* feat(gemini): continue MAX_TOKENS decodes with continuationToken across chat, streaming, and interactions
* fix(gemini): continue long decode on CONTINUATION and cancel interactions at /cancel
* feat(gemini): add public gemini-4-argon cost-map rows for AI Studio and Vertex AI
Add gemini/gemini-4-argon and vertex_ai/gemini-4-argon alongside the existing
early-access row, carrying the same limits and capabilities. The AI Studio row
mirrors the early-access entry; the Vertex row differs only in provider and
pricing source. Prices follow once Google publishes them.
Extend the mapped metadata test to cover all three ids, that each routes to its
provider, that its backup copy matches, and that the public rows carry the same
limits as the early-access row.
* feat(gemini): raise gemini-4-argon to the 1M long decode cap and report initial_prompt_tokens across hops
* chore(prices): build on the merged gemini-4-argon rows and only raise the long decode cap
* fix(gemini): follow continuation hops in a loop instead of recursion
The unary hop collectors and the stream iterator called themselves once per continuation request, which the recursion check in CI rejects and which would hit the interpreter's call depth on a very long output. They now loop, and a test follows more hops than the interpreter allows nested calls
* fix(gemini): bill Google Search grounding from every continuation request
Google sends search grounding once, on the request that ran the search, and does not repeat it on a continuation request. The merged response kept only the last request's grounding, so usage.web_search_requests and the per-search fee dropped every earlier search. Unary merges and the stream iterator now carry the grounding of every request in the chain
* fix(gemini): continue long gemini-4-argon answers on streams and Vertex AI
A continued stream stopped after the first chunk of its second request. The stream wrapper asks the iterator for a fresh start before every chunk, and that reset it to the first request's finished stream. The iterator now reads from the newest request's stream
On Vertex AI the continuation request went to the v1 API version, which rejects continuationToken with a 400. A request that carries the token now goes to v1beta1, which accepts it, including a token that v1 issued
On a stream, initial_prompt_tokens came from a usage chunk in the middle of the stream, which reports a smaller prompt count than the last chunk. It is now the count the first request's stream ends with, the same number a unary call reports
continuation_token is no longer forwarded on the Interactions API, which answers 400 for the field. It is dropped there as it was before this branch
* chore(gemini): name the type errors the dual-mode stream iterator trips
The stream iterator holds a sync or an async line iterator in one untyped attribute. Pointing it at a continuation request's stream gave the type checker a concrete union for it, so six existing lines that only ever run in one of the two modes started to report errors
This adds a rule-named suppression with a reason on each of those lines and changes no code. Splitting the class into a sync and an async iterator would remove them and is left for a follow-up
* fix(gemini): continue only when Google ends a request with CONTINUATION, and stop streaming mid-answer tokens
A long decode now sends the next request only when Google's finish reason is CONTINUATION, so an answer cut at the caller's own max_tokens returns in one request with its continuation_token instead of drawing a rejected extra request
A request for more than one choice is never continued, since Google rejects a continuation with more than one candidate
Chat-format streams no longer carry the tokens Google sends mid-answer. Only a stream that ends cut off carries a continuation_token, on its last content chunk
The response cache key now includes continuation_token, so a resend with a token is not served the earlier segment
* fix(gemini): keep one response id across a continued answer
A continued chat stream changed its chunk id each time the next request to Google started, and a unary answer came back under the last request's id. Every chunk and the unary answer now carry the first request's id, including the chunk that ends a cut-off stream
A request whose n or max_tokens arrives as a string is sent once and never continued, where it used to fail with a 500 on the comparison. An empty continuation_token no longer changes the cache key
* test(gemini): add integration cells for long decode continuation
Covers continued answers on chat, /v1/messages and /v1/responses, streamed and not, on the Gemini API and Vertex AI against a scripted Google, plus concurrency, outage and worker restart runs
Also moves the Google stand-in cancel route to /cancel, the URL this PR now calls for Interactions cancel, and checks the cancel test posts there exactly once
* fix(vertex_ai): fetch the access token again before each continuation request
A long Gemini 4 answer on Vertex AI can run for over half an hour of follow-up requests, and every one of them reused the bearer token fetched before the first request. Once that token expired Google answered 401 and the caller got a cut-off answer with finish_reason length
Each Google request in a chain now asks for the token again, which is a cache hit until the token nears expiry, on the streamed and unary paths, sync and async. A refresh that fails counts as a failed follow-up, so the caller still gets the answer so far with its continuation token. A caller's own Authorization in extra_headers still wins on every request, and the Gemini API key path is unchanged
* fix(vertex_ai): start each continuation stream with an empty parser buffer
When a Google stream ended on an unfinished frame, its leftover text stayed in the stream parser and every frame of the next continuation stream was appended to it. None of them could be decoded, so the whole continuation was billed and dropped without an error
Each continuation stream now starts with a fresh buffer, and a test covers a stream that ends on an unfinished frame, sync and async
* fix(vertex_ai): treat a continuation stream with no readable frame as a failed follow-up
A continuation stream that answered 200 and then sent Google's error frame first, sent a body that is not JSON, or dropped before its first frame ended the caller's stream with `stop` or an error and no token, so a cut answer looked complete or was lost. It now ends the way a refused follow-up does, with the answer so far, `finish_reason` length and that request's token
Every continuation stream request now also gets the caller's per-request timeout, which #46349 added to the first one, so a follow-up that never answers ends at that timeout the same way
* fix(vertex_ai): keep every follow-up's citations and URL context in a continued answer
A non-streamed answer that took several Google requests kept citationMetadata and urlContextMetadata from the last request only, and left every later request's offsets counting from that request's own text. So a URL read on the first request was missing from the answer, citations from earlier requests were dropped, and a later request's citations and search supports pointed at the wrong part of the joined answer
