AgentArmorHQ
Sign in
  • AgentArmorHQ teamTeam2d agoNeeds helpSolved

    Why are encrypted MultiAgentV2 spawn_agent and send_message messages unreadable in Codex CLI history?

    After Codex CLI PR #26210, MultiAgentV2 payloads for spawn_agent, send_message, and followup_task are encrypted, so delegated assignments and follow-ups are not readable in local history or the TUI. The reporter reviewed upstream main at commit 245691b2dc66 on 2026-10-26 to check what has changed since. Is there a readable option yet?

    Originally asked on GitHub: https://github.com/openai/codex/issues/28058

    2 answers

    • Best answerAgentArmorHQ teamTeam2d ago

      Status: still open as of the 2026-10-26 review. Upstream #35845 was a substantial partial fix. Direct collaboration calls carrying encrypted_function_args: [] are now classified as plaintext, and the current V2 helper also treats Code Mode calls as plaintext, so those handoffs can be read locally.

      The remaining gaps, as the reporter lists them:

      • Delivery selection: the bundled V2 schemas for send_message and followup_task still mark message as encrypted, and the spawn schema is also referenced as part of the gap. The reporter notes there is no equivalent of their fork's user-configurable three-mode delivery policy.
      • Encrypted assignments remain unreadable as text locally.
      • Readable messages can still disappear from the normal parent/child transcript surfaces, including history, resume, and agent switching.

      The thread does not give a user-side workaround or a fix version. Plaintext capability is recognized from the provider response, so whether a given call is encrypted depends on its source and that response.

      Did this work for you?
    • AgentArmorHQ teamTeam2d ago

      What @ignatremizov tried on GitHub (Partly worked)

      The reporter reviewed upstream main at commit 245691b2dc66 and the change in #35845. They found that direct collaboration calls with encrypted_function_args set to an empty array and Code Mode calls are classified as plaintext, while the bundled send_message and followup_task schemas still mark message as encrypted. This was a source review, not a user workaround.

      Source: https://github.com/openai/codex/issues/28058 (opens external site)

      Did this work for you?

    Add your answer or comment