diff --git a/.env.example b/.env.example
index 6291cdb..e83cdb3 100644
--- a/.env.example
+++ b/.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
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1fbecbe..639368d 100644
--- a/CHANGELOG.md
+++ b/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)
diff --git a/README.md b/README.md
index 11ba339..d3b0b65 100644
--- a/README.md
+++ b/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
 
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 9b4f79c..ddb3835 100644
--- a/README.zh-CN.md
+++ b/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
 
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index fa9d3b0..6262c57 100644
--- a/docs/ARCHITECTURE.md
+++ b/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 |
diff --git a/docs/ARCHITECTURE.zh-CN.md b/docs/ARCHITECTURE.zh-CN.md
index e10d82b..31d900c 100644
--- a/docs/ARCHITECTURE.zh-CN.md
+++ b/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 |
diff --git a/docs/USER-GUIDE.md b/docs/USER-GUIDE.md
index 55f01f4..05e7a21 100644
--- a/docs/USER-GUIDE.md
+++ b/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)
 
diff --git a/docs/USER-GUIDE.zh-CN.md b/docs/USER-GUIDE.zh-CN.md
index 6e1a82d..ec85495 100644
--- a/docs/USER-GUIDE.zh-CN.md
+++ b/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）
 
diff --git a/package.json b/package.json
index f4bc3c5..eab98c6 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "millwright",
-  "version": "0.2.129",
+  "version": "0.2.130",
   "description": "Open-source AI automation for SolidWorks — talk to your CAD.",
   "keywords": [
     "solidworks",
@@ -30,7 +30,7 @@
     "pack": "npm run build && electron-builder --dir",
     "dist": "powershell -ExecutionPolicy Bypass -File scripts/prepare-python.ps1 && npm run build && electron-builder --publish never",
     "typecheck": "tsc -p tsconfig.main.json --noEmit && tsc -p tsconfig.preload.json --noEmit && tsc -p tsconfig.renderer.json --noEmit",
-    "test": "npm run build:main && node --test tests/agent-loop.test.mjs tests/code-extract.test.mjs tests/env-fallback.test.mjs tests/errors.test.mjs tests/factory.test.mjs tests/generators.test.mjs tests/presets.test.mjs tests/sanitizer.test.mjs tests/sse.test.mjs tests/sw-tools.test.mjs tests/vba-helpers.test.mjs tests/vba-macro-writer.test.mjs tests/housekeeping.test.mjs",
+    "test": "npm run build:main && node --test tests/agent-loop.test.mjs tests/code-extract.test.mjs tests/env-fallback.test.mjs tests/errors.test.mjs tests/factory.test.mjs tests/generators.test.mjs tests/presets.test.mjs tests/sanitizer.test.mjs tests/sse.test.mjs tests/sw-tools.test.mjs tests/vba-helpers.test.mjs tests/vba-macro-writer.test.mjs tests/housekeeping.test.mjs tests/llm-request.test.mjs tests/sidecar-client.test.mjs",
     "lint": "eslint src/ --ext .ts,.tsx",
     "format": "prettier --write src/",
     "clean": "rimraf dist release"
diff --git a/sidecar/sw_agent/server.py b/sidecar/sw_agent/server.py
index 6c73478..0bff7fd 100644
--- a/sidecar/sw_agent/server.py
+++ b/sidecar/sw_agent/server.py
@@ -28,6 +28,7 @@
 import threading
 import time
 from collections import OrderedDict
+from concurrent.futures import TimeoutError as FutureTimeout
 
 from sw_agent import registry, session_log, verify
 from sw_agent.bridge import Context, SWError, hresult, is_dead_connection
@@ -97,6 +98,18 @@ def bump(self) -> int:
 GUARD = _Guard()
 
 
+class _Failed:
+    """P131: a failed call, cached under its op_id like a result. A retry with the same
+    op_id (Node's TIMEOUT follow-up) must get the failure back — before P131 only
+    successes were cached, so the retry RE-RAN a generator that had built half a gear
+    and then raised."""
+
+    __slots__ = ("exc",)
+
+    def __init__(self, exc: Exception) -> None:
+        self.exc = exc
+
+
 class CodedError(Exception):
     def __init__(self, code: str, message: str) -> None:
         super().__init__(message)
@@ -164,6 +177,8 @@ def _advisory(ctx: Context) -> dict | None:
 def _call(ctx: Context, name: str, args: dict, op_id=None, expect_state=None):
     """One tool call on the COM thread. Dead connection → reconnect once and retry."""
     cached = GUARD.get(op_id)
+    if isinstance(cached, _Failed):
+        raise cached.exc
     if cached is not None:
         dup = dict(cached) if isinstance(cached, dict) else {"result": cached}
         dup["_duplicate"] = True
@@ -205,9 +220,14 @@ def work():
         return work()
     except Exception as e:
         if not is_dead_connection(e):
+            GUARD.put(op_id, _Failed(e))
             raise
-        ctx.reconnect()
+    ctx.reconnect()
+    try:
         return work()
+    except Exception as e:
+        GUARD.put(op_id, _Failed(e))
+        raise
 
 
 HEARTBEAT_S = 15.0
@@ -222,7 +242,10 @@ def _call_with_heartbeat(executor, ctx, rid, name, args, op_id, expect):
     while True:
         try:
             return fut.result(timeout=HEARTBEAT_S)
-        except TimeoutError:
+        except (TimeoutError, FutureTimeout):
+            # P131: before Python 3.11 concurrent.futures.TimeoutError is NOT the builtin
+            # TimeoutError — on 3.9/3.10 the first tick escaped as an empty TOOL_FAILED
+            # while the tool kept running (and the agent retried it).
             _write({"id": rid, "progress": True, "tool": name,
                     "elapsed_ms": round((time.perf_counter() - t0) * 1000)})
 
@@ -246,6 +269,27 @@ def _health(ctx: Context) -> dict:
     return info
 
 
+def _parent_alive(ppid: int) -> bool:
+    """P131: is the process that spawned us still running? Checked by PID, never via stdin."""
+    if os.name == "nt":
+        import ctypes
+        from ctypes import wintypes
+        k32 = ctypes.WinDLL("kernel32", use_last_error=True)
+        k32.OpenProcess.argtypes = (wintypes.DWORD, wintypes.BOOL, wintypes.DWORD)
+        k32.OpenProcess.restype = wintypes.HANDLE
+        k32.WaitForSingleObject.argtypes = (wintypes.HANDLE, wintypes.DWORD)
+        k32.WaitForSingleObject.restype = wintypes.DWORD
+        k32.CloseHandle.argtypes = (wintypes.HANDLE,)
+        h = k32.OpenProcess(0x00100000, False, ppid)  # SYNCHRONIZE
+        if not h:
+            return ctypes.get_last_error() != 87  # ERROR_INVALID_PARAMETER: no such process
+        try:
+            return k32.WaitForSingleObject(h, 0) == 0x102  # WAIT_TIMEOUT: still running
+        finally:
+            k32.CloseHandle(h)
+    return os.getppid() == ppid  # POSIX: an orphan is re-parented
+
+
 def serve() -> None:
     ctx = Context()
     executor = ComExecutor("sw-com")
@@ -256,22 +300,23 @@ def serve() -> None:
     # blocks while SolidWorks loads add-ins sits at the HEAD of the queue and delays every
     # probe behind it — that is how the first sw_status of the 16:15 session timed out.
     # Connecting lazily costs a few hundred ms on the first real call and blocks nothing.
+    # If the parent dies mid-COM-call the stdin loop never sees EOF. A helper thread watches
+    # the parent PID: once it is gone, give the current job 5 s and hard-exit.
+    # P131: the P129 version peeked stdin from this thread. Two readers on one pipe — after
+    # ~2 s idle the peek parked inside the buffered reader, and the request that followed the
+    # next one was not read until more input arrived: back-to-back tool calls stalled into
+    # 60 s timeouts. The watchdog must never touch stdin.
+    ppid = os.getppid()
+
     def _watchdog():
-        # If the parent dies mid-COM-call the stdin loop never sees EOF. Poll the pipe from a
-        # helper thread: once stdin is closed, give the current job 5 s and hard-exit.
         try:
-            while not sys.stdin.closed:
+            while _parent_alive(ppid):
                 time.sleep(2)
-                try:
-                    if sys.stdin.buffer.peek(1) == b"":
-                        break
-                except Exception:  # noqa: BLE001 — peek on a closed stdin pipe is the trigger we care about
-                    break
-        except Exception:  # noqa: BLE001 — watchdog must never crash the sidecar itself
-            pass
+        except Exception:  # noqa: BLE001 — a watchdog bug must never kill a healthy sidecar
+            return
         time.sleep(5)
         os._exit(0)
-    threading.Thread(target=_watchdog, name="stdin-watchdog", daemon=True).start()
+    threading.Thread(target=_watchdog, name="parent-watchdog", daemon=True).start()
     for raw in sys.stdin:
         raw = raw.strip()
         if not raw:
diff --git a/sidecar/tests/test_reliability.py b/sidecar/tests/test_reliability.py
index 8eeebcd..f81c63e 100644
--- a/sidecar/tests/test_reliability.py
+++ b/sidecar/tests/test_reliability.py
@@ -15,7 +15,9 @@
 def test_guard_idempotency_and_bump():
     g = server._Guard(cap=2)
     assert g.get(None) is None and g.get("x") is None
-    g.put("a", {"r": 1}); g.put("b", {"r": 2}); g.put("c", {"r": 3})
+    g.put("a", {"r": 1})
+    g.put("b", {"r": 2})
+    g.put("c", {"r": 3})
     assert g.get("a") is None and g.get("c") == {"r": 3}  # LRU cap
     assert g.bump() == 1 and g.bump() == 2
 
@@ -60,3 +62,61 @@ class Ctx:  # never reached: STALE_STATE is checked before any COM access
         server._call(Ctx(), "extrude", {}, op_id=None, expect_state=3)
     assert ei.value.code == "STALE_STATE"
     server.GUARD.state_version = 0
+
+
+def test_failed_call_is_cached_under_op_id():
+    """P131: a TIMEOUT follow-up with the same op_id must get the failure back, not re-run."""
+    from sw_agent import registry
+
+    runs = []
+
+    def flaky(ctx):
+        runs.append(1)
+        raise SWError("gear body built, then the hole failed")
+
+    registry.TOOLS["_p131_flaky"] = registry.ToolSpec("_p131_flaky", "", {}, "", False, True, flaky)
+    try:
+        for _ in range(2):
+            with pytest.raises(SWError):
+                server._call(object(), "_p131_flaky", {}, op_id="p131-op", expect_state=None)
+        assert len(runs) == 1
+    finally:
+        registry.TOOLS.pop("_p131_flaky", None)
+        server.GUARD.done.pop("p131-op", None)
+
+
+def test_heartbeat_survives_futures_timeout(monkeypatch):
+    """P131: on Python < 3.11 Future.result(timeout) raises concurrent.futures.TimeoutError,
+    which is not the builtin — the first heartbeat tick used to escape as TOOL_FAILED."""
+    from concurrent.futures import Future
+    from concurrent.futures import TimeoutError as FutureTimeout
+
+    frames = []
+    monkeypatch.setattr(server, "_write", frames.append)
+
+    class SlowFuture(Future):
+        ticks = 0
+
+        def result(self, timeout=None):
+            SlowFuture.ticks += 1
+            if SlowFuture.ticks < 3:
+                raise FutureTimeout()  # what 3.9 / 3.10 raise
+            return {"ok": True}
+
+    class Exec:
+        def submit(self, fn):
+            return SlowFuture()
+
+    assert server._call_with_heartbeat(Exec(), None, 7, "slow", {}, None, None) == {"ok": True}
+    assert [f["progress"] for f in frames] == [True, True]
+    assert all(f["id"] == 7 for f in frames)
+
+
+def test_parent_watchdog_uses_pid_not_stdin():
+    """P131: the P129 watchdog peeked stdin from a second thread and stalled requests."""
+    import inspect
+
+    assert server._parent_alive(os.getppid()) is True
+    if os.name != "nt":
+        assert server._parent_alive(os.getppid() + 999_999) is False
+    assert "peek(" not in inspect.getsource(server.serve)
diff --git a/src/main/agent/agent-loop-sidecar.ts b/src/main/agent/agent-loop-sidecar.ts
index b3a0996..96d081b 100644
--- a/src/main/agent/agent-loop-sidecar.ts
+++ b/src/main/agent/agent-loop-sidecar.ts
@@ -50,6 +50,9 @@ export interface AgentEvent {
 
 export interface SidecarAgentOptions {
   requestId?: string;
+  /** P131: model id + the system prompt the adapter sends — truncation budgets for them */
+  model?: string;
+  systemPrompt?: string;
   maxRounds?: number;
   signal?: AbortSignal;
   onEvent?: (ev: AgentEvent) => void;
@@ -189,6 +192,10 @@ export async function runSidecarAgent(
   let backupDone = false;
   // P125: surface SolidWorks version advisory once per session
   let advisoryShown = false;
+  // P131: op_ids are namespaced per run. The sidecar's idempotency cache outlives a chat
+  // session, and some providers restart their tool-call ids every conversation (Kimi's
+  // `functions.<name>:<n>`) — a bare call.id could hand a NEW call an OLD session's result.
+  const runNonce = Math.random().toString(36).slice(2, 10);
 