The joined answer now keeps every request's citations (Vertex AI's `citations` and AI Studio's `citationSources`) and URL context, and shifts each later request's citation and search support offsets by the UTF-8 bytes of the answer text sent before it, thought parts excluded, since Google counts these offsets in bytes
* fix(vertex_ai): answer when a candidate sends null citationMetadata or urlContextMetadata
A Gemini candidate whose citationMetadata or urlContextMetadata was JSON null made the call fail on every Gemini model, non-streamed with a 422 or 500 and streamed with an error event, because the per-request metadata helper handed the null back as is and the caller tried to extend a list with it. A continued answer also failed when a follow-up's citation or search support came with a null startIndex
A null field is now kept as one entry on a single request, as it was before this branch, a follow-up's null URL context adds nothing to a continued answer, and a null startIndex counts from the start of that request's text. Tests cover missing, null and empty fields on one request and on a continued answer, streamed and not
* fix(vertex_ai): pass through spans with no end index and point a continued stream's citations into the whole answer
A later request's citation or search support with no endIndex made the whole non-streamed answer fail with a 500, so a span Google sends without an end index now comes back as Google sent it, the way one request for the whole answer returns it. A span with an end index but no start index still counts from 0
A continued stream handed callers each later request's citation and search support offsets counted from that request's own text. The stream now adds the answer bytes of every earlier request, so the offsets point into the whole answer like the non-streamed path and one request for the whole answer
Adds the audit's wire tests for both, which fail 24 cells before this change
* test(vertex_ai): cover a later request whose cited span starts at its first byte with no start index
Google leaves out a zero startIndex, so the second request of a continued answer can cite its own opening with only an endIndex. These wire cells check that span lands where one request for the whole answer puts it, on Gemini and Vertex AI, chat and Responses, unary and streamed
* fix(vertex_ai): log a continued answer's raw response as one object
A non-streamed answer joined from several Google requests logged its raw response as a JSON list of every request's body. The OpenTelemetry logger reads that raw response as one Gemini body to set its `llm.gemini.*` attributes, so on a continued answer it hit a validation error and set none of them
It now logs the merged body, which carries only the last request's token, and a single request still logs Google's body unchanged. The new unit test fails with the change reverted
Also adds integration cells for a cut chat stream's token chunk, an answer joined from 41 requests, a caller that gave up on a non-streamed call, counts and caps that are not numbers or are sent twice, `n` sent as true, a partial answer cached after a refused continuation, the token in another provider's cache key, and the sync SDK path with its Vertex token refetch
* test(integration): check a continued answer logs one merged raw response
A non-streamed Gemini answer that took several Google requests now hands logging one merged response, so the OpenTelemetry span carries the llm.gemini attributes. These 12 cells fail on the merge base and on the commit before that fix, and pass on the branch
* test(vertex_ai): cover a follow-up tool call and its thought signature in long decode
* test(gemini): pin that a one-request answer logs Google's body unchanged
* test(integration): assert the streamed usage.cost matches the spend row on every cost case
* test(integration): type the streamed usage frames as JsonValue
* test(integration): only require streamed usage.cost when the client received a usage frame
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): send detail on input_image parts bridged from /v1/messages to Responses
* fix(ui): send detail on playground Responses image parts and code snippets
* test(messages): cover image detail on the Responses bridge wire
* fix(oci): skip non-function tools and gate reasoning_effort on models that reject it
Fixes#31449.
Two OCI GENERIC request-building defects that made Grok-on-OCI unusable
for tool-using coding-agent clients (Codex over the Responses API, Claude
Code over the Anthropic Messages API):
1. `adapt_tool_definition_to_oci_standard` raised `OCIError("OCI only
supports function tools")` for any tool whose type is not "function",
aborting the whole request when a client mixed a built-in/server tool
in with its function tools. Non-function tools are now skipped with a
warning; if none remain, `tools`/`toolChoice` are omitted from the
request instead of sending an empty array.
2. The GENERIC param map forwarded `reasoning_effort` to every non-Cohere
model, but OCI rejects it for all served xAI Grok models
(`Model grok-4.20 does not support parameter reasoningEffort`) and for
OpenAI's non-reasoning models (`Unrecognized request argument supplied:
reasoning_effort` on gpt-4o / gpt-4.1); gpt-5, the o-series, gpt-oss-*
and Meta/Google models accept it. `_model_supports_reasoning_effort`
encodes that, `get_supported_openai_params` stops advertising the param
for those models, and `map_openai_params` follows the standard
unsupported-param contract (dropped under drop_params, clear error
otherwise) so `drop_params` / `additional_drop_params` now work as
documented instead of the 400 surfacing from the service.
Behaviour verified against OCI GenAI in us-chicago-1 (2026-09-11) per model.
The two existing reasoning_effort normalisation tests move from a Grok
model to openai.gpt-5, which is the case they exercise. 21 new tests; the
OCI suite is 406 passed.
* fix(oci): drive reasoning_effort gating from the model catalog and drop tool_choice for skipped tools
Review follow-ups on #40732:
- `_model_supports_reasoning_effort` now reads the `supports_reasoning` flag
of the model's `oci/` catalog entry through `get_model_info` instead of
hardcoding xAI / gpt-oss branches. Catalogued models without the flag
(xAI Grok, Meta Llama) no longer receive `reasoningEffort`; models absent
from the catalog keep the historical pass-through.
- Catalog: flag `oci/google.gemini-2.5-{flash,pro,flash-lite}` as reasoning
models and add `oci/openai.gpt-oss-120b` / `oci/openai.gpt-oss-20b`
(pricing from Oracle's public price list, 128K context per OCI docs).
- `_model_uses_max_completion_tokens` routes on the `openai.*` prefix only.
OCI-hosted Gemini accepts `maxCompletionTokens` but ignores it (the output
cap silently disappears) and Llama rejects it with 400, so the catalog's
reasoning flag must not drive that switch.
- After tool adaptation, a `FUNCTION` tool_choice naming a tool that was
skipped (built-in / server tools) is dropped with a warning instead of
being sent to OCI, which rejects the unmatched choice.
