agents

How I work with AI agents, and everything that makes it work: the Claude Code skills and hooks I use every day, the memory server they all share, how my phone and both sides of my laptop stay in sync, the two USB sticks behind the Hackintosh work, how I keep up with related projects, and the two working methods (caveman and BMAD) that shape every session.

It's also the install source. A new machine (or a fresh macOS login) gets the whole setup with:

git clone https://github.com/watchthelight-labs/agents ~/src/agents
python3 ~/src/agents/install/install.py          # python on Windows

then the three secrets by hand (docs/new-machine.md). Or just tell Claude Code "set up from watchthelight-labs/agents".

The setup at a glance

 iPhone ─── photos, Claude web ──► watchthelight memory ◄── every Claude Code session (boot hook)
   │                                  (mcp.example.org)          ▲                ▲
   └── SSH over Tailscale ──► Legion 5 15IAX10 ─────────────────┘                │
                                 ├─ Windows 11  ── Claude Code, hooks, skills ────┤
                                 └─ macOS 26.7.1 ─ Claude Code, same skills ──────┘
                                 USB 1: OPENCORE (boot fallback, logs, "For Mac")
                                 USB 2: PROBE1401 (offline hardware scanner)
 GitHub org ── org webhook + Actions ──► Discord channel (commits, releases, upstream news)

What's here

Path	What
claude/skills/	17 Claude Code skills: boot-attempt analysis, upstream scouting, deep research, brainstorming, memory management, caveman, and more (index)
claude/hooks/	The hooks behind every session: memory boot, guardrails, command audit log, per-turn ledger, compaction guard, quota advisory, notifications (how they fit)
claude/statusline/	The status line: model, context use, 5-hour and weekly quota with burn rate
claude-web/skills/	Skills uploaded to claude.ai (the boot-screen transcriber the phone uses)
mcp-server/	watchthelight, my memory MCP server: notes, a searchable archive, an encrypted credential vault, OAuth for claude.ai and ChatGPT, a dashboard. Full setup guide inside
install/	`install.py` puts it all in `~/.claude`; `collect.py` refreshes this repo from my machines (sanitized)
docs/devices.md	Phone, Windows and macOS: which channel carries what
docs/usb.md	The two USB sticks and the routines around them
docs/upstream.md	How I find out what related projects did today
docs/memory.md	The memory protocol every agent follows
docs/methods/caveman.md	Caveman: terse replies by default, and when they switch off
docs/methods/bmad.md	BMAD: interrogate the idea before building, then build in gated phases

What isn't here

Secrets (tokens, keys, the vault, the server's private folder), my permission rules, personal project skills, skills built on code I can't share, and other people's skills I only use (those are linked in THIRD-PARTY.md). Everything published goes through tools/scan.py first, and CI runs it again on every push.

Licence

ISC for my code (LICENSE). The caveman skills and hooks are vendored from JuliusBrussee/caveman under its Apache-2.0 / MIT terms (licenses/). The whatif skill bundles Paged.js (MIT) and the xkcd script font (CC BY-NC 3.0).