   // P19: cache the most recent screenshot so a (pure-text or multimodal) model can
   // ask several follow-up questions about the SAME snapshot without re-capturing.
@@ -236,6 +243,8 @@ export async function runSidecarAgent(
     ...VIRTUAL_TOOLS.filter((t) => t.function.name !== 'run_macro' || !!opts.runMacro),
     ...sidecarTools,
   ].filter((t: any) => !off.has(t?.function?.name));
+  // P131: text that rides along with every request besides the history
+  const overhead = (opts.systemPrompt ?? '') + JSON.stringify(tools);
   // P115: every advertised tool name — the audit scans narration for these.
   const toolNames = new Set(tools.map((t: any) => t?.function?.name).filter(Boolean));
   const destructive = new Set(
@@ -291,7 +300,9 @@ export async function runSidecarAgent(
 
   for (let round = 0; round < maxRounds; round++) {
     if (opts.signal?.aborted) throw new Error('已取消');
-    history = truncateMessages(history, '', opts.requestId ? String(opts.requestId) : '', opts.contextWindow);
+    // P131: the request also carries the system prompt and every tool schema — count them
+    // (they never were: the prompt slot got '' and the model slot got the request id).
+    history = truncateMessages(history, overhead, opts.model ?? '', opts.contextWindow);
     // Thin guard (block-aware truncation should already prevent this)
     while (history.length > 0 && (history[0].role === 'tool' || (history[0].role === 'system' && history[0].toolCalls?.length))) {
       history.shift();
@@ -484,13 +495,21 @@ export async function runSidecarAgent(
         // P130: onStillRunning callback fires every time the sidecar sends a progress
         // heartbeat; we surface a "still running" note in the chat so the user knows the
         // tool is genuinely progressing instead of just hanging.
+        // P131: one note per minute of waiting (heartbeats arrive every 15 s), and Stop
+        // abandons the wait instead of leaving the run stuck until the app restarts.
+        let lastNote = 0;
         const r = await sidecar.call(call.name, call.parameters, {
-          opId: call.id,
-          onStillRunning: (ms) => opts.onEvent?.({ type: 'text', text: `（${call.name} 已运行 ${Math.round(ms / 1000)} s，SolidWorks 仍在执行，继续等待…）` }),
+          opId: `${runNonce}:${call.id}`,
+          signal: opts.signal,
+          onStillRunning: (ms) => {
+            if (ms - lastNote < 60_000) return;
+            lastNote = ms;
+            opts.onEvent?.({ type: 'text', text: `\n（${call.name} 已运行 ${Math.round(ms / 1000)} s，SolidWorks 仍在执行，继续等待…）\n` });
+          },
         });
         // P130: stash the wall-clock duration on the call so the UI card and the session
         // export can show it.
-        (call as any).durationMs = r.durationMs;
+        call.durationMs = r.durationMs;
         // P125: surface unverified SolidWorks version advisory once
         if (r.ok && r.data?._advisory && !advisoryShown) {
           advisoryShown = true;
@@ -517,7 +536,7 @@ export async function runSidecarAgent(
     // from a path not covered by the per-call guard above), the API rejects the next
     // request with "assistant message with 'tool_calls' must be followed by tool
     // messages". Scan backwards from the tail and answer any orphaned call_ids.
-    history = truncateMessages(history, '', '', opts.contextWindow);
+    history = truncateMessages(history, overhead, opts.model ?? '', opts.contextWindow);
     for (let i = history.length - 1; i >= 0; i--) {
       const m = history[i];
       if (m.role !== 'assistant' || !m.toolCalls?.length) continue;
@@ -536,7 +555,10 @@ export async function runSidecarAgent(
       role: 'system',
       content: '(已达到最大工具调用轮数。请不要再调用工具：总结你已完成的操作、当前模型的状态、以及未完成的部分。)',
     });
-    const summary = await adapter.chatWithTools(history, opts.signal, undefined);
+    // P131: pass the real tool list — Anthropic rejects a history containing tool_use /
+    // tool_result blocks when the request declares no tools, so on every Claude model
+    // this summary 400'd and was silently lost. The system note above forbids calls.
+    const summary = await adapter.chatWithTools(history, opts.signal, tools);
     if (summary.content) {
       finalText = summary.content;
       opts.onEvent?.({ type: 'text', text: summary.content });
@@ -775,8 +797,20 @@ function auditClaimedCalls(
  *  auto-nudged, and built a gear from the example values the user never confirmed. */
 export function looksLikeQuestion(text: string): boolean {
   const t = text.trim();
-  if (/[?？]\s*$/.test(t)) return true;                                   // ends with a question mark
+  if (!t) return false;
+  if (/[?？]\s*[)）"”'」]*\s*$/.test(t)) return true;                     // ends with a question mark
   if ((t.match(/[?？]/g) || []).length >= 2) return true;                 // several questions inside
-  return /(请(告诉|提供|给出|确认|回复|选择)|需要(你|您)?(提供|确认|告诉)|告诉我|你想|您想|哪(一)?(种|个)|多少|几个|which|what (size|module|value|diameter)|please (provide|tell|confirm|specify)|let me know|do you want|would you like|could you (tell|confirm))/i.test(t);
+  // P131: judge the ENDING. The P130 pattern matched generic words anywhere (which / 多少 /
+  // 几个 / 哪个 / 你想), so ordinary plans read as questions and the nudge never fired; and
+  // "我还缺少以下参数…" / "I need a few values from you" were missed.
+  const tail = t.slice(-300);
+  if (PLAN_CLOSE.test(tail)) return false;
+  return ASKS_USER.test(tail);
 }
 
+/** "…Starting now." / "…现在开始执行。" — the reply commits to acting. */
+const PLAN_CLOSE = /(现在开始|开始执行|开始建模|马上开始|立即开始|接下来(我)?(会|将)?(调用|执行|开始)|starting now|let me (start|begin)|i(?:'ll| will) (?:now )?(?:start|begin|proceed)|proceeding now)[^。.!！\n]*[。.!！]?\s*$/i;
+
+/** The reply hands the turn back: it asks for input the user has not given. */
+const ASKS_USER = /(请(告诉|提供|给出|确认|回复|选择|补充|说明)|需要(你|您)?(提供|确认|告诉|补充)|告诉我|还?缺少?(以下|这些|下列)?(参数|尺寸|信息|数值)|需要以下|请问|(你|您)希望|是否需要|please (provide|tell|confirm|specify|let me know)|let me know|i need (a few|some|the following|these|you to|more)|(could|can) you (tell|confirm|provide|specify)|do you want|would you like|what (size|module|value|diameter|dimensions?) (do|would|should))/i;
+
diff --git a/src/main/com/sw-sidecar.ts b/src/main/com/sw-sidecar.ts
index dbb0f24..020a937 100644
--- a/src/main/com/sw-sidecar.ts
+++ b/src/main/com/sw-sidecar.ts
@@ -16,6 +16,16 @@
 //       sidecar's idempotency cache hands back the true result when the job finishes
 //     - `inFlight` + `lastStatus` let the UI probe skip the queue while a tool is running
 //     - every rpc logs its duration so the next report carries numbers, not guesses
+//
+// P131 — what P130 left broken:
+//     - heartbeats prove PYTHON is alive, not that SolidWorks progresses: a COM call blocked
+//       on a modal dialog heartbeated forever and nothing timed out. Each call now also has
+//       an absolute cap (3× its budget; sw_status none — the probe must fail fast).
+//     - the sidecar reads stdin one request at a time, so a request queued behind a slow
+//       call burned its budget before it was even read, timed out, and its retry re-ran the
+//       tool. Requests are now sent one at a time; a budget starts when its request is sent.
+//     - `onStillRunning` fired only after a timeout that heartbeats prevented — it now fires
+//       on every heartbeat. A `signal` lets Stop abandon the wait (code CANCELLED).
 
 import { spawn, ChildProcessWithoutNullStreams } from 'child_process';
 import { randomUUID } from 'crypto';
@@ -27,7 +37,7 @@ import { resolvePythonPath, resolveSidecarCwd } from '../python-path';
 
 export type SidecarErrorCode =
   | 'NO_CONNECTION' | 'NO_DOCUMENT' | 'WRONG_DOC_TYPE' | 'UNKNOWN_TOOL'
-  | 'BAD_ARGS' | 'STALE_STATE' | 'COM_ERROR' | 'TOOL_FAILED' | 'TIMEOUT';
+  | 'BAD_ARGS' | 'STALE_STATE' | 'COM_ERROR' | 'TOOL_FAILED' | 'TIMEOUT' | 'CANCELLED';
 
 export interface SidecarResult<T = any> {
   ok: boolean;
@@ -54,18 +64,32 @@ export interface CallOptions {
   expectState?: number;
   /** Idle budget for this call (ms). Defaults: SLOW_TOOLS table, then callTimeoutMs. */
   timeoutMs?: number;
-  /** Fired when the first budget expires and we keep waiting on the same op_id. */
+  /** Fired on every sidecar heartbeat while the call is still running. */
   onStillRunning?: (elapsedMs: number) => void;
+  /** P131: abort → resolve CANCELLED at once. The sidecar finishes the job regardless and
+   *  caches the outcome under the op_id. */
+  signal?: AbortSignal;
+}
+
+interface RpcOptions {
+  /** Idle budget (ms) — heartbeats reset it. */
+  budget?: number;
+  /** P131: absolute deadline (ms) that heartbeats do NOT extend. Defaults to `budget`. */
+  hardCapMs?: number;
+  onProgress?: (elapsedMs: number) => void;
+  signal?: AbortSignal;
 }
 
 interface Pending {
   resolve: (v: SidecarResult) => void;
   reject: (e: Error) => void;
   timer: ReturnType<typeof setTimeout>;
+  hardTimer?: ReturnType<typeof setTimeout>;
   budget: number;
   startedAt: number;
   label: string;
   onTimeout: () => void;
+  onProgress?: (elapsedMs: number) => void;
 }
 
 interface ReadyWaiter { resolve: () => void; reject: (e: Error) => void }
@@ -108,6 +132,8 @@ export class SWSidecar {
   lastStatus: any = null;
   /** P130: name of the tool currently executing (first in-flight `call`), for the UI */
   runningTool: string | null = null;
+  /** P131: requests go out one at a time — the sidecar reads stdin serially anyway */
+  private tail: Promise<unknown> = Promise.resolve();
 
   constructor(opts: SidecarOptions = {}) {
     this.opts = {
@@ -187,11 +213,14 @@ export class SWSidecar {
     if (msg.progress) {
       clearTimeout(p.timer);
       p.timer = setTimeout(p.onTimeout, p.budget);
-      this.opts.onLog?.(`[sidecar] ${p.label} still running (${Math.round((msg.elapsed_ms ?? 0) / 1000)}s)`);
+      const elapsed = Date.now() - p.startedAt;
+      this.opts.onLog?.(`[sidecar] ${p.label} still running (${Math.round(elapsed / 1000)}s)`);
+      try { p.onProgress?.(elapsed); } catch { /* a UI callback must never break the RPC */ }
       return;
     }
     this.pending.delete(msg.id);
     clearTimeout(p.timer);
+    clearTimeout(p.hardTimer);
     this.updateRunning();
     const durationMs = Date.now() - p.startedAt;
     this.opts.onLog?.(`[sidecar] ${p.label} ${durationMs}ms ok=${!!msg.ok}${msg.code ? ' code=' + msg.code : ''}`);
@@ -199,27 +228,53 @@ export class SWSidecar {
   }
 
   private updateRunning(): void {
-    const first = [...this.pending.values()].find((p) => p.label.startsWith('call:'));
+    // sw_status is the UI's own probe — never report it as "a tool is running"
+    const first = [...this.pending.values()].find((p) => p.label.startsWith('call:') && p.label !== 'call:sw_status');
     this.runningTool = first ? first.label.slice(5) : null;
   }
 
-  private rpc(method: string, params?: any, budget?: number): Promise<SidecarResult> {
+  /** P131: queue behind whatever is outstanding; resolve CANCELLED as soon as `signal`
+   *  aborts (a request not yet sent is then never sent). */
+  private rpc(method: string, params?: any, o: RpcOptions = {}): Promise<SidecarResult> {
+    return new Promise<SidecarResult>((resolve) => {
+      let settled = false;
+      const settle = (r: SidecarResult) => {
+        if (settled) return;
+        settled = true;
+        o.signal?.removeEventListener('abort', onAbort);
+        resolve(r);
+      };
+      const onAbort = () => settle({ ok: false, code: 'CANCELLED', error: '已取消' });
+      if (o.signal?.aborted) return onAbort();
+      o.signal?.addEventListener('abort', onAbort, { once: true });
+      const turn = this.tail.then(() => (settled ? undefined : this.send(method, params, o).then(settle)));
+      this.tail = turn.catch(() => undefined);
+    });
+  }
+
+  private send(method: string, params: any, o: RpcOptions): Promise<SidecarResult> {
     if (!this.proc || !this.proc.stdin.writable) {
       return Promise.resolve({ ok: false, code: 'NO_CONNECTION', error: 'Python 组件未运行——请安装 Python + pywin32，或忽略此错误（将自动使用内置 VBS 引擎）' });
     }
     const id = this.nextId++;
     const label = method === 'call' ? `call:${params?.name ?? '?'}` : method;
-    const b = budget ?? this.opts.callTimeoutMs;
+    const b = o.budget ?? this.opts.callTimeoutMs;
+    const cap = Math.max(o.hardCapMs ?? b, b);
     return new Promise<SidecarResult>((resolve, reject) => {
       const startedAt = Date.now();
       const onTimeout = () => {
+        const p = this.pending.get(id);
+        if (!p) return;
+        clearTimeout(p.timer);
+        clearTimeout(p.hardTimer);
         this.pending.delete(id);
         this.updateRunning();
-        this.opts.onLog?.(`[sidecar] ${label} TIMEOUT after ${Date.now() - startedAt}ms (budget ${b}ms, pending=${this.pending.size}, running=${this.runningTool ?? '-'})`);
+        this.opts.onLog?.(`[sidecar] ${label} TIMEOUT after ${Date.now() - startedAt}ms (budget ${b}ms, cap ${cap}ms, pending=${this.pending.size}, running=${this.runningTool ?? '-'})`);
         resolve({ ok: false, code: 'TIMEOUT', error: `Python 组件调用超时：${label}`, durationMs: Date.now() - startedAt });
       };
       const timer = setTimeout(onTimeout, b);
-      this.pending.set(id, { resolve, reject, timer, budget: b, startedAt, label, onTimeout });
+      const hardTimer = cap > b ? setTimeout(onTimeout, cap) : undefined;
+      this.pending.set(id, { resolve, reject, timer, hardTimer, budget: b, startedAt, label, onTimeout, onProgress: o.onProgress });
       this.updateRunning();
       this.proc!.stdin.write(JSON.stringify({ id, method, params: params ?? {} }) + '\n');
     });
@@ -233,7 +288,8 @@ export class SWSidecar {
   }
 
