feat(manager): upgrade Manager runtime to QwenPaw 2.2

[LUOSENGWA]

Summary

Upgrade the Manager image's bundled QwenPaw runtime from 2.0.1 to 2.2.0, following the Worker migration in #1174. No API adaptation is needed — verified across the same three interaction layers (plugin Python API, HTTP API, config schema) and against the v2.2.0 release. The Manager also gains the 2.2.0 stalled-LLM-stream fix (#7150) — the same production defect the Worker PR closes — since the Manager runs qwenpaw too.

Companion to #1174 (Worker): both PRs are independent (zero file overlap), raised separately so each rebases cleanly onto main. Rebased onto origin/main (acaed255).

What's included

1. manager/Dockerfile.qwenpaw — QWENPAW_PIP_SPEC → qwenpaw==2.2.0 (+ doc comments "QwenPaw 2.2"). Unlike the Worker, the Manager has no require_version runtime check — its startup path (copaw_worker.run_copaw_app, API wait in qwenpaw_manager_skill_sync.py) performs no version assertion, so the Dockerfile QWENPAW_PIP_SPEC is the sole version control
2. manager/scripts/init/start-qwenpaw-manager.sh — "QwenPaw 2.0" comments → "2.2"
3. plugins/agentteams-manager-tools/plugin.json — qwenpaw_version → >=2.2.0,<2.3.0. Plugins declare qwenpaw_version: {min, max} with left-closed right-open semantics (>=min,<max); the old max: "2.1.0" rejects 2.2.0. The upper-bound check is currently temporarily disabled upstream (#6532), but the declaration must be correct for when full range checking is re-enabled

No changelog/current.md edit: upstream #1182 removed that file and switched to CI-generated release notes.

Compatibility (verified against v2.2.0 tag, no adaptation)

190 commits in v2.1.0..v2.2.0. The only commit that could touch the config/model-declaration surface is #6302 — audited: the Manager's config-schema output (agent.profiles) stays loadable (new AgentProfileConfig fields all have defaults), and the manager-tools plugin's qwenpaw API surface is unchanged.

• Manager tools (projectflow/taskflow/message/filesync) register via the agentteams-manager-tools plugin, whose qwenpaw API surface (qwenpaw.runtime.hooks/phases, same as the Worker teamharness plugin) is unchanged in 2.2 — HookBase/HookResult present, Phase 8-value enum identical between 2.1/2.2
• Manager startup path uses copaw_worker.bridge (config conversion) + copaw_worker.run_copaw_app (qwenpaw app launch); the config schema it produces (agent.profiles) is backward-compatible
• HTTP API routes: zero destructive deletions to Manager-facing endpoints (same audit as feat(qwenpaw): upgrade QwenPaw Worker runtime to 2.2 #1174)

Status

QwenPaw 2.2.0 is released on PyPI — version declarations match the published version. Rebased onto origin/main (acaed255). Ready for review.

AgentTeams default for LLM stream stall timeouts: 300s (QwenPaw 2.2.0 #7150)

Companion to the Worker-side change in #1174.

QwenPaw 2.2.0 adds LLM stream stall detection (#7150) with a first-content and an idle timeout, both defaulting to 30s and read at process startup. In local-LLM deployments (self-hosted SGLang serving a large model under concurrent agent load), a stream can be starved for well over 30s without being dead, so the 30s default false-kills healthy requests — observed twice in our environment (2026-08-26 and 2026-08-28, single-rank SGLang serving a 27B-FP8 model).

This change sets the AgentTeams container default to 300s in the entrypoint/start scripts, with an overridable priority chain:

1. QWENPAW_LLM_STREAM_FIRST_CONTENT_TIMEOUT / QWENPAW_LLM_STREAM_IDLE_TIMEOUT (native, per-phase, finest control)
2. AGENTTEAMS_LLM_STREAM_TIMEOUT_S (single AgentTeams knob)
3. 300 (new default)

Deployments that want the upstream behavior can set AGENTTEAMS_LLM_STREAM_TIMEOUT_S=30 (or 0 to disable stall detection entirely).

