oh-my-hermes

Overview

Multi-agent orchestration skill suite for Hermes, inspired by oh-my-claudecode and rebuilt on Hermes primitives. Provides composable driver skills covering the full agentic workflow lifecycle: research → interview → consensus planning → verified execution.

Composes end-to-end: deep-research → deep-interview → ralplan (consensus plan) → ralph (verified execution) — with triage and autopilot for routing and hands-off operation.

When to Use

• You want to run deep research on a topic with structured output
• You need consensus planning before executing a complex task (Planner → Architect → Critic)
• You want verified, iterative execution with automatic retry on failure
• You want to triage an incoming task to the right sub-skill automatically
• You want autopilot mode: Hermes routes, plans, and executes with minimal hand-holding
• Don't use for: simple single-turn tasks, or when delegate_task alone is sufficient

Installation

git clone https://github.com/witt3rd/oh-my-hermes
cd oh-my-hermes
bash install.sh

Or install individual skills:

hermes skills install https://raw.githubusercontent.com/witt3rd/oh-my-hermes/main/skills/ralplan/SKILL.md
hermes skills install https://raw.githubusercontent.com/witt3rd/oh-my-hermes/main/skills/ralph/SKILL.md

Skill Suite

deep-research

Structured multi-source research with synthesis:

/skill deep-research
"Research the current state of MoE (Mixture of Experts) architectures in LLMs"

Produces: sources list, key findings, synthesis, open questions, confidence ratings.

deep-interview

Structured elicitation of requirements or knowledge via targeted questions:

/skill deep-interview
"Help me clarify requirements for a multi-tenant SaaS billing system"

Produces: structured requirements doc, edge cases, ambiguities surfaced.

ralplan — Planner → Architect → Critic Consensus

Three-phase planning with internal debate before committing:

1. Planner generates a candidate plan
2. Architect reviews for technical soundness
3. Critic stress-tests assumptions and identifies risks
4. Consensus plan is produced only when all three agree

/skill ralplan
"Plan a migration from a monolith to microservices for a mid-size e-commerce platform"

Best for: high-stakes tasks where a bad plan is expensive to undo.

ralph — Verified Execute → Verify → Iterate

Execution loop with built-in verification:

1. Execute the plan step
2. Verify the output against the expected result
3. Iterate if verification fails (up to N retries)

/skill ralph
"Set up a production-ready Postgres database with connection pooling and monitoring"

Best for: infrastructure tasks, code generation, or any task where partial completion is worse than no completion.

triage

Routes an incoming task description to the most appropriate driver skill:

/skill triage
"I need to build a CI/CD pipeline for a Python monorepo"

Returns: recommended skill(s), rationale, suggested starting prompt.

autopilot

Hands-off mode — Hermes triages, plans, and executes autonomously with checkpoint confirmations at configurable risk levels:

/skill autopilot
"Refactor the authentication module to use JWT refresh tokens"

Autopilot will pause and request confirmation before any destructive or irreversible actions.

Composition Pattern

Full end-to-end pipeline:

1. deep-research   → gather context and background
2. deep-interview  → clarify requirements and constraints
3. ralplan         → build a consensus plan
4. ralph           → execute with verification

Run inline:

"Use the oh-my-hermes pipeline to redesign our API rate limiting system.
Start with deep-research on rate limiting strategies, then interview me
for constraints, then ralplan, then ralph the implementation."

Common Pitfalls

1. Skipping triage on complex tasks. Always run triage first if you're unsure which driver to use — it prevents using ralph on a task that needs ralplan first.
2. ralph retries exhausted. If verification keeps failing, surface the error to the user rather than retrying indefinitely. ralph defaults to 3 retries — increase via config for flaky environments.
3. ralplan critic loop. In rare cases the Planner/Architect/Critic loop doesn't converge. Set a max round limit (default: 5) to prevent runaway token spend.
4. autopilot on destructive tasks. Autopilot will confirm before destructive actions, but make sure the confirmation threshold matches your risk tolerance — configure it before using on production systems.
5. Skill not found after install. Run hermes skills reload after install.sh to force the skill loader to pick up new entries.

Verification Checklist

• Skills installed: hermes skills list | grep -E "ralplan|ralph|triage|autopilot"
• hermes skills reload run after install
• Test triage with a simple task description — returns a recommended skill
• Test ralplan on a small task — produces a 3-phase consensus plan
• Test ralph on a reversible task — executes and verifies output