neohaskell/docs/archive/2026-07-ai-artifacts/claude-skills/integration-pipeline-preview/06-devex-review/02-review-quality

ARCHIVAL KNOWLEDGE — DO NOT USE OR REFERENCE. Superseded by codemap/ + root AGENTS.md. Scheduled for deletion once the pipeline is verified (Phase 6). Manifest: docs/archive/2026-07-ai-artifacts/MANIFEST.md

name: 02-review-quality description: Independently validates the DevEx review against the rubric and gates phase 6 completion. kind: leaf executor: sonnet model: claude-sonnet-4-6

Review DevEx-review quality

Reads the DevEx review and the design draft, applies every check in ../../references/devex-rubric.md, records the per-check verdict to .integration-pipeline/devex-review-rubric.json, and gates phase 6 completion on the result.

This step is the auto-gate that replaces the human PAUSE: when the rubric says pass, the pipeline advances automatically; when it says fail, the pipeline halts here with the rubric record as the audit trail.

Review posture

Assume the artefact under review was produced by a language model (ChatGPT-class output). Treat plausible-looking claims as unverified, expect hallucinated APIs and missed constraints, and refuse to pass anything not directly traceable to the rubric in ../../references/devex-rubric.md and the design draft itself. Strict review is the default — benefit of the doubt goes to the rubric and to the source, never to the producer.

Inputs

• .integration-pipeline/devex-review.md — produced by ../01-produce/.
• .integration-pipeline/integration-design.md — the design draft the review covers.
• ../../references/devex-rubric.md — the eight-check rubric + three carrier rules.
• ../../references/jess-persona.md — for cross-referencing the Jess affordance check.

Plan

1. Load all four inputs → verify: every file exists; refuse on any missing.
2. For each of the eight rubric checks, decide pass / fail / n/a from the evidence in the review + design draft → verify: every check has a verdict + an evidence cite (file path + section/line).
3. For each of the three carrier rules, decide pass / fail → verify: every rule has a verdict.
4. Write the rubric record JSON to .integration-pipeline/devex-review-rubric.json → verify: file exists with checks, carriers, verdict.
5. If verdict == "pass", run pipeline.py complete 6 and print RUBRIC: pass. If verdict == "fail", print RUBRIC: fail plus the failing check names and refuse — do NOT mark phase 6 complete.

Assumptions:

• The reviewer never wrote the DevEx review (independence — different agent invocation).
• Verdicts are derived strictly from the rubric; the reviewer does not invent new criteria.
• Evidence cites must be unambiguous: a section heading + line range or quoted phrase.

If any assumption fails, refuse — do not guess.

Steps

1. Load the DevEx review, the design draft, the rubric, and the Jess persona.

2. Walk each rubric check 1-8 in order. For each, scan the DevEx review for the relevant per-entry verdicts and cross-reference the design draft's Public API. Record pass / fail / n/a plus the supporting cite.

3. Walk the three carrier rules (Simplicity / Surgical / Goal-driven). Record verdicts.

4. Compose the rubric record:

   {
     "phase": 6,
     "checks": [ {"id": "naming-conversions", "verdict": "pass", "evidence": "..."}, ... ],
     "carriers": [ {"rule": "simplicity-first", "verdict": "pass"}, ... ],
     "failing": [ ... ids of failing checks ... ],
     "verdict": "pass" | "fail"
   }
   

5. Write .integration-pipeline/devex-review-rubric.json.

6. If every entry in checks and carriers is pass or n/a, run python3 .claude/skills/integration-pipeline-preview/scripts/pipeline.py complete 6 and print RUBRIC: pass.

7. Otherwise, print RUBRIC: fail followed by the failing check IDs and their cites, then refuse: name the producer step that must re-run and the concrete edits needed.

Output

.integration-pipeline/devex-review-rubric.json written. On pass: phase 6 marked complete; pipeline advances. On fail: pipeline stops; the rubric record names the failures.

Refusals

• Any input file missing → refuse with the missing path.
• A rubric check cannot be decided from the review + design draft alone → refuse: "produce step did not record enough detail to decide check X"; the producer must amend the review.
• Verdict is fail → refuse to call pipeline.py complete 6; print the failure summary and stop.