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

.env.example

@@ -7,12 +7,12 @@
 
 # Anthropic
 # ANTHROPIC_API_KEY=sk-ant-xxxxx
-# ANTHROPIC_MODEL=claude-sonnet-4-20250514
+# ANTHROPIC_MODEL=claude-opus-5-5
 
 # OpenAI / 兼容服务
 # OPENAI_API_KEY=sk-xxxxx
 # OPENAI_BASE_URL=https://api.openai.com/v1
-# OPENAI_MODEL=gpt-4o
+# OPENAI_MODEL=gpt-6-astra
 
 # 百炼
 # DASHSCOPE_API_KEY=sk-xxxxx

CHANGELOG.md

@@ -6,6 +6,103 @@
 
 ## [Unreleased]
 
+## [0.2.130] - 2026-10-01
+
+### Fixed (P131 — current models, and what P129/P130 broke)
+
+**Every current Claude model was unusable.** The Anthropic adapter sent `temperature: 0.3`
+on every request. Fable 5 / 5.1, Opus 4.7+ and Sonnet 5+ reject sampling parameters with a
+400, and the retry only fired for errors that mentioned `thinking`. So every agent turn
+failed on two of the three Claude presets. Reasoning levels were sent as `budget_tokens`,
+which those models also reject (`off` sent `thinking.disabled`, a 400 on Opus 5.5).
+- **Per-model request surface:** `thinking.ts` gains `claudeCaps()` /
+  `anthropicReasoningParams()`. These models now get no sampling fields, and reasoning
+  level maps to `thinking: {type:'adaptive', display:'summarized'}` + `output_config.effort`.
+  `off` on an always-thinking model becomes `effort: low`.
+- **Broader retry:** a 400 naming temperature / thinking / effort is retried once without
+  those fields, so a model id we don't know yet degrades instead of dying.
+- **Refusals:** `stop_reason: "refusal"` now shows a notice instead of an empty reply.
+
+**GPT-6 Astra was unusable on api.openai.com.**
+- The reasoning-model check only matched `gpt-5`, so `gpt-6-astra` got `max_tokens` +
+  `temperature`, both 400s.
+- On `/chat/completions` it also rejects `reasoning_effort` together with tools, and
+  rejects the `minimal` effort. Agent turns now omit the effort (model default), and plain
+  chat maps `off` → `low`.
+- The fallback detector now recognises "… are not supported" errors.
+- "Test connection" sent `max_tokens` to every OpenAI reasoning model, so it failed for
+  GPT-5.x too. It now uses `max_completion_tokens`.
+- The vision path had the same `max_tokens` / `temperature` problem.
+
+**Sidecar stalls (since P129).** The stdin watchdog called `peek()` on stdin from a helper
+thread. After ~2 s idle, the request that followed the next one was not read until more
+input arrived, so back-to-back tool calls stalled into 60 s timeouts. That is likely the
+real source of the "timed out but it finished" reports P130 worked around. The watchdog now
+checks the parent PID and never touches stdin.
+
+**P130 timeout ≠ failure, finished properly.**
+- **Absolute cap:** heartbeats prove Python is alive, not that SolidWorks progresses. A COM
+  call stuck on a modal dialog heartbeated forever, and Stop could not break it (every new
+  message got AGENT_BUSY). Each call now has an absolute cap of 3× its budget (`sw_status`
+  has none), and `call()` takes the abort signal (code `CANCELLED`).
+- **One request at a time:** a request queued behind a slow call burned its budget unread,
+  timed out, and its same-op_id retry re-ran a tool whose first run had failed (only
+  successes were cached). Requests now go out one at a time, budgets start when sent, and
+  failures are cached under their op_id too.
+- **Namespaced op_ids:** op_ids are scoped per agent run. Providers that restart tool-call
+  ids per conversation (Kimi's `functions.<name>:<n>`) could otherwise receive an older
+  session's cached result.
+- **Python 3.9 / 3.10:** `concurrent.futures.TimeoutError` is not the builtin before 3.11,
+  so the first heartbeat tick escaped as an empty TOOL_FAILED.
+- **"Still running" note:** `onStillRunning` only fired after a timeout that heartbeats
+  prevented, so the note never appeared. It now fires on heartbeats, once a minute in chat.
+- **Durations:** tool durations were dropped before reaching the step. They now show on the
+  tool row and in exports.
+- **Amber busy dot:** the busy state never reached `StatusDot`. It is wired through now and
+  serves `connected:true` while a tool runs; concurrent status probes share one request.
+- **`looksLikeQuestion`:** it now judges the reply's ending. Plans mentioning
+  which / 几个 / 哪个 / 多少 no longer read as questions, and
+  「我还缺少以下参数…」 / "I need a few values" no longer get auto-nudged.
+
+**Settings and config**
+- **Protocol switch:** switching to the OpenAI protocol paired `api.openai.com` with
+  `deepseek-v4-pro`. It now uses that endpoint's suggested model.
+- **Quick-fill buttons:** they now apply the provider's model, context window and max
+  output. Those defaults were declared in `presets.ts` but never used.
+- **Number fields:** they clamp on blur, not per keystroke. Typing "2" used to snap to 4096,
+  so values like 200000 could not be entered.
+- **Env-fallback config:** an env-sourced config no longer replaces the saved preferences,
+  and the env key is never persisted to the store.
+- **Config load:** saving before the stored config loaded overwrote it with the defaults
+  and wiped the API key.
+- **Non-streaming timeout:** the 20 s connect-stage timeout no longer acts as a total timeout
+  for non-streaming requests (vision captions), and no longer re-sends finished-but-slow
+  requests up to 3×.
+- **Truncation:** agent-loop truncation now counts the system prompt and tool schemas, and
+  never returns a lone tool result. The max-rounds summary turn passes the tool list, since
+  Anthropic 400s on tool blocks without tools.
+- **IPC channels:** theme / locale channels moved into `ipc-channels.ts`.
+- **CI:** `sidecar/tests/test_reliability.py` failed ruff E702, so the next CI run would
+  have been red.
+
+### Changed
+- **Recommended models:** GPT-6 Astra (`gpt-6-astra`, 1.05M context) for OpenAI;
+  Claude Opus 5.5 (`claude-opus-5-5`) and Claude Fable 5.1 (`claude-fable-5-1`) for
+  Anthropic.
+- **Where they changed:** presets, provider defaults, README / README.zh-CN, USER-GUIDE,
+  ARCHITECTURE, `.env.example`. Saved model ids that left the preset list still load as
+  "Custom model".
+
+### Tests
+- **`tests/llm-request.test.mjs`:** the request bodies each adapter sends, per model, plus
+  truncation and `llmFetch`.
+- **`tests/sidecar-client.test.mjs`:** the Node client against the real sidecar server with
+  fake tools: queueing, failure caching, cap + progress, cancel, and the idle-gap stall.
+  It fails 5/5 on 0.2.129.
+- **Python:** failure caching, the futures timeout on <3.11, and the PID watchdog.
+- **`looksLikeQuestion`:** plan and question cases.
+- **Totals:** 191 JS + 58 Python tests.
+
 ## [0.2.129] - 2026-09-12
 
