Lab Plan

---
name: lab-plan
description: Run a structured planning session — interview the sources of truth, draft, adversarial review (up to 3 reviewers, scaled to the change), approve, update docs. Required for any change to the shipped engine, a skill/template, a hook, recall, security, or a public release, and for features spanning several files or systems; a small change touching none of those doesn't need it.
---

# Planning Session Protocol

You are running a structured planning session. Follow the steps in order, but **scale the ceremony to the change** (see the floor checklist below): a floor-checklist change runs every step in full; a small, self-contained change that touches no floor item may merge or skip steps that don't earn their weight — but never skip the SoT grounding (Step 2), the approval gate (Step 5), the adversarial pass for any floor-checklist change (Step 4), or the tracking updates when DECISIONS/STATUS change (Step 6).

**Floor checklist — run the full ceremony (all 3 adversarial lenses included) whenever the change touches ANY of these:** the shipped engine (`src/`) · a shipped skill / template / assets file · a git-hook · recall scoring or indexing · security / redaction / auth · user data · a public release · or a feature spanning several files or systems. These are objective properties of the diff, not a judgment call about how big the change *feels* — an author who under-rates their own change is the exact failure this pass guards against. Scale down (Step 4) only when every box is unchecked.

## Arguments

`$ARGUMENTS` — short name for the plan (e.g. "browser-automation", "auth-rework").

## Pre-flight

1. Read `Log/STATUS.md` — current project state.
2. Read `Log/DECISIONS.md` (if it exists) — what's already decided.
3. Read the most recent file in `Sessions/` — pick up context.
4. Read the relevant `Source/` genome docs (INTENT / SPEC / USER-STORIES) for the area this plan touches.

## Step 1: Create the plan file

Create `Log/plans/YYYY-MM-DD_<short-name>.md` (copy `Log/plans/TEMPLATE.md` if present) using today's date and the argument. Set status to `draft`.

## Step 2: Sources of Truth (SoT) interview

This is an **active interview driven by you**, not a passive file review. For each source of truth relevant to this plan:

1. Read the source file.
2. Create an **annotation file** at `Log/plans/<plan-name>-sot-notes.md` with your findings and open questions.
3. Present findings to the user with specific questions — do NOT ask verbally without the annotation file.
4. Record the user's answers in the annotation file.

Sources to check (pick the relevant ones):
- `Source/INTENT.md` / `Source/SPEC.md` — what the project intends + specifies for this area.
- Existing implementation — what's already built (grep/glob the codebase).
- `CLAUDE.md` — constraints and protocols.
- `Log/DECISIONS.md` — prior decisions that constrain this plan.

## Step 3: Draft the plan

Fill in the plan file with:
- **Problem statement** — what this solves, why now.
- **Design** — architecture, tradeoffs, why this approach wins.
- **Tasks** — numbered task table with estimates and dependencies.
- **Risks** — what could go wrong, rollback plan.
- **Verification** — specific, checkable acceptance criteria.

Set status to `review`. Tell the user the plan is ready.

## Step 4: Adversarial review

After the user approves the draft for review (or immediately if they say "go"):

Run **up to 3** adversarial reviewer passes — feasibility, risk, scope — **scaled to the change, not fixed at 3.** If the change hits the floor checklist above, run all 3 (mandatory). Otherwise scale down: a moderate change runs **risk plus one other** (risk is never the lens you drop — it's the one with the track record for catching the real breakage); a genuinely small change that touches no floor item may run **risk only**, or **none** if you can fully verify it yourself in a few tool calls. Prefer higher reasoning `effort` on fewer, sharper lenses over spawning agents for their own sake — and don't spawn reviewers to double-check work you can verify yourself; reserve the pass for genuinely independent adversarial perspective on a *plan*, not self-review.

Run each pass in parallel where the harness supports subagents, sequentially otherwise — each with a focused, adversarial domain. Tell each to be skeptical and concrete, and to ground every finding in the actual files (read them) — not vibes:

1. **Feasibility** — can each task be built with the available tools, patterns, and dependencies? Are the estimates realistic?
2. **Risk** — what breaks? Failure modes, data loss, security gaps, regression potential, isolation violations.
3. **Scope** — does the plan match the problem statement? Flag overengineering, scope creep, and under-scoping.

Each dispatch prompt includes: the plan file path; the relevant `Source/` + `Log/STATUS.md` + `Log/DECISIONS.md` paths; any source files the plan proposes to change (so findings are grounded); and an instruction to return structured findings — area/task, issue, severity (critical/important/minor), suggested fix.

### Consolidation
After the reviewer passes return:
1. Collect findings into one numbered list.
2. Categorize: critical (must fix) / important (should fix) / minor.
3. Apply critical + important fixes to the plan.
4. Present the consolidated findings and changes to the user.

## Step 5: Approval

Present the final plan to the user. Wait for explicit approval before proceeding.

## Step 6: Post-approval documentation

Once approved, **immediately** (in the same response):
1. Set plan status to `approved`.
2. Add a decision entry to `Log/DECISIONS.md` (if used) with the plan reference.
3. Update `Log/STATUS.md` with the new plan.
4. Update `Log/TASKS.md` / `Log/PLAN.md` if the plan creates new phases or tasks.
5. Update the session note in `Sessions/`.

Do not proceed to implementation in this session unless the user explicitly asks.