Bayesian interaction-quality classification for Piclaw chat history. Requires Piclaw >=2.1.0.
Open Settings → Add-Ons and install late-night-regrets from the catalog.
scripts/train-interaction-quality-bayes.ts trains a weak-label Multinomial Naive Bayes model and writes classifier artifacts.scripts/classify-recent.ts classifies recent messages.scripts/setup-nightly-task.ts creates the optional scheduled agent task.late-night-regrets skill guides the agent through classification and reflection./regrets currently displays working/status feedback; it does not execute the classifier scripts itself.The extension registers a direct config API but no browser settings pane. Saved config is available to the injected prompt; the scripts still take their own CLI parameters where documented.
By default, scripts write under exports/interaction-quality/:
interaction-quality-weights-latest.jsoninteraction-quality-predictions-latest.jsonlinteraction-quality-attention-latest.jsonlinteraction-quality-report-latest.mdClassification is mechanical and uses no model tokens. The optional scheduled agent task reads the flagged set and may append reflection notes through the normal agent workflow.
Create the nightly task explicitly:
bun addons/late-night-regrets/scripts/setup-nightly-task.ts --cron '30 2 * * *'Installing the add-on alone does not schedule a task.
The classifier emits successful_execution, course_correction, misinterpretation, over_engineering, under_delivery, context_failure, good_proactive, or neutral.
bun addons/late-night-regrets/scripts/train-interaction-quality-bayes.ts
bun addons/late-night-regrets/scripts/train-interaction-quality-bayes.ts --recent-hours 48
bun test addons/late-night-regrets/index.test.ts