* test: move offline proxy auth, guardrail hook, spend, budget reset, GCS payload and pass-through tests into tests/unit
Relocate 53 legacy tests that pass with no network or keys into the tests/unit
files that mirror the code they exercise, drop 2 llm_guard tests already covered
by test_llm_guard_call_type_aliases, and delete the emptied legacy files.
* test: freeze the reset clock and run the real reset methods in the moved budget reset tests
Pin datetime in reset_budget_job and timezone_utils to a fixed instant and await
the service-hook tasks the job spawns instead of sleeping. Drive the per-user
failure with a malformed budget_duration row from the prisma mock so the real
_reset_budget_for_key/user/team run instead of patched replacements.
* test: use fixed timestamps in the moved pass-through and spend log tests
* test: mock the moderation HTTP call and type the moved moderation and spend log tests
Serve the OpenAI moderation response through an httpx MockTransport so the real
async_make_request runs, annotate the moved tests' locals with Final, and pass
get_logging_payload typed kwargs instead of an untyped input_args dict.
* test: drive the moved proxy tests through boundaries and annotate their locals with Final
Serve GCS downloads and the Router moderation call through httpx mocks with fake
Google credentials, let the real pass-through success handler fail on the mocked
upstream response, and fold the Vertex live route endpoint check into the existing
route test. Annotate moved locals with Final, build the budget reset rows without
rebinding, and type the remaining fakes without **kwargs
Moves 58 legacy nodes (cyberark, logging callback manager, langfuse handler and helpers, SQS, app crypto, bedrock nova and invoke, cloudflare, litellm_proxy, nvidia nim, replicate, triton, convert_dict_to_response, get_model_info, vcr live-call probe) into the tests/unit files that mirror the modules they exercise. Calls that reached huggingface.co, langfuse or 0.0.0.0 are refused at the HTTP boundary with respx. Tests that need credentials, a license, a spawned server, sleeps, a subprocess, or assert nothing stay in their legacy files
* 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>
* test: delete 77 legacy tests covered by e2e, unable to fail, or dead in CI
* test: keep router helper tests and coverage ignore list unchanged
* test: keep cohere error handling tests
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test: move the unit half of 126 mixed legacy files into tests/unit
* test: restore litellm globals that moved tests set
* test: finalize migration test cleanup
* test: restore original bodies of moved legacy tests
The move into tests/unit had rewritten 612 test bodies, and some of the rewrites dropped assertions. Each moved test now carries its original body from the legacy file, with only the imports, helpers, fake provider credentials and monkeypatched env it needs to run under tests/unit
test_timeout_streaming goes back to tests/local_testing because it needs the fake OpenAI endpoint server. The image payload fixture moves with its only user, and two tests that leaked global state (a registered model cost entry and queued logging tasks) are now isolated
* test: drop module imports shadowed by restored local imports
* test: assert on LiteLLM output in no-assertion moved tests and isolate leaks
Twenty no-assertion candidates get one assertion on the value LiteLLM returns, with the original lines unchanged. Four tests go back to their legacy files because they only check types or imports, write into the working directory, or cannot assert without a body change
Two moved tests leaked globals into later tests in the same worker, so monkeypatch fixtures now restore the retry-after header parser and the end user cost tracking flags
* test: drain queued logging tasks before the Phoenix span test
The moved Phoenix test counted spans from logging tasks that earlier tests had queued, so the drain fixture moves to tests/unit/conftest.py and both it and the Datadog batch test use it. test_factory_function goes back to its legacy file because its returned wrapper calls the real Assistants API and cannot be asserted on without a body change
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test: delete unconditionally skipped legacy tests
* test: move whole-unit legacy test files into tests/unit
* test: keep moved legacy tests free of import-time global state
* test: keep the module-level invocation scan pointed at tests/local_testing
* ci: drop the agent_testing CircleCI job emptied by the move
* test: fix moved-test isolation and router coverage
* test: add Tinyfish search package marker
* test: isolate moved tests from logger state leaks
* test: isolate Helicone logging fixture state
* test: isolate Vertex pass-through credentials between moved tests
* test: cancel S3 periodic flush tasks started by moved tests
* ci: restore CircleCI assistant test selection after move
* test: prevent Datadog datetime import shadowing
* ci: drop the litellm_assistants_api_testing CircleCI job emptied by the move
* test: deduplicate imports in rebased unit tests
* test: remove duplicate passthrough router patch import
* test: remove moved legacy source files after rebase
* test: align moved tests with rebased main
* test: carry main's legacy-file edits into moved destinations
* test: make the moved cost map fallback tests assert the fetch and the backup
The four fallback cases only checked the result was non-empty, so they still
passed with integrity validation disabled. They now inject a mock client, assert
one fetch happened, and assert the result is exactly the local backup with the
fallback reason recorded.
---------
Co-authored-by: yuneng <yuneng@berri.ai>
* test(integration): isolate Codex catalog and provider discovery fixtures
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): keep model discovery constant import-safe
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: align test keys and CI env with generated master keys
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@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>
* fix(ci): stop five stale or flaky CI reds and retry CyberArk policy-load conflicts
The Langfuse redaction unit test exports to a local OTLP capture instead of
polling Langfuse Cloud through a recorded lookup. The passthrough worker-kill
test only requires spend rows for requests the surviving worker served. The
spend-routes sweep treats the intentional /spend/capture_rate 503 as expected.
CyberArk retries a 409 policy load in Python, Rust and the e2e Conjur helper
instead of reading it as "variable exists". The integration egress guard now
matches the script's own cgroup, so it no longer blocks the CircleCI agent,
which runs as the same user.
* fix(ci): keep the policy-load backoff typed as float
* fix(ci): retry CyberArk policy loads without blocking the event loop and tighten the worker-kill and Langfuse tests
* fix(secrets): load CyberArk policy one request at a time per manager
* test(secrets): pin that non-conflict CyberArk policy failures are not retried
* test(unit): run tests/unit with only an allowlisted host environment
CircleCI's unit job inherits every project env var, so real provider keys,
REDIS_HOST, DATABASE_URL and AWS or Azure credentials reached tests that
assume none are set. Locally, litellm's import-time load_dotenv did the same
from any .env up the tree. The unit conftest now drops every variable outside
a small allowlist and disables dotenv before litellm is imported.
* test(e2e): name a failed search and the stuck batch status instead of misattributing them
The websearch session test read an empty web_search_tool_result_error block as a
successful search, so a failing search tool surfaced as a session billing bug.
The batch cancellation timeout now reports the last status the proxy returned.
* fix(ci): scrub the host environment per unit test instead of for the whole pytest process
GHA shards run tests/unit next to other suites in one process, so the import-time
scrub deleted MCP_TEST_PEER_PYTHON before tests/mcp_tests read it and the MCP
upstream fell back to the SDK2 interpreter. The two websearch tests that called
OpenAI and Perplexity live are removed: tests/unit no longer sees their keys.
