* fix(context_management): count the system prompt toward the clear_tool_uses input_tokens trigger
Anthropic's API compares an input_tokens trigger with the whole prompt, system
included; the polyfill counted messages and tools only, so it cleared late by the
system prompt's size. compact_20260112 already counts it: share its system_to_text.
* fix(context_management): take read-only types in the system prompt count
The type-discipline gate (LIT001) rejects new list and dict annotations.
* chore: rerun CI (proxy-build-from-pip timed out)
* fix(context_management): count the system prompt through compact's shared counter
Finishes the main merge's conflict in clear_tool_uses.py. The input_tokens trigger now adds the system prompt through count_system_tokens, the same counter compact uses, while the count reused for cleared_input_tokens stays messages and tools only, so the reported savings are just the cleared tool results
* perf(context_management): skip the token count when nothing can be cleared
The keep check runs before the trigger count, the system prompt reader accepts any mapping block, and the system prompt test gets short parameter ids
* test(context_management): keep the off-loop count test counting after the keep check moved first
* test(context_management): cover a client compaction summary counting toward the trigger
The trigger also skips the system prompt count when messages and tools alone already pass it
* fix(context_management): move a client compaction summary into the system prompt when clearing runs before compact
* fix(context_management): clear only the history after a client compaction block so compact still replaces its summary
* fix(context_management): keep tool calls beside a client compaction block clearable
---------
Co-authored-by: AlonSoko <25619863+AlonSoko@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(responses): drop the null created_by OpenAI rejects when output items are replayed
* fix(responses): drop a null created_by from replayed input items too
A conversation saved from an older proxy still carries created_by: null on its compaction, apply_patch and shell items, so OpenAI keeps rejecting the replay. Drop the null from input items on the HTTP request transform and the WebSocket response.create rewrite, keeping a created_by that is not null.
* fix(responses): drop a null created_by from counted input items too
* test(responses): cover saved reasoning nulls and re-compaction of a saved conversation
* fix(responses): drop a set created_by from input under drop_params
* fix(types): type the Responses output serializer's parameters
* fix(responses): drop a set created_by on reasoning items and Perplexity presets under drop_params
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(azure_ai): cover Foundry agents on the project Responses API
* feat(azure_ai): reach Foundry agents through the project Responses API
* fix(azure_ai): keep system prompts and project-endpoint auth on Foundry agent calls and follow-ups
* fix(azure_ai): read AZURE_SCOPE before the Foundry audience default on project endpoints
* fix(azure_ai): reject or drop instructions on previous_response_id turns and a caller's other agent for Foundry agents
* fix(azure_ai): apply the other-agent rule to agent_reference sent at the top level of a Foundry agent request
* fix(azure_ai): match the caller's agent by name and version, and hide the agent's own instructions
* fix(azure_ai): keep instructions on plain Foundry models that return a null agent_reference
* fix(azure_ai): keep a Foundry agent version requested at the top level when extra_body also names the agent
* fix(azure_ai): answer 400 for /responses/compact on a Foundry agent instead of calling the project with the agent name
* feat: add StreamLake provider metadata
* fix(streamlake): complete provider discovery and test request flows
* test(streamlake): assert cost and tool-param invariants instead of pinning vendor prices
* fix(streamlake): accept tools and reasoning on endpoint ids and match the current price page
Endpoint ids have no cost-map row, so tools and reasoning_effort were rejected
before reaching StreamLake. The provider now declares both capabilities, and
named non-reasoning rows opt out explicitly.
The rows now follow StreamLake's 2026-09-17 price page: drop the KAT-Coder and
Qwen3 rows it no longer serves, add the served models that had no price, drop
account-level rpm and tpm, and fill missing context windows. StreamLake is also
registered for streamlake/* wildcards and its messages and responses support.
* fix(streamlake): regenerate the endpoint matrix and type the new tests
The backup endpoint matrix advertised chat only while the root map listed
Messages and Responses, so /public/endpoints hid both. Regenerating with
scripts/generate_provider_endpoints_support.py makes the two files match.
Types the StreamLake tests and sets the Add Model test's API key with
fireEvent.change, per the dashboard test guide.
* fix(streamlake): drop the retired max_tokens field from the StreamLake cost-map rows
* fix(streamlake): price the four served models StreamLake publishes rates for and add the Qwen3.8 cache prices
* fix(streamlake): send token caps as max_completion_tokens and mark image models
StreamLake ignores max_tokens on kimi-k2.6, kimi-k2.7-code, qwen3.6-27b and
qwen3.6-35b-a3b, while every served model honors max_completion_tokens, so
the JSON provider now maps max_tokens to max_completion_tokens instead of the
reverse. Rows whose models read an image in a live probe set supports_vision,
and the Qwen3.8 rows drop the explicit cache write price, which this route
never uses since cache_control is dropped
* fix(streamlake): keep cache_control marks, price explicit cache reads, and use API input limits
StreamLake reads cache_control, so the provider keeps the marks through a preserve_cache_control flag. Qwen3.8 explicit cache hits now use their own read price, rows with a cache read price declare prompt caching, three models StreamLake serves without reasoning drop reasoning_effort, and max_input_tokens follows the limits the API reports.