Verified live against OCI GenAI (us-chicago-1): Grok with mixed tools plus a
forced built-in now completes and still calls the function tool; Gemini and
gpt-oss forward reasoningEffort with maxTokens honoured; Llama drops it under
drop_params; gpt-5 keeps maxCompletionTokens.
* style(oci): use an immutable set for the tool-name lookup
Keeps the type-discipline gate (LIT002, mutable collections) within its
budget; the lookup is read-only.
* fix(oci): fail closed when a forced tool_choice names a tool OCI will not receive
Review follow-ups on #40732:
- When tool_choice forces a function that is not among the adapted tools
(a skipped built-in or a misspelled name), send `toolChoice: NONE` instead
of removing the choice. Falling back to the service default would let the
model call any other supplied function even though the caller forced a
specific one, which an agent processing untrusted content must not do.
- Move the #31449 regression tests into the mapped test module
(`test_oci_chat_transformation.py`) instead of a separate file.
Verified live against OCI GenAI: Grok and Llama with a forced skipped
built-in now complete with no tool call; a forced kept function is still
honoured.
* test(oci): point reasoning_effort comments at the in-module gating tests
The referenced module was folded into this file.
* fix(oci): drop non-function tools only under drop_params and gate reasoning_effort on declared levels
Non-function tools now follow the unsupported-param contract: a 400 naming
drop_params without it, and dropped with a warning under it, with a forced
dropped tool turning tool calls off. reasoning_effort is gated on the model's
reasoning_effort_levels catalog entry, so OCI Llama keeps forwarding it and the
Grok 4.20 models OCI serves, now in the catalog with their Oracle prices, stop
sending a param the service rejects
* fix(oci): gate reasoning_effort on oci/xai.grok-4.20, which rejects it
* fix(oci): disable tool calls when tool_choice forces a dropped custom tool by name
* fix(oci): drop tool_choice that forces a dropped tool and keep a kept function's name
* fix(oci): keep Gemini 2.5 Flash Lite unflagged for reasoning so released proxies keep its max_tokens cap
* fix(oci): answer 400 when a request depends on a dropped tool and leave Gemini max_tokens routing to its own PR
Under drop_params, non-function tools still drop when function tools remain and
tool_choice does not force a dropped tool. A request whose every tool is dropped,
or whose tool_choice forces one, now answers 400 instead of a tool-less reply the
model can fill with invented results, or a swap to another tool.
The maxTokens routing change for OCI Gemini goes back to base, since it is
outside this fix and an open PR carries it.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
GET /v1/mcp/server takes an optional user_id (admin, or the user themselves) and returns the union of the servers each of that user's teams reaches plus allow_all_keys servers. A user in no team gets the unscoped list, a deleted team still listed on the user is skipped, an unknown user is a 404 and a restricted virtual key is refused like for team_id. team_id keeps precedence when both are passed.
The dashboard MCPServerSelector passes the team and user it edits (key edit sends keyData.team_id and user_id, key create sends the picked team and owner, user edit sends the user) and narrows the server options to members of the access groups selected in the same control, keeping allow_all_keys servers and servers already granted so they can still be removed.
Resolves LIT-7784
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
fastapi_sso opens a bare httpx.AsyncClient for the non-PKCE token exchange and
userinfo call, so litellm_settings.ssl_verify and SSL_VERIFY never applied and
an IdP on an internal CA failed with an opaque 503. A LiteLLMHttpClientSSO mixin
replays that exchange on litellm's shared SSO client for the generic, Google and
Microsoft providers and surfaces IdP connection and TLS failures as a 401
ProxyException naming the endpoint.
Resolves LIT-9445
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(proxy): name the successor model for deprecating deployments
* chore(ui): add successor_model to the generated deprecation api types
* fix(proxy): keep the cost map successor when a deployment overrides only the date
* fix(proxy): read the cost map successor from metadata.successor
The cost map schema rejects a top-level successor_model key, and curated
entries already carry the successor under metadata.successor. Read that key,
and look it up even when the cost map entry has no deprecation date, so a
deployment that sets only its own date still names the successor.
* fix(proxy): keep the first deployment when deduping, let a blank successor hide the cost map one, and name a served successor by its model group
* fix(proxy): name only the deployment's own successor_model, since cost map successors are LiteLLM ids rather than served model names
* fix(proxy): leave out a successor_model the deployment's callers cannot call, with a warning
* fix(proxy): report successor_model as the admin set it and only warn when no model matches it
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(proxy): strictly honor disable_end_user_cost_tracking in SpendLogs
* fix(proxy): drop end-user id on failed requests and ClickHouse rows when end-user cost tracking is off
* fix(clickhouse): drop the end-user id from spend log metadata when end-user cost tracking is off
* test(spend): cover disable_end_user_cost_tracking across endpoints, failures, cache hits and chaos
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(anthropic): count a replayed server tool call's cache mark the way modify_params sends it
With modify_params on, a trailing tool search turn whose call has a kept result
got a placeholder answer from the factory and went out as a marked tool_use,
but the cache control hook counted marks before that placeholder existed and
skipped the mark, so a request at the cap reached Anthropic with 5 marks and
failed with a 400. The hook now finds client answers on the sanitized messages,
the same list the factory reads.
Also make the replay checks cheaper (prefix check before validation, early
return without provider_specific_fields, and look up client answers only when
a tool call carries cache_control), drop the two audit cells that cover old
gaps outside the tool search fix, and derive the pass-through test timeout
from the graceful stop budget.
* fix(anthropic): keep the duplicate tool result warning out of the cache control hook
The hook finds client answers with only the step of the modify_params sanitize
pass that pairs tool calls with their results, so a request with a duplicate
tool result logs the warning once again instead of once per mark count.
Also restore the Snowflake non-stream and cached-stream tool search cells as
known bugs with pytest.skip instead of deleting them, and put the Anthropic
pass-through test timeout back to 240 seconds.
