AutoCode for Codex

AutoCode turns Codex from a session-based coding agent into a persistent, modular project system.

Coding agents are excellent at completing individual tasks, but longer software projects introduce a different problem: goals drift, context gets lost between sessions, verification becomes inconsistent, and interrupted work is difficult to resume. I built AutoCode to manage that project-level layer.

Inspiration

The idea came from repeatedly switching between Codex, project files, terminal output, testing tools, and separate planning conversations.

A normal workflow was too linear:

1. Explain the goal.
2. Ask an AI to create a plan.
3. Copy the plan into another session.
4. Implement one part.
5. inspect the result manually.
6. Explain the current state again.
7. Decide what should happen next.

I wanted to reduce that coordination overhead without simply giving one enormous prompt to an autonomous coding agent.

AutoCode therefore treats Codex as a powerful worker inside a larger deterministic project process.

What AutoCode does

A user enters a high-level project goal. AutoCode then:

• converts the goal into a bounded project specification;
• identifies modules, constraints, requirements, and success criteria;
• creates a dependency-aware task graph;
• selects an appropriate Codex execution profile for each task;
• compiles only the relevant project, module, and task context;
• starts fresh Codex worker sessions;
• verifies completed work proportionately;
• preserves recoverable checkpoints;
• updates a curated Obsidian project knowledge base;
• evaluates milestones against the original goal;
• converts milestone findings directly into subsequent tasks.

The normal entry point is intentionally simple:

cd /path/to/project
autocode

Built With

• agents
• ai
• automation
• cli
• developer
• git
• gpt-5.6
• json
• local-first
• markdown
• node.js
• obsidian
• openai
• playwright
• testing
• tools
• typescript