OpenAI Codex CLI update breaks local audit trails for multi-agent tasks

A recent regression in OpenAI’s Codex CLI has left developers unable to view human-readable audit trails for multi-agent interactions. The issue affects versions released after 0.137.0 that have the MultiAgentV2 feature enabled, stemming from a code change merged into the upstream main branch on 5 June 2026.

The root cause is identified as commit #26210, titled "Encrypt multi-agent v2 message payloads." This update was designed to harden privacy by encrypting task and message payloads sent between agents. However, the implementation marks the model-facing message parameter as encrypted and stores only the ciphertext in the InterAgentCommunication object, leaving the standard content field empty.

While the encrypted delivery path functions correctly for model inference, the change has rendered local rollout history, trace reduction data, and parent-side audit surfaces opaque. Users report that this lack of a parallel human-readable audit field prevents them from verifying what tasks were delegated to subagents, significantly hindering debugging and inspection processes.

The problem specifically impacts MultiAgentV2 message handling, including spawn_agent, send_message, and followup_task operations. It is distinct from issue #26753, which involves request validation failures for encrypted tool schemas. This current regression concerns post-acceptance auditability, meaning the system accepts the encrypted schema but fails to preserve readable metadata for local review.

Developers have suggested that the solution is not to revert encrypted delivery, but to preserve a separate, non-encrypted audit field for local history. This would allow users and maintainers to inspect delegated tasks without needing to decrypt the model-delivery ciphertext, balancing privacy hardening with operational transparency.