Prevent a coding agent from receiving a cut off reasoning response from Codex.
Route requests to a Responses compatible API through one proxy with configurable authentication.
Inspect proxy metadata to see how many hidden continuation rounds a response took.
| neteroster/codexcont | nvidia-nemo/evaluator | snazzybean/roommind | |
|---|---|---|---|
| Stars | 315 | 314 | 316 |
| Language | Python | Python | Python |
| Last pushed | — | 2026-07-17 | — |
| Maintenance | — | Active | — |
| Setup difficulty | moderate | hard | moderate |
| Complexity | 4/5 | 4/5 | 3/5 |
| Audience | developer | developer | general |
Figures from each repo's GitHub metadata at analysis time.
Requires Python 3.12+ and a config.toml pointed at your upstream Responses API and auth mode.
CodexCont is a small proxy server that sits between a coding AI agent and an API that OpenAI's Codex or a similar Responses style service exposes. Its job is to work around a specific pattern the author noticed: sometimes the model's internal reasoning gets cut off partway through, and this tool detects that pattern and quietly asks the model to keep thinking instead of returning an incomplete answer. The project explains that it is intentionally bypassing an observed truncation behavior in the OpenAI Codex service, and it states plainly that anyone using this tool is responsible for any consequences if that counts as abuse of the service's terms or leads to unexpected costs. Behind the scenes, the proxy streams the model's reasoning to the coding agent as it comes in, but holds onto the actual final answer temporarily. When it sees the signal that a response was cut short, a specific pattern in the token count the API reports, it discards that partial answer, sends the reasoning back with an instruction to continue, and starts another round with the upstream service. If the response finishes normally, it just passes the final answer through, rebuilt into a single coherent reply so the coding agent only ever sees one response, even though several requests may have happened behind the scenes. Any traffic that does not match this specific pattern is passed straight through unchanged. The tool is written in Python, needs version 3.12 or newer, and uses the small web framework Starlette along with the httpx and uvicorn libraries. To run it, a user copies an example configuration file, adjusts settings like the upstream service address and authentication mode, and starts it with a single command. It can either forward a caller's own authentication headers, inject its own configured credentials, or fall back between the two, and it includes a safeguard that refuses to leak configured credentials to a different upstream address supplied in a request. The project includes an offline test suite that does not need any extra testing framework to run, covering things like the truncation detection math, parsing of the streaming responses, and the authentication safeguards.
A Python proxy that sits between a coding agent and Codex-style APIs, detecting truncated AI reasoning and quietly continuing it so the agent sees one complete response.
Mainly Python. The stack also includes Python, Starlette, httpx.
The README does not state a license.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.