implement-plan

name	implement-plan
description	Use when executing an approved plan or a clear single-file edit — TDD for risky paths, surgical edits, bounded delegation, worktrees only when real filesystem isolation is needed. Phase = Build.

Implement Plan

Build-phase entry skill. Execute the approved plan with discipline: TDD where it matters, surgical edits, fresh-context reviewer pattern for delegated work, worktrees only when parallel filesystem isolation is real.

Iron Rule

1. NEVER edit code without reading the plan and the touched files first. 2. NEVER expand scope mid-implementation. New idea → write it down and finish the planned task first. 3. ALWAYS write the failing test first for bug fixes and high-risk-surface work. 4. NEVER delegate to a subagent without a written task scope and a clear done criterion. 5. CONTINUOUS execution between tasks on a multi-task plan — no "should I continue?" check-ins, no progress summaries. Stop only on BLOCKED that the Lead cannot resolve, a spec/plan gap, or scope ambiguity.

When to use

• A plan (formal or informal) exists and the next step is editing
• A small one-file edit is clear from the conversation
• Bounded delegation to a specialist is needed
• Two engineers (or sessions) need filesystem isolation

Skip when:

• The task is a question only
• The plan is still vague — go back to write-plan first

Boundary

Owns:

• Executing the approved plan.
• Reading touched files, making surgical edits, writing/running task-level tests, bounded delegation.

Does not own:

• Changing product scope.
• Redesigning the plan silently.
• Declaring final done.
• Merging or branch fate.

Return / hand off:

• Plan vague / wrong / missing file → return to write-plan.
• Root cause unknown → debug-issue.
• Edits complete → check-work.
• Stuck — BLOCKED survives context / model / scope changes and a re-plan → manage-context (escalate mode).

Inputs to gather

• The plan or task list (file paths, ordered tasks, tests)
• The touched files (read them before editing)
• Style conventions from 2-3 nearby files
• The done criterion for this task

Workflow

1. Read first

Read the touched files end-to-end (or the relevant region with line numbers). Don't pattern-match — verify the symbols and behavior exist where the plan expects.

The plan is the loop contract. Verify each task by running its Command verbatim — never a re-derived guess. When it passes, flip EVERY - [ ] under that task to - [x] in the plan artifact (tasks carry one checkbox per step; all flip together — that is how progress survives compaction). On a failing Command, follow the task's On fail, or the plan's Failure policy when the task has none.

2. TDD-light for risky paths

Bug / new logic / billing / migration / auth / race / security:

• Write the failing test that captures the desired behavior
• Run it (must fail)
• Write the smallest code change that makes it pass
• Run all tests (must stay green)

Pure rename / typo / comment fix: tests-after or skip per the test gate. For the full task-type → discipline matrix, see references/tdd-by-risk.md.

3. Surgical edit + quality reflexes

Touch only what the task requires. No "while I'm here" refactors. No reformatting. No new abstraction for single-use.

While editing, hold these reflexes:

• Comments — default to none. Add one only when the WHY is non-obvious (a hidden constraint, a workaround, a surprising invariant). Never comment WHAT the code does.
• One source of truth — before adding a helper, constant, type, or validation, search for an existing one (rg) and extend it instead of duplicating.
• New dependency — justify it: not already covered by stdlib or an existing dep, maintained, reasonable size, compatible license. Unsure → ask.
• Tests — never mock the database in an integration test; prefer a real dependency over a fake / stub / mock.
• Code intel — if code-intel index available, use it for blast radius before wide edits; otherwise rg + Read.

4. Bounded delegation

First decide whether to delegate. The plan's Owner: line wins when set — Owner: Lead → self-do, a named agent → delegate to it. Only a task with no Owner runs the Q1-Q4 test:

Q1: More than 1 file to edit?        Q2: Run tests / build / server?
Q3: A real design-judgment call?     Q4: More than 3 tool calls total?

All "no" → self-do. Any "yes" → delegate to the closest specialist by path / concern / strategy.

When delegating, fill templates/task-brief.md — it scopes the task to 1-2 files, names allowed / forbidden paths, the test command, the done criteria, and the tool cap. Two rules are absolute: the subagent NEVER commits (it returns a manifest, the Lead commits), and it NEVER expands scope.

Pass the full task text + scene-setting context inline in the brief. Do not point the subagent at the plan file path — controller curates exactly the slice the subagent needs.

Implementer self-review before manifest. The subagent scans its own diff for placeholders, missing tests, and plan coverage before returning the manifest. Self-review is not a substitute for the §6 review pipeline; it is a cheap pre-filter.

