Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Delegated subagent sessions cannot load any skill: built-in worker/explore/verifier roles omit the skill tool

未关闭 适合新手
  • 哪里不对？
#556 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
76/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
typescript
领域
ai-infra-agents, developer-experience

调研方向

阅读 packages/local-runtime-v2/assets/agents/{worker,explore,verifier}/agent.md，并将其中的工具和 skill 设置与根会话进行比较。检查原生 skill 工具和 skill 目录如何提供给委派会话。完成标准是：委派的 worker 能通过该工具加载已安装的 skill，且 explore/verifier 设置允许按预期共享访问 skill。

由索引模型根据 Issue 内容生成。

描述

Product or interface

Subagent delegation — the task tool and the built-in worker / explore / verifier roles — and the native skill tool.

Use case and problem

I use MiniMax Code as a long-running personal assistant and delegate almost all real work to subagents (task → worker / explore / verifier). Third-party skills are how I encode repeatable procedures (TDD, code review, debugging, planning), so that every delegated task follows the same process instead of reinventing it.

Delegated children currently cannot load any skill at all.

The built-in roles are defined in packages/local-runtime-v2/assets/agents/{worker,explore,verifier}/agent.md (verified at 50448abe), and their tools: lists do not contain the native skill tool:

  • worker/agent.md → read, write, edit, bash, task_query, task_output, task_stop, grep, glob, todowrite, web_fetch, website_deploy
  • explore/agent.md → bash, read, grep, glob, web_fetch, plus builtinTools: [] and skills: []
  • verifier/agent.md → read, grep, glob, bash, task_query, task_output, task_stop, web_fetch, plus builtinTools: [] and skills: []

Measured on my machine (desktop app, Windows): I ran four parallel delegated tasks with plain natural prompts — implement a function with tests, fix a failing test, review a diff, and plan a feature. Across 88 tool calls in 4 child sessions the skill tool was invoked 0 times. The tool schema actually offered to the child was:

bash, edit, glob, grep, mcp_invoke, read, tool_search, web_fetch, web_search, website_deploy, write

The only way a skill procedure reached a child was the child reading the SKILL.md file by hand, after tool_search for a skill loader returned {"hits":[],"total":0} — i.e. it went around the platform instead of through it. That works, but it is luck: it depends on the child happening to guess that the file exists, and the child never learns that the native loader exists.

Worth noting: packages/agent-tools/src/desktop/canonical-tool-policy.ts does not mention skill at all (the file is 80 lines and contains no occurrence of the string), so the policy layer is not what blocks it. The only gate is the per-role tools: list, so this looks like a one-line-per-role change rather than a policy decision.

The same gap applies to skills: [] on explore and verifier: even if the skill tool were listed, an explicit empty skills: array means there is no catalog to load from.

Desired behavior
  • A delegated child session can load a skill with the native skill tool, exactly like the root mavis session can.
  • skills: [] in explore/agent.md / verifier/agent.md either exposes the catalog, or is replaced by an explicit opt-out mechanism, so a delegated verifier can run a shared review procedure instead of improvising one.
  • Acceptance criteria: a delegated worker assigned a task that is covered by an installed skill invokes skill({ name }) for it without reading the file path by hand, and the returned SKILL.md content is visible in the child's context.
Platform

Desktop app 3.1.7.195 on Windows 11 x64. Source references are against upstream commit 50448abe (2026-10-29).

Reporter MMX UID: 500063342928973827

Alternatives and additional context

I worked around this today by creating three custom roles under agents/<name>/agent.md that add skill to tools: and drop skills: [], and delegating to those instead of the built-ins. It works, but every user has to re-create the built-in roles to get skill support, and the workaround is not discoverable from the product or from the built-in role descriptions.

Related, and worth checking in the same area: the official plugin packages that are published for MiniMax Code drop the upstream hooks/ directory, so a plugin that relies on a SessionStart hook can never fire in the desktop app. I filed that separately since it is a packaging problem rather than an agent-configuration one.

主要语言
TypeScript
星标
2.1k
派生
271
平均合并
14 小时 23 分钟
30 天内合并 PR
35

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

MiniMax-AI/minimax-code 的其他 Issue

查看 MiniMax-AI/minimax-code 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。