Code review plan create

Create a Code-Review Meta-Plan

First punch-in command of the code-review workflow. Produces the meta-plan that drives slicing — a single document at planning/YYYY-MM-DD-code-review-plan-<slug>.md (frontmatter type: code-review-plan) with a slice table that the rest of the workflow operates on.

Core principle: Git is the source of truth, plans are best-effort context. Commits land in three patterns — clean (matches a plan), revised (matches a post-receive plan version), drifted (matches no plan). This skill maps what it can and Q&As what it can't.

When to Use

• After implementation work has landed on a branch and you're ready to review it.
• When you want to slice a long-running branch into reviewable chunks.
• When you have a branch with mixed plan-driven and ad-hoc work and need to surface boundaries.

Process

Step 1: Identify Branch and Base

Defaults: current branch, base = origin/main. Confirm both with the user before walking commits. Edge cases:

• Branch has no upstream → ask which base to use.
• Working tree dirty → warn but proceed; SHAs in the meta-plan are committed-only.
• Detached HEAD → bail, ask user to check out a branch.

Step 2: Walk Commits

git log --oneline --reverse <base>..HEAD
git log --pretty=format:"%H%x09%ad%x09%s" --date=short --reverse <base>..HEAD

Capture for every commit:

• full SHA, short SHA
• author date (yyyy-mm-dd)
• subject line
• body (for plan-doc references / co-author info)
• file-touch summary (git log --stat -1 <SHA>)

Step 3: Discover Candidate Plan Docs

Scan all of:

~/.claude/plans/*.md
~/.copilot/session-state/*/plan.md
<project>/planning/**/*.md

For each candidate, extract:

• frontmatter type (filter to plan or no type — skip code-review-* types here)
• frontmatter slug if present
• first H1 heading
• file mtime and creation date
• any explicit slug-style file basename

Step 4: Map Commits → Plans

For each plan candidate, score commits by:

1. Plan slug or filename in commit subject — strong match.
2. Plan filename referenced in commit body — strong match.
3. Plan H1 keywords overlapping commit subject (≥3 non-trivial words) — medium match.
4. Time-window heuristic: commits authored within [plan_creation_date, plan_creation_date + 14 days] — weak match, used only when other signals tied.
5. Commit subject prefix match (e.g., commits prefixed Phase B1.x: map to a plan whose body or H1 mentions "Phase B").

A commit may match multiple plans; weakest match wins only if no stronger match exists. Threshold: at least one strong OR two medium matches required to claim a mapping.

Step 5: Surface to User

Present a draft mapping in chat:

Commit mapping draft for <branch> (<base>..HEAD = N commits):

PLAN: <path/to/plan-A.md>
  Score: 23 commits matched (12 strong, 11 medium)
  - <SHA> Phase B1.a: dual-write swarm-phase ...
  - <SHA> Phase B1.b: ...
  ...

PLAN: <path/to/plan-B.md>
  Score: 8 commits matched (5 strong, 3 medium)
  - <SHA> Smart Continue: wire the leader's repair_plan_after_failure tool
  ...

UNMAPPED (12 commits):
  - <SHA> SPA: re-hydrate after recovery actions so SSE reconnects
  - <SHA> Randomize hung-swarm-7e1c8d95.sql by default
  ...

Step 6: Q&A on Unmapped Chunks

For each unmapped chunk:

1. Show the commits with subjects and short SHAs.
2. Ask: "These commits aren't mapped to any plan. Are they:
   • one slice of related work, or
   • multiple slices, or
   • should they be folded into an adjacent mapped plan's slice?"
3. If "one/multiple slices", ask for a name and focus areas for each.
4. If "fold into adjacent", ask which plan and adjust the mapping.

Step 7: Confirm Slicing

Show the user the proposed slice table:

Proposed slices:

| # | Name | Commits | Plan docs | Focus |
|---|---|---|---|---|
| 01 | state-machine-foundation | <base>..<sha> (15 commits) | plan-A.md | dual-write transitions, drop legacy columns |
| 02 | smart-continue            | <sha>..<sha>  (5 commits) | plan-B.md | leader repair tool wiring |
| 03 | recovery-recommendation  | <sha>..<sha>  (8 commits) | plan-C.md, plan-D.md | recommendation surface + orphan defense |
| 04 | demo-seeds-randomization | <sha>..HEAD   (3 commits) | (none — ad-hoc) | seed determinism, GUC pinning |

Iterate with user until they say "yes, write it." Allow:

• splitting a slice
• merging two slices
• reordering (must remain chronologically valid: each slice's base_sha is its predecessor's head_sha)
• renaming
• adjusting focus areas

Step 8: Write the Meta-Plan

Filename: planning/YYYY-MM-DD-code-review-plan-<slug>.md. The <slug> is typically the branch name kebab-cased (feature-state-machine → feature-state-machine), or user-provided.

