Deep Review

Deep Review

Review at CodeRabbit grade with no file cap and one assertive posture: funnel the diff, discover root/nested project instructions and relevant local skills, shard defects and polish into independent cohorts, fan out reviewers, then merge with complete hunk/rule accounting. Defects require causal evidence and control the verdict; advisories require a concrete improvement and always remain visible.

Steps 1–4 drive an idempotent artifact pipeline under <out>: every stage gate is a bundled-script exit 0, valid agent outputs are never re-run, and an interrupted round resumes by re-running the same commands.

<skill-dir> below means the directory containing this SKILL.md; run every bundled command from the repo root.

Inputs (all optional)

Repo config — .deep-review.yaml

Optional repo-root file, the skill-native config standard. Any key absent there falls back to its .coderabbit.yaml counterpart (reviews.*), so repos migrating from CodeRabbit work unconfigured. Top-level keys, all optional:

The manifest builder resolves path_filters into manifest.json; the knowledge stage ingests path_instructions together with project instructions and skills.

Hard rules

• Source is read-only and frozen: the manifest pins worktree_snapshot, and run_jobs.py / render_review.py refuse a drifted checkout. Writes go only to <out>, .deep-review/ state, and — with --publish — the target PR.
• No file-count cap: a large selection means more cohorts, never a skipped or silently truncated review. Every selected file lands in exactly one cohort.
• Every defect starts with Premise → Path → Verdict; every advisory starts with Premise → Improvement → Fix. Investigated rejections remain visible in the suppression ledger.
• Every selected hunk line receives both defect and polish coverage. Every bound rule receives an explicit compliant/violated/not-applicable assessment.
• Run the repo's linters first and record every overlapping candidate as linter-overlap rather than reporting it again.
• Cite rubric rules verbatim with their source path; severity comes from the taxonomy, never inflated.
• Publishing needs --publish or the user's explicit go-ahead in this session; otherwise the review stays local.
• Every review ends with a SHIP / FIX_BEFORE_SHIP / REWORK verdict derived by render_review.py and stated only after that script exits 0.
• External --subagent runtimes spend compozy exec credit.

Procedure

Step 1: Funnel — build the manifest

1. Run the bundled manifest builder (bootstrap helper; reads the repo and gh, writes only under --out):

   python3 <skill-dir>/scripts/build_manifest.py --out <out> \
     [--pr N | --base REF | --staged | --worktree] [--files p1,p2] [--full]
   

   It resolves repo path filters, detects generated / trivial / renamed files, scopes to the incremental delta when prior state exists, and pins the source-freeze snapshot.

2. Read the printed summary. For --pr, the manifest base is the merge-base of the fetched PR base/head, so base-only changes stay outside the review. If the head is missing, run the printed fetch command and retry; if the base/history is missing, fetch it before retrying.

Done when: <out>/manifest.json exists, every changed file is accounted for as selected, ignored(reason), or skipped(reason), and every selected file carries its hunk list (the units of judgment and the publish anchors).

Step 2: Knowledge + plan — project rules, cohorts, walkthrough

1. STOP. Read <skill-dir>/references/context-pack.md and <skill-dir>/references/taxonomy.md in full before extracting rules or defining reviewer lanes. Run the bootstrap helper (reads the repo, writes only under <out>):

   python3 <skill-dir>/scripts/build_knowledge.py --out <out>
   

   Read every source left pending in <out>/rules.template.json in full, including direct references of selected project skills. Write <out>/rules.json with every source marked applied or not-applicable (reason required), then extract verdict-bearing rules verbatim with scope globs. Assemble <out>/context-pack.md and run/fold the detected linter lanes.

2. Read <skill-dir>/references/orchestration.md (cohort rules, sweep triggers) and <skill-dir>/references/output-contracts.md (walkthrough anatomy, effort scale) in full. Write <out>/plan.json — broad cohorts of up to <max-cohort-files> files (default 200) / ~15,000 changed lines plus any sweep whose trigger fires — and <out>/walkthrough.md.

3. Run the bootstrap plan gate (reads repo artifacts, writes only under <out>):

   python3 <skill-dir>/scripts/build_jobs.py --out <out> \
     [--max-cohort-files N]
   

   It rejects incomplete source accounting, proves defect ownership, derives broad polish cohorts (≤200 files / 15,000 changed lines), injects bound rules into every lane and sweep, and materializes <out>/jobs.json.

