during parallel bursts
* test(ci): isolate full proxy lifecycle from gateway imports
* test(ci): preserve first failure responses and sweep evidence
* test(ci): restore documentation URL settings between cases
* test(ci): isolate open MCP tools in private gateways
* test(ci): group private MCP cases and handle exited workers
* test(ci): retain proxy exits and MCP failure details
* test(ci): repair merge interactions with main
Key the schema shard's full clone on artifact-name, since main removed the
unit-flag input. Give each owned Redis its own log file again. Hand the
realtime transcription tests main added a logging object whose dispatch
handlers are awaitable, so the per-test logging drain no longer waits on a
MagicMock task forever
* perf(ci): shard owned integration cases and close lifecycle leaks
* fix(ci): qualify shard execution and harden interrupted cleanup
* fix(ci): bound worker supervision and own fault traffic
* perf(ci): reuse unparametrized fixtures across bounded case groups
* fix(ci): preserve boot fault allowances and complete exit records
* test(ci): tolerate startup retries and retain readiness details
* perf(ci): balance integration shards by recorded duration
* fix(ci): load migration database helpers only for workers
* feat(strands_decider): call a Strands Decider on an AgentCore runtime ARN
A strands_decider deployment whose api_base is a bedrock-agentcore runtime ARN now
calls InvokeAgentRuntime on that runtime, signed with SigV4 from the deployment's aws_*
params or the default chain, reuses one session per runtime, and maps the runtime's
200 error envelopes to 400, 503, or 500
* ci: run the strands_decider unit tests in the llms job
* fix(strands_decider): take the AgentCore session id from the deployment, never the caller
A caller could pick the runtime session through extra_headers, and every new id starts another microVM. The session id now comes from the deployment's agentcore_runtime_session_id (banned in request bodies) or the ARN default, and caller-supplied host, x-amz-*, and x-amzn-bedrock-agentcore-runtime-* headers are not forwarded
* fix(strands_decider): map runtime error replies only on AgentCore ARN calls
* fix(strands_decider): sign plain decisions calls inline, off the AWS signing executor
* test(strands_decider): cover AgentCore runtime ARN calls in integration wire tests
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(realtime): only inject response.create when a transcript guardrail suppressed VAD auto-response
The proxy sent `response.create` to the backend after every
`conversation.item.input_audio_transcription.completed`, whether or not a
`realtime_input_transcription` guardrail had disabled the backend's server-VAD
auto-response. Without such a guardrail the backend already created the turn's
response, so the injected one was a duplicate and the backend answered every
user turn with `conversation_already_has_active_response`.
Gate the re-trigger on the same predicate as the suppress side
(`_has_audio_transcription_guardrails`), via `_proxy_drives_turns()`, in both
the raw-backend and the provider_config paths. Guardrails still run on every
transcript; only the injection is gated. Transcription-only sessions keep
skipping the injection as before.
Flip the unit test that asserted the old behaviour and add a provider-path
regression test.
Fixes#31726
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* test(realtime): assert guardrail-gated response.create on the provider_config path
Mirror of test_realtime_guardrail_allows_clean_transcript for backends that go through
provider_config. With a realtime_input_transcription guardrail the proxy disabled the
backend's auto-response, so a clean transcript must trigger exactly one response.create.
Locks in the other half of the suppress/re-trigger pair introduced for #31726.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* test(realtime): type the passthrough provider_config helper's frame argument
* test(realtime): expect no proxy response.create on an unguarded voice session in the integration suite
* test(realtime): audit cells for client-owned turns without a transcript guardrail
* test(realtime): wait for the killed worker to exit, not for its parent to reap it
* test(realtime): pin each worker-kill session's outcome to the worker that accepted it
* test(realtime): drive the provider_config path through a concrete passthrough realtime config
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(decisions): add the OpenAI Decisions spec types and the System One translation
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(decisions): share one DecisionsModel config and require model and usage on responses
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(decisions): rename the shared pydantic parent to DecisionsObjectBase
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(decisions): dispatch /v1/decisions through provider configs and the shared HTTP handler
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(decisions): map OpenRouter connection failures to APIConnectionError and drop explanatory docstrings
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(decisions): switch /v1/decisions to the OpenAI Decisions schema
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(decisions): build the usage the OpenAI response schema now requires
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(lens): ask signal questions through the OpenAI Decisions schema
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(decisions): refuse safety_identifier for System One providers unless drop_params
* feat(decisions): refuse safety_identifier on providers that cannot take it unless drop_params drops it
* test(decisions): cover hosted_vllm in the safety_identifier and per-provider wire tests
* fix(decisions): check a System One request's safety_identifier against the provider too
* fix(decisions): drop a non-string safety_identifier under drop_params
A malformed safety_identifier now follows the drop_params convention in both body shapes: it answers 400 without drop_params and is dropped before the provider call with it. A string identifier on OpenAI stays on the wire either way.
