code-check

name	code-check
description	Triggers on /code-check. Structural code review - file splits, DRY, dead code. Writes findings to ai_todos. Callable standalone or from /close.
argument-hint	[uncommitted|unpushed|<path>|<hash>]

/code-check

Structural review: file splits, DRY, dead code. Writes findings to ai_todos.

Scope resolution

Determine what to review based on args:

Arg	What to run
uncommitted	git diff HEAD --name-only --diff-filter=ACM
unpushed	git log @{u}..HEAD --name-only --diff-filter=ACM --format=
Looks like a file path	treat as single-file list
Looks like a hash or range (abc1234, HEAD~3..HEAD)	git diff <arg> --name-only
No args	ask using AskUserQuestion (see below)

When asking, options:

• "Uncommitted changes" - files changed but not yet committed
• "Unpushed commits" - commits made since last push
• "Specific file or path" - follow up asking the dev to type the path (plain message, not AskUserQuestion)
• "Commit hash or range" - follow up asking the dev to type the hash or range

After resolving, filter to code files only. Drop: .md, .json, .toml, .yaml, .yml, .gitignore, anything under .for_bepy/ or memory/.

If the filtered list is empty: print "No code files in scope." and stop.

Step 1 - Size check

For each file in scope, get line count: wc -l "path" via Bash tool.

If > 400 lines AND has an obvious split seam (separate concerns, reusable unit, clear module boundary), record:

{ "title": "[file] should be split", "files": ["path"], "problem": "[file] is N lines, mixes [X] and [Y]", "fix": "split at [boundary] into [new-file]" }

If no obvious seam, skip that file.

Step 2 - DRY pass

For each new top-level symbol in scope (function, const, class, interface, type, export, def, func, local function - language-dependent): Grep the repo for equivalents by name, shape, and purpose. For each duplicate found, record:

{ "title": "Duplicate: [symbol]", "files": ["path:line (new)", "path:line (existing)"], "problem": "what duplicates what (one sentence)", "fix": "delete X and import Y / extract shared util to Z" }

Cap: ~3 Grep calls per symbol. Use only Grep, Read, and Glob - no shell pipelines. If scope has zero new top-level symbols (body-only edits), skip this step.

Step 3 - Dead code pass

For each new top-level symbol: Grep pattern: "\\b<symbol>\\b", output_mode: "count". Count <= 1 (definition only) = never called. Also flag: unreachable branches, commented-out blocks left in, imports never read.

{ "title": "Dead code: [symbol]", "files": ["path:line"], "problem": "one sentence", "fix": "delete / uncomment / wire up at X" }

Step 4 - Output

Merge findings from Steps 1-3.

If .for_bepy/ai_todos/ exists: write each finding as a .md file. Filename: zero-padded numeric prefix (scan existing files for max id, add 1) + kebab-case slug. Format:

# [title]

## Goal
[what needs to happen]

## Context
[problem. Cite files as path:line.]

## Approach
[fix]

## Acceptance
[how to verify it is done]

If no .for_bepy/ exists: print all findings inline, one block per finding.

Print a summary line:

code-check: N findings (A size, B DRY, C dead code). M written to ai_todos.

If zero findings: code-check: No structural issues found.