production-code-audit

Audit a codebase the way years of production use would: find the show-stoppers and the small edge-case bugs, prove each one, then fix the ones you approve with regression tests.

The goal is to make code behave like mature software, software that has been run by real users for years and had every discovered bug fixed. Most of those bugs are ordinary: the off-by-one on the last page, the retry that double-charges, the config that differs in production, the error swallowed so the user sees a spinner forever. A single agent reading a repo top-down misses them. This plugin runs the audit as a search instead: enumerate, rank, understand each unit, hunt with single-question lenses, reproduce every candidate, verify adversarially, report with coverage.

Install

Claude Code (one step on 2.1.275 or later):

/plugin install production-code-audit --marketplace mdlockyer/code-quality-plugins

Or in two steps: /plugin marketplace add mdlockyer/code-quality-plugins then /plugin install production-code-audit@mdlockyer-plugins.

Claude Cowork / Claude desktop: download dist/production-code-audit.plugin from a release and open it, or install from the marketplace as above.

OpenAI Codex:

codex plugin marketplace add mdlockyer/code-quality-plugins
codex plugin add production-code-audit@mdlockyer-plugins

Codex defines subagents in its own TOML format. To give the audit its five workers in a repository, copy codex/agents/*.toml into that repository's .codex/agents/ (or into ~/.codex/agents/ for every repository). Without them the skill still runs, using Codex's built-in explorer and worker agents for dispatch.

Google Antigravity (agy):

agy plugin install https://github.com/mdlockyer/code-quality-plugins/tree/main/plugins/production-code-audit

or from a clone: agy plugin install ./plugins/production-code-audit. The five agents load natively from agents/.

What's in the plugin

Skill: production-code-audit — the orchestration procedure. Six phases, each with a coverage-based done condition, three modes (Quick, Standard, Deep), a ledger kept outside the repository, and a sequential fallback for harnesses without subagents. References cover the 14 hunting lenses, the understanding-record format, reproduction and verification rules (including exclusions, precedents, and a severity rubric), the report contract, and dispatch prompt templates. Includes scripts/inventory.py, a dependency-free repo enumerator with risk and git-history heuristics.

Agents (dispatched by the skill, one unit or finding per agent, fresh context each):

The agents are authored once in agents/*.md (Claude Code format, also read by Antigravity). codex/agents/*.toml carries the same instructions in Codex's custom-agent schema.

Usage

Ask for an audit in whatever words come naturally: "audit this repo for production readiness", "find the bugs before we launch", "is this ready to ship", "what will break when real users hit this". The skill picks a mode from the size of the target and your phrasing; say "go deep" or mention money, auth, or a launch to get Deep mode.

The audit never edits the repository until you approve fixes. It writes its working state to a sibling directory (../<repo>-audit/) and presents a report that leads with counts and a coverage section. Then it stops and asks which fixes to apply.

Works with behavioral-testing

Install behavioral-testing from the same marketplace and the audit uses its bug-repro skill for reproductions (three code states: fails on the bug on an assertion, expected value from the spec, passes on any correct fix), its test-verifier for fix review, and its test-auditor for the "can the tests fail?" lens. Without it, the audit falls back to compact versions of the same rules in references/verification.md.

What the fix phase will and won't do

Fixes land on an audit/<date> branch, one commit per approved finding, each with a regression test that was red before the fix. A fix that changes a documented behavior, public API, accepted input, or error shape needs your explicit yes before the edit, and the fix review runs the project's fixtures and recorded requests to see what the change breaks. When a correct fix turns out to need a redesign, the audit stops and reports instead of building it.

Tuning it to your codebase

When you reject a finding ("that's intentional", "not a problem here"), add the reason to the precedents list in skills/production-code-audit/references/verification.md, or to a REVIEW.md in your repository; the verifier reads both. That list is what keeps the signal clean over time.

To measure whether changes to the plugin help, plant known bugs in a representative repo and track recall per iteration. Three eval cases are included under skills/production-code-audit/evals/ in the claude plugin eval layout.

Requirements

• Read, grep, write, and shell access to the target repository and a scratch directory beside it.
• The project's own build and test tooling installed (the audit runs them as oracles).
• Optional: semgrep, bandit, gosec, staticcheck, or the equivalent for the stack; the audit uses whichever are present.

Layout

plugins/production-code-audit/
├── .claude-plugin/plugin.json      Claude Code / Cowork manifest
├── plugin.json                     Portable Agent Plugins 1.0 manifest (Codex)
├── .codex-plugin/plugin.json       Codex compatibility overlay
├── .antigravity-plugin/plugin.json Antigravity manifest
├── skills/production-code-audit/   The orchestration skill, references, script, evals
├── agents/                         Worker agents (Claude Code and Antigravity read these)
├── codex/agents/                   The same workers as Codex custom-agent TOML
└── assets/logo.svg