Nih Audit

NIH Audit — Not Invented Here

Find code reinventing the wheel. Recommend libraries. Score with evidence.

/nih-audit [scope] — scope defaults to repo root.

Phase 0 — Manifest discovery (no LLM)

Find dependency manifests anywhere in scope, excluding node_modules/, vendor/, .git/, build artefacts:

• package.json, Cargo.toml, pyproject.toml, go.mod, Gemfile, requirements.txt, composer.json, build.gradle, pom.xml, mix.exs.

Extract dependency names per ecosystem:

Build a depManifest:

{
  "workspaces": [
    { "root": ".", "ecosystem": "node", "deps": ["express", "zod"] }
  ],
  "primaryLanguages": ["typescript"]
}

If no manifest is found anywhere in scope, abort with a clear message — without an installed-deps list there is nothing to cross-reference.

Phase 1 — Structural NIH scan

Spawn the nih-scanner sub-agent:

inputs:
  languages   = primaryLanguages from Phase 0
  scope       = $ARGUMENTS or repo root
  depManifest = JSON from Phase 0
  slug        = <derived slug>

The scanner returns a JSON candidate list inline plus a one-paragraph summary. Parse the candidates from the response. If 0 candidates, report clean and stop.

Phase 2 — Library discovery (parallel, /research)

Group candidates by category (RETRY, UUID, VALIDATION, DATE, DEBOUNCE, CLONE, ARGPARSE, STRING, HTTP, SERIALIZATION, ERROR, CRYPTO, SECURITY, FORMAT, COMPARE).

For each category, dispatch /research with a focused question shape:

/research "best <category> library for <language>; must be MIT/Apache/BSD;
list weekly downloads or crates.io downloads, GitHub stars, last commit
date, contributor count; flag if functionality is in the standard library"

Cap: max 5 parallel research dispatches. If /research is unavailable, emit candidates with recommendation: null and a note.

For each library returned, capture:

• name + latest version
• license (flag GPL; prefer MIT/Apache-2.0/BSD)
• weekly/monthly download count
• GitHub stars, last commit date, contributor count
• whether it is stdlib, micro-library, or framework
• one-sentence API example showing replacement shape

Drop recommendations whose name is already in depManifest. Stdlib alternatives (no new dep) are the highest-value class — keep them prioritized.

See references/categories.md for the canonical category-to-library map.

Phase 3 — Spec and intent alignment

Find spec directories within scope: **/specs/*.md, .cheese/specs/*.md. Filter out node_modules/, vendor/, .git/.

For each candidate, look for signals that the NIH choice was deliberate:

• In specs: keywords like "intentionally", "we chose to build", "build vs buy", "don't use", "avoid dependency on" near the candidate's concept.
• In the candidate file: comments like intentionally, deliberately, we chose, NOTE:, DECISION:, instead of, rather than.

Apply scoring modifiers:

Phase 4a — Score

For every candidate with a library recommendation, run the 4-step confidence chain:

Step 1 — classify finding type

Step 2 — evidence grounding

Step 3 — context modifiers

Apply Phase 3 modifiers (spec intent, code comments, planned features) plus:

Phase 4b — Validate and size

Step 4 — independent second pass

For every candidate, score independently from the first pass:

1. Drop the first score from working memory.
2. Re-read the NIH code and the library API fresh.
3. Score using Steps 1–3 again.
4. Report both scores in the finding.
5. Final score = average of pass 1 and pass 2.
6. Divergence > 20 points → tag AMBIGUOUS, still emit.

Effort sizing

Phase 5 — Report

Write the full report to .cheese/nih/<slug>.md. Render every candidate above the threshold; do not silently filter. Each finding follows the template in references/finding-template.md.

Top of the report:

# NIH Audit — <scope>

## Summary
- Files scanned: N
- Candidates found: N
- Already using best option: N (filtered)
- Ambiguous (passes diverge >20): N

## All findings (sorted by score, descending)

| # | Score | P1 | P2 | Category | NIH Code | Replace With | Effort |
|---|-------|----|----|----------|----------|--------------|--------|
| 1 | 92 | 90 | 94 | UUID | src/utils/uuid.ts:12 | crypto.randomUUID() (stdlib) | S |

After the table, render every ### Finding #N block inline.

Phase 6 — Hand-off

Print the report path. Do not auto-invoke any follow-up:

NIH report: .cheese/nih/<slug>.md

Next steps (manual):
  • Pick the highest-scored finding and migrate.
  • For S-effort stdlib swaps, /cook is usually the right next step.
  • For L-effort framework swaps, run /mold first to lock down the
    migration interface.

Rules

• Do not modify code or implement migrations — recommend; the human decides.
• Do not recommend GPL libraries without flagging the licence risk.
• Do not override explicit NIH decisions documented in specs or code comments — apply the scoring modifiers and surface the conflict.
• Do not run in scopes without any manifest file — there is nothing to cross-reference.
• Empty results report clean and stop. Never force findings.
• Cap total tool calls across all phases at ~60. After the cap, synthesize from available data and note incomplete coverage.

Gotchas

• ast-grep patterns are approximate. A clearTimeout + setTimeout combo isn't always a debounce. Step 2's "generic pattern, code does more" modifier (-15) catches false positives.
• Stdlib alternatives are the highest value. crypto.randomUUID() replacing a hand-rolled UUID is a no-brainer. Always score these highest (REPLACE_WITH_STDLIB, base 55, cap 100).
• Already-installed is the most common false positive. A codebase that has lodash but hand-rolls deepClone may have done so for bundle size reasons. The spec/comment check (Phase 3) catches this.
• Monorepo dep scoping. A function in packages/api/ may be NIH in that workspace but the library is installed in packages/web/. Each workspace's depManifest is independent.
• Licence compatibility is not just MIT-vs-GPL. Some projects have specific licence requirements. When in doubt, flag the licence and let the human decide.