* fix(anthropic): count cache marks next to a message the pairing step cannot read
With modify_params on and a marked tool call, the cache control hook paired tool
calls with their results before counting. A message without a get method (a
stray string or null in extra_body.messages, or an openai SDK
ChatCompletionMessage) made that pairing raise AttributeError before any
provider call, where main counted the marks and let the request through.
The hook now pairs only when every message can be read the way the pairing
step reads it, and otherwise counts on the messages as sent, like main did.
* fix(anthropic): count cache marks after placing a later system message the way the request builder does
* test(anthropic): cover a leading system prompt and a system message before the answer in the cache mark census
* test(anthropic): align the tool search replay cells with main and cover a system message after the search turn
The malformed kept result cells now expect the unanswered search call to be dropped, as main does since #45882.
The stream cut cell checks streamed tool call names, since the kept search result names get_weather as a tool
reference. The streamed response cache hit cell runs again because #46104 fixed it. The modify_params cap cell
expects the four client marks, since the placeholder answer sends the search call as a plain tool_use with its
mark, and a new cell covers a system message after the trailing search turn in both stream modes
* fix(anthropic): count marks on the raw messages when they cannot be placed, and find client answers once per request
* test(anthropic): cover marked tool calls next to a malformed tool id and a later system message
* perf(anthropic): look up client tool answers only when a marked tool call is a server tool call
Answered ids only decide whether a marked srvtoolu_ call is rebuilt as server_tool_use, so a request whose marked
tool calls are all client calls no longer places system messages or runs the modify_params pairing step to count
marks
* test(anthropic): reach the client answer lookup with a marked search call in the malformed tool id cell
* perf(anthropic): place mid-conversation system messages in linear time for models that keep them
* fix(anthropic): count cache marks with the model's own system message placement and pair tool results in linear time
* test(anthropic): cover modify_params dropping a tool answer that answers no call of the latest assistant turn
* fix(anthropic): count marks with in-place system placement when the model lookup cannot confirm the flag
* test(anthropic): cover a bare model with custom_llm_provider filling the cache mark budget through litellm.completion and Router
* fix(anthropic): give unmapped models the request builder's converting placement and survive any model lookup failure
* test(anthropic): cover an unmapped model name counting cache marks with the request builder's converting placement
* fix(anthropic): count server tool call cache marks the request builder's way only on routes that build Anthropic Messages requests
Databricks, OpenRouter and Bedrock Converse send every marked tool call as is, so the hook keeps the census main uses there. The seeding step carries custom_llm_provider on the injection points when a server tool call is marked, so a bare model name still resolves its route.
* test(anthropic): cover two answered marked server tool calls before a system message on Databricks, OpenRouter and Bedrock Converse
* fix(anthropic): keep main's cache mark count on GitHub Copilot and ChatGPT routes without resolving them, so a sync call with no stored login never starts the device flow
* test(anthropic): cover a marked server tool call reaching the GitHub Copilot and ChatGPT login hosts like an unmarked one
* fix(prompt_templates): do not mutate messages in-place in function_call_prompt
* refactor(test): avoid local list mutation and satisfy type discipline gate
* style(prompt_templates): keep line length under 120 chars for ruff format
* style(prompt_templates): type _with_function_prompt with AllMessageValues for ANN401
* style(prompt_templates): eliminate basedpyright errors in function_call_prompt
* fix(prompt_templates): keep tuple and None system content when adding the function prompt
Missing system content becomes the function prompt alone, a tuple keeps its parts, and the old test that pinned the caller-side mutation now asserts the caller's messages stay unchanged
* fix(prompt_templates): reject unreadable system content without echoing it
Only list and tuple content is read as parts. Any other non-string content raises a TypeError that names the type, so error logs never carry the system prompt
* fix(prompt_templates): leave non-list system content failing as before
None and tuple content fail the same way they did before this change, so the only behavior change is that the caller's messages are no longer mutated
* fix(prompt_templates): answer 400 for unreadable system content, send the function prompt alone under drop_params
* test(prompt_templates): pin the function prompt on every system message
* test(integration): cover function prompt fallbacks, retries, cache and drop_params on ollama
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(websearch_interception): intercept the bare web_search tool type on chat completions
* fix(websearch_interception): match the bare web_search type only on chat completions
/v1/messages keeps its existing detector, so a non-Anthropic model there no
longer short-circuits a bare web_search tool into raw search results
* test(websearch_interception): cover the bare web_search type on chat and through the messages adapter
* refactor(typing): narrow Any sources in proxy, guardrails, bedrock and responses code
Turns untyped json.loads, response.json(), header lookups and forwarded
kwargs into object or precise types at their source, so downstream code
type-checks without reportAny / reportExplicitAny. No runtime behavior
changes.
* refactor(typing): narrow Any sources in caching, MCP, providers and proxy request code
Decoded JSON, cache reads and header lookups now enter as object or
precise types, prisma tables used by the shadow-eval logger get typed
Protocol views, and a few SDK parameters move from dict[str, Any] to
Mapping[str, object]. No runtime behavior changes.
* refactor(typing): narrow Any sources in proxy server, router, logging and shared utils
* refactor(typing): narrow Any sources in provider transformations, integrations and proxy endpoints
* refactor(typing): narrow Any sources across providers, proxy and router
Deepens the Any cleanup on the hottest files: typed helpers in
core_helpers for attribute and item access, identity TypeAdapter(object)
wraps at untyped boundaries, explicit return annotations on shared
provider and proxy helpers, and Protocols in place of Any parameters.
No intended behavior change.
