abildtoft/kramme:code:cleanup-ai

Remove AI Code Slop

This skill uses the kramme:deslop-reviewer agent to identify AI slop in the branch's diff against a base, then applies the agent's recommended fixes.

Sibling: this is the AI-slop-specific pass via kramme:deslop-reviewer; for a general simplification pass on the same post-feature branch, use kramme:code:refactor-pass.

Parse $ARGUMENTS before the preconditions. If --auto is present, set AUTO_MODE=true and remove the flag before base-branch resolution. --auto applies medium-confidence cleanup findings automatically; it does not bypass dirty-worktree protection or behavior/API/test-expectation safeguards.

Preconditions

Run git status --porcelain. If the working tree is dirty and AUTO_MODE=true, stop with MISSING REQUIREMENT: working tree is dirty; rerun without --auto to decide whether to continue. If the working tree is dirty and AUTO_MODE is false, confirm with the user before continuing — the skill's edits will land alongside theirs in git diff and will be hard to separate when reverting.

Process

1. Resolve the base branch — try these sources in order; use the first non-empty value:

   • The argument passed to this skill.
   • gh pr view --json baseRefName --jq '.baseRefName' 2>/dev/null
   • git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@'

   Assign the result to BASE_BRANCH with any leading origin/ stripped, then validate and fetch:

   git check-ref-format --branch "$BASE_BRANCH" && \
     git fetch origin "refs/heads/${BASE_BRANCH}:refs/remotes/origin/${BASE_BRANCH}" && \
     git rev-parse --verify --quiet "origin/$BASE_BRANCH"
   

   On failure, stop and ask the user to re-run with the base branch as the skill argument.

2. Scan for slop

   If git diff origin/$BASE_BRANCH...HEAD is empty, report "no branch changes to review" and stop.

   Launch kramme:deslop-reviewer in code review mode against git diff origin/$BASE_BRANCH...HEAD.

3. Filter the findings

   Discard any finding whose file is generated, vendored, a lockfile, a snapshot, or *.d.ts before deciding what to fix.

4. Apply fixes by confidence

   The agent scores findings 0–100. Apply the agent's specific recommendation per finding — do not pattern-match generically:

   • ≥76: auto-apply.
   • 51–75: summarize the finding to the user and apply on confirmation. If AUTO_MODE=true, apply automatically and count it separately as medium-confidence auto-applied.
   • <51: skip; list in the final report.

   Leave a finding unchanged when applying it would alter test expectations, public APIs, or behavior the agent itself flagged as possibly intentional.

5. Report

   1–3 sentences: which files were edited, plus counts for applied / confirmed / skipped / filtered.