repoc

repoc = repo doctor. A lightweight CLI that inspects a GitHub or local repository before you clone, fork, install, or contribute to it.

It answers questions like:

• Can I trust this repository?
• What language and framework does it use?
• Is it still maintained?
• Are there suspicious scripts, install hooks, or possible secrets?
• Is it safe to run install/build commands?

Disclaimer: repoc does not prove that a repository is safe. It highlights suspicious patterns, metadata, and maintenance signals that deserve manual review.

Installation

pip install repoc-cli

The PyPI package is repoc-cli; the command it installs is repoc.

Or from source:

git clone https://github.com/GuillaumeYves/repoc
cd repoc
pip install -e ".[dev]"

Requires Python 3.11+.

Usage

# Inspect a public GitHub repository
repoc inspect owner/repo

# Or from a full URL
repoc inspect https://github.com/owner/repo

# Inspect the current working directory
repoc inspect . --local

# JSON output (machine-readable)
repoc inspect owner/repo --format json

# Markdown export
repoc inspect owner/repo --format markdown --output report.md

# Deeper scan — also inspects source files, not just manifests
repoc inspect owner/repo --deep

# Check pinned dependency versions against the OSV.dev vulnerability database
repoc inspect owner/repo --check-deps

# SARIF for GitHub code scanning
repoc inspect owner/repo --format sarif --output repoc.sarif

# Use as a CI gate — exit non-zero if a HIGH (or worse) security finding appears
repoc inspect owner/repo --fail-on high

Common options

Exit codes: 0 ok · 1 --fail-on gate tripped · 2 bad target/usage · 3 GitHub/network error or rate limit · 4 authentication error.

GitHub authentication (private repos)

repoc never writes a credential to disk, a keyring, or any config file — a token lives in memory for a single run and is discarded on exit. You can supply one in whichever way fits your trust model:

# 1. Environment variable / flag (raises the rate limit; reads private repos)
export GITHUB_TOKEN=github_pat_xxx
repoc inspect owner/private-repo

# 2. Pipe a token via stdin — keeps it out of shell history and the process list
echo "$GITHUB_TOKEN" | repoc inspect owner/private-repo --token-stdin

# 3. One-time browser device-code login (nothing is persisted)
repoc inspect owner/private-repo --login

For --login, set REPOC_GITHUB_CLIENT_ID to a public GitHub OAuth App client id (enable Device Flow on the app). The client id is not a secret. If you'd rather not register an app, use a fine-grained, read-only PAT with --token-stdin.

Use in CI / pre-commit

# .github/workflows/audit.yml
- run: pip install repoc
- run: repoc inspect ${{ github.repository }} --fail-on high

The gate only counts security findings (committed secrets, install hooks, risky code patterns). Maintenance, documentation, and scan-coverage notes never fail it.

Sample output

# Repoc Report: https://github.com/owner/repo

## Verdict

- Risk level: **Medium**
- Trust score: **72/100**
- Project type: Backend API
- Primary language: Python
- Detected stack: FastAPI, SQLAlchemy, Pydantic

## Summary

owner/repo appears to be a Backend API written in Python using FastAPI,
SQLAlchemy, Pydantic. The scan produced low-severity findings only.

...

What it detects

repoc has two depth layers. Code-logic analysis (the security rule packs that read your source) currently covers 7 ecosystems. Identification (language + framework + secret scanning) is broader. Secret scanning runs on every text file regardless of language.

Statically analyzed (security rule packs)

Rules run against comment/string-masked source (see Keeping signal high below), so they're tuned to flag real calls, not mentions in comments or strings.

Identified but not yet code-analyzed

repoc detects Go, Rust, Java/Kotlin, C#, C/C++ (and labels ~25 languages in total) and parses their dependency manifests for frameworks (Gin/Fiber/Cobra, Actix/Axum/Tokio, Spring/Hibernate, Laravel/Symfony, …). These files are still secret-scanned, but dedicated code-logic rule packs for them are planned for future releases — today a risky call in a .go/.java/.rs file is not flagged.

