oh-my-pi (omp)

A terminal-first AI coding agent — "a coding agent with the IDE wired in." A fork of Mario Zechner's Pi, extended with production coding features. The IDE-integrated, feature-rich end of the agentic-coding-harness spectrum (where claw-code is the minimal end). MIT, a Rust core plus TypeScript/Bun. (This line first said "~27k lines of Rust"; the figure was ~80k by the 2026-08-04 re-verification below and the project's own number today. Written by Can Bölük, copyright Stencil Labs, Inc.)

The upstream is now paged (2026-08-30), and it changes how to read this page. When this page was written the fork was here and Pi was not, so the spectrum was drawn against claw-code. Pi turns out to sit at the minimal pole itself, and deliberately: it refuses sub-agents, permission gates, plan mode and MCP, shipping each as an optional extension. oh-my-pi takes the same base and builds in 32 tools, LSP, debuggers and subagents — so the two are not two points on a spectrum of unrelated projects but the same codebase forked in the direction its author declined to go, which is a sharper comparison than the original framing allowed.

Distinctive design

• Hash-anchored edits — content-hash anchors instead of line numbers; claims up to 61% token reduction and no whitespace conflicts. A concrete bet that edit representation is a lever on cost/reliability.
• Optimized tool harness — 32 built-in tools behind a unified interface; in-process Rust implementations (ripgrep, bash, glob) instead of shelling out to system binaries.
• LSP integration — "everything your IDE knows, the agent knows"; workspace-wide refactors (renames). Debuggers (lldb, dlv, debugpy) as first-class tools — vs competitors' print debugging.
• Persistent Python/JS kernels with bidirectional tool-calling; Puppeteer browser (stealth); Slack/Electron integration.
• Subagents — isolated parallel workers returning structured JSON, not prose to parse (agent-orchestration).
• 40+ LLM providers (see llm-provider, cross-wiki); auto-discovers existing rule files (Cursor/Claude/etc.); preview-before-apply UX; extensible via TypeScript modules.

Why it belongs here

Another agentic-coding-harness instance, and the clearest example yet that the harness competes on engineering depth (LSP, in-process tools, debuggers, edit encoding), not just prompts — the "structure as a substitute for capability" thesis taken to the IDE layer. Its provider-agnosticism (40+) reinforces that the harness, not any one model, is the durable unit. (Caveat: README claims, incl. the 61% figure — unbenchmarked here.)

Related

pi-coding-agent · mario-zechner · agentic-coding-harness · claw-code · agent-orchestration · agentsys · llm-provider

Re-verified 2026-08-04 (staleness queue)

21.8k★ / 2.1k forks, MIT. Scale is larger than this page recorded: 31 built-in tools, 60+ model providers, ~80,000 lines of Rust core, with ripgrep, glob, bash, 46 vendored coreutils and tree-sitter all in-process to avoid fork/exec on hot paths.

Two things bear on this spoke's threads. It ships LSP integration and real debugger support (lldb, dlv, debugpy, DAP) — the agent gets the same code intelligence a human editor has, which is a different answer to grounding than reading files. And it runs five ways: TUI, one-shot CLI (omp -p), Node SDK, stdio RPC, and ACP mode for editor integration — the same headless/embeddable spread claude-agent-sdk represents for claude-code.

Re-seen 2026-09-06 — the product site (omp.sh), and the benchmark that was missing

Arrived again via Telegram, this time as https://omp.sh/ rather than the repository. Read together with the README and the GitHub API the same day. Two things changed that this page had been waiting for, and a third that nobody was waiting for.

It has a documentation site now. omp.sh carries a full docs tree — quickstart, providers, settings, sessions, plan mode, subagents, collab, skills, MCP, slash commands, keybindings, code intelligence, debugging, web, GitHub, CLI, RPC, SDK, extension authoring. Every earlier reading of this project was from a README. The docs read as written for a user rather than a buyer, and they say the quiet part directly: "omp can modify files and run commands in your environment, so review its work as you would a teammate's change." The front page does not read that way.

The 61% claim has a benchmark behind it. This page's standing caveat — "README claims, incl. the 61% figure — unbenchmarked here" — is now partly answered by The Harness Problem, a 16-model × 3-format run with published fixtures and per-run reports. It is the author's own benchmark of his own format and remains T3, so the caveat narrows rather than lifts: hashline beats apply_patch in 14 of 16 models and averages +15 points, the weakest models gain most, and the two strongest gain nothing or lose. See agent-edit-format for the mechanism and the independent work that supports the general claim without endorsing this format.

Windows, natively. The front page's "skip the WSL" is a real claim about the architecture: the in-process approach that started as a fork/exec optimization — ripgrep, glob, find, and now 58 coreutils plus a brush bash with sessions that survive across calls, all compiled into the binary — removes the dependency on a Unix userland, so the same binary runs on macOS, Linux and Windows. It is the only harness in this spoke whose performance argument and whose portability argument are the same argument. The author's background is Windows kernel work, which is where that instinct comes from.

