Resolves the six conflicts main accumulated against the SDK migration.
endpoints_client.py stays deleted (main only added the safety_identifier
kwarg to it), and its two new importers move onto the SDK clients: the
Bedrock safety_identifier forwarding test in test_responses_e2e.py now
sends its /v1/responses leg through the OpenAI SDK and its chat leg
through ProxyClient.chat, and test_messages_azure_foundry_e2e.py drops
the x-litellm-rust header check together with EXPECT_RUST, which main
removed from e2e_config in a84f68b6e3. proxy_client.py keeps both the
rerank and router-settings model imports. uv.lock is main's lock plus
the anthropic e2e-dev entry, and the pricing JSONs are taken verbatim
from main since this PR never meant to change them.
Rotating the Speech-to-Text stream at 240 s no longer ends the active turn:
the turn and its billed seconds continue on the new stream, forced at 280 s.
Bound the request and event queues (64 and 256) so a slow peer applies
backpressure instead of growing memory. Route a model to the Chirp realtime
path from its cost-map entry (mode audio_transcription plus /v1/realtime)
instead of a hardcoded name. Return on every branch of the recv and
transform helpers (CodeQL mixed returns), have the shared protocol helper take
the provider's error class so the Meta tests assert MuseProtocolError again,
and pin google-cloud-speech in the ci group so unit shards import it.
Replace the bespoke dependency-install CI gate with a real migration:
require mcp>=2.2.0,<3 alongside httpx2>=2.5.0,<3 and pydantic>=2.12.0,<3
in the proxy and mcp extras, drop langchain-mcp-adapters (pins mcp<2)
from the dev group, and remove the dependency-install workflow and
tests/mcp_dependency_tests that only exercised the old pins.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Bridge OpenAI Realtime transcription sessions on vertex_ai/chirp_* models to
Google Speech-to-Text v2 StreamingRecognize over gRPC, so partial and final
transcripts stream back while audio is still being sent. Interim results become
delta events, finals become completed events carrying billed seconds, the gRPC
stream rotates at 240 s under Google's five-minute cap with billed time chained
across rotations, and audio is split into 25 KB requests.
The OpenAI transcription protocol helpers move into a shared module that Meta
Muse now uses too, google-cloud-speech ships behind a new stt-vertex-chirp extra
bundled into the proxy runtime, and the cost map lists /v1/realtime for chirp_3.
The bedrock-realtime extra pinned aws-sdk-bedrock-runtime 0.7.x, whose Config and BedrockRuntimeClient surface is gone in 0.11. The handler now resolves AsyncBedrockRuntimeConfig, builds AsyncBedrockRuntimeClient with the awscrt duplex transport, closes the client when the session ends, and tells an absent SDK apart from an installed but unsupported version. Moves the pin to >=0.10.0,<0.12.0 with the awscrt extra
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Re-migrates the e2e tests that main extended through endpoints_client
since this branch was opened (Azure Foundry, mid-conversation system
messages, Bedrock web search, google native streaming) onto the
provider SDK clients, so no endpoints_client reference remains
The sequential version sent one request at a time, so a Redis outage never
reached the concurrency where the failed-tracking alert body actually grows.
This drives the proxy with locust against one model group of three mock
deployments, two failing at order 1 and one serving at order 2, so every
request spends its retries on the failing pair and lands on the serving
deployment through the order-based fallback. Two phases, a healthy baseline
and a CLIENT PAUSE WRITE window, and every request must succeed in both.
Latency, RSS and CPU are reported as p50/p90/p99 per phase rather than
asserted on: RSS and CPU come from psutil on the proxy's process tree, since
a multi-worker proxy serves /metrics from the prometheus multiprocess
collector and that drops the process collector's series. Thresholds stay open
until weekly runs give real baselines.
Co-Authored-By: Claude Code <noreply@anthropic.com>
4.17 was picked on the belief that dnspython only became a core pymongo
dependency there, which is wrong: pymongo has declared dnspython>=1.16.0,<3.0.0
as a core requirement since well before that, so mongodb+srv:// URIs resolve at
4.9 too. The real floor is 4.9, the release AsyncMongoClient landed in, and 4.8
has no AsyncMongoClient at all.
Verified against live Atlas on 4.9: sync and async search, list_search_indexes,
same top hit and score as 4.17. Resolution is unchanged, pymongo 4.17.0 either
way, so this only widens what an existing environment is allowed to bring.
GHSA-jp53-mhqp-8xcg (fixed in 6.16.0), GHSA-23w6-3w8w-8484 and
GHSA-763m-79hh-57f2 (fixed in 6.16.1) flag pypdf 6.15.0 in uv.lock and
keep osv-scan red alongside the tornado advisories. The proxy-runtime
extra now requires pypdf>=6.16.1 and the lock resolves 6.16.2.
Atlas Vector Search has no HTTP query API, since the Data API and HTTPS
Endpoints are end-of-life, so this provider extends BaseDirectVectorStoreConfig
and runs the $vectorSearch aggregation through pymongo rather than shaping an
httpx request. That is the same seam Valkey uses for RESP.
vector_store_id names the Atlas Search index, matching Valkey, with the
database and collection supplied through litellm_params.
pymongo lives in a new optional `mongodb` extra and is imported lazily, so the
base install still pulls no MongoDB driver. The floor is 4.17 because that is
where dnspython became a core dependency instead of the `srv` extra, and Atlas
issues mongodb+srv:// URIs that will not resolve without it.
Clients are cached per connection rather than opened per search. Measured
against Atlas, a fresh client costs ~890ms versus ~80ms warm, so copying the
Valkey open-and-close-per-call pattern would have added ~810ms to every query.
ProxyExtrasDBManager.spend_logs_is_partitioned() (#38452) silently returns
False when psycopg can't be imported, and psycopg was never added to the
extra_proxy install, so every production image lacks it. Schema
reconciliation then generates the unfiltered primary-key rewrite against a
genuinely partitioned LiteLLM_SpendLogs and Postgres rejects it, exactly the
failure the fix was meant to prevent. Ships psycopg via extra_proxy and logs
a warning when it's still missing instead of failing silently.
RestrictedPython 8.3 extended its protected-name validation to cover
positional-only parameters, so sandboxed source can no longer bind a
local named _getattr_, _getitem_, _write_ or _print_ that takes
precedence over the hooks the custom-code guardrail sandbox installs.
8.4 and 8.5 continue in the same direction with safer_getattr and the
Python 3.15 syntax audit.
The floor moves rather than the lock alone so downstream installs of
litellm[proxy] pick up the same behaviour.
The MCP client's HTTP transport needs mcp>=1.24.0 for streamable_http_client,
but a base litellm install declares no mcp constraint and no extra existed to
pin one, so environments carrying an older mcp fail at connect time with
'streamable_http_client is not available. Please install mcp with HTTP
support.', which names no version floor and no installable remedy.
Add a litellm[mcp] extra matching the proxy extra's mcp>=1.28.1,<2.0 and
replace the vague ImportError with one naming the required floor, the
installed mcp version, and the pip commands that fix it.