   /** Invoke a tool. A timeout is retried once with the SAME op_id (P125 idempotency cache):
-   *  if the sidecar finished the job meanwhile, we get the real result, not a re-run. */
+   *  if the sidecar finished the job meanwhile, we get the real result, not a re-run
+   *  (P131: failures are cached too, so a half-built generator is never re-run). */
   async call(name: string, args?: Record<string, any>, opts?: CallOptions): Promise<SidecarResult> {
     const opId = opts?.opId ?? randomUUID();
     const params: any = { name, args: args ?? {} };
@@ -241,25 +297,33 @@ export class SWSidecar {
     if (opts?.expectState != null) params.expect_state = opts.expectState;
     const budget = opts?.timeoutMs ?? SLOW_TOOLS[name] ?? this.opts.callTimeoutMs;
     const t0 = Date.now();
-    let r = await this.rpc('call', params, budget);
+    const o: RpcOptions = {
+      budget,
+      // P131: heartbeats may stretch a call to 3× its budget, never further; the probe not at all
+      hardCapMs: name === 'sw_status' ? budget : budget * 3,
+      onProgress: opts?.onStillRunning ? () => opts.onStillRunning!(Date.now() - t0) : undefined,
+      signal: opts?.signal,
+    };
+    let r = await this.rpc('call', params, o);
     if (!r.ok && r.code === 'TIMEOUT' && name !== 'sw_status') {
       opts?.onStillRunning?.(Date.now() - t0);
-      r = await this.rpc('call', params, budget);   // same op_id → cached result when the job lands
+      // same op_id → cached result when the job lands; one more budget, no extension
+      r = await this.rpc('call', params, { ...o, hardCapMs: budget });
       if (!r.ok && r.code === 'TIMEOUT') {
         r.error = `工具 ${name} 执行超过 ${Math.round((Date.now() - t0) / 1000)} s 仍未返回。SolidWorks 可能在重建、加载插件或弹出了对话框；`
           + `操作可能已经完成——请先用 list_features 核对，不要重复执行。`;
       }
     }
     if (name === 'sw_status' && r.ok) this.lastStatus = r.data;
-    if (r.durationMs == null) r.durationMs = Date.now() - t0;
+    r.durationMs = Date.now() - t0;   // P131: the whole wait, including a follow-up
     return r;
   }
 
-  ping(): Promise<SidecarResult> { return this.rpc('ping', undefined, 10_000); }
-  reconnect(): Promise<SidecarResult> { return this.rpc('reconnect', undefined, 30_000); }
+  ping(): Promise<SidecarResult> { return this.rpc('ping', undefined, { budget: 10_000 }); }
+  reconnect(): Promise<SidecarResult> { return this.rpc('reconnect', undefined, { budget: 30_000 }); }
 
   async health(): Promise<SidecarResult<SidecarHealth>> {
-    const r = await this.rpc('health', undefined, 10_000);
+    const r = await this.rpc('health', undefined, { budget: 10_000 });
     if (!r.ok && /unknown method/.test(r.error ?? '')) {
       const p = await this.ping();
       return { ok: p.ok, data: { connected: p.ok, state_version: 0, tool_count: 0 } as SidecarHealth, error: p.error };
@@ -268,7 +332,7 @@ export class SWSidecar {
   }
 
   async stateVersion(): Promise<number> {
-    const r = await this.rpc('state', undefined, 10_000);
+    const r = await this.rpc('state', undefined, { budget: 10_000 });
     return r.ok ? Number(r.data?.state_version ?? 0) : 0;
   }
 
@@ -286,6 +350,7 @@ export class SWSidecar {
     this.rl = null;
     for (const [, p] of this.pending) {
       clearTimeout(p.timer);
+      clearTimeout(p.hardTimer);
       p.resolve({ ok: false, code: 'NO_CONNECTION', error: err.message });
     }
     this.pending.clear();
diff --git a/src/main/ipc/handlers.ts b/src/main/ipc/handlers.ts
index 11dfb1e..031b5d8 100644
--- a/src/main/ipc/handlers.ts
+++ b/src/main/ipc/handlers.ts
@@ -13,7 +13,7 @@ import { createAdapter, validateConfig } from '../llm';
 import { truncateMessages } from '../llm/context-window';
 import { resolveSystemPrompt } from '../llm/prompts';
 import { getBridge } from '../com/sw-bridge';
-import { getSidecar } from '../com/sw-sidecar';
+import { getSidecar, type SidecarResult } from '../com/sw-sidecar';
 import { collectDocumentContext, formatContextForPromptAsync, invalidateContextCache } from '../com/context-collector';
 import { ScriptEngine } from '../scripts/engine';
 import { validateScript } from '../scripts/sanitizer';
@@ -120,6 +120,7 @@ export function registerIpcHandlers(getMainWindow: () => BrowserWindow | null) {
     return { ok: true };
   });
 
+  let statusProbe: Promise<SidecarResult> | null = null;
   ipcMain.handle(IpcChannels.SW_STATUS, async () => {
     // P73: the sidecar holds the connection the tools actually run through. If it can read
     // ActiveDoc we ARE connected, whatever the separate cscript probe concludes — and it was
@@ -130,10 +131,16 @@ export function registerIpcHandlers(getMainWindow: () => BrowserWindow | null) {
       // P130: a tool is executing on the sidecar's single COM thread — do not queue behind it.
       // Serve the last known status and say we are busy; the UI shows "working" instead of a
       // red/green flicker, and the probe can no longer time out in the queue.
-      if (sidecar.runningTool && sidecar.lastStatus) {
-        return { ...sidecar.lastStatus, source: 'sidecar' as const, busy: true, runningTool: sidecar.runningTool };
+      // P131: also when no status was cached yet (a probe would only queue behind the tool),
+      // and always connected — a tool is running through this very connection, whatever
+      // an older probe said.
+      if (sidecar.runningTool) {
+        return { ...(sidecar.lastStatus ?? {}), connected: true, source: 'sidecar' as const, busy: true, runningTool: sidecar.runningTool };
       }
-      const r = await sidecar.call('sw_status', {});
+      // P131: the UI polls every 3 s; while SolidWorks is slow to answer, share the probe in
+      // flight instead of queueing another one behind it each tick.
+      statusProbe ??= sidecar.call('sw_status', {}).finally(() => { statusProbe = null; });
+      const r = await statusProbe;
       if (r.ok && r.data?.connected) {
         return { ...r.data, source: 'sidecar' as const };
       }
@@ -317,6 +324,8 @@ export function registerIpcHandlers(getMainWindow: () => BrowserWindow | null) {
       if (sidecarReady) {
         const text = await runSidecarAgent(adapter, payload.messages, sidecar, {
           requestId,
+          model: enrichedConfig.model,
+          systemPrompt: enrichedConfig.systemPrompt,
           maxRounds: payload.config.maxRounds ?? 24,
           approvalMode: payload.config.approvalMode ?? 'normal',
           // P115: strong-prompt mode — L3 re-injects tool rules every round +
@@ -477,23 +486,20 @@ export function registerIpcHandlers(getMainWindow: () => BrowserWindow | null) {
   });
 
   // ===== Theme (kept independent of the LLM config) =====
-  // Reusing the CONFIG_ channel-name convention felt too cramped, so we register
-  // two dedicated handlers here. The channel names piggyback on `config:save/load`
-  // for now; we can split them out later if needed.
-  ipcMain.handle('theme:load', async (): Promise<ThemeName> => {
+  ipcMain.handle(IpcChannels.THEME_LOAD, async (): Promise<ThemeName> => {
     return await loadTheme();
   });
 
-  ipcMain.handle('theme:save', async (_e, theme: ThemeName) => {
+  ipcMain.handle(IpcChannels.THEME_SAVE, async (_e, theme: ThemeName) => {
     await saveTheme(theme);
     return { ok: true };
   });
 
-  ipcMain.handle('locale:load', async (): Promise<LocaleName> => {
+  ipcMain.handle(IpcChannels.LOCALE_LOAD, async (): Promise<LocaleName> => {
     return await loadLocale();
   });
 
-  ipcMain.handle('locale:save', async (_e, locale: LocaleName) => {
+  ipcMain.handle(IpcChannels.LOCALE_SAVE, async (_e, locale: LocaleName) => {
     await saveLocale(locale);
     return { ok: true };
   });
diff --git a/src/main/llm/anthropic.ts b/src/main/llm/anthropic.ts
index 322fe3f..f15463e 100644
--- a/src/main/llm/anthropic.ts
+++ b/src/main/llm/anthropic.ts
@@ -23,7 +23,7 @@ import { resolveSystemPrompt } from './prompts';
 import { extractFirstCodeBlock } from './code-extract';
 import { LLMHttpError, extractErrorMessage, toLLMError } from './errors';
 import { parseSSE } from './sse';
-import { splitThinking } from './thinking';
+import { splitThinking, anthropicReasoningParams } from './thinking';
 import type {
   ChatMessage,
   LLMResponse,
@@ -45,6 +45,7 @@ interface AnthropicResponseBody {
   content: AnthropicTextContent[];
   model: string;
   stop_reason: string | null;
+  stop_details?: { category?: string | null } | null;
   usage?: {
     input_tokens: number;
     output_tokens: number;
@@ -67,19 +68,14 @@ export class AnthropicAdapter extends BaseLLMAdapter {
   }
 
   /**
-   * P53: extended thinking. Anthropic takes a budget, not a level, and requires
-   * budget_tokens < max_tokens. 'auto' sends nothing (model default).
+   * P53: reasoning depth. P131: per-model request surface (see anthropicReasoningParams)
+   * — current Claude models reject `temperature` and `budget_tokens` with a 400, so
+   * sampling and thinking fields are chosen together, by model id.
    */
-  private thinkingExtras(): Record<string, any> {
-    const lv = this.config.reasoningLevel ?? 'auto';
-    if (lv === 'auto' || lv === 'adaptive') return {};  // P54: no Anthropic equivalent of 'adaptive'
-    // P55: the Mythos-class models (Fable 5 / Mythos 5) run always-on adaptive thinking
-    // and reject an attempt to disable it — leave them on the provider default.
-    if (/fable|mythos/i.test(this.config.model ?? '')) return {};
-    if (lv === 'off') return { thinking: { type: 'disabled' } };
-    const budgets: Record<string, number> = { low: 1024, medium: 4096, high: 16384 };
-    const budget = Math.min(budgets[lv] ?? 4096, Math.max(1024, this.maxTokens() - 1024));
-    return { thinking: { type: 'enabled', budget_tokens: budget } };
+  private requestExtras(): Record<string, any> {
+    return anthropicReasoningParams(
+      this.config.reasoningLevel, this.config.model, this.maxTokens(), this.config.temperature,
+    );
   }
 
   private buildBody(messages: ChatMessage[], stream: boolean) {
@@ -91,11 +87,10 @@ export class AnthropicAdapter extends BaseLLMAdapter {
     return {
       model: this.config.model,
       max_tokens: this.maxTokens(),
-      temperature: this.config.temperature ?? 0.3,
       system: systemPrompt,
       stream,
       messages: rest.map((m) => ({ role: m.role, content: m.content })),
-      ...this.thinkingExtras(),
+      ...this.requestExtras(),
     };
   }
 
@@ -120,18 +115,25 @@ export class AnthropicAdapter extends BaseLLMAdapter {
    * P53: POST, and if the gateway rejects the `thinking` field with a 400, retry ONCE
    * without it. An Anthropic-compatible proxy that predates extended thinking should
    * degrade to "no reasoning control", never to a dead request.
+   * P131: same for `output_config` (effort) and for sampling — a model id we don't
+   * recognise yet that has dropped `temperature` should degrade, not die.
    */
   private async post(body: any, signal: AbortSignal): Promise<Response> {
     const url = `${this.getBaseURL()}/v1/messages`;
     const send = (b: any) => llmFetch(url, {
       method: 'POST', headers: this.buildHeaders(), body: JSON.stringify(b), signal,
-    });
+    }, undefined, { connectMs: b.stream ? undefined : 0 });
     const res = await send(body);
-    if (res.status !== 400 || !('thinking' in body)) return res;
+    const optional = ['thinking', 'output_config', 'temperature'].filter((k) => k in body);
+    if (res.status !== 400 || optional.length === 0) return res;
     const text = await res.clone().text();
-    if (!/thinking/i.test(text)) return res;
     const stripped = { ...body };
-    delete stripped.thinking;
+    if (/thinking|budget_tokens|effort|output_config/i.test(text)) {
+      delete stripped.thinking;
+      delete stripped.output_config;
+    }
+    if (/temperature|top_p|top_k|sampling/i.test(text)) delete stripped.temperature;
+    if (optional.every((k) => k in stripped)) return res;  // the 400 is about something else
     return send(stripped);
   }
 
