| name | repo-review |
| description | Perform a full-coverage, strict repository evaluation. Supports quick and default report detail; default is the comprehensive strict review. |
| argument-hint | [quick|default] |
| allowed-tools | ["Read","Bash","Grep","Glob","Agent"] |
Repository review
Evaluate an entire repository against concrete engineering evidence. Every section starts at 0
points and earns percentage credit only from files, commands, history, or live configuration that
the reviewer actually inspected. Assign the letter grade only after calculating the percentage.
Modes
default (implicit): full coverage, strict grading, detailed evidence and remediation report.
quick: the same full evaluation and strict grading, but a compact report with only decisive
evidence, blockers, and the top three actions.
There is no lenient mode. “Quick” changes output detail, not coverage or standards.
Host compatibility
Use the host's native subagent mechanism when available. Assign one independent section to each
subagent, up to the host's safe concurrency limit. If delegation is unavailable, evaluate sections
sequentially in the current agent. Never require a branded model or vendor-specific tool call.
Required workflow
1. Establish scope
- Confirm the repository root with
git rev-parse --show-toplevel when Git is present.
- Inventory every tracked and relevant untracked file. Exclude only generated dependency caches,
VCS internals, and build outputs that are clearly non-source.
- Record language, framework, package, deployment, and agent-host surfaces.
- Read all repository guidance (
AGENTS.md, linked host pointers, ADRs, contribution/security
policy) before grading.
2. Read all evidence
Full coverage is mandatory. Partition the inventory across the 15 sections in the criteria reference
so every file is opened by at least one reviewer. Read all source, tests, manifests, workflows, and public docs.
Inspect relevant Git history and live GitHub configuration when access exists. Report any read gap;
do not silently sample.
Read and apply every criterion in references/criteria.md. If a delegated
section fails, times out, or samples its bucket, redispatch it or complete it sequentially before
grading. Reporting an available file as unread is not a substitute for completing the review.
Run safe, repository-defined checks when dependencies are available. Capture commands and exit
codes. A documented command is not evidence until it runs successfully.
3. Score from zero
Begin every section at 0%. Add earned points criterion by criterion, total the percentage, and
only then assign a letter grade. Do not use a provisional letter grade as the scoring baseline.
Use this scale without rounding up:
| Grade | Score | Evidence standard |
|---|
| A | 93–100 | Near exemplary; no critical/major issues and at most two minor issues. |
| B | 80–92 | Strong; no critical issues and at most one major issue. |
| C | 70–79 | Functional but contains several material gaps. |
| D | 60–69 | Poor; fundamental practices or contracts are broken. |
| F | 0–59 | Missing, unsafe, or fundamentally unreliable. |
Absence of evidence earns no credit. Intent, TODOs, filenames, and static badges do not prove a
criterion. Mark a criterion N/A only with a project-specific reason.
Establish the product-maturity baseline before scoring. Versioning, migration, and backwards-
compatibility criteria apply only to established supported releases or public contracts. When the
maintainer explicitly identifies the evaluated state as the first supported release, record that
assumption and treat earlier bootstrap interfaces as N/A for compatibility scoring.
4. Evaluate all sections
Evaluate all 15 authoritative sections in references/criteria.md. That
file owns section names and criteria; do not maintain a second section registry here.
Read and explicitly apply every decision rule in
../../docs/policies/development.md: KISS, YAGNI, TDD, DRY,
SOLID, modularity, least astonishment, durable-artifact discipline, and behavior-first testing.
Flag prose-string/document-count tests, documentation snapshots, flaky implementation-detail
assertions, generated documentation, manual changelogs, duplicated registries/catalogs/inventories,
and unrelated generated files unless a demonstrated product consumer requires them.
5. Calculate the overall score
Weights: Structure 2%, Documentation 7%, Architecture 15%, Source quality 15%, Testing 12%, CI/CD 8%, Dependencies 3%, Security 12%, Reliability 10%, Planning 2%, Agent tooling 5%, Packaging 3%, Developer experience 2%, API/CLI 2%, Governance 2% (100% total).
Verdicts:
- GO: score ≥80, no critical issues, at most three major issues.
- CONDITIONAL GO: score ≥65, at most two critical issues with concrete remediation.
- NO-GO: anything else.
Output contract
Default report
Produce:
- Executive scorecard for all 15 sections and weighted overall score.
- Per section: grade, evidence reviewed, strengths, severity-ranked findings, missing criteria,
and principle compliance with exact
path:line citations.
- Consolidated critical/major/minor list without duplicates.
- Development-principles matrix.
- Coverage report: files inventoried, files read, read errors, commands run.
- GO/CONDITIONAL GO/NO-GO verdict and ordered remediation plan.
Quick report
Produce:
- The complete 15-section scorecard and weighted verdict.
- Critical and major findings with exact citations.
- Coverage gaps and failed checks.
- The top three remediation actions.
Do not omit a section merely because quick mode was requested.