feat(aikit): record thinking levels on assistant messages - #53

Summary

Every assistant message now records two thinking fields, stamped by aikit at the wire:

• thinkingLevel (required): the level the caller requested, before aikit fits it to the model. It is "off" when no level was requested.
• providerThinkingLevel (optional): the provider-native value the request actually carried. For example, OpenAI minimal is recorded as low and OpenAI off as none. It is left empty for token-budget models and when the request carried no level.

providerThinkingLevel is read off the exact request body, not off the options aikit passes to the AI SDK, because the SDK drops OpenAI efforts for model ids it does not recognise:

• For streamed requests, the body comes from the start-step part. streamText is now given include: { requestBody: true }.
• For rejected requests, it comes from APICallError.requestBodyValues.

One reader covers every provider: reasoning.effort, reasoning_effort, Gemini thinkingConfig.thinkingLevel, and Anthropic output_config.effort unless thinking is disabled.

The harness needs no code change, because the message envelope is persisted whole. Messages stored before this change no longer decode, so local session databases need clearing.

Also in this PR

• Codex fetch failures: a fetch that rejects before any response is now wrapped in a retryable APICallError carrying the request body, as AI SDK providers do. It is retried, aborts and timeouts pass through, and the failed message still records its effort.
• Claude Haiku 5: treated as an adaptive thinking model, with xhigh and max mapped. Haiku 5.5 rejects fixed token budgets, takes every effort from low to max, and, unlike Opus and Sonnet 5.5, accepts disabled thinking.
• Live test models:
  • Haiku 5.5 replaces claude-sonnet-4-5-20250929 and Fable 5.1 in the Anthropic suites.
  • Gemini 3.8 Flash leaves the OpenRouter image suite.
  • Haiku 5.5 skips the text-only placeholder test, which it declines to answer.

Test plan

• vp check and vp run typecheck are clean.
• Offline wire tests capture the real HTTP body and assert that the recorded providerThinkingLevel equals the effort on the wire. They cover clamped levels, off, caller overrides, token-budget and disabled thinking, and unknown OpenAI model ids.
• Codex transport tests: a network failure is retried and records its effort, and an abort is not retried.
• Live thinking-level suites assert both fields on real responses for OpenAI, Codex, Anthropic and OpenRouter.
• Full live suites for aikit, harness and codework. The remaining failures come from the organisation's model access and rate limits, plus model variance that passed on rerun.

🤖 Generated with Claude Code