* fix(decisions): let /v1/decisions drop a non-string safety_identifier under drop_params
The route checked the whole body before routing, so a non-string
safety_identifier answered 400 even when the deployment or the body set
drop_params. The route now leaves that field to the Decisions call, which
knows every drop_params source
* refactor(decisions): return the unsupported safety_identifier as a value and raise it in _prepare_call
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(gemini): map OpenAI voice names to Gemini prebuilt TTS voices
* test(gemini): type the voice mapping test helpers
* fix(gemini): leave nova out of the voice mapping since Gemini accepts it as given
* test(gemini): add integration cells for the OpenAI voice mapping on Gemini and Vertex TTS
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(databricks): keep #/$defs refs in json_schema response_format for non-Claude models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(databricks): type the json_schema response_format override
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(databricks): integration cells for json_schema $defs refs across endpoints
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(databricks): split stacked loop in streaming json_schema cell
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(databricks): mark BaseConfig dict contract on json_schema override
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(databricks): cover json_schema refs on every endpoint, edge refs, sad paths and an outage burst
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(proxy): keep the mapped status on assistants and threads route errors
* test(proxy): type and annotate the assistants and threads route error tests
* fix(proxy): redact internal details from assistants and threads route error messages
* test(proxy): integration cells for assistants route error mapping
* test(proxy): drop malformed JSON cases that never reach the assistants route handler
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(azure_ai): support Microsoft-Decision-1 on the decisions API
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(azure_ai): strip project path and API suffix together in decisions base
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(decisions): add Databricks ai_decide as a /v1/decisions provider and auto-router decider
* fix(decisions): reject Databricks endpoint names carrying URL separators and keep the provider wiring under llms/
* fix(decisions): reject dot-segment Databricks endpoint names and add Databricks to the dashboard's OSS classifier editor
* fix(ui): name every OSS classifier provider in the classifier radio copy
* feat(decisions): register the Databricks OpenJev decider as an evaluation-mode cost-map entry
* fix(cost-map): declare zero cache rates on the databricks-openjev-qwen35-4b entry
* fix(proxy): reconcile every DB deployment when a registry miss lands on an auto router
A /model/new for an auto router lands on one replica. A request routed to
that auto router on a sibling replica missed it, the model read-through
loaded only the auto router's own row, and the pre-routing hook then picked
a tier the sibling had not loaded, so the caller got a 400 "no healthy
deployments for <tier>" until the next periodic reload. An auto router miss
now runs the full add_deployment reconcile, tiers included. Found by the
audit's two-worker rig on this PR's routing cells; CircleCI's one-worker
proxy never opens the window.
* test(integration): audit cells for Databricks decisions and the Databricks auto-router classifier
Wire cells for /v1/decisions with databricks/<endpoint> deployments, routing
cells for the complexity router's databricks classifier on all three
endpoints with the Postgres rows read back, management cells for the
classifier's spend writes and for the evaluation-mode health check of a
Databricks serving endpoint.
* chore(github_copilot): sync model catalog with Copilot's served models
* test(github_copilot): assert catalog invariants instead of pinning vendor facts
* fix(github_copilot): keep thinking capabilities on served Claude catalog rows
* fix(github_copilot): keep supports_reasoning on catalog rows that shadow the family fallback
* fix(model-catalog): preserve merged Copilot Haiku metadata
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): list the account's invocable models behind bedrock/* when check_provider_endpoint is on
* fix(bedrock): sign the listing query RFC 3986 style and raise BedrockError on a failed listing
* test(utils): keep test_utils.py at main's formatting, only the Bedrock discovery test is new
* fix(bedrock): chain BedrockModelInfo's constructor so the passthrough config keeps its event parser
* fix(bedrock): list discovered ids without the provider prefix so partial wildcards filter
get_valid_models returns bare vendor ids for every other provider and for the static bedrock catalog, and the proxy's wildcard expansion adds the provider prefix itself. The lister prefixed its ids, so a partial wildcard like bedrock/anthropic.* never matched the proxy's filter and listed uncallable bedrock/anthropic.bedrock/<id> entries
* test(bedrock): integration cells for wildcard discovery through the proxy and the SDK
* fix(proxy): keep a partial wildcard a filter when its deployment repeats the prefix
The wildcard expansion guessed filter-or-alias by whether any provider id carried the prefix. A deployment such as model_name bedrock/anthropic.* over model bedrock/anthropic.* can only ever route names that share the prefix, so when the account has no on-demand anthropic.* id (every Anthropic model behind an inference profile) the guess fell into the alias branch and listed bedrock/anthropic.<every id>, none of them callable. A deployment whose model repeats the suffix now always filters, and lists nothing when nothing matches
* fix(proxy): prefix every discovered id under a custom wildcard prefix that starts a vendor id
* test(bedrock): Final and read-only annotations in the wildcard discovery tests
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(integration): poll the partition lock witness off the event loop
The witness poll ran a blocking psycopg connect and query on the test's event loop, so the partition DDL task only progressed during the 20ms sleeps and missed the 3s deadline on loaded runners. Polling through asyncio.to_thread lets the DDL reach the held lock concurrently, and the deadline is 10s since it only bounds how long the DDL takes to start waiting
