oh-no-harness

Oh No Harness

English | 한국어

Your coding agent does not need another runtime. It needs a workflow it can actually read and follow.

Oh No Harness is that workflow for Claude Code and Codex: 11 stage skills plus 9 role agents that move vague work from interview to ralplan to verified ralph execution, with fusion-rescue for hard stalled problems, without npm, tmux, MCP, or a terminal-only control plane.

It sits between two extremes:

• Not an oh-my-* runtime asking for a corner of your terminal.
• Not a bare skill drawer where every phase feels like one more thing to pick.

It is a text-native workflow harness: stage skills coordinate the handoffs; role agents handle focused passes for exploration, planning, execution, review, security, QA, and verification.

• No npm install -g
• No npx dance
• No tmux window to keep alive
• No custom CLI muscle memory
• No MCP server to wire up before the work can start
• No terminal-only workflow that falls apart in app or plugin UIs

The runtime is deliberately boring: text files your agent can read — platform skill wrappers in skills/ and skills-claude/, shared workflow core in docs/skill-core/, maintenance-only company prompt references in docs/providers/, agents/, thin commands/, plugin manifests, and one optional Claude Code SessionStart hook.

[!NOTE] If you can read Markdown, you can audit the harness. If you can follow a handoff, you can understand the workflow.

Eleven focused workflows help clarify vague work, plan, execute with verification, rescue hard stalled problems, debug, and clean up — without a daemon, background service, or hidden state.

Oh No Harness follows semantic versioning from 1.0.0.

Highlights

🛠 Architecture

• Plain text, not a sidecar. No npm package, no project CLI, no tmux session manager, no MCP server. The behavior is Markdown you can read, diff, fork, and edit.
• Native host install. Claude Code and Codex load the plugin through their own plugin/skill systems; Oh No Harness does not add another thing to supervise.
• Terminal optional. Shell users can install from the terminal, but the daily workflow is not terminal-bound. The same Markdown skills fit Claude Code sessions and Codex App-style plugin UIs.
• Workflow spine. Public skills own the software-development stages; internal agents supply the specialist judgment without becoming extra commands to memorize.
• Skills + agents. 11 workflow skills backed by 9 role agents (explore, analyst, planner, plan-reviewer, executor, debugger, verifier, code-reviewer, fusion-rescue-analyst).
• Slash ↔ skill parity. commands/*.md mirrors all 11 skill names with argument hints, then delegates to the Claude Code wrapper in skills-claude/<name>/SKILL.md; Codex reads the matching wrapper in skills/<name>/SKILL.md.

Too much	Too little	Oh No Harness
Start a sidecar runtime	Keep picking from a loose skill shelf	Install a native Claude Code / Codex plugin
Learn a project CLI	Remember every phase by hand	Use a small stage surface: interview, ralplan, ralph
Debug hooks, HUDs, MCP, tmux	Hope one skill has enough context	Let skills hand off to role agents with explicit evidence gates
Stay in terminal-land	Lose structure in GUI hosts	Use the same text skills through native plugin discovery
Operate a platform	Collect prompts	Review the Markdown that drives the workflow

🔁 Workflow

• Socratic interview. /oh-no-harness:interview routes code facts, research facts, and judgment calls separately — capturing decisions, constraints, and non-goals before any spec.
• Mode-gated execution. Specs and plans size work as LIGHT / STANDARD / THOROUGH; Ralph follows the recorded mode instead of always running the heaviest loop.
• Fusion rescue. /oh-no-harness:fusion-rescue runs exactly three panel lenses, including a Codex-preferred adversarial lens when available, then lets the current host synthesize the next action.
• Auto-routing. /oh-no-harness:auto-routing on nudges Claude to consult the right skill before clarifying or editing — no hidden state, no skipped approval gates.