 ### Changed (P130 — timeout≠failure)

README.md

@@ -28,12 +28,12 @@
 </p>
 
 <p align="center">
-  <img src="https://img.shields.io/badge/version-0.2.37-blue" alt="version" />
+  <img src="https://img.shields.io/badge/version-0.2.130-blue" alt="version" />
   <img src="https://img.shields.io/badge/electron-28-47848F?logo=electron" alt="electron" />
   <img src="https://img.shields.io/badge/react-18-61DAFB?logo=react" alt="react" />
   <img src="https://img.shields.io/badge/typescript-5.3-3178C6?logo=typescript" alt="typescript" />
   <img src="https://img.shields.io/badge/python-3.9%2B-3776AB?logo=python&logoColor=white" alt="python" />
-  <img src="https://img.shields.io/badge/tests-167_JS_%2B_13_Python-brightgreen" alt="tests" />
+  <img src="https://img.shields.io/badge/tests-191_JS_%2B_58_Python-brightgreen" alt="tests" />
   <img src="https://img.shields.io/badge/license-Apache_2.0-orange" alt="license" />
 </p>
 
@@ -86,7 +86,7 @@ Millwright:
 - **Agentic tool loop.** Observe → reason → act. The model chains multiple tool calls, reads structured JSON back from each one, and recovers from errors instead of failing silently.
 - **Visual understanding.** Reorient, rotate, screenshot, and analyze the model — via a multimodal main model or a dedicated vision model.
 - **Resident execution engine.** A persistent Python sidecar holds one COM connection open across an entire multi-step task.
-- **Developer-friendly.** 167 TypeScript/Node tests plus a Python suite (`pytest sidecar/tests`) for the sidecar, a typed IPC boundary, and a `SKIP_SW_CONNECT` mode for UI-only development without SolidWorks installed.
+- **Developer-friendly.** 191 TypeScript/Node tests plus a Python suite (`pytest sidecar/tests`) for the sidecar, a typed IPC boundary, and a `SKIP_SW_CONNECT` mode for UI-only development without SolidWorks installed.
 
