dzshzx/codexcomp

codexcomp is a small local proxy that fixes a specific bug in OpenAI's Codex command line tool. When Codex uses the gpt-5.5 model, its reasoning sometimes gets cut off partway through at a predictable token count, and the tool then answers based on that incomplete thought, which lowers the quality of the response. The project's own data suggests this happens in a large share of responses that reach the affected token count. The tool runs on your own computer and sits between Codex and OpenAI's servers. When it detects a truncated response, instead of passing the broken answer along, it asks the model to keep reasoning, then combines the extra rounds into one complete answer before sending it back to Codex. From the user's point of view, Codex simply receives a full, untruncated response instead of a broken one. Setup is a single command to install the tool, followed by adding one line to a Codex configuration file that points Codex at the local proxy instead of directly at OpenAI. No changes to how Codex identifies itself or organizes its sessions are needed, so features like remote conversation compaction keep working normally. The proxy uses a fast connection method matching what Codex expects, and falls back to a more compatible connection method if needed. The project also includes some resilience features: if the connection to OpenAI's servers gets stuck or times out, the proxy automatically resets the connection instead of failing every request until restarted. It never stores or logs your login credentials, only forwards them, and only listens on your own machine rather than the network. There is also an optional command to make the proxy start automatically when your computer boots.