plan-execution（计划执行）

Plan Execution Skill

This skill auto-activates when the user references an existing plan or wants to continue working from one.

Trigger phrases:

• "continue the plan"
• "next task"
• "run task 2.1"
• "what's the plan status"
• "mark task X done"
• References to plans/*.plan.md files

For creating new plans: Use /superplan <description> command instead.

Core Behaviors

When User Says "next task" or "continue the plan"

Execute the /plan next flow:

1. Find the active plan (plans/*.plan.md with status in-progress)
2. Parse for the first unchecked [] task whose dependencies are all [x]
3. Show task details and ask for confirmation
4. Implement, verify, mark complete, commit

When User Says "run task X.Y"

1. Find the active plan
2. Locate the specific task by ID (e.g., 2.1)
3. Check dependencies are met
4. If dependencies unmet, warn: "Task 2.1 depends on 1.3 which is incomplete. Proceed anyway? (y/n)"
5. Implement, verify, mark complete, commit

When User Says "what's the plan status"

Execute the /plan status flow — show progress table for all plans.

When User Says "mark task X done"

1. Find the task in the plan file
2. Update [] to [x] with completion date
3. Show updated progress

Plan File Parsing

Reading a Plan

Plans follow this structure:

• # Plan: <title> — main heading
• | Status | <value> | — in metadata table
• ## Milestone N: <name> — milestone headings
• ### N.M [] <title> — pending task
• ### N.M [x] <title> *(completed YYYY-MM-DD)* — completed task

Finding Dependencies

Each task has a **Dependencies:** field:

• None — can run immediately
• 1.1 — depends on task 1.1
• 1.1, 1.2 — depends on both tasks 1.1 and 1.2
• Milestone 1 — depends on all tasks in milestone 1

A dependency is met when the referenced task(s) have [x].

Updating Task Status

When marking a task complete, replace:

### 2.1 [ ] Add API endpoint

with:

### 2.1 [x] Add API endpoint *(completed 2026-03-21)*

Updating Plan Status

• When first task starts: change | Status | draft | to | Status | in-progress |
• When all tasks complete: change to | Status | complete |

Task Execution Flow

When implementing a task from the plan:

1. Read Full Context

• Read the entire plan file (not just the current task)
• Understand what was built before (completed tasks)
• Understand what comes next (future tasks)
• Check architecture decisions section for relevant context

2. Implement

• Follow the task's What specification
• Modify only the Files listed (unless the task clearly requires additional files)
• Follow existing codebase patterns and style guides

3. Verify (Acceptance Gate — mandatory)

Go through each bullet in the task's Acceptance criteria one by one and explicitly confirm it passes:

• Run the exact commands or checks described
• For UI changes: use browser automation, take screenshots
• For backend changes: make API calls, check database

If any criterion fails:

1. Fix the issue and re-verify all criteria (retry 1)
2. If still failing, fix again and re-verify (retry 2)
3. After 2 retries still failing: stop, do NOT mark [x], and report to the user exactly which criterion failed and what you observed. Do not guess further.

Only proceed to the next step when all acceptance criteria pass.

4. Run Code Quality Checks

• Run whatever code quality tooling the project provides (linting, type checking, tests)
• If the project has a dedicated code-checks command or script, use it

5. Update Plan

• Mark task [x] with completion date
• Update plan status if needed
• Add notes to Architecture Decisions if you learned something important

6. Commit

Create a commit with message format:

Plan <N.M>: <Task Title>

Include both the code changes AND the plan file update in the same commit.

7. Report Progress

Show:

• What was completed
• Current progress (X/Y tasks, percentage)
• What's next (next executable task)

Multi-Agent Execution

For tasks that have independent subtasks (e.g., frontend + backend work), use parallel agents:

Task 2.1: Add user settings page
+-- Agent 1 (backend-specialist): Create API endpoint
+-- Agent 2 (frontend-specialist): Create React component
+-- Merge results, verify integration

Only parallelize when the task spec explicitly mentions independent frontend/backend work.

Handling Plan Changes

If during implementation you discover the plan needs changes:

1. Minor adjustment (file paths wrong, small scope change): Update the plan file directly and note it
2. Significant change (new task needed, task should be removed, reordering): Flag to user before changing
3. Architecture change: Update the Architecture Decisions section and discuss with user

Never silently deviate from the plan. If you need to do something different from what's specified, say so.

Integration Points

The plan system works with whatever tools and commands the project provides: