auto-codex

The problem

Long autonomous Claude sessions burn tokens fast.

Letting one Claude session research, plan, review, and implement a task means every token of every phase comes out of your Claude budget. For a non-trivial task, that adds up quickly — and the same model doing both the work AND reviewing it doesn't catch much.

Without auto-codex

One model, one budget

Claude plans, Claude implements, Claude reviews itself. No cross-check. Tokens stack up.

Claude tokens used ≈ 100%

With auto-codex

Right model per stage

Codex plans (your ChatGPT account). Opus reviews adversarially. Sonnet swarms implementation. Claude only orchestrates.

Claude tokens used ≈ 30%

Multi-model orchestration

Each model does what it's best at.

Codex is fast, structured, and great at writing plans. Opus is the strongest reviewer — adversarial, finds edge cases. Sonnet is fast and parallelizable for implementation. Putting them together — with cross-checks at every handoff — beats any single model doing all three jobs.

CODEX

latest (gpt-5.5) · planner · gate

Drafts the implementation plan
Final integrity gate (verifies Opus findings resolved)
Iterates the optimize loop

OPUS

adversarial reviewer

Reviews Codex's plan — finds what's missing
Reviews the swarm's implementation
Mandatory gate before code lands

SONNET

parallel implementer

Up to 6 subagents in parallel
One atomic task per agent
Writes code + tests per the plan

How it feels

One command. The pipeline runs itself.

Type /auto-codex and it does the right thing — drafts a scope if there's no session, asks about an existing draft, or resumes a paused session. Stop and resume whenever.

claude code · /tmp/my-project

Pipeline breakdown

Seven phases. Each has a gate.

Hard gates between phases block sloppy work from reaching implementation. No vibes. Each phase produces a verifiable artifact before the next phase starts.

Research

Analyze the problem space, read code, identify constraints. Writes findings to tmp/research.md.

claude

min 50 chars to advance

Plan-Codex

Codex drafts an atomic, gateable plan. Opus reviews it adversarially. Findings logged.

codex + opus

tmp/plan.md

Review (hard gate)

Codex verifies every Opus FAIL item was resolved in the final plan. Approves or rejects.

codex

review_passed flag

Swarm

Up to 6 Sonnet subagents implement in parallel — one atomic unit each. Opus reviews after.

sonnet × N

+ opus review

Verify

Runs the test suite + verification checks from the plan. Failing tests bounce back to swarm.

claude

tests must pass

Optimize (conditional)

If the plan declared a metric, Codex iterates — propose change → eval → keep if improved, revert if not.

codex

auto-abort on 3 fails

Complete

Marks the task done, clears tmp/, picks the next task — or deactivates when all done.

pipeline

→ next task

The killer feature

The optimize loop.

After your code lands, an optional autonomous loop iterates on a single file to improve a metric. Codex proposes a change, your eval script runs, the change is kept if the metric improved or reverted if it didn't. Then it does it again. And again. For 30, 50, 100 cycles — overnight if you want.

Inspired by

karpathy/autoresearch

Andrej Karpathy's 630-line Python script that ran 50 LLM-training experiments overnight, autonomously. We took the keep/revert pattern and wired it as the final phase of the auto-codex pipeline.

Live metric trace

Sharpe ratio · higher is better

cycles: 28
improvements: 9
best: 2.41

baseline

improvement zone

kept (metric improved)

Single-file constraint

Codex can only edit ONE file per cycle. Forces focus, blocks scope creep.

Auto keep/revert

Metric improved? Git-commit the change. Didn't? Revert it. No human in the loop.

History-aware prompts

Codex sees the last 15 experiments — what worked, what didn't. Stops repeating losing ideas.

Safety stops

3 consecutive failures → auto-abort. Hostile eval or broken Codex CLI can't burn your night.

What people optimize

→ Trading strategies

Sharpe ratio, drawdown, hit rate

→ Model training

Val loss, accuracy, perplexity

→ GPU kernels

Throughput, memory bandwidth

→ Template engines

Render time, bundle size

→ Tabular ML

Cross-val score, AUC

→ Anything measurable

If you can print metric=value, Codex can optimize it.

Other lanes

This plugin ships four pipelines, not one.

Everything above describes /auto-codex, the flagship lane. Three more ship in the same plugin — including two that need no Codex at all.

🤖

/makeitso

Claude-only. No Codex, no external CLI. Same phase shape as /auto-codex, self-planned and self-reviewed by fresh Claude subagents.

claude only

no Codex needed

🤖

/makeitso-codex

Codex plans and builds; Claude reads the plan once and the diff once. Minimizes Claude token spend. Offers /makeitso by name if Codex isn't available — never switches silently.

codex

needs Codex

🤖

/makeitso-project

19 gated phases for building whole projects from a spec: requirements, traceability, gap analysis, red-team, cleanup, deploy.

claude only

no Codex needed

Full lane comparison, subcommands, and setup details live in the README.

Install

Three steps. Mostly one-time.

You need Claude Code, Python 3.8+, and git. A Codex / ChatGPT account is only needed for the two Codex-driven lanes (/auto-codex, /makeitso-codex) — skip step 1 if you're starting with /makeitso or /makeitso-project.

Install the Codex CLI (one-time, only for the Codex lanes)

This is the official OpenAI Codex CLI. The plugin uses it under the hood as the MCP server.

npm install -g @openai/codex && codex login

Add the marketplace + install the plugin

Run these inside any Claude Code session. The plugin auto-registers Codex as an MCP server.

/plugin marketplace add neonplants/auto-codex

/plugin install auto-codex@auto-codex-marketplace

Use it

cd to any project dir, then in Claude Code, pick a lane:

/auto-codex

The first run drafts a scope of work, asks you to fill in tasks, waits for your approval, then runs autonomously. Stop anytime with /auto-codex stop. No Codex installed? Run /makeitso instead — same idea, Claude-only.