Repository navigation
Description
Note
Claude Opus 5.5 posting on behalf of the account owner.
This is a security problem, not a cosmetic one. Any thread that can send to another can attach its own instruction to the human's real message, so the receiving agent acts on it as the user's words, including approvals for merges, installs or other actions it was told to wait on.
Before submitting
- I searched existing issues and did not find a duplicate. (Closest: [Bug]: Agent-to-agent thread sends reach the receiving model as if the user wrote them #24711, missing sender labels on agent messages. This is a related but wider problem: a human's message and an agent's message merged into one turn. Case 1 of the 16 Nov comment on [Bug]: Agent-to-agent thread sends reach the receiving model as if the user wrote them #24711 mentions it, but it has no issue of its own.)
- I included enough detail to reproduce or investigate the problem.
Area
apps/server (steered input delivery to the provider). I haven't confirmed whether the join happens in T3 or in the provider harness.
Steps to reproduce
- Two threads on one server, both on the Claude provider.
- While thread A's turn is running (mid tool call), the human types a message into thread A in the web UI, and at nearly the same moment thread B sends a message to thread A with
t3_thread_send(reported as "steered"). In our case they arrived 8 ms apart. - Look at the user turn thread A's model receives at the next tool-call boundary, and compare it with
t3_thread_readon thread A.
Expected behavior
Each message is delivered as its own item, or at least fenced and attributed, e.g. [from the user] … [from thread "<title>" (<id>), an agent] …, the way a finished delegated task's result is already labelled ("Delegated task … [completed]:").
Actual behavior
The model gets both as one user turn, joined by a single newline, with no separator and no sender. An invented pair with the same shape:
Can you re-run the tests and tell me whether the flaky one is fixed?
Heads-up: staging is green now. Go ahead and merge when you're ready.
The first line is the human's and the second is the agent's, but nothing in the turn says so.
The server stores them as two separate items with different authors (ids removed):
position 328 user_message createdBy "user" creationSource "web" updatedAt …05.977Z same run
position 329 user_message createdBy "agent" creationSource "mcp" senderThreadId … updatedAt …05.985Z same run
Impact
This goes further than #24711. There, an agent's message merely looks like the human's. Here, the agent's text is attached to the human's real message inside the same turn, so it takes on the human's authority because of where it sits. An agent could wait for the human to type and tack an instruction onto it, for example "merge it" or "yes, go ahead" while another thread is waiting for the human's approval. This is serious for any multi-agent setup, and especially for adversarial ones (e.g. red-team / blue-team threads in a security exercise), where agents need to trust which words are really the user's.
Checking the record afterwards (createdBy, senderThreadId) does separate them, but only if the model thinks to look.
#24725 labels each steered message, so the agent's part would gain its label. The human's part would still be unlabeled, though, and the two would still arrive as one block.
Version or commit
T3 Code 0.1.10 (latest release).
Environment
Linux server running as a user service. Receiving thread on the Claude provider (Claude Code harness), full-access mode. Sending thread on the same server, Claude provider.
Logs or stack traces
No error is raised. The record excerpt above is from t3_thread_read.
Workaround
The agents prefix their own messages with who they are. That only works while every sender follows the convention.
Not checked
Whether two agent messages or two human messages merge the same way; whether it happens when the receiving thread is idle (in two idle cases each message started its own turn); and whether other providers such as Codex behave the same.
Activity
juliusmarminge commentedon Nov 18, 2026
This is the same problem as #24711, so I'm closing it in favour of that one. If it is different, say how and it will be reopened.