* fix(bedrock): forward x-pass- headers on the /bedrock/model pass-through routes
The /bedrock/model/{model}/invoke, invoke-with-response-stream, converse and
converse-stream routes built the pass-through call from the request body, so
x-pass- headers never reached Bedrock. The processor now hands the caller's
x-pass- headers to the pass-through route, keeps them out of the logged body
snapshot, and on Converse drops the invoke-only guardrail headers when the
deployment sets guardrailConfig and the request metadata header when the proxy
owns request metadata
* fix(bedrock): bill the served service tier on invoke pass-through and scope x-pass- forwarding to the Bedrock routes
* test(bedrock): assert forwarded x-pass- headers on the request Bedrock receives
* test(bedrock): gate the router-model x-pass- forwarding on the request Bedrock receives
* fix(bedrock): bill a caller's forwarded invoke guardrail and keep x-pass- headers out of guardrail vendor payloads
* fix(bedrock): bill a caller guardrail on every invoke model and replace a trace value Bedrock rejects
* fix(bedrock): keep guardrail traces out of chat responses, forward the trace header as sent, and drop credential x-pass- headers
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(proxy): count a prompt-only call_endpoint request with the provider on every counter
POST /utils/token_counter?call_endpoint=true with only a prompt sent the
prompt to the provider count API on Gemini and Vertex only. Anthropic,
OpenAI, Bedrock and Azure AI (Claude) deployments answered a local
estimate instead. Every counter now gets the prompt as one user message,
so a provider error surfaces the same as the messages form, and the
framed stand-in after a failed provider count frames the prompt as one
user turn, matching the messages form
* fix(proxy): count a call_endpoint prompt like a user message on providers without a count API
* fix(proxy): count a call_endpoint prompt as a turn only on chat deployments
A prompt-only /utils/token_counter request with call_endpoint=true now goes
to the count API as one user message only when the deployment serves chat
(mode chat, responses, or unknown). Embedding rows and models with no
deployment keep the raw local count.
The Anthropic, Azure AI, and OpenAI count calls now honor the deployment's
timeout, and a count error with an empty body names its status code.
* fix(proxy): fall back to router_settings.timeout for count calls and answer 408 on a count timeout
Count calls to Anthropic, Azure AI, and OpenAI now pick their timeout the way
a chat call to the same deployment does: the deployment's timeout, else
router_settings.timeout, else request_timeout. A count call that times out
answers 408 instead of a 500 processing error.
* fix(proxy): count a prompt as a turn only where a count API exists and give count calls chat's timeout order
* fix(proxy): read the count call timeout through the router settings snapshot
The token counter now asks llm_router.settings().timeout_for() for the chat timeout order instead of a new Router member, so the router surface does not grow
* fix(router): answer timeout_for on every router settings snapshot
The proxy reads the count call timeout through llm_router.settings().timeout_for(), but only the Rust core's snapshot had that method, so a plain RouterSettings (the unit tests' router stub) raised AttributeError. timeout_for now lives on RouterSettings and reads the core's _core_deployment_call_timeout, like the snapshot's other members
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(decisions): keep a same-model deployment's false box-question flag off the shared cost-map row
* fix(router): take a withdrawn supports_box_questions true off the shared row on update and delete
A deployment's true reached the shared backend row, but a later false no longer
overwrites it, so updating or deleting that deployment left box questions on for
its unflagged siblings until restart. The row now gets the catalog's value back
unless another served deployment of the model still sets true.
* fix(router): count every live router and set_model_list when withdrawing a shared box-questions true
* fix(router): match box-questions siblings by their cost-map row and withdraw on retire
A sibling spelled with its provider registers onto the same litellm.model_cost row, so the withdrawal now compares rows instead of backend keys. Retiring a router takes back the true its deployments put on rows no other live router still sets. Withdrawals are deduped per backend so a large set_model_list stays cheap.
* fix(router): take back a box-questions true an aborted set_model_list registered
A set_model_list that fails partway rolls back the deployment list, but a true the
aborted build already wrote to the shared cost-map row stayed there, so unflagged
same-model deployments kept sending box questions. The rollback now withdraws the
attempted entries too
A withdrawal now only rewrites rows that still hold true, so an entry the failed build
never reached gets no row written for it, even on a model the catalog does not list
The delete wire test now uses a decisions-mode catalog backend whose catalog lists no
box questions, since an unflagged deployment only reads the shared row for a
decisions-mode model
* fix(router): take back an env-named box-questions true a failed set_model_list registered
The rollback took flags back from the raw model_list entries, so a deployment whose model came from os.environ/ was looked up under the reference instead of the model it registered onto, and its true stayed on the shared row
It now takes them back from the entries the failed build reached, whose os.environ/ values are already read
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(responses): replay Claude tool search on follow-up turns
Claude tool search on /v1/responses now returns a tool_search_output after each tool_search_call, streamed and not, listing the tools the search loaded as the caller sent them
A stateless follow-up sends the search back to Anthropic as the server_tool_use and tool_search_tool_result Claude produced, and the Anthropic request drops a replayed search whose loaded tools are no longer sent, so a changed tool list no longer answers 400
The guardrail write-back for SDK output items reads part text through a typed helper, so the new output item type adds no attribute errors
* fix(responses): emit Claude's tool search output in stream order
The streamed tool_search_output item was announced at the end of the stream with an output index reserved earlier, so output_item.added events arrived as 0, 2, 1 and the OpenAI SDK's stream helper raised IndexError. The item is now emitted when Claude's search result arrives.
* fix(anthropic): keep the caller's tool names in tool search results
A tool name Anthropic cannot take gets rewritten on the request, so Claude's
tool_search_tool_result named the rewritten tool. The response now maps those
tool references back to the caller's names, streamed and not, and a replayed
search result maps them forward again on the next turn.
* fix(responses): replay Claude's tool search only on routes that run it
A follow-up sent to a route that cannot run tool search, a Claude model without it included, now leaves out the replayed search the way it did before, instead of sending Claude search blocks that model cannot take. The stateless gate and the stored-turn filter now use the same check that already decides whether the tool_search tool itself is kept
A streamed tool_search_call now gets its output_item.done when Claude's search result arrives, so it finishes before its tool_search_output. Clients that build history from done events, like Codex, kept the output ahead of its call
* fix(responses): keep a replayed search when the caller sent Anthropic's own search tool
A Claude model the cost map does not flag for tool search still runs it when the request carries Anthropic's own tool_search_tool_regex or bm25 tool, which is how turn 1 already decides to keep it. Both replay gates now follow that rule, so a stored follow-up on such a model keeps Claude's search as it did before, and a stateless one replays it
* fix(responses): replay a BM25 tool search under its own name
* fix(responses): keep a replayed tool search visible to guardrails and prune only missing tools
A replayed Claude tool search now reaches pre-call guardrails and the Purview DLP scan with its
search arguments, a stale search drops only the tool references the follow-up no longer sends,
and the stored previous_response_id path keeps the base gate for routes the cost map does not flag
* fix(responses): replay Claude's search on a stateless follow-up only when it still sends a search tool
A store false follow-up that sent no search tool still got the search replayed, and Claude then said it called a tool it no longer had
* fix(responses): replay a stored Claude search only when the follow-up still sends a search tool
A previous_response_id follow-up that sent no search tool still got the stored search, the same gap a store false follow-up had. The stale search pruning now logs the tools it leaves out at debug level.
handler.py reads the reported speed from the usage chunk, and the guardrail write-back skips tool search outputs, so the branch adds no type errors.
* fix(responses): leave out a stored Claude search kept in server_tool_calls
Main stores Claude's server-side tool search on the assistant message in
provider_specific_fields.server_tool_calls and tool_results, not in
tool_calls. The previous_response_id path only looked at tool_calls, so a
follow-up that sent no search tool still replayed the stored search.
The session handler now finds the search in either place and drops it from
both server_tool_calls and tool_results when the follow-up cannot run a
search.