---
type: code-review-plan
branch: <branch>
base_sha: <full SHA>
head_sha: <full SHA>
slug: <slug>
slice_count: <N>
created: YYYY-MM-DD
---

# Code Review Plan — <branch>

## Context

<2-4 paragraphs: what's on the branch, why it's being reviewed now,
known plans and their state (clean / revised / drifted), any constraints.>

## Slice Table

| # | Name | base_sha | head_sha | Plan docs | Focus areas |
|---|---|---|---|---|---|
| 01 | state-machine-foundation | abc1234 | def5678 | planning/<plan-A>.md | • dual-write transitions through IStateTransitionService<br>• drop legacy Phase/Status columns<br>• migration safety |
| 02 | ... | ... | ... | ... | ... |

## Slice Detail

### Slice 01 — state-machine-foundation

**Commits (15):**

<git log --oneline base..head for slice 01>

**Plan docs:**
- [<plan-A.md>](path/to/plan-A.md) — short context on what the plan covered

**Focus areas:**
- Verify all transitions go through IStateTransitionService (no direct entity writes).
- Confirm legacy column drops are accompanied by data backfill or migration safety net.
- ...

**Known context (optional):**
- Plan went through one round of review-receive; the v2 of the plan is what was implemented.
- Phase B1.k landed late and may show up at slice boundary.

### Slice 02 — ...
[same shape]

## Review Order

Default: 01 → 02 → ... → N. Adjust if some slices logically must precede others.

**When `slice_count > 1`, you MUST include a dependency table** instead of prose. Tables are scannable; prose dependency callouts get skimmed past. Example:

| Slice | Depends on | Why |
|---|---|---|
| 01 | (base) | foundation |
| 02 | 01 | builds on the state machine introduced in 01 |
| 03 | 01, 02 | fixes target paths introduced earlier |
| ... | ... | ... |

Every slice from 02 onward should have at least one dependency row — usually `(NN-1)` at minimum, with additional cross-slice deps called out where they exist (e.g., a UI slice depending on an event-broker slice). If a slice is genuinely independent of all others, put `(none)` and explain why.

## Stopping rules

- Pause after any slice that produces ≥3 Critical findings.
- Pause if `code-review-plan-execute` reports >2 failed slices.
- User may always interrupt and re-evaluate.

## Out of scope

- <if any: e.g. "we are not reviewing the docs commits 4a2469d / 1f80579 — they are documentation-only and listed but not sliced.">

Step 9: Confirm and Notify

Report to chat:

• Path of the meta-plan
• Slice count
• Total commits covered
• Any commits intentionally excluded from slicing
• Suggested next step:
  • If you want adversarial review of the slicing strategy: run plan-send-review on the meta-plan.
  • If you trust the slicing: run code-review-plan-execute directly to seed and review.

Heuristics That Help Discovery

• Look for Phase X.y or [ticket-id] prefixes — these are usually plan-owned.
• Co-author trailers (Co-Authored-By:) often correlate with plan-driven work.
• Commits with bodies referencing planning/ paths are giveaways.
• Single-commit "Fix:" / "Chore:" entries are usually drift; group them by file/theme.

Do Not

• Fabricate commits or SHAs.
• Map commits to plans the file system can't confirm.
• Write the meta-plan without user sign-off on the slice table.
• Force every commit into a slice — out-of-scope is fine if the user says so.
• Use a slug already in use elsewhere in planning/ for code-review-plan files.

Red Flags

Thought	Reality
"Just put every commit in slice 01"	A 60-commit slice is unreviewable. Push back.
"I can guess the focus areas"	Ask. Wrong focus areas waste reviewer effort.
"The plan looks close enough"	Confirm with the user — close-enough mappings are how drift accumulates silently.
"Skip the Q&A — auto-cluster the unmapped"	The Q&A is the value. Auto-clustering is what created the mess.