用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/CleanExpo/skills-library --skill sync-claude-sessions命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Dispatch the eight operational roles against a task, each grounded in the vault before it reads any code, with a live view of what is running. Use for work that wants several independent lenses at once — map, plan, build, verify, review, security, CI, release — rather than one agent doing everything.
Thin wrapper around the Pi-CEO Hour-1 provisioner. Given a client slug + email, invokes `swarm/inbox/provisioner.py` to spin up the full Linear project + Supabase portal_content + ContextBot Telegram intake bot + welcome email + Phill ping. Use when the user says "onboard {client}", "spin up portal for", "new client setup", or right after `stripe-milestone-invoice` confirms a deposit cleared.
PM-Core — the first autonomous execution agent. Claims the highest-priority agent-ready ticket from the Unite-Group Linear board, clones the repo, implements the fix on a feature branch, runs CI, and opens a PR. Model: claude-sonnet-5 (senior agent tier). Trigger: /pm-core [ticket-id] or autonomous via Hermes cron.
正在显示 SKILL.md
| 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.