* fix(utils): warn when a custom_provider_map name matches a built-in provider
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(guardrails): scan Responses tool results in texts-mode guardrails
* fix(guardrails): keep guardrail-added tools and non-text parts on Responses tool scans
Tool result texts now come from the same conversion that builds the chat tool message, so items without a call_id, dict outputs, and outputs with no text parts stay aligned for Bedrock and PANW. A changed joined text replaces the first text part and keeps the other parts. scan_only_tool_results appends tools the guardrail returns with new names, like chat, and tool_result and computer_call_output items are scanned through the shared type list.
* fix(guardrails): skip empty and screenshot tool outputs and keep text part types on Responses scans
* fix(guardrails): keep Bedrock latest-turn masking aligned when tool outputs send no text
* fix(guardrails): scan each Responses tool output text part on its own
* fix(guardrails): split a Responses tool output's structured row when only textless outputs share its call_id
* fix(guardrails): pair each Responses tool output with its own structured row
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(responses): drop metadata from compact requests to OpenAI and Azure
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): drop user and drop_params-rejected fields on compact and route Azure compaction to v1
* fix(responses): keep stream on compact and keep non-dated Azure api-versions on the v1 compact route
* fix(responses): drop compact params only for providers verified to reject them
---------
Co-authored-by: mateo <mateo@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(vertex_ai): drop the caller's anthropic-version header on Claude /v1/messages
Vertex reads the API version from the body's anthropic_version and answers 400 when a forwarded anthropic-version header names a different one, which anthropic SDK 2.x always does. Strip the header on the Python and LITELLM_RUST Vertex messages routes.
* fix(vertex_ai): carry the caller's anthropic-version into the Claude /v1/messages body
Vertex reads the API version from the body alone, so the caller's
anthropic-version header now becomes the body's anthropic_version instead
of a pinned vertex-2023-10-16. The 2023-01-01 and 2023-06-01 first-party
versions keep vertex-2023-10-16, and the header itself is still never sent.
Same change on the Python proxy and the Rust gateway.
* fix(vertex_ai): follow drop_params for an anthropic-version Vertex does not accept
Only versions Vertex accepts in the body (vertex-2023-10-16, 2026-10-21) are
carried into anthropic_version. 2023-06-01, 2023-01-01, a blank header, or no
header keep vertex-2023-10-16. Any other value drops to vertex-2023-10-16 under
drop_params and answers a 400 naming drop_params without it, on the Python proxy
and the Rust gateway.
* refactor(vertex_ai): return the unsupported anthropic-version error as a value and keep Rust helpers and tests in place
* fix(rust gateway): resolve os.environ drop_params and accept the same flag words as Python
The gateway read drop_params: os.environ/NAME and words like yes as off, so the drop_params opt-out from the new anthropic-version 400 did not work there
* docs(rust router): name Router::from_config in the crate README
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
OpenAI shut down dall-e-2 and dall-e-3 on 2026-05-12. Drop their image
generation and edit configs, the Azure dall-e-2 create-and-poll flow, and
the silent dall-e-2 fallback when no model is passed.
image_generation without a model or provider now fails locally with
"model is required". image_edit and image_variation leave model out of
the provider request instead of sending dall-e-2, so providers that
ignore model (Topaz) keep working and OpenAI returns its own 400.
* fix(vertex_ai): split Gemini SSE lines on CR/LF only so U+2028 in text keeps the stream
The Gemini and Vertex AI streaming path read the SSE body with httpx
iter_lines()/aiter_lines(). Those follow str.splitlines(), which also breaks
on U+0085, U+2028, U+2029, \x0b, \x0c and \x1c-\x1e. JSON leaves U+0085,
U+2028 and U+2029 unescaped inside strings, so a data: line whose text holds
one of them was cut in two. The tail is not a data: line and the head never
parses, and the accumulate-and-retry fallback then carries the unparsed
prefix forward, so every later event is appended to it and the stream ends
with no more content and no usage.
SSE lines end only at CRLF, LF or CR. The new iter_sse_lines/aiter_sse_lines
split decoded text on exactly those, treating a CR at the end of one chunk and
an LF at the start of the next as one break. The first request and every
long-decode hop now read through them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(streaming): keep merged think tags balanced and split shared SSE lines only on CR/LF
Close an open <think> block on the first content or tool-call delta and at stream end, reopen it when reasoning returns, and route the shared chat completions stream through the SSE line splitter so U+2028 and U+2029 inside JSON no longer drop chunks.
* fix(streaming): exact-match the SSE [DONE] marker in BaseModelResponseIterator
* fix(streaming): split every SSE reader on CR/LF only and keep the logged think block closed
Switch the remaining httpx stream readers (Anthropic chat, Google GenAI,
Interactions, Codestral, openai_like, Predibase, LangGraph, Vertex Agent
Engine, A2A providers, Azure AI agents, sandboxes, the A2A endpoint, and
others) from iter_lines/aiter_lines to iter_sse_lines/aiter_sse_lines, so a
raw U+2028 or U+2029 inside a JSON string no longer splits an event.
Keep the chunk that closes a think block in the logged and cached response
when a stream ends without a finish_reason, close the block before a legacy
function_call, and make sync Codestral FIM streaming send stream=true and use
the text-completion-codestral stream parser.
Test fakes that only mocked iter_lines now stream real httpx bytes and
assert the content they feed in.
* fix(streaming): split buffered SSE parsers only on CR and LF, shard the codestral tests
Buffered SSE and JSONL parsers used str.splitlines(), which also breaks on
U+2028, U+2029 and U+0085 that JSON allows raw inside strings. They now use
split_sse_lines from sse_lines.py, which also gains a str.split fast path.