* fix(ci): scrub only the host variables present before litellm is imported
The per-test scrub also deleted TIKTOKEN_CACHE_DIR, which litellm sets at import to
its bundled encodings, so tokenizer paths tried to download them and hit the
socket guard. The prisma setup test now passes its own database URL instead of
reading one another test leaked into the process environment.
* fix(ci): stop the order-dependent unit reds and settle logging tasks on their own queue
LoggingWorker marked a task done on whichever queue was current when the callback
finished, so a callback that outlived an event-loop change raised "task_done()
called too many times" or undercounted the new loop's queue. It now settles the
queue the task came from.
The rest are test isolation fixes for failures that only appeared when another
file ran first on the same xdist worker: a replaced user_api_key_cache, breaker
metrics unregistered by prometheus tests, semantic_router's health-check filter on
uvicorn.access, logging tasks carried over from bedrock tests, a Router-written
model_cost entry, and a stray post captured by the langflow test. The token
counter check now asserts bounded chunking instead of wall-clock time.
* test(e2e/ui): wait for the logout redirect before visiting a protected page
Logout revokes the session server-side before clearing cookies and navigating, so an immediate page.goto either ran with the cookie still set or was aborted by the logout redirect (net::ERR_ABORTED).
* test(unit): restore the prometheus metrics config per test and settle logs carried from earlier tests in the a2a cost tests
* test(router): pin the router clock in the usage counter tests so a minute rollover cannot empty the read
* test(e2e/ui): wait for logout to clear the token cookie instead of for a login redirect
* test(integration/mcp): answer the model-info probe another test's proxy sends to the model double
* feat(langfuse): migrate the sdk callback to langfuse v4
Replace the v2 trace()/generation()/span() calls with SDK v4 observations exported over OpenTelemetry, with one isolated tracer provider per Langfuse credential set, a discarding exporter for mock mode, and v4 trace and observation id normalization. Keeps the session-header trace provenance logic from main so each call under a session alias still gets its own trace
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): drop the always-true prompt client check now that v4 get_prompt is non-optional
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(langfuse): isolate the e2e sync test from cached clients and log the real sdk major
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): type the slack trace-url lookup and drop dead v2 test shims
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(slack): cover the langfuse trace url built from the logger host
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* build(docker): pin langfuse to the locked 4.15.2 in the pip image
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): hash all-zero trace and observation ids instead of passing them through
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(langfuse): honour caller generation ids and assert v4 OTLP exports in legacy tests
v2 accepted generation(id=...). v4 derives the observation id from the OTel
span id, so the isolated tracer provider now carries an id generator that
hands out the id start_generation asked for through a context variable, and
the callback passes the resolved generation_id metadata into it.
The legacy e2e suite patched httpx.Client.post and compared v2 ingestion
batches; it now patches requests.Session.post, decodes the OTLP protobuf
and compares the exported generation against regenerated fixtures. The
local readback test replaces the removed get_generations() with
api.observations.get_many() and polls Langfuse Cloud instead of sleeping.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): read the sdk version header from package metadata
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): propagate trace_metadata as trace-level attributes in v4
v2 wrote trace(metadata=...) onto the trace object. In v4 the trace only
carries what the observations propagate, so a continuation request with
update_trace_keys=["trace_metadata"] updated the generation's metadata
while the trace kept its stale values. Coerce each entry to the SDK's
string limit and hand it to propagate_attributes(metadata=...).
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): propagate interrupts raised during deferred client teardown
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): honor ssl_verify=False and SSL_VERIFY on the v4 OTLP exporter
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): fall back to the default CA when the configured bundle path is missing
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): renew the client when eviction lands before the callback lease
The cache can evict a logger between handing it to the callback and the callback taking its
lease. Such a lease now hands back a fresh client acquired through the same parameters, so that
callback exports through a live tracer provider instead of one teardown already shut down.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): emit litellm_call_id and response_id as generation metadata
v2 put the provider response id inside the generation id. v4 observation ids are 16 hex chars derived from that string, so the ids move to generation metadata to keep generations searchable by response id
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): read the response id through a typed protocol
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): do not claim trace root when continuing an existing trace
Langfuse derives a trace's name and I/O from any observation flagged
langfuse.internal.as_root, so a request carrying existing_trace_id
renamed the trace to the generation name and replaced the trace input
and output on every continuation. v2 only updated the keys listed in
update_trace_keys. Continuations now export as plain children of the
remote parent and keep the explicit langfuse.trace.* attributes for the
fields they do want changed.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): iterate lease renewal instead of recursing, monkeypatch update_trace_keys flag in tests
The recursive lease fallback tripped tests/code_coverage_tests/recursive_detector.py; the renewal
candidates are now walked with itertools.chain. The six update_trace_keys tests set the litellm
global through pytest monkeypatch so the TQ008 budget stays within its ceiling
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): retry raised OTLP exports and honor LANGFUSE_TIMEOUT
The OTLP http exporter only retries 429 and 5xx; a connect or read timeout
propagates and BatchSpanProcessor drops the batch. Wrap the exporter in
RetryingSpanExporter (three backoff retries, as the v2 consumer did) and
build it on every path so the default and private-CA deployments share the
same channel, timeout and retry behaviour
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): sample on a hash of the full trace id and tolerate bad LANGFUSE_SAMPLE_RATE
TraceIdRatioBased reads the low 64 bits of the trace id. litellm trace ids are
UUIDs, whose variant bits sit at the top of that word, so every fractional rate
up to 0.5 dropped all traces. A SHA-256 of the full id gives an unbiased,
deterministic decision. Values outside [0, 1] or non numeric now warn and export
everything instead of raising during callback construction, which surfaced as a
500 on the first request of each worker
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): put the Langfuse trace link back into Slack alerts
The proxy registers LangfusePromptManagement for callbacks: ["langfuse"], so the alert helper never saw the literal "langfuse" string and returned before looking up the trace id, and the prompt management logger never stored the trace id it got back from log_event_on_langfuse. Recognize LangFuseLogger instances in the callback list, record the returned trace id in the shared service trace id cache, and skip the link when no trace id arrives
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(deps): relock langfuse 4.15.2 and opentelemetry 1.33.1 on current main
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* chore(langfuse): mark the deliberate blind except in client teardown for the strict ruff gate
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): pass the resource attributes mapping straight to Resource.create
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): warn about ignored UPSTREAM_LANGFUSE_* on the shared client init path too
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): normalise the OTLP export path so a trailing host slash never yields a double slash
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): nest guardrail and grounding spans under the generation
Langfuse v4 derives the trace name and I/O from every observation marked as_root, and the one with the latest start time wins. Guardrail and grounding spans used to claim root next to the generation, so a post_call guardrail could replace the model's request and response on the trace with its own. Only the generation claims root now; the sibling spans become its children
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): rebuild the cached bundle when mock mode or sample rate changes
The SDK keys resource bundles on the public key alone, so a bundle built with the discarding exporter for LANGFUSE_MOCK, or with an earlier LANGFUSE_SAMPLE_RATE, was handed back to a client that asked for a live exporter or a different rate. Compare both when deciding whether the cached bundle is still valid
