---
name: implement-plan
description: Use when a plan folder (from create-plan) exists and needs to be executed in an isolated git worktree with iterative cross-model milestone review. ALWAYS invoke when user says "implement the plan", "execute the plan", "start implementation", "resume the plan", or similar execution requests.
---

# Implement Plan (Claude Code)

Execute an existing plan (created by `create-plan`) in an isolated git worktree, with iterative cross-model review at each milestone boundary.

## Prerequisite Check (MANDATORY)

Required:
- Plan folder exists under `ai_plan/` at project root
- `continuation-runbook.md` exists in plan folder
- `milestone-plan.md` exists in plan folder
- `story-tracker.md` exists in plan folder
- Git repo with worktree support: `git worktree list`
- Superpowers execution skills:
  - `superpowers:executing-plans`
  - `superpowers:using-git-worktrees`
  - `superpowers:verification-before-completion`
  - `superpowers:finishing-a-development-branch`

If any dependency is missing, stop immediately and return:

"Missing dependency: [specific missing item]. Ensure all prerequisites are met, then retry."

If no plan folder exists:

"No plan found under `ai_plan/`. Run `create-plan` first."

## Process

### Phase 1: Locate Plan

1. Scan `ai_plan/` for plan directories (most recent first by date prefix).
2. If multiple plans exist, ask user which one to implement.
3. If no plan exists, stop: "No plan found. Run create-plan first."
4. Read `continuation-runbook.md` first (source of truth).
5. Read `story-tracker.md` to detect resume state (`in-dev` or `completed` stories).
6. Read `milestone-plan.md` for implementation details.

### Phase 2: Configure Reviewer

If the user has already specified a reviewer CLI and model (e.g., "implement the plan, review with claude sonnet"), use those values. Otherwise, ask:

1. **Which CLI should review each milestone?**
   - `codex` — OpenAI Codex CLI (`codex exec`)
   - `claude` — Claude Code CLI (`claude -p`)
   - `cursor` — Cursor Agent CLI (`cursor-agent -p`)
   - `skip` — No external review, proceed with user approval only

2. **Which model?** (only if a CLI was chosen)
   - For `codex`: default `o4-mini`, alternatives: `gpt-5.3-codex`, `o3`
   - For `claude`: default `sonnet`, alternatives: `opus`, `haiku`
   - For `cursor`: **run `cursor-agent models` first** to see available models
   - Accept any model string the user provides

3. **Max review rounds per milestone?** (default: 10)
   - If the user does not provide a value, set `MAX_ROUNDS=10`.

Store `REVIEWER_CLI`, `REVIEWER_MODEL`, and `MAX_ROUNDS`. These values are fixed for the entire run.

### Phase 3: Set Up Worktree (REQUIRED SUB-SKILL)

Invoke `superpowers:using-git-worktrees` explicitly.

1. Branch naming: `implement/<plan-folder-name>` (e.g., `implement/2026-03-04-auth-system`).
2. Follow worktree skill's directory priority: `.worktrees/` > `worktrees/` > CLAUDE.md > ask user.
3. Verify `.gitignore` covers worktree directory.
4. Run project setup (auto-detect: `npm install`, `cargo build`, `pip install`, etc.).
5. Verify clean baseline (run tests).

**Resume detection:** If `story-tracker.md` shows `in-dev` or `completed` stories, check if worktree branch already exists (`git worktree list`). If so, `cd` into existing worktree instead of creating a new one.

### Phase 4: Execute Milestones (Loop)

For each milestone (M1, M2, ...):

#### Step 1: Read Milestone Spec

Read the milestone section from `milestone-plan.md`.

#### Step 2: Update Tracker

Mark first story `in-dev` in `story-tracker.md`.

#### Step 3: Implement Stories

Execute each story in order. After completing each story:
1. Mark `in-dev` -> `completed` in `story-tracker.md`
2. Update counts
3. Mark next story `in-dev`

Commit hashes are not available yet — they are backfilled in Step 6 after the milestone is approved and committed.