@@ -162,7 +164,8 @@ export class AnthropicAdapter extends BaseLLMAdapter {
         .join('');
 
       // P53: drop any inline <think> block (OSS models behind Anthropic-compatible proxies)
-      return this.finalize(splitThinking(content).answer, data.usage, data.stop_reason);
+      const answer = withRefusal(splitThinking(content).answer, data.stop_reason, data.stop_details);
+      return this.finalize(answer, data.usage, data.stop_reason);
     } catch (err) {
       throw toLLMError(err, 'Anthropic 请求失败');
     } finally {
@@ -180,6 +183,7 @@ export class AnthropicAdapter extends BaseLLMAdapter {
     let acc = '';
     const usage: { input_tokens?: number; output_tokens?: number } = {};
     let stopReason: string | null = null;
+    let stopDetails: any = null;
 
     try {
       yield { type: 'start', requestId };
@@ -217,6 +221,7 @@ export class AnthropicAdapter extends BaseLLMAdapter {
           }
           case 'message_delta': {
             if (payload.delta?.stop_reason) stopReason = payload.delta.stop_reason;
+            if (payload.delta?.stop_details) stopDetails = payload.delta.stop_details;
             if (payload.usage?.output_tokens != null)
               usage.output_tokens = payload.usage.output_tokens;
             break;
@@ -237,6 +242,11 @@ export class AnthropicAdapter extends BaseLLMAdapter {
         }
       }
 
+      const notice = withRefusal(acc, stopReason, stopDetails).slice(acc.length);
+      if (notice) {
+        acc += notice;
+        yield { type: 'delta', requestId, chunk: notice };
+      }
       yield {
         type: 'done',
         requestId,
@@ -325,15 +335,12 @@ export class AnthropicAdapter extends BaseLLMAdapter {
     const body: any = {
       model: this.config.model,
       max_tokens: this.maxTokens(),
-      temperature: this.config.temperature ?? 0.3,
       system: systemPrompt,
       stream,
       messages: wire,
-      ...this.thinkingExtras(),
+      ...this.requestExtras(),
     };
     if (tools && tools.length > 0) body.tools = this.toAnthropicTools(tools);
-    // Extended thinking requires the default temperature
-    if (body.thinking?.type === 'enabled') delete body.temperature;
     return body;
   }
 
@@ -365,7 +372,7 @@ export class AnthropicAdapter extends BaseLLMAdapter {
       const split = splitThinking(content);
 
       return {
-        content: split.answer,
+        content: withRefusal(split.answer, data.stop_reason, data.stop_details),
         reasoning: [reasoning, split.reasoning].filter(Boolean).join('\n') || undefined,
         toolCalls: toolCalls.length ? toolCalls : undefined,
         finishReason: toolCalls.length ? 'tool_use' : 'stop',
@@ -401,6 +408,7 @@ export class AnthropicAdapter extends BaseLLMAdapter {
     let content = '';
     let reasoning = '';
     let stopReason: string | null = null;
+    let stopDetails: any = null;
     const usage: { input_tokens?: number; output_tokens?: number } = {};
 
     try {
@@ -473,6 +481,7 @@ export class AnthropicAdapter extends BaseLLMAdapter {
 
           case 'message_delta':
             if (payload.delta?.stop_reason) stopReason = payload.delta.stop_reason;
+            if (payload.delta?.stop_details) stopDetails = payload.delta.stop_details;
             if (payload.usage?.output_tokens != null) usage.output_tokens = payload.usage.output_tokens;
             break;
 
@@ -485,10 +494,14 @@ export class AnthropicAdapter extends BaseLLMAdapter {
       }
 
       const split = splitThinking(content);
+      const answer = withRefusal(split.answer, stopReason, stopDetails);
+      if (answer.length > split.answer.length) {
+        yield { kind: 'text', chunk: answer.slice(split.answer.length) };
+      }
       yield {
         kind: 'done',
         response: {
-          content: split.answer,
+          content: answer,
           reasoning: [reasoning, split.reasoning].filter(Boolean).join('\n') || undefined,
           toolCalls: toolCalls.length ? toolCalls : undefined,
           finishReason: toolCalls.length ? 'tool_use'
@@ -573,3 +586,15 @@ export class AnthropicAdapter extends BaseLLMAdapter {
     };
   }
 }
+
+/**
+ * P131: Fable 5.1 / Opus 5.5 run safety classifiers that can decline a turn — HTTP 200,
+ * `stop_reason: 'refusal'`, usually with no text. Without a notice the agent loop
+ * finished the turn with an empty bubble and no explanation.
+ */
+function withRefusal(text: string, stopReason: string | null, details?: { category?: string | null } | null): string {
+  if (stopReason !== 'refusal') return text;
+  const cat = details?.category ? `（${details.category}）` : '';
+  const notice = `⚠️ 模型的安全策略拒绝了这次请求${cat}。可以换个说法重试，或在设置里换用其他模型。`;
+  return text ? `${text}\n\n${notice}` : notice;
+}
diff --git a/src/main/llm/context-window.ts b/src/main/llm/context-window.ts
index a6e7818..af8e24e 100644
--- a/src/main/llm/context-window.ts
+++ b/src/main/llm/context-window.ts
@@ -126,10 +126,10 @@ export function truncateMessages(
         ? MODEL_TOKEN_BUDGETS[budgetKey]
         : DEFAULT_BUDGET) - OUTPUT_RESERVE;
 
-  let availableTokens = totalBudget - estimateTokens(systemPrompt);
-  if (availableTokens <= 0) {
-    return messages.slice(-1);
-  }
+  // P131: an overhead larger than the budget used to return `messages.slice(-1)`, which can
+  // be a lone tool result (a 400). Fall through instead: the hard floor below keeps the last
+  // whole block.
+  let availableTokens = Math.max(0, totalBudget - estimateTokens(systemPrompt));
 
   const blocks = toBlocks(messages);
   const kept: ChatMessage[][] = [];
diff --git a/src/main/llm/net.ts b/src/main/llm/net.ts
index 46e1824..c48c786 100644
--- a/src/main/llm/net.ts
+++ b/src/main/llm/net.ts
@@ -78,7 +78,10 @@ async function fetchWithConnectTimeout(
   connectMs = 20_000,
 ): Promise<Response> {
   const controller = new AbortController();
-  const timer = setTimeout(() => controller.abort(new Error('connect timeout')), connectMs);
+  // P131: connectMs <= 0 → no connect-stage timer (see LlmFetchOptions.connectMs)
+  const timer = connectMs > 0
+    ? setTimeout(() => controller.abort(new Error('connect timeout')), connectMs)
+    : undefined;
   const onAbort = () => controller.abort(signal?.reason);
   if (signal) {
     if (signal.aborted) controller.abort(signal.reason);
@@ -95,6 +98,18 @@ async function fetchWithConnectTimeout(
   }
 }
 
+export interface LlmFetchOptions {
+  /**
+   * P131: connect-stage timeout. The P109 premise — "headers arrive fast, the body is the
+   * outer timeout's job" — holds only for STREAMING. A non-streaming server sends headers
+   * when the whole answer is done, so 20 s became a total timeout for vision captions and
+   * non-streamed turns, and the abort counted as transient: the finished-but-slow request
+   * was re-sent (and billed) up to 3× per stack. Pass 0 for non-streaming requests; the
+   * caller's own total timeout governs them.
+   */
+  connectMs?: number;
+}
+
 /**
  * Fetch that respects the OS system proxy (via Electron's Chromium stack) and
  * retries transient network errors. Same signature as global fetch.
@@ -103,6 +118,7 @@ export async function llmFetch(
   url: string,
   init: RequestInit = {},
   retries = 2,
+  opts: LlmFetchOptions = {},
 ): Promise<Response> {
   let lastErr: unknown;
 
@@ -116,7 +132,7 @@ export async function llmFetch(
 
     // 1) Undici path — no proxy, direct DNS.
     try {
-      return await fetchWithConnectTimeout(fetch, url, init, init.signal ?? undefined);
+      return await fetchWithConnectTimeout(fetch, url, init, init.signal ?? undefined, opts.connectMs);
     } catch (err) {
       lastErr = err;
       if (!isTransient(err)) {
@@ -130,7 +146,7 @@ export async function llmFetch(
     if (typeof net !== 'undefined' && typeof net.fetch === 'function') {
       try {
         return await fetchWithConnectTimeout(
-          (u, i) => net.fetch(u, i as any), url, init, init.signal ?? undefined,
+          (u, i) => net.fetch(u, i as any), url, init, init.signal ?? undefined, opts.connectMs,
         );
       } catch (err) {
         lastErr = err;
diff --git a/src/main/llm/openai.ts b/src/main/llm/openai.ts
index 09d7be4..4a0e43e 100644
--- a/src/main/llm/openai.ts
+++ b/src/main/llm/openai.ts
@@ -20,7 +20,10 @@ import { extractFirstCodeBlock } from './code-extract';
 import { LLMHttpError, extractErrorMessage, toLLMError } from './errors';
 import { parseSSE } from './sse';
 import { buildOpenAITools } from './tools-schema';
-import { ThinkSplitter, reasoningParams, isReasoningParamError, splitThinking, dropsTemperature, detectDialect } from './thinking';
+import {
+  ThinkSplitter, reasoningParams, isReasoningParamError, splitThinking, dropsTemperature, detectDialect,
+  isOpenAIReasoningModel, isGpt6Family,
+} from './thinking';
 import type {
   ChatMessage,
   LLMResponse,
@@ -54,12 +57,24 @@ interface PartialCall {
 
 export class OpenAIAdapter extends BaseLLMAdapter {
   /** P51: extra body fields for reasoning depth — empty when set to 'auto'. */
-  private reasoningExtras(): Record<string, any> {
-    return reasoningParams(
+  private reasoningExtras(withTools = false): Record<string, any> {
+    const extras = reasoningParams(
       this.config.reasoningLevel,
       this.config.reasoningDialect,
       this.config.baseURL,
     );
+    // P131: GPT-6 on /chat/completions 400s on reasoning_effort + tools, and on the
+    // 'minimal' effort we map 'off' to. Agent turns always carry tools, so they run at
+    // the model's default effort; plain chat gets the closest accepted level.
+    if ('reasoning_effort' in extras && isGpt6Family(this.config.model)) {
+      if (withTools) {
+        const rest = { ...extras };
+        delete rest.reasoning_effort;
+        return rest;
+      }
+      if (extras.reasoning_effort === 'minimal') return { ...extras, reasoning_effort: 'low' };
+    }
+    return extras;
   }
 
   // P54: 8192 truncated answers on reasoning models — the scratchpad is spent from the
@@ -83,11 +98,11 @@ export class OpenAIAdapter extends BaseLLMAdapter {
    * 400 and require `max_completion_tokens`; every other OpenAI-compatible gateway
    * still expects `max_tokens`. Detect by host + model id rather than sending both,
    * because strict gateways reject unknown fields.
+   * P131: the pattern only knew gpt-5 — GPT-6 Astra got max_tokens + temperature and
+   * every request 400'd.
    */
   private isOpenAIReasoner(): boolean {
-    const host = (this.config.baseURL ?? '').toLowerCase();
-    if (!host.includes('openai.com') && !host.includes('azure.com')) return false;
-    return /^(o\d|gpt-5)/i.test(this.config.model ?? '');
+    return isOpenAIReasoningModel(this.config.baseURL, this.config.model);
   }
 
   private tokenCap(): Record<string, any> {
@@ -235,10 +250,12 @@ export class OpenAIAdapter extends BaseLLMAdapter {
       const res = await llmFetch(`${this.getBaseURL()}/chat/completions`, {
         method: 'POST',
         headers: this.buildHeaders(),
+        // P131: reasoning models reject `max_tokens` — "Test connection" failed for
+        // every GPT-5.x / GPT-6 model on api.openai.com even with a valid key.
         body: JSON.stringify({
           model: this.config.model,
           messages: [{ role: 'user', content: 'hi' }],
-          max_tokens: 1,
+          ...(this.isOpenAIReasoner() ? { max_completion_tokens: 32 } : { max_tokens: 1 }),
         }),
         signal: s,
       });
@@ -267,7 +284,7 @@ export class OpenAIAdapter extends BaseLLMAdapter {
       tool_choice: 'auto',
       stream,
       ...(stream ? { stream_options: { include_usage: true } } : {}),
-      ...this.reasoningExtras(),
+      ...this.reasoningExtras(true),
     };
   }
 
@@ -358,7 +375,7 @@ export class OpenAIAdapter extends BaseLLMAdapter {
     const url = `${this.getBaseURL()}/chat/completions`;
     const send = (b: any) => llmFetch(url, {
       method: 'POST', headers: this.buildHeaders(), body: JSON.stringify(b), signal,
-    });
+    }, undefined, { connectMs: b.stream ? undefined : 0 });
 
     const extras = this.reasoningExtras();
     const res = await send(body);
diff --git a/src/main/llm/thinking.ts b/src/main/llm/thinking.ts
index e964213..9369950 100644
--- a/src/main/llm/thinking.ts
+++ b/src/main/llm/thinking.ts
@@ -183,6 +183,132 @@ export function reasoningParams(
   }
 }
 