* test(responses): give the replayed search guardrail fixture its own call id
The guardrail test reused toolu_01Weather, an id that streaming and bridge
tests already put in the process-wide TOOL_CALLS_CACHE. When one of those ran
first in the same worker, the guardrail write-back took its rebuild path and
the test failed for a reason unrelated to the replayed search.
* fix(anthropic): keep Claude's replayed tool search on renamed tools and stored follow-ups
A search-only assistant turn now gets its search result renamed to the sanitized
tool names too, so the stale-tool trim no longer drops it. Stored follow-ups on
Claude routes replay the search as base did, trimmed to the tools still sent, and
Bedrock invoke sends the tool search beta whenever the history holds a search
* test(responses): pin that guardrails scan a replayed Claude search on a follow-up without a search tool
* fix(anthropic): keep Anthropic's tool names in chat tool search results
A chat history whose search result used the request's tool names (weather.get) got
Anthropic's 400 when replayed on a proxy still on the old code, during a rolling deploy or a
rollback. Chat completions now returns Claude's search result as main does, and the Responses
bridge maps a rewritten name back to the request's tool when it builds tool_search_output.
That also fills tool_search_output on Bedrock invoke streams, which carry no name map
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Google AI Studio answers generationConfig.candidate_count "" with a 400 before it generates anything, so the truncated 200 this case scripts is a reply Google never gives. The other malformed count and cap cases stay.
* fix(router): merge default metadata under the request's instead of replacing it
* fix(router): join default metadata tags after the request's tags
A router default_litellm_params.metadata.tags list replaced the request's tags, so the deployment and credential tags the router adds never reached the logged request tags. Default tags now follow the request's tags, skipping ones already present, like the proxy's key tag merge
* fix(router): keep masking a router default turns on
Merging default metadata under the request let a caller's mask_input or mask_output false switch off masking an admin set in default_litellm_params.metadata. A masking flag the default turns on now stays on, and a caller can still turn masking on. Also gives this PR's test suppressions a reason and adds a missing blank line
* fix(router): keep the router's cache namespace and drop malformed caller tags and dicts
A caller's metadata.redis_namespace no longer beats a router default, so it
cannot pick another tenant's cache partition. A caller tags value that is not a
list, or a non-dict value under a dict default, is treated as absent the way the
proxy's merge_tags treats malformed tags, so the router defaults still apply.
* test(router): check that mask_output stays on when the router default turns it on
* fix(router): keep a router default guardrail whatever guardrails the request lists
* fix(router): keep the admin's settings for a router default guardrail when the caller lists it
* test(router): check that a router default guardrail entry keeps the admin's settings
* fix(router): keep a router default private Langfuse trace private whatever the caller sends
* fix(router): keep a router default function such as langfuse_masking_function over the caller's value
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* test(integration): cover claude-fable-5-2 on Vertex AI and Bedrock Converse
Scripted-upstream wire tests for the Fable capability flags on both routes, the
vertex_ai and us.anthropic cost cases every other Fable row has, Fable 5.2 in
BEDROCK_CONVERSE_MODELS, and bedrock/claude-fable-latest repointed to 5.2 like
the Anthropic and Vertex aliases already are.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(cost-map): add GovCloud Bedrock rows for claude-fable-5-2
bedrock/us-gov-west-1 and bedrock/us-gov-east-1 rows copy the Fable 5.1 field
set with the GovCloud rates from the Bedrock price list.
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): cover claude-fable-5-2 capability flags and prompt-cache pricing over Anthropic direct
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): replace the fable-5-2 e2e suite with scripted-upstream wire and cost cases
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): add claude-fable-5-2 wire and cost cases
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>
* fix(completion): accept a tool_choice that forces an OpenAI hosted tool
litellm.completion rejected tool_choice={"type": "web_search"} before any request, while /v1/responses took the same choice. The chat validator now forwards a hosted tool type unchanged and a provider gate turns it into a 400 off openai, azure and litellm_proxy, where the choice would otherwise be dropped or mangled.
Fixes#50566
* fix(utils): keep a non-string tool_choice type a 400, not a TypeError
A list or dict in tool_choice["type"] is unhashable, so the hosted tool lookup needs the string check first. The bridge test that used file_search as an untranslatable choice now uses a type that really is unknown.
* refactor(completion): forward any Responses API tool_choice type instead of keeping a hosted tool allowlist
Chat completions now accepts any tool_choice dict with a string type it has no shape for and sends it as is on openai, azure and litellm_proxy, where OpenAI validates the type itself (a typo is still a 400, from OpenAI). The provider gate keeps the 400 everywhere else. The HOSTED_TOOL_CHOICE_TYPES list is gone, so a new OpenAI hosted tool needs no change here.
* Revert "refactor(completion): forward any Responses API tool_choice type instead of keeping a hosted tool allowlist"
This reverts commit 5a6d5a8e08.
* fix(utils): add tool_search to the hosted tool_choice types
OpenAI lists it among the supported tool_choice.type values (live error on 2026-11-08), alongside the ones already in the set.
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): surface mid-stream upstream errors that arrive without an event type
Bedrock's native Responses route can end a stream with a bare {"error": {...}}
chunk. It parsed as a typeless GenericEvent and crashed the iterator with
AttributeError, so clients saw that instead of the upstream error. Such chunks
are now read as error events, which raise the mapped upstream error and reach
the client as response.failed.
* refactor(responses): inline the untyped error chunk check to keep LIT001 at the base count
* fix(anthropic): preserve streamed tool argument fragments verbatim
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(anthropic): send the schema-less stream test the null properties form providers accept
* fix(anthropic): unwrap values when a schema-less stream sends its whole tool input in one piece
* fix(anthropic): unwrap only a one-piece values input and never emit Infinity in schema-less streams
* fix(anthropic): keep JSON mode answers verbatim when they cannot be written back as strict JSON
A tool input holding a number past a float or an integer past Python's digit limit now passes through as the model wrote it on the stream and non-stream paths, instead of coming back with Infinity or raising. One-piece values rows for false, null, a plain object and too-deep nesting pin the remaining unwrap branches.
* fix(anthropic): reject non-finite floats through pydantic instead of a raising parse hook
* fix(anthropic): keep split JSON mode answers whole when a field cannot be written back as strict JSON
* fix(anthropic): keep out-of-range numbers as the model wrote them in JSON mode answers
Decode tool input with parse hooks that keep a number past a float or an integer past the digit limit as its source text, and write it back unchanged, so held, wrapped and name and value split answers stay whole instead of falling back to raw blocks.
* fix(anthropic): keep numbers that round to zero as the model wrote them in JSON mode answers
* refactor(anthropic): encode JSON mode answers without a recursive helper
* fix(anthropic): decode deeply nested JSON mode answers without slow recursive validation
* fix(anthropic): type the decoded JSON mode answer as object and test non-object answers
json.loads output now goes through the existing TypeAdapter(object), which returns it unchanged, so no Any reaches the object guard. New stream and non-stream cases check that a list, string or number answer comes back as written.
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(proxy): preserve Responses SSE event framing
* style(proxy): format Responses stream terminator condition
* fix(proxy): validate Responses guardrail event names
* fix(proxy): keep Responses first-event errors and keepalive failures in Responses framing
* test(proxy): use controlled clocks for Responses keepalive failures
* fix(proxy): end late Responses failures with an error event the SDK raises on
Errors that arrive after keepalive pings, and guardrail error strings yielded
mid-stream, now end /v1/responses with a named `error` event carrying the
same error object chat sends, so the OpenAI SDK still raises APIError with a
clean message and provider_specific_fields. Background polling passes the
same framing and records an `error` event as a failed response. The first
data line of a frame is found by splitting on newlines only, so a U+2028 in
an error message no longer hides a first-event error.
* fix(proxy): match OpenAI's Responses error event and stream late passthrough blocks as block responses
* fix(proxy): keep the reason when a background Responses request fails
A background poll of a request a post-call guardrail blocked saved a
null code and message, because the guardrail's error event nests them.
A request whose first chunk was an error polled as completed with no
error. Both now poll failed with the same code and message a late
failure gets.
