| name | audit-codebase |
| description | Use when someone wants a whole codebase audited and fixed as tracked, resumable work — "audit this codebase", "find and fix everything wrong in this repo", "go through the app and clean it up", "set up an audit plan", "what audit phase is next" — rather than a one-shot look at the current diff. Routes to the /audit:* commands, which own the actual procedure. |
Auditing a codebase with the audit plugin
This skill exists to be found, not to explain anything. Everything it knows is one of
three files; read the one that matches and follow it exactly.
Do not use this for
A one-shot review of the working diff or a pull request. /review, /security-review and
the native code-review tooling are the right answer there, and reaching for a manifest to
look at ten changed lines is pure overhead. This is for work that spans sessions and has to
survive being interrupted.
Route
Check for a manifest first — docs/audit/audit-plan.json by default, or manifestPath in
.claude/audit.config.json.
| Situation | Command |
|---|
| No manifest yet | /audit:init — generates one from the codebase |
| A manifest exists, and the question is "where are we" | /audit:status |
| A manifest exists, and the question is "what now" | /audit:next |
| Run the next phase end to end | /audit:phase <id> |
| Something is not working and it is not obvious why | /audit:doctor |
Read ${CLAUDE_PLUGIN_ROOT}/commands/<name>.md and do what it says. Those files are the
source of truth; nothing here restates their procedure, because two copies of a procedure
is one copy and one lie.
The one thing worth knowing before starting
/audit:init spawns several agents and is the most expensive command here. If the person
has not committed to the pipeline yet, /audit:usage --backfill costs nothing, reads
transcripts already on disk, and shows them what their existing work has cost — see the
audit-spend skill. That is a better first move than a manifest they did not ask for.