🪴 Brain

oh-my-pi

6 min read

oh-my-pi

🗒️ can1357/oh-my-pi (omp) is a terminal coding agent with the IDE wired in, built by Stencil Labs. It is a fork of Pi by Mario Zechner, rewritten as a coding-first surface: sessions, subagents, slash commands and extensions, all TypeScript on top of a Rust native core. MIT. README headline numbers: 60+ providers · 31 built-in tools · 14 LSP ops · 28 DAP ops · ~80k lines of Rust.

🚀 The thesis, from the project’s own post The harness problem: the tool surface, not the model, is what makes agents fail. If the edit format stops eating the model alive, the same weights do much better.

Description

🧩 What stands out (all as described in the README, none verified by me):

  • Hashline edits. The model points at content-hash anchors instead of retyping the lines it wants to change. A stale file makes the anchors diverge and the patch is rejected before it corrupts anything. Claimed effects: Grok Code Fast 1 from 6.7% to 68.3%, Grok 4 Fast using 61% fewer output tokens, MiniMax 2.1x pass rate. These are the vendor’s own numbers.
  • LSP wired into every write. A rename goes through workspace/willRenameFiles, so re-exports, barrel files and aliased imports update before the file moves. “Everything your IDE knows, the agent knows.”
  • Drives a real debugger (DAP). Attaches lldb, dlv or debugpy, steps to the bad pointer and reads the frame, instead of sprinkling print statements.
  • Time-traveling stream rules (TTSR). Rules sit dormant until the model goes off-script. A regex match aborts the stream mid-token, injects the rule as a system reminder and retries from the same point, so you only pay context for a rule when it is needed. Injections survive compaction.
  • First-class subagents. task fans out into isolated worktrees; each worker yields a schema-validated object the parent reads directly. Alt+A opens an Agent Hub to watch, steer, revive or kill any worker.
  • Advisor model. A second model assigned the advisor role reads every turn and injects notes (aside, concern or hard blocker) on its own context.
  • Nine model roles (default, smol, slow, plan, commit, vision, task, advisor, tiny), fallback chains on 429 or quota walls, path-scoped models, and round-robin credentials across stacked API keys.
  • Everything is a path. Sixteen internal schemes (pr://, issue://, agent://, skill://, ssh://, conflict://…) resolve inside the normal read / grep / write tools, so a PR or a merge conflict is read and resolved like a file.
  • Native, even on Windows. ripgrep, glob, find, a bash implementation (brush) and 58 coreutils-style utilities run in-process, with no fork/exec and no WSL bridge. Prebuilt for linux, darwin and win32 on x64 and arm64.
  • Memory the agent curates (retain, recall, reflect, learn), with memory.backend of local, Hindsight or Mnemopi, scoped per project by default.
  • Inherits what other tools wrote. On first run it reads rules, skills and MCP servers from .claude, .cursor, .windsurf, .gemini, .codex, .cline, .github/copilot and .vscode, with no migration script.
  • Four entry points: TUI, one-shot omp -p, a Node SDK (@oh-my-pi/pi-coding-agent), and RPC / ACP over stdio. ACP lets Zed drive it with the editor’s own buffer and terminal.
  • Also: /review with P0 to P3 priorities and a verdict, omp commit that splits changes into atomic commits, ast_edit previews staged until you accept, a web_search that walks 23 providers, browser and desktop-control tools, /collab live session sharing over an end-to-end encrypted relay, and prompt keywords ultrathink, orchestrate and workflowz.

Download or use

bun install -g @oh-my-pi/pi-coding-agent   # recommended, needs bun >= 1.3.14

Also brew install can1357/tap/omp, mise use -g github:can1357/oh-my-pi, and a Nix flake. The macOS / Linux and Windows one-liners (curl … | sh, irm … | iex) pipe a remote script straight into the shell, so I would read them first or stick to bun. Custom OpenAI-compatible providers go in ~/.omp/agent/models.yml; roles in ~/.omp/agent/config.yml.

Reasoning for

I have not run it. It is on the radar because it is the most complete worked example of harness engineering done in the tool layer rather than the prompt layer, which is the lens in Harness Engineering and Harness Engineering (Ryan Lopopolo). Three ideas are worth stealing even if I never switch:

  • Edit format as a first-class design problem. Content-hash anchors avoid the string-not-found retry loop that burns tokens in every str_replace-style harness.
  • Just-in-time rules (TTSR). Dormant guardrails that fire on a regex match are the “shift interventions right” principle from the Lopopolo note made concrete, and a cheaper alternative to a bloated always-on CLAUDE.md.
  • A reviewer model on its own context (advisor) is the separate-judge pattern from 4 Claude Code Upgrades for Making Money (Nate Herk) built into the harness.

⚠️ What I would check before trusting it:

  • The benchmark gains are the vendor’s own, on specific models, and pair with a blog post I have not read in full.
  • A huge feature surface for one project, and the README says pull requests are temporarily open to everyone as a trial, with the vouch system possibly returning. Watch the contribution policy and supply chain.
  • It reads rules, skills and MCP servers from other tools’ folders automatically, which is convenient but means the trust boundary is wider than one config file. See AI Agent Security.

Alternatives considered

  • Claude Code: the incumbent in this vault; extension model is skills, MCP, hooks and now mods (Claude Code Mods (RoboNuggets)). Its edit tool is exact-string replacement, the failure mode hashline targets, but it is tightly integrated with Anthropic’s models.
  • Cursor and GitHub Copilot: IDE-first rather than terminal-first.
  • Hermes Agent and DeepSeek Harness: other open, hackable agent harnesses, with a different emphasis (messaging gateway, plugin-everything).
  • Serena and MCP Language Server: add LSP-grade code intelligence to an existing agent over MCP, which is the lighter way to get part of what omp bakes in.
  • OmniRoute and LiteLLM: gateways that do the model routing outside the agent. omp already ships fallback chains and round-robin credentials in the harness, so for one agent the gateway layer is optional.
  • Pi itself (badlogic/pi-mono): the upstream project omp forks, if the batteries are not needed.

Resources

Template: tool

Graph View

Backlinks