--- 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 finalizing each feature PR, inspect what its implementation makes obsolete: calculations/data generation, model or transport fields, database columns, caches, flags/settings, jobs/watchers/listeners, compatibility paths, UI state, tests, and docs. Re-check the plan's cleanup assessment and add newly discovered causal cleanup to durable plan truth.

Implement small, safe cleanup directly caused by the feature when it keeps the PR coherent. Split or ask first when cleanup is a considerable refactor; defer with an explicit compatibility/migration/risk reason when removal is not yet safe. Do not retain dead artifacts solely as an audit trail when Git history is sufficient, and do not use cleanup as a reason for unrelated scope expansion.

## Plan Review

Use `architecture-plan-review` only for a new architecture-bearing production plan that has not already been reviewed. Ask a sub-agent to perform the review using the skill. Apply valid findings directly without re-reviewing those fixes. A too-vague rejection may be reviewed once more after clarification; if it is rejected as too vague again, ask the user how to proceed. Considerable changes caused by new findings or user requests may also be reviewed again.

## Implementation Review

Use `architecture-implementation-review` only when production changes alter actual architecture: new or moved classes/files, dependency or DI ownership, public or persisted contracts, cross-layer flow, lifecycle ownership, or shared boundaries. Ask a sub-agent to perform the review using the skill. It is not a general implementation-correctness reviewer; do not call it for localized logic changes, bug fixes, tests, formatting, or tooling work.

Prefer a Git-defined scope, normally the current branch against `main`, an explicit commit or commit range, the last N commits, or a PR. File or directory scopes are also acceptable when they are more useful. In that case, make the current change clear and let the reviewer use Git history and diffs to avoid mistaking pre-existing code for new code.

Run up to two implementation-review passes before seeking user guidance:

1. Run one complete review after implementation and focused verification. 2. Fix valid findings that are clearly within the current request. 3. Use a second review only when useful after those fixes.

Avoid a review loop. If the second review still rejects the implementation, ask the user how to proceed before another review. If rejection is based only on a decision the user explicitly approved, that approval supersedes the review; do not re-review or re-litigate it.

Do not let review trigger a broad cleanup. If a finding asks to move, rename, or refactor pre-existing files, classes, or architecture beyond the current request, stop before making that expansion and ask whether the user wants it in scope. Explain the impact and any smaller in-scope alternative. A reviewer does not authorize scope expansion, and a user waiver or decision must not be re-litigated.

## Working Style

Be pragmatic and flexible. Preserve unrelated work, avoid speculative abstractions, keep recovered failures observable, never hand-edit generated files, and finish the requested work end to end whenever feasible. Add tests only when they provide meaningful confidence.

Edge cases are infinite and completeness is not the goal. Before adding a guard, name the concrete flow that reaches the bad state and the damage if it does; if you cannot name a real caller or sequence, leave the case unhandled. "An API technically accepts it", "a misbehaving client might", and "a reviewer raised it" are not flows. Guarding an unreachable state puts new code on the path that runs constantly to defend one that never runs, so the guard becomes a failure point in exchange for nothing.

Keep defensive depth proportional to damage, and enforce an invariant once, at the place that owns it, on the entity the caller named — never widened to parents, children, or related entities in case something reaches them another way. When review pressure keeps pushing a gate outward, that is a signal to stop and ask the user, not to keep widening it.

Apply the cleanup rules above pragmatically and keep unrelated refactors out of the current PR.