ReActV2 finish: :submit: a thought-only step continues before submit is forced

[deepfates]

User story

A host runs a program with finish: :submit, so the person sees only what goes through submit. Today, when the model writes a thought and calls no tool, Imp immediately sends the forced submit with every other tool removed. This goes wrong in three ways:

• It cuts off what the model said it would do. "Let me read the thread first." must now be answered without reading the thread.
• It adds a request that can lose the reply. An ordinary provider error on the forced request goes to last_request_failed with no fallback, and so does a spend stop. Neither does the forced submit check whether the deadline has passed, as the last text request does.
• The extractor's output gets published. When the forced request falls through to the extractor, its instructions ask for an honest report of limitations. A host then publishes that report as the answer.

Change

A thought-only step continues the loop (finish: :submit only).

• The next request is an ordinary step: every tool and tool_choice: "auto".
  • With native tools, the thought is the request's last user message under [[ ## next_thought ## ]] (the rendering from ReActV2: finish: :submit opt-in; forced submit never ends on the model's turn #343).
  • With written tool calls, the request ends on the tool listing, after the thought.
• A second thought-only step in a row is the :empty_tool_calls interruption, and the forced submit follows as before. So is a step that says nothing, or only whitespace: there is no thought to show, and the same request would be sent again.
• A step that calls a tool resets the count.
• max_iters still ends the turn with the forced submit.
• Signatures that get submit under finish: :auto keep DSPy's ending, which forces at once. This scope is deliberate: the measurement below covers single-text final steps only.
• The bound is one continued step, and it is not an option. One ordinary step lets the model take the action it announced, or submit what it wrote. In the measurement, a model that writes text a second time is about one sample in eight, and continuing again would mostly spend more requests before the same forced submit. If a host needs a different bound, it can become an option then.

No request after the deadline. If the process's Imp.Deadline has passed, none of these is sent:

• the continued step;
• the forced submit;
• the "required" retry;
• the extraction request.

The turn ends :incomplete with termination_cause: :deadline_exceeded, as the last text request already does.

Extracted outputs are recorded in the history. On an :extracted turn, the history used to end on the forced step's thought. It now ends with the extracted outputs, recorded as a step that called no tool, together with any inputs no step spent. The host's record then holds what the turn produced, and the episode is delimited by its outputs as a submit turn's is.

• This step renders no message on replay. thought_as_request therefore looks past it.
• Without that, resuming after an :extracted turn would end on the forced step's thought as an assistant message. Mistral and Gemini refuse such a request.

What changes under finish: :auto. The default is not entirely unchanged. For every signature with submit:

• the deadline checks above apply;
• an :extracted turn's history now ends with the extracted outputs.

The continuing step applies only under finish: :submit.

Evidence

Probe 5 (2026-10-16): 42 real final steps from 11 models, sent through OpenRouter in the request shape above. On the ordinary step after a thought-only step:

• Most samples called submit with the thought's reply, copied or paraphrased. Meta's models did so 6 of 6 with nothing forced, which matters because their provider cannot force a tool. Kimi, GPT and Gemini models did the same.
• A thought that was a closing note was submitted as an empty answer.
• A DeepSeek model went on to fetch what it had said it would read.
• 5 of 42 wrote text again.

The measurement is recorded in decisions.md, with react.md sections 2 and 3 and the termination tables updated.

Open

On replay, the extracted outputs are in the record but not shown to the model. On a continued conversation, the model reads the forced step's thought, shown as a user message, but not the answer the extractor produced. A submit turn shows its answer through the submit call.

Rendering the extracted outputs would show the model an answer it did not write itself. It would also change the prompt for every replayed history that holds one. I chose to look past the step instead, and the trade-off is left here for review.

Tests

test/react_v2_thought_continues_test.exs, plus one updated test in test/react_v2_finish_submit_test.exs. Each fails on main:

• thought, then the next ordinary step submits (:submit; the second request is auto with every tool and ends on the thought);
• thought, thought, then forced submit (:forced_submit, :empty_tool_calls);
• thought, tool call, thought, submit: every request is auto, and the tool call resets the count;
• the step limit after a thought still forces submit (:max_iters);
• a deadline that has passed sends no forced submit (:incomplete, :deadline_exceeded, one request);
• a deadline that passes before the continued step, the "required" retry or the extractor sends no further request;
• a whitespace-only thought is not continued;
• extracted outputs end the history;
• resuming after an :extracted turn does not end on the model's turn.

mix check and mix quality.check pass.