* test(proxy): type the first-chunk poll test parameters
* fix(proxy): keep first-chunk error detection to lone error frames and handle string poll errors
* fix(proxy): keep the HTTP status on late Responses errors and end Responses streams on error events
A late error body without a code now takes the response's status, an upstream error event counts as the stream's terminal event, and a Responses stream that fails before its first chunk ends with an error event instead of data: [DONE].
* test(proxy): type the first-chunk failure test's parameters
* fix(proxy): keep cursor chat framing and JSON-encode Responses hook errors
* fix(proxy): encode non-dict Mapping Responses chunks as JSON objects
* fix(proxy): encode mappings and models nested in Responses stream chunks as JSON objects
* test(proxy): freeze the hook-output fixture
* fix(proxy): keep masked mappings and odd keys safe in Responses stream frames
Only read-only dict views and pydantic models nested in a hook chunk or error event expand into JSON objects. Other mappings, such as httpx.Headers, keep their own string form so masked values stay masked. Keys JSON cannot take become strings, and a chunk JSON still cannot encode goes out as the plain data frame it was before.
* fix(proxy): never fail a Responses stream frame on a value JSON cannot encode
Only dicts and read-only dict views expand into JSON, at the top level too, so a header mapping yielded as a whole chunk keeps its masked string form. Any failure while encoding a chunk falls back to the plain data frame the base sent, and an error event that cannot encode still goes out with its code and message, so a late block never drops the connection.
* fix(proxy): keep what a nested value's repr hides out of Responses stream frames
* fix(proxy): print values in Responses stream frames the way each base path did
* fix(proxy): send string and number subclasses whose repr shows their value as that value in Responses stream frames
* fix(proxy): print container subclasses and SDK models in Responses stream frames the way base did
* fix(proxy): read hook container storage the way repr does and send error-event models as str()
* fix(proxy): frame streamed Responses reads and send unreadable hook values as base did
Track pre-framed and raw SSE frames so a late failure never follows a terminal event, accept CRLF and CR line breaks in a first named error, frame GET /v1/responses/{id}?stream=true as Responses events, print a whole hook proxy chunk once and fall back to its text, ignore a non-integer status_code in a hook error frame, read a late-block dict subclass through items() when that view differs from its storage, and keep the order and inline form base's repr printed for nested mappings
* fix(proxy): match base on Responses stream edge cases
Read a decimal-string or integral-float status_code as its integer again, so
"429" and 429.0 still map to rate_limit_exceeded. Print each hook value's repr
once and reuse it, so a repr that changes between calls cannot reach the
client. Track only pre-framed sequence numbers from 0 to 2**63-1, so a huge
one cannot break the failure event. Build nested JSON with one frame per level
again, which restores the nesting depth sent as JSON. Count a pre-framed frame
as terminal from its data type, and a named error frame only when its data
carries an error, the way the OpenAI SDKs read them.
* fix(proxy): print hook proxies once and keep base's chunk tracking
Read-only mappings in hook chunks now emit the text from their first print, hook
chunks are read without their error field, only a non-empty error object makes a
framed error event terminal, and framed sequence numbers stay below 2**63-1 while
hook chunks keep base's unbounded tracking
* fix(proxy): print each read-only mapping item once and poll framed error events as failed
A MappingProxyType in a hook chunk or error detail now expands from the
mapping it wraps, so each item is printed once, the way base printed it.
A value whose repr changes between calls stays masked, a value whose repr
works once is kept, and deep nesting keeps one stack frame per level
The background reader takes each frame's type from the same terminal rule
the stream uses, so an event: error frame with an error object polls
failed even when its data names another type or none
* fix(proxy): print self-holding hook values once and poll how the stream ended
A container that holds itself now goes out as the marker its repr prints, like {...}, [...] or mappingproxy({...}), and the dict behind a read-only mapping joins the cycle check, so no item is printed a second time. A dict whose item repr adds a key keeps going after the items already shown, the way dict's repr does, instead of printing the whole dict again
A whole read-only mapping chunk that cannot go out as JSON is printed once by the formatter, so a print that fails fails the stream as it did on the base
The background poll reads the event type from the block it parses, as on the base, and keeps the first error frame the stream treats as terminal anywhere in a chunk. That frame fails the poll only when no response.completed, response.failed or response.incomplete follows it
* fix(proxy): name hook chunk events from the sent JSON and keep polls on frames they cannot read
A mapping chunk now takes its event: line from the JSON it sends, so a masked type value never reaches that line, and a chunk the event model cannot read goes out as a plain data frame instead of failing the stream
A dict whose item repr deletes or swaps keys keeps printing every key not yet shown, and an error code past the integer digit limit reads as no status
The background poll skips a frame whose shape it cannot read, as base skipped frames it could not parse, stops reading a chunk at [DONE], and lets only the last terminal frame in a chunk decide its status
A Responses stream's first chunk answers as a JSON error only for codes 400 and up
* fix(proxy): read every frame of a polled chunk and name model chunks from the sent JSON
A background poll parsed only the first data frame of each chunk, so a terminal event behind an error or delta frame in the same chunk was lost, and the poll stored a status its own stream never ended with. It now reads every frame up to [DONE]
A terminal event whose response is a string or null kept no status and polled completed. Its status now comes from the event type, and final output items that are not objects are skipped, so the rest of the final output is still stored
Model chunks now take their event line from the JSON they send, as mapping chunks already do, so a serializer that drops type no longer sends event: response.completed over {}
* fix(proxy): read polled chunks as SSE events and keep draining past a mid-chunk [DONE]
A background poll split a chunk into data lines, so one event with two data lines parsed as two payloads, and a frame with an integer past the digit limit raised a ValueError the parse guard did not catch, which failed the poll and dropped its spend row. The poll now joins each event's data lines the way an SSE reader does and skips any frame it cannot parse
A [DONE] behind another frame in one chunk ended the poll and left the stream unbilled. Only a chunk that opens with [DONE] ends the poll now, as on main, and the poll stops reading a chunk at its first terminal event, so two terminal events in one chunk poll the first one as main did
* fix(proxy): JSON-encode only plain hook chunks on Responses streams
A hook chunk whose whole tree is builtin JSON types streams as its named
event. Any other chunk keeps the base data frame, so nothing a custom
repr hides reaches the wire. Error event details encode the way the chat
error frame does.
* fix(proxy): skip containers already visited when checking hook chunks for JSON
* fix(responses): keep streaming past an event with no type
The Responses stream iterator read event.type directly when it noted a
yielded event, so an event whose JSON had no type, like a custom hook's
dict that the proxy now sends as JSON, raised AttributeError and ended
the stream. Read it with getattr, as the iterator's other checks do.
* fix(proxy): keep hook dicts with no type on the base Responses frame
A hook dict with no type is not a Responses event, so it gets no event
name either way. Sending it as JSON only made litellm SDKs that read
type on every event stop the stream. Keep such dicts on the frame the
base sends, so every SDK version reads them as before.
* fix(responses): keep a background poll's first terminal status and take a first-chunk error code from its HTTP status
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(bedrock): forward Codex agent messages that carry encrypted content intact
Codex multi-agent v2 sends worker messages as agent_message items whose
encrypted_content block only the model server can decrypt. Bedrock Mantle and
Bedrock Runtime used to 400 on agent_message, so LiteLLM rewrote every one into
assistant text, pasting the ciphertext into the prompt and dropping the sender
and recipient. Bedrock now decrypts these items itself, so the rewrite left the
recipient model reading an opaque token instead of its task.
Agent messages with an encrypted_content block now go to Bedrock untouched on
both endpoints. The Bedrock configs declare supports_encrypted_agent_messages,
and the shared normalizer keys off that flag so the two cannot drift. Mantle's
/v1/responses route for open-weight models still rejects agent_message in any
form, so it keeps the old rewrite. Plaintext-only agent messages keep the
rewrite on every route.
* fix(bedrock): only keep worker messages locked for OpenAI GPT models
Gemma and Grok on Mantle's OpenAI route never lock Codex worker messages, so
Codex puts plain text in the locked slot and Bedrock rejected it with
invalid_encrypted_content once the route alone decided to forward it intact.