+// ===== P131: model capability detection =====
+
+/**
+ * OpenAI's own reasoning models (o-series, GPT-5.x, GPT-6.x) reject `max_tokens`
+ * (they need `max_completion_tokens`) and reject `temperature` / `top_p`. Only on
+ * OpenAI's / Azure's own hosts — every other OpenAI-compatible gateway still expects
+ * `max_tokens`, and strict gateways reject unknown fields.
+ */
+export function isOpenAIReasoningModel(baseURL?: string, model?: string): boolean {
+  const host = (baseURL ?? '').toLowerCase();
+  if (!host.includes('openai.com') && !host.includes('azure.com')) return false;
+  return /^(o\d|gpt-([5-9]|\d{2,}))/i.test(model ?? '');
+}
+
+/**
+ * P131: GPT-6 (Astra and later) on /chat/completions rejects `reasoning_effort` whenever
+ * `tools` are present ("Function tools with reasoning_effort are not supported …
+ * use /v1/responses"), and rejects the 'none' / 'minimal' efforts outright.
+ */
+export function isGpt6Family(model?: string): boolean {
+  return /^gpt-([6-9]|\d{2,})/i.test(model ?? '');
+}
+
+/** What a Claude model's request surface accepts. */
+export interface ClaudeCaps {
+  /** temperature / top_p / top_k are rejected with a 400 — never send them. */
+  noSampling: boolean;
+  /** Reasoning depth is `thinking: {type:'adaptive'}` + `output_config.effort`;
+   *  `{type:'enabled', budget_tokens}` is removed or deprecated. */
+  adaptive: boolean;
+  /** Thinking runs even when `thinking` is omitted. */
+  thinksByDefault: boolean;
+  /** Thinking cannot be switched off — `{type:'disabled'}` is a 400. */
+  alwaysThinks: boolean;
+}
+
+/**
+ * P131: request-surface capabilities by Claude model id. Matches first-party ids
+ * (`claude-opus-5-5`), dotted proxy ids (`claude-opus-4.8`) and prefixed ids
+ * (`anthropic.claude-fable-5-1`). Unrecognised ids — old dated models, OSS models
+ * behind Anthropic-compatible proxies — get the legacy surface.
+ *
+ *   Fable 5 / 5.1, Mythos   always thinks; sampling + budget_tokens → 400
+ *   Opus 5.5                always thinks; sampling + budget_tokens → 400
+ *   Sonnet 5.5              disabled → 400; non-default sampling → 400
+ *   Opus 5, Sonnet 5        thinks by default; sampling + budget_tokens → 400
+ *   Opus 4.7 / 4.8          adaptive only on-mode (off by default); sampling → 400
+ *   Opus 4.6 / Sonnet 4.6   adaptive recommended; sampling allowed
+ *   older (Haiku 4.5, …)    budget_tokens thinking; sampling allowed
+ */
+export function claudeCaps(model?: string): ClaudeCaps {
+  const legacy: ClaudeCaps = { noSampling: false, adaptive: false, thinksByDefault: false, alwaysThinks: false };
+  const id = (model ?? '').toLowerCase();
+  if (/fable|mythos/.test(id)) {
+    return { noSampling: true, adaptive: true, thinksByDefault: true, alwaysThinks: true };
+  }
+  const m = /(opus|sonnet|haiku)[-_ ]?(\d{1,2})(?!\d)(?:[-.](\d)(?!\d))?/.exec(id);
+  if (!m) return legacy;
+  const family = m[1];
+  const v = Number(m[2]) + Number(m[3] ?? 0) / 10;
+  if (family === 'opus') {
+    if (v >= 5.5) return { noSampling: true, adaptive: true, thinksByDefault: true, alwaysThinks: true };
+    if (v >= 5) return { noSampling: true, adaptive: true, thinksByDefault: true, alwaysThinks: false };
+    if (v >= 4.7) return { noSampling: true, adaptive: true, thinksByDefault: false, alwaysThinks: false };
+    if (v >= 4.6) return { noSampling: false, adaptive: true, thinksByDefault: false, alwaysThinks: false };
+    return legacy;
+  }
+  if (family === 'sonnet') {
+    if (v >= 5.5) return { noSampling: true, adaptive: true, thinksByDefault: true, alwaysThinks: true };
+    if (v >= 5) return { noSampling: true, adaptive: true, thinksByDefault: true, alwaysThinks: false };
+    if (v >= 4.6) return { noSampling: false, adaptive: true, thinksByDefault: false, alwaysThinks: false };
+    return legacy;
+  }
+  return legacy;
+}
+
+/**
+ * P131: Anthropic request fields for reasoning depth + sampling, by model.
+ *
+ * Before P131 every request carried `temperature: 0.3` — a hard 400 on Fable 5 / 5.1,
+ * Opus 4.7+ and Sonnet 5+ (every Claude preset except Sonnet 4.6) — and reasoning
+ * levels were sent as `budget_tokens`, which those models also reject.
+ *
+ * Adaptive-surface models: level → `output_config.effort`; thinking summaries are
+ * requested (`display: 'summarized'`) whenever thinking runs, because the default
+ * ('omitted') streams empty thinking blocks and the UI would show nothing while the
+ * model works. 'off' on a model that cannot stop thinking becomes the lowest effort.
+ */
+export function anthropicReasoningParams(
+  level: ReasoningLevel | undefined,
+  model: string | undefined,
+  maxTokens: number,
+  temperature: number | undefined,
+): Record<string, any> {
+  const lv = level ?? 'auto';
+  const caps = claudeCaps(model);
+  const adaptiveOn = { thinking: { type: 'adaptive', display: 'summarized' } };
+
+  if (caps.adaptive) {
+    let out: Record<string, any>;
+    if (lv === 'off') {
+      out = caps.alwaysThinks ? { ...adaptiveOn, output_config: { effort: 'low' } } : { thinking: { type: 'disabled' } };
+    } else if (lv === 'low' || lv === 'medium' || lv === 'high') {
+      out = { ...adaptiveOn, output_config: { effort: lv } };
+    } else if (lv === 'adaptive' || caps.thinksByDefault) {
+      // 'auto' on a model that thinks anyway: same behaviour, readable summaries.
+      out = { ...adaptiveOn };
+    } else {
+      out = {};  // 'auto' on Opus 4.6–4.8 / Sonnet 4.6: provider default (no thinking)
+    }
+    // Sampling is rejected outright on noSampling models, and incompatible with thinking on.
+    if (!caps.noSampling && (!out.thinking || out.thinking.type === 'disabled')) {
+      out.temperature = temperature ?? 0.3;
+    }
+    return out;
+  }
+
+  // Legacy surface: budget_tokens thinking; temperature only while thinking is off.
+  // P55: an unrecognised Mythos-class id is handled above (always thinks).
+  if (lv === 'auto' || lv === 'adaptive') return { temperature: temperature ?? 0.3 };
+  if (lv === 'off') return { temperature: temperature ?? 0.3, thinking: { type: 'disabled' } };
+  const budgets: Record<string, number> = { low: 1024, medium: 4096, high: 16384 };
+  const budget = Math.min(budgets[lv] ?? 4096, Math.max(1024, maxTokens - 1024));
+  return { thinking: { type: 'enabled', budget_tokens: budget } };
+}
+
 /** P54: providers that ignore sampling params while thinking — we drop them to avoid noise. */
 export function dropsTemperature(
   level: ReasoningLevel | undefined,
@@ -205,7 +331,10 @@ export const REASONING_FIELDS = [
 export function isReasoningParamError(body: string): boolean {
   const b = (body || '').toLowerCase();
   if (!b) return false;
+  // P131: 'not supported' — GPT-6 Astra phrases its tools+reasoning_effort 400 as
+  // "Function tools with reasoning_effort are not supported …", which fell through.
   const complains = b.includes('unknown') || b.includes('unsupported') || b.includes('invalid')
-    || b.includes('not allowed') || b.includes('unrecognized') || b.includes('extra');
+    || b.includes('not allowed') || b.includes('unrecognized') || b.includes('extra')
+    || b.includes('not supported');
   return complains && REASONING_FIELDS.some((f) => b.includes(f));
 }
diff --git a/src/main/llm/vision.ts b/src/main/llm/vision.ts
index da8046c..a40d38d 100644
--- a/src/main/llm/vision.ts
+++ b/src/main/llm/vision.ts
@@ -4,6 +4,7 @@
 import type { VisionConfig, LocaleName } from '../../shared/types';
 import { toLLMError } from './errors';
 import { llmFetch } from './net';
+import { isOpenAIReasoningModel } from './thinking';
 
 export interface AnalyzeImageInput {
   question: string;        // Question drafted by the main model = image-to-text prompt
@@ -41,11 +42,24 @@ export async function analyzeImage(input: AnalyzeImageInput): Promise<string> {
         ],
       },
     ],
-    temperature: 0.2,
-    max_tokens: 1024,
+    // P131: OpenAI reasoning models (GPT-5.x / GPT-6) as the vision model 400'd on
+    // max_tokens + temperature. Their cap also covers reasoning tokens, so 1024 would
+    // leave an empty caption — give them room.
+    ...(isOpenAIReasoningModel(config.baseURL, config.model)
+      ? { max_completion_tokens: 8192 }
+      : { temperature: 0.2, max_tokens: 1024 }),
     stream: false,
   };
+  // P131: a caption is one non-streamed answer — its headers arrive only when it is done, so
+  // the connect-stage timer must not apply (it killed and re-sent slow captions at 20 s).
+  // VisionConfig.timeoutMs (default 120 s) is the total budget instead.
+  const controller = new AbortController();
+  const onAbort = () => controller.abort(signal?.reason);
+  if (signal?.aborted) controller.abort(signal.reason);
+  else signal?.addEventListener('abort', onAbort, { once: true });
+  const timer = setTimeout(() => controller.abort(new Error('视觉模型响应超时')), config.timeoutMs ?? 120_000);
   let res: Response;
+  let text: string;
   try {
     res = await llmFetch(`${base}/chat/completions`, {
       method: 'POST',
@@ -54,12 +68,15 @@ export async function analyzeImage(input: AnalyzeImageInput): Promise<string> {
         Authorization: `Bearer ${config.apiKey}`,
       },
       body: JSON.stringify(body),
-      signal,
-    });
+      signal: controller.signal,
+    }, undefined, { connectMs: 0 });
+    text = await res.text();
   } catch (err) {
     throw toLLMError(err, '视觉模型网络请求失败');
+  } finally {
+    clearTimeout(timer);
+    signal?.removeEventListener('abort', onAbort);
   }
-  const text = await res.text();
   if (!res.ok) {
     throw toLLMError(new Error(text), `视觉模型请求失败 (HTTP ${res.status})`);
   }
diff --git a/src/main/store/config.ts b/src/main/store/config.ts
index 11ba15e..a795c9d 100644
--- a/src/main/store/config.ts
+++ b/src/main/store/config.ts
@@ -25,6 +25,10 @@ export interface StoredConfig {
 
 const SCHEMA_VERSION = 1;
 
+// P131: the API key currently supplied by the environment (.env / process.env). It is used,
+// never persisted — saveConfig drops it so a later save cannot copy it into the store.
+let envSourcedKey = '';
+
 // P107: live approval-mode cache — updated on every saveConfig so an in-flight agent
 // session can re-read it per tool call (issue #1: settings changes should apply to
 // the running conversation immediately).
@@ -99,7 +103,18 @@ export async function loadConfig(): Promise<LLMConfig> {
       console.info(
         `[Millwright] 使用 .env fallback 配置: protocol=${envFallback.protocol}, model=${envFallback.model}`,
       );
-      return envFallback;
+      // P131: take only the CONNECTION from the environment. Returning the env config whole
+      // threw away every saved preference (approval mode, disabled tools, context window…),
+      // and the next save then encrypted the env key into the store.
+      envSourcedKey = envFallback.apiKey;
+      return {
+        ...DEFAULT_CONFIG,
+        ...llm,
+        protocol: envFallback.protocol,
+        baseURL: envFallback.baseURL,
+        model: envFallback.model,
+        apiKey: envFallback.apiKey,
+      };
     }
   }
 
@@ -119,7 +134,7 @@ export async function saveConfig(config: LLMConfig): Promise<void> {
   const { apiKey, ...rest } = config;
 
   let encryptedApiKey = '';
-  if (apiKey) {
+  if (apiKey && apiKey !== envSourcedKey) {
     if (safeStorage.isEncryptionAvailable()) {
       encryptedApiKey = safeStorage.encryptString(apiKey).toString('base64');
     } else {
diff --git a/src/preload/index.ts b/src/preload/index.ts
index 96eade7..324a01a 100644
--- a/src/preload/index.ts
+++ b/src/preload/index.ts
@@ -28,12 +28,12 @@ const api = {
     save: (config: LLMConfig) => ipcRenderer.invoke(IpcChannels.CONFIG_SAVE, config),
   },
   theme: {
-    load: (): Promise<ThemeName> => ipcRenderer.invoke('theme:load'),
-    save: (theme: ThemeName) => ipcRenderer.invoke('theme:save', theme),
+    load: (): Promise<ThemeName> => ipcRenderer.invoke(IpcChannels.THEME_LOAD),
+    save: (theme: ThemeName) => ipcRenderer.invoke(IpcChannels.THEME_SAVE, theme),
   },
   locale: {
-    load: (): Promise<LocaleName> => ipcRenderer.invoke('locale:load'),
-    save: (locale: LocaleName) => ipcRenderer.invoke('locale:save', locale),
+    load: (): Promise<LocaleName> => ipcRenderer.invoke(IpcChannels.LOCALE_LOAD),
+    save: (locale: LocaleName) => ipcRenderer.invoke(IpcChannels.LOCALE_SAVE, locale),
   },
   sw: {
     connect: (): Promise<{ ok: boolean; status: SWStatus }> =>
diff --git a/src/renderer/App.tsx b/src/renderer/App.tsx
index b2bf66d..931f453 100644
--- a/src/renderer/App.tsx
+++ b/src/renderer/App.tsx
@@ -35,8 +35,17 @@ export default function App() {
 