* test(e2e): retry the reasoning turn until the model emits a reasoning item
Whether gpt-5.4-mini emits a reasoning item next to a forced function call is up to the model, and it skipped it on four scheduled runs while a same-SHA rerun passed. The first turn now retries up to three times before the unchanged reasoning-item assertion
* test(e2e): retry a Bedrock Converse model once when it reports no cache tokens
Opus on Bedrock Converse reported zero cache tokens on half of the scheduled runs while a same-SHA rerun passed. A model that comes back uncached is asked once more inside the 5 minute window, and the non-zero cache token assertion is unchanged
* test(e2e): poll until a deleted stored response stops being retrievable
Azure kept serving a just-deleted streamed response for a moment, so a single retrieve did not raise. Both delete tests now poll retrieve until it returns an error and still require a 4xx
* test(e2e): give the Together prompt cache five primed attempts
Together documents its prefix cache as best-effort, and three primed attempts came back with zero cached tokens once while a same-SHA rerun passed
* test(e2e): allow 30s for the idle RSS reading at session start
A loaded router replica took longer than 10s to answer the session-start memory read, which pytest reruns cannot recover. The RSS budget assertion is unchanged
* test(e2e): scope the MCP submission check to its own card
The register call timed out at 15s on a loaded stack and leaked a submitted server, so the retries matched two "3 passing, 1 failing" cards. Registration gets 60s and the check reads only the submitted server's card
* test(router): wait for the primary success record before counting it
The shadow fan-out test waited only for the two shadow success events and then asserted exactly one primary success, which the logging worker sometimes had not delivered yet. It needed a rerun on 5 of 59 main runs
* test(proxy): give the spend-log cleanup run a 1s budget
A 0.25s budget could expire before the first batch on a loaded xdist worker, leaving rows_deleted at 0. One second still stops the 50-batch, 5-second loop on the deadline, which is what the test proves
* test(autorouter): release the slow token count at 2.05s instead of 2.3s
The count only has to finish after a quarter of the 8s worker budget, and the planner gives it 3s, so releasing at 2.3s left 0.7s of slack that a loaded runner used up. 2.05s is still past the quarter mark with almost 1s of slack
* test(xai): run the reasoning-effort tests on grok-4.6 with a prompt that needs reasoning
grok-4.7 now returns reasoning tokens without any reasoning text, so reasoning_content was never set and both tests failed on every scheduled run. Called directly, grok-4.6 returned reasoning text 6 of 6 times on a step-by-step arithmetic prompt but only some of the time on a bare greeting
* test(integration): give the two worker-kill chaos tests the owned-proxy time budget
Each test spends about 50s on the burst and then up to graceful_stop_seconds() stopping its owned proxy, which overran the 90s default pytest timeout on loaded runners. They now use the same 2 * graceful_stop_seconds() + 120 budget as other owned-proxy tests
* test(integration): compare regional image responses without their created timestamp
Two identical generations that straddle a second boundary differ only in created, which failed the byte-for-byte comparison. Every other field and both spend rows are still compared
* test(integration): count only the S3 logger's flush task
The set of asyncio tasks created while the logger starts also picked up client close finalizers left by earlier tests, so the count ranged from 1 to 8. The test now counts the periodic_flush task it owns and cancels
* test(integration): send guardrail timeout probes eight at a time
All 44 probes went at once to a two-worker proxy, so on a loaded runner some guardrails used their whole 1s timeout before their request left the proxy, and the sink never saw them. A different set of providers failed on each run. Eight in flight still overlaps the waits
* test(integration): count a killed Arize chaos worker as gone once it is a zombie
psutil reports an unreaped zombie as still running, so the 10s death check failed whenever the supervisor was slow to reap the SIGKILLed worker, and the stop then overran the 90s default timeout. The test now accepts a zombie or a missing process and has the owned-proxy time budget
* test(integration): send the Typesafe connection test its real model and endpoint
The test passed the proxy alias as litellm_params.model, and /health/test_connection lays the request over the stored deployment, so the alias replaced the real typesafe/ model and the check failed with "LLM Provider NOT provided" on every run since it landed in #45481. It now sends the provider model, api_base and api_key directly
* test(integration): keep DB-stored models out of the usage-routing Redis read test
The owned proxy inherited store_model_in_db and the job's shared database, so a model another test left behind joined the router and added its cooldown key to the MGET the test compares exactly. Reproduced locally with one /model/new model present (3 failed), and green with model loading from the database turned off
* test(vertex_ai): prove the batch upload streams by laziness instead of peak memory
The two tracemalloc ratio tests flaked on unrelated PRs because peak memory on a shared xdist worker includes other threads' allocations and garbage from earlier tests. They are replaced by deterministic checks of the same property: the upload stream parses and maps a row only when it is pulled, so a body whose tail is not JSON yields its valid rows first, and a Path source reflects a row rewritten on disk after the upload started. Making the parse, the output, or the file read eager fails these tests
* test(integration): keep the alias test_connection call as a known bug
* test: type the counting mapper and the Bedrock rerun helper
* test: annotate the new test locals as Final and build them as tuples
* test(vertex_ai): prove pull-driven transforms with the garbage tail alone
* feat(decisions): add hosted_vllm provider
Adds HostedVLLMDecisionsConfig so hosted_vllm/<served model> works on
POST /v1/systemone and POST /v1/decisions against a vLLM server that
serves the System One body (vllm-project/vllm#59299, after v0.31.0).
