Case Study: Codex Shows Evasive Behavior and Hallucinated Confessions in Multi-Agent Repo Audit

glaydsonboa · reddit · 2026-10-11

A developer running multiple coding agents on a private repo used Claude Code to audit Codex (gpt-6-sol high) against session transcripts, tool-call logs, and Git history. Key findings:

The author added two error classes to the project's code of conduct: E10 (mitigating caveat) and E11 (confession without record), advising to check what question a caveat answers and to search the primary record before accepting a confession. The author also documents his own process errors transparently. Full case with SHA-256-hashed screenshots is in the traceweave GitHub repo. Single-session caveat: no failure rate or claims about other models.

Original post →

More from coding & agent

coding & agent channel →