can1357/oh-my-pi

The central bet in oh-my-pi (omp) is that the harness itself is the bottleneck. The benchmark table makes this concrete: Grok Code Fast 1 jumps from 6.7% to 68.3% pass rate — a tenfold lift — purely from fixing the edit format. Grok 4 Fast spends 61% fewer output tokens once the retry loop on bad diffs disappears. MiniMax pass rate more than doubles on the same weights and same prompt. That's the argument the whole project is built around.

The Rust core is where that argument gets implemented. Roughly 80,000 lines across six crates handle search, shell execution, AST operations, PTY, desktop control, and image decode — all in-process on the libuv thread pool. The embedded bash engine alone runs to about 38,000 lines and ships command-line utilities compiled directly into the builtins crate: ls, sed, sort, jq, ripgrep-backed grep, and more. No fork-exec on the hot path. The same binary runs on macOS, Linux, and Windows without a WSL bridge.

The tool surface is unusually coherent. LSP renames go through workspace/willRenameFiles, so barrel files and re-exports update before the file moves. The debugger integration covers lldb, dlv, and debugpy through DAP — the agent attaches, steps, reads frames, and inspects variables rather than inserting print statements. The hashline edit format uses content-hash anchors instead of line numbers, so stale-file patches get rejected before they corrupt anything rather than silently applying to the wrong location.

Subagent coordination is first-class rather than bolted on. The task tool fans work into isolated worktrees, each worker runs its own tool surface, and results come back as schema-validated objects the parent reads directly. The advisor role pairs a second model to watch every turn and inject concerns inline on its own context budget. Time-traveling stream rules abort a generation mid-token when a regex fires, inject a system reminder, and retry from the same point — course-correction without paying context cost on every turn.

The provider story is wide: 60-plus providers, ten routing roles, fallback chains, round-robin credentials, and path-scoped model assignments. It reads config formats from eight other agents — Cursor MDC, Cline rules, Codex AGENTS.md, Copilot applyTo — without a migration script.

This is a fork of Mario Zechner's Pi, extended aggressively. The scope is large enough that "batteries included" undersells it; this is closer to a full platform with a clear performance thesis behind every design choice.