feat(llm): capture Claude/OpenAI reasoning as a real subtype:reasoning channel (ENG-1109)

[torrmal]

Summary

Anton has no way to distinguish the model's own reasoning/narration ("let me verify that first...", "confirmed — this changed, let me re-check...") from the final answer — both travel through StreamTextDelta, the only text-carrying event that exists. That means narration gets permanently baked into the persisted answer instead of showing as a live, disappearing "current thought." The frontend already supports this correctly (built for cowork's ENG-1108) via a thought.progress + subtype: reasoning|thinking event — the Hermes harness already emits it; Anton just never did.

Confirmed via direct code reading, not assumption:

• StreamTextDelta is the only text event in provider.py's StreamEvent union.
• Anthropic's extended-thinking content blocks were never parsed (misclassified as plain text, or thinking_delta/signature_delta silently dropped).
• OpenAI's Responses-API reasoning-summary events were never requested (reasoning.summary was missing from the request kwargs) and never handled in the streaming loop.
• The MindsHub-managed gateway path (mdb.ai/mindshub.ai) proxies even Claude calls over an OpenAI chat.completions-shaped envelope (FLAVOR_MINDS_PASSTHROUGH), not the Responses API — so a correct fix needs 3 code paths, not just "enable Claude thinking."

Changes

• provider.py: new StreamReasoningDelta event, added to StreamEvent.
• anthropic.py: output_config.effort (verified real — Claude 4.6/4.7's adaptive-thinking parameter, already sent whenever reasoning_effort is configured for a compatible model) already triggers server-side thinking; the gap was purely on the parsing side. Added content_block_start handling for thinking/redacted_thinking block types, and content_block_delta handling for thinking_delta → StreamReasoningDelta. signature_delta is intentionally a no-op (cryptographic verification signature, not user-facing text). No new request parameter — reuses the existing, safely-gated reasoning_effort config.
• openai.py (Responses API / FLAVOR_OPENAI): added "summary": "auto" alongside the existing effort in the reasoning kwarg (only sent when reasoning_effort is already configured, same gating as today), plus a new response.reasoning_summary_text.delta branch.
• openai.py (chat.completions / FLAVOR_MINDS_PASSTHROUGH + FLAVOR_OPENAI_COMPATIBLE_GENERIC): reads delta.reasoning_content defensively via getattr — the de facto convention several OpenAI-compatible reasoning gateways (DeepSeek, vLLM's reasoning parser) use. This one is best-effort — flagging clearly for review since I couldn't verify the actual mdb.ai gateway's wire format from this repo alone.
• session.py: the synthetic "Thinking..." timer (reasoning_start/reasoning_done) now also closes on the first StreamReasoningDelta, matching its existing behavior for text/complete.

Companion PR in cowork-server (maps StreamReasoningDelta → the wire subtype: reasoning event, matching hermes_harness's existing shape exactly): mindsdb/cowork-server#241 (opening next).

No frontend changes needed — cowork's ENG-1108 already built the client-side handling for this exact subtype: reasoning|thinking shape.

Linear: https://linear.app/mindsdb/issue/ENG-1109/anton-capture-claudeopenai-reasoning-as-a-real-subtypereasoning

Test plan

• New tests: TestAnthropicProviderReasoningStream (thinking_delta → StreamReasoningDelta, signature_delta ignored, effort passed via extra_body), TestResponsesAPIReasoningSummary (kwargs gating, reasoning_summary_text.delta mapping), TestChatCompletionsReasoningContent (reasoning_content read + graceful absence)
• Full suite: 1377 passed, 17 skipped — 2 pre-existing failures in test_chat_scratchpad.py (subprocess sandbox execution issue), confirmed identical on unmodified staging via git stash, unrelated to this change
• ast.parse sanity check on all 4 edited modules; ruff check clean (one pre-existing, unrelated E741 elsewhere in session.py)

🤖 Generated with Claude Code