/review-it

The QA front door — prove the build obeys the plan

plan-it plans. fable-it builds. review-it verifies.
The independent verification leg of the plan → build → review triangle.

Install  ·  Why  ·  How it works  ·  The gate catalog  ·  What's bundled  ·  Changelog

Two real incidents define this plugin (the Airtable postmortem):

1. A run shipped "VERIFIED" UI with 4 operability bugs — the tests checked that a combobox rendered text, never that it could be typed into, opened, or used to select anything.
2. A third-party write whose record rendered empty in Airtable's own UI while every API GET looked correct — verified in the wrong layer.

Both were caught by a human, after the report said green. review-it makes those two failures — and nine sibling classes — mechanically un-shippable, with a gate catalog where every rule is written as trigger → test → action and every verdict is a lookup into an evidence ledger, not a judgment call.

And the quiet failure underneath them all: a QA run with no authored test contract that quietly reverse-engineers its expectations from the code it is testing — and then grades itself green. review-it tags every verdict's oracle AUTHORED or DERIVED; a DERIVED green means "self-consistent," never "obeys the plan." No self-graded green.

# 1. Add the DevOtts marketplace (once — shared with fable-it / plan-it)
/plugin marketplace add DevOtts/review-it

# 2. Install
/plugin install review-it@devotts

# Or as a standalone skill package
npx skills add DevOtts/review-it

Then hand it a target:

/review-it qa/test-plan-master.md      # contract-qa: run the plan-phase Test Contract
/review-it PR #42                      # pr-review: severity-tiered, evidence-cited
/review-it prod                        # deploy-verify: is it ACTUALLY live + ready?
/review-it <feature dir with third-party writes>   # side-effects verification

/review-it <target>
   │
   ├─ MODE contract-qa    → PRIMARY: run the plan-phase Test Contract against the build
   ├─ MODE side-effects   → third-party writes read back from the target's OWN API + UI
   ├─ MODE deploy-verify  → deployed-code ladder + [REAL] re-runs + release checklist
   └─ MODE pr-review      → SECONDARY: severity-tiered findings, blocking vs advisory
   ▼
  GATES  — 11 rules, applied in every mode
   ▼
  REPORT — one format, shared with fable-it's evidence ledger

• Preflight (R9) — prove which app/branch/checkout is under test before any verdict.
• Verifiability precheck — unreachable target ⇒ honest IMPLEMENTED-NOT-VERIFIED with a named blocker, never a manufactured green. [REAL] cases are never verified on a mock.
• No-contract ladder — invoked with no Test Contract, it never refuses and never self-grades: locate an authored oracle (plan-it DoDs/goals count) → derive only if none → confirm → label provenance → persist the derived contract as durable coverage.
• Routing, not re-implementation — execution goes to the existing specialists by name: full-qa (functional/CDP QA), iterate (fix loops), chrome-cdp-control (authenticated real Chrome), make-eval (LLM evals), parallel-lifecycle (isolation).
• The honesty layer — every mode's verdict passes an evidence adapter (VERIFIED is a ledger lookup), a system-of-record adapter (subagent narration is never evidence), and a fresh-context verifier that reads the pixels, not just the prose.

Full specs (each trigger → test → action): plugins/review-it/skills/references/gate-catalog.md.

review-it consumes plan-it's Test Contract 1:1 and feeds its verdict rows into fable-it's evidence ledger — one report format, hosted here, both plugins point at it.

review-it is read-mostly by design: it writes only its report/ledger conventions in consumer repos, never proposes credential storage, human-gates credential operations and destructive steps (safety ladder), and routes authenticated browser work to chrome-cdp-control's per-write confirmation gate. Details in SKILL.md.

MIT — see LICENSE.

Authored by DevOtts.