Secret scanning (all languages)

Regex-based and conservative: GitHub PATs, AWS access keys, PEM private keys, JWTs, DB URLs with credentials, Slack/Discord webhooks, Stripe secret keys, committed .env files. Matched values are redacted.

Dependency vulnerabilities (--check-deps)

Opt-in. repoc parses pinned versions from lockfiles (requirements.txt, poetry.lock, Pipfile.lock, package-lock.json, Gemfile.lock, Cargo.lock, composer.lock, go.mod) and batch-queries OSV.dev for known advisories. Findings are capped at HIGH severity — repoc can't assess whether a (possibly transitive) vuln is reachable, so it never escalates a whole repo to Critical on a dependency alone. Needs network; skipped under --no-network.

Keeping signal high

• Comment/string aware. Code-pattern rules run against a copy of each file with comments and string literals blanked out, so an eval( mentioned in a docstring or a # curl … | bash example line is not reported. (Secret rules still run on the raw text — secrets live inside strings.)
• Path context. Findings in tests/, examples/, fixtures/, docs/, vendor/, and *.example / *.sample files are down-ranked one severity tier — a token in a test fixture is rarely a live leak.
• Inline suppression. Add # repoc: ignore, # repoc: ignore PY001, or # repoc: ignore-file (any comment syntax) to silence a known-safe match.

Scan coverage

Every report states how many files were actually inspected. If the --max-files cap is hit, or oversized/binary files are skipped, repoc emits explicit COV findings and a Partial scan banner — a capped scan can never look like a clean bill of health.

Default remote scans are metadata-only. repoc inspect owner/repo reads manifests, lockfiles, install hooks, workflows, Dockerfiles, and docs — it does not download the source tree. When a repo has source files that weren't inspected, repoc says so (COV003) and tells you to add --deep, which downloads the repo once and scans the code. Local scans (--local) always read the full source tree.

Maintenance & docs

Repository age, archive status, license presence, open-issue backlog, plus the usual suspects in the repo root: README, LICENSE, CONTRIBUTING, SECURITY, CODE_OF_CONDUCT, CHANGELOG, docs/, examples/.

Scoring

The trust score is a weighted average:

Findings are weighted by category. Committed secrets and npm install hooks are treated as "stop and look" signals and weigh heavily; risky code patterns (eval, subprocess, curl | bash) are "worth reviewing" signals that barely move the score on their own and taper with repetition — twenty subprocess calls in a large codebase do not zero the score the way twenty private keys would.

Risk levels are Low, Medium, High, Critical, Unknown. Any CRITICAL finding, or a HIGH/CRITICAL committed secret or install hook, escalates the risk on its own; a lone HIGH code pattern (e.g. one eval) does not. (Popularity is excluded for local scans, where stars/forks are unknown.)

The full breakdown is included in every report so the score is auditable.

Development

git clone https://github.com/GuillaumeYves/repoc
cd repoc
pip install -e ".[dev]"

ruff check .
pytest

The project layout:

repoc/
  cli.py            Typer entry point
  auth.py           Ephemeral GitHub auth (device flow, stdin token)
  github_client.py  Minimal GitHub REST client + tarball fetch
  models.py         Pydantic models
  scoring.py        Score aggregation + risk classification
  analyzers/        One module per analysis dimension (incl. coverage)
  rules/            Regex rule packs + comment/string masking
  renderers/        terminal / markdown / json
tests/
.github/workflows/  ci.yml, release.yml

Releasing

Bump __version__ in repoc/__init__.py, commit, then:

git tag v1.0.0
git push origin v1.0.0

The release workflow builds wheels, creates a GitHub release, and publishes to PyPI via Trusted Publishing.

Roadmap

• Code-logic rule packs for Go, Rust, Java, and C# — the languages repoc identifies today but doesn't yet analyze for risky calls.
• Heuristics for typosquat detection on package names.
• Resolve --check-deps for range-only manifests (e.g. a bare package.json).
• Pluggable rule packs (drop-in .toml/.py rules).
• Caching layer for repeated GitHub queries.

License

MIT — see LICENSE.