メニュー
B.U.I.L.D. Inflow pipeline
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/CleanExpo/skills-library --skill sync-claude-sessionsコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
| name | sync-claude-sessions |
| description | B.U.I.L.D. Inflow pipeline |
Turns your own terminal history — the richest training data you own — into knowledge the self-improving loop can act on. One redacted digest per session.
Walks ~/.claude/projects/**/*.jsonl (808+ files, ~442 MB) and for each session
extracts: the opening intent, follow-up user corrections/decisions, the
tools used (histogram), the conclusions (last assistant output), and any
shipped PRs (pr-link events). Every text field is redacted before it
touches disk. Output is an OKF-formatted session-digest per session in
~/2nd Brain/2nd Brain/process/sessions/.
This is the content pass. It does NOT duplicate the FABLE distiller
(~/Fabel Prompt Engineer/scripts/fable-distill.mjs), which mines the same lake
for working-rhythm metrics. Run both; they are orthogonal.
cd ~/Pi-Dev-Ops
python3 scripts/sync_claude_sessions.py --dry-run --limit 5 # safe sample → scratch
python3 scripts/sync_claude_sessions.py --since-last # incremental (routine default)
python3 scripts/sync_claude_sessions.py # full re-scan
Flags: --dry-run (writes to /private/tmp/.../scratch-session-digests, never
updates the marker), --since-last (skip files unchanged since last run, via
process/.sync-claude-sessions.json), --limit N, --out-dir PATH.
The redactor (redact()) strips full-length secret shapes BEFORE any write,
including the two that scripts/secrets_check.py misses: sk-ant-oat
(Anthropic OAuth) and AIza (Google), plus Anthropic API, GitHub PAT/fine,
Linear, Slack, AWS, Stripe live, OpenAI, JWT, Bearer, and password/secret
assignments. Verify after any change to the pattern bank:
python3 scripts/test_sync_claude_sessions.py # 23 assertions, must be ALL PASS
Gate (redactor as oracle): re-applying redact() to every output file must
change nothing. If it would, a secret leaked — fix the bank, re-run.
session_id, so their digests dedup to the latest
(744 sessions → 583 distinct digests on first run). Digests are derived data —
safe to delete and re-mine anytime.process/sessions/index.md lists
the new digests: python3 ~/pi-seo-workspace/unite-group/apps/workspace/scripts/okf-index.py "$HOME/2nd Brain/2nd Brain".~/Pi-Dev-Ops/docs/specs/2026-06-29-self-improving-system-build.md.improve-system reads these digests + the vault and proposes
graded improvements; /data-ingestion orchestrates this + source-ingest + wiki sync.