Skip to content

Adonis0123/adonis-skills / skills/open-code-review-loop

open-code-review-loop

Unverified

Run a bounded OCR-delegation review-fix-re-review loop on a Git workspace until current evidence validates NO_FINDINGS. Use only when the user wants verified findings fixed and independently re-reviewed until clean with codex, claude-code, grok-build, or cursor-cli. Do not use for one-shot open-code-review-delegate or read-only review. A full loop request asks at most once for an unnamed Reviewer (recommending a product other than the host), then runs host mutation checks without further confirmation. Fail closed on missing ocr, reviewer mutation, skipped files, malformed output, stale evidence, or exhausted rounds.

Open Code Review Loop

ocr delegate is LLM-free: it selects files and resolves rules. Only a validated reviewer result with full selected coverage may produce NO_FINDINGS. OCR never decides that code is clean.

Fast Path

Do these in order. Later references stay closed until a step needs them.

  1. Git repository? Otherwise stop.
  2. which ocr — missing → MISSING_DEPENDENCIES. Do not open adapter-contracts / review-contract / fix-contract, and do not install OCR without authorization. Stop before any Reviewer call.
  3. Resolve Reviewer/Fixer product IDs, then open only those product sections in adapter-contracts.md.
  4. Open review-contract.md + review-schema.json only in Phase 2.
  5. Open fix-contract.md + fix-schema.json only in Phase 4 for independently verified Fix rows.
  6. Historical --commit / --from/--to → HUMAN_GATE before any fix; do not load the fix contract.

Trigger and exclusions

Use this Skill only for an OCR-delegation loop (ocr delegate / open-code-review) on a Git workspace with a supported reviewer (codex, claude-code, grok-build, cursor-cli) when the request includes both fixing verified findings and independent re-review until validated NO_FINDINGS or CLEAN. Merely asking for one ocr delegate / open-code-review-delegate pass is insufficient; route that request to open-code-review-delegate. Skip other one-shot or read-only review, non-Git files, commit/push/deploy, and destructive work.

Wrong skill. Packet protocol, .review-handoff, review-loop run, or a Grok consult → agentic-review-handoff. Redirect; do not start ocr delegate. Named path + architecture scan-fix-rescan until no architecture findings → architecture-hardening-loop. Do not claim OCR CLEAN. Do not copy those skills' packet or scanner logic.

Inputs and defaults

Resolve before the first model call:

  • Repository: current Git root; never infer a broader repo
  • Target: workspace changes; also OCR --from/--to and --commit
  • Reviewer: user-selected product; when none is named, ask once and recommend a product other than the visible host; always use a real independent read-only product session
  • Fixer: current visible host. An external Fixer needs a user-authorized isolated checkout that becomes $REPO before round 1
  • Paths/excludes: OCR preview; preserve user exclusions every round. default_path is OCR's built-in non-review scope and user_exclude records a supplied selector, so neither needs a second confirmation. OCR-excluded .md/.mdx files become mandatory supplemental Reviewer scope. Other reasons remain unaccepted unless the user already accepted that exact reason
  • Round budget / deadline: 3 rounds (a ceiling, not success permission) plus only the user grants recorded in the round ledger; 10 minutes per Reviewer/Fixer call. Run the loop through this Skill: calling a review CLI repeatedly outside it loses the round ceiling
  • Background: user requirement or none; pass through preview/rule/prompt

Once the user has requested the full fix-and-re-review loop, explicit skill invocation authorizes the reversible local workspace loop. Do not ask whether to start, reconfirm OCR scope, or accept supplemental Markdown review. One named product without roles → that product is Reviewer, current host is Fixer. No named product → ask once with the host question tool, recommending an installed capability-safe product other than the visible host (a model reviewing its own host's work is the weakest independence); the host product stays a valid answer in a separate session. Without a question tool, or when the user asked for defaults, pick that recommendation and record it. If an auto-selected Reviewer cannot satisfy the read-only adapter, try the next installed capability-safe product and record the fallback; never replace a product the user explicitly named. Missing ocr, historical-target fixes, external actions, and destructive or security-sensitive choices still fail closed. Same product in both roles always means separate sessions.

