oh-my-agent

oh-my-agent is a multi-agent harness for AI coding agents whose central premise is that agents narrate success while the harness checks the artifacts. It is distributed as an npm package and can be installed either as a skills pack (via npx skills add or Microsoft's Agent Package Manager) or as a full harness with workflows, rules, configuration, keyword-detection hooks, and a CLI. Core verification mechanisms are described as mechanical rather than LLM-judged. A Stop hook blocks session termination while a persistent workflow is active and runs a configured gate script before allowing a stop; only typecheck, test, and lint are executable, and the gate is capped at five reinforcements. An anti-circumvention gate checks for artifacts a shortcut cannot fake, including phase records, plan JSON, and result files from distinct QA and refactor agents. An independent judge is spawned with fresh context, briefed only on criteria, and re-verifies every criterion each round including prior passes. Every gate pass, failure, and decision is appended as JSON lines to a session event log that is append-only and cross-vendor. Additional tooling includes per-agent verification batteries covering scope violations, charter alignment, hardcoded secrets, TODO scans, and declared outputs, plus type-specific checks. A skill evaluation harness measures utility lift on held-out tasks and keeps only edits that improve measured lift. Budgets cap tokens, spawn counts, and per-vendor spend, with the orchestrator refusing further spawns when limits are exceeded. The project ships role-based agents modeled as an engineering team, including architecture, backend, brainstorm, database, debug, deep security, design, dev workflow, docs, explanation, frontend, mobile, observability, orchestration, project management, QA, refactor, source control, search, and Terraform infrastructure agents. Separate content and research pipelines cover academic writing, HWP and PDF conversion, image generation, market research, conversation recaps, scholarly search, slide generation, translation, video generation, and voice generation. Work is driven by chat or slash commands such as /brainstorm, /architecture, /plan, /work, /orchestrate, /ultrawork, /ralph, /review, /deepsec, /debug, /docs, and /scm. Keyword auto-detection activates workflows across multiple languages, and detection accuracy is measured against a labeled prompt corpus with CI gating. The harness supports many agent runtimes from a single .agents/ directory, including Claude Code, Codex CLI, Antigravity, Cursor, Qwen Code, OpenCode, GitHub Copilot, and others, with per-agent model selection via configuration. It is MIT licensed.