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): keep trace_public true when a guardrail span is exported
Langfuse folds langfuse.trace.public across every observation in the trace and reads a missing attribute as false, so a guardrail child span without the flag turned a trace_public: true request private on Langfuse Cloud. Child spans now repeat the generation's value
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): emit observations as plain OTel spans, keep the SDK for prompts and auth
The callback now owns an isolated TracerProvider and OTLP exporter and builds generation and child spans with public OpenTelemetry APIs plus the LangfuseOtelSpanAttributes constants. Caller trace ids, generation ids, parent observation ids and historical start and end times are honoured through the OTel id generator, remote SpanContext and explicit span timestamps, so no private Langfuse SDK tracing handle is used any more. The Langfuse client stays only for get_prompt and auth_check
This also resolves the gauntlet findings on the previous draft: fresh traces start from an empty context so caller application spans are never stamped, the Slack trace link is read from the request logging state instead of constructing a logger per alert, a truthy non-mapping trace_metadata is serialized instead of raising, trace_input and trace_output land on the root generation, discarding a cached client is done under the lock, and the prompt cache no longer leaks a task manager because the client cache no longer tears down shared providers
Fixtures under tests/logging_callback_tests lose the SDK-private langfuse.internal.as_root marker; every other exported attribute is unchanged
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): hand the SDK client a validated sample rate so an unusable LANGFUSE_SAMPLE_RATE no longer breaks the callback
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): gate the SDK version before importing the OTel module in prompt management
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): flush every export channel on proxy shutdown and use the callback's host in Slack trace links
The shutdown hook imported litellm.utils.langFuseLogger, a global the callback registry never assigns, so a graceful restart dropped the spans still queued in the batch processors. Shutdown now calls flush_langfuse_tracing, which force-flushes every acquired channel. The Slack alert link falls back to the registered LangFuseLogger's langfuse_host when the request carries no dynamic host, and the export endpoint tests pin that scheme-relative or absolute LANGFUSE_OTEL_TRACES_EXPORT_PATH values stay on the configured host
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): store resolved credentials on LangfusePromptManagement
The Slack alert trace link reads langfuse_host from every registered LangFuseLogger. Prompt management subclasses it without calling the parent constructor, so it never set the attribute and the alerting handler crashed before posting
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): flush every export channel concurrently under one shutdown deadline
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): flush export channels on daemon threads so a stuck channel cannot hold up interpreter exit
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): own the tracer config and drop the SDK client for prompts and auth
The callback's TracerProvider now sets its sampler, span limits and id generator explicitly so unrelated OTEL_* variables no longer change what Langfuse receives, and trace metadata is written once on the trace instead of folded into the generation, which kept input and output under the attribute cap. Spans are emitted under the langfuse-sdk scope so Langfuse renders them natively, the batch processor queues 100k spans and honors LANGFUSE_FLUSH_AT, and the proxy shutdown flush runs off the event loop with a 10s deadline and logs a miss.
Prompts, auth_check and the project id now go through LangfuseAPI directly with a litellm-owned TTL cache, so no Langfuse() client is built and a host application's client on the same public key is left alone. Dead attributes, the unreachable exporter branch and the export list are cleaned up, and the client-budget eviction behavior is documented.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): export OTLP spans and fetch prompts through litellm's HTTPHandler instead of a private requests session
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): gate the SDK version before importing the tracing module and retire unheld export channels
An installed v2 SDK used to fail inside the langfuse_sdk import and surface as "Langfuse not installed"; the version check now runs first so v2 users get the upgrade message, and only PackageNotFoundError means the package is missing
Export channels are now leased per credential set: acquire adds a holder, LangFuseLogger.stop (called by DynamicLoggingCache on expiry) releases one, and a channel with no holders is flushed and shut down after a 60 s grace, so rotating key or team credentials no longer grows one batch thread per credential set for the life of the process
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): end the generation when a child span fails, take the client slot last, keep prompt cache keys structured
Generation spans now end in a finally block so a bad guardrail or provider entry cannot strand the trace. The logger acquires its export channel and REST client before counting a client slot and releases the channel synchronously if the REST client fails to build, so retries after a bad config do not exhaust the budget. LANGFUSE_TIMEOUT accepts decimals for the REST client like it already did for OTLP export. The prompt cache keys on (name, version, label) so a missing label and the literal label None stay apart
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): claim the cache entry before releasing its slot and channel hold on eviction
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): coerce generation names, keep v2 release, timeout and retry defaults, refresh stale prompts off the loop
A non-string metadata generation_name reached the OTLP encoder and took the whole batch down; it is now exported as its text and the exporter drops only the span the encoder rejects. LANGFUSE_RELEASE falls back to the deploy platform's commit variable again, the export deadline is back to the v2 default of 20 s and LANGFUSE_MAX_RETRIES sizes the retry ladder. An expired prompt is served at once while one background thread refreshes it, a re-acquired export channel cancels the pending retire timer, and flush reports delivery rather than a drained queue
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(proxy): assert the current Langfuse shutdown flush warning
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): keep host OTel resource out, carry big metadata ints, tolerate bad flush and TTL env, stamp trace I/O under a parent
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): name a malformed prompt cache TTL before the SDK import, keep metadata ints JSON safe, retry every 5xx export
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): name the auth check failure, split a 413 export, wire LANGFUSE_DEBUG, stamp error output under a parent, send the ingestion version header
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): honor LANGFUSE_DEBUG on the callbacks path, cap retry backoff, name the auth failure status and body
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): cap LANGFUSE_MAX_RETRIES at 1000 so an absurd value cannot stall callback init
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): fold 413 halving into bounded rounds instead of recursion
The code-quality recursive-function gate flagged LangfuseSpanExporter.export. A batch of n spans settles within n.bit_length() halving rounds, so the split is a reduce over a frozen round state with the same posts, logs and results. The TTL gate test now asserts the gate returns without raising
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): truncate a single oversized span like v2 instead of dropping it, no retries on REST auth and project lookups
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): write the metadata truncation marker under a flattened key so Langfuse keeps it
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(langfuse): patch the HTTPHandler export path and sync the metadata fixture and lease registry with the v4 callback
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(langfuse): give the 413 split helpers a single explicit return path
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): url-encode prompt names and fetch cold prompts without client retries
A cold get_prompt runs inline on the event loop; the generated v4 client's default two retries slept through
Retry-After (up to 60 s per attempt) and held the loop. The wrapper also passed the raw name into
api/public/v2/prompts/{name}, so 'what?' fetched prompt 'what' and folder names left the route. Quote the
name with safe='' like the v4 SDK's own get_prompt and pass max_retries=0 like the projects.get calls
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(langfuse): retry a cold prompt miss once and drop upstream headers from prompt errors
A cold prompt fetch makes one immediate second attempt after a 5xx or a
transport failure, as the v2 client did, still with the generated client's
sleeping retries and Retry-After handling off so the event loop never stalls.
