fix(anthropic): honor adaptive thinking config for newer Claude models

[syf2211]

Summary

Parse and forward Anthropic thinking / output_config completion options from config.yaml instead of always mapping the legacy reasoning toggle to thinking.type: enabled.

Motivation

Claude Opus 4.x models reject thinking.type.enabled and require thinking.type.adaptive with output_config.effort. Users who configured adaptive thinking in YAML still got API errors because Continue stripped unknown completion-option fields and the chat UI forced reasoning: true.

Fixes #12908

Changes

File	Change
packages/config-yaml/src/schemas/models.ts	Add thinking and output_config to completion options schema
core/index.d.ts	Extend BaseCompletionOptions types
core/llm/llms/Anthropic.ts	Prefer explicit thinking config over legacy reasoning boolean
gui/src/redux/thunks/streamNormalInput.ts	Skip legacy reasoning toggle when model config already sets thinking.type
core/llm/llms/Anthropic.vitest.ts	Regression tests for adaptive thinking request bodies

Tests

• Added Anthropic.vitest.ts cases for adaptive thinking + output_config, and for config precedence over legacy reasoning: true
• Full vitest suite could not be executed locally due to monorepo workspace package resolution in this environment; CI should run project tests

Notes

• Legacy reasoning: true behavior is unchanged when no thinking block is configured
• output_config.effort enum currently covers low/medium/high; API may add more values later
• Bedrock provider may need a similar follow-up for Opus 4.x on Bedrock

---

Summary by cubic

Honor Anthropic thinking and output_config from YAML and per-request options, preferring them over the legacy reasoning flag. Fixes Claude Opus 4.x errors by supporting thinking.type: adaptive with output_config.effort, and forwards output_config even when set without thinking (fixes #12908).

• Bug Fixes
  • Parse thinking and output_config in packages/config-yaml; add types in core/index.d.ts.
  • Build Anthropic requests with explicit thinking (budget_tokens for enabled) and always include output_config; fall back to legacy reasoning only if thinking isn’t set.
  • In gui, skip adding the legacy reasoning toggle when thinking.type is configured.
  • Add tests for adaptive thinking/output_config, output_config-only, output_config on the legacy reasoning path, and config precedence over legacy reasoning; include required model in completionOptions.

^{Written for commit 69f7f0f. Summary will update on new commits.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

1 issue found across 5 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[syf2211]

I have read the CLA Document and I hereby sign the CLA

[syf2211]

recheck

[syf2211]

recheck

[itxaiohanglover]

Solid implementation! The buildThinkingParams method cleanly handles the three thinking modes (enabled/adaptive/disabled) with proper budget_tokens fallback. Adding thinking and output_config to the type definitions is important for API compatibility.

[syf2211]

Addressed cubic P2 review in 9d1784e: output_config is now forwarded independently of thinking.type, including when only output_config is configured or on the legacy reasoning: true path. Added regression tests for both cases. Composer-2.5 review: APPROVE.