analyzing-all-conversations

Analyzing All Conversations

Core principle: Numbers come from the deterministic engine (memory-bank analyze), meaning comes from fact/ontology search. Never fill gaps with guesses — report actual coverage numbers.

Step 0 — Locate the plugin root

PLUGIN_ROOT is two directories up from this skill's base directory (the directory containing cli/, scripts/, skills/).

Step 1 — Run the deterministic analysis (read-only)

node "$PLUGIN_ROOT/cli/memory-bank.js" analyze

Options: --top <n> (projects, default 15), --months <n> (timeline, default 12), --json (for scripting), --out <file> (save report).

This returns, without any LLM calls:

• Coverage — conversations / sessions / exchanges / projects / date range
• Pipeline coverage — fact extraction (done/pending %) and summaries (done/missing %)
• Facts — active count, by category, by scope
• Top knowledge domains (ontology)
• Per-project rollups — conversations, sessions, exchanges, facts, activity range
• Monthly activity timeline
• Recommendations — which backfills to run

Step 2 — Fill analysis gaps (backfill, run in background)

If the report recommends backfill:

Gap	Action
Pending fact extraction	node "$PLUGIN_ROOT/scripts/backfill-extract-worker.js" --max 500 — run in background. Requires ANTHROPIC_API_KEY. Lock-protected and idempotent: safe to start, exits immediately if another worker is already running.
Missing summaries	node "$PLUGIN_ROOT/cli/memory-bank.js" sync — generates up to 10 summaries per run; repeated syncs (each session start) drain the queue gradually.

Do NOT wait for backfills to finish. They are long-running background work. Report them as "진행 중 (백그라운드)" with the log location (~/.config/superpowers/conversation-index/backfill-extract.log), never as "완료".

Step 3 — Synthesize the organized report

Enrich the numbers with meaning using MCP tools (do not re-derive numbers by hand):

• graph_stats — knowledge graph overview (domains, relations)
• search_facts with category: "decision" / "pattern" / "constraint" — representative facts for the top projects
• cross_project_insights — transferable lessons relevant to the current project

Then present the final report in the user's language (Korean → 존댓말) with this structure:

1. 전체 개요 — 대화/세션/exchange 수, 기간, 프로젝트 수
2. 분석 커버리지 — extraction/summary 커버리지 표. 백필을 시작했다면 "완료된 것"과 "백그라운드 진행 중인 것"을 명확히 구분
3. 프로젝트별 정리 — top 프로젝트 표 + 프로젝트별 핵심 결정/패턴 1–2줄
4. 핵심 지식 — 도메인 분포 + 대표 facts (category별)
5. 활동 타임라인 — 월별 추이 요약
6. 남은 공백과 조치 — 미분석 세션 수, 실행한/권장하는 backfill

Rules

• Never claim full coverage unless pending == 0 and missing summaries == 0. Always report actual N/M numbers.
• Backfill started ≠ analysis complete — report scope honestly (foreground done / background running).
• Everything here is additive: analyze is read-only; backfill workers only add facts/summaries, never delete.
• Keep the report to roughly one page unless the user asks for more depth. Detail belongs in --out files.