omc

name	omc
description	oh-my-claudecode (OMC) harness — Claude Code orchestration. Use when user says ralph, team, autopilot, ultrawork, ultrapilot, or wants Claude Code session with hooks/skills/agent catalog. Routes to local `omc` CLI binary.

OMC harness skill

When to use

User says or implies:

• "ralph" → sequential persistence loop until task done with verification
• "team" → spawn N coordinated agents on shared task list
• "autopilot" → full autonomous execution from idea to working code
• "ultrawork" → high-throughput parallel task completion
• "ultrapilot" → autopilot + ultrawork combination
• "claude code session", "use claude code", "interactive claude" → omc launch
• "ask claude about X" with code context → omc ask claude -p "X"
• "deepinit", "deepsearch", "deep interview" → OMC skill triggers

Commands

Intent	Command	Cost class
Sequential persistence loop	omc ralph "<task>"	HIGH
Parallel coordinated agents	omc team N:executor "<task>"	HIGH
Autonomous build	omc autopilot "<task>"	HIGH
High-throughput parallel	omc ultrawork "<task>"	HIGH
Provider advisor prompt	omc ask claude -p "<prompt>"	LOW
Interactive TUI	omc launch	INTERACTIVE
Diagnose OMC install	omc doctor	FREE
Show version	omc --version	FREE
Configure stop callbacks	omc config-stop-callback <type>	FREE
Session inspection	omc session list	FREE

Execution

Run via Hermes terminal / code_execution toolset. Working directory defaults to user's project root (allow override via cwd: arg in skill invocation).

Pre-flight: ensure omc --version returns ≥4.13.7. If absent:

npm install -g oh-my-claudecode
# or
bun install -g oh-my-claudecode

Environment

OMC subprocess MUST run with:

• HOME=$HOME (the PAI user's home directory, e.g. /home/pai on a dedicated VPS user)
• CLAUDE_CONFIG_DIR=$HOME/.claude

This ensures PAI statusline + UpdateCounts.hook + Spinner assets resolve. Without these, OMC launches a vanilla Claude Code session without PAI integration.

Output handling

• version, doctor, ask (non-streaming): capture full stdout, return to Hermes context.
• ralph, team, autopilot, ultrawork: streaming. Use code_execution_tool streaming mode if Hermes config allows. Otherwise tail output file.
• launch: interactive — spawn in tmux pane or sandbox terminal. Do not block Hermes main loop.

Cost discipline

ralph, autopilot, team, ultrawork burn Claude API tokens at scale (multi-agent, multi-turn). Confirm with user before launching when task scope >5 files OR estimated duration >30 min.

If pai-cost-tracker reports 5h window ≥80% (ALERT) or ≥95% (BLOCK per tools/cost_check.py DEFAULT_THRESHOLDS), refuse to launch high-cost OMC commands until window resets or user explicitly overrides.

Caveats

• omc ralph may run for hours. Show progress via tail or skill chain to pai-pulse for completion notification.
• omc team spawns subagents that consume tokens independently of main agent.
• OMC native rate_limits read from Claude OAuth — same usage cache pai-cost-tracker reads.
• Mutations made by OMC are NOT auto-recorded in pai-anywhere install manifest.

Triggers in Hermes natural language

• "ralph the test failures" → omc ralph "fix failing tests"
• "spawn a team of 3 to refactor auth" → omc team 3:executor "refactor auth module"
• "build me X with autopilot" → omc autopilot "<X>"
• "claude opinion on this commit" → omc ask claude -p "review HEAD commit"