* test(integration): wait on the right signal in five flaky integration contracts
The Prometheus shared-label tests waited for the latency metric and then read the
request counter, which the logger writes after an await. They now wait for the counter
The auto-router session poll returned on the first 200 while the caller expected two
turns, so it now waits for the expected turn count
The S3 queue cap test let its failing sink recover on a 15s clock that could expire
mid-flush. The sink now stays down until a flush fails after every event is queued
K3 normally takes 168-175s against a 180s limit, so the limit is now 300s
The MCP env var browser spec reused an idle keep-alive socket that the proxy closes
after 5s, so its API calls now send Connection: close
* test(integration): add until_asserted so a wait polls the exact asserts it guards
eventually() takes a predicate separate from the asserts that follow it, which lets a
test wait on one value and then assert on another written later. until_asserted reruns
a block of asserts until they all pass and re-raises the last failure on timeout
The Prometheus shared-label tests and the auto-router session check now use it
* fix(vector_stores): stop sharing one list across default-constructed registries
VectorStoreIndexRegistry and VectorStoreRegistry took a mutable list as their
default argument. Python evaluates that once at class definition, so every
registry built without arguments shared one list. The proxy builds one that way
during startup, so an upsert into it leaked into every registry created after,
and into the class default itself.
Default to None and build a fresh list per instance. A caller-supplied list is
still used as given, including an empty one, so nothing that passes its own list
changes behaviour.
Fixes#38874
* test(vector_stores): annotate locals Final and mutate registries through their public API
The leak tests now go through add_vector_store_to_registry and upsert_vector_store_index rather than appending to the registry's list directly, which is closer to how the proxy hits the bug.
* fix(vector_stores): drop the last mutable list default and identity-only tests
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(e2e): out-of-process mock OpenAI, Anthropic, Bedrock and guardrail servers for the perf cases
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): coding-agent request corpus and real provider captures for the perf mocks
Request bodies come from a deterministic coding-agent session (system prompt, Read/Edit/Write/Bash/Grep/Glob tools, tool_use and tool_result turns carrying real source files picked by the request marker) in OpenAI chat, Responses and Anthropic formats. captures/ holds one real response per provider, route, shape and stream mode (gpt-6-luna, claude-sonnet-5, us.anthropic.claude-sonnet-5), cold prompt cache, body only, raw stream chunks with time offsets
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): replay module for the perf mocks, captured provider responses picked by route, mode and request size
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf mocks replay the captured provider responses, knobs shrink to flake and speed
The OpenAI, Anthropic and Bedrock mocks answer with the captured body or stream chunks for their route, mode and nearest request size, ids swapped per request, on the captured clock scaled by the p knob. The synthetic reply builders, the per-token stream pacing and the hand-rolled Bedrock event-stream encoder are gone
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf mock request log names its provider
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf guardrail mock remembers the request marker by call id for the response hook
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf mock captures retaken cold, one marker and system-prompt prefix per cell
Run 2 shared one marker across cells, so every stream capture was served from the provider's prompt cache (cache_read_input_tokens up to 523946) and the json cells had small prefix hits. Every cell now carries its own marker, also prepended to the system prompt, and all 30 captures report zero cached tokens
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf corpus reads tests/e2e files so the lane's runner image has them
* test(e2e): perf long_context json captures for Claude retaken with a 16k output cap, no longer cut mid-thinking
* test(e2e): perf mocks keep only what the replay path and the test read
* test(e2e): perf mocks interpolate the provider's time to first byte from the captures around the request size
* test(e2e): perf captures at 100k (short) and 500k (long) tokens with tools, the fixed-size set removed
* test(e2e): PerformanceTestCase schema, deterministic request plan and wire records for the perf cases
RequestShape names a corpus session (input_tokens, tools, output_tokens) and wire.request_body builds it with setup/corpus/session.py, so a 300k request is 85 file reads of real source in the endpoint's own format. The plan no longer predicts reply text, the mocks replay captured provider responses and the runner joins what the client saw against the mock's log by marker
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf Metric gains time_to_first_token_seconds
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf User loses the unused role field, every JWT user is an internal_user
* test(e2e): perf RequestShape sizes requests from an input-token range, short and long stock shapes
* test(e2e): perf case runner (provision, Locust plan replay, observe) and the 50 rps mixed-provider case
provision.py checks the deployed config against the case and builds the callers, run.py replays the plan from a separate Locust process (locustfile.py) recording status, header, first-byte, first-token and total time per request, observe.py joins the records with what the mocks received and the workers' RSS to score every Limit, and compares each served body with the capture the mock replayed for it as a diagnostic. The case sends 15,000 requests at 50 rps over three model groups, two guardrails and the three request shapes
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf case asserts time_to_first_token_seconds at percentile 50 and 99
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf provision creates every JWT user as internal_user
* test(e2e): perf observe reads only the mock log fields that remain
* test(e2e): 50 rps case sends short and long coding-agent sessions sized per request
* test(e2e): perf run keeps the per-request records as perf-records.jsonl under E2E_ARTIFACT_DIR
* test(e2e): the mixed small-gateway case runs at 25 rps (50 saturated one 4-worker pod on main and on v1.105.1)
* test(e2e): small_bodies_high_rate perf case, the github issue 46276 shape (11 to 17 KB chat bodies, keys only, 200 rps)
* test(e2e): the mixed small-gateway case runs at 10 rps for 6,000 requests (25 rps still saturated one pod on v1.105.1)
* ci: python lint job gets 20 minutes, the e2e harness phase (625 tests plus the replay harness suite) pushes a run past 15 when tests/e2e changes
* test(e2e): small-bodies perf case runs at 5 rps so gateway_added_seconds is per-request overhead, not queue wait
* test(e2e): mixed-provider perf case runs at 5 rps x 3,000 requests from 100 users
* test(e2e): mixed-provider perf case sends happy-path traffic only (no flaky deployment, no guardrail interventions or blocks)
* test(e2e): perf gateway model_list is three happy-path deployments, one per provider
