AgentSkillsCN

jules-swarm-orchestrator

为代码库协调并行的Jules代理集群。派遣8个专业代理(架构师、Bolt、导航员、调色板、哨兵、验证者、清洁工、书记员),管理文件所有权以避免合并冲突,并定义评审与合并的工作流程。当用户说“运行Jules集群”“派遣代理”“Jules批量处理”“运行维护代理”或“进行并行代码评审”时,该技能便派上用场。

SKILL.md
--- frontmatter
name: jules-swarm-orchestrator
description: >-
  Orchestrates parallel Jules agent swarms for codebases. Dispatches 8 specialized
  agents (Architect, Bolt, Navigator, Palette, Sentinel, Verifier, Janitor, Scribe),
  manages file ownership to prevent merge conflicts, and defines review/merge workflow.
  Use when user says "run jules swarm", "dispatch agents", "jules batch",
  "run maintenance agents", or "parallel code review".
license: MIT
metadata:
  author: Semester AI
  version: 1.0.0
  category: workflow-automation
  tags: [jules, google, multi-agent, swarm, code-quality]

Jules Swarm Orchestrator

Dispatch and coordinate parallel Jules coding agents against a GitHub repository.

How It Works

Eight specialized agents run simultaneously in isolated Jules sessions. Each agent owns a specific concern and set of files. After all sessions complete, results are reviewed and merged in priority order.

Agent Roster

AgentConcernPrompt
Architect 🧭Structure, routing, module boundariesassets/prompts/architect.md
Bolt ⚡Performance, lazy loading, memoizationassets/prompts/bolt.md
Navigator 🧭UX flows, page logic, navigationassets/prompts/navigator.md
Palette 🎨Visual consistency, theming, a11yassets/prompts/palette.md
Sentinel 🛡️Error handling, validation, safetyassets/prompts/sentinel.md
Verifier 🧪Test coverage, correctnessassets/prompts/verifier.md
Janitor 🧹Dead code, lint, cleanupassets/prompts/janitor.md
Scribe 📝Docs, comments, changelogassets/prompts/scribe.md

Step 1: Dispatch

Use the bundled dispatch script or run manually:

bash
# All 8 agents
./scripts/dispatch.sh owner/repo

# Specific agents only
./scripts/dispatch.sh owner/repo sentinel verifier

# Manual single agent
jules new --repo owner/repo "$(cat assets/prompts/architect.md)"

Dispatch all agents in one batch so they work against the same base commit.

Step 2: Monitor

bash
jules remote list --session

Step 3: Review and Merge

bash
# List sessions
./scripts/review.sh

# Inspect a session's diff
./scripts/review.sh <session-id>

# Apply changes locally
./scripts/review.sh <session-id> apply

Merge in this order (higher priority first): Architect > Sentinel > Navigator > Palette > Bolt > Verifier > Janitor > Scribe

If two PRs conflict, merge the higher-priority agent first, then rebase the other.

File Ownership

Each agent may only edit files in its owned scope. See references/file-ownership.md for full matrices (Web, Android, Backend).

The pattern:

  • Config files (package.json, build configs, CI): human-only
  • Entry/routing: Architect
  • API/services: Sentinel
  • Pages: Navigator (Palette may adjust styling)
  • Components: Palette (Navigator may adjust layout)
  • Tests: Verifier
  • Docs: Scribe
  • Dead code: Janitor (except test/doc files)

Agents log out-of-scope issues as // TODO: [AgentName] ... comments.

Agent Rules

  1. Read AGENTS.md before starting — agent prompts supplement, never override
  2. Only edit files in your owned scope
  3. One concern per PR
  4. Never add dependencies, change auth, or modify deploy config
  5. Run the repo's build/test command before finishing
  6. Log out-of-scope issues as TODOs

Customizing

Drop agents that don't apply to your repo:

Repo TypeRecommended
Frontend (React/Vue)Architect, Navigator, Palette, Sentinel, Verifier, Janitor
Backend (API)Architect, Sentinel, Bolt, Verifier, Janitor, Scribe
Mobile (Android/iOS)All 8
Library/SDKArchitect, Sentinel, Verifier, Scribe

To customize prompts: copy assets/prompts/<agent>.md to your repo's .prompts/jules/ directory and edit.