Implementer return status. Manifest declares one of: COMPLETED with no concerns → §6 pipeline; COMPLETED with Concerns listed → address scope/correctness first, then §6; PARTIAL → review the done slice, redispatch the remainder narrowed; BLOCKED → change a variable (context / model / scope / escalate), never retry blind. Deep handling in references/subagent-dispatch.md.

5. Worktrees only for real parallel

Use a git worktree when two sessions actually need the same files at the same time. A branch is enough for sequential work.

6. Per-task review pipeline

After a subagent returns COMPLETED, dispatch two reviewers in order on the diff alone — no context from the implementer's session:

1. Spec compliance — does the diff match the task spec exactly? No missing requirements, no extras. Issue → implementer fixes → re-review. Approve before stage 2.
2. Code quality — patterns, DRY, smell, test strength. Issue → implementer fixes → re-review.

Both stages mandatory for delegated work. Lead-executed tasks: §6 collapses to one self-review read cold — run git diff, then re-read each touched file from disk (not from memory) before judging (fresh-context simulation).

After all tasks pass per-task review, dispatch a final whole-implementation review on the cumulative diff to catch cross-task drift (type/symbol/contract mismatch, unowned files in parallel layouts). Hand off to check-work only after the final review clears.

7. Model selection by task complexity

Use the least powerful model that can handle each role — cost compounds across N tasks × M reviews. The task-type → tier table lives in references/subagent-dispatch.md.

If a matching child plugin skill is available

Prefer sibling edit primitives over hand-rolled writes when the domain matches (Extension Protocol v1 — spec in the rolepod source repo):

• rolepod-uiproof /scaffold-e2e — e2e test scaffold from scenario + replay (playwright / vitest+playwright / pytest+selenium)
• rolepod-wplab /wp-edit-{design,plugin,theme}, /wp-scaffold — WP edit primitives + boilerplate inside wp-content/

Evidence auto-routes to <git-root>/.rolepod/evidence/ under parent; standalone path otherwise. check-work aggregates.

If a matching Rolepod agent is available

Delegate the bounded task to the closest specialist:

• frontend-developer / ui-ux-designer — UI / interaction
• backend-developer — API / business logic / DB models
• mobile-developer — iOS / Android / RN / Flutter
• billing-engineer — billing / credits / subscription
• ai-ml-engineer — LLM / RAG / Anthropic SDK / prompt cache
• data-scientist — analytics / pipelines / dashboards
• content-strategist — written output; pass audience: dev|user|prospect

Brief: spec + plan + files + tests + done criterion + handoff partner.

If no matching agent is available

Execute as Lead with this minimum viable checklist:

1. Read the plan + touched files; write failing test first for risky paths
2. Make the smallest change that satisfies the test; run the full test suite (or the touched module's)
3. Match local style — naming, error handling, imports; do not invent new patterns unless the plan requires it
4. Flag dead code adjacent to your change; do not delete without asking
5. Verify before claiming done — see check-work

Output

The implementation manifest is the canonical artifact: templates/implementation-manifest.md. It carries files changed, tests, commands, evidence, a scope check, and the status. A subagent returns this; the Lead commits. Do not restate the manifest shape here; the template is the single source.

Examples

Non-blocking — read only when unsure about scope or whether to trust a subagent:

• examples/execution-examples.md — a surgical-vs-scope-creep execution and an accept-vs-reject subagent manifest, each a good/bad pair with a "why good wins" table. Read the whole file; the contrast is the lesson.

References

Load only when the task needs it:

• references/tdd-by-risk.md — task type → test discipline: test-first vs evidence-after
• references/subagent-dispatch.md — implementer status taxonomy (COMPLETED ± concerns / PARTIAL / BLOCKED) and handling, two-stage review prompt scaffolds (spec compliance / code quality), model selection table, continuous-execution rationale

Hard stops

• A planned file does not exist where expected → stop, verify or re-plan
• A test you wrote passes before you added the code → assertion is too weak; tighten
• Subagent returns COMPLETED with failing tests, or BLOCKED re-dispatched unchanged → reject / change a variable (context / model / scope); never accept unchanged retry
• A subagent's diff accepted without spec-compliance + code-quality reviews → stop, run the §6 pipeline before building further
• Scope creep beyond the task list → stop, write a follow-up, finish current task
• About to write "should I continue?" between tasks on a multi-task plan → don't; Iron Rule 5
• About to verify a task with a self-invented check while the plan names a Command → run the plan's Command

Full Rolepod enhancement

Adds cohesion contracts for parallel work, model-tier-aware agent routing, hooks that block subagent commits, the qa-tester floor, and the two-stage fresh-context review pattern.

Next phase

• If check-work is available, continue there to prove the change works.
• If check-work is not available, run tests / build / curl / browser yourself and report evidence inline.