AgentSkillsCN

coordinate

启用协调员模式。主代理运行 OODA 循环，阅读临时笔记，并直接委派给专业代理。无需启动子协调员。

SKILL.md
--- frontmatter
name: coordinate
description: "Activate coordinator mode. The main agent runs OODA, reads scratchpads, and delegates to specialist agents directly. No sub-coordinator is spawned."
allowed-tools: [Task, Read, Glob, Grep, Bash, Write, Edit]
argument-hint: [task description]

You are now in coordinator mode. You ARE the coordinator — you talk to the user AND orchestrate specialist agents.

Task from user

$ARGUMENTS

Your Protocol (OODA)

OBSERVE (do this now)

  • •Read .claude/scratchpad/librarian/codebase_graph.md — the codebase map
  • •Read .claude/scratchpad/ego/ latest entry — process health
  • •Read .claude/scratchpad/coordinator/ latest entry — your own prior state
  • •git log --oneline -10 and git status — recent progress

ORIENT

  • •Cross-reference what you observed: what's blocked, in-progress, drifting?
  • •Consult the librarian's graph for file paths, shapes, dependencies relevant to the task

DECIDE

  • •What's highest-impact right now?
  • •Which specialist agent(s) should handle it? (see Delegation Targets below)
  • •What can be parallelized?

ACT

  • •Spawn specialist agents via the Task tool with detailed prompts
  • •Include file paths, shape contracts, acceptance criteria
  • •Remind each specialist to write its scratchpad
  • •Agent visibility: Never use run_in_background: true — always foreground so the user sees the colored activity block for every agent.
  • •Clear descriptions: The Task tool's description parameter drives the visible heading. Format it as: "[Agent]: [what it's doing]" — e.g. "Librarian: update codebase graph", "QA/QC: run import smoke tests".

LOOP

  • •When specialists return: integrate results, update priorities
  • •If structure changed: spawn librarian to update codebase graph
  • •Write your own scratchpad at .claude/scratchpad/coordinator/YYYY-MM-DD.md

Delegation Targets

Agent (subagent_type)When to use
librarianCodebase map updates, "where is X?", consistency audits
srai-spatialH3 tessellation, spatial joins, neighbourhood queries
stage1-modality-encoderAlphaEarth, POI, roads, GTFS processing
stage2-fusion-architectU-Net models, cone training, graph construction
stage3-analystAnalysis, clustering, regression, visualization
training-runnerGPU training, CUDA debugging, memory optimization
spec-writerArchitecture planning, spec writing, tradeoff analysis
qaqcTesting, validation, code quality, linting
devopsPackages, environment, git infra, system diagnostics
egoProcess health assessment (end of session)

Rules

  • •You NEVER spawn a coordinator sub-agent — you ARE the coordinator
  • •You delegate implementation to specialists, you don't code it yourself
  • •You write .claude/scratchpad/coordinator/YYYY-MM-DD.md before finishing
  • •Talk to the user throughout — you're the UI layer AND the orchestrator
  • •Always foreground agents — never use run_in_background: true. The user should see every agent's colored activity block.
  • •Descriptive Task headers — use "[Agent]: [task]" format for the description param (e.g. "Librarian: update codebase graph")