| name | late-night-regrets |
| description | Nightly Bayesian interaction-quality classifier — flags behavioral patterns from chat history and writes self-improvement reflections without spending model tokens on classification. |
| distribution | public |
Late Night Regrets
A lightweight, zero-token-cost interaction-quality classifier that runs nightly to identify behavioral patterns worth improving.
What it does
- Trains a Multinomial Naive Bayes classifier on the full chat message history
- Classifies every user message that follows an agent turn into quality categories
- Flags attention-worthy messages (corrections, misinterpretations, under-deliveries)
- Triggers a short agent reflection pass that reads flagged messages, identifies patterns, and writes learnings to notes
Categories
| Label | Meaning |
|---|
successful_execution | Agent fulfilled the request correctly; user approved or moved on |
course_correction | User had to steer, clarify, or redirect |
misinterpretation | Agent misread intent; user explicitly corrected |
over_engineering | Agent did too much; user asked to simplify |
under_delivery | Agent gave too little; user pushed for more |
context_failure | Agent forgot/lost context; user had to repeat |
good_proactive | Agent anticipated a need; user approved |
neutral | Normal flow, no strong signal |
How classification works (no model tokens)
- Messages are tokenized (lowercased, code/URLs stripped, bigrams, structural features)
- Sequential context features: previous message sender, length, turn-pair detection, self-repetition ratio
- Weak labels derived from pattern matching on user follow-ups (approval words, correction phrases, repetition detection)
- Standard MNB training with 80/20 deterministic split
- Confidence-based filtering: only non-neutral predictions above threshold get flagged
Nightly flow
The scheduled task runs at 02:30 UTC (configurable):
1. bun run <addon>/scripts/train-interaction-quality-bayes.ts
→ retrains on full history, writes weights + predictions + attention file
2. Agent reads attention file, filters last 24h
3. For each flagged message:
- Reads surrounding context from messages DB
- Identifies what the user wanted vs what agent did
4. Writes patterns to notes/memory/interaction-reflections.md
- Date, top patterns, behavioral adjustments
5. Appends new steering cues to notes/memory/feedback.md if warranted
Manual trigger
/regrets
Shows working/status feedback only; it does not execute the classifier scripts.
Configuration
The direct config API stores these fields; the add-on does not include a browser settings pane:
| Field | Default | Description |
|---|
| Enabled | true | Master switch |
| Cron schedule | 30 2 * * * | When to run nightly |
| Confidence threshold | 0.55 | Minimum confidence for attention-worthy |
| Reflections path | notes/memory/interaction-reflections.md | Where learnings are written |
| Exports dir | exports/interaction-quality | Classifier artifacts |
| Recent hours | 24 | Window for the reflection pass |
Artifacts
exports/interaction-quality/
interaction-quality-weights-latest.json # trained model weights
interaction-quality-predictions-latest.jsonl # all predictions
interaction-quality-attention-latest.jsonl # flagged messages only
interaction-quality-report-latest.md # human-readable report
Design principles
- Zero token cost for classification: pure Bayes on bag-of-words + structural features
- Sequential context matters: features include prior message characteristics and self-repetition
- Conservative flagging: only high-confidence non-neutral predictions surface for reflection
- Model tokens only for understanding: the nightly agent pass reads a small set of flagged messages to extract why
- Incremental improvement: each night's reflection appends patterns; Dream consolidates over time