vLLM only supports choice questions and rejects others, so the base
decisions config gains a per-provider health_check_questions attribute
and the evaluation-mode health probe sends a choice question for
hosted_vllm while every other provider keeps the noul probe.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(decisions): move the hosted_vllm placeholder key to constants
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): count a leading system run through count_tokens' system parameter
Anthropic's count_tokens rejects role "system" at the head of messages, so
/v1/responses/input_tokens with instructions, and /v1/messages/count_tokens
with a system-role message, fell back to the local tokenizer. The shared
Anthropic count_tokens transformation now lifts the leading run of system
messages into the top-level system parameter, the way the chat path sends
it, after any system the caller set. Anthropic direct, Azure AI Anthropic,
and Bedrock Mantle share that transformation.
* test(integration): cover the count_tokens leading-system lift across Anthropic, Azure AI and Bedrock Mantle
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(gemini): accept file content blocks with video_metadata on multimodal embeddings
* test(gemini): annotate the embedding mock helper and keep the type comment short
* fix(gemini-embeddings): keep file blocks through a partial embedding cache hit and validate video_metadata strictly
* test(caching): type the partial-hit cache test and pin its clock
* fix(gemini-embeddings): accept the Files API URI /v1/files returns as a file block's file_id
* fix(gemini-embeddings): 400 on bad input shapes, drop unknown block keys under drop_params, cache multi-block inputs
A bare object `input` answered 500 from both the caching handler and the
transformation; both now answer 400. An unresolved `files/` reference on
vertex_ai/ answered a ValueError 500; it now answers 400 naming the gemini/
provider. An empty `format` passed through to the provider; it now answers 400
naming file.format. Unknown block keys (`detail`, an unknown video_metadata
key, a top-level block key) are dropped under drop_params, global or
per-request, and still answer 400 without it. The embedding cache counted
file blocks against `max_messages`, so a request with 5 or more blocks was
never cached; file blocks no longer count.
* fix(caching): answer 400 for an object embedding input on the cache lookup
* test(gemini-embeddings): annotate the new embedding tests with return types and Final
* test(gemini-embeddings): annotate the transformation and files batch tests with return types and Final
* test(gemini-embeddings): audit file content blocks on the wire, in the cache, in batch uploads and under chaos
Integration cells for gemini/ and vertex_ai/ embedding file blocks with video_metadata: the batchEmbedContents and embedContent wire shapes, format overrides, nested and repeated inputs, every malformed block answering 400 before any provider call, drop_params at the request, deployment and YAML levels, the OpenAI SDK sync and async clients, Redis cache fills, hits, partial hits and the metadata in the key with zero-spend hit rows, chat, responses and messages controls, the vertex_ai/ batch JSONL upload, a mixed fast/slow/malformed/dropped burst and a worker kill. The wire peer now reads chunked request bodies, which the streamed GCS media upload sends
* test(integration): treat an upload the client abandons before its terminating chunk as a disconnect, never a stored request
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(voyage): rebrand to VoyageAI by MongoDB, refresh model list, support both contextual input shapes
* fix(voyage): send auto-chunking params for flat contextual inputs
A flat list[str] or bare str for the contextualized embeddings endpoint is
only valid as documents with enable_auto_chunking=True and input_type=document,
or as queries with input_type=query. Default those params for non-query flat
inputs so the request matches the live API contract, while letting caller-set
values win. Nested list[list[str]] still passes through unchanged.
Tests now assert the auto-chunking params instead of only echoing inputs.
* feat(voyage): route MongoDB-issued keys to ai.mongodb.com
Mirror voyageai.util.get_default_base_url from the official SDK: a key with
the `al-` prefix is issued by MongoDB and is only valid on ai.mongodb.com,
every other key on api.voyageai.com. The choice now lives in one shared
helper that the embedding, contextual, multimodal, and rerank configs all
use for both the default base URL and the Authorization header, so the host
and the key always agree.