Decide from the model instead, and restore the full-value rewrite assertion.
* fix(bedrock): run realtime guardrails on Nova Sonic sessions
* fix(bedrock): quote the spoken violation message and skip it after a blocked Nova Sonic tool output
* fix(bedrock): screen every message role, end Nova Sonic sessions on blocked transcripts, drop unparseable frames
* fix(realtime): apply guardrail masks, keep block reasons out of spoken notices, fail closed on guardrail crashes
* fix(realtime): screen joined text parts, keep the documented notice, honor key opt-outs
Text parts of one message are joined before screening again, so a phrase split
across parts is still blocked, and a mask collapses them into one part. The
spoken notice is realtime_violation_message or the block reason, quoted only on
Nova Sonic. Client-written output_text and text parts are screened. Key and team
guardrail opt-outs reach regular and translation realtime sessions. A non-string
upstream transcript skips screening as before.
* fix(realtime): honor guardrail role scope, clear stale response.create drops, screen split keywords, report Nova guardrail outages
* fix(realtime): keep every per-part mask when a joined text masks differently
* fix(realtime): screen one joined text per item and pass the session's tenant context to guardrails
Multi-part items reach each guardrail once as the space-joined text, as on main, and only an item a
guardrail rewrites gets its parts screened again to keep per-part masks. Guardrails now receive the
session request data with its metadata, so key and team policies apply, and a ModifyResponseException
block is reported as a block. Nova scopes every message item as the user turn it sends to Bedrock.
* fix(realtime): catch phrases split mid-word across text parts without re-screening single-part items
* fix(realtime): screen each joined text in its own guardrail call and keep a Nova opening message for the retry while it is screened
* fix(realtime): screen text parts in a loop so large messages keep their masks
* fix(realtime): fail closed on a non-string transcript under a guardrail and give each turn its own guardrail metadata
* fix(realtime): fail a non-string transcript only when a guardrail would screen the user's turn
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(bedrock): enable native structured output by default for selected Claude models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): keep govcloud and bare model rows off native output
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): retry rejected native schemas with the JSON tool on default-native models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): keep the rejection retry on the unforceable-tool native path
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): retry rejected schemas on the Messages path for default-native models
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(bedrock): format Messages native output retry
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(bedrock): allow profile-only native output metadata
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): only disable default-native output on schema-rejection retries
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(bedrock): type fixture parameters in native output retry tests
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): retry with the JSON tool when Bedrock rejects a schema's grammar size
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): drop strict tools after a native-output retry hits the grammar limit
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* style(bedrock): format native output fallback
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): honor the deployment native output opt-out and keep wide batch schemas on the JSON tool
A deployment's supports_bedrock_native_structured_output: false now keeps the JSON tool on rows that default to native output. Bedrock batch has no retry, so a record whose schema has more than 64 properties keeps the JSON tool there too.
* fix(bedrock): retry a native structured output stream as a buffered stream when Bedrock denies the stream action
A role allowed only bedrock:InvokeModel got a 403 on schema streams once these
models went native, since native streams call converse-stream while the JSON
tool path buffered through converse. The retry now sends that request again
as a buffered JSON tool stream, the shape the base sent.
The wire test moves its neither-key and opt-in cells to Opus 4.7, its
Bedrock-key-only cells to Sonnet 5.5, and its unforceable-tool cells to the
bare Opus 5.5 id, since this PR makes Opus 5.5 and Haiku 5.5 always native.
The batch wire test now checks a thinking schema record on both sides of the
contract the batch unit test already pins, the JSON tool on opt-in Sonnet 5.5
and native output on always-native Haiku 5.5, since its Haiku 5.5 and Opus 4.8
cells assumed those models were opt-in.
* fix(bedrock): move the Claude 5 native output default to a key released proxies do not read
Released proxies load the cost map from main and have none of the native
output retries, so setting supports_native_structured_output on the Opus
5.5, Fable 5.1, Sonnet 5 and Haiku 5.5 profile rows turned working wide
or min/max schemas into 400s on v1.108.1.
Those rows keep supports_native_structured_output false and gain
bedrock_native_structured_output_by_default true, which get_model_info
reports as supports_native_structured_output, so this version behaves as
before while released versions keep the JSON tool.
* fix(bedrock): rename the Claude 5 native output default key to supports_bedrock_native_structured_output_by_default
The cost map guard runs main's model catalog on the PR's cost map, and that catalog rejects any undeclared key that does not start with supports_. Declaring the old name on main first needs its own PR, so the key takes the supports_ prefix. Released proxies still never read it
* fix(bedrock): honor registered and batch opt-outs, move Opus 5 and 4.8 to the default key
A registered supports_bedrock_native_structured_output false now beats the
native output default, and batch uploads read the deployment's model_info, so
an opted-out deployment keeps the JSON tool on every record. The Opus 5 and
Opus 4.8 profile rows move to supports_bedrock_native_structured_output_by_default,
so released proxies on main's cost map keep the JSON tool on them. The 64
property batch cutoff now applies only to rows with that key, so models that
were already native keep their batch output shape.
* fix(proxy): report a deployment's Bedrock native output opt-out on /model/info
/model/info merged the deployment's model_info over the backend row, so a
deployment that set only supports_bedrock_native_structured_output: false
still showed the backend's derived supports_native_structured_output: true.
The generic flag is now derived from the merged info unless the deployment
set it itself, through the same helper get_model_info uses.
* fix(bedrock): keep native output when a denied schema stream is resent buffered
A role without bedrock:InvokeModelWithResponseStream had its native output schema stream resent as a buffered JSON tool call. The resend now keeps the native params with fake_stream set, and falls back to the JSON tool only if Bedrock then rejects native output.
* fix(bedrock): drop rejected strict tools on the JSON tool retry after a buffered native resend
* fix(bedrock): send a buffered schema stream to /invoke and keep the by-default key on its deployment
The shared HTTP handler now builds the non-streaming URL for a fake-streamed request, so the invoke route's buffered resend after a denied stream calls /invoke instead of invoke-with-response-stream. supports_bedrock_native_structured_output_by_default is deployment scoped, so setting it on one deployment no longer reaches sibling deployments of the same model
* fix(bedrock): answer a buffered invoke stream from the non-streaming reply
* fix(bedrock): keep the JSON tool retry buffered after a denied stream's schema is rejected
* fix(bedrock): count only a true fake_stream and keep the deployment by-default key out of /model/info
A fake_stream sent as "true" or 1 picked the non-streaming URL but kept the streaming reader, so Bedrock invoke streams answered 500 after three billed calls. Only a real True counts now, as on main.
/model/info read supports_bedrock_native_structured_output_by_default from the deployment, while routing reads it only from the cost map, so it reported native output for deployments that use the JSON tool.
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): do not surface server_tool_use as client tool_calls
Anthropic server-executed tools (web search, tool search, code execution)
were returned in chat completions tool_calls with srvtoolu_ ids, so
OpenAI-protocol agent clients treated them as pending work and re-invoked
the model until their recursion limit.
Move them to provider_specific_fields.server_tool_calls on both the
non-streaming and streaming paths. History replay, the /v1/responses
bridge and the /v1/messages streaming guardrail read that field, so
multi-turn reconstruction, web_search_call output items and the
guardrail's fail-closed check on server tool blocks keep working.
* fix(responses): return Claude web_fetch calls as web_search_call open_page items
* fix(responses): keep Claude server web calls without a result as hosted items
A non-streamed Claude turn that calls web_search or web_fetch next to a
client tool can return the server call before its result. Build an
in_progress web_search_call for it, so the Responses bridge no longer
returns it as a client function_call
* fix(anthropic): keep pending server calls native and hide server calls from client tool paths
Replaying a server call that has no result yet now sends it back as
server_tool_use, so Anthropic accepts the next turn. The tool permission
guardrail checks server calls again and rewrite mode removes a denied one.