Where it has grown to

Also from the API: 2,391 open issues, 21,374 commits, 780 branches, 841 tags, TypeScript as the reported top language, MIT, repository created 2025-12-31 and pushed the same hour this page was written. Eight months old and past 29k stars — the fastest accumulation of any harness here except Pi itself, and the open-issue count is a real number to sit beside the star count.

Governance detail worth keeping: pull requests are "temporarily open to everyone as a trial." The project previously required a vouch before accepting a PR, and the README says the vouch system may return depending on how the trial goes. A 29k-star MIT project that gates contribution by invitation and is experimenting with opening it is an unusual posture, and this spoke has no other example of it.

The feature surface has left the IDE framing behind

This page has called omp "the IDE-wired end of the harness spectrum" since June. That framing no longer covers it. Beyond LSP and debuggers, the README now documents:

• Curated memory across sessions — the agent writes facts mid-run (retain), captures lessons (learn), searches them back (recall), and compresses each session into a mental model loaded on the next session's first turn. Backend is switchable: local, Hindsight, or Mnemopi. Project-scoped by default. This is agent-memory as a first-class built-in rather than a convention, and it is gated off by default.
• An advisor model watching every turn — a second model on its own context reads what the main agent does and injects inline notes, from an aside to a hard blocker. Two models, different roles, same session.
• Time-traveling stream rules — a regex match aborts generation mid-token, injects the rule as a system reminder, and retries from that point, so a rule costs context only when it fires. The injection survives compaction.
• Live session sharing — /collab puts the session on a relay and returns a link and a QR code; a teammate joins from another terminal or a browser, read-write to pair on the same agent or read-only to watch. Frames are sealed client-side, so the relay never sees provider keys.
• Sixteen internal URL schemes resolved inside the ordinary filesystem tools — read pr://1428 returns the same shape as read src/foo.ts, grep walks a diff like a directory, agent://<id>/findings.0.path pulls one field out of a subagent's structured output, and each merge conflict becomes a conflict://N you resolve by writing @ours, @theirs or @base. The stated reason is to avoid teaching the model a new tool per integration.
• Desktop and browser control — a Chromium/Electron tab runtime with stealth on by default, plus computer helpers that enumerate windows, screenshot, send native input and walk the OS accessibility tree.
• It reads other tools' config in place — Cursor MDC, Cline .clinerules, Codex AGENTS.md, Copilot applyTo and four more, in their native shape, with no migration step. A harness competing on not asking you to port anything.

Set against Pi, the 2026-08-30 reading holds and sharpens: the upstream ships its refusals as optional extensions, and this fork built every one of them in and kept going. What is new is that the fork now has features Pi never refused because nobody had proposed them — a relay for shared sessions, an advisor model, a desktop AX tree. The two projects are no longer the same question answered twice.

Still unmeasured: everything above except the edit-format numbers. Provider count, tool count, LSP and DAP op counts, the memory and advisor designs, the collab relay's key handling — all are the project's own description, and the only figures independently obtained are the GitHub counts. (Re-checked 2026-09-14: editorial head-to-heads now exist — Composio's Pi vs OMP and StandardCompute's scoring — but they are ratings, not measurements, so the sentence stands.) Corrected 2026-10-01: the 09-14 re-check misread Composio. Its Pi vs OMP post (Shrijal, 2026-09-01, composio.dev/content/pi-vs-omp) is a measurement: eight harnesses, one model (DeepSeek V4 Flash), 30 tool-use tasks with the same verifier and a 900 s limit. Pi passed 20/30, OMP 17/30; OMP's median task took 272.4 s against Pi's 132.2 s, used 742,283 tokens per task against 558,885, and cost $0.103 per success against $0.028. Composio flags its own caveats (Pi ran on high rather than max reasoning; most Pi trials went through DeepSeek's own API), and 30 tasks is one small run. StandardCompute's scoring is still a rating. The front page also teases "omp², soon!" with no detail attached.

Re-seen 2026-09-07 — the repository itself, one day later

The curator sent the GitHub repository, which is the URL this page has carried since June, the day after the omp.sh reading above. Checked against the GitHub API rather than assumed: 29,934★ / 3,047 forks / 2,389 open issues, pushed the same hour. That is +156 stars and +12 forks in a day, with the open-issue count down by two — the accumulation rate the 2026-09-06 entry described is holding, not a one-off spike around the site launch. Nothing in the README moves any claim above, and "omp², soon!" still has nothing attached.

Two of the project's own counts that earlier readings did not record: 23 web-search backends (Perplexity through DuckDuckGo) and 50+ tree-sitter grammars behind the AST operations. Both sit with everything else in the "Still unmeasured" list — they are README figures.

Related (2026-09-06)

harness-problem-hashline · agent-edit-format · can-boluk · stencil-labs · pi-coding-agent · agent-memory · agents-md · agentic-coding-harness