#### Step 4: Verify Milestone (REQUIRED SUB-SKILL)

Invoke `superpowers:verification-before-completion` explicitly.

```bash
# Lint changed files
# Typecheck
# Run tests (targeted first, then full suite)
```

All must pass before proceeding. If failures: fix, re-verify. Do NOT proceed to review with failures.

#### Step 5: Milestone Review Loop

Send to reviewer for approval **before committing**. See Phase 5 for details. The review payload uses working-tree diffs (`git diff` for unstaged, `git diff --staged` for staged changes).

**Skip this step if reviewer was set to `skip`.** When skipped, present the milestone summary to the user and ask for approval directly.

#### Step 6: Commit & Approve

Only after the reviewer approves (or user overrides at max rounds):

```bash
git add <changed-files>
git commit -m "feat(<scope>): implement milestone M<N> - <description>"
```

Do NOT push. After committing:
1. Backfill the commit hash into the Notes column for all stories in this milestone in `story-tracker.md`.
2. Mark milestone as `approved` in `story-tracker.md`.
3. Move to next milestone.

### Phase 5: Milestone Review Loop (Detail)

**Skip this phase entirely if reviewer was set to `skip`.**

#### Step 1: Generate Session ID

```bash
REVIEW_ID=$(uuidgen | tr '[:upper:]' '[:lower:]' | head -c 8)
```

Use `REVIEW_ID` for all milestone review temp file paths:
- `/tmp/milestone-${REVIEW_ID}.md`
- `/tmp/milestone-review-${REVIEW_ID}.md`
- `/tmp/milestone-review-${REVIEW_ID}.json`
- `/tmp/milestone-review-${REVIEW_ID}.stderr`
- `/tmp/milestone-review-${REVIEW_ID}.status`
- `/tmp/milestone-review-${REVIEW_ID}.runner.out`
- `/tmp/milestone-review-${REVIEW_ID}.sh`

Resolve the shared runtime helper path before writing the command script:

```bash
REVIEWER_RUNTIME=~/.claude/skills/reviewer-runtime/run-review.sh
```

Set helper success-artifact args before writing the command script:

```bash
HELPER_SUCCESS_FILE_ARGS=()
case "$REVIEWER_CLI" in
  codex)
    HELPER_SUCCESS_FILE_ARGS+=(--success-file /tmp/milestone-review-${REVIEW_ID}.md)
    ;;
  cursor)
    HELPER_SUCCESS_FILE_ARGS+=(--success-file /tmp/milestone-review-${REVIEW_ID}.json)
    ;;
esac
```

#### Step 2: Write Review Payload

Write to `/tmp/milestone-${REVIEW_ID}.md`:

```markdown
# Milestone M<N> Review: <title>

## Milestone Spec (from plan)
[Copy milestone section from milestone-plan.md]

## Acceptance Criteria
[Copy acceptance criteria checkboxes]

## Changes Made (git diff)
[Output of: git diff -- for unstaged changes, or git diff --staged for staged changes]

## Verification Output
### Lint
[lint output]
### Typecheck
[typecheck output]
### Tests
[test output with pass/fail counts]
```

#### Review Contract (Applies to Every Round)

The reviewer response must use this structure:

```text
## Summary
...

## Findings
### P0
- ...
### P1
- ...
### P2
- ...
### P3
- ...

## Verdict
VERDICT: APPROVED
```

Rules:
- Order findings from `P0` to `P3`.
- `P0` = total blocker, `P1` = major risk, `P2` = must-fix before approval, `P3` = cosmetic / nice to have.
- Use `- None.` when a severity has no findings.
- `VERDICT: APPROVED` is allowed only when no `P0`, `P1`, or `P2` findings remain. `P3` findings are non-blocking.
- The calling agent should still try to fix `P3` findings when they are cheap and safe.