Adds tests/unit/llms/codestral to the unit workflow shard.
* fix(streaming): keep U+2028 in restamped Anthropic frames and fold CRLF before splitting SSE lines
* test: type the new stream tests, drop a comment, and benchmark the SSE line splitter
* test: type the iterator class parameters in the line-separator stream tests
* fix(responses): stream tool call arguments opened in a text chunk
* fix(responses): hold a tool call that shares a chunk with text until the text ends
* test: type the a2a SSE handler helper and the sandbox fake responses
---------
Co-authored-by: Tim Vink <vinktim@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: michelligabriele <gabriele.michelli@icloud.com>
* fix(gemini): preserve every grounding web source
* test(gemini): cover all grounding sources per passage
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Zabir Raihan <zabirraihan@icloud.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(types): replace Any with proven types in 52 files
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(types): use typing_extensions.TypedDict for validated seams
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(types): revert strict Anthropic response validation
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(types): accept a null finish reason from Bedrock Cohere and TwelveLabs Pegasus
* test(bedrock): annotate the Pegasus null finish reason test
---------
Co-authored-by: mateo <mateo@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(tencent): add Seedream image generation and editing through TokenHub
* feat(tencent): price Seedream image generation and edits from TokenHub usage
* test(tencent): cover Seedream image generation, edits, and pricing through TokenHub
* fix(tencent): keep non-Seedream image models on the OpenAI-compatible route
Only seedream- models use TokenHub's wand route, so Hy-Image and other
tencent image models keep working as on main. The response now carries
TokenHub's image size and format, Pro's flat fallback bills its default
2K size, and the new tests are typed.
* fix(tencent): bill other Tencent image models as before and forward user and TokenHub params on edits
* fix(tencent): keep dropped edit params out and parse form-string watermark and seed
* fix(tencent): send extra_body edit fields to TokenHub as given
* fix(tencent): keep the routed model on Seedream edits and bill a deployment's per-token price
* fix(tencent): accept string drop params on edits, rebuild nested form fields, and honor a per-image deployment price
An image edit with a non-list additional_drop_params answered 500 for every provider after this branch validated it as a list, so a non-list value now drops nothing again, as chat treats it.
Seedream edits sent through the OpenAI SDK arrive as multipart form fields, so TokenHub-only booleans and integers are parsed and bracketed keys like optimize_prompt_options[mode] are rebuilt into objects. A deployment that sets only output_cost_per_image is billed that price, and the TokenHub URL uses the shared api_base helpers.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(ui): let the per-model budget field take a model pattern or access group
* feat(proxy): let model_max_budget entries name a model family or access group, shared across replicas
* fix(proxy): open each shared model budget window once for every replica, and match patterns in linear time
Without Redis, the first replica to charge an entry in a new window now opens it in the window spend table and
the others join it, so no flush replaces spend another replica already added. Budget patterns match by splitting
on `*` instead of a regular expression, a refused request names the entry that refused it, and the key form no
longer offers a typed pattern another row already budgets.
* fix(proxy): read the shared budget window from the writer, and keep concurrent first charges on one replica
Opening a shared window now reads the row from the writer, since a lagging read replica could hand a replica its
own start and let a flush replace the others' spend. A charge that waited on the open no longer resets a local
window another charge on the same replica opened meanwhile. Window tests use fixed timestamps.
* fix(proxy): stop a non proxy admin from lifting a team model cap with a looser group or pattern entry
Group and pattern entries now take precedence over other entries, so a team update by anyone but a proxy admin
must keep each new or changed entry within every existing cap it can take precedence over. A pattern can match a
provider-prefixed or wildcard-routed name, so any two patterns count as overlapping, and a model or group entry
counts as outranking every pattern.
* fix(proxy): narrow model budget pattern overlap, keep team caps under key patterns, and accept only typed patterns in the UI
* fix(proxy): keep model family budgets through database blips, team models and key entries
A failed window open no longer restarts the window at this replica's clock, so it cannot wipe shared spend. Spend charged meanwhile stays local and joins the shared window once the database answers
A key entry named after a model replaces only the team entry of that name, so team pattern and group caps still apply. Team-scoped models are gated and charged under their public name, and group lookups see team deployments
The team update guard compares caps by window length and only ranks patterns that can match the same name, including Bedrock vendor patterns behind a bedrock wildcard route
* fix(proxy): refuse team entries that take precedence over a capped model_max_budget entry
A non-proxy-admin could add a carve-out (an exact model, a narrower pattern or
a group) at or under the family cap, and its own counter then multiplied the
team's spend on that family. Any new entry that takes precedence over a capped
entry now needs a proxy admin, and group members include the team's own public
model names so team deployments cannot hide a group overlap.
* test(router): cover team_public_model_names_of
* docs(proxy): name pattern and access group keys in model_max_budget docstrings
* fix(proxy): name the entry holding the cap in the team model budget raise refusal
* fix(proxy): check precedence when a team edit makes an unenforced model_max_budget entry take effect
* fix(proxy): let an org admin give a legacy model_max_budget row a budget_duration at the same or a lower cap
* fix(proxy): keep a legacy model_max_budget row's stored cap when an org admin gives it a budget_duration
* fix(aiohttp_transport): keep the owned session open while a response stream is still being read (LIT-10237)
* test(aiohttp_transport): annotate the stream regression test locals with Final and drop its docstring
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(pricing): add databricks priority pay per token prices
* fix(pricing): bill gemini 3.6 flash cache writes at its databricks input rate
* fix(pricing): drop image token priority rate the cost map schema does not carry
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
A Responses API stream that fails partway now bills the usage it already consumed, ends with an error instead of a clean stop, and honors stream_timeout between chunks. Bridged /v1/messages calls report those failures to the proxy's failure hooks, so spend, budgets and logs see them
* fix(aiohttp_openai): forward service_tier in transform_response
* chore(aiohttp_openai): mark service_tier assignment rebind-ok for the type gate
* fix(aiohttp_openai): set service_tier with setattr like other providers
ModelResponse has no declared service_tier field, so attribute assignment
added a reportAttributeAccessIssue error to the type gate.
