@blackbelt-technology/pi-dashboard-session-distiller
v0.7.0
Published
Offline miner: distill reusable artifacts from pi session JSONL logs
Downloads
645
Readme
@blackbelt-technology/pi-dashboard-session-distiller
Offline miner that distills reusable, verified knowledge from pi session JSONL
logs. It is the deterministic engine behind the distill-session-knowledge pi
skill; use the skill package
@blackbelt-technology/pi-dashboard-distill-session-knowledge
to load the discipline into a pi session, or use this engine directly as a CLI /
library.
What it does
Walks the pi sessions for a project (newer than a persisted watermark), segments each trajectory, and extracts five signal classes anchored on objective outcomes:
- fault — a tool error that a later same-tool retry resolves (
isErrorflip). - user_correction — a human correction after an assistant action.
- ask_user_decision — a recorded human decision.
- procedure — an episode of >5 tool calls that ends verified-good.
- documentation — a recurring how-to summary.
Only patterns that recur across >= N sessions (default 3) are promoted. Each
promoted cluster is distilled into an artifact with provenance and a
confidence score that decays with age and model change, then routed to a sink
(skill_manage, memory, or docs).
Privacy
The miner is dry-run by default — nothing is written until you pass
--apply. State (the watermark and below-threshold candidate store) is written
to ~/.pi/agent/distill-session-knowledge/ on your machine; it is never
bundled or transmitted. Distilled artifacts land only in your own local sinks
after review. The emitted plan carries slugged signatures and provenance
(session ids, model), not raw session payloads.
CLI
# dry-run over the current project (no writes)
npx distill-session-knowledge --cwd "$(git rev-parse --show-toplevel)"
# change the recurrence threshold and emit the routing plan as JSON
npx distill-session-knowledge --cwd "$(pwd)" --n 3 --json
# persist the watermark + candidate store and print the final plan
npx distill-session-knowledge --cwd "$(pwd)" --apply --jsonFlags: --cwd <dir>, --n <k> (recurrence threshold), --sessions-dir <dir>,
--apply, --json.
Library
import { run } from "@blackbelt-technology/pi-dashboard-session-distiller/main.js";
const result = run({ cwd: process.cwd(), n: 3 }); // dry-run
console.log(result.plan);License
MIT