Also drops the voyage-4-nano and voyage-multilingual-2 model map entries:
voyage-4-nano is not served on the Voyage API and voyage-multilingual-2 is
an older model, so neither belongs in this change.
* fix(voyage): restore MongoDB routing on contextual endpoint and re-add voyage-4-nano after upstream merge
The upstream merge replaced the contextual config's get_complete_url with a
Voyage-only host and dropped voyage-4-nano from the model map. Reapply the
al- key routing via get_default_base_url and add voyage-4-nano (open-weight,
per docs.voyageai.com) so the branch keeps its task changes on top of upstream.
* fix(voyage): drop voyage-4-nano and the contextual tests upstream already covers
voyage-4-nano is not served on the Voyage API, so it does not belong in the
model map. The contextual input tests duplicate
tests/test_litellm/llms/voyage/test_voyage_contextual_embedding.py, which
landed upstream with the auto-chunking fix this branch was carrying.
* test(voyage): drop sys.path.insert from the voyage common utils test
* fix(voyage): route rerank on the key it authenticates with
get_complete_url picked the rerank host from the environment while the auth
header carried the request key, so an explicit MongoDB-issued key was posted
to api.voyageai.com. validate_environment now hands the resolved key to the
config instance and get_complete_url reads it back, so the host and the
credential always come from one key. The config is built per request, so
nothing carries over between them.
* test: allow ultrafast pricing keys in model prices schema
* test: drop ultrafast schema keys now added upstream
* test(integration): prove voyage key-based host routing on the wire
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(vertex_ai): forward the inline-tools-2026-09-15 beta to Vertex and Anthropic
* test(vertex_ai): trim the inline-tools beta test docstrings and type their locals
* test(vertex_ai): add audit cells for inline-tools beta forwarding on Vertex Anthropic
* test(vertex_ai): send legal whitespace in the hostile beta header audit cell
* test(vertex_ai): close the SDK clients in the inline-tools audit cells
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(integration): run the response cache tool-call tests on a proxy without the message cap
Since #43878 the response cache skips any request past 4 messages by default, so the three providers-lane tests that post a seven-message tool conversation twice never saw a cache hit. They now boot one owned proxy from the lane config with max_messages set to null and keep their assertions; the shared proxy and the caching-lane cap test are unchanged
* test(integration): give the uncapped-proxy cache tests the owned-proxy timeout budget
The providers lane runs pytest with a 90 second per-test timeout that includes fixture setup, and the first of the three tests to run pays the owned proxy boot before the spend-log test can poll for up to 70 seconds. The sibling owned-proxy tests already carry pytest.mark.timeout(240), so these three take the same budget
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test: make legacy live tests in spend, batches, openai endpoints and audio dirs offline (partial)
* test: migrate wave-1b live tests offline (guardrails, images, ocr, search, openai endpoints)
* test: fix wave-1b review items, add responses/ocr integration tests and firecrawl unit test
* test: anthropic messages router/bedrock/openai-bridge unit tests for wave-1b nodes
* test: anthropic messages logging, prompt-caching and tool-search unit tests; drop migrated base nodes
* test: finish pass_through_unit_tests nodes, logging drain fix and mutations
* test: migrate anthropic passthrough tests to integration wire tests
* test: fix passthrough migration wire spend row lookup and wildcard config
* test: migrate hosted vllm and openai file passthrough tests offline
* test: move assemblyai and vertex passthrough nodes to in-process unit tests
* test: restore unlisted router node and fix logging worker drain in passthrough unit tests
* test: drop spend-row BUG skip and sharpen non-streaming skip reason for anthropic messages
* test: use public presidio alias after merge
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: restore the batch and file unit tests the migration rewrote away
* test: restore full legacy intent in anthropic messages router unit tests
Drop the false BUG skip on non-streaming aanthropic_messages logging (success
callbacks do fire; the skipped body filtered on the wrong model), assert the
logged model_group, messages, cost and usage, cover streaming logging for both
Anthropic and Bedrock invoke, assert dict content blocks for Anthropic, Bedrock
invoke and the OpenAI bridge, add the Bedrock invoke leg of the router test,
fall back from a real 401, test system-prompt caching and streaming
message_start cache fields on converse and invoke, send the legacy tool-search
tools and beta header, remove the type: ignore and bare dict helper, and add