* test(aiohttp_openai): verify serialized served service tier
* fix(aiohttp_openai): copy response when preserving served tier
---------
Co-authored-by: tin-berri <tin@berri.ai>
* feat(providers): add ZeroGPU as an OpenAI-compatible provider
Registers zerogpu through the JSON provider registry so zerogpu/<model> routes to
https://api.zerogpu.ai/v1/chat/completions with ZEROGPU_API_KEY and an optional ZEROGPU_API_BASE.
Adds LlmProviders.ZEROGPU, the constants.py provider and endpoint lists, cost map entries for the
10 ZeroGPU text generation models, the endpoint matrix, the Add Model form fields, the README row
and the dashboard provider entry
* feat(zerogpu): add capability flags verified against the live API
Live calls to https://api.zerogpu.ai/v1/chat/completions confirmed tools, tool_choice and json_schema
response_format on 8 of the 10 catalog text models, cached token reporting on deepseek-v4.1-flash,
gpt-4.1-mini, gpt-5.4-nano and gpt-5.6-luna, and reasoning_effort on gpt-5.6-luna. LFM2.5-1.2B-Instruct
answers stream requests with plain JSON instead of SSE and LFM2.5-1.2B-Thinking returns 500 on every
request, so both stay out of the cost map until the API serves them correctly
* feat(zerogpu): price LFM2.5-1.2B-Instruct
Adds zerogpu/LFM2.5-1.2B-Instruct to the cost map with the catalog prices and context window. It
gets no capability flags because the live API ignores tools, response_format and
max_completion_tokens for this model, so LiteLLM drops tool params for it instead of sending them.
LFM2.5-1.2B-Thinking stays out of the cost map
* fix(zerogpu): reject transcription requests before calling the upstream API
ZeroGPU only serves chat completions, so drop it from the OpenAI transcription transport set. A zerogpu transcription request now fails locally instead of hitting an /audio/transcriptions endpoint that does not exist
* test(zerogpu): cover the responses and messages bridges
Both endpoints reach ZeroGPU through chat completions, so assert the request each one sends upstream and the reply it returns
* refactor(zerogpu): move zerogpu out of the provider list tails
New upstream providers append to the end of openai_compatible_endpoints, openai_compatible_providers and providers.json, so an entry at the tail conflicts with every one of them. Placing zerogpu next to long-stable entries keeps those merges clean. No behavior change
* fix(zerogpu): price cached input for qwen3-30b-a3b-fp8
The ZeroGPU model catalog (https://docs.zerogpu.ai/docs/model-catalog, checked 2026-10-09) lists cached input for qwen3-30b-a3b-fp8 at $0.05 per 1M tokens. Without cache_read_input_token_cost, cached tokens were billed at the full $0.10 input rate
* revert(zerogpu): leave the audio transcription provider set unchanged
Per review, keep OPENAI_AUDIO_TRANSCRIPTION_PROVIDERS as upstream defines it. This PR only covers chat completions, and ZeroGPU now serves /audio/transcriptions anyway, so the local refusal and its test are removed
* fix(zerogpu): add max_tokens so trim_messages can size ZeroGPU history
trim_messages reads max_tokens as the fallback for max_input_tokens and evaluates it eagerly, so without it the lookup raised, got swallowed, and the full history came back untrimmed. max_tokens matches each model's maxTokens in ZeroGPU's models API (https://api-dashboard.zerogpu.ai/api/models, checked 2026-10-10)
* test(zerogpu): record mocked requests on a recorder object
The mock handler appended to a list the caller passed in, which AGENTS.md
forbids. A recorder now owns the captured requests as a tuple
* fix(zerogpu): cap max_tokens at each model's output limit
Each cap is the max_output_tokens ZeroGPU stores for the model in its production models table (read 2026-10-20), the same column its /v1/models serves as max_length. That gives 65536 for deepseek-v4.1-flash, 32768 for gpt-4.1-mini and gpt-oss-120b, 128000 for gpt-5.4-nano and gpt-5.6-luna, 8192 for LFM2.5-1.2B-Instruct and llama-3.1-8b-instruct-fast, and 16384 for qwen3-30b-a3b-fp8. glm-5.3-flash is stored as 32786, which reads as a typo for 32768, so that row uses 32768
trim_messages sizes history from max_input_tokens, so lowering max_tokens leaves it unaffected
* fix(zerogpu): flag prompt caching on glm-5.3-flash
Repeating the same ~4.3K-token prompt against ZeroGPU on 2026-10-20 returned cached_tokens of 4320 on some glm-5.3-flash calls. gpt-oss-120b and llama-3.1-8b-instruct-fast returned 0 on all nine calls, and qwen3-30b-a3b-fp8 never returned cached_tokens, so they stay unflagged. Their cached prices stay as ZeroGPU publishes them, so any cache hit is still billed at the right rate
* fix(zerogpu): set gpt-5.4-nano context window to 272000
ZeroGPU accepts a 271001-token gpt-5.4-nano prompt and rejects 273000 tokens as over the context window (checked 2026-10-21), so max_input_tokens is 272000 rather than the 400000 its catalog lists
* fix(zerogpu): list zerogpu models for zerogpu/* wildcards
zerogpu had no models_by_provider entry, so get_provider_models("zerogpu") returned None and a zerogpu/* wildcard listed no models. zerogpu_models is now filled from the cost map rows like openinfer_models and registered in models_by_provider
* fix(zerogpu): flag gpt-5.4-nano as a reasoning model
ZeroGPU honors reasoning_effort on gpt-5.4-nano (checked 2026-10-21). Without it the model spends 0 reasoning tokens, and with low or high it spends 32 or more. LiteLLM only treated it as a reasoning model by falling back to OpenAI's bare gpt-5.4-nano row, so get_model_info and /model/info reported supports_reasoning as null. The row now carries the flag itself
* fix(zerogpu): flag deepseek, glm, gpt-oss and qwen3 as reasoning models
