fix: P131 current-model request bodies, sidecar stall + P130 follow-ups; recommend GPT-6 Astra / Claude Opus 5.5 / Fable 5.1

[raylanlin]

What Changed

v0.2.130 (P131) fixes three problems: current Claude models and GPT-6 Astra were unusable, the sidecar stalled after idle gaps, and several P130 features never took effect. It also updates the recommended models.

Model support

• Anthropic:
  • temperature was sent on every request, a 400 on Fable 5/5.1, Opus 4.7+ and Sonnet 5+ (every Claude preset except Sonnet 4.6). It is no longer sent to those models.
  • Reasoning level now maps to thinking: {type:'adaptive'} + output_config.effort. The removed budget_tokens form is no longer sent.
  • Always-thinking models never get thinking.disabled.
  • A 400 that names sampling, thinking or effort fields is retried once without them.
  • stop_reason: "refusal" now shows a notice instead of an empty reply.
• OpenAI:
  • gpt-6* is now treated as a reasoning model: max_completion_tokens, no temperature.
  • On GPT-6 /chat/completions, reasoning_effort is omitted whenever tools are present.
  • The reasoning-param fallback now recognises "… are not supported" errors.
  • Test connection and the vision path send max_completion_tokens to reasoning models.

Sidecar

• Stall after idle: the P129 stdin watchdog peeked stdin from a helper thread. After about 2 s idle, back-to-back requests stalled until they timed out. It now checks the parent PID instead.
• Request handling:
  • Requests go out one at a time, and a budget starts only when its request is sent.
  • Heartbeats can stretch a call to at most 3× its budget.
  • Stop now cancels the wait (code CANCELLED).
• No double runs: failed results are cached under their op_id, so the TIMEOUT follow-up never re-runs a half-built generator. op_ids are namespaced per agent run.
• Python < 3.11: the heartbeat loop now catches concurrent.futures.TimeoutError.
• P130 features now wired up:
  • The "still running" note appears.
  • Tool durations reach the tool row and the exports.
  • The amber busy dot shows while a tool runs.
• looksLikeQuestion: it now judges the end of the reply.

Settings / config

• Protocol switch and quick-fill: they apply a matching model and the provider's defaults.
• Number fields: they clamp on blur instead of on every keystroke.
• Env fallback: it keeps the saved preferences, and the env key is never persisted.
• Config load: nothing is saved before the stored config has loaded.
• Non-streaming requests: they no longer hit the 20 s connect timeout, and finished requests are not re-sent.
• Truncation: it now counts the system prompt and tool schemas.
• Max-rounds summary: it passes the tool list.
• IPC channels: theme and locale channels move into ipc-channels.ts.
• Lint: fixes the ruff E702 error in test_reliability.py.

Docs

• Presets, README / README.zh-CN, USER-GUIDE, ARCHITECTURE and .env.example now recommend gpt-6-astra for OpenAI and claude-opus-5-5 / claude-fable-5-1 for Anthropic.
• Version bumped to 0.2.130, with a CHANGELOG entry.

Related Issue

N/A. The issues came from review of v0.2.129 and from the user's request to update the recommended models.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 📝 Documentation
• ♻️ Refactor
• 🧪 Tests
• 🔧 Build / CI

Testing

• npm test passes: 191 tests, including the new tests/llm-request.test.mjs and tests/sidecar-client.test.mjs.
• npm run lint passes. npm run typecheck, ruff check sidecar/, pytest sidecar/tests (58 passed), compileall, the renderer build and scripts/precommit-check.sh 0.2.130 also pass.
• Manual test, Linux only: no Windows or SolidWorks was available.
  • tests/sidecar-client.test.mjs runs the Node client against the real sidecar server with fake tools. It fails 5/5 on 0.2.129 and passes 5/5 here.
  • Stubbed fetch captured the request bodies for Fable 5.1, Opus 5.5 and GPT-6 Astra.
  • The heartbeat fix was checked under Python 3.10.
  • The PID watchdog exits a busy sidecar 7 s after its parent dies.
• Not verified: the Windows branch of the PID watchdog, and real SolidWorks sessions. These need a test of the packaged build.

Checklist

• Code follows project conventions
• Self-reviewed
• Docs updated (if needed)
• CHANGELOG.md updated (if needed)

🤖 Generated with Claude Code

https://claude.ai/code/session_015nT6r6mw2MoBrJhEfWJbQD

---

Generated by Claude Code