#### Liveness Contract (Applies While Review Is Running)

- The shared reviewer runtime emits `state=in-progress note="In progress N"` heartbeats every 60 seconds while the reviewer child is alive.
- The calling agent must keep waiting as long as a fresh `In progress N` heartbeat keeps arriving roughly once per minute.
- Do not abort just because the review is slow, a soft timeout fired, or a `stall-warning` line appears, as long as the `In progress N` heartbeat continues.
- Treat missing heartbeats, `state=failed`, `state=completed-empty-output`, and `state=needs-operator-decision` as escalation signals.

#### Step 3: Submit to Reviewer (Round 1)

Write the reviewer invocation to `/tmp/milestone-review-${REVIEW_ID}.sh` as a bash script:

```bash
#!/usr/bin/env bash
set -euo pipefail
```

**If `REVIEWER_CLI` is `codex`:**

```bash
codex exec \
  -m ${REVIEWER_MODEL} \
  -s read-only \
  -o /tmp/milestone-review-${REVIEW_ID}.md \
  "Review this milestone implementation. The spec, acceptance criteria, git diff, and verification output are in /tmp/milestone-${REVIEW_ID}.md.

Evaluate:
1. Correctness — Does the implementation match the milestone spec?
2. Acceptance criteria — Are all criteria met?
3. Code quality — Clean, maintainable, no obvious issues?
4. Test coverage — Are changes adequately tested?
5. Security — Any security concerns introduced?


Return exactly these sections in order:
## Summary
## Findings
### P0
### P1
### P2
### P3
## Verdict

Rules:
- Order findings from highest severity to lowest.
- Use `- None.` when a severity has no findings.
- `P0` = total blocker, `P1` = major risk, `P2` = must-fix before approval, `P3` = cosmetic / nice to have.
- End with exactly one verdict line: `VERDICT: APPROVED` or `VERDICT: REVISE`
- `VERDICT: APPROVED` is allowed only when there are no `P0`, `P1`, or `P2` findings. `P3` findings are non-blocking."
```

Do not try to capture the Codex session ID yet. When using the helper, extract it from `/tmp/milestone-review-${REVIEW_ID}.runner.out` after the command completes (look for `session id: <uuid>`), then store it as `CODEX_SESSION_ID` for resume in subsequent rounds.

**If `REVIEWER_CLI` is `claude`:**

```bash
claude -p \
  "Review this milestone implementation using the following spec, acceptance criteria, git diff, and verification output:

$(cat /tmp/milestone-${REVIEW_ID}.md)

Evaluate:
1. Correctness — Does the implementation match the milestone spec?
2. Acceptance criteria — Are all criteria met?
3. Code quality — Clean, maintainable, no obvious issues?
4. Test coverage — Are changes adequately tested?
5. Security — Any security concerns introduced?


Return exactly these sections in order:
## Summary
## Findings
### P0
### P1
### P2
### P3
## Verdict

Rules:
- Order findings from highest severity to lowest.
- Use `- None.` when a severity has no findings.
- `P0` = total blocker, `P1` = major risk, `P2` = must-fix before approval, `P3` = cosmetic / nice to have.
- End with exactly one verdict line: `VERDICT: APPROVED` or `VERDICT: REVISE`
- `VERDICT: APPROVED` is allowed only when there are no `P0`, `P1`, or `P2` findings. `P3` findings are non-blocking." \
  --model ${REVIEWER_MODEL} \
  --strict-mcp-config \
  --setting-sources user
```

**If `REVIEWER_CLI` is `cursor`:**

```bash
cursor-agent -p \
  --mode=ask \
  --model ${REVIEWER_MODEL} \
  --trust \
  --output-format json \
  "Read the file /tmp/milestone-${REVIEW_ID}.md and review this milestone implementation.

Evaluate:
1. Correctness — Does the implementation match the milestone spec?
2. Acceptance criteria — Are all criteria met?
3. Code quality — Clean, maintainable, no obvious issues?
4. Test coverage — Are changes adequately tested?
5. Security — Any security concerns introduced?


Return exactly these sections in order:
## Summary
## Findings
### P0
### P1
### P2
### P3
## Verdict

Rules:
- Order findings from highest severity to lowest.
- Use `- None.` when a severity has no findings.
- `P0` = total blocker, `P1` = major risk, `P2` = must-fix before approval, `P3` = cosmetic / nice to have.
- End with exactly one verdict line: `VERDICT: APPROVED` or `VERDICT: REVISE`
- `VERDICT: APPROVED` is allowed only when there are no `P0`, `P1`, or `P2` findings. `P3` findings are non-blocking." \
  > /tmp/milestone-review-${REVIEW_ID}.json
```

