implement-plan
Implements existing Markdown plans exhaustively through delegated task loops, validation, and review closure. Use this skill when asked to implement, execute, carry out, or continue an existing plan. Do not use for creating a plan from scratch or only reviewing a plan without implementing it.
maxedapps · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add maxedapps/agent-skills --skill implement-plan技能说明
name: implement-plan description: >- Implements existing Markdown plans exhaustively through delegated task loops, validation, and review closure. Use this skill when asked to implement, execute, carry out, or continue an existing plan. Do not use for creating a plan from scratch or only reviewing a plan without implementing it. license: MIT compatibility: >- Requires project write access. Delegation requires a safely available subagent capability. Browser-visible work requires browser automation. metadata: short-description: Implement plans via delegated task loops and review
Implement Plan
Hard rules
- Stay on the plan’s required outcome. Discovery/review do not add scope.
- Prefer the smallest behavior-preserving change. Skip unnecessary abstractions/tests/safeguards.
- Delegate by default when safe. “Small/easy” is not a reason to skip subagents.
- Leverage subagents — built-in, extensions/plugins, or skills. Follow
use-subagentspolicy; use native subagents, extensions / plugins, or other skills - Main agent owns: tracker, integration, dispositions, acceptance, user comms, cleanup.
- Child handoffs = evidence, never acceptance. Inspect diffs; rerun checks.
- Findings /
decomplexrecs never auto-create work — disposition first. - Material doubt, scope/risk choices, or stuck review loops → ask user.
- Never fake
Complete.
Startup (blocking)
- Read the full plan (offset if truncated).
- Read
assets/progress-tracker-template.mdwith the file tool (memory doesn’t count). - Leverage subagents — built-in, extensions/plugins, or via skills.
- Copy/adapt tracker into the plan or a standalone file.
- Map every actionable plan requirement → tracker rows (split compound work into tasks/subtasks).
- Record
Template loaded from: implement-plan/assets/progress-tracker-template.md. - Only then implement.
On resume: repeat startup and reconcile. Missing template → stop.
Task loop
For each dependency-ready task/subtask (delegate by default):
- Analyze — read starts-at paths, callers, tests; research only if needed (
web-researchwhen third-party/current behavior is uncertain). - Implement — smallest change that satisfies the row; stay in ownership; record deviations.
- Check — targeted tests + applicable lint/typecheck/build/migration/browser (
agent-browserfor UI). - Review — independent
code-reviewwhen the task crosses a real boundary, or when the batch/phase completes; always for final full-plan review. - Disposition findings:
Fix now·Validate·Reject·Ask user·Block- Fix accepted items (delegate); rerun checks; focused re-review until
Clear - Complexity-increasing fix →
decomplextriage if available, else built-in gate; doubt → ask user - Two failed rounds / recurrence / no progress → ask user
- Update tracker —
Verifiedneeds evidence;Descopedneeds user approval. - Cleanup lane resources per
use-subagents(and launcher runtime cleanup, e.g. Piclean, if used) — no unaccounted workflow-owned resources. - Next ready task/subtask.
Parallelize only independent tasks with isolated writers. Parent is sole tracker writer under concurrency.
When to review
- After major boundaries: integration, migration, public contract, security/data invariant, risky dep, delivery milestone
- Plan-authored checkpoints
- Final full-plan review (plan-backed)
Fresh reviewers; don’t share conclusions pre-handoff. One reviewer default. Parent review only if independent review is unavailable — record the limit.
Reviewer states: Clear · Changes required · Human decision required · Blocked
Done
- Reread full plan — no missed requirements.
- No row
Pending/In progress/Blocked. - Diff hygiene — drop unjustified scope/complexity.
- Final checks + final plan-backed implementation review (+
decomplexAudit of this diff if proportionate). - When repository files changed, use
create-changes-reportafter those checks to generate/QA a candidate (creator-QA-complete exact bytes). Material issue in evidence gathering → disposition, checks, and implementation review before regenerating. - After candidate QA, launch one fresh read-only subagent that did not implement, perform final implementation review, or author the report. Give frozen scope/evidence, the candidate, and
create-changes-reportreferences/artifact-review.md. No edits or recursive delegation. Require candidateClearbefore completion. Keep report-review closure in the tracker/final handoff, never in the reviewed HTML.- Report-only → regenerate, re-QA, re-review with a fresh reviewer.
- Underlying implementation → reopen affected rows, checks, final implementation review, scope freeze, report generation, and candidate review.
- Two failed rounds / recurrence / no progress → existing user-escalation rule.
- No safe subagent → ask if the user accepts a disclosed parent fallback; without approval remain
Partial/Blocked. Never claim independence for fallback review.
- Final cleanup — worktrees, branches, processes, runtime state, including anything created by a report-triggered fix cycle. Make the final changes-report path the primary handoff. If the skill is unavailable, record that and continue with the concise report below.
- Report truthfully.
Complete only when all rows are Verified or approved Descoped, validation passed, final implementation review is Clear, the changes-report candidate is Clear when a report is produced, and nothing material remains open. Else Partial or Blocked.
Report
- changes-report path, or why no report was created
- plan/tracker paths · status · remaining IDs
- what was delegated vs parent-owned (and why)
- checks run / skipped
- implementation-review and candidate-review outcomes + dispositions
- decisions, deviations
- worktrees created/integrated/removed ·