review-llm-artifacts

What this skill does

# LLM Artifacts Review

Detect common artifacts left behind by LLM coding agents: over-abstraction, dead code, DRY violations in tests, verbose comments, and defensive overkill.

## Hard gates (sequence)

Advance only when each **pass condition** is objectively true (prevents “review complete” without artifacts):

| Gate | Pass condition |
|------|----------------|
| **G1 — Scope** | File list is non-empty *or* you exit with exactly the Step 1 message; `scope` is set to `all` or `changed`. |
| **G2 — Four categories** | Tests, dead code, abstraction, and style are each reviewed (four parallel subagent runs when supported, or four sequential passes covering the same categories). **Stop** if any category did not complete; do not write JSON or a summary that implies a full pass. |
| **G3 — JSON before summary** | `.beagle/llm-artifacts-review.json` exists and is valid JSON **before** Step 6 markdown. |
| **G4 — Integrity** | Step 7 checks pass before treating the run as complete. |

## Arguments

Parse `$ARGUMENTS` for flags and optional path:

| Flag | Effect |
|------|--------|
| *(default)* | **Changed-files scope** — only files changed since `git merge-base HEAD main` (PR-style scope) |
| `--all` | Full project scan — all matching source files under the target path |
| `--parallel` | Force parallel execution where subagents are supported (default when 4+ files in scope) |
| Path | Root directory to scan (default: current working directory) |

## Step 1: Determine Scope

**A. Changed files only (default):**

Resolve the base ref explicitly and fail loudly if none exists — **do not** wrap the `git merge-base` call in `|| true`, which would silently swallow a missing `main`/`master` ref and report "no files to scan" on repos that only have `origin/main` or use `master`. If no base ref is found, suggest the user pass `--all` instead of silently falling back.

```bash
BASE=$(for ref in main origin/main master origin/master; do
         git rev-parse --verify "$ref" >/dev/null 2>&1 && { echo "$ref"; break; }
       done)
if [ -z "$BASE" ]; then
  echo "error: no main/master ref found (checked main, origin/main, master, origin/master). Pass --all for a full-project scan." >&2
  exit 1
fi
MERGE_BASE=$(git merge-base HEAD "$BASE") || {
  echo "error: git merge-base HEAD $BASE failed." >&2
  exit 1
}
git diff --name-only "$MERGE_BASE..HEAD" | grep -E '\.(py|ts|tsx|js|jsx|go|rs|java|rb|swift|kt)$' || true
```

(The trailing `|| true` on the `grep` is intentional — zero source-file matches is a legitimate empty-scope result, distinct from a failed base-ref resolution.)

**B. Full project (`--all`):**

From `TARGET` (default `.`), list source files and **prune** excluded dependency/build trees so `find` never descends into them. `! -path "*/foo/*"` only filters the output; `find` still walks the tree (minutes of wasted I/O on large `node_modules`, `target`, etc.). Use `-prune` instead:

```bash
find "$TARGET" \
  \( -type d \( \
       -name node_modules -o -name .git -o -name vendor -o -name __pycache__ \
    -o -name .venv        -o -name venv -o -name dist   -o -name build \
    -o -name target       -o -name .next -o -name coverage -o -name .turbo \
  \) -prune \) -o \
  \( -type f \( \
       -name "*.py" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" -o -name "*.jsx" \
    -o -name "*.go" -o -name "*.rs" -o -name "*.java" -o -name "*.rb" \
    -o -name "*.swift" -o -name "*.kt" \
  \) -print \)
```

**Large repos:** The `--all` path can produce huge file lists. If file count exceeds **400**, warn and suggest narrowing: pass a subdirectory as `TARGET`, or drop `--all` to fall back to the default changed-files scope. Still proceed unless the user explicitly cancels. (This warning does **not** fire on the default changed-files scope, which is already bounded by the PR diff.)

If no files are found, exit with:

`No files to scan. Check the path, branch, or pass --all for a full-project scan.`

Set `scope` in the report: `"all"` for `--all`, `"changed"` for the default changed-files scope.