For `cursor`, the command script writes raw JSON to `/tmp/milestone-review-${REVIEW_ID}.json`. Do not run `jq` extraction until after the helper or fallback execution completes. If `jq` is not installed, inform the user: `brew install jq` (macOS) or equivalent.

Run the command script through the shared helper when available:

```bash
if [ -x "$REVIEWER_RUNTIME" ]; then
  "$REVIEWER_RUNTIME" \
    --command-file /tmp/milestone-review-${REVIEW_ID}.sh \
    --stdout-file /tmp/milestone-review-${REVIEW_ID}.runner.out \
    --stderr-file /tmp/milestone-review-${REVIEW_ID}.stderr \
    --status-file /tmp/milestone-review-${REVIEW_ID}.status \
    "${HELPER_SUCCESS_FILE_ARGS[@]}"
else
  echo "Warning: reviewer runtime helper not found at $REVIEWER_RUNTIME; falling back to direct synchronous review." >&2
  bash /tmp/milestone-review-${REVIEW_ID}.sh >/tmp/milestone-review-${REVIEW_ID}.runner.out 2>/tmp/milestone-review-${REVIEW_ID}.stderr
fi
```

Run the helper in the foreground and watch its live stdout for `state=in-progress` heartbeats. If your agent environment buffers command output until exit, start the helper in the background and poll `/tmp/milestone-review-${REVIEW_ID}.status` separately instead of treating heartbeats as post-hoc-only data.

After the command completes:
- If `REVIEWER_CLI=cursor`, extract the final review text:

```bash
CURSOR_SESSION_ID=$(jq -r '.session_id' /tmp/milestone-review-${REVIEW_ID}.json)
jq -r '.result' /tmp/milestone-review-${REVIEW_ID}.json > /tmp/milestone-review-${REVIEW_ID}.md
```

- If `REVIEWER_CLI=codex`, extract `CODEX_SESSION_ID` from `/tmp/milestone-review-${REVIEW_ID}.runner.out` after the helper or fallback run. If the review text is only in `.runner.out`, move or copy the actual review body into `/tmp/milestone-review-${REVIEW_ID}.md` before verdict parsing.
- If `REVIEWER_CLI=claude`, promote stdout captured by the helper or fallback runner into the markdown review file:

```bash
cp /tmp/milestone-review-${REVIEW_ID}.runner.out /tmp/milestone-review-${REVIEW_ID}.md
```

Fallback is allowed only when the helper is missing or not executable.

#### Step 4: Read Review & Check Verdict

1. Read `/tmp/milestone-review-${REVIEW_ID}.md`
2. If the review failed, produced empty output, or reached helper timeout, also read:
   - `/tmp/milestone-review-${REVIEW_ID}.stderr`
   - `/tmp/milestone-review-${REVIEW_ID}.status`
   - `/tmp/milestone-review-${REVIEW_ID}.runner.out`
3. Present review to the user:

```
## Milestone Review — Round N (reviewer: ${REVIEWER_CLI} / ${REVIEWER_MODEL})

[Reviewer feedback]
```

4. While the reviewer is still running, keep waiting as long as fresh `state=in-progress note="In progress N"` heartbeats continue to appear roughly once per minute.
5. Check verdict:
   - **VERDICT: APPROVED** with no `P0`, `P1`, or `P2` findings -> proceed to Phase 4 Step 6 (commit & approve)
   - **VERDICT: APPROVED** with only `P3` findings -> optionally fix the `P3` items if they are cheap and safe, then proceed
   - **VERDICT: REVISE** or any `P0`, `P1`, or `P2` finding -> go to Step 5
   - No clear verdict but `P0`, `P1`, and `P2` are all `- None.` -> treat as approved
   - Helper state `completed-empty-output` -> treat as failed review attempt, surface stderr/status, fix invocation or prompt handling, then retry
   - Helper state `needs-operator-decision` -> surface status log and decide whether to extend the timeout, abort, or retry with different helper parameters
   - Max rounds (`MAX_ROUNDS`) reached -> present to user for manual decision (proceed or stop)