A failed fetch raises LangfusePromptError carrying only the status and body,
so the proxy no longer forwards Langfuse's response headers to its client
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(langfuse): stub the logger in the health auth_check test instead of dialing a closed port
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(langfuse): integration test for OTLP v4 delivery and prompt fetch through a real proxy
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* build(docker): keep the pip image's langfuse and otel pins on the v2 line its litellm 1.83.0 wheel expects
The image validates the published PyPI artifact, whose langfuse callback still
reads langfuse.version, so the 4.15.2 pin broke that callback. The pins move
together with the next LITELLM_VERSION bump. Also rewords the trace_version
precedence test docstring: v2 carried two version fields, v4 has one per span
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test: point CircleCI-only suites at models still in the cost map
#42435 removed cost map entries past their deprecation date and #42437 added
litellm_uisettings to the config-synced tables, but both only updated
tests/test_litellm. The CircleCI-only suites (local_testing, llm_translation,
logging_callback_tests, litellm_utils_tests, unit) kept using the removed
models or the old table list and went red on main.
Each test keeps its assertions and swaps the removed model for a current one
with the same provider and capabilities. The fireworks tests pick a vision
model from the cost map because #34941 set supports_vision false on
minimax-m3, and the vertex image provider test injects the image model set
because #42435 removed every vertex_ai-image-models entry.
* test(vertex_ai): register the image model through add_known_models in the provider test
Regenerated every touched file from origin/main applying only the B1 test deletions and the unused import and helper cleanup they leave behind, without running the formatter across untouched code. CI only checks ruff format under litellm/, so the earlier reflows of test files were pure diff noise for reviewers
Also drops the tests/local_testing/test_prompt_caching.py entry from the caching-local shard in test-unit.yml since that file is deleted
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The repo rule is that a test must only fail when litellm code changes, never when a vendor updates a price, renames a field, or drops a model. These tests asserted shipped catalog entries directly, comparing lookup results to literals copied from model_prices_and_context_window.json or requiring named entries to exist or be absent, so every cost map sync could break them without any litellm code changing
Tests that exercise real litellm behavior with an injected local model_cost, invariants like backup parity, and assertions on non-lookup code paths are untouched
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
#41607 registered the typesafe pass-through with a route that only accepted
GET and POST, so a PUT, DELETE or PATCH to /typesafe/... came back 405
before reaching the upstream. CircleCI's pass-through method test caught it,
but that lane does not run on the PR gate, so the mapped unit test now
covers the same invariant for typesafe
The same CircleCI run also failed test_models_by_provider because typesafe
is not a key of models_by_provider. Registering it there would satisfy the
assertion without changing behaviour: typesafe has no LlmProviders member,
so a typesafe/* deployment never loads and get_valid_models returns nothing,
and its spend is priced straight from model_cost. The test already skips
search-mode providers for that reason, so it now skips evaluation mode too
The budget-tier reset read every customer linked to an expiring tier into
one result set before the write, then invalidated their caches one key at
a time. Both of those scale with the customer count, so a large enough
deployment can OOM the proxy pod on the read, and the tail of the
population sits on a stale spend counter while the per-key invalidations
drain
PR #40639 moved the reset write itself to a link-based UPDATE, so that
pre-commit read no longer feeds the write. It only fed cache invalidation
and the service-logging counts, which means it can move after the commit.
This replaces it with a keyset walk over litellm_endusertable ordered by
user_id, taking RESET_BUDGET_JOB_BATCH_SIZE rows per page, the same shape
_reset_windows_for_source already uses, with no per-run page cap for the
same reason that walk has none: the cursor cannot survive the run, so a
cap would restart at the first customer on every tick and never reach the
tail
Each page's counter and cache keys now go out as one batched delete
through a new DualCache.async_delete_cache_keys, which drops the
in-memory entries and chunks the Redis DELETE at
DEFAULT_MAX_REDIS_BATCH_CACHE_SIZE
num_endusers_found and num_endusers_updated now report the customers
whose caches were invalidated after the commit rather than the rows read
before it, so both read 0 when the cascade write fails
The reset job moved from zeroing spend to an atomic decrement of the amount
it cleared, so every batched write now carries {"decrement": <cleared>}
instead of 0. Four tests still pinned 0 and had been failing since, which
also meant they no longer checked the amount at all. Assert the decrement
equals each row's own pre-reset spend, so a wrong amount fails the test.
The cascade zeroed end-user spend with a single update_many whose where
clause enumerated every dependent user id. Prisma compiles that IN-list
into one prepared statement carrying one bind variable per customer, and
PostgreSQL caps a statement at 32,767 of them. Once a shared budget had
more dependents than that the statement could not be parsed at all, so
the atomic cascade rolled back, budget_reset_at never advanced, and the
tier stayed due on every later tick forever. Customers sitting at their
cap were blocked indefinitely with only a recurring log line to show for
it.
End users now match on budget_id like every other gated table, plus a
NULL-budget_id branch for the implicitly created rows that carry no link
and ride the default tier. The statement's bind count now tracks the
number of expiring tiers rather than the customer population, so a reset
costs the same whether a budget has ten dependents or a million.
Fixes#40564
Claude-Session: https://claude.ai/code/session_01Hn5E8Jz1LjGLFyiYxBRcBW
* test: drop the cwd-relative sys.path.insert calls from the test suite
TQ003 stands at 1,077 across 1,058 files, and 1,015 of them are the same shape:
sys.path.insert(0, os.path.abspath("../..")) and its deeper siblings. The
argument resolves against the working directory rather than the file, so from
the repo root, where every job runs pytest, it inserts the directory two levels
above the checkout. It has never pointed at litellm. The package is installed
into the environment anyway, which is what actually makes the import work, and
what the rule's message has said all along.
Removing them leaves 1,634 imports of sys and os with no remaining reference,
and those go too, except where another test module imports the name back out of
the file. The rest of TQ003 is 62 call sites that resolve against __file__ or a
variable, which are a different question and are left alone.
Collection is identical either way: 45,871 tests and the same 51 pre-existing
collection errors before and after, and ruff reports no new undefined name.
* test: drop the duplicate imports the sys.path sweep exposed to F811
* test(pre-call-utils): restore the os import the new bedrock tests need
A test that asserts on the error inside its own except block passes when the
call stops raising, because nothing runs the handler. That is the exact case
the test exists to catch, so the regression lands green.
Rewrites all 111 such blocks into pytest.raises, which fails when the call
succeeds, and selects PT017 in ruff-tests.toml so no new one lands.
A name bound twice keeps only the second binding. In `tests/` that is nearly
always a repeated import, harmless but misleading, and the same rule is what
catches the cases that are not harmless: a local that shadows an import the
module still calls, and a second `def test_x` that quietly replaces the first.
311 of the 344 sites were repeated imports and came out with ruff's own fix.