   // —— Config ——
   const [config, setConfig] = useState<LLMConfig>(DEFAULT_CONFIG);
+  const configLoaded = useRef(false);
   useEffect(() => {
-    window.api.config.load().then(setConfig);
+    window.api.config.load()
+      .then((c) => { configLoaded.current = true; setConfig(c); })
+      .catch((err) => console.error('[config] load failed', err));
+  }, []);
+  // P131: never persist before the stored config has loaded — an approval or tools toggle
+  // in that window saved DEFAULT_CONFIG over the real one and wiped the API key.
+  const persistConfig = useCallback((next: LLMConfig) => {
+    setConfig(next);
+    if (configLoaded.current) void window.api.config.save(next);
   }, []);
 
   // —— SolidWorks status ——
@@ -322,11 +331,7 @@ export default function App() {
               onCancel={cancel}
               isGenerating={isGenerating}
               approvalMode={config.approvalMode ?? 'normal'}
-              onApprovalChange={(m) => {
-                const next = { ...config, approvalMode: m };
-                setConfig(next);
-                void window.api.config.save(next);
-              }}
+              onApprovalChange={(m) => persistConfig({ ...config, approvalMode: m })}
               placeholder={
                 !config.apiKey
                   ? tr('input.placeholderNoKey')
@@ -343,11 +348,7 @@ export default function App() {
           <ToolsPanel
             t={t}
             disabled={config.disabledTools ?? []}
-            onChange={(next) => {
-              const cfg = { ...config, disabledTools: next };
-              setConfig(cfg);
-              window.api.config.save(cfg);
-            }}
+            onChange={(next) => persistConfig({ ...config, disabledTools: next })}
           />
         )}
       </main>
diff --git a/src/renderer/components/SettingsModal.tsx b/src/renderer/components/SettingsModal.tsx
index 159b6be..21f24a0 100644
--- a/src/renderer/components/SettingsModal.tsx
+++ b/src/renderer/components/SettingsModal.tsx
@@ -75,11 +75,27 @@ export function SettingsModal({
   };
 
   const handleProtocol = (p: 'anthropic' | 'openai') => {
+    // P131: pair the default URL with a model that endpoint actually serves. The first
+    // OpenAI-protocol preset is DeepSeek, so switching gave api.openai.com + deepseek-v4-pro.
+    const official = OPENAI_COMPATIBLE_PROVIDERS.find((x) => x.url === DEFAULT_URLS[p]);
     setDraft((d) => ({
       ...d,
       protocol: p,
       baseURL: DEFAULT_URLS[p],
-      model: MODEL_PRESETS[p][0].value,
+      model: official?.suggestedModel ?? MODEL_PRESETS[p][0].value,
+    }));
+    setTestStatus({ kind: 'idle' });
+  };
+
+  // P131: a quick-fill button sets the whole provider, not just its URL — the suggested
+  // model and the per-provider context / output defaults were declared but never applied.
+  const applyProvider = (p: (typeof OPENAI_COMPATIBLE_PROVIDERS)[number]) => {
+    setDraft((d) => ({
+      ...d,
+      baseURL: p.url,
+      ...(p.suggestedModel ? { model: p.suggestedModel } : {}),
+      ...(p.contextWindow ? { contextWindow: p.contextWindow } : {}),
+      ...(p.maxTokens ? { maxTokens: p.maxTokens } : {}),
     }));
     setTestStatus({ kind: 'idle' });
   };
@@ -243,7 +259,7 @@ export function SettingsModal({
         >
           <span style={{ color: t.textSecondary, fontSize: 13 }}>{tr('settings.swConnection')}</span>
           <div style={{ display: 'flex', alignItems: 'center' }}>
-            <StatusDot connected={swStatus.connected} />
+            <StatusDot connected={swStatus.connected} busy={swStatus.busy} runningTool={swStatus.runningTool} />
             <span
               style={{
                 color: swStatus.connected ? '#4caf72' : '#d45454',
@@ -295,7 +311,7 @@ export function SettingsModal({
               {OPENAI_COMPATIBLE_PROVIDERS.map((p) => (
                 <button
                   key={p.name}
-                  onClick={() => update('baseURL', p.url)}
+                  onClick={() => applyProvider(p)}
                   style={{
                     padding: '3px 8px', borderRadius: 4,
                     border: `1px solid ${t.cardBorder}`,
@@ -416,12 +432,12 @@ export function SettingsModal({
 
         {/* P30: Max agent rounds */}
         <label style={labelStyle}>{tr('settings.maxRounds')}</label>
-        <input
-          type="number"
+        <ClampedNumber
           min={4}
           max={100}
+          fallback={24}
           value={draft.maxRounds ?? 24}
-          onChange={(e) => update('maxRounds', Math.max(4, Math.min(100, Number(e.target.value) || 24)))}
+          onCommit={(v) => update('maxRounds', v)}
           style={{ ...fieldStyle, marginBottom: 4, width: 120 }}
         />
         <p style={{ color: t.textMuted, fontSize: 11, margin: '2px 0 16px 1px' }}>{tr('settings.maxRoundsHint')}</p>
@@ -503,13 +519,13 @@ export function SettingsModal({
 
         {/* P54: Context window + max output tokens — a matched pair, so show them together */}
         <label style={labelStyle}>{tr('settings.contextWindow')}</label>
-        <input
-          type="number"
+        <ClampedNumber
           min={4096}
           max={2000000}
           step={4096}
+          fallback={128000}
           value={draft.contextWindow ?? 128000}
-          onChange={(e) => update('contextWindow', Math.max(4096, Math.min(2000000, Number(e.target.value) || 128000)))}
+          onCommit={(v) => update('contextWindow', v)}
           style={{ ...fieldStyle, marginBottom: 4, width: 160 }}
         />
         <p style={{ color: t.textMuted, fontSize: 11, margin: '2px 0 14px 1px' }}>
@@ -518,13 +534,13 @@ export function SettingsModal({
 
         {/* P51: Max output tokens */}
         <label style={labelStyle}>{tr('settings.maxTokens')}</label>
-        <input
-          type="number"
+        <ClampedNumber
           min={1024}
           max={512000}
           step={1024}
+          fallback={32768}
           value={draft.maxTokens ?? 32768}
-          onChange={(e) => update('maxTokens', Math.max(1024, Math.min(512000, Number(e.target.value) || 32768)))}
+          onCommit={(v) => update('maxTokens', v)}
           style={{ ...fieldStyle, marginBottom: 4, width: 140 }}
         />
         <p style={{ color: t.textMuted, fontSize: 11, margin: '2px 0 18px 1px' }}>
@@ -705,3 +721,40 @@ export function SettingsModal({
     </div>
   );
 }
+
+/**
+ * P131: a number field that clamps when editing ends (blur / Enter), not on every keystroke.
+ * Clamping per keystroke snapped a half-typed "2" to the minimum, so a value such as 200000
+ * could not be typed at all.
+ */
+function ClampedNumber({ value, min, max, step, fallback, onCommit, style }: {
+  value: number;
+  min: number;
+  max: number;
+  step?: number;
+  fallback: number;
+  onCommit: (v: number) => void;
+  style?: React.CSSProperties;
+}) {
+  const [text, setText] = useState(String(value));
+  useEffect(() => { setText(String(value)); }, [value]);
+  const commit = () => {
+    const n = Number(text);
+    const v = Math.max(min, Math.min(max, text.trim() !== '' && Number.isFinite(n) ? Math.round(n) : fallback));
+    setText(String(v));
+    if (v !== value) onCommit(v);
+  };
+  return (
+    <input
+      type="number"
+      min={min}
+      max={max}
+      step={step}
+      value={text}
+      onChange={(e) => setText(e.target.value)}
+      onBlur={commit}
+      onKeyDown={(e) => { if (e.key === 'Enter') commit(); }}
+      style={style}
+    />
+  );
+}
diff --git a/src/renderer/components/Sidebar.tsx b/src/renderer/components/Sidebar.tsx
index e51d56d..8cee572 100644
--- a/src/renderer/components/Sidebar.tsx
+++ b/src/renderer/components/Sidebar.tsx
@@ -98,7 +98,7 @@ export function Sidebar({
           }}
         >
           <div style={{ display: 'flex', alignItems: 'center', minWidth: 0 }}>
-            <StatusDot connected={swStatus.connected} />
+            <StatusDot connected={swStatus.connected} busy={swStatus.busy} runningTool={swStatus.runningTool} />
             <span
               style={{
                 fontSize: 11,
diff --git a/src/renderer/components/ToolCallGroup.tsx b/src/renderer/components/ToolCallGroup.tsx
index 80370be..3a0be60 100644
--- a/src/renderer/components/ToolCallGroup.tsx
+++ b/src/renderer/components/ToolCallGroup.tsx
@@ -159,7 +159,9 @@ export function ToolCallGroup({ steps, t }: { steps: AgentStep[]; t: ThemeTokens
                     return null;
                   })()}
                   {target && <span style={{ fontSize: 11.5, color: t.textMuted, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{target}</span>}
-                  <span style={{ fontFamily: "'Consolas', monospace", fontSize: 10, color: t.textMuted, marginLeft: 'auto', flexShrink: 0 }}>{name}</span>
+                  <span style={{ fontFamily: "'Consolas', monospace", fontSize: 10, color: t.textMuted, marginLeft: 'auto', flexShrink: 0 }}>
+                    {name}{typeof s.durationMs === 'number' ? ` · ${(s.durationMs / 1000).toFixed(1)}s` : ''}
+                  </span>
                   {hasDetail && <span style={{ fontSize: 9, color: t.textMuted, flexShrink: 0 }}>{rowOpen ? '▲' : '▼'}</span>}
                 </button>
                 {rowOpen && hasDetail && (
diff --git a/src/renderer/hooks/useLLM.ts b/src/renderer/hooks/useLLM.ts
index 60f81c6..f2b6a20 100644
--- a/src/renderer/hooks/useLLM.ts
+++ b/src/renderer/hooks/useLLM.ts
@@ -135,11 +135,11 @@ export function useLLM({ config, initial }: UseLLMOptions) {
       for (let i = steps.length - 1; i >= 0; i--) {
         const s = steps[i];
         if (s.kind === 'tool' && (s.id === key || s.name === tc?.name) && s.status === 'running') {
-          steps[i] = { ...s, status, result: tc?.result, params: s.params ?? tc?.parameters };
+          steps[i] = { ...s, status, result: tc?.result, params: s.params ?? tc?.parameters, durationMs: tc?.durationMs };
           return { ...m, steps };
         }
       }
-      steps.push({ kind: 'tool', id: key, name: tc?.name, params: tc?.parameters, status, result: tc?.result });
+      steps.push({ kind: 'tool', id: key, name: tc?.name, params: tc?.parameters, status, result: tc?.result, durationMs: tc?.durationMs });
       return { ...m, steps };
     });
   }, [updateAssistant]);
diff --git a/src/renderer/session-export.ts b/src/renderer/session-export.ts
index 7d2d8a8..f91c439 100644
--- a/src/renderer/session-export.ts
+++ b/src/renderer/session-export.ts
@@ -50,7 +50,7 @@ function statusMark(status?: string): string {
 function toolToMarkdown(s: AgentStep): string {
   // P130: surface wall-clock duration next to the tool name so a long gear/batch is
   // visible at a glance ("- ✓ `extrude` · 1.2s").
-  const dur = (s as any).durationMs;
+  const dur = s.durationMs;
   const durStr = typeof dur === 'number' ? ` · ${(dur / 1000).toFixed(1)}s` : '';
   const lines = [`- ${statusMark(s.status)} \`${s.name ?? 'tool'}\`${durStr}`];
   if (s.params && Object.keys(s.params).length) {
@@ -135,7 +135,7 @@ export function sessionToJSON(messages: ChatMessage[]): string {
                 ...(s.status ? { status: s.status } : {}),
                 ...(s.result ? { result: s.result } : {}),
                 // P130: persist the wall-clock duration alongside the tool step.
-                ...((s as any).durationMs !== undefined ? { durationMs: (s as any).durationMs } : {}),
+                ...(s.durationMs !== undefined ? { durationMs: s.durationMs } : {}),
               })),
             }
           : {}),
diff --git a/src/shared/ipc-channels.ts b/src/shared/ipc-channels.ts
index cc3114e..baf0853 100644
--- a/src/shared/ipc-channels.ts
+++ b/src/shared/ipc-channels.ts
@@ -50,6 +50,12 @@ export const IpcChannels = {
   UPDATE_SKIP: 'update:skip',
   UPDATE_OPEN_RELEASE: 'update:open-release',
   UPDATE_EVENT: 'update:event',
+
+  // Theme / UI language (kept independent of the LLM config)
+  THEME_LOAD: 'theme:load',
+  THEME_SAVE: 'theme:save',
+  LOCALE_LOAD: 'locale:load',
+  LOCALE_SAVE: 'locale:save',
 } as const;
 
 export type IpcChannel = (typeof IpcChannels)[keyof typeof IpcChannels];
diff --git a/src/shared/presets.ts b/src/shared/presets.ts
index 6f260ad..ad0c2a7 100644
--- a/src/shared/presets.ts
+++ b/src/shared/presets.ts
@@ -2,6 +2,9 @@
 // Model presets / default URLs / default parameters.
 // P54/P55: preset IDs and contextWindow/maxTokens aligned with current provider docs (2026-07).
 // DeepSeek's old `deepseek-chat` / `deepseek-reasoner` aliases were retired 2026-07-24.
+// P131: GPT-6 Astra (2026-09-03), Claude Opus 5.5 and Claude Fable 5.1 become the
+// recommended OpenAI / Anthropic models. Saved ids that left the list still load — the
+// Settings dropdown shows any non-preset id as "Custom model".
 
