Distill Session Knowledge

distill-session-knowledge

Offline miner. Turns past pi session traces into durable artifacts. No new infrastructure — routes into skill_manage, the memory tool, and docs/ (indexed by context-mode FTS5). Dry-run by default; nothing is written until you review the plan and explicitly apply.

Engine: the published @blackbelt-technology/pi-dashboard-session-distiller package, invoked through its distill-session-knowledge bin (this skill package declares it as a runtime dependency).

When to use

• "Mine / distill my pi sessions", "what recurring mistakes do I keep making", "extract reusable procedures from my history", "turn session logs into skills".
• After a stretch of work, to harvest verified lessons before they decay.

Procedure

1. Dry-run the miner (no writes) for this project:

   npx --no distill-session-knowledge --cwd "$(git rev-parse --show-toplevel)"
   

   Add --n <k> to change the recurrence threshold (default 3). Add --json to emit the full routing plan as JSON for programmatic review.

2. Review the routing plan. Each entry shows signal -> sink [action] conf=…:

   • procedure -> skill_manage — candidate reusable SKILL.md.
   • fault -> memory(failure/tool-quirk) — a wrong-way→right-way tool fix.
   • user_correction -> memory(failure/correction) — +AGENTS.md flag means it establishes a rule and should also patch the relevant AGENTS.md row.
   • ask_user_decision -> memory(project/convention) — a recorded human decision.
   • documentation -> docs — a recurring how-to summary.
   • STALE flag = confidence below floor; skip or prune, do not write.

3. Apply once the plan looks right — persists the watermark + below-threshold candidates store and prints the final plan:

   npx --no distill-session-knowledge --cwd "$(git rev-parse --show-toplevel)" --apply --json
   

4. Execute the routed writes from the plan, using your tools (this is the haiku-class subagent distillation step). For each non-stale entry:

   • skill_manage → mcp__pi__skill_manage create/patch (query existing first).
   • memory → mcp__pi__memory add with the entry's target + category (memory_search first to dedup; merge into a match instead of duplicating).
   • docs → delegate the write to a docs subagent in caveman style per the Documentation Update Protocol, then ctx_index the new doc so it is FTS5-searchable (task 5.2).
   • +AGENTS.md entries → patch the relevant AGENTS.md rule via the docs subagent (≤200 chars/row, caveman style) in addition to the memory write.

Guardrails

• Dedup before every write — re-running over the same corpus must create zero duplicates (the planner marks merge when a sink already holds the artifact; still query the live sink, the planner's exists is conservative).
• Never write a STALE artifact.
• Stay within this project's sessions (the miner already scopes to the cwd dir).
• Respect the doc protocol: main agent orchestrates, a subagent writes under docs/.

Verification

• npx --no distill-session-knowledge --cwd "$(pwd)" prints a routing plan and mutates nothing (dry-run default; no watermark advance).
• Re-running the dry-run over the same corpus yields the same plan (idempotent).
• After apply + routed writes, ctx_search returns a freshly distilled doc.