The /v1/messages adapter keeps Claude's server blocks, and the Responses
bridge never returns a server call as a client function_call.
* fix(anthropic): let guardrails, tool usage, tracing, and Responses replay see Claude's server calls
Claude's server tool calls moved out of tool_calls into provider_specific_fields.server_tool_calls,
so every reader of tool_calls stopped seeing them. The chat guardrail handler now passes them to
output guardrails and writes rewrites back to server_tool_calls (stream and non-stream), tool usage
and registry discovery count them, otel, Datadog and the Logs Tools panel list them, and a pending
Claude web search or fetch item replayed to /v1/responses goes back to Anthropic as its own
server_tool_use instead of running again
* fix(anthropic): let guardrails and loggers read Claude's server tool calls again
Guardrails, tool policies, tracing loggers and the Responses guardrail
translation read only client tool calls after server calls moved to
provider_specific_fields. They now read server calls first, then client
calls, through one shared helper, and the Responses guardrail treats
Claude's web_search_call items as web_search or web_fetch calls.
Non-streaming splits server calls by block type, as streaming does.
* fix(anthropic): keep Claude server calls visible to guardrails, redaction and tracing
A guardrail rewrite of a Claude server call now also updates its hosted web_search_calls copy,
streamed or not. Redaction with message logging off hides server call arguments and hosted web
call actions on chat and Responses. Pillar and Pangea scan server calls, transform-mode streaming
guardrails keep server call chunks, and OTEL v2 Responses spans list Claude's web calls.
* fix(anthropic): keep the legacy hook view of streamed /v1/messages web search after the main merge
Main's legacy post-call hooks see Claude's server call as tool_use with its search results first. The streamed /v1/messages response now keeps that view while native scans keep server_tool_use.
* fix(guardrails): let Cisco AI Defense scan Claude's web calls on Responses
Claude's web search and fetch now come back from /v1/responses as web_search_call items, so Cisco AI Defense reads and clears their query or URL the way it did for the old function_call arguments. The helpers that read and rewrite those items move to the shared guardrail utils so the Responses guardrail handler and Cisco use one copy.
* test(guardrails): inject Pillar's HTTP client and type the Cisco web call actions
* test(guardrails): type the fake Pillar client like its Protocol
* fix(anthropic): keep a pending server call and its thinking when the turn also waits on a client tool
Only a turn with no client tool_use is a paused turn. A tool_use turn that also started web_fetch or
web_search still needs that call and its thinking replayed, or Claude loses the fetch result
* fix(guardrails): scan replayed Claude server calls and hosted web calls
Request guardrails now see and rewrite server_tool_calls replayed in chat
messages. Cato and Purview scan Claude web_search and web_fetch calls on
Responses output, and a function tool call with an extra custom key is no
longer dropped from Cisco, Purview, Repello and PANW scans.
* fix(guardrails): scan replayed Claude server tool calls in Lasso, Lakera and Qualifire request checks
* fix(anthropic): apply streamed guardrail tool call rewrites to their own block when a server tool call follows
* test: annotate the new server tool call test locals with Final and type their parameters
* test: type the new guardrail test helpers with litellm's tool call types instead of Any
* test: narrow the remaining new test helper parameters from bare dict and object
* fix(anthropic): keep server calls in Responses streams, guardrail masks, OTEL and Logs
Streamed /v1/responses now emits Claude's tool search and paused web search calls, text guardrail masks rewrite replayed server calls, OTEL gen_ai messages and the Logs pretty view list server calls next to client calls, and integration tests read server calls from server_tool_calls
* fix(guardrails): keep hosted web call copies in sync on request masks and denied server calls, and accept null tool_calls in Lakera
* fix(anthropic): serialize masked hosted web call copies through their adapter so a replay before the OpenAI model is built does not 500
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(bedrock/converse): omit reasoning signature for non-Anthropic models (#49969)
* test(bedrock): cover signed reasoning replay in mapped tests
* fix(bedrock): drop another model's signed thinking when replaying to a non-Claude Converse target
A signed thinking block from Claude replayed to a Converse model known not to be Claude, such as GPT-6 Luna or Nova 2, now leaves the request instead of going out as reasoningContent. Bedrock rejects that field on those models, so the whole request answered 400
A target that may still be Claude, such as an application inference profile ARN with no base_model, keeps the signed block, since Claude needs it on a tool-use continuation. Unsigned blocks keep the existing text path
The thinking-block error matcher added earlier on this branch is removed, because Converse and non-Claude Bedrock requests never reach that retry
* fix(bedrock): keep signed thinking for Claude behind a region prefix litellm does not list
ca. and in. inference profiles name Claude after a region prefix get_bedrock_cross_region_inference_regions does not have yet, so they counted as non-Claude and lost their signed thinking. The thinking-only turn test now pins the exact merged and continue-message shapes.
* fix(bedrock): limit the region-prefix Claude match to signed thinking and drop a dropped block's cache point
* fix(bedrock): drop Claude's bridged reasoning items before a Bedrock Responses API replay
* fix(bedrock_mantle): drop litellm-minted reasoning items before a Mantle Responses call
* test(integration): expect Bedrock and Bedrock Mantle Responses to drop minted reasoning
Bedrock and Bedrock Mantle answer 400 on a reasoning item litellm minted, so the
wire test now expects both to drop it, like OpenAI and Azure
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
The primary stream's mid-stream stall reaches failure_exits as a mapped litellm.Timeout (408), which cools the deployment on the Rust core. Keep that behavior and update the integration test that expected the Python core's exemption. Fallback hops stay exempt through is_stream_stall.
* feat(interactions): continue requires_continuation long decodes with continuation_token hops
A foreground Interactions request stops at Google's per-hop output cap with
status requires_continuation and a continuation_token. litellm now re-posts the
request with previous_interaction_id and continuation_token until the status
moves on, merging the hops into one interaction (last id, every hop's steps,
summed usage with initial_input_tokens from hop 1), for non-streaming and
streaming calls. continuation_token is accepted as a request param, and
requires_continuation is a known status for billing.
* refactor(gemini interactions): validate streaming chunks with model_validate
* fix(interactions): keep the held completion across stream hops, stop on a repeated continuation_token, follow the spec redirect
* fix(interactions): add grounding_tool_count per tool type, stop on any seen continuation_token, keep streaming past a [DONE] frame when a hop is pending
* ci(interactions): point the spec canary at aistudio.google.com, where Google moved the Interactions OpenAPI spec
* test(interactions): scripted Google ends each long decode hop with a [DONE] frame
* refactor(interactions): pass requires_continuation and continuation_token through instead of looping hops in the proxy
The caller drives the continuation, the same as against Google directly. /interactions accepts continuation_token in place of input so a resume request needs no input.
* test(proxy): assert the accepted /interactions bodies pass the required-param check
* fix(proxy): keep the router default fallback for input on /interactions, continuation_token stands in for it
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): stop forcing the response_format json tool when Claude will think
A forced tool call skips Claude's thinking. The json_schema emulation forced
json_tool_call unless thinking.type was enabled, so adaptive thinking and the
models that think by default (Opus 5, Sonnet 5, Haiku 5.5, and the always-on
families) lost their thinking on every structured-output request.
AnthropicModelInfo.response_format_tool_can_be_forced is now the one check
the Anthropic, Bedrock Converse (including the native output fallback) and
Databricks paths use. It leaves the tool unforced when thinking is requested
in any form or the model thinks by default, and keeps forcing it when the
request turns thinking off (thinking disabled or reasoning_effort none) on a
model that accepts forced tools.
The new thinking_on_by_default cost map key marks Opus 5, Sonnet 5 and
Haiku 5.5 on the providers that reach these paths, and a miss-only fallback
rule covers unmapped Claude 5 and later ids.
* fix(bedrock): route native output and thinking token counts by whether Claude will think
Bedrock's native output opt-in now uses the same will-think check as the json tool, so Claude 5 models that think by default with no thinking param go native. Converse also asks for /usage/output_tokens_details on those models, so reasoning tokens are reported for them.
