Phase 2: ds-plan -> PLAN.md written
  -> [THIS SKILL] Dispatch plan reviewer subagent
  -> For plans with >15 tasks: review per-chunk
  -> Issues found? Fix PLAN.md -> re-dispatch reviewer
  -> Approved? -> Phase 3: ds-implement

Agent(
  subagent_type="general-purpose",
  description="Review DS plan document",
  allowed_tools=["Read", "Glob", "Grep", "Bash(read-only)"],
  prompt="""
You are a data science plan document reviewer. Verify this plan is complete, matches the spec, and is ready for implementation.

**Tool Restrictions:** The plan reviewer is READ-ONLY. It reads `.planning/PLAN.md` and `.planning/SPEC.md`, evaluates against checklist, returns verdict. It MUST NOT use Write or Edit.

**Plan to review:** .planning/PLAN.md [-- Chunk N only, if chunked]
**Spec for reference:** .planning/SPEC.md

Read BOTH files, then evaluate the plan against ALL categories below.

## What to Check

| Category | What to Look For |
|----------|------------------|
| Completeness | TODOs, placeholders, incomplete tasks, missing steps |
| Spec Alignment | Plan covers ALL spec requirements, no scope creep, no requirements silently dropped |
| Data Profiling | Data profile section present with shape, types, quality issues documented |
| Task Decomposition | Tasks atomic enough for a single subagent, clear boundaries, steps actionable |
| Task Ordering | Dependencies correct (cleaning before analysis), no circular dependencies |
| Intermediate Outputs | Each task defines what it produces and what proves completion |
| Output-First Verification | Each task includes verification steps (print shape, check nulls, sample output) |
| ETL Strategy | If data > 1M rows or multiple sources: filter strategy, parallelism plan, caching documented |
| Reproducibility | Random seeds, package versions, data snapshots documented where relevant |

## CRITICAL - Look Especially Hard For:

- Any TODO markers or placeholder text
- Steps that say "similar to X" without actual content
- Tasks missing intermediate output definitions (what does this task produce?)
- Tasks missing verification steps (how do you know it worked?)
- Missing data profiling tasks (should always come before analysis)
- Data cleaning tasks that lack strategy for each quality issue found in profiling
- Spec requirements not covered by ANY task (silently dropped)
- Tasks too large for a single subagent (>100 lines of change or multiple distinct operations)
- ETL strategy missing when data is large (>1M rows) or from multiple sources
- Missing output verification plan section

## Output Format

## Plan Review

**Status:** APPROVED | ISSUES_FOUND

**Issues (if any):**
- [Task X, Step Y]: [specific issue] - [why it matters for implementation]

**Spec Coverage Check:**
- [Requirement 1]: Covered by Task N | NOT COVERED
- [Requirement 2]: Covered by Task N | NOT COVERED

**Recommendations (advisory - don't block approval):**
- [suggestions for improvement that aren't blocking]
""")

"Plan reviewer has flagged issues 5 times. Remaining issues:
[list issues]
Should I: (A) Fix these, (B) Proceed with known gaps, (C) Rethink the plan?"

1. IDENTIFY: `.planning/PLAN.md` exists
2. DISPATCH: Send to reviewer subagent (per-chunk if >15 tasks)
3. READ: Reviewer returns APPROVED or ISSUES_FOUND
4. VERIFY: If ISSUES_FOUND, fix and re-dispatch (max 5)
5. CLAIM: Only proceed to ds-implement when ALL chunks APPROVED