in-process native /anthropic passthrough spend logging tests
* test: fix passthrough migration wire tests and drop false native spend BUG skip
The native spend row test read rows by the shared master-key digest, so it
matched other tests' rows; read each row by its own message id instead and
assert tokens, total, spend, tags, provider, api_base and end_user on both the
non-streaming and streaming native routes. Merge the streaming test that never
checked spend, assert exact tags from litellm_metadata, stop rebinding a Final
in a loop, require cost > 0, and add the chat-completions bridge cost case the
legacy test covered
* test: harden the spend, OCR, image, search and guardrail migration replacements
The OCR spend tests now build fresh kwargs per case instead of mutating a
shared fixture, and every payload case asserts the exact logged spend. The
OCR wire test reads its spend row by request id and checks exact page
pricing. Image edit, Nova Canvas, DuckDuckGo, Firecrawl, Bedrock guardrail
and Presidio replacements now fake only the provider HTTP boundary (respx or
an in-process aiohttp server) and assert the outbound request, so the
DuckDuckGo limit, Azure base_model pricing and guardrail masking are proven
rather than assumed
* test: cover Exa and Perplexity search structure and max_results offline and retire the two base search methods
* test: drive batch and file replacements through the provider HTTP boundary and real logging callback
Replace monkeypatched litellm.afile_content and AsyncHTTPHandler doubles with respx routes,
read batch logging metadata from a registered success callback instead of get_logging_payload,
use the real managed-files hook for the GEN-2166 regression, assert outbound request bodies,
pin poller ownership explicitly in the migrated DB-sync tests, and require the scripted
upstream to be hit in the responses error-status wire tests
* test: assert file content download headers pass through the proxy
* test: point spend coverage references at the tests that replaced the retired spend job
* test: assert passthrough identity, spend and route dispatch from the code under test
The AssemblyAI non-admin test asserted metadata it wrote itself and leaked a
background poll to the real AssemblyAI host. It now drives assemblyai_proxy_route
with a real Request and waits for the success callback for its own transcript id.
The Vertex spend test matches its log by call id instead of taking the first event.
The OpenAI files wire test hit the native /{provider}/v1/files route; it now calls
/openai/files so the passthrough is what forwards the upload and delete.
* test: mock only the HTTP boundary in the migrated audio tests
Vertex TTS tests no longer replace _ensure_access_token or AsyncHTTPHandler.post;
the token comes from a mocked Google OAuth endpoint and the synthesize call from
respx. Speech tests assert the outbound body, the transcription cache test polls
for the cache write instead of relying on test ordering, and the model pass-through
test checks the multipart model field per model.
* test: wait on a logger event instead of polling the clock in anthropic messages unit tests
Recorders keep payloads in a rebound tuple and set an asyncio.Event; tests
await it with asyncio.wait_for instead of a sleep-and-deadline poll loop
* test: freeze module-level batch and file response fixtures as Final MappingProxyType
* test: fake the presidio analyzer with an in-memory aiohttp connector
The blocked-entity tests started an aiohttp TestServer, which binds a local
socket. They now hand the guardrail a ClientSession whose connector answers
/analyze and /anonymize in process, so no socket is opened and the outbound
analyze text and entities are still asserted
* test: type anthropic messages router test helpers with LiteLLM's Anthropic TypedDicts
Messages, cached system blocks and tool-search tools now use
AnthropicMessagesUserMessageParam, AnthropicMessagesTextParam,
AnthropicToolSearchToolRegex and AnthropicMessagesTool instead of bare
dict shapes; tools are converted to plain dicts only at the acreate call,
whose tools parameter is list[dict]
* test: type batch limiter helpers with TypedDicts and wait on the logging callback event instead of polling
* test: give the migrated OCR, image and presidio helpers precise types
OCR spend helpers take ReadOnly TypedDicts for kwargs and responses and use
LiteLLM's OCRResponse/OCRUsageInfo instead of local pydantic stand-ins; spend
metadata is validated with a TypeAdapter. The presidio fake uses LiteLLM's
PresidioAnalyzeRequest/ResponseItem types, and the image-edit logger validates
the logged payload instead of storing an untyped dict
* test: signal callback and cache events instead of polling
Recorders keep tuples and set an asyncio.Event, thread-safely, when the payload
for this test's transcript id or upstream URL arrives. The transcription cache
test waits on a Cache subclass that signals after async_add_cache. No clock
polling or sleeps remain in these tests.
