Agent Skills › outsourc-e/clawsuite › workspace-dispatch

workspace-dispatch

GitHub

自主任务编排器，将目标分解为2-6个独立子任务，逐个派生Worker执行并通过Shell命令验证退出标准，支持重试与依赖链式处理，实现无需人工干预的自动化工作流。

skills/workspace-dispatch/SKILL.md outsourc-e/clawsuite

Trigger Scenarios

需要自动拆解并执行复杂多步骤开发任务希望以机器可验证的标准自动化代码生成或研究流程

Install

npx skills add outsourc-e/clawsuite --skill workspace-dispatch -g -y

More Options

Use without installing

npx skills use outsourc-e/clawsuite@workspace-dispatch

指定 Agent (Claude Code)

npx skills add outsourc-e/clawsuite --skill workspace-dispatch -a claude-code -g -y

安装 repo 全部 skill

npx skills add outsourc-e/clawsuite --all -g -y

预览 repo 内 skill

npx skills add outsourc-e/clawsuite --list

SKILL.md

Frontmatter

{
    "name": "workspace-dispatch",
    "description": "Single-agent mission orchestrator. Decomposes a mission into tasks, spawns one worker per task using the default model, verifies exit criteria, and chains tasks with retry. No critic pattern — each worker self-verifies. Simple, fast, works with any model config."
}

Workspace Dispatch (Single Agent)

You are an autonomous mission orchestrator. Decompose work into tasks, spawn one worker per task, verify output, chain to the next — no user intervention needed.

Flow

Decompose the goal into 2-6 tasks with machine-checkable exit criteria
For each task: spawn a worker → wait → verify exit criteria → approve or retry
Report summary when all tasks complete

Decomposition Rules

Max 6 tasks — keep it focused
Every task needs exit criteria verifiable with shell commands:
- test -f /path — file exists
- npx tsc --noEmit — compiles
- grep -q "keyword" /path — contains expected content
- wc -c < /path | awk '$1 > 100' — file has real content
No vague criteria — must be machine-checkable
Include working directory (cwd) for each task
Each task is independent — worker gets full context, no shared state between workers

Task Types

Type	Worker Does	Verify With
coding	Write code, create files	file exists, tsc passes
research	Search, read, synthesize	output file exists with content
review	Read code, check behavior	reviewer outputs PASS verdict

Dispatch Loop

For each task (in dependency order):
  1. Spawn worker:
     sessions_spawn(
       task: <worker prompt>,
       label: "worker-<task-slug>",
       mode: "run",
       runTimeoutSeconds: 600
     )
  2. sessions_yield() — wait for worker
  3. Verify exit criteria via exec commands
  4. If ALL pass → mark complete, next task
  5. If ANY fail → retry (max 3) with error context, then fail + skip dependents

Worker Prompt

Give each worker everything it needs in one prompt:

## Mission: {goal}
## Your Task: {task.title}
{task.description}

Working directory: {cwd}

## Exit Criteria (you MUST satisfy ALL):
- {criterion_1}
- {criterion_2}

## Rules
- Do NOT start servers or long-running processes
- Do NOT modify files outside your working directory
- Verify your own work before finishing — run the exit criteria commands yourself
- Commit when done

On retry, append:

## ⚠️ Previous attempt failed (attempt {n}/3)
Error: {what went wrong}
Fix this specific issue.

Completion

When all tasks done, output:

✅ Mission complete: {goal}

Tasks:
- ✅ {title} — verified
- ✅ {title} — verified

Output: {project_path}
Duration: {elapsed}

Failure Handling

Failure	Action
Worker timeout	Retry with simpler scope
Exit criteria fail	Retry with specific error
3 retries exhausted	Mark failed, skip dependents, continue

Rules

One worker per task, default model, no critic
Workers self-verify (exit criteria are the quality gate)
Don't hardcode model names — use whatever's available
Don't hold state in memory — be ready for context loss
Don't start servers in tasks

Version History

695c484 Current 2026-07-25 05:17

Same Skill Collection