AgentSkillsCN

workspace-task-execute

通过调度工作员与评审员子智能体，统筹安排工作项的任务执行。当用户希望执行任务、运行实施流水线，或开始按任务列表逐一推进时，可优先使用此技能。它会从 .claude/workitems/{type}/{name}/tasks/ 中提取待办任务，在工作树中循环执行工作员与评审员的协作流程，并标记任务为已完成。支持多实例并行执行，所有输出日志都会被记录在 .claude/workitems/{type}/{name}/logs/ 中。

SKILL.md
--- frontmatter
name: workspace-task-execute
description: Orchestrate task execution for a workitem by dispatching worker and reviewer subagents. Use when the user wants to execute tasks, run the implementation pipeline, or start working through a task list. Picks up pending tasks from .claude/workitems/{type}/{name}/tasks/, runs worker+reviewer cycles in the worktree, and marks tasks completed. Supports parallel execution via multiple instances. Logs all output to .claude/workitems/{type}/{name}/logs/.

Workspace Task Execute

Dispatch worker and reviewer subagents to execute tasks from a workitem's task list. Processes tasks sequentially, respecting dependencies and blockers, until all tasks are completed or remaining tasks are blocked.

Prerequisites

  • •Task files in .claude/workitems/{type}/{name}/tasks/ (via workspace-task-generate)
  • •Claude CLI (claude) available in PATH
  • •The executor validates that task files exist before starting

Usage

bash
# By workitem path (type/name)
bash <skill-path>/scripts/task_executor.sh feature/auth-middleware

# Interactive selection (no argument)
bash <skill-path>/scripts/task_executor.sh

Set MAX_TASK_ROUNDS to control max worker-reviewer rounds per task (default: 5):

bash
export MAX_TASK_ROUNDS=3
bash <skill-path>/scripts/task_executor.sh feature/auth-middleware

How It Works

  1. •Validates that task files exist (fails with guidance if workspace-task-generate hasn't been run)
  2. •Recovers any in-progress tasks from a previous crashed run back to pending
  3. •Finds the next pending task with all dependencies completed and no blockers
  4. •Marks the task as in-progress
  5. •Resolves the working directory:
    • •Task 01 (worktree setup): runs from workspace root
    • •All other tasks: reads .claude/workitems/{type}/{name}/worktree.path and runs from the worktree
  6. •Passes the absolute path to the task file in the subagent prompt
  7. •Subagents read the task file directly, do their work, and write progress/feedback back to the same file
  8. •Runs worker → reviewer cycle (up to MAX_TASK_ROUNDS)
  9. •If all acceptance criteria are checked: marks completed, moves to next
  10. •If max rounds reached: leaves as in-progress for manual review
  11. •Stops when no more tasks are available or remaining tasks are blocked
  12. •Prints execution summary with stats

Task Status Flow

code
pending → in-progress → completed
              ↑
              └── reviewer returns needs-work → worker iterates

On crash recovery: in-progress → pending (automatic on next run)

Logging

All output is logged to .claude/workitems/{type}/{name}/logs/execution-{timestamp}.log. Each run creates a new log file. Output goes to both stdout and the log file.

Execution Summary

After all tasks are processed, the executor prints:

  • •Tasks completed, failed, skipped (blocked)
  • •Remaining pending tasks
  • •Total duration
  • •Path to the log file

Parallel Execution

Multiple executor instances can run simultaneously. Each instance locks a task by marking it in-progress before starting. Other instances skip in-progress tasks.

Cross-Platform

The script runs on both macOS and Linux. It uses a sedi wrapper for cross-platform sed -i and avoids grep -P (PCRE) in favor of grep -oE (POSIX ERE).