Capability preflight

Fast Path already covers Git, ocr, and product IDs. Before editing, also inspect installed product help: a binary name does not prove read-only review, structured output, scoped writes, or session recovery. Reviewer inspects without editing. External Fixer: verify the isolated checkout holds the intended snapshot, freeze that Git root as the only $REPO, and never review one worktree while fixing another. No isolated target → HUMAN_GATE (no patch-transfer protocol). Do not auto-create a worktree.

Record allowed source/test paths, prohibited Git/external actions, and already accepted exclusion reasons. OCR default_path entries are not Reviewer coverage and do not need a user gate; keep them in the evidence report. Run relevant tests as host verification even when OCR excludes test files, but never count test execution as Reviewer coverage. The builder promotes unsupported_ext Markdown to supplemental_reviewable_files, captures its content, and applies supplemental_rules; it remains visible in raw OCR exclusions but no longer needs a question. Other unsupported extensions, binary, or a missing reason remain incomplete coverage. Partition sessions by repository, loop ID, role, and canonical product ID — never resume a Reviewer as a Fixer. Reviewer invocations use the strongest available read-only controls plus dontAsk; configured hooks/MCPs or lack of an OS-perfect sandbox do not alone create a user gate. Freeze the subject Git identity immediately before the call and recompute it afterward. Any mutation caused during review returns UNVERIFIED: REVIEWER_MUTATED_SUBJECT, with the actual diff reported and no automatic reset or retry. If a named product lacks structured output or cannot run non-interactively, return HUMAN_GATE; an automatic selection tries the next installed product without asking.

Loop contract

Freeze before round 1:

OCR Review Contract
- Repository / target:
- Reviewer / session policy:
- Fixer / session policy:
- Scope / exclusions:
- Accepted exclusion reasons:
- Background / acceptance checks:
- Write boundary:
- Round budget:
- Per-call deadline:
- Git and external actions: none

Reviewer owns findings and the clean verdict. Fixer owns source edits and verification. Visible host owns OCR evidence, schema validation, triage, round accounting, and the final result.

Phase 1: Build current evidence

ROUND_DIR="$(mktemp -d)"
python3 <skill-dir>/scripts/build_review_bundle.py \
  --repo "$REPO" \
  --output "$ROUND_DIR/bundle.json" \
  [--from-ref <ref> --to-ref <ref>] \
  [--commit <hash>] \
  [--exclude <patterns>]... \
  [--allow-excluded-reason <reason>] \
  [--rule <rule.json>] \
  [--background <text> | --background-file <path>]

Keep ROUND_DIR outside the repo so the next preview cannot select the bundle and permanently drift evidence. --exclude may be repeated; the script combines values into OCR's list and keeps that selection every round. --allow-excluded-reason is not authority: pass it only after recording the user's explicit acceptance of that exact non-default reason. The builder accepts default_path and user_exclude automatically and routes Markdown unsupported_ext entries into mandatory supplemental scope; do not pause at round 0 to reconfirm either behavior. Zero OCR reviewable files, zero supplemental reviewable files, and zero unaccepted exclusions → CLEAN with 0/0 coverage. Any remaining unaccepted exclusion → UNVERIFIED: INCOMPLETE_COVERAGE. Do not invoke an AI just to manufacture a verdict when the combined selected scope is empty.

Builder already fail-closes on (do not re-implement; read scripts/build_review_bundle.py only if the builder fails):

  • mid-build HEAD / binary-diff / untracked fingerprint drift
  • missing captured content vs a real zero-byte file (empty_file: true)
  • first-parent merge selection and immutable ref resolution
  • in-repository output
  • background argv redaction and a private 0600 snapshot

Prefer --background-file for non-trivial or sensitive background. Do not commit the bundle. validate_round.py recomputes the canonical digest, so editing bundle content (repository included) while keeping an old evidence_id also fails closed.

Phase 2: Invoke the Reviewer

Give the Reviewer the frozen contract, background, whole bundle.json (including evidence_id), nearby read-only context, non-mutating checks, and the JSON contract from Fast Path step 4. The selected review set is the union of reviewable_files and supplemental_reviewable_files; apply both rules and supplemental_rules. Every selected (path, status) appears exactly once as reviewed or skipped. Emit only evidence-backed, actionable findings. Tests that need temporary writes belong to host verification and are not a reason to weaken the Reviewer sandbox.