 ## Cross-version compatibility
 
@@ -160,17 +160,19 @@ A `Millwright-*-x64.zip` is also published alongside the Setup installer for use
 
 | Provider | Protocol | Base URL | Suggested model |
 |---|---|---|---|
+| OpenAI | OpenAI | `https://api.openai.com/v1` | `gpt-6-astra` (GPT-6 Astra) |
+| Anthropic | Anthropic | `https://api.anthropic.com` | `claude-opus-5-5` (Opus 5.5, default) / `claude-fable-5-1` (Fable 5.1, most capable) |
 | DeepSeek | OpenAI-compatible | `https://api.deepseek.com` | `deepseek-v4-pro` |
 | Kimi / Moonshot | OpenAI-compatible | `https://api.moonshot.cn/v1` | `kimi-k3` |
 | MiniMax | OpenAI-compatible | `https://api.minimaxi.com/v1` | `minimax-m3` |
-| Anthropic | Anthropic | `https://api.anthropic.com` | `claude-sonnet-5` / `claude-opus-5` |
-| OpenAI | OpenAI | `https://api.openai.com/v1` | `gpt-5.6` |
 | Alibaba Bailian (Qwen) | OpenAI-compatible | `https://dashscope.aliyuncs.com/compatible-mode/v1` | `qwen-3.8max` |
 | Zhipu (GLM) | OpenAI-compatible | `https://open.bigmodel.cn/api/paas/v4` | `glm-4.6` |
 | SiliconFlow | OpenAI-compatible | `https://api.siliconflow.cn/v1` | — |
 | Ollama (local) | OpenAI-compatible | `http://localhost:11434/v1` | — |
 
-> Model IDs move fast — check your provider's docs for the current lineup. Agentic tool calling requires a model that supports function calling; DeepSeek V4, Kimi K3, MiniMax M3, and GLM-4.6 are first-class targets.
+> Model IDs move fast — check your provider's docs for the current lineup. Agentic tool calling requires a model that supports function calling; GPT-6 Astra, Claude Opus 5.5 / Fable 5.1, DeepSeek V4, Kimi K3, MiniMax M3, and GLM-4.6 are first-class targets.
+>
+> The newest models reject request fields that older ones accepted: GPT-6 Astra needs `max_completion_tokens`, takes no `temperature`, and refuses `reasoning_effort` alongside tools on `/chat/completions`; Claude Opus 5.5 and Fable 5.1 reject `temperature` and `budget_tokens` and always think (depth is set with `effort`). Millwright detects these models by ID and sends the right fields — the **Reasoning depth** setting maps onto each model's own controls.
 
 ## Examples
 
@@ -244,7 +246,7 @@ Contributions welcome — see [CONTRIBUTING.md](docs/CONTRIBUTING.md). We especi
 
 - [x] **v0.1** — MVP: Electron shell, LLM adapters, COM bridge, first tool set
 - [x] **v0.2** — Python sidecar, agentic tool loop, dual-engine fallback, vision feedback, confirmation cards, Apache-2.0 open source
-- [x] **v0.2.4 → v0.2.37** — Extensive hardening against real SolidWorks installs ← *current*: the sketch → feature → cut → visual-verification loop now runs end to end on real hardware
+- [x] **v0.2.4 → v0.2.130** — Extensive hardening against real SolidWorks installs ← *current*: the sketch → feature → cut → visual-verification loop now runs end to end on real hardware
 - [ ] **v0.3** — Streaming tool calls, sketching on model faces (not just reference planes), hole wizard, sheet metal, drawing annotations, remaining `#VERIFY` parameters confirmed
 - [ ] **v1.0** — MCP server, multi-CAD support
 

README.zh-CN.md

@@ -27,12 +27,12 @@
 </p>
 
 <p align="center">
-  <img src="https://img.shields.io/badge/version-0.2.37-blue" alt="version" />
+  <img src="https://img.shields.io/badge/version-0.2.130-blue" alt="version" />
   <img src="https://img.shields.io/badge/electron-28-47848F?logo=electron" alt="electron" />
   <img src="https://img.shields.io/badge/react-18-61DAFB?logo=react" alt="react" />
   <img src="https://img.shields.io/badge/typescript-5.3-3178C6?logo=typescript" alt="typescript" />
   <img src="https://img.shields.io/badge/python-3.9%2B-3776AB?logo=python&logoColor=white" alt="python" />
-  <img src="https://img.shields.io/badge/tests-167_JS_%2B_13_Python-brightgreen" alt="tests" />
+  <img src="https://img.shields.io/badge/tests-191_JS_%2B_58_Python-brightgreen" alt="tests" />
   <img src="https://img.shields.io/badge/license-Apache_2.0-orange" alt="license" />
 </p>
 
