ResponsesToAnthropicRequest only recognized opus/sonnet 5.5, so Haiku
requests fell into the generic enabled+budget_tokens path (upstream 400)
and lost signed thinking history (preserveThinking=false). Mirror the
Messages-side rules: adaptive+medium default, none->disabled+low,
temp/top_p/combo guards, forced tool_choice allowed.
Also preserve Haiku55 thinking signatures in AnthropicToResponsesResponse
and streaming state, and name the model correctly in the between_tools
rejection.
Closes#7919. Wire claude-haiku-5-5 (released 2026-10-07) across model
registry, effort catalog, pricing, billing, validation and Codex
descriptors, following the opus-5-5/sonnet-5-5 template and the official
migration guide + overview specs.
- registry: DefaultModels entry (fixed ID, no date suffix per guide)
- effort: full low/medium/high/xhigh/max family + IsHaiku55 +
claude-haiku-5.5 dotted normalization (official default: medium)
- pricing catalog: 1M/128k limits, base (<=100k) prices and
*_above_100k_tokens absolute prices; the existing ladder derivation
yields threshold=100000 with uniform 5x multipliers, matching the
official two-tier table exactly (cache tiers follow input x5)
- billing/pricing fallbacks carry the same tier so a missing catalog
entry can never silently bill >100k prompts at base rate
- validation: adaptive-only thinking (enabled always 400; disabled 400
at xhigh/max), sonnet-style sampling locks, temp+top_p combo rejected,
forced tool_choice explicitly allowed (differs from opus/sonnet)
- oauth mimicry: skip temperature:1 top-up when top_p is present on
sampling-locked models (sonnet-5-5 included), otherwise the gateway
manufactures the temp+top_p 400 itself
- toolset beta filter + Codex 1M context cover haiku-5-5
- Bedrock/Antigravity mappings intentionally left out (no 5.5 entries
exist there at all, same as opus-5-5); Vertex passes the fixed ID
through untouched
AnthropicContentBlock / antigravity ClaudeContentItem serialized `signature`
with omitempty, so thinking blocks without an upstream signature (Chat
Completions reasoning_content, Responses reasoning without encrypted_content,
every streamed thinking content_block_start) went out with no `signature` key.
Anthropic always sends the key (`""` on content_block_start). Strict clients
such as Grok Build deserialize it as a required field and abort the response
or SSE stream with "missing field `signature`".
- apicompat: thinking MarshalJSON always emits `thinking` and `signature`
- antigravity: ClaudeContentItem MarshalJSON does the same; streamed thinking
content_block_start carries `signature: ""`
- tests: wire-shape regression tests; update TestStreamingReasoning expectation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Responses refusal arrives as a {"type":"refusal","refusal":...} content
part (streamed as response.refusal.delta), not as output_text. The
Responses → Chat Completions converters only read output_text, and
ResponsesContentPart had no refusal field, so Chat clients received an
empty assistant message when the model refused.
Map refusal parts to choices[0].message.refusal, response.refusal.delta
to delta.refusal, and let BufferedResponseAccumulator keep refusal deltas
so non-streaming replies rebuilt from deltas keep them too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Legacy Chat function calling has no call IDs: the assistant turn carries
function_call and the role=function result refers to it only by name.
ChatCompletionsToResponses ignored the assistant function_call and used
the function name as the result's call_id, so the Responses upstream got
a function_call_output with no matching function_call and rejected the
replayed history.
Emit a function_call item for each legacy assistant call and give the
following same-name result the same call_id. IDs follow message order
(call_legacy_N), so the same history converts to the same input across
turns, and skip IDs already used by tool_calls. A result with no
preceding call keeps the previous name-based call_id.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
chatMessageToResponsesItems only mapped "system" explicitly and sent every
other unknown role through the user converter, so Chat Completions
developer messages reached the Responses upstream as user turns and lost
their instruction priority.
The Responses API accepts role "developer" in input messages; convert it
like a system message and keep the original role. Downstream converters
already understand it (ResponsesToAnthropicRequest folds developer into
the system prompt, the Codex transform forwards it unchanged).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chat Completions names a forced function as
{"type":"function","function":{"name":"X"}}, while the Responses API
expects {"type":"function","name":"X"}. ChatCompletionsToResponses passed
the Chat shape through unchanged, so OpenAI API-key accounts on
/v1/chat/completions forwarded the nested shape to /v1/responses and the
upstream rejected the request.
Flatten the named Chat choice; strings and objects already in the
Responses shape are kept as-is.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Chat Completions file parts already become Responses input_file, but the
Antigravity path dropped them when converting to Anthropic and then to
Gemini. Map data-URI input_file to a document block and emit it as
inlineData so Gemini receives the PDF.
Fixes#7583
The canonical Anthropic stream leaves tool_use input empty on
content_block_start and streams the arguments as input_json_delta.
Anthropic-compatible relays may instead put the complete arguments on
content_block_start and send no delta. The Anthropic-to-Responses stream
converter ignored that input, so function_call_arguments.done carried ""
and the closed item "{}", and clients executed the tool with no arguments.
Keep the inline input as a seed. A real input_json_delta supersedes it, so
canonical streams keep their exact event sequence and two JSON documents are
never concatenated. When no delta arrived by content_block_stop, emit the
seed as a single arguments delta before the done event, which still repeats
exactly what the deltas streamed. Empty, "{}" and null inputs keep the
existing "{}" fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
BufferedResponseAccumulator.SupplementResponseOutput rebuilt the output only
when the terminal response had an empty output array. Upstreams can also end
with a non-empty output whose message carries no (or only blank) output_text.
The buffered Chat Completions and Anthropic Messages paths then returned an
empty reply although the text had streamed, while usage still billed the
terminal output_tokens.
When the terminal output has no usable message text, fill the first empty
output_text part (or append one, or append a message item) from the
accumulated deltas. Non-empty terminal text stays authoritative, and nothing
is synthesized when no text streamed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
isReasoningModel matched the literal "gpt-5" prefix plus the GPT-6 Sol/Luna
spellings, so gpt-6 and gpt-6-astra (both already in the model catalog and
price list) fell into the sampling branch. Anthropic Messages and Chat
Completions requests converted to the Responses API then forwarded
temperature/top_p, which the Responses API rejects for these models with
"Unsupported parameter".
Key the check on the parsed GPT generation number instead, keeping the
existing Sol/Luna spelling match, so later families are covered without
another edit. Non-generation GPT families (gpt-image-1, gpt-audio) and
gpt-4.x are unaffected.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Wesley Liddick