Add /plan-review command and plan-reviewer subagent

[kubiknyc]

What

Adds tooling to review an implementation plan from fresh context before it gets built. A plan reviewed inside the context that produced it is rationalization — the model is anchored on its own reasoning. Two new files: a plan-reviewer subagent and a /plan-review slash command that fans it out.

plan-reviewer (.claude/agents/plan-reviewer.md) is read-only (Read, Grep, Glob, Bash). It takes a lens — factual-grounding, adversarial, or invariants — and returns a JSON verdict. It is told to default to refuted, to treat "I could not verify this" as a valid finding, and to check claims about existing code by opening the file rather than trusting the plan's description of it. It complements worklog-reviewer, which reviews diffs; this one runs while throwing the plan away is still cheap. The invariants lens carries the plan-stage version of this repo's hard rules — OTA safety, SELECT-only report tables, schema parity, platform split, sync purity, the mutationQueue.ts coverage pin, tests never under app/, testID coverage, and open decisions in docs/architecture/00-README.md.

/plan-review (.claude/commands/plan-review.md) writes the plan and the original request to .plan-review/, spawns the three lenses in parallel with only those two file paths — deliberately withholding the author's reasoning — then merges: drops findings with neither a file nor a concrete failure scenario, settles reviewer disagreements by reading the code rather than averaging, and reports surviving assumptions before the user approves.

.plan-review/ is gitignored. It also gives a future PreToolUse hook on ExitPlanMode a findings file to gate on, if the advisory version turns out not to be enough.

Milestone / plan task

None — tooling, not a milestone task.

Gates

• npm run verify green locally — not run: node_modules is absent in this container and npm run typecheck fails at expo/tsconfig.base, which is environmental, not caused by this diff. The change is outside everything verify covers: no TS, no tests, and .prettierignore excludes both .claude/ and *.md, so format:check does not read either new file. CI will confirm.
• Backend schema changed? → no
• New native dependency? → no
• New screen or element an E2E flow will drive? → no
• Touches an open decision in docs/architecture/00-README.md? → no
• New user-facing copy? → no UI copy; the agent and command prose is agent-facing
• Changes SQLite schema or mutation payload shape? → no

Open decisions touched

None.

Adversarial review

Not run — worklog-reviewer checks app/, src/, scripts/, and .maestro/, and this diff touches none of them. Two agent-configuration markdown files and one .gitignore line.

One judgement call worth flagging for review: the three lenses are fixed in the command rather than being chosen per plan. That is deliberate — letting the author pick the lenses reintroduces the bias the fan-out exists to remove — but it means a plan whose real risk sits outside all three gets a clean verdict. The unverifiable field is the partial mitigation, and the merge step is instructed to treat it as amber rather than green.

---

Generated by Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: ffcff689-54f7-4be3-bc3e-f2d2232d8b61

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}