## Step 2: Detect Languages

Extract unique file extensions from the file list:

```bash
echo "$FILES" | sed 's/.*\.//' | sort -u
```

Map extensions to language names for the report:
- `.py` -> Python
- `.ts`, `.tsx` -> TypeScript
- `.js`, `.jsx` -> JavaScript
- `.go` -> Go
- `.rs` -> Rust
- `.java` -> Java
- `.rb` -> Ruby
- `.swift` -> Swift
- `.kt` -> Kotlin

## Step 3: Review the Four Categories

Cover all four categories below. **If the agent supports subagents** and file count >= 4 (or `--parallel` is set), dispatch one subagent per category in parallel. **Otherwise**, run the four category reviews sequentially yourself, producing the same findings. Either way:

1. Load the [llm-artifacts-detection](../llm-artifacts-detection/SKILL.md) skill
2. Review each category (one per subagent when parallel, one pass at a time when sequential)
3. Collect findings in the structured format below

### Category 1: Tests

**Focus:** Testing anti-patterns from LLM generation

- DRY violations (repeated setup code, duplicate assertions)
- Testing library/framework code instead of application logic
- Wrong mock boundaries (mocking too much or too little)
- Overly verbose test names that describe implementation
- Tests that just mirror the implementation

### Category 2: Dead Code

**Focus:** Unused or obsolete code

- Unused imports, variables, functions, classes
- TODO/FIXME comments that should have been resolved
- Backwards compatibility code for removed features
- Orphaned test files for deleted code
- Commented-out code blocks
- Feature flags that are always on/off

### Category 3: Abstraction

**Focus:** Over-engineering patterns

- Unnecessary abstraction layers (interfaces for single implementations)
- Copy-paste drift (similar code that diverged slightly)
- Over-configuration (configurable things that never change)
- Premature generalization
- Factory/Builder patterns for simple object creation
- Deep inheritance hierarchies

### Category 4: Style

**Focus:** Verbose or defensive patterns

- Verbose comments explaining obvious code
- Defensive overkill (null checks on non-nullable values)
- Unnecessary type hints (dynamic languages with obvious types)
- Overly explicit error messages
- Redundant logging
- Self-documenting code with documentation

## Step 4: Consolidate Findings

**Prerequisite:** **G2** satisfied (all four category reviews finished successfully).

Once all four category reviews have completed (parallel subagents or sequential passes), then:

1. Merge all findings into a single list
2. Assign unique IDs (1, 2, 3...)
3. Group by category for display

**Echo before write (anti-confabulation):** Every finding written to JSON MUST come from a category review's `[FILE:LINE] ISSUE_TITLE` output, not from the branch name, directory, or your own inference. After assigning ids, echo the consolidated table — `id | category | file:line | description` — and confirm each row traces to a specific category result. Do not add findings that no category review reported.

**ID lock:** Ids are contiguous `1..N` with no gaps or duplicates. This `1..N` set is the **locked id set** that downstream skills ([verify-llm-artifacts](../verify-llm-artifacts/SKILL.md), [fix-llm-artifacts](../fix-llm-artifacts/SKILL.md)) bind to 1:1. `summary.total` MUST equal `N`, and `summary.by_category` counts MUST sum to `N`. State the id set before writing JSON.

## Step 5: Write JSON Report

Create `.beagle` directory if it doesn't exist:

```bash
mkdir -p .beagle
```

Write findings to `.beagle/llm-artifacts-review.json`:

```json
{
  "version": "1.0.0",
  "created_at": "2024-01-15T10:30:00Z",
  "git_head": "abc1234",
  "scope": "all" | "changed",
  "target": ".",
  "files_scanned": 42,
  "languages": ["Python", "TypeScript", "Go"],
  "findings": [
    {
      "id": 1,
      "category": "tests" | "dead_code" | "abstraction" | "style",
      "type": "dry_violation" | "unused_import" | "over_abstrac