#### Step 5: Address Feedback & Re-verify

1. Address the reviewer findings in priority order (`P0` -> `P1` -> `P2`, then `P3` when practical) (do NOT commit yet).
2. Re-run verification (lint/typecheck/tests) — all must pass.
3. Update `/tmp/milestone-${REVIEW_ID}.md` with new diff and verification output.

Summarize revisions for the user:

```
### Revisions (Round N)
- [Change and reason, one bullet per issue addressed]
```

If a revision contradicts the user's explicit requirements, skip it and note it for the user.

#### Step 6: Re-submit to Reviewer (Rounds 2-N)

Rewrite `/tmp/milestone-review-${REVIEW_ID}.sh` for the next round. The script should contain the reviewer invocation only; do not run it directly.

**If `REVIEWER_CLI` is `codex`:**

Resume the existing session:

```bash
codex exec resume ${CODEX_SESSION_ID} \
  -o /tmp/milestone-review-${REVIEW_ID}.md \
  "I've addressed your feedback. Updated diff and verification output are in /tmp/milestone-${REVIEW_ID}.md.

Changes made:
[List specific changes]

Re-review using the same `## Summary`, `## Findings`, and `## Verdict` structure as before.
Keep findings ordered `P0` to `P3`, use `- None.` when a severity has no findings, and only use `VERDICT: APPROVED` when no `P0`, `P1`, or `P2` findings remain. `P3` findings are non-blocking."
```

If resume fails (session expired), fall back to fresh `codex exec` with context about prior rounds.

**If `REVIEWER_CLI` is `claude`:**

Fresh call with accumulated context (Claude CLI has no session resume):

```bash
claude -p \
  "You previously reviewed milestone M<N> and requested revisions.

Previous feedback summary: [key points from last review]

I've addressed your feedback. Updated diff and verification output are below.

$(cat /tmp/milestone-${REVIEW_ID}.md)

Changes made:
[List specific changes]

Re-review using the same `## Summary`, `## Findings`, and `## Verdict` structure as before.
Keep findings ordered `P0` to `P3`, use `- None.` when a severity has no findings, and only use `VERDICT: APPROVED` when no `P0`, `P1`, or `P2` findings remain. `P3` findings are non-blocking." \
  --model ${REVIEWER_MODEL} \
  --strict-mcp-config \
  --setting-sources user
```

**If `REVIEWER_CLI` is `cursor`:**

Resume the existing session:

```bash
cursor-agent --resume ${CURSOR_SESSION_ID} -p \
  --mode=ask \
  --model ${REVIEWER_MODEL} \
  --trust \
  --output-format json \
  "I've addressed your feedback. Updated diff and verification output are in /tmp/milestone-${REVIEW_ID}.md.

Changes made:
[List specific changes]

Re-review using the same `## Summary`, `## Findings`, and `## Verdict` structure as before.
Keep findings ordered `P0` to `P3`, use `- None.` when a severity has no findings, and only use `VERDICT: APPROVED` when no `P0`, `P1`, or `P2` findings remain. `P3` findings are non-blocking." \
  > /tmp/milestone-review-${REVIEW_ID}.json
```

If resume fails, fall back to fresh `cursor-agent -p` with context about prior rounds.

Do not run `jq` extraction until after the helper or fallback execution completes, then extract `/tmp/milestone-review-${REVIEW_ID}.md` from the JSON response.

After updating `/tmp/milestone-review-${REVIEW_ID}.sh`, run the same helper/fallback flow from Round 1.

Return to Step 4.

