Model echoes the internal task-routing wrapper as its first reply instead of acting on the prompt

On the first turn of a session, instead of answering, the assistant printed the internal task-routing envelope back to me as its reply. The visible text started with "You have received a new task from the user. Here is the full contents of the task:", then my prompt wrapped in <user_task> tags, then the routing instructions about deciding whether to handle the task directly or hand it to a subagent with the Agent tool.

So my whole prompt got quoted back at me inside internal scaffolding, together with instructions that are clearly meant for the model and not for me to read.

I checked the session transcript (.jsonl) to localize it. The wrapper is stored as an assistant text block with model claude-opus-4-8, so the model generated it. It is not a rendering glitch where an injected block showed up in the wrong place. The model reproduced its own routing prompt as output on the first turn.

Expected: the routing envelope stays internal. I see a normal assistant reply, or a normal Agent/Task tool call if it delegates.

Actual: the full envelope (the <user_task> block plus the "use your best judgement ... use the Agent tool" instructions) came back as the visible assistant message.

Repro: intermittent, on the first turn of a session. No deterministic trigger found yet. It happened on a session that had several system-reminder blocks injected before my first message (deferred tool list, skills list, memory context).

Environment:

• Claude Code 2.1.201
• OS: Linux
• Model: Opus 4.8 (1M context)
• Node v22.23.1