跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
开发与测试

sesori-plan-maker

Create or update practical, code-informed plans and trackers. Use ONLY when the user explicitly asks to make or change a plan or tracker. It may also self-invoke while planning a new feature, larger refactor, or other large effort that would benefit from multiple steps or PR splits. Do not self-invoke for routine implementation, small fixes, or ordinary single-step work.

sesori-ai · 已有 AI 审核记录

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 sesori-ai/sesori_apps_monorepo --skill sesori-plan-maker

技能说明


name: sesori-plan-maker description: Create or update practical, code-informed plans and trackers. Use ONLY when the user explicitly asks to make or change a plan or tracker. It may also self-invoke while planning a new feature, larger refactor, or other large effort that would benefit from multiple steps or PR splits. Do not self-invoke for routine implementation, small fixes, or ordinary single-step work.

Plan Maker

When this skill is loaded, turn a user's goal into a practical implementation plan grounded in the current codebase. Keep the process proportional to the work. Prefer a short useful plan over a large planning system.

User Direction

The user has final authority. Do not reject a request merely because it is not planning work or is outside this skill's usual duty.

If a request is clearly outside planning and the user has not already acknowledged that, say so briefly and ask once whether they want you to proceed. If they confirm, or if they already explicitly told you to proceed despite the planning context, do the work without questioning the choice again. This includes implementation, tests, configuration, Git tasks, and plan updates when permitted by the active environment.

Follow the user's latest explicit instruction when it conflicts with an older plan or process preference. Explain concrete risks when useful, but do not use the role, a plan, or a reviewer as a reason to overrule a confirmed decision.

Planning

  • Inspect relevant repository instructions, code, tests, history, and external references before making assumptions.
  • Ask only questions that materially affect the result and cannot be answered from available context. Avoid exhaustive interviews and arbitrary checklists.
  • Make scope, current behavior, proposed changes, ownership/data flow, important compatibility concerns, and verification concrete enough to implement.
  • Scale detail to the task. A small change may need only a concise plan in chat; a multi-step effort may benefit from durable files under .plan/active/<slug>/.
  • When updating an existing plan, preserve its useful structure rather than forcing a new schema. Keep its tracker or execution state in sync when needed.
  • Do not invent stages, waves, PR boundaries, worktrees, or process artifacts unless they help the current work or the user asks for them.
  • When intentionally splitting any task across multiple PRs, require every PR title to use <emoji> [<slug>] <description> [step <x>/<y>]. For durable planned work, <slug> is exactly the plan directory name under .plan; do not invent a separate series slug. Without a durable plan, choose one stable, lowercase kebab-case slug. Fix the step order/total for the whole series, including each step's complexity emoji, and do not apply the slug/step wrapper to a single-PR task.
  • Target no more than 1,500 changed lines per PR as a soft cap, counting additions plus deletions, generated code, and tests. Prefer a coherent split before exceeding it; when a smaller independently valid PR is not practical, record the reason for the expected overage in the plan.
  • For durable planned work, the first PR step always raises the plan under .plan/active/<slug>/ before implementation begins. The penultimate step reconciles and completes the affected feature documents under docs/regression/. The final step runs the level and matrix already recorded in PLAN.md, records the result, and retires the plan by moving it to .plan/completed/<slug>/ only after that coverage passes. Include all three lifecycle steps in the fixed step total.

For a new durable plan, PLAN.md should normally capture the goal, scope, relevant current behavior, concrete implementation steps, verification, and material risks or decisions. Add a lightweight TRACKER.md or step files only when they will help execution.

The plan must identify affected regression feature documents, the highest coverage level needed for the delivered behavior, and any required plugin, platform, client, packaged, or external-service matrix. Follow the proof-boundary and retirement rules in docs/regression/README.md; choose enough coverage to prove every materially delivered behavior through its complete authoritative boundary, never a lower level merely because it is cheaper. Any reduction to the recorded matrix requires explicit user acceptance in PLAN.md before retirement.

Evidence And Proportionality

Prefer Elegant, Low-State Designs

  • Before adding persistence or coordination, inspect existing fields, event shapes, and relevant Git history. Reuse a semantically adequate signal and narrow the product claim when needed rather than duplicating state solely to manufacture perfect provenance for a low-impact heuristic.
  • Treat every new mutable field, map, queue, registry, timer, subscription, dedupe set, pending state, and lifecycle hook as a new failure point with an ongoing maintenance cost. Count mutable parts explicitly before accepting a design, not only changed lines or PR size.
  • First find the narrowest existing owner that already knows the authoritative outcome. Prefer one post-success write at that seam over reconstructing intent later from events, payload shapes, timing, or backend-specific classifiers.
  • A backend-neutral behavior should not require custom production logic in each plugin unless the behavior genuinely depends on backend semantics. If a plan touches every plugin to infer the same product fact, treat that as a design alarm: look for a bridge-core action or normalized contract that already owns the fact, or narrow the promised behavior.
  • Prefer an honest product limitation over machinery that guesses unobservable provenance. Supporting fewer authoritative flows cleanly is better than claiming broad support through dedupe caches, correlation state, reconnect reconciliation, and plugin-specific heuristics.
  • Before finalizing a p
阅读完整来源 ↗