[HTExplicit]

Problem

1. No reasoning selector on many API-key accounts. The account test dialog shows the effort select only for model rows carrying reasoning_efforts. Only OpenAI rows were ever annotated: a single OpenAI-only loop in openAIAccountTestModels. The resolver had no Anthropic wire, and none of the Claude test payloads carried an effort. In production that covered all 12 Anthropic API-key accounts plus the OAuth account (e.g. 16510).
2. OpenRouter accounts showed no selector even for gpt-5.5. OpenRouter's /models states reasoning as an object ({mandatory, default_enabled, supported_efforts, default_effort}). Decoding it into *bool stored false, which hid efforts in the test dialog and in the Codex levels for those models.

Change

• One resolver, one annotation point. Model rows of every platform are annotated in one place (accountTestReasoningRows), and the raw and plan views share it. Levels come from one wire-gated resolver evaluated on the upstream ID the test actually sends (accountTestClaudeUpstreamModel, also used by the tester).
• Claude (API key, third-party base URL, OAuth, Vertex):
  • levels and default from claude.EffortLevelsForModel / DefaultEffortForModel (Opus 5.5 medium, others high);
  • writes output_config.effort, plus thinking:{type:adaptive} where the family pairs them (not Opus 4.5);
  • drops temperature and sets max_tokens 64000;
  • runs validateClaude55Request like forwarding does.
• Bedrock: efforts only where InvokeModel keeps output_config.effort, via one predicate shared with PrepareBedrockRequestBodyWithTokens; the effort body goes through that same preparation.
• Grok and Chat tests:
  • Grok offers efforts only where grokSupportsReasoningEffort holds on the forwarding-canonical ID, and the body goes through normalizeGrokResponsesReasoningEffort;
  • Chat tests apply NormalizeGLMOpenAIReasoningEffort;
  • the effective effort shown to the dialog is read from the final wire body.
• Upstream metadata: reasoning is decoded as raw JSON. Booleans work as before; the OpenRouter object supplies its efforts (lightest first) and default; any other shape leaves reasoning unknown, never false. Stored snapshots refresh through the existing API-key 24h catalog refresh.
• Capacity catalog: official rows for claude-opus-5-5 and claude-sonnet-5-5 (1M / 128K, checked 2026-10-02 on platform.claude.com). Advertised values do not change.
• Unchanged: a test with no effort selected sends exactly the same body as before.

Verification

Run locally:

• gofmt, go build ./..., and go vet on the changed packages;
• targeted tests: effort catalog, resolver per wire (Anthropic mapping, OAuth, Bedrock gate, Grok gate, Gemini/CN-anthropic still none), Claude/Bedrock/Grok/GLM wire bodies, handler row annotation, OpenRouter parsing, official catalog integrity.

Not run: real model calls.

🤖 Generated with Claude Code