Done when: build_jobs.py exits 0, every discovered source has an audited decision in rules.json, context-pack.md lists applied source/rule and linter outcomes without copying the full registry, and walkthrough.md satisfies its contract.

Step 3: Fan-out — parallel review

Execute <out>/jobs.json with the mutating runner and engine contract loaded in Step 2. When --subagent is not native, read <skill-dir>/references/subagent-runtimes.md in full before execution. Completion is engine-independent — re-dispatch whatever is listed as pending/invalid until exit 0:

python3 <skill-dir>/scripts/run_jobs.py --out <out> --validate-only

Done when: run_jobs.py --validate-only exits 0 — every defect, polish, and sweep output matches the schema and completely accounts for assigned hunks and rules.

Step 4: Merge + report

Run the bootstrap merger, mutating state/report renderer, and bootstrap HTML hydrator:

python3 <skill-dir>/scripts/merge_findings.py --out <out>
python3 <skill-dir>/scripts/render_review.py --out <out> [--rework "<structural rationale>"]
python3 <skill-dir>/scripts/render_html.py --out <out>

merge_findings.py emits <out>/findings.json plus <out>/review-stats.json, deduplicates both result classes, reconciles rounds, and fails unless every selected hunk line has defect and polish coverage. render_review.py derives the verdict from defects only. render_html.py shows defects, advisories, suppressions, and coverage separately in <out>/review.html.

When ReportFindings is available, report defects first and every advisory afterward. The user-facing summary states the verdict, defect/advisory counts, every Critical/Major defect, coverage status, and artifact paths.

Done when: render_review.py and render_html.py exit 0 and the final message states the verdict, every Critical and Major defect, and the review.html path.

Step 5: Publish (only with --publish)

1. Read <skill-dir>/references/publish-github.md in full and execute its recipes: upsert the walkthrough, publish every anchorable in-diff defect and advisory inline, keep only unanchorable/outside-diff results in the body, and edit resolved prior-round comments.

Done when: the PR shows the updated walkthrough and the new review, and both URLs are cited in the final message.

Step 6: Learnings

1. state.json was already written at Step 4. When the user — or a PR reply — rebuts or dismisses a result, read <skill-dir>/references/state-and-learnings.md in full, distill the correction into .deep-review/learnings.md, and mark that fingerprint dismissed in the state ledger.

Done when: every user correction from the session is captured as a learning or explicitly declined.

Incremental rounds

With prior state (or fingerprints recovered from the PR thread), Step 1 scopes to commits since the last reviewed head and archives the prior round's artifacts under <out>/rounds/. Unresolved prior results re-surface once under Duplicates; dismissed fingerprints stay suppressed; resolved ones receive the ✅ edit in publish mode. --full reviews the whole diff again. Each round's Step 4 regenerates <out>/review.html, so a browser tab left open on it tracks the rounds by itself.

Error handling

• --pr or --publish without a passing gh auth status → stop and name the gap; publishing by any other transport is out of scope.
• Workflow tool unavailable → automatic Agent fallback; record the mode in walkthrough.md's Review details.
• External --subagent failure (model not available, missing/invalid output file, non-zero exit) → apply the failure handling loaded in Step 3.
• Empty selection after the funnel → report "nothing reviewable" with the manifest counts; write no findings.
• A linter lane unavailable → proceed and state in review.md that overlap suppression did not run for that lane.
• A bootstrap gate failing (build_manifest.py, build_knowledge.py, build_jobs.py, merge_findings.py) → stop and surface stderr. Missing knowledge accounting or incomplete defect/polish coverage is a review failure, not a warning.
• run_jobs.py exit 2 (blocked) → a provider limit interrupted the fan-out; valid outputs are preserved and <out>/run-blocker.json lists the pending jobs — resume by re-running the same command once the limit clears. Providers that signal limits differently need extra --block-on patterns.
• run_jobs.py exit 3 or a render_review freeze failure → the checkout drifted mid-round; findings would anchor to stale lines. Restart from Step 1 — the round increments and prior artifacts are archived.
• More than 75 publishable results → use the Step 5 batching contract.

Bundled implementation

assets/PROMPT.md, assets/findings.schema.json, and assets/REVIEW_UI.html are author-tooling sources consumed by the bundled scripts; agents use their rendered prompt/schema/report artifacts rather than loading these assets directly. <skill-dir>/scripts/_common.py is a read-only library imported by the CLIs and is never invoked directly.