@@ -85,7 +85,7 @@ Millwright：
 - **Agent 工具循环。** 观察 → 推理 → 执行。模型串联多次工具调用，读取每次返回的结构化 JSON，出错能自愈而不是静默失败。
 - **视觉理解。** 可翻转、旋转、截屏，再做分析——既支持多模态主模型，也支持独立视觉模型。
 - **常驻执行引擎。** 常驻 Python 边车在一整个多步任务中复用同一条 COM 连接。
-- **开发者友好。** 167 个 TS/Node 单元测试，另有独立的 Python 测试套件（`pytest sidecar/tests`），类型化 IPC 边界，`SKIP_SW_CONNECT` 纯 UI 开发模式（无需 SolidWorks）。
+- **开发者友好。** 191 个 TS/Node 单元测试，另有独立的 Python 测试套件（`pytest sidecar/tests`），类型化 IPC 边界，`SKIP_SW_CONNECT` 纯 UI 开发模式（无需 SolidWorks）。
 
 ## 跨版本兼容
 
@@ -159,8 +159,8 @@ npm run dev
 
 | 服务商         | 协议        | Base URL                                            | 推荐模型                                |
 | ----------- | --------- | --------------------------------------------------- | ----------------------------------- |
-| OpenAI      | OpenAI    | `https://api.openai.com/v1`                         | `gpt-5.6-sol`                       |
-| Anthropic   | Anthropic | `https://api.anthropic.com`                         | `claude-fable-5`（旗舰）/ `claude-opus-4-8`（更稳） |
+| OpenAI      | OpenAI    | `https://api.openai.com/v1`                         | `gpt-6-astra`（GPT-6 Astra）         |
+| Anthropic   | Anthropic | `https://api.anthropic.com`                         | `claude-opus-5-5`（Opus 5.5，默认）/ `claude-fable-5-1`（Fable 5.1，最强） |
 | DeepSeek    | OpenAI 兼容 | `https://api.deepseek.com`                          | `deepseek-v4-pro`（强） / `deepseek-v4-flash`（快） |
 | Kimi / 月之暗面 | OpenAI 兼容 | `https://api.moonshot.cn/v1`                        | `kimi-k2.5`                         |
 | MiniMax     | OpenAI 兼容 | `https://api.minimaxi.com/v1`                       | `minimax-m3`（512K 上下文）         |
@@ -169,7 +169,9 @@ npm run dev
 | 硅基流动        | OpenAI 兼容 | `https://api.siliconflow.cn/v1`                     | —（用户自填）                       |
 | Ollama（本地）  | OpenAI 兼容 | `http://localhost:11434/v1`                         | —（用户自填）                       |
 
-> 各家型号更新很快，请以服务商官方文档为准。Agent 工具调用需要模型支持 function calling；DeepSeek V4、Kimi K2、MiniMax M3、GLM-4.6 是一等公民。OpenAI `gpt-5.x` / o 系列和 Anthropic `claude-fable-5` 需要特定字段名，Millwright 已自动识别。
+> 各家型号更新很快，请以服务商官方文档为准。Agent 工具调用需要模型支持 function calling；GPT-6 Astra、Claude Opus 5.5 / Fable 5.1、DeepSeek V4、Kimi K2、MiniMax M3、GLM-4.6 是一等公民。
+>
+> 新一代模型会拒绝老模型能接受的请求字段：GPT-6 Astra 要求 `max_completion_tokens`、不接受 `temperature`，并且在 `/chat/completions` 上不允许 `reasoning_effort` 与工具同时出现；Claude Opus 5.5 和 Fable 5.1 拒绝 `temperature` 和 `budget_tokens`，并且始终开启思考（深度用 `effort` 控制）。Millwright 按模型 ID 自动识别并发送正确的字段——设置里的「推理深度」会映射到各模型自己的控制参数上。OpenAI `gpt-5.x` / o 系列同样已自动识别。
 
