@codex

Thanks for the review. I addressed both blocking regressions.

Changes made:

I left ThinkingBlock.signature out of this PR since it was called out as acceptable follow-up work.

Validation:

• ✅ npm test -- tests/structured-output.test.ts tests/anthropic-adapter.test.ts tests/openai-adapter.test.ts
• ✅ npm run lint
• ✅ npm test (48 test files, 689 tests)

Implemented the review follow-ups for native reasoning support.

This update keeps the original reasoning streaming design, while fixing the merge-blocking regressions and token-estimation gap:

Validation:

• ✅ npm run lint
• ✅ npm test -- tests/structured-output.test.ts tests/anthropic-adapter.test.ts tests/openai-adapter.test.ts
• ✅ npm test -- tests/tokens.test.ts tests/context-strategy.test.ts tests/structured-output.test.ts
• ✅ npm test (49 test files, 690 tests)

Thanks, the latest round addresses the structured-output and SDK-narrowing regressions cleanly. One outbound-serialization concern before merge.

anthropic.ts:67, openai-common.ts:143/179, and gemini.ts:88 wrap reasoning text as <think>...</think> and push it back into provider history as ordinary assistant/user text. The content was real, but the format is non-roundtrip: what the model emitted (e.g. an Anthropic ThinkingBlock plus TextBlock) gets flattened into a single text block on the next turn, and the model sees a different shape than what it actually produced.

Concrete costs:

• Context pollution: <think>...</think> becomes part of assistant text on subsequent turns, where the model has no contract that says "ignore this region."
• Token waste: reasoning traces are typically the largest part of a turn, and they now ride along on every subsequent request.
• Blocks the future Anthropic native path: extended-thinking continuation across tool-use turns expects ThinkingBlockParam (with signature) in history, not text. Once we want that, the text-wrap codepath has to come out anyway.

Simplest fix: skip reasoning blocks when serializing outbound. Stream events still fire, LLMResponse.content still retains reasoning for consumers, history just stops carrying it.

// anthropic.ts toAnthropicContentBlockParam
-    case 'reasoning': {
-      const param: TextBlockParam = { type: 'text', text: `<think>${block.text}</think>` }
-      return param
-    }

Same delete in toOpenAIUserMessage, toOpenAIAssistantMessage, and toGeminiContents (and the toThinkTag helper goes with them). Anthropic native thinking continuation (with signature) is a separate follow-up. Main currently degrades thinking blocks to a fallback text block anyway, so this PR doesn't introduce a regression on that front.

One caveat worth flagging: local <think>-tag models (DeepSeek R1, QwQ) emit reasoning as part of their normal text output, not as a separate channel, so they're unaffected by the outbound rule above; their tags will keep appearing in text blocks the way they do today. If we want a different policy for those later, it can be a separate issue.

Process note for next time, not blocking this PR: #169 explicitly says "not a green light to implement until the StreamEvent extension is designed," and lists three open design questions. This PR ended up answering all three (retain reasoning in LLMResponse.content, unify Anthropic + OpenAI into one event/block, count toward token budget). The answers themselves are reasonable, but the agreement should land in the issue before the implementation PR opens, so reviewers can push back on the design separately from the code. I'll split those into their own issues so they're tracked, not buried in this thread:

• Retain reasoning in LLMResponse.content vs. stream-only
• Unify or split Anthropic thinking vs. OpenAI reasoning_content block types
• Token budget treatment for reasoning text

Drop those four cases and this is good to go.

Thanks, addressed the outbound serialization concern.

What changed:

This means provider history no longer gets polluted with reasoning text serialized as ordinary assistant/user text. Local models that emit literal <think> tags inside normal text blocks are unchanged, as noted.

Validation:

• ✅ npm run lint
• ✅ npm test -- tests/openai-adapter.test.ts tests/anthropic-adapter.test.ts tests/gemini-adapter-contract.test.ts tests/tokens.test.ts tests/structured-output.test.ts
• ✅ npm test (49 test files, 690 tests)

LGTM, merging.