AgentSkillsCN

planner-exec

执行单个计划文件，进行全面依赖验证、配置选项（自动提交、更新README、更新CLAUDE.md）、冲突检测和PROGRESS.md跟踪。如果依赖未满足则阻止执行。适用于用户想要运行特定计划、执行计划文件或实施单个计划时。

SKILL.md
--- frontmatter
name: planner-exec
description: Execute a single plan file with full dependency validation, configuration options (auto-commit, README update, CLAUDE.md update), conflict detection, and PROGRESS.md tracking. Blocks execution if dependencies are not met. Use when user wants to run a specific plan, execute a plan file, or implement a single plan.
allowed-tools: Task, TaskOutput, Read, Glob, AskUserQuestion
user-invocable: true
context: fork

Execute Single Plan

You will now execute the single plan workflow. Do NOT just report what you'll do - actually execute each step using the tools available.

EXECUTE NOW: Step 1 - Find Plan File

Use the Glob tool NOW to find the plan file. Parse user's arguments to identify the plan name.

  • •If partial name (e.g., "01"): Glob("plans/*01*.plan.md")
  • •If full name: Glob("plans/[name]")
  • •If multiple matches: Use AskUserQuestion to clarify

EXECUTE NOW: Step 2 - Read Plan

Use the Read tool NOW to read the plan file found in Step 1.

Extract:

  • •The full plan content
  • •The depends_on field from # Configuration section (if present)

EXECUTE NOW: Step 3 - Check Dependencies

If depends_on exists in the plan:

  1. •Use Read tool to read plans/PROGRESS.md
  2. •Check if each dependency has status COMPLETED
  3. •If any dependency is NOT completed:
    • •Report: "⛔ BLOCKED: Dependency [name] is not COMPLETED"
    • •STOP execution immediately

EXECUTE NOW: Step 4 - Read Config

Use Read tool NOW to read plans/planner.config.json.

Extract these values (use defaults if not found):

  • •auto_commit (default: false)
  • •auto_commit_standard (default: "no_standard")
  • •auto_update_claude_md (default: false)
  • •replan_on_exec (default: false)

EXECUTE NOW: Step 5 - Spawn Agent

You MUST immediately call the Task tool with these exact parameters:

code
description: "Execute plan: [plan_name]"
subagent_type: "planner:plan-executor"
prompt: |
  plan_name: "[actual plan filename]"

  plan_content: |
    [FULL PLAN FILE CONTENT from Step 2]

  skip_dependency_check: true

  config:
    auto_commit: [value from Step 4]
    auto_commit_standard: [value from Step 4]
    auto_update_claude_md: [value from Step 4]
    replan_on_exec: [value from Step 4]

  additional_instructions: [user's extra instructions or "None"]

  BEGIN EXECUTION.

Call the Task tool NOW. Do not delay. Do not just describe what you would do.

Step 6: Report Results

After the agent completes, report to the user:

code
════════════════════════════════════════
## Execution Result

**Plan**: [plan_name]
**Status**: SUCCESS | FAILURE | BLOCKED
**Summary**: [Brief description of what was done]
**Files Modified**: [List of changed files]
**Config Actions**: [Actions taken: committed, updated CLAUDE.md, etc.]
**Issues**: [Any problems encountered, or "None"]
**Next**: [Suggested next plan from PROGRESS.md, or "All plans complete"]
════════════════════════════════════════

Reference Information

What This Skill Does

When executing a single plan, this skill:

  1. •Verifies Dependencies: Checks that all required plans are COMPLETED
  2. •Reads Configuration: Gets settings from plans/planner.config.json
  3. •Checks for Conflicts: Warns if other plans are IN PROGRESS
  4. •Executes the Plan: Follows plan steps with fresh context
  5. •Updates Tracking: Marks plan as IN PROGRESS → COMPLETED/FAILED
  6. •Applies Configuration: Handles post-execution tasks based on config

Safety Features

  • •Dependency Blocking: Refuses to execute if dependencies aren't met
  • •Conflict Detection: Warns about other plans currently running
  • •Resume Awareness: Detects if a plan was already completed
  • •Structured Results: Consistent SUCCESS/FAILURE/BLOCKED reporting

Configuration Options

Configuration is read from plans/planner.config.json (set during planner-setup):

OptionDescription
Auto-commitCreate a git commit after successful execution
Auto-commit standardCommit message format: "conventional_commits" or "no_standard"
Auto-update CLAUDE.mdAnalyze and update project CLAUDE.md if changes made it inaccurate
Re-plan on ExecutingRe-analyze and draft fresh implementation before executing (slower)

These options are only applied if execution succeeds. Failed plans skip configuration actions.

Configuration Actions (Applied After Success Only)

Actions are applied in this order:

  1. •

    Auto-update CLAUDE.md (if enabled):

    • •Read current CLAUDE.md
    • •Analyze if changes made content inaccurate
    • •Update if needed
  2. •

    Auto-commit (if enabled):

    • •Uses the planner-commit skill (@skills/planner-commit/SKILL.md)
    • •If auto_commit_standard is "conventional_commits": Uses Conventional Commits format
    • •If auto_commit_standard is "no_standard": Uses simple format feat(planner): Complete [plan_name] - [summary]
    • •DO NOT push (user does that)

PROGRESS.md Updates

The plan-executor agent manages PROGRESS.md automatically:

Before execution:

markdown
| auth-02-api.md | NOT STARTED | |

During execution:

markdown
| auth-02-api.md | IN PROGRESS | |

After success:

markdown
| auth-02-api.md | COMPLETED | 2026-01-09 |

After failure:

markdown
| auth-02-api.md | FAILED | |

Status Values

  • •NOT STARTED: Plan hasn't been executed yet
  • •IN PROGRESS: Plan is currently executing (shows in progress, warns of conflicts)
  • •COMPLETED: Plan finished successfully (dependencies satisfied for dependent plans)
  • •FAILED: Plan encountered an error (blocks dependent plans)