late-night-regrets
extension
Nightly Bayesian interaction-quality classifier — trains on chat history, flags behavioral patterns, and writes self-improvement reflections
"What did I get wrong today?"
A nightly Bayesian interaction-quality classifier for PiClaw. Trains on chat history, flags behavioral patterns, and writes self-improvement reflections — without spending model tokens on classification.
┌─────────────────────────────────────────────────────────────────┐
│ Chat message history │
│ (SQLite: messages.db, all chats, all sessions) │
└──────────────────────────────┬──────────────────────────────────┘
│
┌──────────▼──────────┐
│ Multinomial Naive │
│ Bayes classifier │ ← zero model tokens
│ (weak-label trained)│
└──────────┬──────────┘
│
┌────────────────▼────────────────┐
│ Attention-worthy messages │
│ (corrections, misinterpretations,│
│ over-engineering, under-delivery,│
│ context failures) │
└────────────────┬────────────────┘
│
┌──────────▼──────────┐
│ Agent reflection │ ← model tokens here only
│ (reads context, │
│ extracts patterns) │
└──────────┬──────────┘
│
┌────────────────▼────────────────┐
│ notes/memory/interaction- │
│ reflections.md │
│ notes/memory/feedback.md │
└─────────────────────────────────┘Open Settings → Add-Ons and install late-night-regrets from the catalog.
The settings pane exposes:
| Field | Default | Description |
|---|---|---|
| Enabled | true |
Master switch |
| Cron schedule | 30 2 * * * |
Nightly run time (UTC) |
| Confidence threshold | 0.55 |
Min confidence for attention-worthy |
| Reflections path | notes/memory/interaction-reflections.md |
Output file |
| Exports dir | exports/interaction-quality |
Classifier artifacts |
| Recent hours | 24 |
Lookback window for each reflection |
The classifier labels user messages that follow agent turns:
| Category | Signal |
|---|---|
successful_execution |
"perfect", "thanks", "exactly", short affirmatives |
course_correction |
"not that", "I meant", "try again", "to clarify" |
misinterpretation |
"wrong", "misunderstood", "that's not what I asked" |
over_engineering |
"too much", "simpler", "just do", "keep it simple" |
under_delivery |
"you forgot", "also need", "continue", "the rest" |
context_failure |
"I already said", "remember", high self-repetition |
good_proactive |
"good idea", "nice catch", "smart" |
neutral |
Normal flow, no strong signal |
/regrets — manually trigger a reflection passAfter each run:
exports/interaction-quality/
├── interaction-quality-weights-latest.json
├── interaction-quality-predictions-latest.jsonl
├── interaction-quality-attention-latest.jsonl
└── interaction-quality-report-latest.md# Run the classifier manually
bun run scripts/train-interaction-quality-bayes.ts
# Run with recent-only classification
bun run scripts/train-interaction-quality-bayes.ts --recent-hours 48
# Test
cd /workspace/piclaw-addons/addons/late-night-regrets && bun testMIT