 import type { LLMProtocol, ModelPreset, LLMConfig } from './types';
 
@@ -12,8 +15,8 @@ export const DEFAULT_URLS: Record<LLMProtocol, string> = {
 
 export const MODEL_PRESETS: Record<LLMProtocol, ModelPreset[]> = {
   anthropic: [
-    { label: 'Claude Fable 5 (1M ctx, always-on thinking)', value: 'claude-fable-5' },
-    { label: 'Claude Opus 4.8', value: 'claude-opus-4-8' },
+    { label: 'Claude Opus 5.5 (1M ctx, recommended)', value: 'claude-opus-5-5' },
+    { label: 'Claude Fable 5.1 (1M ctx, most capable)', value: 'claude-fable-5-1' },
     { label: 'Claude Sonnet 4.6', value: 'claude-sonnet-4-6' },
     { label: 'Custom model', value: 'custom' },
   ],
@@ -26,6 +29,7 @@ export const MODEL_PRESETS: Record<LLMProtocol, ModelPreset[]> = {
     { label: 'Qwen 3.7 Max (阿里百炼)', value: 'qwen3.7-max' },
     { label: 'GLM-4.6 (智谱)', value: 'glm-4.6' },
     // —— OpenAI official ——
+    { label: 'GPT-6 Astra (1M ctx, recommended)', value: 'gpt-6-astra' },
     { label: 'GPT-5.6 Sol', value: 'gpt-5.6-sol' },
     { label: 'GPT-4.1', value: 'gpt-4.1' },
     { label: 'GPT-4o Mini', value: 'gpt-4o-mini' },
@@ -50,8 +54,8 @@ export const OPENAI_COMPATIBLE_PROVIDERS: Array<{
   /** P54: per-provider recommended max output (tokens). Used as the default in Settings. */
   maxTokens?: number;
 }> = [
-  // OpenAI official — GPT-5.x / o series REQUIRE max_completion_tokens (handled in adapter)
-  { name: 'OpenAI', url: 'https://api.openai.com/v1', supportsTools: true, suggestedModel: 'gpt-5.6-sol', contextWindow: 1_000_000, maxTokens: 32_768 },
+  // OpenAI official — GPT-6 / GPT-5.x / o series REQUIRE max_completion_tokens (handled in adapter)
+  { name: 'OpenAI', url: 'https://api.openai.com/v1', supportsTools: true, suggestedModel: 'gpt-6-astra', contextWindow: 1_050_000, maxTokens: 32_768 },
   { name: 'DeepSeek', url: 'https://api.deepseek.com', supportsTools: true, suggestedModel: 'deepseek-v4-pro', contextWindow: 1_048_576, maxTokens: 32_768 },
   { name: 'Kimi / Moonshot', url: 'https://api.moonshot.cn/v1', supportsTools: true, suggestedModel: 'kimi-k2.5', contextWindow: 262_144, maxTokens: 32_768 },
   { name: 'MiniMax', url: 'https://api.minimax.io/v1', supportsTools: true, suggestedModel: 'minimax-m3', contextWindow: 512_000, maxTokens: 32_768 },
diff --git a/src/shared/types.ts b/src/shared/types.ts
index a6738a7..3d2188b 100644
--- a/src/shared/types.ts
+++ b/src/shared/types.ts
@@ -107,6 +107,9 @@ export interface AgentStep {
   status?: 'running' | 'ok' | 'error' | 'rejected';
   /** tool step: formatted result / error message */
   result?: string;
+  /** P131: tool step: wall-clock duration of the sidecar call (P130 computed it, but it
+   *  was dropped before reaching the step, so the UI and exports never showed it) */
+  durationMs?: number;
   /** P28: confirm step's parent agent requestId (for click → IPC 回执) */
   requestId?: string;
   /** P51: 该推理块是否还在流式接收中 */
diff --git a/tests/agent-loop.test.mjs b/tests/agent-loop.test.mjs
index 275e486..01c7389 100644
--- a/tests/agent-loop.test.mjs
+++ b/tests/agent-loop.test.mjs
@@ -40,3 +40,28 @@ test('looksLikeQuestion: asking for parameters ends the turn', async () => {
   assert.equal(looksLikeQuestion('What module and tooth count do you want?'), true);
   assert.equal(looksLikeQuestion('Plan: top plane, 80×50 plate, extrude 10 mm, then four M6 holes. Starting now.'), false);
 });
+
+// P131: P130's pattern matched generic words anywhere (which / 多少 / 几个 / 哪个 / 你想), so
+// ordinary plans read as questions (nudge skipped → user must type "继续" again), and
+// common parameter requests without a question mark were missed.
+test('looksLikeQuestion: plans that mention which/几个/哪个 still proceed', async () => {
+  const { looksLikeQuestion } = await import('../dist/main/main/agent/agent-loop-sidecar.js');
+  const plans = [
+    'Plan: I will sketch an 80×50 rectangle on the Top Plane, which is the XZ plane in SolidWorks, extrude it 10 mm, then cut four Ø6.6 through holes at (±30, ±15). Starting now.',
+    '方案：先确定几个关键尺寸——板 80×50×10，四角 M6 通孔距边 10 mm。步骤：1) 上视基准面画矩形并拉伸 10；2) 在顶面画 4 个圆并切除。现在开始执行。',
+    '整体方案如下：齿轮模数 2、齿数 20、齿宽 15、内孔 Ø10。先说明一下在哪个基准面建模：用前视基准面，这样齿轮轴线沿 Z。接下来调用 create_spur_gear 生成齿轮，再切键槽。',
+    '根据你想要的规格（法兰外径 120、厚 12、6 个 M8 孔均布在 PCD 90 上），方案：1) 前视基准面画 Ø120 圆拉伸 12；2) 顶面画 Ø9 孔，圆周阵列 6 个；3) 中心孔 Ø40。开始建模。',
+    '方案：壁厚统一 2 mm，不管外形多少个圆角，最后统一 fillet R1。先画 100×60 轮廓拉伸 30，再 shell 2 mm 开口朝上，最后倒角。现在开始。',
+  ];
+  for (const p of plans) assert.equal(looksLikeQuestion(p), false, p);
+});
+
+test('looksLikeQuestion: parameter requests without a question mark end the turn', async () => {
+  const { looksLikeQuestion } = await import('../dist/main/main/agent/agent-loop-sidecar.js');
+  const asks = [
+    '要生成这个齿轮，我还缺少以下参数：\n1. 模数 m\n2. 齿数 z\n3. 齿宽\n4. 内孔直径\n收到后我会立即建模。',
+    '为了准确建模，请补充下列尺寸：法兰外径、厚度、螺栓孔数量与孔径、分布圆直径。拿到这些数值后我会一次性完成建模，不会使用示例值代替。',
+    'Before I build the gear I need a few values from you: the module, the number of teeth, the face width and the bore diameter. Once I have them I will create it in one go.',
+  ];
+  for (const a of asks) assert.equal(looksLikeQuestion(a), true, a);
+});
diff --git a/tests/llm-request.test.mjs b/tests/llm-request.test.mjs
new file mode 100644
index 0000000..7de541e
--- /dev/null
+++ b/tests/llm-request.test.mjs
@@ -0,0 +1,220 @@
+// tests/llm-request.test.mjs
+//
+// P131: the request bodies the adapters actually put on the wire, per model.
+// Current Claude models (Fable 5 / 5.1, Opus 4.7+, Sonnet 5+) reject `temperature`
+// and `budget_tokens` with a 400; GPT-6 Astra rejects `max_tokens`, `temperature`
+// and `reasoning_effort` alongside tools. A stubbed global fetch records each body.
+
+import { test, afterEach } from 'node:test';
+import assert from 'node:assert/strict';
+import { AnthropicAdapter } from '../dist/main/main/llm/anthropic.js';
+import { OpenAIAdapter } from '../dist/main/main/llm/openai.js';
+import {
+  claudeCaps,
+  isOpenAIReasoningModel,
+  isReasoningParamError,
+} from '../dist/main/main/llm/thinking.js';
+
+const realFetch = globalThis.fetch;
+afterEach(() => { globalThis.fetch = realFetch; });
+
+/** Stub fetch: answers each request with the next queued [status, json] pair. */
+function stubFetch(...replies) {
+  const bodies = [];
+  globalThis.fetch = async (_url, init) => {
+    bodies.push(JSON.parse(init.body));
+    const [status, json] = replies.length > 1 ? replies.shift() : replies[0];
+    return new Response(JSON.stringify(json), {
+      status, headers: { 'content-type': 'application/json' },
+    });
+  };
+  return bodies;
+}
+
+const CLAUDE_OK = { content: [{ type: 'text', text: 'ok' }], stop_reason: 'end_turn' };
+const OPENAI_OK = { choices: [{ message: { content: 'ok' }, finish_reason: 'stop' }] };
+const MSGS = [{ role: 'user', content: 'hi' }];
+const TOOLS = [{ type: 'function', function: { name: 'sw_status', description: 'x', parameters: { type: 'object', properties: {} } } }];
+
+const claude = (model, extra = {}) => new AnthropicAdapter({
+  protocol: 'anthropic', baseURL: 'https://api.anthropic.com', apiKey: 'k', model, temperature: 0.3, ...extra,
+});
+const openai = (model, extra = {}) => new OpenAIAdapter({
+  protocol: 'openai', baseURL: 'https://api.openai.com/v1', apiKey: 'k', model, temperature: 0.3, ...extra,
+});
+
+test('claudeCaps: model families', () => {
+  for (const id of ['claude-fable-5-1', 'claude-fable-5', 'claude-mythos-5-1', 'claude-opus-5-5', 'claude-sonnet-5-5']) {
+    assert.equal(claudeCaps(id).alwaysThinks, true, id);
+    assert.equal(claudeCaps(id).noSampling, true, id);
+  }
+  for (const id of ['claude-opus-5', 'claude-opus-4-8', 'claude-opus-4.7', 'anthropic.claude-sonnet-5']) {
+    assert.equal(claudeCaps(id).noSampling, true, id);
+    assert.equal(claudeCaps(id).alwaysThinks, false, id);
+  }
+  assert.equal(claudeCaps('claude-opus-4-8').thinksByDefault, false);
+  assert.equal(claudeCaps('claude-opus-5').thinksByDefault, true);
+  assert.deepEqual(claudeCaps('claude-sonnet-4-6'), { noSampling: false, adaptive: true, thinksByDefault: false, alwaysThinks: false });
+  for (const id of ['claude-haiku-4-5-20251001', 'claude-3-7-sonnet-20250219', 'claude-sonnet-4-20250514', 'MiniMax-M3', '']) {
+    assert.equal(claudeCaps(id).adaptive, false, id);
+    assert.equal(claudeCaps(id).noSampling, false, id);
+  }
+});
+
+test('anthropic: Fable 5.1 / Opus 5.5 never get temperature or budget thinking', async () => {
+  for (const model of ['claude-fable-5-1', 'claude-opus-5-5']) {
+    for (const reasoningLevel of ['auto', 'off', 'adaptive', 'low', 'medium', 'high']) {
+      const bodies = stubFetch([200, CLAUDE_OK]);
+      await claude(model, { reasoningLevel }).chatWithTools(MSGS, undefined, TOOLS);
+      const b = bodies[0];
+      assert.equal('temperature' in b, false, `${model}/${reasoningLevel} sent temperature`);
+      assert.notEqual(b.thinking?.type, 'disabled', `${model}/${reasoningLevel} disabled thinking`);
+      assert.equal(b.thinking?.budget_tokens, undefined, `${model}/${reasoningLevel} sent budget_tokens`);
+      assert.equal('tool_choice' in b, false, 'forced tool_choice is a 400 on these models');
+    }
+  }
+});
+
+test('anthropic: reasoning level maps to adaptive thinking + effort', async () => {
+  let bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-opus-5-5', { reasoningLevel: 'high' }).chat(MSGS);
+  assert.deepEqual(bodies[0].thinking, { type: 'adaptive', display: 'summarized' });
+  assert.deepEqual(bodies[0].output_config, { effort: 'high' });
+
+  bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-opus-5-5', { reasoningLevel: 'off' }).chat(MSGS);
+  assert.deepEqual(bodies[0].output_config, { effort: 'low' }, 'off → lowest effort on an always-thinking model');
+
+  bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-opus-4-8', { reasoningLevel: 'off' }).chat(MSGS);
+  assert.deepEqual(bodies[0].thinking, { type: 'disabled' });
+  assert.equal('temperature' in bodies[0], false);
+
+  bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-opus-4-8').chat(MSGS);
+  assert.equal('thinking' in bodies[0], false, 'auto on Opus 4.8 = provider default');
+  assert.equal('temperature' in bodies[0], false);
+});
+
+test('anthropic: older / proxied models keep temperature and budget thinking', async () => {
+  let bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-sonnet-4-6').chat(MSGS);
+  assert.equal(bodies[0].temperature, 0.3);
+
+  bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-haiku-4-5', { reasoningLevel: 'medium' }).chatWithTools(MSGS, undefined, TOOLS);
+  assert.deepEqual(bodies[0].thinking, { type: 'enabled', budget_tokens: 4096 });
+  assert.equal('temperature' in bodies[0], false, 'budget thinking requires the default temperature');
+
+  bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('MiniMax-M3', { baseURL: 'https://api.minimax.io/anthropic' }).chat(MSGS);