The round ledger is not a file you choose. validate_round.py keeps one per repository checkout at ~/.local/state/open-code-review-loop/ledgers/<repository-hash>.json and prints the path as ledger. That directory is fixed under the home directory that the system account database lists for the current user: no argument or environment variable (XDG_STATE_HOME, HOME) moves it, and a sandboxed host needs it as an extra writable directory. When the account database has no entry for the user, or Python has no fcntl file lock (native Windows), the script returns UNVERIFIED (exit 2) and records nothing; there is no fallback to HOME and no unlocked mode. bundle.repository is part of evidence_id, and the script asks Git which checkout it is and whether bundle.base_sha is its HEAD, so a stale bundle, or one edited to point at another checkout (a worktree or clone on the same HEAD included), is refused instead of getting another ledger. Each checkout has its own ledger and the script does not link them, so build every bundle in the frozen $REPO; never build in another worktree or clone to continue a loop. A handoff, a resumed session, a new ROUND_DIR, or a new commit lands on the same ledger and the rounds already used; never move, delete, edit, or recreate it. Omit --history: any other path is refused.

Rounds add up across commits; the ledger lists every base_sha it saw. Only a validated NO_FINDINGS closes the loop, and the next evidence then starts a new budget. A loop that ended any other way stays open: the next loop in that checkout continues its count until the user grants rounds.

Before every Reviewer call, run the gate and chain the call with &&, never ; or a new line, so a refused gate (exit 2) cannot be followed by a review:

python3 <skill-dir>/scripts/validate_round.py --check-call \
  --bundle "$ROUND_DIR/bundle.json" \
  && <Reviewer command from the product's adapter section>

A passed gate records the round as started before the Reviewer runs. Run it only immediately before a real Reviewer call, never to look at the count (earlier output and the ledger file show it): a call that starts and is then interrupted has used its round, and a relay session sees it. Passing the gate again for the same evidence ID stays one round; a refused gate writes nothing.

A fresh evidence ID consumes one round of the budget; one same-session format correction uses the same round. A timeout is UNVERIFIED and its round stays used. Validate each raw envelope so the same ledger entry gets the result and extraction failures are counted:

python3 <skill-dir>/scripts/validate_round.py \
  --bundle "$ROUND_DIR/bundle.json" \
  --review "$ROUND_DIR/raw-review.json" --product "$REVIEWER_PRODUCT" \
  --output "$ROUND_DIR/validation.json" \
  [--session-id "$RECORDED_SESSION_ID"]

--output must not be the ledger, its .tmp or .lock file, or any path inside the ledger state directory: that is refused (exit 2) and nothing is written. Keep it in ROUND_DIR. An error raised before that check finishes (an unreadable bundle, no account home) is printed to stdout only, so a validation.json left by an earlier run may be stale: trust the exit code and this run's stdout, not the file, when the validator exits non-zero.

The validator reuses extract_product_output.py. Its gate stops a second malformed result as UNVERIFIED and refuses a round beyond the budget: format correction supplies no new evidence. A review obtained after a refused gate is rejected the same way and never recorded. Every refusal carries a fix or errors text; follow it instead of working around the ledger.

Phase 3: Detect evidence drift

Rebuild the bundle with the same target and exclusions. Compare evidence_id. Equal → the result covers the current snapshot. Different → discard the verdict, record evidence-drift, and review the new bundle. Never carry findings or NO_FINDINGS across snapshots.

Phase 4: Triage and fix findings

For a valid FINDINGS result:

  1. Independently open each cited path and verify the evidence.
  2. Classify each item as Fix, Reject, or Human decision.
  3. Give the Fixer only verified Fix rows, target paths, required fix, acceptance check, current evidence ID, and the write boundary. Load the fix contract only now (Fast Path step 5).
  4. Extract with extract_product_output.py, then validate_fix_result.py. One same-session correction for malformed output; a valid result is still only a claim.
  5. Fixer re-reads current code and applies the smallest coherent change.
  6. Host recomputes Git status/diff, derives actual changed paths, rejects out-of-scope writes, and runs listed checks. Never trust claimed FIXED, changed_paths, or the Fixer's test summary.
  7. Record rejected findings with counter-evidence. Do not edit code to satisfy a known false positive.

