[fengyue-xve]

Summary

Fixes #1849: in云端协同 mode, the hub's 「专家」 module could delete a peer (bridge) expert.

Two layers caused it:

1. Tunnel policy let the delete through: DELETE /api/agents/{id} matched _AGENT_RESOURCE and was tunneled to the peer, where it ran as the connection owner — the peer-side owner check (_assert_agent_owner) cannot stop it, because the request identity is the owner themselves.
2. UI offered the affordance: list_remote_agents maps bridge shadow experts with is_owner: true, so the delete button rendered for them too.

Changes:

• tunnel_policy.py: deny tunneling a bare agent-resource DELETE /api/agents/{id} (local or bridge: shadow id). Agent sub-surfaces (threads / files / avatar / skills / …) stay deletable, and editing (PATCH /api/agents/{id}) still tunnels.
• Dashboard: AgentExpertsTable and AgentCard hide the delete button for bridge experts, with a new render test (AgentExpertsTable.bridgeDelete.test.tsx).

Target branch

• Base is develop (feature / fix — default)
• Base is main (release/* or hotfix/* only)

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation
• Refactor / chore
• Release / hotfix

Test plan

• Added/updated tests:
  • tests/unit/bridge/test_tunnel_policy.py — new test_denies_deleting_the_agent_itself: bare DELETE (local id, bridge: id, trailing slash) denied; sub-path DELETE and PATCH still allowed. Ran locally: 7 passed.
  • dashboard: AgentExpertsTable.bridgeDelete.test.tsx renders a local + a bridge expert and asserts exactly one Delete button remains (the local row). Ran locally with vitest: 1 passed.
• make all passes locally — ran the targeted backend test module, ruff (check + format), and the targeted vitest file; full CI will cover the rest.

Checklist

• Updated CHANGELOG.md (if user-facing)
• README / docs updated (if needed)

[jubaoliang]

GitHub PR 提交规范

✅ PR 必备清单

1. 📸 提供视觉证明（截图/录屏）

必须贴截图，无论放在 PR 描述中还是评论里——让 reviewer 看到"人"，证明是经过人工检查、用心完成的。

2. 🧪 测试用例完整 & CI 通过

• ✅ 新增功能 → 新增对应测试
• ✅ Bug 修复 → 添加回归测试
• ✅ 所有 CI Workflow 必须通过 → 不要期望 reviewer 去 fix 你的 CI

3. 🔄 与 AI 协作：确认理解一致

• 需求明确后，让 AI 重复一遍需求，确认理解无误
• 代码生成后，主动追问：
  • "这个实现是否满足所有需求点？"
  • "是否有遗漏的场景？"

4. 🎯 代码质量：最小修改 & 追问优化

• 坚守最小修改原则，一个 PR 只解决一个问题
• AI 生成代码后，必问：
  • 是否还有优化空间？
  • 所有异常是否妥善处理？
  • 是否有冗余代码？

---

📋 Quick Checklist（可直接放进 PR 模板）

- [ ] 📸 已提供 UI 截图/功能演示
- [ ] 🧱 测试用例已添加并通过
- [ ] ✅ CI Workflow 全部通过
- [ ] 🤖 已与 AI 对齐需求理解
- [ ] 🎯 遵循最小修改原则，代码结构清晰
- [ ] 📝 已询问并处理代码优化建议

核心原则：PR 是沟通，不是交接。让 reviewer 轻松，就是让自己更快合入。

---

🔍 本 PR 自动检查发现的问题

• ❌ CI Workflow 未通过：当前存在失败的 check，请先修复到全绿
• 📸 本 PR 涉及前端改动，但描述中未提供截图/录屏等视觉证明

以上为 automated 检查提示，请补充后 @ 我重新评审。