feat(agent): support dynamic skill visibility

[Voitamin]

What

把 SkillManager 构造时的 Skill 列表拆成两个概念：

• allowed_skills：权限/启用边界，运行中不可越过
• skills：初始可见集合

新增：

• expose_skills(ids)：在当前 session 内动态把 allowed skill 加入可见
• list_skill_metadata(scope='allowed'|'visible')：对外暴露 descriptor catalog（不含 SKILL.md 正文）
• 动态 exposed 存 session-local workspace，session 结束自动清理，不跨会话泄漏
• ReactAgent/LazyLLMAgentBase 支持传入预构建 skill_manager

Why

LazyMind 需要在 AgentLoop 前做候选筛选、Loop 内二次发现，但原 SkillManager 把「允许范围」和「当前可见范围」混为一谈，无法动态扩大可见范围。

Tests

• test_skills.py + test_agent_base.py：allowed/visible 分离、aliases/tags 归一化、disable-model-invocation 不进 allowed、动态 expose 边界、Agent 接受预构建 manager（25 passed）
• 既有 test_react_agent_with_skills 需下载 Qwen2.5-32B + ModelScope，与本次无关，本地 deselect

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 286514d8a0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Refresh the model prompt after exposing skills

When expose_skills() is called after a ReactAgent has already been built, these newly appended keys never reach the model: FunctionCall eagerly evaluates skill_manager.build_prompt() when constructing its ChatPrompter (functionCall.py:90-94), and the once-wrapped agent build is not repeated. The manager's tools will accept the exposed skill, but the model-facing skill list remains unchanged, so the model cannot discover it; the prompter should resolve the skill prompt dynamically or be refreshed after exposure.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Preserve exposed skills for the entire agent session

When an agent invocation produces its final string, FunctionCall.forward() pops the entire workspace (functionCall.py:235-240), including _exposed_skills stored here. Thus exposing a skill once and invoking the same agent again within the same new_session silently revokes that skill, contrary to this API's documented session scope; store exposure outside the transient invocation workspace or explicitly preserve it during cleanup.

Useful? React with 👍 / 👎.

[Yuang-Deng]

感谢这个 PR 对 Skill progressive disclosure 的支持。allowed skills 和当前可用 Skill 的区分，以及运行过程中动态 expose Skill 的能力，为后续 LazyMind 支持数百个 Skill 提供了比较重要的底层能力。

架构上建议后续重点考虑以下几点：

• 进一步拆分 SkillManager 的职责。 当前 SkillManager 同时承担 Skill 目录/元数据管理、可见状态、Prompt 渲染和 Skill 执行等职责。后续更适合拆分为共享的 SkillCatalog 和一次用户 Query 级别的 SkillRuntime，避免 Catalog 状态和运行时状态混在一起。
• 避免 LazyMind 直接依赖具体的 SkillManager 实现。 LazyMind 需要的是“本次 Query 可以使用哪些 Skill、当前加载了哪些 Skill”的运行时能力，建议通过更稳定的 Runtime/Protocol 接口传递，避免当前 skill_manager 参数过早成为固定接口。
• Skill 内容建议作为 Runtime Context 管理。 SkillRuntime 的生命周期是一整个用户 Query，覆盖内部多轮 LLM 请求和工具调用。get_skill 加载的完整 Skill 内容后续更适合作为 Runtime 临时上下文持续注入，而不是作为普通 Tool Result 写入 Conversation History。