ZeroGPU's /v1/models marks these four as reasoning models and accepts reasoning_effort on all of them (checked 2026-10-21), but LiteLLM rejected the param with UnsupportedParamsError because their rows had no supports_reasoning flag. glm reports reasoning_tokens under reasoning_effort, and gpt-oss and qwen3 spend hidden reasoning tokens on every reply. The gpt-oss test case now expects reasoning_effort to reach ZeroGPU
* fix(zerogpu): set gpt-oss-120b context window to 128000
ZeroGPU routes gpt-oss-120b across several upstream backends, and at least one of them returns 413 past a 128000-token window, so max_input_tokens uses that as the safe floor. The backends reached on 2026-10-21 accepted up to about 131K prompt tokens, so this only gives up a little headroom on the larger ones
* feat(zerogpu): call ZeroGPU's Responses API natively
ZeroGPU serves /v1/responses for all nine catalog models, so LiteLLM no longer bridges Responses requests through chat completions. Native calls return ZeroGPU's reasoning output items for glm, gpt-oss and qwen3, which the bridge dropped. Checked live on 2026-10-22 through litellm.responses for every model, non-streaming and streaming, plus a gpt-oss tool call. LFM2.5-1.2B-Instruct streams correctly now that ZeroGPU's text_generation lane emits Responses SSE
* fix(zerogpu): stop inferring the zerogpu provider from api_base alone
Listing https://api.zerogpu.ai/v1 in openai_compatible_endpoints rerouted bare model names on that api_base from openai to zerogpu. A bare gpt-5.6-luna call with verbosity used to return 200 and started raising UnsupportedParamsError. Drop the entry and its api_base-alone test, so zerogpu is only selected through the zerogpu/ prefix
* fix(zerogpu): price gpt-6-luna, llama-3.2-3b, LFM2.5-1.2B-Thinking and qwen3.6
A zerogpu/* wildcard served these four chat models with no cost-map row, so their calls logged $0 spend. Add rows at ZeroGPU's catalog prices from https://api-dashboard.zerogpu.ai/api/models, with context windows and output caps from ZeroGPU's models table. LFM2.5-1.2B-Thinking has no cached price, like LFM2.5-1.2B-Instruct. gpt-6-luna is flagged supports_reasoning because LiteLLM already routes gpt-6 names as reasoning models. Checked live on 2026-10-28, all four now report nonzero response_cost
* fix(zerogpu): drop the retired max_tokens field from the zerogpu rows
Upstream #48471 retired max_tokens from the cost map in favor of max_input_tokens and max_output_tokens, and test_no_row_carries_the_retired_max_tokens_field failed on all 13 zerogpu rows. Every row already carries the same value in max_output_tokens, so nothing is lost. The zerogpu limits test now checks that each row's output limit fits its context window
* chore(zerogpu): regenerate provider endpoint support for zerogpu
Upstream #48619 generates provider_endpoints_support.json from the cost map and checks it in CI, and the hand-written zerogpu entry no longer matched. Regenerated with scripts/generate_provider_endpoints_support.py. Only the zerogpu entry changed, and it now matches what the generator emits for other JSON-registry providers such as openinfer
* fix(zerogpu): flag tools, json_schema and prompt caching on gpt-6-luna
ZeroGPU serves function calling, tool_choice, json_schema response_format and prompt caching on gpt-6-luna, so set supports_function_calling, supports_tool_choice, supports_response_schema and supports_prompt_caching on its row in both cost maps. Checked live on 2026-10-28, a weather tool call came back with the right arguments, json_schema output validated, and repeated prompts reported 3260 of 3263 tokens cached
* fix(zerogpu): set qwen3.6-35b-a3b context to 262144 and flag its tools and reasoning
ZeroGPU accepts up to 262,144 input tokens on qwen3.6-35b-a3b and serves tools, tool_choice, json_schema response_format and reasoning_effort on it. Raise max_input_tokens from 16384 and set supports_function_calling, supports_tool_choice, supports_response_schema and supports_reasoning in both cost maps. Before this, LiteLLM rejected tools and reasoning_effort for the model. Checked live on 2026-10-28, a 209,905-token prompt succeeded, an oversized one was refused with a 262144-token limit, and tool calls, json_schema output and reasoning_effort all worked
* fix(zerogpu): flag vision on gpt-6-luna and qwen3.6-35b-a3b
ZeroGPU reads image inputs on gpt-6-luna and qwen3.6-35b-a3b, so set supports_vision on both rows in both cost maps. Checked live on 2026-10-28, both models named the animal in an image_url photo, and the image showed up in their prompt token counts
* test(azure): audit #47348
Add proxy integration tests for #47348's cached Entra token providers and
off-loop token refresh, and fix the token herd the audit found: once the
fetch moved to a worker thread, concurrent cold requests each fetched their
own token. Every cached bearer token provider now serializes its calls so
the first fetch fills the cache and the rest reuse it
* fix(azure): build each cached Entra token provider once under concurrent first calls
lru_cache does not hold back concurrent misses, so worker threads that miss together each built their own credential and token provider, and each fetched its own token. A lock around each provider cache makes the first caller build it and the rest reuse it
* fix(azure): share Entra token fetches only while no valid token is held
Concurrent callers now share one result (or one error) from the cold
fetch, while callers inside the refresh window keep the still-valid
token. The oidc/azure provider build and the client-assertion exchange
are single-flighted per audience, and the shared helpers no longer call
subscripted generic aliases, which crashed on Python 3.10.
