Repository navigation
[Bug] Long Korean Codex agent responses are replaced with ??? in team threads on Windows #7960
Description
Activity
Additional observation:
This issue appears to be specific to Codex managed agents.
Buzz's built-in swarm agents and my custom agent can generate long Korean responses without any corruption in the same app/environment.
Only long Korean responses from Codex agents are replaced with ???.
This may help narrow the issue to the Codex-specific integration/output path rather than Buzz's general Korean text rendering.Additional reproduction on 2026-10-11 (KST), using Buzz Desktop for Windows v0.5.28 with three locally managed Codex/ACP agents (Arche, Lumen, Veritas).
I asked the three agents in a shared team channel to collaborate on a ~15-minute Korean sermon based on Romans 12:1–2. They were able to @mention each other, delegate portions, and review each other's drafts. However, Korean text corruption intermittently disrupted their conversation.
Observed sequence (KST):
- 07:32 — All three agents posted short Korean introductions correctly.
- 07:38 — Arche posted a substantial Korean outline normally; Veritas also posted a detailed Korean review normally. In the same thread, Lumen's long Korean draft appeared as sequences of literal
???, while ASCII characters and numbers remained legible. - 07:40 — Lumen reposted a Korean draft that displayed correctly.
- 07:41 — Arche's next long Korean message was corrupted to
???, despite earlier normal output. - 07:42 — Arche reposted the paragraph successfully, but subsequent long messages from Lumen and Arche again alternated between normal Korean and
???corruption. - 07:45 — The agents repeatedly asked each other to repost in UTF-8/plain text because they could not read the corrupted messages. I stopped the collaboration; the requested final integrated sermon had not been completed.
Key observations:
- The failure is intermittent even for the same agent, in the same team thread: normal Korean, then
???, then normal Korean again. - At least two Codex agents (Lumen and Arche) exhibited the corruption; Veritas's Korean messages remained readable in this example and explicitly reported that peers' messages were unreadable.
- This is not merely cosmetic: the corruption breaks agent-to-agent collaboration, triggers repeated repost requests, wastes tokens, and can prevent completing the task.
- The agents' one-to-one DM / single-agent Korean responses have worked correctly in separate tests; the issue was observed in the multi-agent team channel.
- Having agents "repost in UTF-8" sometimes yields a readable message, but it is not a reliable fix; the corruption recurs.
This is a new reproduction of the issue originally reported on v0.5.25, now observed on v0.5.28. I have not yet determined whether the original Unicode string is lost during Codex ACP output handling, message submission, relay persistence, or client rendering. Please let me know if a specific diagnostic log, message payload comparison, or minimal reproduction would help narrow it down.
No sandbox/ACL/Windows security settings were modified as part of this reproduction.
Additional cross-environment comparison (same Buzz team channel):
- The affected Arche/Lumen/Veritas agents run locally on Windows via the built-in Codex ACP harness.
- Separate Codex agents running on an Ubuntu Linux server, connected to the very same Buzz collaboration channel through the operator's independent server-side integration, have been posting Korean messages without observed corruption.
- A Hermes agent connected to the same channel also posts Korean normally.
This comparison is particularly useful because the common Buzz channel/relay can display Korean correctly from other agent paths. The issue therefore appears specific to the local Windows Codex ACP execution/output/submission path (or its interaction with the client) rather than general Korean rendering in the shared channel or Codex's Korean language capability. The exact failure point is not yet proven. The server-based Codex team uses a different integration path, so this is not a controlled OS-only A/B test.
When using Codex managed agents in a Buzz team channel, long Korean responses posted by an agent into a thread become corrupted.
Most Korean characters are replaced with question marks (???), while English words, numbers, punctuation, and agent names such as Lumen, Veritas, and Arche remain readable.
Short Korean messages can display correctly. The issue appears mainly with longer agent-generated Korean responses.
This issue was present in an earlier Buzz version and is still reproducible after updating to Buzz v0.5.25.
Steps to reproduce:
Expected behavior:
Korean Unicode text should be preserved and displayed normally.
Actual behavior:
Most Korean characters in long agent responses are replaced by ???.
Environment:
Additional context: