breezewish/CodexPotter

What it solves

CodexPotter is designed to replace the standard /goal command in AI coding assistants. It solves the problem of "context rotting"—where an LLM's performance degrades as a conversation grows longer—by continuously reconciling a codebase toward a specific goal using fresh contexts for each iteration.

How it works

The tool implements a reconciliation loop (the "Ralph Wiggum pattern"). Instead of one long session, it uses a main agent and subagents to review and polish code over multiple rounds. It treats the file system as memory, storing instructions and plans (e.g., in a HANDOFF.md file) to ensure details are preserved and not lost to context window limits. It also maintains a local knowledge base to help the AI quickly understand the project in these clean-room environments.

Who it’s for

Developers using Codex CLI or Codex Desktop who want to automate complex, multi-step coding tasks without having to manually clean up unfinished work or manage long, degrading chat histories.

Highlights

• Clean-room iterations: Uses fresh contexts for every round to maximize AI "IQ" and avoid context poisoning.
• File-system memory: Stores goals and instructions in files to resist compaction and preserve detail.
• Tiny footprint: Uses fewer than 1,000 tokens for its internal logic, leaving more room for business logic.
• Unattended execution: Allows users to queue multiple $loop tasks that execute sequentially while the user is away.
• Local Knowledge Base: Includes a built-in index to accelerate project learning across different sessions.