Fixer does not commit, push, merge, rebase, deploy, send messages, or edit outside the frozen boundary. External Fixer writes only the authorized isolated $REPO. Direct writes to the user's original worktree need path-scoped enforcement, not a later diff check; the four documented CLI adapters do not meet that bar. Range/commit review always returns HUMAN_GATE before fixing — another writable branch cannot prove those edits against the immutable target. Any source or test edit invalidates the prior bundle, validation, tests, and verdict; return to Phase 1.

Fixer timeout or lost delivery after it could have written → recompute the real Git diff and return UNVERIFIED: DELIVERY_UNKNOWN_WITH_MUTATION. Do not retry blindly or reset user changes. FIXED with unchanged Git/evidence and a still-reproducing finding → FIXER_NO_MUTATION. Delivery is known, so allow one same-session correction. A repeated no-mutation result is UNVERIFIED, never CLEAN.

Phase 5: Re-review and stop

Resume the Reviewer when the product supports reliable session recovery; otherwise start a fresh read-only session with the finding ledger and new bundle. The Fixer never supplies the terminal verdict.

  • CLEAN: validated NO_FINDINGS with 100% combined OCR + supplemental coverage, zero skipped files, zero unaccepted exclusions, and no drift; or the bundle proves reviewable_files, supplemental_reviewable_files, and unaccepted_excluded_files are all empty (no Reviewer invoked).
  • HUMAN_GATE: real product decision, writable-target choice, unavailable non-interactive capability, repeated disagreement, or round extension.
  • MISSING_DEPENDENCIES: OCR or a required product capability unavailable before edits.
  • UNVERIFIED: ambiguous delivery, malformed output after one correction, failed checks, repeated drift, or stale required evidence.

Round-budget exhaustion is never CLEAN, and only the user extends the budget. The extension exists once it is recorded, with the user's words verbatim:

python3 <skill-dir>/scripts/validate_round.py --bundle "$ROUND_DIR/bundle.json" \
  --grant-rounds <N> --granted-by "<the user's own words about extra rounds>"

The budget is 3 plus the grants recorded in the open loop. A host's unattended prompt (for example "go on with the recommended option" / 「按推荐继续」), a handoff note, or your own judgment is not a grant; only what the user said about adding rounds is. The script refuses the obvious non-quotes: a placeholder (<...>, TODO, TBD), blank or invisible text, and a bare go-ahead such as 「按推荐继续」 or 「都按推荐来」. Text that passes that check is still a grant only when the user said it. A run is unattended when sideby is on PATH and sideby unattended status --json reports "on": true, or the host has stated that nobody is present and questions are refused. A missing or failing sideby settles nothing; without that host statement the user is present. Probe when the last budgeted round returns FINDINGS, before Phase 4, and again when the gate refuses a call.

  • User present: report the remaining ledger and ask whether to authorize another bounded set of rounds; record a yes with the command above before the next gate.
  • Unattended: do not ask. Skip Phase 4 for the last budgeted round's findings, since no round is left to re-review an edit; a fix already applied when the gate refuses is reported as unreviewed. Record one item for the user with sideby unattended defer "<N> OCR findings open on <branch> after <rounds_used> rounds" (only the finding count, the branch or the short SHA on a detached HEAD, and rounds_used from the gate output); without sideby the final report carries it alone. Code with open findings is not committed or merged by the surrounding task, which continues only with work that does not depend on it and lists the item in its final report.

Output contract

OCR Review Loop Result
- Result: CLEAN | HUMAN_GATE | MISSING_DEPENDENCIES | UNVERIFIED
- Repository / target:
- Scope / exclusions:
- Reviewer / Fixer:
- Rounds used / budget: <rounds_used>/<round_budget>; grants: none | <N> by "<granted-by words>"
- Round ledger:
- Final evidence id:
- Coverage: <reviewed>/<OCR + supplemental>; skipped: <count>; unaccepted excluded: <count>
- Fixed / Rejected / Open findings:
- Deferred user decision: none | <item and where it was recorded>
- Verification:
- Sessions / recovery:
- Git and external actions: none

Reviewer-backed CLEAN includes the final validate_round.py result and the same-evidence comparison. Zero-file CLEAN includes the bundle summary proving OCR scope, supplemental scope, and unaccepted exclusions are all empty; no review response exists on that path. Every other state names the exact stop point and does not claim OCR or the AI approved the current code.

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 467kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall