claudit

name	claudit
description	Perform a thorough code audit of the current branch vs main, writing findings to docs/claudit/.
disable-model-invocation	true

Claudit — Branch Audit

Perform a thorough code review of the current branch's changes against main.

Input

No arguments required. The branch name and counter are determined automatically.

Procedure

0. Branch check — Get the current branch name from git branch --show-current. If the branch is main, stop immediately and tell the user: "Claudit audits compare against main — running on main itself would produce an empty diff. Please switch to a feature branch first." Do not proceed.

1. Preparation — Read CLAUDE.md for project context. Read existing files in docs/claudit/ to understand prior audit history and resolved findings for this branch.

2. Gather the diff — Run git diff main and git diff main --stat. Note the approximate line count and files changed.

3. Read source files — Read all changed/new source files in full (not just the diff hunks). Understanding surrounding context is critical for finding issues the diff alone won't reveal.

4. Audit — Review for:

   • Correctness — logic errors, off-by-one, missing error handling, invariant violations
   • Security — injection, TOCTOU, information disclosure, fail-open vs fail-secure, unbounded recursion/allocation
   • Performance — unnecessary allocations, N+1 patterns, missing batching, redundant computation
   • API design — inconsistent interfaces, missing validation, confusing semantics, encapsulation violations
   • Style — import ordering (per CLAUDE.md), hardcoded strings that should use enums, pub vs pub(crate) violations

   Focus on new findings not covered in prior claudit rounds. Check whether prior open findings are still open or have been resolved.

5. Classify findings by priority:

   • High — correctness or security bugs that could cause wrong behavior, data loss, or security vulnerabilities
   • Medium — design issues, performance problems, or gaps that could cause subtle bugs
   • Low — style issues, minor inefficiencies, documentation gaps, encapsulation concerns

Output format

Determine the output filename:

1. Get the current branch name from git branch --show-current.
2. List existing files in docs/claudit/ matching *-<branch>-*.md to find the highest counter. If none exist, the counter is 1. Otherwise increment the highest by 1.
3. Write to docs/claudit/<YYYY-MM-DD>-<branch>-<counter>.md.

Use this exact structure:

# Branch Audit: <Branch Name> (Round N) — <YYYY-MM-DD>

<One-line scope description: branch name, diff size, file count, focus areas.>

## Summary

| Priority | Open | Resolved |
|----------|------|----------|
| High     | X    | Y        |
| Medium   | X    | Y        |
| Low      | X    | Y        |

---

## High Priority

### 1. <Finding title>

**File:** `path/to/file.rs:line-range`

<Description of the issue. Be specific about what's wrong and why it matters.>

**Suggested fix:** <Concrete suggestion, not vague.>

### ~~2. <Resolved finding title>~~ — RESOLVED

**File:** `path/to/file.rs:line-range`

~~<Original description, struck through.>~~

**Resolution:** <What was done to fix it.>

---

## Medium Priority

<Same pattern as above.>

---

## Low Priority

<Same pattern as above.>

---

## Positive Observations

- **<Observation title>.** <Why it's good — be specific about the design decision or implementation quality.>

Rules

• Number findings sequentially across all priority levels (don't restart numbering per section).
• Open findings get a plain ### N. Title. Resolved findings get ### ~~N. Title~~ — RESOLVED.
• Include file paths with line numbers where applicable.
• Include 4-6 positive observations — acknowledge good design decisions, not just problems.
• Each round is a clean document with only new findings. Resolved findings are updated inline in their original document (strikethrough + resolution note), NOT carried forward into new round documents. A new round document should only contain new findings discovered in that round. The summary table at the top of each document reflects that document's findings only.
• Do NOT fix any issues — this is a read-only audit. The user will decide what to fix.
• If the branch has had prior claudit rounds, note the total resolved findings count in the summary.