Inspect the suggestions.
Keep the decision.
Slow Triage is a local prototype for reviewing an issue export. A real Mastra workflow and MiniLM embeddings suggest labels and possible duplicates, with references you can inspect and changes you can undo.
Reviewed code + CLI demo, prepared for public release.Actual commands.
Readable evidence.
A 60-second, silent walkthrough of captured CLI output, with formatted excerpts and reading holds. It shows local inference, repeated decisions and an apply/undo call through the real loopback API.
CLI demonstration with synthetic data. Reading holds are added; playback is not real time. The browser interface was not exercised. The final frame describes the publication status at recording time, October 19, 2026 UTC.
Read the visual transcript ↗- 01
Validate the export
Normalize issue IDs, titles, bodies and labels within bounded input limits.
- 02
Embed locally
The pinned q8 MiniLM model runs on CPU and produces 384-dimensional vectors.
- 03
Compare and abstain
Compare against 24 synthetic references. Low similarity or a small margin sends a case to review.
- 04
Build the review queue
Expose references and scores. Label changes need explicit selection and affect a volatile copy.
The baseline did better.
The conservative classifier underperformed a keyword baseline in both small synthetic evaluations. These results support an inspectable demonstration; they do not establish real-world effectiveness.
| Synthetic fixture | Embedding policy | Keyword baseline |
|---|---|---|
| Developer-authored 20 cases | 45%9 / 20 correct | 90%18 / 20 correct |
| Independently authored 40 cases | 37.5%15 / 40 correct | 60%24 / 40 correct |
References and thresholds were frozen before scoring. Independent fixture: 8 answered, 31 abstained, 1 blank-title input rejected. Similarity is a heuristic, not a calibrated probability. Inspect the retained metrics.
Duplicate suggestions are still suggestions.
The independent synthetic duplicate test found 7 true-positive pairs, 1 false positive and 1 missed pair. Precision and recall were each 7/8 (87.5%); nearest-candidate matches were 15/16 positive queries.
The sample is small. Related topics can look like duplicates. No automatic issue closing or merging exists, and no real-maintainer study has been run.
The source is the handoff.
Read the exact reviewed source, synthetic fixtures and retained measurements. Download the frozen archive to run the prototype locally after setup.
Start with the parts that matter.
- prototype/README.md Setup, mechanism and limitations
- prototype/src/workflow.mjs The four real Mastra stages
- prototype/src/core.mjs Validation and decision policy
- prototype/src/model.mjs Pinned local model execution
- review/metrics.json Independent synthetic results
- prototype/models/manifest.json Model source, sizes and hashes
Setup is a separate step.
Requires Node.js ≥22.13 and Python 3. The archive includes the dependency lockfile; installed packages and model weights are excluded. Initial uncached setup needs npm and Hugging Face downloads. No paid inference API is used.
Local and review-first.
The model truncates at 256 wordpieces. English short-text, ambiguous and domain-specific issues can fail. Working copies and undo state are memory-only. The original issue export stays unchanged.
Verification scope: actual CLI inference, loopback API, source checks and simulated-DOM flows were reviewed. Rendered-browser end-to-end interaction, keyboard behavior, payload execution and actual browser downloads remain unverified. No UI screenshot, real-world benchmark, human time-saving result, prize eligibility or organizer acceptance is claimed.