oh-my-god

oh-my-god

The sidecar harness that catches Claude when it lies.

Don't ask Claude if it's done. Make it prove it.

Quick Start • Commands • How it works • Configuration • Uninstall • Architecture

---

Why oh-my-god?

Every Claude Code plugin observes AI behavior and hopes the model does better next time. oh-my-god is different — it enforces.

Three failure modes plague every AI coding workflow:

#	What breaks	How oh-my-god fixes it
1	Claude says "done" when tests are still red	Stop hook runs your tests independently and sends Claude back to work if any fail
2	Claude picks a plausible bug instead of the one you meant	UserPromptSubmit hook scores prompt clarity and forces Claude to ask clarifying questions on vague input
3	Claude reaches for rm -rf when you asked it not to	PreToolUse hook enforces per-role tool allowlists at the Claude Code runtime level

All three guarantees run at the Claude Code hook layer, using the same mechanism Claude Code uses internally for its built-in Stop hook enforcement. The AI cannot accidentally bypass them — the runtime fires the hooks regardless of what the model's prompt says.

Tagline: Your cynical PR reviewer, in plugin form.

---

Quick Start

Prerequisites: Node.js ≥ 20, Claude Code CLI, git.

Step 1 — Install as a Claude Code plugin

From inside a running Claude Code session:

/plugin marketplace add https://github.com/ckdwns9121/oh-my-god
/plugin install oh-my-god

Or manually (dev mode):

git clone https://github.com/ckdwns9121/oh-my-god
cd oh-my-god
npm install
npm run build
# then point Claude Code at this directory as a local plugin install

Step 2 — Initialize the state DB

/omg-setup

This creates ~/.omg/state.db and verifies the plugin is wired correctly. Run this once per machine.

Step 3 — Start using Claude Code normally

$ claude
> Read and explain src/core/types.ts

That's it. oh-my-god runs in passive mode by default. You won't see it unless something triggers an intervention.

To enter strict mode for a specific session, invoke the lock skill:

> /omg-lock executor

Now Claude's tool use is restricted to the executor role's allowlist, and Stop-hook gate judgment will block false completions. When you're done, release with:

> /omg-release

---

Commands

Slash commands (invoked inside a Claude Code session)

Command	What it does	Example
/omg-setup	Initialize ~/.omg/state.db and verify plugin install	/omg-setup
/omg-lock <role>	Enter strict mode — enforce tool allowlist for the current session	/omg-lock executor
/omg-release	Exit strict mode — session returns to passive	/omg-release

Roles (used with /omg-lock)

Role	Allowed tools	Typical use
default	all	Passive mode, plugin is invisible
executor	Read, Grep, Glob, Edit, Write, Bash, TodoWrite, NotebookEdit	Code implementation, refactoring
explorer	Read, Grep, Glob	Codebase tours, "explain this" sessions
architect	Read, Grep, Glob	Analysis + design thinking (same as explorer for now)
reviewer	Read, Grep, Glob, Bash	Code review with git log, git diff, git show

Edit roles.json to customize the allowlist per role or add new roles.

Shell scripts (rarely needed directly)

If you want to lock/release from the terminal instead of inside Claude:

# lock the most recent session to a role
node "$CLAUDE_PLUGIN_ROOT"/scripts/omg-lock.mjs executor

# release
node "$CLAUDE_PLUGIN_ROOT"/scripts/omg-release.mjs

# target a specific session id
node "$CLAUDE_PLUGIN_ROOT"/scripts/omg-lock.mjs explorer --session <session-id>

Both scripts exit 0 on success, 1 on error, 2 on missing arguments. Stdout carries the status line and stderr carries warnings.

---

How it works

oh-my-god wires three hooks into your Claude Code session. All three are registered via hooks/hooks.json and fire automatically on the matching Claude Code lifecycle event.