I work on repo-scoped hooks, verification gates, lifecycle reminders, and local analytics for AI coding agents. The center of gravity is practical infrastructure for Claude Code, Codex CLI, and GitHub Copilot CLI: tools that make agent work easier to trust in real repositories.
- Agent hook runtimes that work across shells, operating systems, and coding-agent clients.
- Verification loops that turn "the agent says it is done" into something the machine checks.
- Local-first analytics and guardrails for developers who already use AI coding tools heavily.
| Project | What it is for | Start here |
|---|---|---|
| lattice | Repo-scoped AI client hook runtime for Claude Code, Codex CLI, and GitHub Copilot CLI. Shared policy gates, lifecycle hooks, and provider integrations for real repos. | pnpm add @lzong.tw/lattice |
| clawback | Verification loops for Claude Code and Codex: protected files, post-edit format/lint, stop-time typecheck, and context-compaction reminders. | npx @lzong.tw/clawback |
| claude-code-usage-advisor | Read-only local analyzer for Claude Code history. It recommends practical model, effort, permission, sandbox, and prompt-cache profiles. | npx claude-code-usage-advisor |
| pdf2epub-pro | High-fidelity PDF to EPUB pipeline with chunked parsing, restored PDF links, offline referenced-content appendices, and clean cover generation. | Clone and run from source |
| readwise-reader-management | Python scripts for Readwise Reader automation: tags, article operations, exports, and batch account maintenance. | Clone and configure locally |
These badges show public registry download counts, not validated user adoption. I treat them as early distribution telemetry and pair them with issues, stars, public references, and install feedback before making traction claims.
If one of these tools saves you a debugging session, a star on that repo is the cleanest signal that the experiment is worth keeping alive.
I do not want AI-generated volume, drive-by PR noise, or generic personal branding. The useful work is the shipped tool, the repro, the release note, and the issue that documents a real failure mode.