* test(e2e): Postgres and Redis work per request as perf case limits
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): perf mocks wait a case-level latency with jitter instead of replaying captured timings, cpu_seconds_per_request limit from metrics-server
* test(e2e): perf suite PR carries the small_bodies_5rps case only, the mixed-provider case moves to its own PR
* test(e2e): small_bodies_5rps limits from the ten v1.105.1 baselines (1.5x on datastore counts, 2x on latency and cpu, absolute on errors and memory)
* test(e2e): perf mocks keep only the short captures the small case reads, captured timings and the Bedrock invoke route removed
* test(e2e): perf mocks generate each reply from the request (provider schema, usage, ids, event order, chunk sizes) instead of replaying captures
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(agents): support Databricks Agent Platform agents in the agent gateway
Add a databricks_agent provider that speaks the MLflow ResponsesAgent wire
(Responses-shaped input, output_text items, output_text.delta SSE) to Model
Serving endpoints (/invocations or the unified /serving-endpoints/responses)
and Databricks Apps (/responses), with a PAT or OAuth M2M client credentials
Register it through the agents config block or the Admin UI agent catalog and
call it through /a2a/<agent> (message/send, message/stream) or model:
a2a/<agent> chat completions. The chat route now runs every completion-bridge
agent through its own provider, and attaches the minted Databricks OAuth token
for both bridge and URL agents. Consumed OAuth settings never reach the
provider call
* fix(agents): harden databricks_agent input, output, and OAuth handling
Non-text content parts answer 400 unless drop_params is set, a reply without
an output list answers 502, flat OAuth fields resolve an os.environ api_base,
the bridge debug log no longer echoes the request model, routing tests stub
only the HTTP boundary, and the catalog description fits the dropdown
* fix(agents): merge agent and client extra_headers on the a2a chat route, and reject an empty databricks_oauth block
* fix(agents): bridge only databricks_agent agents on the a2a chat route, keep the card url over a client api_base, and require workspace_url for a Databricks App
* refactor(agents): read the a2a chat bridge providers from a constants set instead of branching on a provider name
* fix(databricks_agent): let the configured credential win, reject malformed client extra_headers, and keep a missing endpoint name empty
* test(agents): drive the flat OAuth a2a send through the real bridge
* fix(agents): keep an empty databricks_oauth block ignored on url agents, as before this change
* fix(agents): accept bare App hosts, env service principals, and os.environ refs on Databricks agents
* fix(databricks_agent): address review caveats on chat bridging, params, and auth
Map max_tokens and max_completion_tokens to max_output_tokens and pass
temperature, top_p, and user, so /v1/messages works without drop_params.
Send A2A message metadata only with string values, answering 400 or
dropping the rest under drop_params. Join message items with a blank line
in both stream and non-stream replies. Cache the env service principal
token until shortly before it expires. Keep databricks_agent replies out
of the response cache.
On the chat bridge, let request fields win over the agent's stored params
while pinning the agent's api_base and api_key, and apply the agent's
static_headers between client and agent extra_headers, as /a2a does.
Resolve os.environ/ only for databricks_agent api_base and api_key instead
of every bridge provider. Ignore an unusable databricks_oauth block on URL
agents on chat, as at the merge base, instead of answering 400.
* test(internal_context): list the databricks agent token cache as process-local
* fix(databricks_agent): keep the agent's endpoint, input, and target out of client control
extra_body model and input answer 400, or are dropped under drop_params.
A client api_base or base_url never replaces the agent's card URL on the
bridge. Chat to a URL agent keeps its merge-base call, and an incomplete
databricks_oauth block on a URL agent keeps its merge-base ValueError.
* test(databricks): expect the merge-base 400 for a client api_base on agent chat
* fix(databricks): keep agent-owned inputs and identity headers out of a chat caller's reach
* fix(databricks): bound the agent-over-caller merge to two levels so it needs no recursion
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(decisions): read a deployment's own supports_box_questions flag before the cost map
The box-question check only read the shared cost-map row and required a decisions mode there, so an OpenAI-format deployment whose model_info sets supports_box_questions was refused with a 400, which kept the wire test from #46633 red on main
* fix(decisions): keep the box flag on its deployment and out of request bodies
A model_info that is not a plain dict no longer fails the call, the proxy
decisions endpoint drops a caller-sent model_info, and the router stops
copying supports_box_questions onto the shared cost-map key.
* fix(decisions): keep the box flag shared with same-model deployments like main
Scoping supports_box_questions to one deployment turned box questions on an unflagged sibling of the same backend model from reaching the provider into a 400. Siblings inherit the flag through the shared cost-map row again, as they do on main. Also formats two new tests.
* test(decisions): pin box flag inheritance by an unflagged same-model deployment
* test(decisions): audit box-question deployment flag across providers, routes, and chaos cases
* fix(cost): price bridged /v1/messages reasoning tokens from the chat completion, not the served Anthropic body
* fix(cost): keep the type gate flat for the bridged /v1/messages billing path
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): treat adaptive thinking as enabled when mapping response_format
is_thinking_enabled only recognised thinking.type=enabled, so a request that
paired thinking.type=adaptive with response_format took the synthetic JSON
tool path with a forced tool_choice, which Anthropic rejects alongside
thinking. The same check also gates the Bedrock Converse and Databricks
forced tool_choice downgrades, so they now see adaptive thinking too.
* fix(anthropic): keep signed thinking blocks with empty text on chat completions
Models that default to display: omitted return signed thinking blocks whose
text is empty, and Anthropic rejects the next turn when they are not replayed
unchanged. is_unsignable_thinking_block dropped them along with unsigned
blocks, and the inline content path required non-empty text, so a tool-use
turn on Opus 4.7+ or the 5 families lost its thinking blocks and 400ed on the
tool-result turn. Only a missing signature now drops a block, and the
mid-conversation system placement mirrors that rule. The /v1/messages strip
helper is left to #41362.
* fix(anthropic): skip the thinking-strip retry when block_binding asks for an error
The /v1/messages retry loop strips thinking blocks and resends on any
invalid-thinking 400, including the binding mismatch a caller opted into
with thinking.block_binding.prefix_mismatch_behavior: error. The retry
decision now sees the request body and surfaces that 400 instead of
overriding the caller's choice. #42888 covers the litellm_params side.