The remaining 33 needed a decision. Four modules imported a name they never
used because a local definition below already shadowed it. Two comprehensions
bound `call` over `unittest.mock.call`, which those modules import and use.
One test rebound the two module handles its nested reload closure had captured.
One class attribute shadowed an unused `status` import.
The load-test fixtures move to a conftest, which is how pytest is meant to share
them, so the test module no longer imports three fixture names it never calls.
The nine `prisma_client` parameters keep a narrow `noqa`: pytest resolves that
fixture by name before the body runs, so the parameter never shadows anything.
`pytest.raises(Exception)` with no `match=` passes on any error that broad. A
TypeError from a refactor, a botched fixture, an import that moved: all of them
read as the rejection the test claims to police, so the test goes green for the
wrong reason and stays green after the behaviour it guards is gone.
PT011 closes that gap for the 317 sites B017 could not reach, because B017 only
fires on a single-statement body with no `as e` binding. Each pattern here is the
message the code actually raised, recorded by running the sites under a plugin
that logged the concrete type and text per call site, so the assertions describe
observed behaviour rather than a guess. Where a site raises more than one message
across its parametrize cases, the pattern is an alternation of what was seen;
where the exception carries an empty `str()` and puts the text on `.message`, the
site keeps a narrow `noqa` with the reason.
PT014 removes four parametrize cases that were listed twice. The duplicate re-runs
an assertion that already passed, and it usually marks a case someone meant to
vary and forgot to edit.
* test(lint): ban blind pytest.raises(Exception) with ruff B017
A bare pytest.raises(Exception) accepts whatever the body throws. The TypeError
a refactor introduces satisfies it exactly as well as the rejection the test was
written for, so the crash reads as a pass and the test never goes red.
All 111 existing sites are narrowed here. A runtime probe recorded the concrete
exception each one actually catches, and each site now names that type. Where
the code under test genuinely raises a bare Exception, the site pins a stable
slice of the message with match= instead.
Two sites tell on themselves. The shared responses-API cancel test raises
"custom_llm_provider is required but passed as None" rather than talking to a
provider at all, because cancel_responses takes a provider, not a model. And
test_bedrock_guardrails_with_streaming was the only test in its file still
passing without AWS credentials, because the NoCredentialsError boto3 raised
long before the guardrail ran satisfied the blind raises.
* fix(test): widen the openai batch-dispatch assertion to OpenAIError
The narrowed NotFoundError only holds where OPENAI_API_KEY is set. Without one
the SDK raises OpenAIError while building the client, long before any 404, so CI
went red. OpenAIError covers both and still rejects a TypeError from a refactor.
Every pod schedules the budget reset job, so a fleet re-read the whole due
population and wrote it back against one Postgres at the same calendar
boundary, multiplying a single sweep by its replica count. The job now takes
the shared PodLockManager lease, so one pod sweeps per tick. A deployment with
no Redis keeps its previous behavior, and a Redis that cannot answer sweeps
unguarded rather than stranding every expired budget at its cap.
The per-window scan read every row carrying budget_limits in one statement, so
its cost grew with the deployment's key count. It is now keyset-paginated and
walks to the end of the table on every sweep. A per-run cap would need a resume
position, and no pod can hold one because the lease rotates between ticks, so
the strictly advancing cursor is what terminates the walk.
Found and updated rows were also JSON-serialized into the service hook's
metadata and into debug lines on every chunk, on the event loop, whether or not
anything consumed them. The hooks now carry counts, and the debug payload is
deferred until a record is actually emitted.
Resolves LIT-4793
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
`assert False` inside a `try:` raises AssertionError, which the `except
Exception` right below it catches, so several tests reported green no matter
what the code did. `pytest.fail` raises Failed, a BaseException, and escapes.
A bare `a == b` statement is evaluated and discarded. Nine of those sat in
tests, and one was comparing against a model name the router never produces.
Selects B011, B015, B018, PT015, PLR0133 and PLW0127 in ruff-tests.toml
alongside F821, with all 50 existing violations fixed, so no budget file or
ratchet is needed. CI already runs this config over tests/.
* fix(reset_budget_job): advance budget_reset_at atomically with the spend cascade
A postgres timeout mid-cascade previously left LiteLLM_BudgetTable rows
stamped for the next window while team member, enduser, org and tag spend
stayed at cap, so every later tick skipped them until the window rolled
over. All cascade writes and the budget_reset_at advance now share one
prisma batch transaction; a failed run persists nothing and the rows stay
due for the next ~10 minute tick. Cache and counter invalidation runs only
after commit, and the catch-all enduser log line now names the cascade.
* fix(reset_budget_job): elect one runner per tick and chunk the reset scans
Every pod and worker previously ran the reset job every ~10 minutes,
each fetching every expired row with no limit and writing one giant
transaction at the same calendar-aligned boundary; that concurrency is
what piled up postgres lock contention and timeouts. The job now takes
the shared PodLockManager redis lock (no redis keeps the old behavior),
and each phase walks its due rows in 500-row chunks, one transaction per
chunk, stopping when a chunk is short, makes no forward progress, or
hits the per-run cap; leftovers wait for the next tick.
* chore(lint): ratchet budget ceilings down for fixed violations
* fix(reset_budget_job): harden chunk loop, fail open on redis errors, heartbeat the lock
Review fixes on the two prior commits. Reset scans now skip rows with no
budget_duration, so permanently due rows can neither starve a phase nor
have a lifetime cap zeroed every tick. Chunk progress counts rows whose
new budget_reset_at actually cleared the cutoff, so a zero-length
duration cannot burn the per-run chunk cap. A failed lock acquire only
skips the run when another pod verifiably holds the lock; a broken redis
runs unguarded instead of silently disabling resets fleet-wide. Partial
row failures report real progress and fire the failure hook without
killing the phase. The leader re-asserts the lock between phases and
stops if another pod took over, and the budget window advance uses
update_many so a tier deleted mid-chunk cannot abort the transaction.
Lint budget ceilings re-ratcheted for the net-fixed violations.
* fix(reset_budget_job): renew the leader lease and reject non-positive budget durations
Bot review follow-ups. PodLockManager now extends the lock TTL when the
holding pod re-acquires, via an atomic compare-and-expire script with a
plain SET fallback, so a run longer than the TTL keeps its lease instead
of silently sharing the job with another pod. The positive-duration
validation that team member endpoints already had is hoisted to
management common_utils and applied to key, internal user, budget,
customer and team intake, so a tenant can no longer create zero-duration
budgets whose permanently due rows starve other tenants' resets. Such
durations now return 400 at intake; existing rows are untouched.
* refactor(reset_budget_job): defer leader election to a follow-up PR
* fix(reset_budget_job): satisfy strict lint gates
String defaults for the two getenv calls (PLW1508) and the chunk
outcome returns moved to try/else (TRY300).
