AgentSkillsCN

multi-agent-orchestrator

在管理者主导的多智能体执行过程中，使用此技能进行计划管控、文件所有权边界划定、审核人/验证人检查，以及计划收尾工作。

SKILL.md
--- frontmatter
name: multi-agent-orchestrator
description: Use this skill to run manager-led multi-agent execution with plan gating, file ownership boundaries, reviewer/verifier checks, and plan closure.
version: 1.0.0

Multi-Agent Orchestrator

Use this skill when a task should run with a manager + specialist workflow instead of a single-agent change.

Trigger intent

  • •multi-agent-orchestrator <ISSUE-ID>
  • •orchestrate <ISSUE-ID>
  • •execute parent issue <ISSUE-ID>

If the input is an issue ID (for example DIM-14), treat it as the root orchestration issue.

Default operating contract

  1. •Follow nearest AGENTS.md and docs/agent/*.
  2. •If input is a root issue:
    • •load parent issue details
    • •discover child issues
    • •load dependency edges (blockedBy / blocks)
    • •build dependency waves (topological order)
    • •execute same-wave issues in parallel when ownership is disjoint
    • •gate each wave on reviewer/verifier checks
    • •when running in GitHub Copilot CLI:
      • •enable /fleet for same-wave parallel subagents
      • •use /tasks to track and wait for subagent completion
      • •if /fleet is unavailable, execute the same waves sequentially
  3. •Use Linear parent/child issues as the primary orchestration record.
  4. •Only for exceptional high-risk/audit/incident cases, create a plan in docs/agent/exec-plans/active/YYYYMMDD-<slug>.md.
  5. •Decompose by file ownership:
    • •client: {{CLIENT_PATH}}
    • •server: {{SERVER_PATH}}
    • •shared: {{SHARED_PATH}}
  6. •Assign mode per task: delegate, review, or own.
  7. •Run reviewer pass against:
    • •docs/agent/security.md
    • •docs/agent/quality.md
  8. •Run verifier checks and capture command evidence.
  9. •If formal plan was used, move plan to docs/agent/exec-plans/completed/.