* test: assert the batch limiter hook updates the caller's request in place
* test: tolerate model-list probes and read native passthrough rows by owned key
The router's OpenAI-compatible model-info refresh (litellm/router.py:10710)
sends GET /v1/models to configured openai api_bases, so the wire answers it
with an empty list and excludes it from the provider-call assertions. Native
/anthropic spend rows are now read by a per-request virtual key digest and
call_type, then the row's request_id is checked against the message id
* test: expect the OCR alias in the proxy response model
The proxy restamps every OpenAI-compatible response model to the name the
client requested (_override_openai_response_model), so /v1/ocr returns the
scenario alias. The upstream model is now checked on the drained request body
instead of inside the peer, where a failed assert never reached the test
---------
Co-authored-by: yuneng <yuneng@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: make 77 legacy live tests offline in litellm_utils, router_unit and responses dirs
* test: restore lost coverage in litellm_utils offline replacements
Bound live health-check tasks to the concurrency limit so eager task
creation behind a semaphore fails, route the Langfuse trace-id check
through completion -> Logging.get_trace_id with all four metadata
combinations, run the callback dedup checks through acompletion, pass
the dynamic key via litellm_params, cover the env-inferred default
model list, and add the per-provider audio transcription config lookup
* test: restore router coverage lost in the offline move
Add the missing non-stream LIT-3058 header/count-once test, pin the UTC minute on every
usage-counter test, assert router-level client reuse for transcription, and replace the
private selector-attribute checks with routing outcomes per strategy. Assert outbound bodies
for speech, rerank, image, assistants and moderation, add timeouts to event waits, and drop
the router_unit_tests husk files that no longer hold tests
* test: restore sync stream, router sync stream, error event and field-type coverage for migrated responses tests
Add the sync streaming logging and sync Router.responses streaming cases the
offline replacements dropped, port the legacy per-event and response field-type
validation, assert raw headers, the float created_at conversion, MCP tool headers,
the search_context_size input, and add an offline replacement for the in-stream
context-window error event. Remove helpers left dead in the legacy file.
* test: run numpydoc-backed unit tests in GHA and drop a test that pinned a crash
* test: restore sequence number, item id and content part checks in the responses stream validator
* test: match router logging events to the test's own deployment so late events from other tests are ignored
* test: restore the s3 cold storage history test to its original assertions
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* fix(cache_control): let a hosted deployment opt into the OpenAI cache dialect
_targets_openai_prompt_cache_breakpoint gated on custom_llm_provider == "openai"
unconditionally, so an OpenAI-shaped model served by another provider could never
qualify, even with supports_prompt_cache_breakpoint set explicitly on its own
cost-map entry. bedrock_mantle honours prompt_cache_breakpoint end to end, and
had no way to say so.
model_cost is keyed per exact deployment string, so a flag on the deployment's own
entry states the dialect more precisely than a provider name can. Consult it
before the provider check, and require the entry's litellm_provider to match the
serving provider so an openai entry cannot license another provider. Entries for
the openai provider keep their api_base check, so an OpenAI-compatible third-party
host is still not assumed to speak the dialect.
Fixes#38666
* test: drop a laziness assertion the existing suite already makes
test_provider_lookup_skipped_for_models_below_gpt_5_6 already pins that
_resolve_provider is not called for an unflagged model, and it covers the new
flag lookup unchanged. The duplicate patched a litellm internal for no added
coverage, which the test-quality gate counts against TQ008.
* fix(bedrock_mantle): flag GPT-5.6 and newer rows for OpenAI explicit prompt cache breakpoints
* fix(responses): predict the chat-completions bridge with the model name the dispatch resolves
* fix(cache_control): read a region-prefixed Mantle GPT deployment through its region-free price-map row
* fix(cache_control): let a bare hosted deployment name read its own provider's price-map row
* fix(responses): hand the hook the provider the router resolved for a Foundry GPT deployment
* fix(cache_control): read the deployment breakpoint flag strictly and default a null prompt_cache_options
A cost-map or deployment `supports_prompt_cache_breakpoint` that is not the boolean `true` (the string "true",
the integer 1, "false", 0, a long string) no longer opts a deployment into the OpenAI prompt cache dialect; only
`true` does, the same reading Bedrock Converse applies to its own flag.
A client that sends `"prompt_cache_options": null` on `/v1/chat/completions` or `/v1/messages` now gets the
implicit default the hook already applied on `/v1/responses` when it placed a breakpoint; before, the null
suppressed the default and the request left with a breakpoint and no options.
* test(integration): audit cells for the Mantle GPT prompt cache breakpoint dialect
Deterministic cells for the configured-breakpoint path on Bedrock Mantle GPT and Azure AI Foundry GPT-6
deployments: every endpoint, streaming and not, sync and async SDKs and raw httpx, the hostile option shapes,
flag precedence, the response cache twin, a two-worker burst, a worker kill and a graceful restart.
* fix(cache_control): ignore a malformed cache_control_injection_points value instead of failing the request
A deployment or client `cache_control_injection_points` that is not a list of points (a string, an
integer, a bare point dict, a list of strings) raised inside the prompt hook (`'str' object has no
attribute 'get'`, `'int' object is not iterable`) and turned every request to that deployment into a
500 on `/v1/chat/completions`, `/v1/messages` and `/v1/responses`. Every entry point now reads such a
value as no configured points, the way a `null` already read, and the request leaves without a
breakpoint; entries of a list that are not points are dropped and the point entries kept.
* test(integration): cover int, dict and string-list injection point shapes in the Mantle audit cells
The D9 cell now also sends an integer, a bare point dict and a list of strings as the deployment's
`cache_control_injection_points`, which the merge base answered with a 500 on every request.
