ac-grind.js — $0 Local-Model Code Grinder (per-finding)
What it is
A wrapper around aider + the autocoder.js differential-tsc-gate logic that fixes one finding/file per pass using a local $0 model (Ollama qwen2.5-coder:32b-instruct-q8_0 @ 10.0.0.2:11434, or MLX qwen3-coder-30b @ 11437). Implements the measured recipe from /tmp/evidence-mlx-autocoder-fix/verification.md.
Tool: ~/system/tools/ac-grind.js · MC: #103050 · Status: proven (2026-06-06)
The recipe (7 levers)
- Scope = 1 finding/file per pass (broad batches choke the local model).
- Auto-attach relevant DEF files as read-only aider context — biggest accuracy lever (measured 5→1). Resolves the target's
@/…imports to real files via tsconfig paths. - Qwen sampling: temp 0.7 / top_p 0.8 / top_k 20 / rep_penalty 1.05.
- Repo-map (tree-sitter) + diff edits (whole-file fallback for small files on aider 0.86.x).
- DIFFERENTIAL touched-file-only tsc gate — repo is ~615 errors RED at baseline, so only NEW errors in the touched file count.
- Reflection cap ≥ 6 (aider 0.86.x caps internal loop at 3 → wrapper does ceil(N/3) outer passes, stops on tsc-delta stall).
- E2E browser gate (vite + Playwright + MSAL fixture) as the real DONE signal, not tsc-only.
Usage
node ~/system/tools/ac-grind.js --file <abs path> [--desc "what to fix"]
[--engine ollama|mlx] [--max-reflections 6]Proof (MC #103050, EVVRecordsList.tsx, LumisCare)
- 8 DEF files auto-attached.
- 5 → 0 differential tsc errors in one aider pass (independently confirmed:
npx tsc -bclean on fixed file). - Diff validated vs ground-truth APIs:
PaginationIndicator(real export, props page/totalPages/rowCount/pageSize/onPageChange),Badge size="sm"(valid cva variant). OriginalPagination currentPage=+size={16}were genuine type errors. - E2E Playwright PASS (results.json expected:1 unexpected:0) + real screenshot — component mounts in a real browser without crash.
- $0 (local model), no deploy, proof edit local-only & reverted.
Proven scope / honest limits
- Reliable for mechanical type-error findings (prop/API mismatches) where DEF files carry ground-truth types.
- NOT proven for open-ended feature generation (model must invent new logic).
- E2E gate currently proves mount-not-crash, not full data render — backend must run to prove real rows.
Hardening & real-batch results
Guard A — file-size guard
DEFAULT_MAX_LINES = 450 (calibrated: proven-good 397 lines < 450 < proven-bad 585 lines). Files over the threshold emit a loud SKIPPED_TOO_LARGE box to stderr with exit code 2 — before any model call is made, so no tokens are wasted. CLI flags: --max-lines N to override, --force-large as escape hatch.
- EVVRecordDetail.tsx (586 lines) → exit 2, SKIPPED_TOO_LARGE (correctly blocked).
- QualificationsTab.tsx (397 lines) → allowed by default, flows to aider (proven-good path confirmed).
Guard B — artifact lint
Diff-scoped JSX-comment and whitespace-expression detection runs after every aider edit. Detects patterns such as {" "} whitespace expressions, {/* narrative comments */}, and inline narrating comments on JSX tags. On detection, ac-grind auto-strips and re-validates; if the stripped result still fails, it reverts the file and exits with ARTIFACT_REJECTED (exit 3). Legitimate {" "} (e.g. dash separators) are preserved because detection is diff-scoped, not file-scoped. Never ships dirty output.
- Pattern unit test: 12/12 PASS.
- Integration test (MC #103058 stash artifacts): 5 artifact lines detected, auto-stripped cleanly, legitimate separator preserved.
Real backlog sweep (MC #103063)
Ran across the backoffice module of LumisCare. 15 files edited by ac-grind sweep; 13 kept (clean error reduction), 4 reverted (quote-style-only changes, zero error reduction).
| Metric | Value |
|---|---|
| Files edited by sweep | 15 |
| Kept clean (real error reduction) | 13 |
| Reverted (quote-style only, no improvement) | 4 |
| Baseline tsc errors (origin/dev) | 612 |
| After sweep | 578 |
| Net reduction | −34 errors (5.5% of backlog) |
| New errors introduced | 0 |
Shipped as PR #55 to dev (branch ac-grind-sweep-103063, commit b17f606), pending Proveo review. Not merged.
Honest takeaway: ac-grind reliably clears mechanical prop/API type-errors on files at or below 450 lines. Structural errors (e.g. EVVRecordsList.tsx — 5 errors untouched) and over-threshold files still require manual work.
Guard evidence
/tmp/evidence-103061/verification.md— Guard A + B hardening + threshold recalibration to 450/tmp/evidence-103063/verification.md— real-batch sweep triage table + PR
Evidence
/tmp/evidence-103050/verification.md(+ John independent verification section)/tmp/verify-103050/evidence/(build-tsc, test-e2e, playwright-results.json, screenshot)- Recipe origin:
/tmp/evidence-mlx-autocoder-fix/verification.md - Related memo:
project_local_model_coding_setup_2026-06-06.md