Oh My Pi

You will like this if you have watched an agent sprinkle print statements into a segfaulting binary and wished it had a debugger. The daily trait is that the IDE's machinery is in the tool: renames go through the language server's willRenameFiles so imports and barrels move first, and debugging attaches lldb, dlv or debugpy and steps to the bad frame. A second model can sit in the advisor role and interrupt with a note.

Where it hurts is the size and the pace of change. Pick it for real work on a compiled codebase. Pick Pi if you want the parent with less in it.

The risk is reach. The browser tool ships with stealth on, so sites see a normal user, and the same API attaches to any Electron app; the README's own example points it at Slack so the agent reads your direct messages. A relay extension lets it adopt Chrome tabs you are logged into. A lot of authority for a tool that also sends native input to the desktop.

The consequence: run it under an account with nothing to leak. Governance is unsettled; pull requests were vouch-gated and are open only as a trial. What it does right: ast_edit stages a proposed change and applies it only after an Accept, atomically.

The headline number, Grok Code Fast from 6.7% to 68.3%, comes from the author's post, and the methodology is stated. 1. Fixtures are real files with a mechanical mutation, such as a removed guard clause, to be restored. 2. Three runs of 180 tasks, sixteen models, three edit formats, a fresh session and four tools each time. 3. Success is comparison against the original file after formatting. Total spend is disclosed.

This is reproducible and self-reported, not independent, and the fixtures favour edits that reverse a change exactly, which hashline is good at. The observation: the post's own table shows one model losing ground, and reports it anyway.

The copyright line reads Zechner, then the author, then Stencil Labs, Inc., the story of an individual's fork becoming a company. Distribution is other people's subscriptions: the provider list signs into Cursor, GitHub Copilot, Devin and a row of coding plans and routes work through them. The author's own post records Anthropic cutting a rival harness off from subscription access; that is the business risk, stated by the founder.

Moat: the harness; the post argues harnesses matter more than models, a thesis with no revenue attached. Likely acquirer: Zed, which it already speaks ACP with, or a model vendor that wants a harness. Position: long the code, flat the company.

The demo is an agent attaching a debugger. For sixty seats the software is free, the meter is whichever of sixty-plus providers each engineer signs into, and one command hands out a link and a QR code to a live session that anyone holding the link can watch, the sentence security will underline. There is no SSO, no SCIM and no audit trail. CI fit is fine: a one-shot flag runs a prompt and exits.

The fleet story is good: a Nix flake and a Home Manager module can pin the version and the settings declaratively for every machine. Onboarding is heavy: thirty-one tools to learn. Not yet.

This is the one. MIT, and ~/.omp/agent/models.yml takes any provider that speaks openai-completions, anthropic-messages, bedrock-converse-stream or six other API shapes; Ollama, llama.cpp and vLLM run with the key left blank. retry.fallbackChains fails over per role when a provider throws 429s, and stacked keys rotate with per-credential backoff. It reads the rules and MCP servers already in .claude, .cursor, .windsurf and the rest, so nothing migrates. An extension is a TypeScript module with the same tool API the built-ins use, loaded by /reload-plugins.

The eval kernel calls back into the agent's own tools from inside Python. No grudge available. I would fork it just to keep it.