AgentSkillsCN

mpx-execute-task

执行下一项任务。系统会提示您选择阶段,并将任务委派给执行代理。

SKILL.md
--- frontmatter
name: mpx-execute-task
description: Execute next task. Prompts for phase selection, delegates to executor agent.
disable-model-invocation: true
allowed-tools: Read, Write, Task, Bash, AskUserQuestion

Execute Next Task

Execute a task by selecting a phase and delegating to the executor agent.

Usage

code
/mpx-execute-task           # Select phase, execute next task

Workflow

Step 1: Find Available Phases

  1. Read .mpx/STATE.md for overall status
  2. Read .mpx/ROADMAP.md for phase dependencies
  3. Find all phase folders in .mpx/phases/
  4. For each phase, check:
    • Status from ROADMAP.md (Not Started / In Progress / Completed)
    • Dependencies from ROADMAP.md
    • Whether dependencies are satisfied (all dependency phases = Completed)

Step 2: Prompt for Phase Selection

Use AskUserQuestion to let user choose which phase to work on.

Rules:

  • Only offer phases whose dependencies are satisfied
  • First option = first incomplete phase (Recommended)
  • Next 3 options = next eligible phases
  • "Other" allows manual phase number entry

Example prompt:

code
Question: "Which phase to execute?"
Header: "Phase"
Options:
  1. "Phase 1: Foundation (Recommended)" - "First incomplete phase, 3/10 tasks done"
  2. "Phase 2: Core API" - "Dependencies met, 0/8 tasks done"
  3. "Phase 3: UI Components" - "Dependencies met, 0/12 tasks done"
  4. "Phase 4: Integration" - "Blocked by Phase 2, 3"

If only one phase available: Skip prompt, proceed with that phase.

If user selects "Other": Parse phase number from input.

Step 3: Find Next Task in Selected Phase

  1. Read selected phase's CHECKLIST.md
  2. Find first unchecked task (- [ ])
  3. Read phase's SPEC.md for context

Step 4: Prepare Context for Agent

Gather essential context to pass to the executor agent:

  • Project name and tech stack (from SPEC.md)
  • Current phase name and objectives (from phase's SPEC.md)
  • Task description
  • Any relevant decisions (from global or phase STATE.md)

Step 5: Delegate to Executor Agent

Use the Task tool to spawn a subagent:

code
Task tool:
  subagent_type: "mpx-executor"
  model: opus
  description: "Execute task: [task_name]"
  prompt: |
    You are an executor agent with fresh context.

    ## Project Context
    [Include relevant SPEC.md content]

    ## Your Mission
    Execute this task from Phase N: [Task Description]

    ## Task Details
    [Task description and any relevant context]

    ## Instructions
    1. Implement the task
    2. Commit after completing (use descriptive message)
       - Format: "phase-N: description"
    3. Mark task complete: change `- [ ]` to `- [x]` in phase CHECKLIST.md
    4. If you encounter blockers, document them and stop
    5. Report summary when done

    ## Working Directory
    [Current working directory]

    ## Important
    - Do NOT modify files outside the project scope
    - Do NOT skip the task without documenting why
    - Do NOT make architectural changes not in the spec

Step 6: Handle Agent Response

When the agent completes (or stops due to blocker):

On Success:

  1. Task is already marked complete by agent in phase's CHECKLIST.md
  2. Check if all phase tasks are complete
  3. If phase complete:
    • Update phase's STATE.md (Status = Completed)
    • Update .mpx/STATE.md (global)
    • Update .mpx/ROADMAP.md (Status column = Completed)
  4. Report completion to user

On Blocker:

  1. Update STATE.md:
    • Add blocker to Blockers section
    • Add session note in Session Handoff section
  2. Report to user what happened and what's needed

Step 7: Report Results

code
Task Completed: [Task Description]
Phase: N - [Phase Name]

Phase Progress: X/Y tasks complete
[If phase complete: "Phase N Complete!"]

Commits Made: N

Next Task:
  [] [Next task description]

Run `/mpx-execute-task` to continue.
Run `/mpx-show-project-status` for full progress overview.

Error Handling

  • No phases found: "No project found. Run /mpx-init-project or /mpx-parse-spec first."
  • All phases complete: "All phases complete! Project finished."
  • No eligible phases: "All remaining phases are blocked. Check dependencies in ROADMAP.md."
  • Agent fails: Report error and suggest manual intervention

Parallel Execution Note

Multiple /mpx-execute-task commands can run in parallel on different phases if:

  • The phases have no mutual dependencies
  • Both phases' dependencies are already satisfied

Example: Phase 2 and Phase 3 can both run if they only depend on Phase 1 (completed).

Notes

  • The subagent gets fresh 200k context, preventing degradation
  • Each task should be atomic and completable in one execution
  • STATE.md provides continuity between tasks and sessions
  • Agent commits after each task to preserve progress
  • ROADMAP.md is the source of truth for phase completion status