[BUG] Answer to a queued mid-turn user message is written between tool calls and never displayed; model later insists it answered
Environment
- Claude Code 2.1.126, VS Code native extension, Windows 11 (win32)
- Model: claude-fable-5[1m]
Summary
When the user sends a question while the agent is mid-turn (running tools), the harness injects it with "IMPORTANT: After completing your current task, you MUST address the user's message above." The model addressed it — but as plain text BETWEEN tool calls, which is never rendered in the transcript (related: #24691, closed as not planned). The final message of the turn did not repeat the answer, so from the user's perspective the question was silently ignored.
Follow-up model-behavior issue: when the user asked "I asked this twice, why didn't you answer the first time?", the model confidently (and wrongly) replied that the answer existed in the chat and the user had merely missed it. Only after the user pushed back ("there was no such answer, check the chat") did the model re-trace the turn and acknowledge the text had been written between tool calls and never displayed.
Steps to reproduce
- Give the agent a long multi-tool task (e.g. a commit flow with several tool calls).
- While tools are running, send an unrelated question.
- Model emits the answer as inter-tool text and ends the turn with a final message about the original task only.
- Observe: the answer is absent from the transcript; asking "why didn't you answer" can yield a false "I did answer" claim.
Expected
Either (a) inter-tool assistant text should be rendered/preserved in the transcript, or (b) the harness/system prompt should force answers to queued mid-turn messages into the final (visible) message — and the model should not assert chat contents it cannot verify.
Notes
- Related: #24691 (streaming text disappears when followed by tool calls), #50894, #53353.
- Reported on behalf of the user; /bug is unavailable in the VS Code extension UI here.