Codex Agents Judged Files They Had Not Fully Read Due to Hidden Output Limits
A developer running a team of LLM agents on Codex CLI found that tool outputs were being silently truncated, with 31 of 33 sessions on one day losing parts of their reads. The truncation dropped the middle of long files, including reviewer findings and work plans, and the agents still made decisions from the incomplete text. Raising the configured token limit did not help, because a separate per-cell budget in Codex's JavaScript exec mode capped returned output at 10,000 tokens. The model rarely set the required pragma that would lift that cap, and the problem is tracked as an open issue in the openai/codex repository. The developer responded by hardcoding a fixed budget into every call and instructing agents to reread any cut sections before relying on them.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in