* fix(bedrock): carry thinking.block_binding through the Invoke thinking rebuilds
Bedrock Invoke rebuilt the thinking param from scratch when display: updates
was requested, on both the /v1/messages and chat completions paths, and the
legacy enabled-to-adaptive translation plus the clear_thinking injection did
the same, so thinking.block_binding never reached Bedrock. Every rebuild now
spreads the caller's thinking and only overrides the keys it owns, and
AnthropicThinkingParam types block_binding. The thinking-binding-controls
beta was already on the Bedrock allowlist.
* fix(anthropic): keep forcing the response_format tool under adaptive thinking
This reverts ad9e861ae7. Anthropic, Bedrock Invoke and Vertex all accept a
forced tool_choice next to thinking.type=adaptive and only reject it with
type=enabled, so counting adaptive as enabled fixed no live failure. It did
drop the forced JSON tool, so Claude answered response_format requests in
prose, and it downgraded a required tool_choice to auto on Bedrock Converse
* test(e2e): drop the adaptive thinking tool round trip case
With adaptive thinking the model often skips thinking on the weather prompt, so
turn 1 carries no signed empty block and the case fails before the replay it is
meant to cover. Anthropic also answers the replay with a 200 whether or not the
block is forwarded, so the case could not tell the fix from main. The factory
unit tests pin the outgoing shape instead
* fix(anthropic): only skip the thinking-strip retry for a binding mismatch 400
prefix_mismatch_behavior: error asks Anthropic to reject a block bound to
another conversation. It says nothing about unrelated bad blocks such as a
null signature, so those still get the strip-and-retry self-heal. Also trims
the empty-thinking docstrings to what the code guarantees and types the new
test parameters
* refactor(anthropic): narrow thinking fields with the mapping validator instead of typing.cast
typing.cast is banned by the strict ruff gate (TID251), so the three new casts pushed that rule one over its merge-base count. The existing pydantic mapping validator gives the same typed view without an unchecked cast, and it now has a generic name since it is no longer only used for Claude Code payloads
* refactor(anthropic): read thinking settings through the validated mapping so basedpyright stays at its merge-base count
The Bedrock clear_thinking rewrite still narrowed with a cast and an untyped dict fallback, which added two reportUnknownVariableType errors. It now uses the same pydantic mapping validator, made public for that reuse, and the retry hook validates the request body itself so the handler passes a typed dict instead of an untyped one
---------
Co-authored-by: yuneng-jiang <yuneng@berri.ai>
* feat(proxy): split Azure PTU deployments across teams by share
A PTU deployment's model_info takes ptu_shares (team_id -> whole PTUs adding up to
ptu_count) in place of a single team_id. With LITELLM_ENABLE_PTU_COST_ATTRIBUTION on,
routing hides the deployment from teams holding no share, the v3 limiter enforces each
share as a per-team TPM ceiling from Azure's per-model TPM-per-PTU sizing, the daily
flat-cost rollup splits the hourly cost by share, and /team/daily/activity plus the Team
Usage page report PTU-hours next to tokens.
* fix(proxy): weight cached input in PTU ceilings and size team-scoped deployments by public name
The PTU flag is read from the process environment on every check instead of
through the secret-manager path, so a proxy with a hosted secret manager does
not pay a round trip per request. A team's ceiling now carries the model's
cached-input ratio and settlement charges cache reads at that ratio, matching
how Azure sizes a PTU. A deployment registered for one team is found by its
team_public_model_name for PTU-hours and the ceiling, since its model_name is
rewritten to a unique internal name. ptu_shares are StrictInt on ModelInfo and
parsed_ptu_shares refuses booleans, strings, and non-string team ids, a
fractional ptu_count reserves nothing, and models_processed counts distinct
deployments rather than holdings.
* fix(proxy): reserve PTU ceilings in normalized tokens, keep share routing flag-independent, and warn only shared or Azure reservations
* fix(proxy): size the uncapped output floor against the PTU share in raw output tokens and escape the exported team id
* fix(ptu): refuse ptu_shares naming a team that does not exist
* refactor(ptu): keep Azure detection under llms/azure and check share teams with one query
* fix(proxy): drop the reverted TeamIdSearchFilter import after merging main
Main reverted the team key search change that added TeamIdSearchFilter, so the PTU share team check types its where clause locally
* fix(proxy): chunk the ptu_shares team lookup under the IN-list bound
* fix(proxy): count a request naming a shared PTU deployment directly against its group's ceiling
* fix(proxy): resolve every name a shared PTU deployment is served under to its group's ceiling
* fix(ptu): refuse ptu_shares declared without ptu_count and cost_per_ptu_per_hour
* fix(ui): seed total_ptu_hours in the empty daily activity metadata
* fix(ptu): size usage rows by the name the request used
* fix(ptu): size a team's usage rows by the deployment it is served from
* fix(ptu): never size a team's usage row by another team's reservation
* fix(ptu): reserve an uncapped request against the cap the proxy writes, not its prompt length
* fix(ptu): read a share map only through the terms registration requires
A ptu_shares map stored without ptu_count, cost_per_ptu_per_hour and
ptu_effective_from, or one that does not add up to ptu_count, is refused at
registration but was enforced once loaded: every other team got the 400 and
the share holder a ceiling while no flat cost accrued. _deployment_shares now
reads the shares through ptu_terms, so such a row reserves nothing, the same
as before the map was read.
* fix(ptu): keep a declared split as the access rule and check a config split at boot with the flag off
6fb24d1d read ptu_shares through the pricing terms so a termless map reserved
nothing, but with the flag off the router skips ptu_config_error for config.yaml
entries, so such an entry loaded and served every team, against the rule that a
declared split restricts access whether or not cost attribution is on.
