Implement Plan

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-subagents policy; 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 / decomplex recs never auto-create work — disposition first.
• Material doubt, scope/risk choices, or stuck review loops → ask user.
• Never fake Complete.

Startup (blocking)

1. Read the full plan (offset if truncated).
2. Read assets/progress-tracker-template.md with the file tool (memory doesn’t count).
3. Leverage subagents — built-in, extensions/plugins, or via skills.
4. Copy/adapt tracker into the plan or a standalone file.
5. Map every actionable plan requirement → tracker rows (split compound work into tasks/subtasks).
6. Record Template loaded from: implement-plan/assets/progress-tracker-template.md.
7. Only then implement.

On resume: repeat startup and reconcile. Missing template → stop.

Task loop

For each dependency-ready task/subtask (delegate by default):

1. Analyze — read starts-at paths, callers, tests; research only if needed (web-research when third-party/current behavior is uncertain).
2. Implement — smallest change that satisfies the row; stay in ownership; record deviations.
3. Check — targeted tests + applicable lint/typecheck/build/migration/browser (agent-browser for UI).
4. Review — independent code-review when the task crosses a real boundary, or when the batch/phase completes; always for final full-plan review.
5. Disposition findings:
   • Fix now · Validate · Reject · Ask user · Block
   • Fix accepted items (delegate); rerun checks; focused re-review until Clear
   • Complexity-increasing fix → decomplex triage if available, else built-in gate; doubt → ask user
   • Two failed rounds / recurrence / no progress → ask user
6. Update tracker — Verified needs evidence; Descoped needs user approval.
7. Cleanup lane resources per use-subagents (and launcher runtime cleanup, e.g. Pi clean, if used) — no unaccounted workflow-owned resources.
8. 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

1. Reread full plan — no missed requirements.
2. No row Pending / In progress / Blocked.
3. Diff hygiene — drop unjustified scope/complexity.
4. Final checks + final plan-backed implementation review (+ decomplex Audit of this diff if proportionate).
5. When repository files changed, use create-changes-report after those checks to generate/QA a candidate (creator-QA-complete exact bytes). Material issue in evidence gathering → disposition, checks, and implementation review before regenerating.
6. 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-report references/artifact-review.md. No edits or recursive delegation. Require candidate Clear before 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.
7. 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.
8. 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 · retained resources + why