---

将 Manager 镜像内置的 QwenPaw 运行时从 2.0.1 升级到 2.2.0，与 #1174 的 Worker 迁移一致。无需 API 适配——已在三个交互层（插件 Python API / HTTP API / config schema）和 v2.2.0 release 上验证。Manager 同样获得 2.2.0 的 stalled-LLM-stream 修复（#7150）——与 Worker PR 关闭的是同一生产缺陷，因为 Manager 也跑 qwenpaw。

Companion to #1174（Worker）：两个 PR 独立（零文件重叠），分开提交以便各自干净 rebase 到 main。已 rebase 到 origin/main（acaed255）。

改动清单

1. manager/Dockerfile.qwenpaw — QWENPAW_PIP_SPEC → qwenpaw==2.2.0（+ 注释同步 "QwenPaw 2.2"）。与 Worker 不同，Manager 无 require_version 运行时强校验——启动路径（copaw_worker.run_copaw_app、qwenpaw_manager_skill_sync.py 等 API）不做版本断言，版本控制完全靠 Dockerfile 的 QWENPAW_PIP_SPEC
2. manager/scripts/init/start-qwenpaw-manager.sh — "QwenPaw 2.0" 注释 → "2.2"
3. plugins/agentteams-manager-tools/plugin.json — qwenpaw_version → >=2.2.0,<2.3.0。插件声明 qwenpaw_version: {min, max}（左闭右开 >=min,<max），原 max: "2.1.0" 会拒绝 2.2.0。上限检查上游目前暂时禁用（#6532），但声明必须正确以防恢复检查后失效

无 changelog/current.md 改动：上游 #1182 删除了该文件，改为 CI 从 merged PRs 自动生成 release notes。

兼容性（对照 v2.2.0 tag 验证，无需适配）

v2.1.0..v2.2.0 共 190 commit。唯一可能动 config/模型声明面的是 #6302——已审计：Manager 的 config-schema 输出（agent.profiles）保持可加载（新 AgentProfileConfig 字段均有默认值），manager-tools 插件的 qwenpaw API 面零变化。

• Manager 工具（projectflow/taskflow/message/filesync）通过 agentteams-manager-tools 插件注册，其 qwenpaw API 依赖（qwenpaw.runtime.hooks/phases，与 Worker 的 teamharness 插件相同）在 2.2 零变化——HookBase/HookResult 存在，Phase 8 值枚举 2.1/2.2 完全一致
• Manager 启动路径用 copaw_worker.bridge（配置转换）+ copaw_worker.run_copaw_app（启动 qwenpaw app）；生成的 config schema（agent.profiles）向后兼容
• HTTP API 路由：Manager 面向的端点零破坏性删除（与 feat(qwenpaw): upgrade QwenPaw Worker runtime to 2.2 #1174 相同审计）

状态

QwenPaw 2.2.0 已发布到 PyPI——版本声明与发布版本一致。已 rebase 到 origin/main（acaed255）。可以 review。

LLM 流 stall 超时：AgentTeams 容器默认 300s（QwenPaw 2.2.0 #7150）

配套 #1174（Worker）的对应改动。

2.2.0 新增 LLM 流 stall 检测（#7150）：首内容 / 空闲双超时，均默认 30s，进程启动时读取。本地 LLM 部署（自托管 SGLang 并发服务大模型）下，流可能被并发负载饿死远超 30s 但并未死亡，30s 默认会误杀健康请求——我方环境 8/26、8/28 两度实锤（单 rank SGLang 服务 27B-FP8）。

本变更在容器启动脚本中将 AgentTeams 默认调为 300s，优先级链（可覆盖）：

1. QWENPAW_LLM_STREAM_FIRST_CONTENT_TIMEOUT / QWENPAW_LLM_STREAM_IDLE_TIMEOUT（原生变量，分相位细调，最细粒度）
2. AGENTTEAMS_LLM_STREAM_TIMEOUT_S（AgentTeams 单旋钮）
3. 300（新默认）

需要上游行为的部署：设 AGENTTEAMS_LLM_STREAM_TIMEOUT_S=30（或 0 完全禁用 stall 检测）。