* fix(cache_control): stamp the OpenAI dialect for a bare deployment name served by a flagged provider
A deployment written as `model: openai.gpt-5.6-sol` with `custom_llm_provider: bedrock_mantle` has no cost-map row
of its own and no openai row of the same name, so the dialect stamp's cheap gate (`supports_openai_prompt_cache_breakpoint`)
returned early and the configured point was never stamped. On `/v1/responses` and on the chat seeding path the hook
sees no provider, so it fell back to the Anthropic `cache_control` marker, which Bedrock Mantle strips.
The gate now also passes a model whose serving provider is already known and whose provider-keyed row carries the
flag, which is the same row the dialect resolution reads and costs no provider lookup. A bare name without a provider
is still left alone, so models below GPT-5.6 keep their points untouched and resolve nothing.
* test(integration): cover a bare Mantle deployment name with its provider in the audit cells
A deployment configured as `model: openai.gpt-5.6-sol` plus `custom_llm_provider: bedrock_mantle` sends the
breakpoint and the implicit options on all three endpoints; the merge base leaves it on the Anthropic dialect.
* fix(cache_control): keep a configured point that sits beside junk entries on every chat seed path
`configured_injection_points` kept the point dicts of a mixed `cache_control_injection_points` list as a tuple,
but only read a `list` back. The chat seed and the `/v1/responses` dialect stamp write the normalized value back
onto the request, and on a deployment the OpenAI dialect does not stamp (an Anthropic model, Claude on Bedrock
Mantle) that value is the tuple itself, so the hook then read it as no configured points and the valid point was
silently dropped. The stamped OpenAI dialect and the `/v1/messages` path kept it only because they build a new list.
The normalizer now reads back the tuple it wrote, so the point reaches the wire on every path; an all-dict list
still passes through as the same object.
* test(integration): cover a configured point beside junk entries on a Claude Mantle deployment
A deployment configured with `cache_control_injection_points: ["system", {system point}, 3]` marks the system
block on the Anthropic dialect on all three endpoints; the merge base answers 500 on the junk entry.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(responses): map input_audio blocks in the chat-to-Responses bridge
* fix(responses): honor base_model when deciding to drop input_audio under drop_params
* fix(responses): carry a dropped audio part's cache breakpoint to the preceding part
Dropping an input_audio part under drop_params lost the prompt_cache_breakpoint
riding on it, so an injected system marker landing on a trailing audio block left
prompt_cache_options explicit with no breakpoint on the wire. The marker now moves
to the nearest preceding kept part when that part has none.
Update the CircleCI-only Responses bridge wire tests that pinned input_audio to
its stringified input_text shape: video_url becomes the stringified kind so the
marker-on-stringified-block coverage survives, and dedicated audio tests assert
the forwarded part and the drop_params carry-over.
* refactor(responses): carry dropped audio markers in one pass
Replace the per-block slice that re-read the rest of the content list for
every retained part with a single reverse scan, and skip the copy for a
message that carries no audio part. Type the unit test helper's messages
as a sequence of mappings instead of bare lists.
* test(responses): add integration cells for input_audio parts through the bridge
Cover the chat-to-Responses bridge's input_audio handling on the real proxy: the part is forwarded
as input_audio without drop_params and dropped under drop_params unless the model or its base_model
supports audio input, across httpx, the OpenAI SDK (sync stream and async), the Anthropic SDK, native
/v1/responses, tool and assistant messages, hostile input_audio values, duplicate parts, an audio-only
message, marker carry-over, a null drop_params, a global litellm_settings.drop_params, a mixed burst,
and a burst with dropped upstream connections
* test(responses): check every upstream attempt of a dropped connection
The dropped-connection cell required exactly one upstream POST per call,
which tied it to the retry policy instead of the fix. It now groups the
recorded POSTs by marker, requires one attempt per answered call, and
checks that every attempt of a doomed call carries the dropped audio shape.
* feat(decisions): serve System One format at /v1/systemone and OpenAI format at /v1/decisions
The System One request format moves to /v1/systemone and /systemone. /v1/decisions and
/decisions now accept the OpenAI Decisions API format, translate it into a System One
request, route it through the same pipeline, and translate the answers back.
* fix(decisions): import assert_never from typing_extensions for Python 3.10
* fix(decisions): cap OpenAI-format questions at the System One limit
/v1/decisions accepted up to 200 questions, but the System One request it translates to takes at most 128, so 129 to 200 questions failed with an internal validation error. Both now share MAX_DECISION_QUESTIONS.
* fix(decisions): return 400 for bodies that are not JSON
* test(decisions): send System One bodies to /v1/systemone in integration tests
* fix(decisions): return guardrail_information on OpenAI-format decisions when requested
* refactor(decisions): validate proxy request data before reading guardrail settings
