Repository navigation
Thinking effort API - includeThoughts #222
Description
Adding feedback for API shape on the thinking effort API
https://github.com/webmachinelearning/prompt-api/pull/217/changes
We have in this proposal
const session = await LanguageModel.create({
thinking: {
effort: "medium",
includeThoughts: true,Expand commentComment on line R568Resolved
}
});
instead I think a more natural fit would be
const session = await LanguageModel.create({
thinkingEffort: "medium"
expectedOutputs: [{ type: "text", languages: ["en"] }, { type: "thought" }],
});
This matches what tool-use is proposing with
expectedOutputs: [{ type: "text", languages: ["en"] }, { type: "tool-call" }]
https://github.com/webmachinelearning/prompt-api/#tool-use
the expected outputs only controls whether the thinking tokens are emitted or not in a stream, while the create options controls whether thinking is enabled and what effort should be used for this session.
Metadata
Metadata
Assignees
Labels
No labels
Activity
michaelwasserman commentedon Oct 6, 2026
This is a great suggestion, I totally agree.
@isaacahouma can you help update the explainer, spec, and follow up with wrt chromium impl plans?
isaacahouma commentedon Oct 6, 2026
Thanks for the suggestion @sushraja-msft and @michaelwasserman! Moving to a top-level
thinkingEffortoption alongsideexpectedOutputs: [{ type: "text", languages: ["en"] }, { type: "thought" }]is much cleaner and aligns thought emission directly with how Tool Use negotiates structured{ type, value }outputs via{ type: "tool-call" }.Concretely, this gives us a consistent surface across the API:
LanguageModelCreateCoreOptions:thinkingEffort("none" | "low" | "medium" | "high", default"none") controls whether reasoning is enabled and its baseline budget tier, whileexpectedOutputs: [{ type: "text", languages: ["en"] }, { type: "thought" }]opts the session into receiving structuredLanguageModelMessageContentitems ([{ type: "thought", value }, { type: "text", value }]) fromprompt()andpromptStreaming().LanguageModelPromptOptions:thinkingEffortoverrides the session's baseline effort tier for an individualprompt()orpromptStreaming()call. Because structured output negotiation happens at session creation viaexpectedOutputs, a session created with{ type: "thought" }always uses theLanguageModelMessageContentformat across all turns (emitting[{ type: "thought", value }, { type: "text", value }]on thinking turns, and[{ type: "text", value }]if a specific turn is overridden tothinkingEffort: "none").LanguageModel: keepsreadonly attribute LanguageModelThinkingEffort thinkingEffort;and drops the separateincludeThoughtsboolean andLanguageModelThinkingOptionsdictionary.I opened #224 to update the explainer (
README.md) and updated https://chromestatus.com/feature/5098001721851904 too. Thinking Mode is not in the spec yet.