fix: resolve #746

[lightweight-agent]

Issue for this PR

Closes #746

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Fix tool_choice: 'required' incompatibility with thinking-enabled Anthropic models when using structured output.

When structured output (json_schema format) is requested, the system sets toolChoice: 'required', which the AI SDK maps to tool_choice: { type: 'any' } in Anthropic's API. Anthropic prohibits this when extended thinking is enabled, only allowing 'auto' or a specific tool by name. The fix intercepts toolChoice in llm.ts before sending to streamText() and rewrites 'required' to { type: 'tool', toolName: 'StructuredOutput' } for Anthropic-backed providers with thinking enabled, preserving the structured output guarantee while satisfying the API constraint.

How did you verify your code works?

• ✅ Research analyzed (Gemini 3.1 Pro)
• ✅ Plan generated (Sonnet 4.6)
• ✅ Fix implemented (Opus 4.6)
• ✅ TypeScript type-checking passed (post-patch)
• ✅ Quick validation passed (GPT-5.3 Codex)
• ✅ Changes are minimal and focused
• ✅ Self-review passed (Opus 4.6 Fast)

Changed files

packages/opencode/src/session/llm.ts | 52 +++++++++++++++++++++++++++++++++++-
 1 file changed, 51 insertions(+), 1 deletion(-)

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

---

Generated by Lightweight Agent — 5-model team: Gemini 3.1 Pro → Sonnet 4.6 → Opus 4.6 → GPT-5.3 Codex → Opus 4.6 Fast

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

[bSi (brunusansi)]

Auto-approved: bot PR passed all checks.

[bSi (brunusansi)]

Auto-approved: bot PR passed all checks.