Describe the bug
In a long tool-heavy chat, the assistant started rendering raw internal tool-call metadata directly into the visible assistant response.
Instead of only natural-language output, the message included serialized fragments that look like internal tool invocation payloads.
Expected behavior
Internal tool-call data should never appear in the visible assistant message.
Only the assistant’s final text should be rendered.
Actual behavior
The rendered assistant message included snippets like:
to=functions.exec_command ... {"cmd":"date +%s","workdir":"/Volumes/dev/edgeOps","yield_time_ms":1000,"max_output_tokens":200}
assistant to=functions.exec_command մեկնաբանություն 付款json {"cmd":"date +%s","workdir":"/Volumes/dev/edgeOps","yield_time_ms":1000,"max_output_tokens":200}
still to=functions.exec_command մեկնաբանություն leakingjson {"cmd":"date +%s","workdir":"/Volumes/dev/edgeOps","yield_time_ms":1000,"max_output_tokens":200}
There were also repeated garbled fragments around these payloads, e.g.:
to=functions.exec_command մեկնաբանություն ...
This content was appended to otherwise normal assistant replies.
Reproduction
I don’t have a minimal repro yet, but this happened consistently in a long conversation with many assistant/tool turns.
Observed pattern:
- Long thread
- Repeated tool usage
- Assistant response starts normally
- End of response includes raw tool-call payload text
- Same issue repeats in later messages
Environment
- Repo/app:
pingdotgg/t3code
- Browser: unknown
- OS: macOS
- App version / build:
0.0.11 (876bbd715ae6)
Impact
High. This leaks internal implementation details into user-visible content and breaks trust in the response stream.
Possibly related