* fix(azure): refresh Entra tokens once while other requests keep the valid token
MSAL treats a token with less than five minutes left as expired, so every
request in that window fetched a new one. Now one request refreshes while
the others keep using the still-valid token.
Requests that waited on a fetch that failed retry once through a fresh
fetch instead of all failing with the first error.
Credentials that only implement get_token keep working, and OIDC token
exchanges are cached per scope.
* fix(azure): bound Entra token waits during an outage
Requests waiting on a failed token fetch now retry only when the fetch
before it succeeded, so during an Entra outage later bursts fail after one
fetch instead of two.
Requests during a refresh only get the held token while it has more than
60 seconds left, and a failed refresh falls back to that token instead of
failing its request.
* fix(azure): share one retry per failed Entra token fetch and raise on a failed refresh
Waiters of a failed fetch now retry the flight they joined, so a fast-failing Entra makes one retry instead of one per waiter. A failed refresh raises again instead of returning the held token, which Postgres token auth read as a fresh one. The DefaultAzureCredential fallback reuses one provider per scope.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(responses): audit #47363
Adds integration cells for the input token count paths #47363 changed: the provider id of a managed file and the 403 for another team's file on OpenAI, file_search and additional_tools parity plus cache-mark stripping and untranslatable tools on Claude, and the Bedrock runtime count leaving out the options it refuses.
* fix(count-tokens): reject a lone tool with no chat equivalent before the Claude count call
A Claude input token count whose only tool was mcp or computer_use_preview skipped the pre-call check and went to the provider, so the caller got the provider error instead of the 400 naming the tool. The check now runs whenever a tool would get a placeholder name.
* fix(responses): count a lone Anthropic toolset and name the uncountable tool
The lone-tool 400 now fires only for OpenAI Responses built-in tool types, so a nameless Anthropic toolset still counts. The error says LiteLLM cannot count that tool type for the model. Adds a Bedrock unit cell for the lone-tool 400 and pins the runtime wire test's tool shape per request type.
* fix(responses): take the uncountable tool types from the OpenAI SDK and keep the region hint off LiteLLM's 400
The Responses built-in tool types now come from openai's ToolParam union, so a type the SDK adds is refused before the count call too. The AWS region hint no longer lands on an error LiteLLM raised before calling the provider. Tests cover two more lone types and the hint.
* fix(responses): keep LiteLLM's own list of uncountable tool types and pass a non-string tool type through
* fix(responses): count tool_search and namespace tools the way Claude chat inference sends them
The count refused a lone tool_search with LiteLLM's 400 and sent a lone
namespace raw or squashed it into an empty tool, while Claude chat
inference maps tool_search to Anthropic's tool search tool and unpacks
a namespace into its function tools. Both now go through the
deployment's own map_openai_params, so the count sends what inference
sends, and tool_search leaves the list of uncountable types.
* fix(bedrock): count a Claude request with its deployment's own chat translation
Bedrock invoke and Mantle inference turn a chat tool_search tool into Bedrock's
tool search tool, but the Bedrock counter always translated with the Anthropic
config, so the count answered 400 for a request inference serves.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(mistral): replay assistant reasoning as a Mistral thinking chunk
When a client sends back an earlier Mistral assistant turn, its reasoning
(reasoning_content, thinking_blocks, or Mistral's own thinking chunk in
content) now reaches Mistral as a ThinkChunk inside content. The top-level
LiteLLM fields stay stripped because Mistral rejects them with 422.
* fix(mistral): replay reasoning only to Mistral models that accept it
Mistral answers 400 to a thinking chunk on models without reasoning support,
so a mid-chat model switch, a fallback, or Vertex AI Mistral broke once the
replay was added. Gate the replay on the same check reasoning_effort uses and
keep dropping the reasoning for the other models, as before.
