Mid-turn message injection tells the model to "continue this turn", which produces bundled answers when the injected message raises a separate topic
@
When I send a message while Claude is still working, the harness injects it into the running turn wrapped in a system-reminder ending with words to the effect of "Address the message above as you continue this turn."
That instruction is fine when my message refines what Claude is already doing. It misfires when my message raises a different topic. Claude then answers both in one message, interleaved, because it has been told to address the new message as a continuation rather than as a separate thing needing its own turn.
Concretely: Claude was part-way through answering a question of mine. I interrupted with an unrelated request. The reply that came back merged the tail of the first answer with the start of the second — two topics, one message, neither cleanly delivered. I had to re-read it to work out which parts belonged to which.
This is not the same as the known issues where mid-turn messages are lost (#79609, #79656) — the message was received and acted on. Nor is it about steering mechanics (#64624, #30492). The delivery works; the instruction attached to the delivery is what causes the bad shape.
Why it matters beyond tidiness. Interrupting is what a user does when something is urgent or has gone off track. That is precisely the moment a clear, single-topic reply matters most, and the current wording makes a jumbled one more likely. It also cuts against the one-thing-at-a-time behaviour many users configure their agents for — a project-level instruction to answer one thing at a time is silently overridden by a harness instruction to fold the new message into the current answer.
Suggested fix. Have the injected wording distinguish the two cases rather than assuming continuation. Something closer to: "The user sent this while you were working. If it refines the current task, fold it in. If it raises a separate topic, finish or wind up what you are doing, then address it as its own reply." The model can tell the difference; it just is not currently asked to.
Alternatively, leave the wording alone and let the model choose the ordering explicitly — the failure is that no choice is presented, so continuation is the default.
@