Agent Skills
› uw-syfi/vibesys
› run-roles
run-roles
GitHub定义运行中各代理角色的职责，指导在编排任务时选择正确的目标角色或查询工作分配情况。
触发场景
需要指派任务给特定代理角色
询问哪个角色完成了某项工作
安装
npx skills add uw-syfi/vibesys --skill run-roles -g -y
SKILL.md
Frontmatter
{
"name": "run-roles",
"description": "What each agent role of a run does, so a steer or an answer names the right one. Use before choosing a target for a steer, or when asked who did what."
}
What each role does
A run's agents have roles. A role id is the search's name and the role (for example
<search>-implementer); the run declares its own, so call list_targets for the exact ids
and use those, never one you guessed. The kinds of role a search may have:
- planner: plans the next ideas; works in no workspace. Steer it to change what gets tried.
- implementer: implements one idea in an isolated copy of the code and takes its own measurements. Steer it to change how an idea is built.
- judge: reviews a candidate and the evaluations the implementer took, and gives the verdict. Steer it to change what counts as acceptable.
- profiler: looks at a revision to inform later plans; produces no candidate.
- merger: merges candidates into the main line.
- outliner: outlines the steps of a task's plan once, at the start, when the task has none.
A role the run does not use does nothing: aim a steer at a role that works on what the operator cares about. When unsure, steer the whole run or ask which role did the work, using the tools to find it. Evaluation is not steered: a steer changes what agents are told, never how a candidate is measured or accepted.
版本历史
- 3a50e8b 当前 2026-11-02 13:56
