/fix

name: fix
description: "Auto-fix fixable findings in NL artifacts — missing fields, heading gaps, field renames"
argument-hint: "[path]"
allowed-tools: Read, Write, Edit, Glob, Grep, Task

User Input

$ARGUMENTS

Workflow

Step 1: Score First

Parse `$ARGUMENTS` for a path (default: cwd if empty). Dispatch the `nlpm:scorer` agent via Task tool to score all artifacts at that path. Collect all findings.

If no artifacts are found → "No NL programming artifacts found." and stop.

Step 2: Classify Findings

Separate findings into auto-fixable and not-auto-fixable:

**Auto-fixable (will be applied):**

• Missing `user-invocable: false` on shared partials → add to frontmatter
• `tools` field in command frontmatter where `allowed-tools` is expected → rename field (note: `tools` in agent frontmatter is valid and not auto-fixed)
• Heading hierarchy gaps (## → ####) → insert missing ### level
• Trailing whitespace → strip
• Missing `description` in shared partial frontmatter → generate from filename and first heading
• Missing `name` in skill frontmatter → derive from directory name
• Missing `argument-hint` on commands that reference `$ARGUMENTS` in body → add `"[args]"`

**Suggested fixes (shown with diff, applied only if user approves):**

• Agent using opus for mechanical task → suggest downgrade to sonnet/haiku with rationale
• Agent with no `<example>` blocks → generate one skeleton example from the description (one is full R09 credit); when the description also lacks an exclusion, apply the next fix in the same edit
• Agent whose description names nothing it is not for → add one "Not for …" sentence (the other R09 row, -5). Derive the exclusion from the agent's own body (its Boundaries or "Do NOT" lines) or from a sibling agent that owns the excluded work ("Not for X; use <sibling>"). If neither source names one, report the finding instead of writing a sentence — never invent a boundary
• Skill >500 lines → suggest split points based on H2 sections

**NOT auto-fixable (reported but not modified):**

• Vague descriptions (what should the trigger phrases be?)
• Behavioral contradictions (which artifact is right?)
• Scope boundaries, output formats, error handling

Step 3: Apply Fixes

For each auto-fixable finding: 1. Read the file 2. Apply the fix using Edit tool 3. Log: "Fixed: {file}:{line} — {what was changed}"

Step 4: Re-Score

Re-run scoring on all fixed files to compute new scores.

Step 5: Report

NLPM Fix Report

Fixed {N} findings in {M} files:
  {file}:{line} — added user-invocable: false (+25 points)
  {file}:{line} — renamed tools → allowed-tools (+0 points, consistency)
  {file}:{line} — inserted missing ### heading (+0 points)

Not auto-fixable ({N} findings remain):
  [{SEVERITY}] {file}:{line} — {finding} (requires human judgment)

Score changes:
  {file}   {old_score} → {new_score}  (+{delta})

Overall: {old_avg} → {new_avg}

**Error handling:**

• No fixable findings → "All remaining findings require human judgment. Run /nlpm:score to see details."
• File not writable → skip with warning