* fix(anthropic): tell Claude to answer through the unforced response_format tool
Moves the integration tests that check the forced JSON tool from Claude 5 models to Opus 4.8, since Claude 5 models now keep their default thinking and take the unforced tool or native output
* fix(anthropic): keep open-object schemas on the json tool when Claude will think
Native output closes every object, so an open object came back empty. Also declare thinking_on_by_default in the Rust model catalog and expect the Converse thinking-token field in test_main
* fix(anthropic): hold streamed prose until the JSON tool decides the answer
In json mode with the unforced JSON tool, Claude can write its working before it calls the tool. The stream now holds that text, drops it when the JSON tool answers, and sends it when Claude calls a caller tool or skips the JSON tool, matching the non-streamed reply.
* test(anthropic): build the stream test events without nested comprehensions
* fix(anthropic): keep always-on thinking models unforced, open objects native when thinking, and the tool instruction Claude-only
* test(anthropic): keep main's open-object integration cases and cover the unforced tool on a default-thinking model
* fix(anthropic): treat thinking False as thinking off so the JSON tool stays forced
Cover thinking False on the Anthropic and Converse forcing tests, and pin that a Converse
json_schema stream with the unforced JSON tool never sends Claude's working text as content
* fix(anthropic): drop the thinking False forcing case that completion never reaches
litellm.completion turns thinking False into no thinking param before any route sees it,
so the case was dead on the request path and False keeps meaning the model's default
* test(bedrock): collect the Converse json_schema stream chunks in one assignment
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(nvidia_riva): send the api key once per call by talking to Riva over grpc directly
Drops nvidia-riva-client from the stt-nvidia-riva extra. Depending on its version it sent
the bearer token zero times on plaintext or twice on TLS, so token-checking Riva deployments
rejected every request (#50059). The provider now declares the few Riva messages it uses,
opens its own grpc channel and puts authorization and function-id on each call exactly once
* test(nvidia_riva): run the provider against a real grpc server built from nvidia's own definitions
* fix(nvidia_riva): validate endpointing settings with pydantic before building the riva message
* test(nvidia_riva): check the key is sent once over tls, and drop a comment that restated the code
* feat(decisions): add EmpirioLabs AI as a decisions provider
empiriolabs/aplomb-1 now works with litellm.decisions and litellm.systemone, the proxy /v1/decisions and /v1/systemone routes, the decision model guardrail and the Admin UI. Requests go to POST /v1/decisions on the EmpirioLabs AI API with the System One body, keyed by EMPIRIOLABS_API_KEY with an optional EMPIRIOLABS_API_BASE
Chat models keep routing through the existing OpenAI-compatible JSON provider, and a test pins that
* ci: run the EmpirioLabs AI decisions tests in the llms unit shard
* fix: align EmpirioLabs registration with CI gates
---------
Co-authored-by: Adam Dalloul <47503782+Adam-Dalloul@users.noreply.github.com>
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): answer a 503 as 529 overloaded_error on /v1/messages
Anthropic's error contract has no 503, so a client built against it (Claude Code) treats a 503 from /v1/messages as a plain 5xx and retries it ten times instead of taking its overload path. A Bedrock 503 now leaves /v1/messages and /v1/messages/count_tokens as 529 overloaded_error, for both a JSON answer and a stream that fails before its first byte.
An Anthropic 529 overloaded_error now maps to litellm.ServiceUnavailableError (503) instead of InternalServerError (500), so OpenAI-shaped routes answer it as 503, the status OpenAI documents for an overloaded engine, and router retry and allowed-fails policies key it as ServiceUnavailableError.
* test(proxy): expect overloaded_error on a 503 stream error after a keepalive ping
* fix(proxy): record the 529 the Anthropic routes answer and type the chat stream error frame
The messages route stamps the answered status on the exception before post_call_failure_hook runs, so the spend log row, the OTel server span, the failed-request metric, and alerting record the 529 the client saw instead of the deployment's 503, while the deployment failure metric and cooldown keep the deployment's own status. A chat completions stream error frame names the type its status stands for instead of the null .type ServiceUnavailableError carries. An Anthropic-shaped 503 envelope passes through as 529 overloaded_error through a validated TypedDict, and count_tokens answers every ProxyException as the Anthropic envelope. Five integration tests that pinned 503 on the messages route now expect 529
* test(proxy): expect the top-level Anthropic envelope from count_tokens errors
count_tokens now answers a ProxyException as a JSONResponse carrying the
Anthropic error envelope instead of raising HTTPException, which FastAPI
wrapped in {"detail": ...}. The three legacy error-format tests now read
the response body and check that no detail wrap is left.
* test(proxy): cover the 529 overloaded_error answer across routes, cooldown, and retry budgets
* fix(anthropic): keep Anthropic overloads as InternalServerError and answer 529 from the provider status
An upstream Anthropic 529 stays litellm.InternalServerError (status 500) so
InternalServerErrorRetries, InternalServerErrorAllowedFails, SDK handlers,
and the OpenAI-shaped routes behave as on main. The error now carries
provider_status_code, and the /v1/messages routes answer 529
overloaded_error from it.
* refactor(proxy): unwrap a mid-stream fallback error once instead of recursing for the provider status
* fix(anthropic): keep count_tokens' detail body, end keepalive streams with an Anthropic error event, and log any 529 as an overload
* test(integration): cover the audit round 3 cells and expect 529 on the count_tokens 503 cases
* fix(proxy): re-raise an unchanged stream HTTPException as itself and walk nested mid-stream fallbacks for the provider status
* fix(proxy): walk nested mid-stream fallbacks without recursion
* fix(anthropic): preserve 503s and answer upstream 529s on messages routes
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(anthropic): answer an upstream Anthropic 529 as 529 on count_tokens
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(anthropic): use integration coverage for keepalive errors
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>
* chore(cost-map): remove the 52 rows whose provider retired them before 2026
* test(utils): pin the proxy function-calling parity case to azure/gpt-4.1
* test(batches): price the Azure batch-logging regression case against gpt-4.1
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(logging): keep the caller's call_type when chat completions are bridged to the Responses API
* fix(logging): log bridged chat completions with the chat response they returned
* fix(logging): log bridged text completions and /v1/messages as served, with the request's options
Bridged rows were translated without the request's options, so structured
output and replay fields could differ from what the client got. Text
completions and non-streamed /v1/messages through the chat bridge also kept
a Responses API object or the encoded id.
* fix(logging): redact reasoning summaries on bridged rows and log bridged text completions as text completions
Redaction now drops reasoning_items on chat messages and deltas, so a bridged row keeps no reasoning summary when message logging is off. A non-streamed bridged text completion logs the text completion the client got, priced like the chat response it came from. The bridge handler's input TypedDict fields are ReadOnly.
* fix(logging): redact bridged text completions for opted-out callbacks and pin the sync logged shape and usage ack
* test(integration): look up bridged chat spend rows by the id the client got
Bridged chat spend rows are keyed by the client-facing id and logged with the caller's call type, so the helpers that decoded the old resp_<base64> key or pinned call_type responses now assert the new contract
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(cache_control): read the routed deployment's own prompt-cache flag, not a sibling's shared cost-map entry
Two deployments that share one model string share one key in the live
litellm.model_cost map, and router registration writes each deployment's
supports_prompt_cache_breakpoint into that key. The cache control hook read
the live map, so a flagless deployment picked up a sibling's opt-in and sent
prompt_cache_options and prompt_cache_breakpoint upstream, which the provider
rejects with a 400
When a call carries the routed deployment's model_info, the prompt cache
checks now use its declared supports_prompt_cache_breakpoint, and fall back
to the loaded catalog instead of the live map. This covers the hook on
completion, acompletion, Responses and the /v1/messages route, the azure_ai
chat transform, the OpenAI Responses transform under drop_params, and the
/v1/messages to Responses adapter
* fix(prompt caching): read each deployment's own breakpoint flag in the openai and bedrock readers
A deployment's own supports_prompt_cache_breakpoint now picks OpenAI's
prompt_cache_breakpoint only where the request goes out OpenAI-shaped on a
provider other than openai, such as azure_ai OpenAI models and bedrock_mantle.
