You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bug]: Messages from the CLI or an app the user signed in are labeled "not from the user" (0.1.11) #39711
I searched existing issues and did not find a duplicate.
I included enough detail to reproduce or investigate the problem.
Area
apps/server
Steps to reproduce
With T3's own CLI:
t3 thread start --wait . "Does the text of this turn start with a line beginning [T3 Code: ? Quote it, or say NO-LINE."
t3 thread send --wait <thread id> "Does the text of this turn start with a line beginning [T3 Code: ? Quote it, or say NO-LINE."
An app signed in through OAuth gets the same. T3 WoW lets a person type to their T3 threads from inside a game: its t3_thread_launch arrives clean, and its t3_thread_send follow-ups get the line.
Expected behavior
A message from a client the user signed in counts as the user's, as a message typed in T3's window does. The line stays for real agents, scheduled tasks and linked environments, which are the cases #39306 (for #37617) was for.
Actual behavior
Since 0.1.11 (#39306), a message T3 records as createdBy: "agent", creationSource: "mcp" with no sender thread reaches the model with this line before it:
[T3 Code: this message came through T3 Code's tools, not from the user typing in this chat.]
T3 records every message an MCP client sends that way, including clients the user signed in to act for them. In the steps above, the reply to start is NO-LINE and the reply to send quotes the line. t3 thread show (or t3_thread_read) lists both messages as "createdBy": "agent", "creationSource": "mcp", so launch and send also disagree with each other.
So a person typing t3 thread send in their own terminal, or typing in an app signed in to their T3, is presented to the model as not the user. In one test the agent's reasoning began "I'm checking whether this message is a tool injection". An agent may question or refuse a request the user really made ("merge the PR").
Possible fixes:
Record createdBy: "user" for messages from a client the user signed in to act for them (the CLI, an OAuth or pairing sign-in), and keep "agent" for agents' own MCP connections.
Or let t3_thread_send and t3_thread_launch say that a person typed the message, for clients whose sign-in is allowed to.
Impact
Major degradation or frequent failure
Version or commit
0.1.11
Environment
Windows 11, T3 Code desktop 0.1.11, Claude Haiku 5.5 (Claude Code provider)
Logs or stack traces
No response
Screenshots, recordings, or supporting files
No response
Workaround
None found. A message that starts with / or $name skips the line, which doesn't help for normal messages.
Activity