fix(buzz-acp): raise MODELS_TIMEOUT to 60s so slow-start agents don't fail models/probe subcommands

[paramiao]

UPSTREAM-PR-DESC — buzz-acp models 超时修复 PR 草稿

• 草稿时间: 2026-08-23 · itops · huanxi-buzz v1 收尾 D2
• 目标仓库: github.com/block/buzz（main 分支，草稿基线 = commit f995325，2026-08-22 HEAD）
• 配套 diff: 同目录 UPSTREAM-PR-DRAFT.patch（干净可提交形态；本地 fork 版含标注注释，在 /home/mk/buzz-huanxi/build/models-timeout.patch）
• ⚠️ 状态: 草稿，未提交。实际 push 到 GitHub 前需 main 最终确认（外部发布动作）。

---

PR Title（建议）

fix(buzz-acp): raise MODELS_TIMEOUT to 60s so slow-start agents don't fail `models`/probe subcommands

PR 描述（建议正文）

Problem

buzz-acp lightweight helper subcommands (models, methods, and the initialize probe path) are
guarded by a hardcoded MODELS_TIMEOUT = 10s (crates/buzz-acp/src/lib.rs, const MODELS_TIMEOUT).
Any ACP agent whose process spawn + initialize handshake takes longer than 10s cold fails these
subcommands with error: agent timed out (10s), even though the agent is fully functional.

Reproducible case: openclaw acp (Node.js-based ACP server, v2026.7.1) takes ~11.4s from spawn to
first initialize response on a cold start. Result:

$ BUZZ_ACP_AGENT_COMMAND=openclaw BUZZ_ACP_AGENT_ARGS=acp,--session,<id> buzz-acp models
error: agent timed out (10s)

Meanwhile the runtime agent-pool path already uses a 60s timeout for the exact same
client.initialize() call, so the runtime bridge works fine while the helper subcommands fail —
inconsistent behavior for the same operation.

Proposed fix (minimal)

Raise the helper-subcommand timeout constant from 10s to 60s, matching the runtime agent-pool
initialize timeout. Single-constant change, no behavior change for fast agents beyond a later
timeout error.

Alternative (happy to switch if preferred): make it configurable, e.g.
BUZZ_ACP_MODELS_TIMEOUT env var with a 60s default — useful for agents with even slower cold
starts (JVM, heavy Python envs). We kept the minimal diff here since 60s already matches the
runtime path constant.

Evidence

• Cold-start timing: initialize first response at 11.44s, session/new at 12.19s (measured against
  openclaw-acp v2026.7.1-2, stdio probe).
• Official binary: models → error: agent timed out (10s) (reproduced twice).
• Patched binary (only this constant changed): models → exit 0, agent identity reported correctly;
  runtime agent-pool behavior unchanged.

Testing

• buzz-acp models against openclaw ACP agent: timeout → PASS after patch
• buzz-acp models --json: structured agent name/version output correct
• 25s full bridge run: agent_pool_ready agents=1 (runtime path unaffected)
• cargo test suite on CI (upstream)

---

提交前 checklist（itops 执行，待 main 确认后进行）

1. fork block/buzz 到可发布账号（不使用个人机器凭据之外的新凭据；gh auth 现有身份）
2. 基于最新 upstream main 重做分支（草稿基线 f995325 可能已漂移；单 hunk 补丁上下文足够小，重放低风险）
3. git diff 重新生成（UPSTREAM-PR-DRAFT.patch 的 index hash 是草稿基线的，正式提交前需重算）
4. PR 正文 = 上面建议内容（去掉内部路径/内部项目名，替换为通用表述）
5. push + 开 PR，回填 PR URL 到 CLOSEOUT 文档

内部注意

• 提交内容不得包含任何密钥/私钥/内部域名/内部路径（本文件与 patch 已审查：无）
• patch 仅含 1 行常量变更 + doc comment 上下文，无敏感信息