Agents: the "finished without a text answer" UI placeholder is persisted and replayed to the model as the agent's own turn (follow-up to #517)

[juanlentino]

Split from #517 per maintainer request (#517 (comment)).

Problem

When an agent run fails to produce answer text, the chat UI renders the placeholder "The agent finished without a text answer." — and that placeholder string is persisted as the agent's conversation turn. Every subsequent run in that conversation replays it to the model as something the agent actually said. Captured from a live composed prompt:

Conversation so far, oldest first:
User: Propose a revision to post 1716 …
You: The agent finished without a text answer.
User: Try again
You: The agent finished without a text answer.

Effects

• The model is told it previously "answered" with a meta-sentence about UI state, which it may imitate, apologize for, or reason about — context pollution that compounds with every failed run in the thread.
• Once AI: surface an empty final answer as an error instead of a silent success #530 lands (empty answers become errors), the equivalent question becomes how error turns are persisted/replayed — probably as nothing, or as a clearly-marked system note rather than an agent utterance.

Suggested direction

Persist failure turns as a distinct row type (or not at all) so the prompt composer either skips them or renders them as bracketed system context, never as You: <ui placeholder>.