* fix(mistral): drop native thinking chunks for models without reasoning and mark the devstral, code agent, small and leanstral aliases as reasoning
* fix(mistral): keep the leanstral rows as on main, since their reasoning support cannot be checked live
* fix(mistral): send only the text items of a native thinking chunk, so a reference or image item no longer gets Mistral's 422
* fix(mistral): keep closed false only on the last chunk of a final prefix turn, so an earlier turn no longer gets Mistral's 400
* fix(mistral): drop closed false when merged text follows the thinking chunk, so a prefix turn with text first no longer gets Mistral's 400
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(chatgpt): derive Responses session id from prompt_cache_key
ChatGPT partitions the Codex prompt cache by the Responses session-id
header (the Codex CLI sends a stable per-conversation key derived from
prompt_cache_key). The provider only consulted explicit session ids
and otherwise stamped a per-request uuid4, so no turn ever reused the
prompt cache -- 0% cache hits on repeat turns, an order-of-magnitude
quota/cost multiplier on subscription accounts.
Resolution order after this change:
explicit session config (session_id / litellm_session_id /
metadata.session_id) > prompt_cache_key (sanitized with
_safe_header_value) > litellm-internal ids > uuid4.
Two guards, mirroring get_fireworks_session_id:
- proxy-generated session ids (missing_session_id: "generate") are
per-request and are skipped, checked in both "metadata" and
"litellm_metadata" (LITELLM_METADATA_ROUTES, responses included)
- internal per-request ids (litellm_trace_id / litellm_call_id) lose
to a stable conversation key
Supersedes the body-key forwarding of #37630: a random session header
overrides the body key on the provider side, so the header must carry
the caller's key. Related: #29978, #29993.
Verified against a real ChatGPT Codex account: stable session id ->
97-100% cached on repeat turns; random per-request -> 0%.
* style: ruff format
* fix(chatgpt): address review - hash unsafe cache keys, move tests, drop comments
- hash prompt_cache_keys containing header-unsafe chars instead of
mapping them through _safe_header_value, whose replacement char could
collapse distinct keys onto one cache shard and serialize unrelated
requests (greptile review)
- strip the explanatory comment blocks per AGENTS.md comment policy
- move the tests into the existing mapped responses transformation test
file per AGENTS.md test-placement rule
* fix(chatgpt): keep session typing and test diff scoped
* fix(chatgpt): hash surrogate cache keys, type session params, flatten tests
* test(chatgpt): record upstream requests with a mock instead of appending to a list
* fix(chatgpt): keep proxy-mapped session headers ahead of body session_id and ignore non-string cache keys
* fix(chatgpt): hash cache keys with leading or trailing whitespace so the session header stays valid
* test(chatgpt): type the session id test parameters
* fix(chatgpt): keep trace ids copied into litellm_session_id from outranking prompt_cache_key
* fix(chatgpt): mark session ids copied from trace_id so explicit ones keep precedence
* fix(chatgpt): keep the trace-id session marker out of provider params
The marker now uses the _litellm_ internal prefix, so bridges that pass litellm_params back into completion() drop it instead of sending it as a provider param
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(huggingface): cover TEI rerank truncate passthrough and server default
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(huggingface): stop forcing truncate=false on TEI rerank and pass truncate params through
Co-authored-by: Luke Schaefer <225965040+luke-e-schaefer@users.noreply.github.com>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(huggingface): keep TEI truncate options provider-specific
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(huggingface): use MockTransport for rerank truncation
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>
Co-authored-by: Luke Schaefer <225965040+luke-e-schaefer@users.noreply.github.com>
* fix(streaming): exact-match the SSE [DONE] marker in BaseModelResponseIterator
* test(streaming): cover reasoning text and error frames that mention [DONE]
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(router): avoid assistant prefill in Responses fallbacks
* fix(router): keep prefill for Claude fallbacks that take it and skip constrained continuations
A Responses fallback to a Claude model that takes assistant prefill now gets
the partial response as prefill again, so it continues mid-word without the
prompt-shape drift. Every other fallback keeps the request to continue.
A stream cut after text no longer continues when the request constrains its
output (a json_schema or json_object text format, or a forced tool_choice),
since a continued JSON document or tool call would not parse. The client gets
the primary's error, as on the chat path.
The fallback's re-sent leading space at the seam is dropped, and the prefill
shaping runs per deployment after the encrypted reasoning strip, so stripped
reasoning is never replayed.
* fix(router): guard SDK text_format, narrow forced tool_choice, read Claude base_model, and cap Responses continuation output
Cut Responses streams that asked for structured output through the SDK's text_format now surface the primary's
error instead of stitching two models' JSON. Only a required or named tool_choice counts as forced, so allowed_tools
in auto mode and hosted tool choices still continue. A Claude deployment named only through base_model keeps its
prefill continuation, and the continuation's max_output_tokens is lowered by what the cut stream already emitted.
