[Fix] normalize agent skill lookup and tool filtering

[dingyi222666]

description: |-

This pr fixes extension-agent skill materialization, case-insensitive skill and sub-agent resolution, and delegated tool filtering so agent reloads and handoffs stay consistent.

New Features

None.

Bug fixes

• reset the remote skill materialization directory before rewriting skill resources so stale files do not leak across reloads
• make skill name matching case-insensitive in permission checks, activation, rendering, and active-skill lookups
• make sub-agent lookup fall back to case-insensitive name matching when the exact name differs only by casing
• stop relying on a prebuilt tool mask for sub-agent tool exposure and instead filter tools through the current permission rules
• clear cached materialized skills when the skill catalog is disposed or reloaded so later runs do not reuse stale state

Other Changes

• remove the automatic disabled-skill config insertion during skill import
• normalize visible skill indexing to lowercase for consistent lookups across the agent services
• keep fallback sub-agent context compatible by using the incoming tool mask when present and a permissive default otherwise

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[coderabbitai]

总体信息

Walkthrough

该PR围绕技能匹配、权限管理和远程技能物化进行优化。具体包括：引入技能物化器清理机制、实现远程技能目录重置逻辑、移除技能导入时的自动映射、将技能名称匹配改为不区分大小写，以及调整子代理工具掩码的权限处理流程。

Changes

Cohort / File(s)	Summary
Remote Skill Materialization packages/extension-agent/src/computer/materialize.ts	新增 SkillMaterializer.clear() 方法重置物化状态；引入 resetRemoteSkillDir() 函数在物化前远程清理目录，包括递归移除现有目录或强制移除非目录路径。
Skill Import Simplification packages/extension-agent/src/service/index.ts	移除 ChatLunaAgentService.importSkills() 中遍历导入技能名称并插入 skills.items 映射的代码；简化配置持久化逻辑。
Case-Insensitive Permission Matching packages/extension-agent/src/service/permissions.ts	新增 matchRuleIgnoreCase() 辅助函数，将技能名称过滤和技能启用决策改为不区分大小写的匹配；工具和MCP检查保持原有区分大小写逻辑。
Skill Visibility & Lifecycle packages/extension-agent/src/service/skills.ts	将 _visibleByName 映射改为小写键存储实现不区分大小写查询；强化技能子代理可用性检查使用不区分大小写比较；在 stop() 和 reload() 方法中新增物化器清理逻辑。
Sub-agent Tool Authorization packages/extension-agent/src/service/sub_agent.ts, packages/extension-agent/src/sub-agent/run.ts	移除子代理创建时的工具掩码计算；改为通过权限检查动态过滤工具；从 CreateSubAgentOptions 接口移除 mask 属性；调整工具选择为权限驱动的列表过滤方式。

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~65 minutes

Possibly related PRs

• feat(extension-tools): add OneBot file sender tool and improve agent/open-terminal robustness #794: 修改 packages/extension-agent/src/sub-agent/run.ts 中的子代理运行时创建和代理执行器行为。
• fix(extension-agent): restore keep-alive sandbox reconnects and skill bin checks #791: 涉及远程技能物化和远程移除逻辑的添加，与本PR的 resetRemoteSkillDir 功能直接相关。
• fix(extension-agent): keep sub-agent tool permissions independent #813: 修改 packages/extension-agent/src/service/sub_agent.ts 中的工具掩码处理，与本PR移除 permission.createSubAgentToolMask 的更改相关。

Poem

🐰 技能匹配不分大小，远程目录焕然一新，
权限检查更显通透，工具筛选更加灵活，
物化清理保持整洁，代理协作更添顺畅～

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	标题准确地反映了该拉取请求的主要目的——规范化代理技能查找和工具过滤。
Description check	✅ Passed	描述详细说明了修复内容，包括技能物理化、不区分大小写的解析和工具过滤的改进。

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/agent-system

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[gemini-code-assist]

Code Review

This pull request introduces case-insensitive matching for skills and sub-agents, refactors sub-agent tool management to use permission checks instead of explicit masks, and adds a mechanism to reset remote skill directories. A review comment suggests simplifying the shell command for directory cleanup by removing redundant existence checks, as rm -rf is idempotent.

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

packages/extension-agent/src/service/skills.ts (1)

123-130: ⚠️ Potential issue | 🟠 Major

把物化缓存清理提前到 reload() 开头。

现在 materializer.clear() 只会在 scanSkills() / buildSkillCatalog() 成功后执行。只要 reload 因坏 skill 或 I/O 错误中途抛错，旧的远端物化映射就会继续保留，后续仍可能复用陈旧状态，这和本次修复目标相反。

💡 建议修改

 async reload() {
+    this.ctx.chatluna_agent?.computer.materializer.clear()
     await syncBundledSkills(this.ctx)
     const local = await scanSkills(this.ctx, this.config)
     const scanned = local
     this._skills = new Map(scanned.map((s) => [s.id, s]))
     this._catalog = buildSkillCatalog(scanned, this.config.skills.items)
-    this.ctx.chatluna_agent?.computer.materializer.clear()
     this._visibleByName = new Map(
         this._catalog
             .filter((s) => s.visible)
             .map((s) => [s.name.toLowerCase(), this._skills.get(s.id)!])
     )

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@packages/extension-agent/src/service/skills.ts` around lines 123 - 130, The
materializer.clear() call should happen at the start of reload() so cached
remote materializations are cleared even if
syncBundledSkills/scanSkills/buildSkillCatalog fail; move the
this.ctx.chatluna_agent?.computer.materializer.clear() invocation to the top of
reload() (before awaiting syncBundledSkills/scanSkills) and keep the existing
optional chaining to avoid null deref; ensure reload() still proceeds to call
syncBundledSkills, scanSkills, buildSkillCatalog and then sets
this._skills/_catalog as before.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@packages/extension-agent/src/service/skills.ts`:
- Around line 123-130: The materializer.clear() call should happen at the start
of reload() so cached remote materializations are cleared even if
syncBundledSkills/scanSkills/buildSkillCatalog fail; move the
this.ctx.chatluna_agent?.computer.materializer.clear() invocation to the top of
reload() (before awaiting syncBundledSkills/scanSkills) and keep the existing
optional chaining to avoid null deref; ensure reload() still proceeds to call
syncBundledSkills, scanSkills, buildSkillCatalog and then sets
this._skills/_catalog as before.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 9697354d-fa01-406d-b51d-f6e8b501f204

📥 Commits

Reviewing files that changed from the base of the PR and between 3303710 and 14c699c.

📒 Files selected for processing (6)

• packages/extension-agent/src/computer/materialize.ts
• packages/extension-agent/src/service/index.ts
• packages/extension-agent/src/service/permissions.ts
• packages/extension-agent/src/service/skills.ts
• packages/extension-agent/src/service/sub_agent.ts
• packages/extension-agent/src/sub-agent/run.ts