repodocs
Generates a wiki of your repository where every claim points at the source it came from. Runs with OMP, Claude Code or Codex CLI.
uv tool install repodocsOpen source and AI · Ary Rabelo
AI gives you speed. Direction is still engineering work.
Coding agents write code fast. Understanding the system, reviewing the change and proving it works is still the job. Here is what I publish for that: itaruby, a type checker for Ruby, PRs merged into Rails, Discourse and fastlane, and tools for working with coding agents.
Open source
itaruby: a Ruby type checker that finds bugs with no annotations
Written in Rust, open source (MIT) and still alpha. I run it on the code of large projects and only send a fix for what I can prove:
- itaruby reads the project with no type annotations and points at where the code will break.
- A warning only counts if I reproduce the real error, a NameError, NoMethodError or ArgumentError, by running the project’s own code.
- The fix only goes out when that same run passes. False positives get written down, not hidden.
Merged into other people’s projects
30 pull requests merged and 11 open, across 12 projects, since May 2026.
- rails/rails itaruby 3 merged
- discourse/discourse itaruby 6 merged
- fastlane/fastlane itaruby 1 merged
- gastownhall/beads 1 merged
- coleam00/Archon 1 merged
- Chachamaru127/claude-code-harness 1 merged
- serpapi/serpapi-mcp 1 merged
- getpaseo/hub 1 merged
- nicolasmelo1/software-factory 15 merged
Marked itaruby: fixes that came out of a bug hunt with it.
See every contribution →AI tools I publish
What I built to work better with coding agents, open with the code on GitHub. Where the README has an install command, it is on the card.
For coding agents
See everything in Projects →next-action-protocol
Every agent response ends in exactly one clear next action: [DONE], [DECIDE], [HUMAN] or [WAIT].
npx skills add aryrabelo/next-action-protocol --agent claude-code -gserpapi-agent-toolkit
SerpApi search for agents, as a CLI and a Claude Code skill, using far less context than an MCP server. I measured the difference.
searchapi-agent-toolkit
The same idea for SearchApi: a zero-dependency CLI and a Claude Code skill that returns only the fields you ask for.
jev-accept
Reads the issue and epic behind a pull request and judges whether it delivers what was asked. Local CLI and GitHub Action, same contract.
OMP plugins
See everything in Projects →omp-burnrate
Per-account subscription burn rate when OMP opens: which account is about to hit its cap before the window resets.
omp plugin install github:aryrabelo/omp-burnrateomp-green-gh
Your repository's pull requests in the OMP statusline: is the current branch green, what is blocking it, and what else is open?
PlanQueue
A per-session task queue: write the plan once, then dispatch it to the agent one line at a time.
omp plugin install github:aryrabelo/planqueueSharing what I learn
What I learn, I leave open.
“I got AI all wrong! I did more, built more and cared less, until I started leaving behind what mattered. Focus on understanding, on being the human AI is missing.”
-
Notes
Every note starts from a real problem, with the number I measured and what I changed because of it.
Read the blog → -
The gauntlet method
One agent builds. Another agent, which does not know which side is mine, compares the result against an outside reference and names the biggest gap. The round repeats until my side wins. That is how I let an agent work without taking its word about its own work.
-
Your first day with Claude, no steps skipped
claude-101: from your first prompt to automating one task of yours. 5 short modules, in English and Portuguese, at your pace.
Start the free course →
Notes on what I measured
Every note starts from a real problem, with the number I measured and what I changed because of it.
- Article Why I built aryrabelo.com from scratch A personal hub that is a lead engine, a self-hosted link-in-bio, a bilingual portfolio, and a blog. All self-hosted on Cloudflare, no data outsourced.
- Article I measured MCP vs a CLI for agent search. The MCP used 17x more tokens per call. The same Google search through SerpApi's serpapi-mcp and through serp, a small MIT CLI. The MCP put 771 tokens into context before searching and 6,047 per call; the CLI, 0 and 351. The measurement, and why it matters for agents.
Ask for a course
I want to learn ___
Em Paralelo · one issue a week
What broke this week, and what I changed
Measurements, mistakes and decisions from my daily work with agents. No hype, no news roundup.
Hi, I’m Ary
Software engineer for twenty years, Rails since 2012, and a platform serving 1.8 million visits a month in production. Today I write open source in Ruby and tools for coding agents, and I use those agents every day.
Everything I publish here is open, or a course you take on your own.
Read the full story →