Clients send tool_choice as {"type": "auto"} (Cursor on chat completions,
Claude Code's Anthropic tool_choice shape). validate_chat_completion_tool_choice
recognized that shape but returned it verbatim, and the chat -> Responses API
bridge only normalized {"type": "function"}, so the wrapper reached OpenAI and
the whole call failed with:
Invalid value: 'auto'. Supported values are: 'code_interpreter', ...,
'web_search_preview', ... (param: tool_choice.type)
That broke every tool call, web search included, on responses-mode models.
Unwrap {"type": "auto"|"none"|"required"} to the bare string at both layers:
the chat completions validation boundary where the shape is first accepted,
and the Responses API bridge that owns the Responses tool_choice contract.
No OpenAI surface accepts the object form for these values, so the previous
passthrough only deferred the 400 to the provider.
Budgets reset at midnight in the configured timezone with no way to control
the time of day, so a drained daily budget surfaces as an overnight incident.
Add a litellm_settings.budget_reset_time option (e.g. "12:00") that shifts
day/week/month resets to a configurable wall-clock time in the existing
timezone, so the end of the budget window lands during business hours.
The reset time is parsed once into an immutable BudgetResetSettings and
injected into the reset job (constructor) and computation, rather than read
from a module-level global at call time. A malformed value fails fast at
startup. Sub-day durations ignore the offset. Unset preserves midnight resets.
key_alias can become the secret name used by external secret manager
integrations (HashiCorp Vault, CyberArk Conjur) when store_virtual_keys is
enabled. Add raise_if_unsafe_secret_name, a shared validation check applied
unconditionally before a secret name reaches either integration or the
/key/generate, /key/update, and /key/regenerate API boundary, independent
of the existing enable_key_alias_format_validation opt-in flag.
Also hardens the Vault URL builder to percent-encode reserved characters
in secret_name (preserving "/" and "@"), and switches the Conjur policy
body to a real YAML serializer instead of raw string interpolation.
* feat(fireworks_ai): sync chat completions endpoint with full API surface
Add 23 missing request parameters to get_supported_openai_params():
seed, top_logprobs, min_p, typical_p, repetition_penalty,
mirostat_target, mirostat_lr, logit_bias, echo, echo_last, ignore_eos,
prompt_cache_key, prompt_cache_isolation_key, raw_output,
perf_metrics_in_response, return_token_ids, safe_tokenization,
service_tier, metadata, speculation, prediction, stream_options,
sampling_mask. Also add reasoning_history gated on supports_reasoning.
Fix prompt_truncate_length to prompt_truncate_len to match the actual
API parameter name. The old name was never in
DEFAULT_CHAT_COMPLETION_PARAM_VALUES, so it always went to extra_body
and was rejected by Fireworks; it never actually worked.
Normalize reasoning_effort boolean values to strings: True becomes
"medium", False becomes "none". The Fireworks OpenAPI schema
documents these as accepted types, but the server rejects non-string
values with HTTP 400 in practice. Integers pass through as-is since the
server is expected to validate them.
Auto-inject stream_options.include_usage=true when stream=true and the
user has not explicitly set stream_options. Without this, Fireworks
returns null usage in all streaming chunks, which is inconsistent with
the non-streaming behavior where usage is always present. If the user
explicitly sets include_usage=false, it is preserved.
Capture Fireworks-specific response fields in transform_response():
perf_metrics, prompt_token_ids, raw_output, and token_ids are now
extracted from the response and stored in response._hidden_params
(fireworks_perf_metrics, fireworks_prompt_token_ids,
fireworks_raw_outputs, fireworks_token_ids) so they are accessible to
logging, the proxy, and downstream consumers when the corresponding
request parameters are enabled.
Remove deprecated document inlining logic. Document inlining was
deprecated on 2025-06-30
(https://docs.fireworks.ai/updates/changelog#-document-inlining-deprecation).
This removes _add_transform_inline_image_block(), the file-to-image_url
migration in _transform_messages_helper(), and the
disable_add_transform_inline_image_block lookup. Current models that
support image input do so natively as VLMs. cache_control,
provider_specific_fields, and thinking_blocks stripping is retained.
Update get_provider_info() to look up supports_vision and
supports_pdf_input from the model cost map instead of hardcoding both
to True (which was based on the now-deprecated document inlining).
supports_prompt_caching remains True.
API docs: https://docs.fireworks.ai/api-reference/post-chatcompletions
Reasoning guide: https://docs.fireworks.ai/guides/reasoning
Prompt caching: https://docs.fireworks.ai/guides/prompt-caching
* fix fireworks chat api surface gaps
* Scope Fireworks thinking param to reasoning models
* style: fix black formatting
* fix(test): update minimax-m3 expected_vision to True
* test: cover non-dict content branch in transform_messages_helper
* fix(fireworks_ai): remove metadata from supported params to prevent internal metadata disclosure
* test(fireworks_ai): replace stale document-inlining capability test
The CircleCI-only litellm_utils_tests suite still asserted the old
behavior where document inlining made every Fireworks model report
supports_pdf_input and supports_vision as True. That premise was removed
in this change, so the test now reflects cost-map-driven capabilities:
unmapped models no longer advertise vision/PDF support while mapped VLMs
like minimax-m3 still do.
* test(fireworks_ai): add end-to-end regression for native OpenAI params
The existing coverage for the newly supported OpenAI-native params asserted
list membership in get_supported_openai_params or called map_openai_params
with a hand-built dict, both of which bypass the get_optional_params gate
(DEFAULT_CHAT_COMPLETION_PARAM_VALUES). That gate is what previously raised
UnsupportedParamsError for seed, top_logprobs, logit_bias, prompt_cache_key,
service_tier and prediction when drop_params=False. Assert the full path so a
revert of the supported-params additions fails the test instead of passing a
shallow membership check.
* test(fireworks_ai): fix test isolation in vision/inlining tests
Use monkeypatch in test_fireworks_ai_vision_capability_from_cost_map so the
LITELLM_LOCAL_MODEL_COST_MAP env var and litellm.model_cost are restored after
the test instead of leaking global state into the rest of the process.
Switch the document-inlining integration tests off deepseek-v3p1, whose
supports_vision is null in the cost map, onto minimax-m3 which is explicitly
supports_vision:true. The pass-through assertions no longer depend on a model
incidentally not being marked non-vision.
* fix(fireworks_ai): gate image rejection on exact vision capability
The image_url rejection read supports_vision via _get_model_cost_capability,
which falls back to hyphen-boundary substring matching when no exact cost-map
entry exists. A custom or fine-tuned model id that merely contains a known
non-vision model's short name (e.g. an id ending in -glm-5p2) inherited that
entry's supports_vision:false and hard-failed valid image_url blocks on a
vision-capable deployment.
Split the exact candidate-key lookup into _get_model_cost_capability_exact and
use it for the hard rejection so a fuzzy match can never block images; the
substring fallback stays a soft signal for capability reporting. Also rewrites
the fallback as a comprehension + max instead of an accumulating loop.
