feat(responses): translate legacy <think> blocks into native reasoning field

[weselben]

TL;DR

Response-path translation of legacy <think>...</think> (and configured equivalents) into each API surface's native reasoning field. The canonical chat response hook covers OpenAI chat completions and Anthropic messages because both dialect converters already surface ExtraFields["reasoning_content"] as their native reasoning payload. The native OpenAI Responses API has its own stream transformer that synthesizes the full reasoning item lifecycle.

What

A model that emits <think>reasoning</think>answer returns, after translation:

• OpenAI chat completions — message.reasoning_content carries reasoning, message.content carries answer.
• Anthropic messages — a thinking content block whose thinking text is reasoning, plus a text content block whose text is answer. Policy-gated: default off, unsigned, or redacted per config.
• OpenAI responses API (non-stream) — a reasoning output item whose text is reasoning, plus a normal message output item whose text is answer.
• OpenAI responses API (stream) — the full reasoning item lifecycle synthesized on the wire: response.output_item.added (type=reasoning, status=in_progress, empty summary) → response.reasoning_text.delta events → response.output_item.done (status=completed, full reasoning in one reasoning_text part). The cleaned text delta continues on the original message item.

Request bodies are never modified. Translation is lossless on the wire: no model-visible character is dropped. A <think> block whose closing tag has not yet arrived is buffered and waits; a never-closed block is forwarded verbatim with no rewrite applied, so a chunk-boundary cut can never strand user content as reasoning.

Why

Modern models (Kimi K3, Claude Fable 5, o-series, …) and Kimi Code-class harnesses expect reasoning in the structured field, not as XML inside the message text. A client that reads reasoning_content gets nothing from a model that emits <think> inline; a client that reads the text alone sees the XML tags as noise. The gateway translating on the response path is the only place that can reconcile both sides without touching the model or the client.

How

Files to review (28, +~3300 / -~80):

File	Why
internal/thinkextract/thinkextract.go (start here)	Tag scanner, multi-pair defaults, buffer cap, single-pass Extract, per-surface gate
internal/thinkextract/stream.go	Chat SSE transformer: per-choice state, [DONE] flush, unknown-field preservation
internal/thinkextract/responses_stream.go	Native Responses SSE transformer: reasoning item lifecycle synthesis
internal/thinkextract/responses.go	Non-stream Responses output transform
internal/thinkextract/chat.go	Non-stream ChatResponse / ResponseMessage rewrite via MergeUnknownJSONFields
internal/thinkextract/surface.go	Surface context tag for per-surface gating
internal/thinkextract/*_test.go	Table-driven tests, 89.7% package coverage
config/thinkextract.go	ThinkExtractConfig: env-gated global + per-surface opt-outs, tag pairs, buffer cap, messages policy
internal/anthropicapi/response.go	FromChatResponseWithPolicy + redacted_thinking block
internal/anthropicapi/stream.go	NewStreamConverterWithPolicy for messages SSE
internal/anthropicapi/policy_test.go	Policy matrix tests
internal/gateway/inference_orchestrator.go	Single canonical hooks on chat and responses provider calls
internal/gateway/inference_execute.go	Gate check before transform
internal/gateway/thinkextract_integration_test.go	End-to-end orchestrator tests across all three surfaces
internal/server/translated_inference_service.go	Chat and responses surface tags on dispatch
internal/server/messages_handler.go	Messages surface tag + policy dispatch
internal/server/handlers.go	Pass-through of thinkExtractOptions
internal/server/http.go	ServerConfig wiring
internal/app/app.go	Config -> options conversion
config/config.go	ThinkExtractConfig registration + defaults

The single canonical-chat hook (line 402 and 432 of inference_execute.go) covers chat and messages surfaces because both flow through chatCompletionProviderCall / streamChatCompletionProviderCall; each dispatch function tags its request context with its Surface identifier so the gate knows which config flag to check. The native Responses path has its own hook at responsesProviderCall and streamResponsesProviderCall.

Reviewer notes

• Default on for chat + responses, default off for messages. Chat and responses are lossless on the wire (content is just moved). Messages default off because Anthropic's thinking block contract normally expects a signature; synthesized reasoning has none, so operators opt in explicitly.
• Messages policy is a three-way enum. off (default) drops synthesized reasoning, unsigned emits a thinking block with no signature, redacted emits a redacted_thinking block with the text in Data. Provider-native reasoning is never affected.
• Per-surface opt-out. THINK_EXTRACT_CHAT_ENABLED, THINK_EXTRACT_RESPONSES_ENABLED, THINK_EXTRACT_MESSAGES_POLICY (off/unsigned/redacted). Global off (THINK_EXTRACT_ENABLED=false) is authoritative — a per-surface true cannot resurrect it.
• Multi-pair defaults. Ships with 7 evidence-backed tag pairs (T9 research): <think>, <thinking>, <reasoning>, <reason>, <thought>, <|begin_of_thought|>, Kimi K2 ◁think▷ brackets, Mistral Magistral [THINK], gpt-oss harmony markers. <reflection> and the <|think|>/<|reasoning|> pipe forms are refuted by primary sources and excluded. Custom pairs via THINK_EXTRACT_TAG_PAIRS (env) or think_extract.tag_pairs (config).
• Streaming chunk-boundary safety. Per-choice state holds partial tags in a bounded buffer until the closing tag arrives. An unclosed block at stream end is re-emitted as ordinary content, never dropped.
• Upstream reasoning wins. When the upstream already sends reasoning_content on the delta or message, the translator leaves it alone — no stomping of structured data with extracted text.
• Synthesized marker is wire-safe. thinkextract_synthesized is set on the messages surface only; the Anthropic converters strip it before the wire, so it never reaches clients. Chat-surface responses never carry it.

Deliberately out of scope

• Anthropic thinking-block signatures on /v1/messages. The dialect converter already emits thinking blocks from ExtraFields["reasoning_content"] today (pre-existing). The signature policy is tracked in #46 (closed, off/unsigned/redacted).
• Responses reasoning.summary population. The source is raw text only, not a model-generated summary; a summary would require an LLM call. The synthesized reasoning item always carries an empty summary array.
• Fast-path chat passthrough (tryFastPathStreamingChatPassthrough) and the native Anthropic messages passthrough (dispatchMessagesNative) deliberately bypass translation today and are documented as excluded.

Tests

go build ./... — ok
go test ./internal/thinkextract/ -v — all pass
go test ./internal/anthropicapi/ -v — all pass
go test ./internal/gateway/ -run TestOrchestrator -v — all pass
go test ./internal/gateway/ -run ThinkExtract -v — all pass
go test ./... — all packages pass
go test ./internal/thinkextract/ -cover — 90.7%
go test ./internal/anthropicapi/ -cover — 85.3%
go test ./config/ -cover — 80.6%

Coverage gaps: json.Marshal fallbacks that cannot fire for map[string]any,
pipe-write error branches that need a closed reader to reach, and a reasoning-pending
Flush branch that tryAdvance makes structurally unreachable.

AI disclosure

Generated by an AI assistant (kimi-code CLI) under human direction. Reviewed and committed by the repository owner.

Wayfinder

Map: #34. All tickets resolved: #35, #36, #37, #38, #39, #40 (awaiting user verdict on hook-point), #42, #43, #44, #45, #46.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}