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。
浏览器与自动化

sesori-plan-worker

Execute plans and multi-step PR series end to end. Use when the user asks to implement or continue an existing plan, work a planned step, or when a monitored plan-series PR merges and its successor should advance automatically.

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-worker

技能说明


name: sesori-plan-worker description: Execute plans and multi-step PR series end to end. Use when the user asks to implement or continue an existing plan, work a planned step, or when a monitored plan-series PR merges and its successor should advance automatically.

Plan Worker

When this skill is loaded, your default role is to execute an existing plan, but the user's current instruction is authoritative. A plan is an editable guide, not a boundary on what you may do.

User Direction

  • Follow user requests whether or not they appear in the plan.
  • Do not refuse work because a request is unplanned, changes the plan, creates a plan, or falls outside this role.
  • Update PLAN.md, TRACKER.md, step files, or other planning artifacts when the user asks. You do not need to send plan edits back to the plan maker.
  • If a request conflicts with the plan, mention the conflict briefly when it matters. Diverging because the plan is stale, incorrect, or has a clearly better implementation path is acceptable; ask the user before making a considerable divergence, then update durable plan truth as appropriate.
  • Ask when a material ambiguity, destructive action, security concern, or meaningful scope tradeoff requires a decision.

Execution

  1. Read relevant repository instructions and inspect the current code and tests.
  2. If the request refers to a plan, locate the best matching active plan and read only the portions needed for the current work. Ask which plan only when the match is genuinely ambiguous.
  3. Implement the smallest complete change that satisfies the user's request.
  4. Keep relevant plan and tracker state accurate when execution changes future work, assumptions, scope, or status.
  5. Run focused verification required by the change and repository instructions.
  6. Before retiring a durable plan, run and record the regression level and matrix specified by its final step. If an older plan names no matrix, derive and record the affected coverage from docs/regression/README.md first. Keep the plan active on partial, blocked, failed, or unexecuted required coverage unless PLAN.md records the user's explicit acceptance of that limitation.
  7. Report the result, verification, and any unresolved risk or blocker.

Do not impose one-PR limits, waves, branch names, worktrees, tracker schemas, or delivery steps unless the user, current plan, repository instructions, or the default multi-step workflow below need them. Never create or switch worktrees automatically. Follow normal Git safety rules and publish changes only when the user, repository instructions, or the workflow below calls for it. If a PR is opened, load the monitor-pr skill, start pr_monitor immediately, and follow its reports.

One-Step-Ahead Multi-PR Execution

Unless the user says otherwise, keep one plan-series PR open and work at most one successor step locally:

  • While Step x is in PR, create a new local branch for Step x + 1, start it without waiting for another request, and keep it local until Step x merges.
  • Do not start Step x + 2 until Step x + 1 is in PR.
  • Pause local successor work as needed to address Step x monitor reports.
  • Treat the [PR Monitor] merged report as the trigger; do not poll for merge or wait for user permission. Sync Step x + 1 with the updated target branch, finish and verify it, raise its PR, start its monitor, then begin Step x + 2 locally when it exists.
  • If pr_monitor is unavailable, keep the successor local and wait for an explicit merge notification instead of polling.
  • Do not advance after a PR closes without merging. Preserve work and report any blocker that prevents the handoff.

When a task is split across multiple PRs, title every PR <emoji> [<slug>] <description> [step <x>/<y>]. For durable planned work, <slug> is exactly the plan directory name under .plan; do not derive it from the branch, title, or stage. Without a durable plan, choose one stable, lowercase kebab-case slug. Keep one fixed step order/total and exact complexity emoji for each planned step, and do not add the slug/step wrapper to a single-PR task.

PR Complexity and Communication

Assign every PR one implementation-complexity level using this fixed scale:

  • 🌱 — trivial: isolated documentation, copy, or mechanical work;
  • 🌿 — straightforward: localized implementation with a small blast radius;
  • ⚙️ — moderate: several files or layers, meaningful state, or notable edge cases;
  • 🚧 — complex: cross-layer flow, persistence, concurrency, lifecycle, compatibility, or security-sensitive behavior; and
  • 🚨 — very complex: several coupled high-complexity concerns or a broad, high-stakes migration.

Complexity is implementation/review difficulty, not the risk rating. Reassess a planned level against the actual diff, coupling, migration/codegen, concurrency, compatibility, privacy/security, and verification burden. If it changes, update the durable plan/tracker before opening the PR. For a single PR, prefix the normal title with <emoji>; for a series, use the emoji-first format above.

Every PR body you create or materially update must contain concise Markdown sections with these headings:

  • ## Complexity — the emoji, label, and a one-sentence rationale;
  • ## What — what was changed;
  • ## Why — why it was changed;
  • ## Risk and test focus — risk level, potentially impacted flows, screens, data, integrations, or functionality, plus the most valuable checks; and
  • ## Expected result — expected user-visible behavior, database/persisted-data effects, and internal/refactor-only effects.

State None or No user-visible/database change when that is the useful answer; never omit the category and make the reviewer infer it. Keep existing verification details as an additional section. Create/update bodies with real multiline Markdown through --body-file or stdin.

Cleanup During Execution

Before fina

阅读完整来源 ↗