Review

Multi-model review without runtime interception.

• One shared contract per review: .reviews/<review_id>/contract.json
• Rounds accumulate: .reviews/<review_id>/rounds/rNNN/

Interaction Order

1. Run init. Report detected dev tool and confidence.
2. Report risk analysis.
3. Report recommended reviewers.
4. Ask user to confirm or override.
5. Run review.

Low-confidence detection: say so and continue.

Workflow

1) Init

python3 skills/nobs-review/scripts/init_review.py --task "<what to review>"

Override tool detection:

python3 skills/nobs-review/scripts/init_review.py --task "<what to review>" --dev-tool claude

Refresh existing review (keep history):

python3 skills/nobs-review/scripts/init_review.py --id <review_id> --task "<updated scope>" --refresh

Init detects the current dev tool, analyzes diff risk, pre-enables recommended reviewers in contract, and sets .reviews/.active_review_id.

2) Confirm reviewers

Check .reviews/<id>/contract.json:

• reviewers[].enabled
• reviewers[].command_template

Accept recommendation or pick your own.

3) Run round

python3 skills/nobs-review/scripts/run_review.py --id <review_id> --round auto

Or use active id:

python3 skills/nobs-review/scripts/run_review.py --round auto

Limit reviewers:

python3 skills/nobs-review/scripts/run_review.py --id <review_id> --round auto --reviewers codex_high,claude_opus

4) Merge findings

python3 skills/nobs-review/scripts/merge_review.py --id <review_id> --round latest

--round auto treated as latest in merge.

Outputs per round: merged.md, resolution.md in .reviews/<id>/rounds/rNNN/. Latest pointers at review root (merged-latest.md, resolution-latest.md) are ephemeral — overwritten on next merge.

For persistent checklist edits, use round-scoped resolution.md.

5) Review auto-fix candidates

After merge, check auto-fix-candidates.md in the round directory (or auto-fix-candidates-latest.md at review root).

Findings listed there satisfy both:

• fix_determinism = high — exactly one right fix, no judgment needed
• fix_scope = local — change confined to 1–3 lines in one function

Apply each fix, mark Applied: [x], commit, then run a new round to verify:

python3 skills/nobs-review/scripts/run_review.py --round auto
python3 skills/nobs-review/scripts/merge_review.py --round latest

Findings that do not appear in auto-fix-candidates.md require human judgment and are tracked in resolution.md.

The auto_fix_threshold in contract.json → policy controls the boundary:

"auto_fix_threshold": { "determinism": "high", "scope": "local" }

Raise determinism to medium or scope to cross-file to widen the auto-fix set.

6) Fix and re-review

python3 skills/nobs-review/scripts/run_review.py --id <review_id> --round auto
python3 skills/nobs-review/scripts/merge_review.py --id <review_id> --round latest

Creates r001 -> r002 -> ... preserving history.

Rules

• Missing reviewer CLI: fail, don’t fake.
• Invalid JSON output: keep raw, mark parse failure.
• No single round is final truth; compare across rounds.

Output Contract

• references/schema.md
• references/providers.md

Notes

• Deterministic logic lives in scripts; this skill is thin by design.
• High-risk diffs: use at least two strong reviewers.
• User decides final reviewer set; recommendation is advisory.