OpenAI Codex team fixes OpenClaw authentication vulnerability, OpenAI model Agent behavior greatly improved

ME News report, April 23 (UTC+8), according to Beating monitoring, the OpenAI Codex team is focusing on optimizing the experience of OpenAI models within OpenClaw. Codex engineering lead Tibo Sottiaux said the team is working with OpenClaw creator Peter Steinberger, teasing more progress next week. Codex product lead Nik Pash discovered a critical issue during debugging: when OpenClaw is configured to use the Codex harness with OpenAI models, the authentication flow is faulty, and the system silently falls back to the Pi harness, causing users to think the Codex harness is working normally when it actually isn't. Pash submitted two PRs, one to fix the authentication bridge and the other to prevent silent fallback. The agent behavior differs significantly before and after the fix. When using the Pi harness, the agent only performs shallow polling on each heartbeat: reads the heartbeat file, checks Discord, returns HEARTBEAT_OK, ignores other instructions, and sometimes infers actions to execute but does not issue tool calls. After switching to the Codex harness, the agent enters a full workflow loop: reads workspace context, parses the task list, checks the repository, performs edits, attempts validation, and subsequent heartbeats can pick up from progress rather than repeating work. Pash stated that the agent's prompt and upper-level workflow logic were unchanged; the improvement came entirely from switching the underlying runtime adapter (harness), i.e., how OpenClaw communicates with the model API changed. (Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned