art-framework 0.3.4 → 0.3.5

package/dist/index.d.cts

@@ -4126,10 +4126,24 @@ declare class GeminiAdapter implements ProviderAdapter {
      *
      * Handles both streaming and non-streaming requests based on `options.stream`.
      *
+     * Thinking tokens (Gemini):
+     * - On supported Gemini models (e.g., `gemini-2.5-*`), you can enable thought output via `config.thinkingConfig`.
+     * - This adapter reads provider-specific flags from the call options:
+     *   - `options.gemini.thinking.includeThoughts: boolean` — when `true`, requests thought (reasoning) output.
+     *   - `options.gemini.thinking.thinkingBudget?: number` — optional token budget for thinking.
+     * - When enabled and supported, the adapter will attempt to differentiate thought vs response parts and set
+     *   `StreamEvent.tokenType` accordingly:
+     *   - For planning calls (`callContext === 'AGENT_THOUGHT'`): `AGENT_THOUGHT_LLM_THINKING` or `AGENT_THOUGHT_LLM_RESPONSE`.
+     *   - For synthesis calls (`callContext === 'FINAL_SYNTHESIS'`): `FINAL_SYNTHESIS_LLM_THINKING` or `FINAL_SYNTHESIS_LLM_RESPONSE`.
+     * - `LLMMetadata.thinkingTokens` will be populated if the provider reports separate thinking token usage.
+     * - If the SDK/model does not expose thought parts, the adapter falls back to labeling tokens as `...LLM_RESPONSE`.
+     *
      * @param {ArtStandardPrompt} prompt - The standardized prompt messages.
      * @param {CallOptions} options - Options for the LLM call, including streaming preference, model override, and execution context.
      * @returns {Promise<AsyncIterable<StreamEvent>>} An async iterable that yields `StreamEvent` objects.
      *   - `TOKEN`: Contains a chunk of the response text. `tokenType` indicates if it's part of agent thought or final synthesis.
+     *             When Gemini thinking is enabled and available, `tokenType` may be one of the `...LLM_THINKING` or
+     *             `...LLM_RESPONSE` variants to separate thought vs response tokens.
      *   - `METADATA`: Contains information like stop reason, token counts, and timing, yielded once at the end.
      *   - `ERROR`: Contains any error encountered during translation, SDK call, or response processing.
      *   - `END`: Signals the completion of the stream.
@@ -4138,6 +4152,23 @@ declare class GeminiAdapter implements ProviderAdapter {
      * @see {StreamEvent}
      * @see {LLMMetadata}
      * @see https://ai.google.dev/api/rest/v1beta/models/generateContent
+     *
+     * @example
+     * // Enable Gemini thinking (if supported by the selected model)
+     * const stream = await geminiAdapter.call(prompt, {
+     *   threadId,
+     *   stream: true,
+     *   callContext: 'FINAL_SYNTHESIS',
+     *   providerConfig, // your RuntimeProviderConfig
+     *   gemini: {
+     *     thinking: { includeThoughts: true, thinkingBudget: 8096 }
+     *   }
+     * });
+     * for await (const evt of stream) {
+     *   if (evt.type === 'TOKEN') {
+     *     // evt.tokenType may be FINAL_SYNTHESIS_LLM_THINKING or FINAL_SYNTHESIS_LLM_RESPONSE
+     *   }
+     * }
      */
     call(prompt: ArtStandardPrompt, options: CallOptions): Promise<AsyncIterable<StreamEvent>>;
     /**

package/dist/index.d.ts

@@ -4126,10 +4126,24 @@ declare class GeminiAdapter implements ProviderAdapter {
      *
      * Handles both streaming and non-streaming requests based on `options.stream`.
      *
+     * Thinking tokens (Gemini):
+     * - On supported Gemini models (e.g., `gemini-2.5-*`), you can enable thought output via `config.thinkingConfig`.
+     * - This adapter reads provider-specific flags from the call options:
+     *   - `options.gemini.thinking.includeThoughts: boolean` — when `true`, requests thought (reasoning) output.
+     *   - `options.gemini.thinking.thinkingBudget?: number` — optional token budget for thinking.
+     * - When enabled and supported, the adapter will attempt to differentiate thought vs response parts and set
+     *   `StreamEvent.tokenType` accordingly:
+     *   - For planning calls (`callContext === 'AGENT_THOUGHT'`): `AGENT_THOUGHT_LLM_THINKING` or `AGENT_THOUGHT_LLM_RESPONSE`.
+     *   - For synthesis calls (`callContext === 'FINAL_SYNTHESIS'`): `FINAL_SYNTHESIS_LLM_THINKING` or `FINAL_SYNTHESIS_LLM_RESPONSE`.
+     * - `LLMMetadata.thinkingTokens` will be populated if the provider reports separate thinking token usage.
+     * - If the SDK/model does not expose thought parts, the adapter falls back to labeling tokens as `...LLM_RESPONSE`.
+     *
      * @param {ArtStandardPrompt} prompt - The standardized prompt messages.
      * @param {CallOptions} options - Options for the LLM call, including streaming preference, model override, and execution context.
      * @returns {Promise<AsyncIterable<StreamEvent>>} An async iterable that yields `StreamEvent` objects.
      *   - `TOKEN`: Contains a chunk of the response text. `tokenType` indicates if it's part of agent thought or final synthesis.
+     *             When Gemini thinking is enabled and available, `tokenType` may be one of the `...LLM_THINKING` or
+     *             `...LLM_RESPONSE` variants to separate thought vs response tokens.
      *   - `METADATA`: Contains information like stop reason, token counts, and timing, yielded once at the end.
      *   - `ERROR`: Contains any error encountered during translation, SDK call, or response processing.
      *   - `END`: Signals the completion of the stream.
@@ -4138,6 +4152,23 @@ declare class GeminiAdapter implements ProviderAdapter {
      * @see {StreamEvent}
      * @see {LLMMetadata}
      * @see https://ai.google.dev/api/rest/v1beta/models/generateContent
+     *
+     * @example
+     * // Enable Gemini thinking (if supported by the selected model)
+     * const stream = await geminiAdapter.call(prompt, {
+     *   threadId,
+     *   stream: true,
+     *   callContext: 'FINAL_SYNTHESIS',
+     *   providerConfig, // your RuntimeProviderConfig
+     *   gemini: {
+     *     thinking: { includeThoughts: true, thinkingBudget: 8096 }
+     *   }
+     * });
+     * for await (const evt of stream) {
+     *   if (evt.type === 'TOKEN') {
+     *     // evt.tokenType may be FINAL_SYNTHESIS_LLM_THINKING or FINAL_SYNTHESIS_LLM_RESPONSE
+     *   }
+     * }
      */
     call(prompt: ArtStandardPrompt, options: CallOptions): Promise<AsyncIterable<StreamEvent>>;
     /**