PatchWarden

01 / Context

Why It Exists

In the collaboration between ChatGPT and Codex/Claude Code, the discussion phase and execution phase are disconnected. ChatGPT provides solutions but cannot directly operate on local codebases; local agents have strong execution capabilities but lack discussion context. PatchWarden bridges this gap.

Problem

When AI agents operate directly on codebases, risks exist: unbounded workspace scope, opaque command execution, and non-traceable output. There is a missing bridge between ChatGPT-discussed solutions and actual execution, making results difficult to audit.

Solution

PatchWarden establishes a controlled channel between discussion and execution: extracts tasks from ChatGPT conversations, limits the agent's workspace and command scope, and returns unified diffs and verification reports after execution, forming complete task lineage evidence. This is a controllable task channel, not a general-purpose remote shell.

02 / System

Core Capabilities

• Extract tasks from ChatGPT conversations, bridge to local coding agents (Codex CLI, Claude Code, OpenCode)
• Limit workspace and command scope to control risk boundaries
• Return structured diffs, verification reports, and audit records
• Support task lineage tracing — fully traceable from discussion to output
• Support Core mode and optional Direct mode; Direct provides independent, bounded direct editing and verification

03 / Workflow

Workflow

1. Discuss approach in ChatGPT and generate task description
2. PatchWarden extracts the task, limits workspace scope
3. Local agent (Codex/Claude Code/OpenCode) executes in controlled environment
4. Generate diff, verification results, and audit log
5. Human reviews output, decides whether to merge or revert

04 / Boundaries

Security & Privacy

PatchWarden itself runs locally and does not store conversation content or code snippets. Code changes occur within the local workspace and are not uploaded through PatchWarden. PatchWarden calls LLMs via public APIs for solution understanding; the data handling policies of the connected agents and model providers depend on their respective configurations and policies. High-risk actions and Direct modifications require explicit boundaries and confirmation.

Current Limitations

• Requires local agent to be installed and configured
• ChatGPT Tunnel has additional prerequisites
• Windows installer is currently unsigned
• High-risk actions and Direct modifications require explicit boundaries and confirmation

05 / Roadmap

Roadmap

06 / Participate

How to Participate

Issues and pull requests are welcome on GitHub. The project is under active development — try it out and share feedback. Latest releases are available from GitHub Releases.