INTP. Camus & Foucault enjoyer. AI native coding, agent infrastructure, and harness engineering — the scaffolding that keeps agents useful after the demo ends.

Preferred harness: Codex. Occasional companion: Hermes agent.

Every project here answers the same question: what breaks when you try to run agents for real?

Document Ingestion — agents can't read PDFs without hallucinating

• docling-skill — Manifest-based quality gating so the agent knows what's actually ready before it starts guessing. 4 releases, multi-OCR backends, Claude Code + Codex.

Agent Output — agents produce Markdown, humans need something readable

• md-for-human — Deterministic renderer: Markdown in, navigable HTML site out. The source stays clean, the artifact is disposable. Has a --verify flag and it's not afraid to use it.

Digital Worker Runtime — agents as operating roles, not chatbots

• knot-agent + knot-skills — Local-first runtime scaffolding for enterprise digital workers: workspace, knowledge, permissions, skills, deliverables, and multi-user IM workflows. Knot asks what an agent needs before it can act like a durable employee inside a company.

Infrastructure — agents that don't crash, because I taught them not to

• hermes-gateway-watchdog — Gateway + Cloudflare tunnel health monitor with staged recovery.
• openclaw-gateway-watchdog — Same pattern, different gateway. Cold-start tolerance so it doesn't mistake "waking up" for "dead."
• watchdog-command-receiver — Feishu IM command receiver. No public URL, no shell, audit-logged. Poke your infra from your phone.

The goal is not to make agents look magical. The goal is to make them useful, inspectable, and boring enough to trust.

"Simple enough to work, complex enough to be interesting."