wovol-lab

wovol-lab

A personal "ops lab" for Claude Code. What makes it distinctive (and what you don't usually see together in one Claude Code project):

• mnemos — a neuroscience-inspired memory layer modelling encoding, hippocampus-to-cortex consolidation, active recall, SM-2 spaced repetition, and Targeted Memory Reactivation (cwd / git-remote cues injected as additionalContext at session start). Claude remembers what a human teammate would, across sessions and projects, instead of starting cold each time.
• 4-profile browser isolation — school / work / personal / sensitive, each pinned to its own Chrome user-data-dir and CDP port. The sensitive profile (banking, health, government, formal applications) is hard-blocked from every orchestrator and every scheduled job, so automation literally cannot reach those tabs.
• 10-hook bilingual safety net — every write scanned against 30+ PII / secret patterns covering both Western identifiers and Japanese ones (マイナンバー, 保険証, 戸籍); every external send / publish / merge gated by explicit confirmation; all sends hard-blocked between 23:00 and 06:00; every tool call logged to a redacted audit trail. Submit-button matching covers 送信 / 提出 / 公開 / 出願 alongside English.

The actual ops the lab runs (morning digests, meeting prep, follow-up nudges, file organization, dev-flow phases) are downstream of those three choices. Outputs land as drafts you review on your phone with coffee — never auto-sends, never auto-clicks Submit.

MIT licensed. Published as an architecture reference; fork-and-adapt for your own setup.

What you'd actually see

Every morning at 07:30 a Gmail draft appears titled Morning Brief — YYYY-MM-DD:

# Morning Brief — 2026-05-12

## Now (decide today)
- review owner/repo#42 (PR sitting 3d, build green) — link
- reply to Alice re. Q2 milestones (no response in 5d) — link

## Today
- inbox: 4 newsletter unsubscribe candidates — link
- school: new announcement in <course> — link

## FYI (count only)
- 12 fyi items (newsletter / promo / read-only)

It's assembled by a morning-brief orchestrator that fans out to Teams, Gmail, GitHub, school portals, and ops health checks in parallel, then combines the results into a single page.

How it works (90 seconds)

• Skills (under .claude/skills/) are Markdown files that tell Claude Code how to perform an op (Gmail triage, GitHub notification triage, school portal scan, ...). Each writes structured JSON to a gitignored audit/ directory.
• Agents (under .claude/agents/) are subagents — isolated context, read-only by default, single-purpose. Skills call them for things like classification or summarization.
• Hooks (under .claude/hooks/) form a safety net. PreToolUse hooks scan content for secrets, block destructive shell, and gate every external send / publish / merge. PostToolUse appends a redacted JSONL entry per tool call.
• Profiles isolate the user-data-dirs: school / wovol / personal / sensitive each have their own Chrome profile + CDP port + domain allowlist. The sensitive profile is hard-blocked from orchestrators and scheduled jobs.

Prerequisites

• macOS. The scripts/launchd/ scheduler and the default browser paths in .env.example assume macOS. The skills themselves are platform-agnostic, but the scheduler and browser launcher are not.
• Python 3.10+ — python3 --version. setup.sh uses python3 for the command-wrapper generator and to pip install the browser-driven deps.
• Claude Code CLI — claude --version (install guide).
• Arc Browser at /Applications/Arc.app — this is the default browser for the personal profile and recommended. If you'd rather use Chrome, edit config/profiles.yml to set browser: chrome for the profiles you care about, and set CHROME_BROWSER_PATH in .env if Chrome isn't at the standard /Applications/Google Chrome.app location.
• Optional MCP servers — install via Claude Code's /mcp command (MCP setup docs) when you need the skills that depend on them:
  • Gmail MCP — required for digest/* and inbox/* skills
  • Linear MCP — required for dev-flow/* skills
  • GitHub MCP (or just the gh CLI) — required for github-inbox-triage and delivery/* skills
  • Playwright MCP — required for radar/* and form-fill-draft

You don't need all four to start. /profile-manager and mnemos/* work with zero MCPs.

Install as a Claude Code plugin (recommended)