analyze-prompt fails with "does not support thinking" when alwaysThinkingEnabled is true

[cosminfuica]

Bug: analyze-prompt fails with "does not support thinking" error when using Haiku model

Environment

• vibe-log-cli: v0.7.2
• Claude Code SDK: @anthropic-ai/claude-code v1.0.128
• Claude Code settings: alwaysThinkingEnabled: true

Issue

The analyze-prompt command fails when called from the UserPromptSubmit hook with exit code 1:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"'claude-3-5-haiku-20241022' does not support thinking."}}

Root Cause

The SDK query uses haiku as the default model, but when Claude Code has alwaysThinkingEnabled: true in settings, the SDK enables thinking mode globally. Haiku doesn't support thinking, causing the API request to fail.

Reproduction

1. Set alwaysThinkingEnabled: true in ~/.claude/settings.json
2. Configure UserPromptSubmit hook: npx vibe-log-cli analyze-prompt --stdin --silent
3. Submit any prompt in Claude Code
4. Hook fails with "does not support thinking" error

Proposed Fix

In src/lib/prompt-analyzer.ts, either:

Option 1: Use Sonnet instead of Haiku for analysis

const selectedModel = options.model || 'sonnet'; // Supports thinking

Option 2: Explicitly disable thinking in queryOptions

const queryOptions = {
  maxTurns: 1,
  model: selectedModel,
  disallowedTools: ["*"],
  maxThinkingTokens: 0, // Disable thinking for Haiku compatibility
  cwd: tempAnalysisDir
};

Option 3: Check Claude Code settings and choose model accordingly

const claudeSettings = readClaudeSettings();
const selectedModel = claudeSettings.alwaysThinkingEnabled 
  ? 'sonnet'  // Use Sonnet if thinking is enabled
  : 'haiku';  // Use Haiku otherwise

Workaround

Users can manually patch the installed package or set alwaysThinkingEnabled: false in their Claude Code settings (though this affects the entire Claude Code experience).

[kierr]

I have not observed this, seems fixed/fine now?