 ## 使用示例
 
@@ -243,7 +245,7 @@ SolidWorks
 
 - [x] **v0.1** — MVP：Electron 骨架、LLM 适配器、COM 桥接、首批工具
 - [x] **v0.2** — Python 边车、agent 工具循环、双引擎降级、视觉反馈、确认卡片，Apache-2.0 开源
-- [x] **v0.2.4 → v0.2.37** — 大量真机加固 ← *当前*：草图 → 特征 → 切除 → 视觉核验的完整闭环已在真机上端到端跑通
+- [x] **v0.2.4 → v0.2.130** — 大量真机加固 ← *当前*：草图 → 特征 → 切除 → 视觉核验的完整闭环已在真机上端到端跑通
 - [ ] **v0.3** — 流式工具调用、在模型面上画草图（而非仅基准面）、孔向导、钣金、工程图标注、剩余 `# VERIFY` 参数完成核验
 - [ ] **v1.0** — MCP server、多 CAD 支持
 

docs/ARCHITECTURE.md

@@ -213,8 +213,8 @@ const DEFAULT_SYSTEM_PROMPT = `You are a SolidWorks automation specialist.
 
 | Provider | Protocol | Base URL | Example model |
 |---------|----------|----------|----------|
-| Anthropic | anthropic | https://api.anthropic.com | claude-sonnet-4-20250514 |
-| OpenAI | openai | https://api.openai.com/v1 | gpt-4o |
+| Anthropic | anthropic | https://api.anthropic.com | claude-opus-5-5 |
+| OpenAI | openai | https://api.openai.com/v1 | gpt-6-astra |
 | Bailian | openai | https://dashscope.aliyuncs.com/compatible-mode/v1 | qwen-coder-plus |
 | MiniMax | openai | https://api.minimax.chat/v1 | MiniMax-Text-01 |
 | DeepSeek | openai | https://api.deepseek.com | deepseek-chat |

docs/ARCHITECTURE.zh-CN.md

@@ -213,8 +213,8 @@ const DEFAULT_SYSTEM_PROMPT = `你是一个 SolidWorks 自动化专家助手。
 
 | 服务商 | 协议 | Base URL | 模型示例 |
 |--------|------|----------|----------|
-| Anthropic | anthropic | https://api.anthropic.com | claude-sonnet-4-20250514 |
-| OpenAI | openai | https://api.openai.com/v1 | gpt-4o |
+| Anthropic | anthropic | https://api.anthropic.com | claude-opus-5-5 |
+| OpenAI | openai | https://api.openai.com/v1 | gpt-6-astra |
 | 百炼 | openai | https://dashscope.aliyuncs.com/compatible-mode/v1 | qwen-coder-plus |
 | MiniMax | openai | https://api.minimax.chat/v1 | MiniMax-Text-01 |
 | DeepSeek | openai | https://api.deepseek.com | deepseek-chat |

docs/USER-GUIDE.md

@@ -42,7 +42,7 @@ If you have an Anthropic API key:
 | API Protocol | Anthropic |
 | Base URL | https://api.anthropic.com |
 | API Key | Your `sk-ant-...` key |
-| Model | `claude-sonnet-4-20250514` (recommended) |
+| Model | `claude-opus-5-5` (recommended) or `claude-fable-5-1` (most capable) |
 
 Get your API key by signing up at console.anthropic.com and creating a key.
 
@@ -53,7 +53,7 @@ Get your API key by signing up at console.anthropic.com and creating a key.
 | API Protocol | OpenAI-compatible |
 | Base URL | https://api.openai.com/v1 |
 | API Key | Your `sk-...` key |
-| Model | `gpt-4o` |
+| Model | `gpt-6-astra` (recommended) |
 
 ### Option 3: Bailian (Alibaba Cloud)
 

docs/USER-GUIDE.zh-CN.md

@@ -42,7 +42,7 @@
 | API 协议 | Anthropic |
 | Base URL | https://api.anthropic.com |
 | API Key | 你的 sk-ant-... 密钥 |
-| 模型 | claude-sonnet-4-20250514（推荐） |
+| 模型 | `claude-opus-5-5`（推荐）或 `claude-fable-5-1`（最强） |
 
 获取 API Key：前往 console.anthropic.com 注册并创建密钥。
 
@@ -53,7 +53,7 @@
 | API 协议 | OpenAI 兼容 |
 | Base URL | https://api.openai.com/v1 |
 | API Key | 你的 sk-... 密钥 |
-| 模型 | gpt-4o |
+| 模型 | `gpt-6-astra`（推荐） |
 
 ### 方式三：使用百炼（阿里云）