You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Three delta kinds beside message and function_call: commentary (text
a model writes for the person while it works), reasoning_summary, streamed
only when the envelope sets reasoningSummaries, and reasoning (raw),
streamed only when it sets rawReasoning. A typed consumer that does not
know them refuses them; a consumer that joins every delta's text into the
answer keeps message only.
PROTOCOL §8: a new delta kind is a breaking change.
Added
reasoningSummaries and rawReasoning on the envelope.
auth.summaries: whether a credential's account may ask for reasoning
summaries, on a provider that allows them only to verified organizations
(openai, chatgpt). Passed with a per-call key as configuration.summaries;
for the key mohdel finds itself, reasoningSummaries in that provider's providers.json entry, read by the gate's LocalAuth and in-process.
phase on a text message part, and commentaryParts on AnswerResult.
Chat-completions providers stream raw reasoning as reasoning deltas —
from reasoning_content, reasoning, or Mistral's thinking chunks — and
OpenRouter's reasoning.summary details as reasoning_summary.
The Responses providers (openai, chatgpt, xai, meta) stream a
commentary-phase message as commentary and return it in commentaryParts; reasoning summaries as reasoning_summary, asked for when
the envelope sets reasoningSummaries — on openai and chatgpt only when auth.summaries is true; raw reasoning text as reasoning.
SUMMARIES_REFUSED: a request refused over reasoning.summary, naming
the setting that stops asking.
Gemini thought summaries stream as reasoning_summary; a Gemini 3 or later
call asks for them (includeThoughts) only when the envelope sets reasoningSummaries.
Anthropic: a thinking call that sets reasoningSummaries gets summarized
thinking (thinking.display: 'summarized'), streamed as reasoning_summary,
progress notes included; any other gets 'omitted'. Thinking tokens are
taken from usage.output_tokens_details.thinking_tokens when the response
reports them.
Fixed
mo ask --stream printed delta objects; it prints the answer's text.
OpenRouter: an effort was sent as reasoning_effort, which Anthropic models
ignore; it goes as reasoning: { effort }.
A Responses provider's commentary was part of output, and went back
without its phase; it goes back as its own messages with their phase,
and the answer beside it as final_answer.
Anthropic thinking blocks were never sent back: a turn with thinking goes
back to the model that produced it as its blocks came, unmodified and in
order. A prompt-cache breakpoint never lands on a thinking block.
A Mistral content array was streamed as one delta; its text is the answer
and its thinking is reasoning.
An assistant turn's reasoning goes back as each provider takes it: as reasoning to Cerebras, as its reasoning_details unmodified to the same
OpenRouter model, as its thinking chunks inside content to Mistral.