Bedrock on every route and providers that build Anthropic messages keep
cache_control, which Bedrock Converse turns into cachePoint
The openai chat reader now passes the routed model_info and checks base_model.
The Bedrock cache point reader takes the deployment's flag first and, for a
routed call with no flag, reads the catalog as loaded instead of the live map
that sibling deployments write into. model_info is threaded from the converse
transform, the factory message and tool builders, and the responses bridge
* fix(prompt caching): keep gateways on cache_control and read register_model opt-ins on routed calls
A declared supports_prompt_cache_breakpoint picks OpenAI's breakpoint only on providers hosting OpenAI's API or for a model the cost map lists under the provider, so OpenRouter, LiteLLM proxy, and vLLM deployments keep cache_control as before. Routed calls read the loaded catalog plus the user's own register_model entries, never router registrations.
* fix(prompt caching): honor a deployment's own supports_prompt_caching on Bedrock and its declared litellm_provider on gateways
* fix(prompt caching): keep wildcard cache flags off matched models, read deployment flags for auto caching and base_model aliases
* test(anthropic): cover auto caching on an unmapped Claude deployment
* fix(cache_control): keep the merge base's reads where the per-deployment flag changed them
A register_model entry under a provider-prefixed name reaches routed calls again, since the
overlay now resolves the bare and prefixed names the way runtime registrations are stored.
The hook and the openai chat reader no longer judge base_model, as at the merge base, so a
client base_model sent under allow_client_pricing_override cannot pick OpenAI's breakpoint
and get a 400.
A Bedrock deployment flag that is not a real true reads as false again, as at the merge base.
An azure_ai wildcard's supports_prompt_cache_breakpoint speaks for each model it matches again,
as at the merge base. Other wildcards still leave both cache flags to the catalog.
The s10/s11 Bedrock wire test asserted that a flagless deployment inherits its sibling's flag
on a shared ARN, which is the leak this PR removes. It now asserts that each deployment keeps
its own flag and that a flagless unmapped ARN keeps its cache points after the delete.
* fix(cache_control): let the later register_model entry win between a bare and a provider-prefixed name
Routed calls read register_model entries through an overlay that merged the bare
name after the provider-prefixed one whatever order they were registered in, so a
later opt-out under one name lost to an earlier opt-in under the other. Each
registration now moves its key to the end of the kept registrations, and the
overlay merges both names oldest first, which matches what SDK calls read from
the live model map.
Also lists the wildcard cache flag helper in the router coverage script, since
the hook's wildcard tests drive it rather than calling it by name.
* fix(cache_control): keep register_model entries visible while a cost map reload replays them
The previous commit moved a re-registered name to the end of the saved
registrations by removing and re-adding it, and the overlay walked that dict to
order a bare and a provider-prefixed name. A reload replays every saved entry,
so a request on another thread could briefly miss an entry or fail on a dict
that changed size mid-walk. Entries are written in place again, a per-name
sequence records which registration came last, and both the overlay and the
reload replay order by it, so the later registration still wins after a reload.
* fix(cache_control): read a string or numeric supports_prompt_caching as the bool it spells and find register_model entries under any route prefix
A deployment flag stored as "true", 1, "false" or 0 (POST /model/new and quoted YAML keep strings) now means that bool on the Bedrock cachePoint reader and Anthropic auto caching, and any other value falls to the catalog. register_model records which names wrote to each catalog row, so an opt-out under bedrock/ or bedrock/converse/ reaches routed calls like one under anthropic/.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(videos): serve custom_provider_map BaseVideoConfig providers on /v1/videos
A provider registered in custom_provider_map with a BaseVideoConfig handler now resolves on every
video call type (create, status, content, list, remix, edit, extension, characters), the same way
CustomLLM handlers already resolve on /v1/images/generations. The video entry points and the HTTP
handler stop casting the provider name to LlmProviders, which raised before any lookup ran, and a
provider without a video config gets a 400 instead of a 500.
* fix(videos): type the custom provider lookups so the type gate holds
Pick the CustomLLM for a custom provider through ProviderConfigManager.get_custom_llm_handler,
which narrows the now wider custom_provider_map handler union, read custom_provider_map instead
of the private _custom_providers list, and annotate the untyped custom_llm_provider parameters
on the video entry points
* test(videos): cover a CustomLLM shadowing a built-in video provider name and type the new tests
Mark custom_handler ReadOnly on CustomLLMItem, type the new tests, and assert that a CustomLLM
registered under openai still leaves openai video on the built-in config
---------
Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): reproduce comparison response ID ownership and wire failures
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): guard stale retries and malformed comparison envelopes
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): authorize and unwrap comparison response ids
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): require comparison IDs to follow response security policy
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): deny raw comparison response ids by default
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): reject raw comparison response ids by default
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(responses): enforce managed ownership with unmanaged ID opt-in
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>
* fix(bedrock): send native structured output for json_schema on Claude models that cannot force a tool
Opus 5.5 and Sonnet 5.5 on Bedrock cannot force the json tool, so a json_schema request without thinking
sent an unforced json tool the model ignored. The opt-in native output key now also applies when the model
map says the model cannot force a tool, and the Opus 5.5 commercial Bedrock rows carry the opt-in key. The
GovCloud Opus 5.5 rows carry it as false, like the Sonnet 5.5 GovCloud rows
* fix(bedrock): keep json tool with thinking on Opus 5.5 and Fable 5.1, and GovCloud on the tool
Native output without thinking now applies to every Bedrock Claude model that cannot
force the json tool unless supports_bedrock_native_structured_output is false, so Opus 5.5
needs no opt-in row and keeps the json tool when thinking is requested, as #46870 chose.
The invoke chat path keeps the native output_format it built instead of inlining it, and
GovCloud region-path ids read their own false rows.
* fix(bedrock): clean caller output_format schemas before sending them natively on invoke
* fix(bedrock): keep Mantle Claude on the json tool and send a caller's output_config.format natively on invoke
* fix(bedrock): keep the faked Converse stream for native output that replaces the json tool
A json_schema stream that goes native only because the model can't force the json tool now keeps the json tool's faked stream, so it still needs only bedrock:InvokeModel. /v1/messages builds a native-output retry only where it sends native output, and the Mantle test drops its mutable capture.
* test(bedrock): cover native output and the opt-out in the invoke json-mode refusal tests
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(gemini): resend once with the placeholder signature when the other backend rejects a replayed thought signature
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(integration): gemini tool-call signature rejected by the backend is resent once with the placeholder
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(gemini): keep the placeholder resend inside the cache rebuild so either rejection order ends in an accepted body
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(gemini): resend a cached-prefix signature inline and cover native generateContent
When the rejected signature sits inside a cachedContents prefix, the
placeholder resend now carries the history inline, since replacing the
signature in the request body cannot reach the cache. The native
generateContent handlers get the same one-time resend.
* fix(gemini): send the history inline when Vertex rejects a signature while creating the cache
* fix(gemini): keep the context cache on a signature resend and go inline only when that is rejected too
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(responses): log the encrypted response id on streamed /v1/responses spend rows
With response id security on, the stream hook encrypts the resp_ id on every
frame, but logging had already received the SDK's own copy of the terminal
response, so the spend row, the standard logging payload and callbacks got the
managed id (or the chatcmpl id on a bridged stream) the client never saw. The
hook now records the exact id it advertises on the request's logging object,
and the logging layer stamps that id on the logged streamed response
* fix(responses): return the id the client sent on GET and cancel, and keep read rows off the create row's id
Every GET /v1/responses/{id} re-encrypted the response id, so a client that continued a bridged conversation with the
id a GET returned referenced a ciphertext no spend row was filed under. The follow-up lost turn 1 and opened a new
session. GET and cancel now answer with the exact id the client sent when it is an id this proxy issued and it decodes
to the response the provider returned. A create with previous_response_id still gets a fresh id for its own response,
and a raw provider id still gets a managed id
Because a read now carries the create's id, a GET or cancel spend row is filed as <id>_retrieve_<call id> or
<id>_cancel_<call id>. That keeps one row per read, as before, and a read row can never take the create row's
request_id, so the create row survives whichever row is written first and a follow-up replays turn 1 once. The
background settlement row keeps <id>_background_cost, and the cost poll's own status checks keep the response id
* fix(responses): log a fresh id for each read and cancel, echo the held id only to its owner
A GET or cancel of a stored response now logs a new encrypted id of its own,
so callbacks and spend rows never reuse the create's id, while the owner of
the id still gets back the exact id they sent. An admin reading someone
else's response gets an id bound to the admin, as before. The spend row id
suffix for reads and cancels is gone, so with response id security off the
rows and session history match the default branch. A streamed GET follows
the same rules.