#### Step 7: Cleanup Per Milestone

```bash
rm -f \
  /tmp/milestone-${REVIEW_ID}.md \
  /tmp/milestone-review-${REVIEW_ID}.md \
  /tmp/milestone-review-${REVIEW_ID}.json \
  /tmp/milestone-review-${REVIEW_ID}.stderr \
  /tmp/milestone-review-${REVIEW_ID}.status \
  /tmp/milestone-review-${REVIEW_ID}.runner.out \
  /tmp/milestone-review-${REVIEW_ID}.sh
```

If the round failed, produced empty output, or reached operator-decision timeout, keep `.stderr`, `.status`, and `.runner.out` until the issue is diagnosed instead of deleting them immediately.

### Phase 6: Completion (REQUIRED SUB-SKILL)

After all milestones are approved and committed:

1. Invoke `superpowers:finishing-a-development-branch` explicitly.
2. Run full test suite one final time — all must pass.
3. Merge the worktree branch into the parent branch:

```bash
# From the main repo (not the worktree)
git merge implement/<plan-folder-name>
```

4. Delete the worktree and its branch:

```bash
git worktree remove <worktree-path>
git branch -d implement/<plan-folder-name>
```

5. Mark plan status as `completed` in `story-tracker.md`.

### Phase 7: Final Report

Present summary:

```
## Implementation Complete

**Plan:** <plan-folder-name>
**Milestones:** <N> completed, <N> approved
**Review rounds:** <total across all milestones>
**Branch:** implement/<plan-folder-name> (merged and deleted)
```

### Phase 8: Telegram Completion Notification (MANDATORY)

Resolve the Telegram notifier helper from the installed Claude Code skills directory:

```bash
TELEGRAM_NOTIFY_RUNTIME=~/.claude/skills/reviewer-runtime/notify-telegram.sh
```

On every terminal outcome for the implement-plan run (fully completed, stopped after max rounds, skipped reviewer, or failure), send a Telegram summary if the helper exists and both `TELEGRAM_BOT_TOKEN` and `TELEGRAM_CHAT_ID` are configured:

```bash
if [ -x "$TELEGRAM_NOTIFY_RUNTIME" ] && [ -n "${TELEGRAM_BOT_TOKEN:-}" ] && [ -n "${TELEGRAM_CHAT_ID:-}" ]; then
  "$TELEGRAM_NOTIFY_RUNTIME" --message "implement-plan completed for <plan-folder-name>: <status summary>"
fi
```

Rules:
- Telegram is the only supported completion notification path. Do not use desktop notifications, `say`, email, or any other notifier.
- Notification failures are non-blocking, but they must be surfaced to the user.
- If Telegram is not configured, state that no completion notification was sent.

## Tracker Discipline (MANDATORY)

**ALWAYS update `story-tracker.md` before/after each story. NEVER proceed with stale tracker state.**

Before starting any story:
1. Open `story-tracker.md`
2. Mark story `in-dev`
3. Add notes if relevant
4. Then begin implementation

After completing any story:
1. Mark story `completed`
2. Review pending stories
3. Update Last Updated and Stories Complete counts

Note: Commit hashes are backfilled into story Notes after the milestone commit (Step 6), not per-story.

## Verification Checklist

- [ ] Plan folder located and all required files present
- [ ] Reviewer configured or explicitly skipped
- [ ] Max review rounds confirmed (default: 10)
- [ ] Worktree created with branch `implement/<plan-folder-name>`
- [ ] Worktree directory verified in .gitignore
- [ ] Baseline tests pass in worktree
- [ ] Each milestone: stories tracked (in-dev -> completed)
- [ ] Each milestone: lint/typecheck/tests pass before review
- [ ] Each milestone: reviewer approved (or max rounds + user override)
- [ ] Each milestone: committed locally only after approval
- [ ] Each milestone: marked approved in story-tracker.md
- [ ] All milestones completed, approved, and committed
- [ ] Final test suite passes
- [ ] Worktree branch merged to parent and worktree deleted
- [ ] Story tracker updated with final status
- [ ] Telegram completion notification attempted if configured