codejury 0.16.0

AI code security review: an adversarial diff-audit engine and an agent-driven whole-repo review methodology, with security knowledge as rich rules

codejury

AI code security review, in two paths matched to their nature:

• Diff review (coded): audit a pull request's diff for newly introduced, exploitable risks. A single balanced LLM call (the default), or an adversarial Finder/Challenger/Judge pass that trades roughly 3x the cost for extra recall on subtle, cross-cutting flaws.
• Whole-repo review (agent-driven): a methodology an interactive agent (Claude Code, Codex) runs to map a codebase's attack surface, trace inputs to sinks across files, verify issues with a real PoC, and iterate over rounds with a persistent memory. Too large for a single LLM call, so codejury ships the methodology and scaffolds the workspace rather than running a pipeline.

Security knowledge lives in rich rules (codejury/data/rules/*.md, with per-language vulnerable/secure examples), injected into the audit prompt, not buried in code.

Install

pip install codejury                 # core
pip install "codejury[anthropic]"    # or [openai] / [litellm] for a backend

Diff review

# audit a diff file
codejury review diff --diff-file changes.diff

# audit a git range in a repo
codejury review diff --repo /path/to/app --git-range origin/main...HEAD

# from stdin
git diff HEAD~1 | codejury review diff

# adversarial mode: Finder + Challenger + Judge (extra recall on subtle flaws, ~3x cost)
codejury review diff --diff-file changes.diff --mode adversarial

# CI gate + SARIF
codejury review diff --diff-file changes.diff --format sarif --fail-on high

Configure a backend with --provider/--model/--api-key/--api-base or the CODEJURY_API_KEY / CODEJURY_MODEL / CODEJURY_API_BASE environment variables. codejury review diff --dry-run exercises the engine with a mock provider and no key (it uses a built-in demo diff when you do not pass one).

Choosing a model and mode

Detection quality is dominated by the model first, then the mode. On real-diff probes:

• A strong model (Claude Sonnet tier) in standard mode caught every planted vulnerability with near-zero false positives. A weaker model raised false positives in both modes, so the model is the lever that matters most.
• Adversarial mode did not lower false positives over standard on those probes and costs ~3x. Reach for it for extra recall on subtle, cross-file logic, not as a false-positive reducer.

Default to standard mode with a strong model (set it with --model or CODEJURY_MODEL). False positives are held down by the do-not-report list and the post-filter, not by the mode.

Use in CI (GitHub Actions)

Audit every pull request and surface findings in the code scanning tab. Copy examples/codejury-pr-review.yml into .github/workflows/, add a CODEJURY_API_KEY repo secret, and it will:

1. diff the PR against its base (--git-range origin/<base>...HEAD),
2. write SARIF and upload it with github/codeql-action/upload-sarif,
3. fail the check on a HIGH or CRITICAL finding (--fail-on high).

The job makes one model call per PR (standard mode); the SARIF is uploaded even when the gate fails, so findings always show up on the PR.

Whole-repo review

codejury review repo /path/to/your/repo

This scaffolds a review workspace (entrypoints/, issues/, analysis/, and a security-review-memory.md), seeds the entrypoint inventory from a deterministic scan, and prints the methodology. Run it with an interactive agent: it reads the methodology and the rules, maps the attack surface, traces inputs to sinks across files, records high-confidence issues with a PoC, and asks you to confirm credentials or false positives along the way. Nothing runs against production.

Findings

Each finding carries a file and line, a severity and category, a concrete exploit scenario, a recommendation, and a confidence. A false-positive filter drops test/mock-path and low-confidence noise; the model is also told not to report dependency CVEs, style notes, speculation, or config-leak-only risks.

Extending

Add a vulnerability class by dropping a new codejury/data/rules/<class>.md with the standard frontmatter (title, impact, tags, triggers) and vulnerable/secure examples. It is data; no code change needed.