AgentSkillsCN

execute

在一次会话中完整运行APEX全流程——从调研到规划、再到实施与交付——的统筹协调者。适用于那些希望在不频繁切换上下文的情况下高效完成的任务。

SKILL.md
--- frontmatter
name: execute
description: Orchestrator that runs the full APEX workflow (research → plan → implement → ship) in a single session. Use for tasks you want to complete without context switches.
argument-hint: [task-description|ticket-id|file-path]
<skill name="apex:execute" phase="orchestrator"> <overview> Meta-skill that chains all 4 APEX phases in sequence:
  1. •/apex:research - Intelligence gathering
  2. •/apex:plan - Architecture design
  3. •/apex:implement - Build and validate
  4. •/apex:ship - Review and reflect

Use this for single-session task completion. For multi-session work, invoke individual skills. </overview>

<when-to-use> - Small to medium tasks that fit in one session - When you don't need to pause between phases - When you want full workflow without manual skill invocation </when-to-use> <when-not-to-use> - Large complex tasks that need human review between phases - When you want to pause after research to think - When context overflow is likely - When you want to run phases in separate sessions </when-not-to-use> <initial-response> <if-no-arguments> I'll run the full APEX workflow. Please provide: - Task description (e.g., "implement dark mode toggle") - Ticket ID (e.g., "APE-59") - Path to task file

Example: /apex:execute "add user authentication" </if-no-arguments> <if-arguments>Begin full workflow.</if-arguments> </initial-response>

<workflow> <step id="1" title="Research phase"> <invoke skill="apex:research"> Pass the original input argument. </invoke> <verify> - Task file created at ./apex/tasks/[ID].md - Frontmatter shows phase: research - `<research>` section populated </verify> <extract> Store `identifier` for subsequent skill invocations. </extract> <on-failure> Stop and report: "Research phase failed. See task file for details." </on-failure> </step> <step id="2" title="Plan phase"> <invoke skill="apex:plan"> Pass the identifier from step 1. </invoke> <verify> - Frontmatter shows phase: plan - `<plan>` section populated - 5 mandatory artifacts present </verify> <on-needs-input> Plan phase is interactive. If user input needed: 1. Present the question to user 2. Get response 3. Continue plan phase with response </on-needs-input> <on-failure> Stop and report: "Plan phase failed. Run `/apex:plan [identifier]` to retry." </on-failure> </step> <step id="3" title="Implement phase"> <invoke skill="apex:implement"> Pass the identifier. </invoke> <verify> - Frontmatter shows phase: implement - `<implementation>` section populated - All tests passing </verify> <on-failure> If tests failing after max retries: 1. Report current state 2. Ask user: "Implementation has issues. Continue to ship (will document issues) or stop here?" 3. If stop, leave task in implement phase for manual intervention </on-failure> </step> <step id="4" title="Ship phase"> <invoke skill="apex:ship"> Pass the identifier. </invoke> <verify> - Frontmatter shows phase: complete, status: complete - `<ship>` section populated - Git commit created - Reflection recorded </verify> <on-failure> Stop and report: "Ship phase failed. Run `/apex:ship [identifier]` to retry." </on-failure> </step> <step id="5" title="Final report"> <template> ## APEX Workflow Complete ✅

Task: [Title] File: ./apex/tasks/[identifier].md

Phases Completed:

  1. •✅ Research - [summary]
  2. •✅ Plan - [chosen architecture]
  3. •✅ Implement - [files changed, tests status]
  4. •✅ Ship - [commit SHA, reflection status]

Metrics:

  • •Total patterns applied: [N]
  • •Tests: [passed]/[total]
  • •Review findings: [N] ([dismissed]% false positives)
  • •Commit: [SHA]

Key Learning:

[From reflection]

Task complete. Full history in ./apex/tasks/[identifier].md </template> </step>

</workflow> <error-handling> <phase-failure> If any phase fails: 1. Report which phase failed 2. Report current state of task file 3. Suggest manual intervention: "Run `/apex:[phase] [identifier]` to retry" 4. Do NOT continue to next phase </phase-failure> <user-input-needed> If any phase needs user input: 1. Present the question 2. Wait for response 3. Continue that phase with response 4. Do NOT skip the interaction </user-input-needed> <context-overflow> If context is getting large: 1. Warn user: "Context is large. Consider continuing in new session." 2. Report current phase and identifier 3. User can restart with individual skill from current phase </context-overflow> </error-handling> <success-criteria> - All 4 phases completed successfully - Task file shows phase: complete, status: complete - Git commit exists - Reflection recorded - Final report displayed to user </success-criteria> </skill>