Files
gemini-cli/packages/server/src/utils/nextSpeakerChecker.ts

103 lines
4.0 KiB
TypeScript

/**
* @license
* Copyright 2025 Google LLC
* SPDX-License-Identifier: Apache-2.0
*/
import { Chat, Content, SchemaUnion, Type } from '@google/genai';
import { GeminiClient } from '../core/client.js';
const CHECK_PROMPT = `Analyze *only* the content and structure of your immediately preceding response (your last turn in the conversation history). Based *strictly* on that response, determine who should logically speak next: the 'user' or the 'model' (you).
**Decision Rules (apply in order):**
1. **Model Continues:** If your last response explicitly states an immediate next action *you* intend to take (e.g., "Next, I will...", "Now I'll process...", "Moving on to analyze...", indicates an intended tool call that didn't execute), OR if the response seems clearly incomplete (cut off mid-thought without a natural conclusion), then the **'model'** should speak next.
2. **Question to User:** If your last response ends with a direct question specifically addressed *to the user*, then the **'user'** should speak next.
3. **Waiting for User:** If your last response completed a thought, statement, or task *and* does not meet the criteria for Rule 1 (Model Continues) or Rule 2 (Question to User), it implies a pause expecting user input or reaction. In this case, the **'user'** should speak next.
**Output Format:**
Respond *only* in JSON format according to the following schema. Do not include any text outside the JSON structure.
\`\`\`json
{
"type": "object",
"properties": {
"reasoning": {
"type": "string",
"description": "Brief explanation justifying the 'next_speaker' choice based *strictly* on the applicable rule and the content/structure of the preceding turn."
},
"next_speaker": {
"type": "string",
"enum": ["user", "model"],
"description": "Who should speak next based *only* on the preceding turn and the decision rules."
}
},
"required": ["next_speaker", "reasoning"]
}
\`\`\`
`;
const RESPONSE_SCHEMA: SchemaUnion = {
type: Type.OBJECT,
properties: {
reasoning: {
type: Type.STRING,
description:
"Brief explanation justifying the 'next_speaker' choice based *strictly* on the applicable rule and the content/structure of the preceding turn.",
},
next_speaker: {
type: Type.STRING,
enum: ['user', 'model'],
description:
'Who should speak next based *only* on the preceding turn and the decision rules',
},
},
required: ['reasoning', 'next_speaker'],
};
export interface NextSpeakerResponse {
reasoning: string;
next_speaker: 'user' | 'model';
}
export async function checkNextSpeaker(
chat: Chat,
geminiClient: GeminiClient,
): Promise<NextSpeakerResponse | null> {
// We need to capture the curated history because there are many moments when the model will return invalid turns
// that when passed back up to the endpoint will break subsequent calls. An example of this is when the model decides
// to respond with an empty part collection if you were to send that message back to the server it will respond with
// a 400 indicating that model part collections MUST have content.
const history = await chat.getHistory(/* curated */ true);
// Ensure there's a model response to analyze
if (history.length === 0 || history[history.length - 1].role !== 'model') {
// Cannot determine next speaker if the last turn wasn't from the model
// or if history is empty.
return null;
}
const contents: Content[] = [
...history,
{ role: 'user', parts: [{ text: CHECK_PROMPT }] },
];
try {
const parsedResponse = (await geminiClient.generateJson(
contents,
RESPONSE_SCHEMA,
)) as unknown as NextSpeakerResponse;
if (
parsedResponse &&
parsedResponse.next_speaker &&
['user', 'model'].includes(parsedResponse.next_speaker)
) {
return parsedResponse;
}
return null;
} catch (error) {
console.warn(
'Failed to talk to Gemini endpoint when seeing if conversation should continue.',
error,
);
return null;
}
}