* feat(fireworks_ai): surface response fields on streaming responses
The Fireworks-specific response fields (perf_metrics, prompt_token_ids,
per-choice raw_output and token_ids) were only captured into _hidden_params in
transform_response, which runs for non-streaming completions; streaming chat
went through the default OpenAI chunk handler and dropped them.
Add a FireworksAIChatCompletionStreamingHandler that the provider now returns
from get_model_response_iterator. It reuses one extraction helper with
transform_response and attaches the fields to each streamed chunk's
provider_specific_fields, which is the channel litellm preserves when it
rebuilds streamed chunks (per-chunk _hidden_params is not carried through).
Per-choice token_ids/raw_output ride the content chunks; response-level
perf_metrics/prompt_token_ids ride the final usage chunk. Covered by an
end-to-end streaming test through litellm.completion(stream=True).
---------
Co-authored-by: Ahmad Shahzad <ahmad@shahzad.dev>
Co-authored-by: Graham Neubig <398875+neubig@users.noreply.github.com>
* fix(guardrails): stop re-initializing DB guardrails on every poll
InMemoryGuardrailHandler._has_guardrail_params_changed compared the
in-memory LitellmParams against the raw dict loaded from the DB. The
in-memory side carries every field default and coerces enums via
model_dump(), while the DB side only holds the keys originally stored,
so the two shapes never compared equal and the guardrail was rebuilt on
every poll cycle.
Each rebuild created a fresh instance, but delete_in_memory_guardrail
only removed the old callback from litellm.callbacks. Request handling
promotes guardrail callbacks into the success/failure/async lists, so
the previous instance stayed referenced there and instances accumulated.
Normalize both sides through LitellmParams(...).model_dump() before
diffing, and purge the callback from every callback list on delete.
* refactor(guardrails): narrow params-normalization fallback to ValidationError
The comparison normalizer caught a bare Exception and silently fell back
to the raw dict, which hid the cause and quietly degraded the affected
guardrail back to re-initializing on every poll. Catch only the
ValidationError that LitellmParams construction can raise, log a warning
so the offending row is diagnosable, and let any other error surface
instead of being swallowed.
* refactor(callbacks): add remove_callback_from_all_lists helper to manager
Move the knowledge of which callback lists a callback can be promoted
into out of the guardrail registry and into LoggingCallbackManager, where
the rest of the callback-list bookkeeping already lives. delete_in_memory_guardrail
now delegates to the new helper instead of iterating the lists itself.
* Fix remaining VCR live-call leaks
* test(vcr): dedupe live-test helpers and drop spurious kwargs
Extract the duplicated isVertexQuotaError/runVertexRequestOrSkip Vertex
quota-skip helpers into tests/pass_through_tests/vertex_test_helpers.js and the
duplicated _skip_live_prompt_caching_test guard into tests/_live_test_helpers.py
so each lives in one place. In test_aarun_thread_litellm, build a separate
message_data carrying role/content for add_message and a thread_data without
them for run_thread/run_thread_stream/get_messages, which no longer receive the
spurious message fields.
* test(overhead): assert mock transport is exercised in non-streaming and stream tests
* fix(reset_budget): write only {spend, budget_reset_at} and stop pre-zeroing counter
ResetBudgetJob's batched update_data path shipped the full key/user/team
model on each reset. Prisma rejects object_permission_id and budget_limits
on the update input type, so any row carrying those fields detonated the
entire batch -- spend never reset, budget_reset_at never advanced. After
v1.84.0 started populating object_permission_id on UI-created keys, this
fires routinely.
_reset_budget_common also zeroed the cross-pod spend counter before the
DB write, so failed resets left enforcement reading 0 from the counter
while the DB still held the over-budget spend, admitting requests past
the cap until the counter naturally re-saturated from new reservations.
Switch the write to per-row narrow updates ({spend, budget_reset_at})
via db.batch_, and move the counter invalidation out of
_reset_budget_common so it only fires after the DB write commits. On
DB-write failure the counter is left untouched, enforcement continues
to block, and the next scheduler tick can retry without leaving a
bypass window.
Fixes#27730.
* fix(reset_budget): address Greptile review on #29358
- Strengthen the bypass-half regression test: replace the for-loop over
call_args_list (vacuously true when empty) with assert_not_called(),
so the test would actually flag a re-introduction of counter-zeroing
via any code path.
- Add the same explanatory docstring on _write_user_reset_updates and
_write_team_reset_updates that _write_key_reset_updates already has,
so all three helpers point future maintainers at #27730.
* test(reset_budget): update test_proxy_budget_reset for new batch-write path
Same shape as the previous test_reset_budget_job.py update: keys/users/teams
now write through prisma.db.batch_().<table>.update, not update_data, so the
tests need a batcher mock and updated assertions. Adds:
- _wire_batcher_for_test helper that returns a list which accumulates per-row
batch updates captured from prisma_client.db.batch_().
- _attrify helper that wraps dict fixtures so getattr(item, "token") works
alongside the dict item-access the fake_reset_* mocks rely on. The new
narrow-write helpers use getattr to pull out the row's id, and would
silently skip plain dicts otherwise.
- Updates 3 partial_failure tests to assert against the batch-call list
(rows by id, payload contains only {spend, budget_reset_at}) instead of
update_data.assert_awaited_once + data_list inspection.
- Updates test_reset_budget_continues_other_categories_on_failure: only
budget + enduser still flow through update_data; key/user/team go through
the batch path now.
- Wires the batcher mock into 3 service_logger_*_success tests so commit()
is actually awaitable and the success hook fires.
These tests were silently passing locally only because the editable install
in .venv pointed at the main repo, not the worktree — running pytest with
PYTHONPATH overridden to the worktree (matching CI) reproduces the failures.
This reverts the Bedrock CI account migration (#28728). The original account
(888602223428) was put under an AWS security restriction after a leaked key
and has since been reactivated, while the replacement account (941277531214)
lacks access to several models the suites exercise (legacy Bedrock Claude 3
models, Cohere, Nova Canvas image gen, Bedrock batch inference, and flagship
Opus). Pointing CI back at the reactivated account restores that coverage.
This is the exact inverse of #28728: all hardcoded 941277531214 references go
back to 888602223428 (provisioned/imported-model ARNs, AgentCore runtime ARNs
and their suffixes, batch execution role ARN, and the example proxy config),
the S3 buckets revert to litellm-proxy and load-testing-oct, the guardrail IDs
revert to wf0hkdb5x07f and ff6ujrregl1q, the SageMaker endpoint and Knowledge
Base revert to their original ids, and the live-call tests go back to the
legacy model strings. The grid_spec fail_reason workaround for the unentitled
Opus cells is dropped while keeping the unrelated bedrock_effort_ceiling field
added after the migration.
The CircleCI AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY env vars still point at
941277531214 and must be set to the reactivated account's fresh credentials
separately via the CircleCI API; AWS_REGION_NAME stays us-west-2.
