ReActV2: a plain-text answer step is never a thought; plain-text replay keeps roles alternating

[deepfates]

User story

A host changes a program's ending mid-conversation, from finish: :auto to finish: :submit, and passes the same history on.

Under finish: :auto, a step that answered in plain text is recorded with its text and the signature's outputs. In 0.14.0 the Chat adapter took that step for a thought-only step under finish: :submit. So "Your message above was not sent. To reply, call submit with …" followed every answer that had in fact been delivered. Models then sent those answers again. In one rehearsal a model re-filed an answer it said it had "composed yesterday and never cleared the send queue".

Change

An answer is never a thought.

• A step that carries any of the signature's outputs is an answer. It gets no unsent line, and thought_as_request never moves it to the user side.
• A submit step's answer is its call, so it was already excluded.
• Only a step with text, no calls and no outputs is a thought.
• The adapter takes the output names from the loop's guidance.

The reverse case: a submit-era history replayed in a plain-text loop. A thought-only step used to render as a bare assistant message, often right before the answer's. That is two assistant messages in a row, and in a plain-text loop it reads as a message that was sent.

• Each genuine thought-only step (text, no calls, no outputs) is now followed by a user message: "Your message above was not sent."
• This is the factual half of the line. The submit instruction is left out because a plain-text loop has no submit, so naming it would be false.
• Roles alternate (thought, line, answer), and nothing is merged or rewritten.
• The sentence comes from ReActV2's guidance for a plain-text signature under finish: :auto.

Plain-text-only histories are unchanged. A history written only in a plain-text loop has no thought-only step: its text steps carry the answer. So its requests are byte-for-byte what 0.14.0 sent.

• A test asserts the exact replayed request.
• I also compared the full request term against origin/main. Both SHA-256 hashes match.

Still two assistant messages in a row, outside this rule: one existing case still replays that way under :auto. A step whose only call was a submit the loop rejected loses the call and its error result on replay, so its text appears bare. That step had a call, so it is not a thought-only step under the definition here. Its existing test (react_v2_test.exs, "a rejected or id-less recorded submit…") is restored unchanged.

Known gap

Under finish: :submit, consider a forced request with nothing new to ask (no inputs, no note) on a history whose last step is a plain-text answer. It still ends on that assistant message, because an answer is not moved. Resuming with new inputs or a note ends on a user turn. This is stated in react.md section 3 and in the CHANGELOG.

Tests

test/react_v2_ending_switch_test.exs. The first two fail on main; the third passes on main and here by design:

• Two answered turns from finish: :auto, replayed under finish: :submit: user, answer, user, answer, user. There is no line.
• A submit-era thought plus answer, replayed under finish: :auto: thought, "Your message above was not sent.", answer, then the new ticket.
• A plain-text-only history (answers and a tool step) replays exactly as before, with no "not sent" anywhere.

There are CHANGELOG entries under Fixed, including the known gap. mix check and mix quality.check pass.