Skip to content

correctionguy/correctionguy

v5.0.1MIT

Agents often drift and waste time on the wrong work. Correction Guy watches from the sidelines and steers them back on track.

GitHubOpen Website

A second pair of eyes for your coding agent. Correction Guy watches from the sidelines and steers the work back on track. It never touches the work itself. Packaged as an Agent Plugins 1.0.0 plugin (portable skills in skills/ behind root plugin.json) with host adapters for Claude Code and Pi.

An agent grading its own work is the weakest check there is. It is biased toward declaring victory: it papers over stubs, asserts things it never verified, asks permission instead of delivering, and never checks what it changed. A sharper self-prompt doesn't fix this. The agent is still inside its own story.

So Correction Guy hands the review to someone else. A different model (Codex, through the Codex SDK, no other plugin required) reads the session as an outsider, with no stake in the agent's narrative, and calls out what the agent talked itself past. It is deliberately narrow: it flags six failures, and it raises naming, formatting, and structure only when your instruction files or your recorded preferences name them. It steps in at the moments that matter: as work begins, while work is underway, and the instant the agent tries to call it done.

  • Unverified assumption: builds on a guess about the system instead of checking it in the workspace.
  • Missed requirement: leaves out behavior the instruction requires: the user's ask to its end gate (a fix that is fixed, a release that is released) and whatever the instruction files the host loads require.
  • Integration error: right idea, wired into the surrounding system incorrectly.
  • Regression: breaks existing behavior while making the change.
  • Wrong file: delivers the change somewhere the running application never calls, such as a one-off script.
  • No reviews: never reviewed nor verified what it changed, programmatically (tests, a smoke run) or empirically (a reviewer agent, when the project defines one).

What it does

  • Session prelude: sets expectations up front: recall memory and preferences, restate the task, then the six failures to avoid.
  • Live monitor: every few tool batches, an outside pass catches the six failures before they compound.
  • Stop check: when the agent tries to stop, the reviewer can block a premature "done" (a missed requirement, a stub, an unrun test, a fix that is still a proposal) and feed the correction back. A non-blocking nudge repeats at most once per 30 minutes per session; a block is never held back. Once a review fails (a usage limit, a capacity error, an oversize context, or malformed JSON), the live monitor and the stop check stay quiet for the rest of the session instead of paying that failure on every stop; the next session start (a new session, a resume, or a compaction) turns them back on.
  • On-demand: the correctionguy skill (/correctionguy) restates the discipline whenever you want it.
  • Actually: /correctionguy:actually is user-triggered when Correction Guy rooted on an old, outdated, or wrong convention: the agent records the override in the memory store your instruction files define and applies it from then on.

The review prompts, and the corrections that come back, are written in compressed "caveman" style to save tokens.

Memory

Correction Guy prescribes no memory location or format. The agent recalls and follows whatever memory and preference store the instruction files the host loads define, and records facts you state there. The reviewer treats a recorded owner-stated fact as checked, never as unsourced. Rules about what to record and where belong in your instruction files, where a missed one is a missed requirement.

Requirements

  • An Agent Plugins client for the portable skills path (ChatGPT/Codex, GitHub Copilot, VS Code, Kiro, and others that load the open format), or Claude Code / Pi for the host adapters below
  • Codex authenticated: run codex login once, or set an API key (needed for the live/stop reviewer hooks on Claude Code and Pi)
  • Bun, only when installing the Claude Code hooks

Install

Agent Plugins (recommended)

Correction Guy ships as an Agent Plugins 1.0.0 package: root plugin.json plus portable skills under skills/. Agent Plugins leaves installation to each client. Install correctionguy/correctionguy through your client's plugin install flow; the compatible clients page links each client's setup instructions.

Use /correctionguy on demand whenever you want the discipline restated; use /correctionguy:actually when Correction Guy rooted on a wrong convention.

What you get from the portable core: the on-demand discipline skills. Session prelude, live monitor, and stop check are host adapters (below). Clients that only load Agent Plugins ignore those layers.

Host adapters (hooks)

Use these when you want the sidecar reviews (session prelude, live monitor, stop check), not only the portable skills.

Claude Code

/plugin marketplace add correctionguy/correctionguy

/plugin install correctionguy@correctionguy

Pi

pi install git:github.com/correctionguy/correctionguy

Or try it for a single run with pi -e git:github.com/correctionguy/correctionguy.

Pi loads the skills from the package manifest and exposes them as /skill:correctionguy and /skill:actually.

Configuration

Everything is tuned through CORRECTIONGUY_* environment variables: model, reasoning effort, review cadence, and a CORRECTIONGUY_YOLO switch that drops the read-only sandbox for the reviews. Defaults live in scripts/codex.ts. You're going to ask an agent to do it, so I'm not going to write it out here 😏

On a host with tokenmaxxing pooling your Codex logins, the Claude Code reviews draw from the pool instead of the one ambient login: each review borrows a usable pooled account (tokenmaxxing seat --codex) for its own run and passes its store as CODEX_HOME, so a maxed-out ambient account no longer silences Correction Guy. The borrow needs tokenmaxxing on the hook process's PATH; when it is not there, or no pooled account is usable, the review runs on the ambient login as before. CORRECTIONGUY_TOKENMAXXING=0 keeps the ambient login. Pi keeps the ambient login by default, because a borrow keyed to its long-lived process would pin one pooled account for the whole chat; CORRECTIONGUY_TOKENMAXXING=1 opts in.

Licensed under MIT.