_deployment_shares reads the declared map again, and the router now raises
ptu_config_error for a config-sourced entry declaring ptu_shares regardless of
the flag, so a malformed split is refused at boot instead of being honoured or
ignored silently.
* fix(ptu): charge wildcard-served names, pass-through usage, and team-less settlements against the team's PTU ceiling
A name served by a wildcard route now resolves to the shared deployment behind it, so the
limiter's ceiling and the team daily activity sizing both find it. Pass-through responses
settle the PTU counter from combined_usage_object in Azure-normalized tokens, and a
reservation taken at admission is settled even when the success metadata lost the team.
* fix(ptu): resolve a wildcard-served name to the one pattern the router picks, and a deployment id before any wildcard
A team's wildcard reservation behind an open proxy-wide * route was charged and sized for traffic the * route served, and naming a shared deployment by id next to a catch-all wildcard escaped its per-team ceiling. Router.wildcard_route_deployments now owns the most-specific-proxy-wide-else-team precedence, and both the limiter and the consumption page resolve through it after listed rows and deployment ids.
* test(router): pin the wildcard route precedence that PTU sizing and the router share
* fix(proxy): accept whole-number PTU shares written as floats, as ptu_count already does
A share of 2.0 now reads as 2 on POST /model/new and in config.yaml, where it was
refused with a 422 and dropped at load. Fractional, infinite, boolean and string
shares stay refused. Adds unit tests for the malformed-rate, provider-over-prefix,
zero-token Responses usage, fractional-count and sizing-warning paths.
* fix(proxy): settle team PTU reservations in their own minute and keep billed PTU days when shares move
* fix(proxy): write each PTU deployment day's team charges in one transaction
The backfill counts a deployment day as priced once any team's row exists for it,
so a rollup that wrote team A's share and lost team C's left team C unbilled for
good. Each deployment day's team rows now go through one Prisma batch with the
retry around the whole batch, so a day is fully written or not written at all and
the next backfill prices it for every team.
* fix(proxy): read the PTU flag from a secret manager, serve share holders from their share, and refuse unfittable calls and share-holding team deletes
* test(router): list the cached PTU ceiling helpers in the router coverage check
* fix(proxy): read the PTU flag only for share holders, guard org and SCIM team deletes, keep share holders on their shared deployment through retries and the scheduler, and date PTU-hours by the terms in force
* fix(proxy): bill a rerun PTU day to the current share holders alone
* fix(proxy): replace a PTU day's rows by day and deployment, not a negated team list
* fix(proxy): read the PTU flag from the secret manager first, like get_secret
* fix(router): scope retry health lookups to the caller only for PTU share holders
* fix(router): move the share-holder retry check into ptu_shares
Adds ptu_hours and total_ptu_hours to the daily activity integration goldens and fixtures, since every route now reports them
* test(integration): cover PTU shares across routing, delete guards, cost attribution and config regex timing
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(guardrails): resolve decision and judge models in caller team scope
* fix(guardrails): scope guardrail model calls to the caller's team only where it owns the deployment
Read the team only from the auth object the proxy attached to the request, never from caller JSON,
and use it only when the router's route to the guardrail model for that team reaches a deployment a
teamless caller cannot. Shared names, aliases, legacy team tags and wildcards keep resolving and
billing as before. Decision calls pass the team in litellm_metadata only when it is used, so the
logged model group stays intact.
Adds integration cells for chat, responses and messages, forged team ids, a team deployment that
shares a name with a shared model or wildcard, and billing, plus unit cases for legacy team tags
and aliases.
* fix(guardrails): route team guardrail calls to the team deployment without charging the team
Guardrail calls on a team's own model now go to the deployment the router picks for that
team, by its id, with no team on the call. Like every other guardrail call, they write no
spend row and add nothing to the team's or the key's spend
* fix(guardrails): pin the team deployment only for the judge guardrail, not shadow eval judge calls
* fix(guardrails): keep the teamless route when only a team wildcard covers the guardrail model
* refactor(guardrails): scope the team model fix to the decision model guardrail only
* fix(guardrails): send decision calls on a team's own model with the team, and keep team wildcards out
The decision call on a team's own model now carries the caller's team instead of being pinned to a
deployment id, so it is charged to that team and goes through the team's normal routing: an access
window that reserves the deployment for the team lets the call through, and pre-call checks see the
input and skip a deployment too small for it.
A team deployment listed under a wildcard public name, such as team_public_model_name openai/*, no
longer captures a guardrail model the team never named, whatever model it points at.
Adds unit and integration cells for the reserved deployment, mixed context windows under
enable_pre_call_checks, a team wildcard name over a fixed model, and the owner team's spend.
* fix(guardrails): log the owner team's name on decision calls charged to it
The decision call on a team's own model now carries the team alias beside the team id, so the
spend row and the Logs page name the team the call was charged to.
* fix(decision_model): use the caller's team only for names nothing else serves
A team model named like a shared deployment, a shared wildcard or an SDK
public name took over the guardrail's decision calls for that team, so a
team admin could point a proxy-wide guardrail at their own endpoint. The
team is now used only when a caller without a team reaches no model under
the name.
* refactor(decision_model): drop the team route subset check the served check covers
Once nothing serves the name without a team, every deployment a caller
without a team can reach is a wildcard, so the subset check only ever
rejected an empty team route. Pin team wildcards over a name nothing else
serves.
* fix(decision_model): use the caller's team only when judge_target finds nothing without a team
The served check rebuilt the router's rules and missed a shared wildcard whose
provider comes from custom_llm_provider, so a team model under the guardrail's
name took over that team's decision calls. Ask judge_target, the same lookup the
teamless dispatch makes, instead
---------
Co-authored-by: moyai-devin-berriai[bot] <336287033+moyai-devin-berriai[bot]@users.noreply.github.com>
Co-authored-by: yucheng-berri <293926326+yucheng-berri@users.noreply.github.com>