* test(vcr): make Redis-backed cassettes replay deterministically across runs
- Pin LITELLM_LOCAL_MODEL_COST_MAP=True in the shared VCR harness so the
per-test importlib.reload(litellm) no longer fetches the model cost map
from raw.githubusercontent.com. That live fetch was being recorded into
cassettes; for tests that subsequently skip it was the only recorded
episode, so the persister refused to save it (skipped tests don't persist)
and the test re-recorded it live every run (MISS:NOT_PERSISTED).
- Compare-time symmetric matcher tolerance for Google OAuth (ya29.*) tokens,
observability/telemetry payloads, credential-exchange bodies, and volatile
UUID/timestamp tokens, so existing cassettes select a recorded episode
instead of growing past the 50-episode cap and re-recording live.
- Don't record fire-and-forget telemetry (langfuse/arize/otel/...) into
non-telemetry tests' cassettes. Several modules set litellm.success_callback
at import time, so observability logging is globally enabled and an async
flush from the background logging worker lands in an unrelated test's VCR
window, saved as a spurious MISS:RECORDED (observed: a Langfuse batch from
another completion landing on test_lowest_latency_routing_buffer). Such a
request now passes through live (telemetry hosts aren't real-spend hosts);
tests that actually assert on telemetry keep recording it.
- Dedupe + cap the VCR diagnostic dump so the classification summary survives
CircleCI's ~400KB step-output truncation.
- Stabilize a non-deterministic rate-limit test body; mark AWS Secrets Manager
lifecycle tests VCR-incompatible (uniquely-named secrets can't be replayed).
- Mark test_router_text_completion_client VCR-incompatible: it fires 300
identical requests to verify async-client reuse, but vcrpy patches the HTTP
transport so replay never exercises the real connection pool the test
validates, and recording 300 near-identical episodes overflows the
50-episode cap (MISS:OVERFLOW every run). It hits a free mock endpoint.
- Mark the Vertex AI MaaS Mistral OCR tests (vertex_ai/mistral-ocr-2505)
VCR-incompatible: the MaaS model is not provisioned in the CI GCP project,
so the live :rawPredict call fails and the test skips every run, leaving no
cassette to record (MISS:NOT_PERSISTED every run). Sibling direct-Mistral
and Azure OCR tests are unaffected and still replay from cache.
* fix(tests/vcr): refresh cassette TTL on read so replayed cassettes don't expire
The Redis VCR persister loaded cassettes with a plain GET, which does not
touch the key's TTL. A cassette that is only ever replayed (HIT/NOOP, never
re-recorded) therefore expired exactly 24h after its last *write*, no matter
how often it was read. Whichever CI run happened to cross that boundary
re-recorded the cassette live and surfaced a spurious VCR MISS on otherwise
deterministic cassettes — the residual per-run flakiness floor (a different
random subset of read-only cassettes expiring each run).
Slide the expiry forward on every successful load (best-effort EXPIRE), so
any cassette used at least once per TTL window stays alive indefinitely and
the 2nd/3rd run of a day replays cleanly.
* fix(tests/vcr): recover from spurious GET-None for existing cassette keys
Under concurrent CI load, the persister's load GET was observed returning
None for a cassette key that demonstrably existed on the (single, non-
clustered) Redis master — an external monitor saw the key present with a
healthy TTL at the same instant the in-process client read None. Because
None is a valid GET result (not a RedisError), the retry-on-error client
config never engaged, so the cassette re-recorded live (a phantom
MISS:RECORDED); for flaky/networked tests the failed live call then
triggered a pytest rerun, which is why a rotating subset of otherwise
deterministic tests missed each run.
On a None result, re-check EXISTS and re-read once. If the key really
exists, use the recovered value and log [vcr-transient-miss-recovered]
(also counted in cassette_cache_health). A genuinely absent key (a new
cassette) still falls through to CassetteNotFoundError.
* chore(tests/vcr): TEMP diagnostic for persistent-miss cassette load path
Logs GET/EXISTS at load time for the three cassettes that re-record every
run despite being present in Redis, to capture what the in-process client
sees. To be reverted before merge.
* chore(tests/vcr): write load diagnostic to Redis (truncation-proof)
CI stdout truncates to the last ~400KB, dropping the early loaddbg lines
for the alphabetically-first failing test. Push the load probe to a Redis
list instead so it survives. To be reverted before merge.
* fix(tests/vcr): don't drop stored telemetry episodes during cassette load
Root cause of the residual per-run misses on present cassettes: vcrpy's
Cassette._load() replays each *stored* interaction through Cassette.append(),
which runs before_record_request on it — and a None return there silently
drops that episode. The telemetry-leak suppressor (_should_drop_telemetry_record)
returns None for telemetry requests, so when a non-telemetry-named test (or the
alphabetically-first test in a worker, whose _current_test_nodeid is still empty)
loaded a cassette containing a Langfuse ingestion episode, the episode was
dropped on read — forcing an endless live re-record (a phantom MISS:RECORDED on
a cassette that was demonstrably present in Redis). Verified by reproducing
Cassette._load() against the real cassette: empty/non-telemetry nodeid -> 0
episodes survive; with the guard -> 1 survives.
Fix: guard the suppressor with a thread-local set around Cassette._load (via a
small idempotent monkeypatch), so the drop only ever stops *new* incidental
telemetry from being recorded and never filters the existing cassette on read.
Also drops the speculative GET-None recovery + its diagnostics from the previous
commits: the load diagnostic showed GET returns the cassette bytes fine
(get=1440B), so the persister never returned a spurious None — the loss happened
later in vcrpy's append. The proven TTL-refresh-on-read fix is retained.
* fix(tests/vcr): drop incidental telemetry export POSTs to stop rotating async-flush misses
litellm's observability loggers flush on a background thread, so a Langfuse
ingestion POST scheduled by one telemetry test can fire mid-way through a
*later* telemetry-named test (after that test's own httpx mock has exited) and
be recorded by VCR as a phantom episode — a non-deterministic MISS:RECORDED /
PARTIAL that rotates onto a different telemetry test from run to run.
Telemetry export POSTs are fire-and-forget; no test asserts on a *recorded*
export response except the pass-through proxy test (which forwards a client POST
to Langfuse ingestion and replays its 207). So _should_drop_telemetry_record now
drops incidental export POSTs for every test except that one. Dropping returns
None (live fire-and-forget, never stored), so it can only turn a phantom miss
into a harmless live call, never the reverse; recorded read-back GETs that
telemetry tests assert on are matched by method and left untouched.
* fix(tests/vcr): restore assertion in test_banner_silent_when_vcr_disabled
The assertion that the banner is suppressed when VCR is disabled was
inadvertently moved into test_diagnostic_log_silent_when_no_dir when
the diagnostic-log tests were added, leaving the disabled-VCR test
verifying nothing.
Co-authored-by: Yassin Kortam <yassin@berri.ai>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Yassin Kortam <yassin@berri.ai>