* fix(responses): echo the sent id to every reader that passes the response access check
A GET, streamed GET or cancel of a response now answers with the id the caller sent whenever the caller passed
the access check, not only when the caller is the user or team the id was issued to. An admin, the master key
or a teammate who read someone else's streamed response got a fresh id back, and the session lookup could not
reach the first turn from it, so a bridged follow-up through that id lost the conversation history
Every read and cancel still logs a fresh id bound to the reader, never the create's id
* fix(responses): keep a read or cancel row in session history only when it brings new output
A GET, streamed GET or cancel sent with a session header joins the
create's session, and its stored response repeats the create's output.
The replay then sent the same assistant turn once per read.
Session history now keeps an aget_responses or acancel_responses row
only when its stored Responses output has an item id that no earlier
kept row carried. A polled background response keeps the first poll
that returns the finished output, since its queued create row has none,
and later polls of the same output are dropped. Every other row keeps
today's rule
* fix(responses): replay each history item once, echo the sent id on delete, keep background cost files apart
A create row now owns its output items in session history, and a read or cancel row adds only items no
create and no earlier read carried, in one linear pass, so a background response polled while it grows,
or a read that ends before its streamed create, replays the answer once and after its prompt
DELETE /v1/responses/{id} answers with the id the caller sent, on a copy, and logs a fresh encrypted id
like a read does
Azure Data Lake Storage per-event files and unbatched GCS objects are named after the event's spend row
id, so a streamed background create and its cost event, which now share the client's id, keep separate
files
The integration tests that decoded the managed id now match the exact id the client received
* fix(responses): keep main's answer for a raw-id delete and main's storage names outside background cost events
A DELETE by a raw provider id returns the provider's id as main did, and logging still records a fresh
encrypted id. A streamed GET matches the held id by upstream response id, like a plain GET
Only a background response's cost event gets the _background_cost suffix on its Azure Data Lake Storage
file and unbatched GCS object, so batch retrieves and events with no id keep main's names
An item a session read replays takes the newest read's copy at the first read's place, so an item a
poll caught half written replays finished
Tests pin the client's id over the router's continued response id, and the storage test no longer reads
the clock after the upload
* fix(spend): give background cost events their own OpenMeter and DynamoDB ids, replay id-less read items
A streamed background create and its later cost event share the client's
id, so OpenMeter dropped the second event and DynamoDB overwrote the
create's item. Both now key the cost event as <id>_background_cost, like
its spend row and storage file.
Session history again replays a read's output items that carry no id,
as main did.
* fix(integrations): keep OpenMeter and DynamoDB logging working without the proxy extras
* fix(integrations): give background cost events their own ClickHouse request id and DeepEval span id
A streamed background create and its later cost event now share the id the
client received, so ClickHouse and DeepEval saw both events under one id.
Route both through logged_event_key, the way the Azure storage, GCS,
OpenMeter and DynamoDB loggers already do, so the cost event's ClickHouse
request_id matches its spend row's `<id>_background_cost` and its DeepEval
span gets its own uuid. ClickHouse response_id stays the raw provider id.
* test(deepeval): annotate the background cost span test's locals with Final
* test(responses): run the DELETE wire test that pins the echoed client id
DELETE /v1/responses/{id} now answers with the id the caller sent, so the
skip that marked it as a known bug no longer applies. The test passes on
the /v1, /openai/v1 and bare routes.
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(count_tokens): count grok-4-1, grok-4.3 and DeepSeek V4.1 tool requests like their chat templates
Render grok-4-1, grok-4.3 and DeepSeek V4.1 tool schemas the way each chat
template does, and add a per-deployment token_count_tool_framing next to
token_count_framing for the hidden tool preamble a host bills. The Azure AI
Foundry grok-4-1, xAI grok-4.3 and Fireworks DeepSeek V4.1 cost map rows carry
the measured framing, overridable in model_info
* fix(count_tokens): count Azure grok-4-1 Responses requests with their own framing and use the tool framing in place of the base framing
* fix(count_tokens): keep operator framing on tool and Responses counts, frame bridged Responses like chat, follow base_model for tool rendering, count grok tool_choice
* fix(token_counter): count Azure Responses strict default, skip false strict and built-in-only tools, per-kind operator framing
* fix(token_counter): count an explicit null strict like false on Azure Responses
* fix(count_tokens): follow the chat completions bridge flag and Azure's strict default on azure/ deployments
Keep Azure grok-4-1 chat tool counts at or over the bill by raising its tool framing to 302
* fix(token_counter): keep strict false for non-grok templates, coerce the bridge flag, frame DeepSeek API counts
* fix(proxy): frame Responses counts for the api_base router defaults send inference to
* fix(proxy): keep the api_base helper's name so main's /v1/messages count caller resolves
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(bedrock): resolve Invoke Claude default max_tokens from the base model id
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(bedrock): build Invoke Claude request params without mutating
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): type tool-search normalizer return to clear unknown-type gate
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): resolve Invoke and Mantle Claude default max_tokens from the deployment id, ARN base_model, and Mantle row
* fix(bedrock): default Invoke Claude max_tokens from the deployment's own cap and honor AnthropicConfig
Read the called deployment's model_info.max_output_tokens before the base model's
cost row, skip the shared regional key the router fills from every deployment, and
leave the default to litellm.AnthropicConfig(max_tokens) when it is set.
* fix(bedrock): send the mapped default max_tokens on bedrock_mantle Claude chat
* fix(bedrock): bound a global AnthropicConfig max_tokens by the deployment's own cap
* test(bedrock): annotate the new Mantle default max_tokens tests with Final
* refactor(bedrock): inject the AnthropicConfig max_tokens reader into the Invoke Claude config
---------
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(databricks): join plain text blocks for non-Claude json_object requests
* fix(databricks): join only the last message for json_object and drop cache_control under drop_params
Databricks checks for "json" only in the last message, so earlier turns now go out as sent and only the
last message, plus any system messages merged into it, is joined into a string. Under drop_params a
text block's cache_control is dropped so that message can be joined too, and without it the request
still goes out as a list.
* fix(databricks): skip dropped assistant turns and blank parts when joining for json_object
The json_object join picked the last message before the request transform drops bare assistant turns,
so a trailing thinking-only or reasoning-only turn (a Responses input ending in a reasoning item included)
left the real last user message as a list and Databricks still returned its 400. The pick now skips the
turns that transform drops.
The join also ran when only one part had text, which sent whitespace-only parts the existing empty block
filter used to drop on requests that already worked. It now joins only two or more non-blank parts, so a
request with one non-blank part goes out exactly as before. Without drop_params, a message with a
message-level cache_control now goes out as sent instead of being rewritten into one cached block.
* fix(databricks): join json_object text parts only on models Databricks hosts
Databricks checks for the word "json" only on the models it hosts
(databricks-* pay per token names and system.ai.* AI Gateway names). An
external model endpoint skips that check and takes a list of text parts
today, so its request now goes out as sent instead of joined.
The drop_params read now goes through drops_unsupported_params, and a
test pins that a text block with a truthy non-string text is kept.
* fix(databricks): reject a text part holding a non-string locally with a clear 400
* refactor(databricks): reject a non-string text part at the transform_request boundary
* fix(databricks): read pydantic message objects in the non-string text check
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