* fix(router): quote the partial's ending in the continue request and surface the primary's error after a cut tool call
* fix(router): stop Responses fallbacks from continuing content-filter cuts and client-run tool calls
A content filter that cuts a /v1/responses stream mid-answer now returns its error, as on chat, instead of letting a content_policy_fallbacks backup finish the filtered text. Shell, apply_patch, computer, local_shell, and MCP approval items count as a started tool call. Chat continuation goes back to treating any tool_choice object as constraining, as on main
* fix(responses): count a stream of tool arguments as a started tool call
* fix(router): let Responses fallbacks continue under allowed_tools auto and after server-run tool calls
A Responses tool_choice of allowed_tools in auto mode no longer blocks the continuation, since the model may still answer in text. Hosted MCP calls and the bridge's server-run tool search no longer count as a started client tool call.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(anthropic): let drop_params drop speed on claude-opus-4-7 after Anthropic removed its fast mode
* test(anthropic): cite the Claude Code fast-mode page for the Opus 4.7 removal date
* test(anthropic): cover the drop_params speed drop for claude-opus-4-7
* test(anthropic): type the model parameter of the speed drop test
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(bedrock): retry without strict when Bedrock rejects strict tool schemas under drop_params
* test(bedrock): cover the strict tools 400 retry under drop_params
* fix(bedrock): anchor the strict tools rejection match and strip toolSpec strict
* test(bedrock): return captured requests from the strict tools test helper
* fix(bedrock): retry without strict only when a tool sent strict to Bedrock
A strict-limit 400 on a request whose tools carry strict: false, or on a model litellm never sends strict to, was sent again with an identical body. The fallback now takes the model and returns nothing unless a tool puts strict on the wire. Also types the fallback test parameter and records the mock transport's requests on the recorder instead of a mutated list
* fix(bedrock): decide the strict retry on the profile's base model and retry grammar timeouts
An application inference profile with base_model sends strict on the wire, so the retry check now resolves the
capability model the way the request transformation does. Bedrock's "Grammar compilation timed out" 400 is a
strict-grammar rejection too, so it takes the same single retry under drop_params.
* test(bedrock): annotate the strict-tools retry test helper's values as Final
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(proxy): audit #46815
Add integration tests for every count route #46815 changed, and count a Claude
subscription caller's tokens with their forwarded token when the deployment has
no configured key, which #46815 turned into a 401
* test(proxy): accept the forwarded subscription token in the count endpoint mocks
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat: add aiand provider
* feat: expose aiand provider in admin UI
* feat(aiand): add ai& model pricing to the cost map
* feat(aiand): list ai& models for aiand/* wildcards and mark /v1/messages supported
* feat(aiand): send Anthropic /v1/messages to ai& natively
* fix(aiand): drop the retired max_tokens field and regenerate endpoint support
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Fixes#46977: Cohere chat API reference expects stop_sequences to be a list of strings. Convert string stop parameter to a single-element list for CohereChatConfig and CohereV2ChatConfig while preserving lists.
Co-authored-by: Praising Harris Ratnam <praisingharris@gmail.com>
* fix(tavily): return raw_content, score, favicon and the response envelope from Tavily search
Tavily search dropped every field except title, url and content, so
include_raw_content, include_answer, include_images and include_favicon
reached Tavily but their output never reached the caller. Each result now
carries raw_content, score and favicon when Tavily returned them, and the
top-level fields (answer, images, request_id, usage, ...) come back as
provider_metadata, same as the Exa fix in #47423
* test(tavily): move the Tavily search wire test's sample data to module constants, like the Exa test
* feat(morph): support open-weight chat models and cached input pricing
* chore(morph): preserve existing catalog number formatting
* test(morph): include provider regressions in CI shard
* fix(morph): satisfy catalog and test package checks
* fix(morph): declare abliterated model output limit
* fix(chatgpt): send system input items to Codex as developer
The Codex backend rejects {"type": "message", "role": "system"} input
items with 400 "System messages are not allowed", while it accepts the
same item with role=developer. The chat completions bridge emits that
shape for every non-leading system message, and typed Responses API
clients send it directly, so both paths failed. Rewrite system items to
developer in place, which keeps their position and content.
Fixes#48562
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(chatgpt): annotate system-item rewrite with read-only types
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(chatgpt): rewrite system items on the typed input before the base transform
Keeps the rewrite on known types so basedpyright sees no new unknowns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
An OpenAI or Azure SDK client litellm built and evicted from the client
cache is usually garbage collected inside the closer's 900 second grace
window, so the closer's weak reference dies and the client's aiohttp
session is never closed. LiteLLMAiohttpTransport.__del__ now hands an
owned, still-open ClientSession to _close_recycled_session, which already
covers the current-loop, foreign-loop and closed-loop cases.
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: BangerSoul <bangersoul@users.noreply.github.com>
* fix(proxy): build proxy-only failure logs from an allowlist and ignore request-time GCS credentials
* fix(proxy): drop caller metadata from proxy-only failure logs
metadata and litellm_metadata leave the proxy-only failure-log allowlist, so a rejected request can no longer point its failure log at its own Langsmith or Langfuse host through nested metadata.
* fix(proxy): keep allowlisted metadata in proxy-only failure logs
* test(proxy): isolate proxy-only failure log metadata tests
* fix(logging): ignore non-list request tags when building the logging payload
A guardrail-rejected request with metadata tags set to a non-list value crashed
the standard logging payload, so its spend row lost model attribution.
* fix(proxy): keep router attribution and drop unauthenticated tags in proxy-only failure logs
A request rejected before routing now logs the configured deployment and model group even when its metadata or litellm_metadata carries its own values. Requests rejected at auth no longer write caller tags into failure logs and tag spend.
* fix(proxy): always strip client guardrail info and mirror failure attribution into litellm_metadata
allow_client_pricing_override no longer keeps a caller-supplied standard_logging_guardrail_information on the chat or pass-through paths, since its guardrail_cost is billed. Proxy-only failure logs now also carry key and team attribution in litellm_metadata.
* refactor(proxy): keep the proxy-only failure-log fixes inside the type gates
Types the picked helpers so basedpyright and the type-discipline gate stay at main's counts. No runtime change.
* fix(proxy): keep the WS frame guardrail-info strip and string interaction tags as sent
The guardrail-info key left the pricing strip list, so the Responses WS frame guard now drops it by name. get_request_tags no longer raises on string tags, so background interaction polling checks for them before keeping them as sent.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
