Skip to main content

​
omp (oh-my-pi)

Fork of badlogic/pi-mono by can1357. Positions itself as the batteries-included coding agent: same TypeScript + TUI shell as Pi, but with a ~55k-line Rust native core and 32 built-in tools. MIT license. GitHub: https://github.com/can1357/oh-my-pi Docs: https://omp.sh Install: curl -fsSL https://omp.sh/install | sh · Homebrew · Bun · Windows PowerShell · mise

​
Why It Diverges From Pi

Pi’s thesis: minimal harness beats feature-heavy harness (validated by Terminal Bench). omp accepts the data but identifies a different failure class: tool quality failures — string-match edits that fail on whitespace drift, missing debugger forcing print-statement debugging, no LSP integration for cross-file renames, fork/exec overhead on every search. omp’s answer: replace unreliable primitives with native ones, then add the tools that eliminate entire categories of retry loops.

​
Tool Surface (32 tools)

​
Runtime

​
Code Intelligence

​
Coordination

​
Outside the Box

​
Memory & State

​
Misc

​
Differentiating Primitives

​
Hashline Edit Format

Content-hash anchors replace string matching. Stale anchors (file changed since read) are rejected before corruption. 61% fewer output tokens on Grok 4 Fast. 2.1× pass rate lift on MiniMax. +5pp over str_replace on Gemini 3 Flash.

​
LSP Wired Into Writes

Every write goes through workspace/willRenameFiles. Re-exports, barrel files, aliased imports update before the file moves. No post-rename breakage without IDE.

​
Time-Traveling Stream Rules (TTSR)

Regex match aborts mid-token stream → injects rule as system reminder → retries from same point. No full-context tax. Injections survive compaction. Qualitatively different from PreToolUse hooks: fires mid-stream, not between turns.

​
DAP Debugger

Drives lldb (C/C++/Rust), dlv (Go), debugpy (Python) via DAP protocol. Attach, step, inspect frame. Not a REPL workaround.

​
Task (First-Class Subagents)

Worktree-isolated workers with typed schema-validated return objects. irc enables in-process inter-agent prose. No prose parsing of subagent output.

​
Eval Kernels

Persistent Python + Bun cells that share a prelude and can call back into agent tools. Multi-kernel, multi-cell within one session.

​
Conflict Resolution Scheme

conflict://N URL → write @theirs/@ours/@base → file resolves atomically. conflict://* for bulk.

​
Preview/Accept Workflow

ast_edit returns a staged “(proposed)” card. Agent calls resolve to commit atomically. Structural rewrites have a gate between proposal and application.

​
Hindsight Memory

SQLite memory engine, project-scoped. retain (write) / recall (search) / reflect (synthesize). Survives sessions. Qualitatively different from manually maintained MEMORY.md.

​
Snapcompact

Bitmap-frame context compression — distinct from summarization-based compaction. Renders discarded conversation history into dense PNG frames of pixel-font glyphs that vision models read back directly. Local and deterministic — no LLM call, no API key. Key detail for tool outputs: snapcompact already truncates tool results during archival compaction:
  • Default cap: 2000 chars per tool result (toolResultMaxChars)
  • Head/tail truncation with 60/40 split (truncateHeadRatio)
  • “Useless” tool results skipped entirely
  • Tool arguments capped at 500 chars per value, 2000 chars per call
This is conversation-level archival compression (old turns → PNG frames). It complements but does not replace per-turn tool output truncation. See Omp Snapcompact Rpc for full research.

​
Native Rust Core (~55k LoC)

In-process on libuv thread pool. No fork/exec on hot path.

​
Provider Surface (40+)

Roles: default, smol (cheap fan-out), slow (deep reasoning), plan, commit. Cycle with Ctrl+P, swap mid-session with /model. Features: fallback chains per role, path-scoped model overrides (different model set per repo prefix), round-robin credential rotation with per-key backoff. Notable: Cursor (oauth), GitHub Copilot (oauth), Kimi Code, MiniMax Coding Plan, Alibaba Coding Plan, Qwen Portal, Xiaomi MiMo — plus all OpenAI-compatible local servers (Ollama, LM Studio, llama.cpp, vLLM, LiteLLM).

​
Config Inheritance

Inherits from .claude, .cursor, .windsurf, .gemini, .codex, .cline, .github/copilot, .vscode on first run. Existing rules, skills, MCP servers carry over without migration.

​
Entry Points

​
RPC Mode (Deep Dive)

omp --mode rpc is a JSON-lines protocol: commands on stdin, responses/events on stdout. Key capabilities for bus integration:
  • set_host_tools — register host-side tools that the agent can call (language-agnostic: Python, Rust, Go, bash)
  • set_host_uri_schemes — custom URI schemes (e.g., commandr://events.jsonl)
  • host_tool_call / host_tool_result — bidirectional structured RPC
  • extension_ui_request — agent requests UI: selector, confirm, input, OAuth
  • subagent_lifecycle / subagent_progress / subagent_event — native subagent monitoring
  • Real-time event streaming: message_update, tool_execution_*, agent_start/agent_end
Why it matters for Commandr: RPC mode enables Level 2+ integration without writing TypeScript plugins. The host (commandr-omp-runner) can be any language — it just speaks JSON lines. See Omp Snapcompact Rpc for full protocol research. RPC vs MCP:

​
Plugin Architecture

omp bundles skills, commands, hooks, custom tools, MCP servers, and themes in a single installable package. The surface is Claude-Code-compatible — existing .claude-plugin/ catalogs work as-is. Install: omp install ./path (symlinks + watches), omp install github:user/repo, omp install @scope/plugin. Project-scoped with -l.

​
Headroom Compression Plugin (Local)

A working example lives at pi-headroom/ in this repo:
  • hooks/pre/headroom-compress.ts — tool_result hook that compresses or truncates large tool outputs before the model sees them using headroom-ai
  • tools/headroom-retrieve/index.ts — custom tool that lets the model fetch original uncompressed content on demand
Install:
Tested on omp v16.0.9: The plugin uses tool_result instead of context because omp 16.0.9 does not support message-array replacement via context hook returns. This still covers the bulk of token bloat (tool outputs). See pi-headroom/README.md for full test results and configuration.

​
Role in Commandr Stack

omp is the strongest practical L2 runner candidate: a high-quality worker, not the bus. Its value is execution quality inside a task: hashline edits, LSP/DAP, typed subagent outputs, internal schemes, persistent eval kernels, and provider routing. Commandr should still own claim/progress/approval/complete lifecycle. Integration ladder: Level 1 is complete with commandr-omp-runner (see commandr-omp-runner/ in this repo). Level 2 becomes useful once task/event/approval schemas stabilize. Level 3 (plugin-based bidirectional sync) is where the headroom plugin pattern also lives — a plugin can both intercept context and register tools that talk back to Commandr’s bus.

​
Internal Schemes (://)

pr://, issue://, agent://, skill://, rule:// and others resolve through the same FS-shaped tools. read pr://owner/repo/1428 returns the same shape as read src/foo.ts. agent://<id>/findings.0.path extracts a field from subagent output.

​
Monorepo Packages (omp-specific additions)