+  assert.equal(bodies[0].temperature, 0.3);
+});
+
+test('anthropic: a 400 naming temperature is retried once without it', async () => {
+  const bodies = stubFetch(
+    [400, { type: 'error', error: { type: 'invalid_request_error', message: 'temperature: is not supported for this model' } }],
+    [200, CLAUDE_OK],
+  );
+  const r = await claude('claude-unknown-9').chat(MSGS);
+  assert.equal(bodies.length, 2);
+  assert.equal(bodies[0].temperature, 0.3);
+  assert.equal('temperature' in bodies[1], false);
+  assert.equal(r.content, 'ok');
+});
+
+test('anthropic: an unrelated 400 is not retried', async () => {
+  const bodies = stubFetch([400, { type: 'error', error: { message: 'messages: roles must alternate' } }]);
+  await assert.rejects(claude('claude-sonnet-4-6').chat(MSGS));
+  assert.equal(bodies.length, 1);
+});
+
+test('anthropic: refusal stop_reason surfaces a notice instead of an empty reply', async () => {
+  stubFetch([200, { content: [], stop_reason: 'refusal', stop_details: { type: 'refusal', category: 'cyber' } }]);
+  const r = await claude('claude-fable-5-1').chatWithTools(MSGS, undefined, TOOLS);
+  assert.match(r.content, /拒绝/);
+  assert.match(r.content, /cyber/);
+});
+
+test('openai: GPT-6 Astra is a reasoning model on OpenAI hosts only', () => {
+  assert.equal(isOpenAIReasoningModel('https://api.openai.com/v1', 'gpt-6-astra'), true);
+  assert.equal(isOpenAIReasoningModel('https://api.openai.com/v1', 'gpt-5.6-sol'), true);
+  assert.equal(isOpenAIReasoningModel('https://api.openai.com/v1', 'o4-mini'), true);
+  assert.equal(isOpenAIReasoningModel('https://api.openai.com/v1', 'gpt-4.1'), false);
+  assert.equal(isOpenAIReasoningModel('https://api.siliconflow.cn/v1', 'gpt-6-astra'), false);
+});
+
+test('openai: GPT-6 Astra gets max_completion_tokens and no temperature', async () => {
+  const bodies = stubFetch([200, OPENAI_OK]);
+  await openai('gpt-6-astra').chatWithTools(MSGS, undefined, TOOLS);
+  const b = bodies[0];
+  assert.ok(b.max_completion_tokens > 0);
+  assert.equal('max_tokens' in b, false);
+  assert.equal('temperature' in b, false);
+});
+
+test('openai: GPT-6 Astra drops reasoning_effort when tools are present', async () => {
+  let bodies = stubFetch([200, OPENAI_OK]);
+  await openai('gpt-6-astra', { reasoningLevel: 'high' }).chatWithTools(MSGS, undefined, TOOLS);
+  assert.equal('reasoning_effort' in bodies[0], false);
+
+  bodies = stubFetch([200, OPENAI_OK]);
+  await openai('gpt-6-astra', { reasoningLevel: 'off' }).chat(MSGS);
+  assert.equal(bodies[0].reasoning_effort, 'low', "'minimal' is rejected by GPT-6");
+
+  bodies = stubFetch([200, OPENAI_OK]);
+  await openai('gpt-5.6-sol', { reasoningLevel: 'high' }).chatWithTools(MSGS, undefined, TOOLS);
+  assert.equal(bodies[0].reasoning_effort, 'high', 'GPT-5.x keeps its effort with tools');
+});
+
+test('openai: test() uses max_completion_tokens for reasoning models', async () => {
+  let bodies = stubFetch([200, OPENAI_OK]);
+  await openai('gpt-6-astra').test();
+  assert.equal('max_tokens' in bodies[0], false);
+  assert.ok(bodies[0].max_completion_tokens > 0);
+
+  bodies = stubFetch([200, OPENAI_OK]);
+  await openai('deepseek-v4-pro', { baseURL: 'https://api.deepseek.com' }).test();
+  assert.equal(bodies[0].max_tokens, 1);
+});
+
+test('isReasoningParamError: GPT-6 tools + reasoning_effort rejection is recognised', () => {
+  assert.equal(isReasoningParamError(JSON.stringify({ error: { message:
+    "Function tools with reasoning_effort are not supported for gpt-6-astra in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'." } })), true);
+  assert.equal(isReasoningParamError('{"error":{"message":"Incorrect API key provided"}}'), false);
+});
+
+test('truncateMessages: an overhead larger than the budget keeps the last whole block', async () => {
+  const { truncateMessages } = await import('../dist/main/main/llm/context-window.js');
+  const history = [
+    { role: 'user', content: 'make a plate' },
+    { role: 'assistant', content: '', toolCalls: [{ id: 'c1', name: 'new_part', parameters: {} }] },
+    { role: 'tool', toolCallId: 'c1', content: 'ok' },
+  ];
+  const kept = truncateMessages(history, 'x '.repeat(200_000), 'claude-opus-5-5', 8192);
+  assert.notEqual(kept[0]?.role, 'tool', 'a lone tool result is a 400');
+  assert.equal(kept[0]?.role, 'assistant');
+  assert.equal(kept.length, 2);
+});
+
+test('llmFetch: connectMs 0 lets a slow non-streaming answer finish', async () => {
+  const { llmFetch } = await import('../dist/main/main/llm/net.js');
+  let calls = 0;
+  globalThis.fetch = (_u, init) => new Promise((resolve, reject) => {
+    calls++;
+    const t = setTimeout(() => resolve(new Response('{}', { status: 200 })), 80);
+    init.signal?.addEventListener('abort', () => { clearTimeout(t); reject(init.signal.reason); });
+  });
+  await assert.rejects(llmFetch('https://x.test', {}, 0, { connectMs: 20 }), /connect timeout/);
+  calls = 0;
+  const res = await llmFetch('https://x.test', {}, 2, { connectMs: 0 });
+  assert.equal(res.status, 200);
+  assert.equal(calls, 1, 'a finished-but-slow request must not be re-sent');
+});
+
+test('anthropic: Sonnet 4.6 keeps temperature when thinking is switched off', async () => {
+  const bodies = stubFetch([200, CLAUDE_OK]);
+  await claude('claude-sonnet-4-6', { reasoningLevel: 'off' }).chat(MSGS);
+  assert.deepEqual(bodies[0].thinking, { type: 'disabled' });
+  assert.equal(bodies[0].temperature, 0.3);
+});
diff --git a/tests/sidecar-client.test.mjs b/tests/sidecar-client.test.mjs
new file mode 100644
index 0000000..28ab96a
--- /dev/null
+++ b/tests/sidecar-client.test.mjs
@@ -0,0 +1,122 @@
+// tests/sidecar-client.test.mjs
+//
+// P131: the Node sidecar client against the REAL sidecar server (sidecar/sw_agent/server.py),
+// with a few fake tools registered by a throwaway bootstrap — no SolidWorks needed.
+//   - requests go out one at a time; a queued call's budget starts when it is sent
+//   - a failed call is never re-run by the same-op_id TIMEOUT follow-up
+//   - heartbeats cannot stretch a call past its absolute cap; onStillRunning fires on them
+//   - an aborted signal returns CANCELLED at once
+//   - after an idle gap, back-to-back requests are all answered (P129 stdin-watchdog stall)
+// Skipped when no python3 is on PATH.
+
+import { test, before, after } from 'node:test';
+import assert from 'node:assert/strict';
+import { createRequire } from 'node:module';
+import { spawnSync } from 'node:child_process';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const require = createRequire(import.meta.url);
+const Module = require('module');
+const origResolve = Module._resolveFilename;
+Module._resolveFilename = function (req, ...rest) {
+  if (req === 'electron') return 'electron-stub';
+  return origResolve.call(this, req, ...rest);
+};
+const tmp = fs.mkdtempSync(path.join(os.tmpdir(), 'mw-sidecar-'));
+require.cache['electron-stub'] = { id: 'electron-stub', filename: 'electron-stub', loaded: true,
+  exports: { app: { requestSingleInstanceLock: () => true, on() {}, getPath: () => tmp } } };
+const { SWSidecar } = require('../dist/main/main/com/sw-sidecar.js');
+
+const PY = ['python3', 'python'].find((p) => spawnSync(p, ['--version']).status === 0);
+const SIDECAR = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..', 'sidecar');
+
+fs.writeFileSync(path.join(tmp, '_bootstrap.py'), `
+import os, sys, time
+os.environ["SW_AGENT_SESSION_LOG"] = "0"
+sys.path.insert(0, ${JSON.stringify(SIDECAR)})
+from sw_agent import server
+from sw_agent.registry import tool
+from sw_agent.bridge import SWError
+
+server.HEARTBEAT_S = 0.1
+RUNS = {}
+
+def _bump(k):
+    RUNS[k] = RUNS.get(k, 0) + 1
+    return RUNS[k]
+
+@tool("t_sleep", "sleeps", params={"secs": {"type": "number"}}, category="query", internal=True)
+def t_sleep(ctx, secs=0.1):
+    time.sleep(secs)
+    return {"slept": secs}
+
+@tool("t_fail", "fails after working", params={"secs": {"type": "number"}}, category="query", internal=True)
+def t_fail(ctx, secs=0.1):
+    n = _bump("fail")
+    time.sleep(secs)
+    raise SWError(f"half-built, then failed (run {n})")
+
+server.serve()
+`);
+
+let sc;
+before(async () => {
+  if (!PY) return;
+  sc = new SWSidecar({ pythonPath: PY, cwd: tmp });
+  await sc.start();
+});
+after(() => { sc?.stop(); fs.rmSync(tmp, { recursive: true, force: true }); });
+
+const opts = { skip: !PY && 'python3 not available' };
+
+test('sidecar: a queued call is not timed out while it waits its turn', opts, async () => {
+  const slow = sc.call('t_sleep', { secs: 0.8 }, { timeoutMs: 5000 });
+  const queued = sc.call('t_sleep', { secs: 0.05 }, { timeoutMs: 400 });   // < slow's run time
+  const [a, b] = await Promise.all([slow, queued]);
+  assert.equal(a.ok, true);
+  assert.equal(b.ok, true, `queued call failed: ${b.code} ${b.error}`);
+});
+
+test('sidecar: the TIMEOUT follow-up never re-runs a failed call', opts, async () => {
+  const r = await sc.call('t_fail', { secs: 0.6 }, { timeoutMs: 300, opId: 'fail-once' });
+  assert.equal(r.ok, false);
+  assert.match(String(r.error), /run 1\)/, 'the follow-up must return the cached failure of run 1');
+  const again = await sc.call('t_fail', { secs: 0 }, { timeoutMs: 2000, opId: 'fail-once' });
+  assert.match(String(again.error), /run 1\)/);
+});
+
+test('sidecar: heartbeats cannot stretch a call past its cap; progress is reported', opts, async () => {
+  let beats = 0;
+  const t0 = Date.now();
+  const r = await sc.call('t_sleep', { secs: 3 }, { timeoutMs: 300, onStillRunning: () => beats++ });
+  const took = Date.now() - t0;
+  assert.equal(r.code, 'TIMEOUT');
+  assert.ok(took < 2500, `took ${took} ms — the 900 ms cap (+300 ms follow-up) was not enforced`);
+  assert.ok(beats >= 3, `onStillRunning fired ${beats}×`);
+  assert.ok(r.durationMs >= took - 50, 'durationMs covers the follow-up too');
+  await sc.call('t_sleep', { secs: 0 }, { timeoutMs: 5000 });   // drain: wait for the 3 s job
+});
+
+test('sidecar: an aborted signal returns CANCELLED at once', opts, async () => {
+  const ac = new AbortController();
+  setTimeout(() => ac.abort(), 100);
+  const t0 = Date.now();
+  const r = await sc.call('t_sleep', { secs: 1 }, { timeoutMs: 5000, signal: ac.signal });
+  assert.equal(r.code, 'CANCELLED');
+  assert.ok(Date.now() - t0 < 600);
+  const p = await sc.ping();   // queued behind the abandoned job, then answered
+  assert.equal(p.ok, true);
+});
+
+test('sidecar: back-to-back requests after an idle gap are all answered', opts, async () => {
+  await new Promise((r) => setTimeout(r, 2500));   // P129's stdin watchdog peeked every 2 s
+  for (let i = 0; i < 3; i++) {
+    const t0 = Date.now();
+    const p = await sc.ping();
+    assert.equal(p.ok, true, `ping ${i}: ${p.code}`);
+    assert.ok(Date.now() - t0 < 1000, `ping ${i} took ${Date.now() - t0} ms`);
+  }
+});