rlaope/oh-my-hermes

oh‑my‑hermes (OM H)

What it is – OM H is a plug‑in/operating layer for the open‑source Hermes Agent (an LLM‑driven AI assistant). It keeps Hermes as the natural‑language front‑end but adds a structured workflow engine, explicit evidence‑gating, long‑term memory, and per‑model routing. In short, it turns a plain Hermes request into a reproducible, auditable “capability” that runs a series of specialized skills while tracking cost, model choice, and outcomes.

Key ideas

• Mixture‑of‑Models routing – each request is scored and dispatched to a model‑category (e.g., ultrabrain, quick, writing). The routing table is editable and includes dozens of modern LLM families.
• Parallel work lanes – work is split into independent units that run in separate Git work‑trees, allowing safe concurrent tool calls.
• Evidence gates – every step produces typed results (process exit, schema validation, verification receipt). A gate must approve the result before it is considered “done”.
• Long‑term memory store – OM H records decisions as review cards, timestamps them, and recalls a budgeted set of memories for future sessions. Hermes’ own memory is never altered.
• 108 specialist omh‑* skills – ready‑made tool calls for front‑end, back‑end, Rust, inference serving, security review, performance budgets, refactoring, etc. The router injects the needed skill automatically.
• Unified terminal UI – the Hermes TUI shows a HUD with per‑lane rows (model, effort, tokens, cost, evidence state) and a phase‑structured TODO list above the prompt.

Installation – one‑line installers for macOS/Linux, Windows PowerShell, Homebrew, Bun, npm, or via the Hermes skill tap. After installing, run omh setup to seed the routing config and register the plug‑in with Hermes.

Typical workflow

1. Ask Hermes (or omh) for a high‑level task.
2. OM H routes the request, selects a model chain, and loads any matching omh‑* specialist skills.
3. The request is split into parallel work units (ulw‑work), each executed in its own work‑tree.
4. Results are typed and pass through verification gates; evidence is recorded.
5. A memory card may be created for any decision that should be remembered across sessions.
6. The terminal HUD continuously displays progress, cost, and evidence status.

Why it matters – By adding explicit routing, parallelism, and audit trails, OM H makes LLM‑driven coding and research more reliable, cheaper, and easier to debug. It is especially useful for developers who already use Hermes Agent and want a production‑grade operating layer without rewriting their existing workflows.

Resources – Website, docs, model‑chain editor (omh model), and a Discord community are linked from the README.