@tanstack/ai 0.42.0 → 0.43.1

package/dist/esm/activities/summarize/chat-stream-summarize.js

@@ -1,212 +1,262 @@
-import { EventType } from "@ag-ui/core";
 import { toRunErrorPayload } from "../error-payload.js";
-import { MAX_TOKENS_KEYS } from "../../utilities/sampling-keys.js";
 import { BaseSummarizeAdapter } from "./adapter.js";
-const MAX_TOKENS_KEY_BY_ADAPTER = {
-  openai: "max_output_tokens",
-  anthropic: "max_tokens",
-  grok: "max_tokens",
-  groq: "max_completion_tokens",
-  gemini: "maxOutputTokens",
-  openrouter: "maxCompletionTokens"
+import { MAX_TOKENS_KEYS } from "../../utilities/sampling-keys.js";
+import { EventType } from "@ag-ui/core";
+//#region src/activities/summarize/chat-stream-summarize.ts
+/**
+* Provider-native max-output-tokens key per summarize-adapter `name`. summarize
+* is provider-agnostic and forwards `modelOptions` opaquely to the wrapped text
+* adapter, so `maxLength` must be written under the exact key the underlying
+* provider reads — no adapter reads a generic `maxTokens`. Ollama is the one
+* exception: it nests sampling under `options`, so it has no entry here and is
+* handled as a special nested case in `applyMaxLength`/`applyDefaultTemperature`.
+*
+* Keep in sync with each adapter's wire mapping:
+* - OpenAI (Responses): `max_output_tokens`
+* - Anthropic / Grok: `max_tokens`
+* - Groq: `max_completion_tokens`
+* - Gemini: `maxOutputTokens`
+* - OpenRouter: `maxCompletionTokens`
+* - Ollama: nested `options.num_predict` (no entry — see `applyMaxLength`)
+*/
+var MAX_TOKENS_KEY_BY_ADAPTER = {
+	openai: "max_output_tokens",
+	anthropic: "max_tokens",
+	grok: "max_tokens",
+	groq: "max_completion_tokens",
+	gemini: "maxOutputTokens",
+	openrouter: "maxCompletionTokens"
 };
-const KNOWN_MAX_TOKENS_KEYS = MAX_TOKENS_KEYS;
+/**
+* Every flat key any supported provider uses to cap output tokens (plus the
+* generic `maxTokens` spelling no adapter reads). Used to detect a
+* caller-supplied token limit so the summarize default never overrides an
+* explicit caller value. Shared with the OTel middleware via
+* `MAX_TOKENS_KEYS` so the two spelling sets cannot drift.
+*/
+var KNOWN_MAX_TOKENS_KEYS = MAX_TOKENS_KEYS;
+/**
+* Whether `applyMaxLength` knows how to place a token limit for this adapter
+* `name` (either the nested Ollama shape or a flat provider-native key).
+* Used to surface a warning when `maxLength` would otherwise be silently
+* dropped for an unrecognised adapter name.
+*/
 function isKnownMaxTokensAdapter(adapterName) {
-  return adapterName === "ollama" || MAX_TOKENS_KEY_BY_ADAPTER[adapterName] !== void 0;
+	return adapterName === "ollama" || MAX_TOKENS_KEY_BY_ADAPTER[adapterName] !== void 0;
 }
+/**
+* Apply the low-temperature summarize default to a working copy of the
+* caller's `modelOptions`, placed where the wrapped provider actually reads
+* it (nested under `options` for Ollama, flat otherwise). The caller always
+* wins: if they already set `temperature` in that location, it is untouched.
+*/
 function applyDefaultTemperature(adapterName, temperature, modelOptions) {
-  const merged = { ...modelOptions };
-  if (adapterName === "ollama") {
-    const existing = merged.options && typeof merged.options === "object" ? merged.options : void 0;
-    if (existing && "temperature" in existing) return merged;
-    merged.options = { temperature, ...existing };
-    return merged;
-  }
-  if ("temperature" in merged) return merged;
-  merged.temperature = temperature;
-  return merged;
+	const merged = { ...modelOptions };
+	if (adapterName === "ollama") {
+		const existing = merged.options && typeof merged.options === "object" ? merged.options : void 0;
+		if (existing && "temperature" in existing) return merged;
+		merged.options = {
+			temperature,
+			...existing
+		};
+		return merged;
+	}
+	if ("temperature" in merged) return merged;
+	merged.temperature = temperature;
+	return merged;
 }
+/**
+* Resolve `maxLength` to the provider-native max-output-tokens key for the
+* given summarize-adapter `name` (this wrapper's OWN `name`, not the wrapped
+* text adapter's) and merge it into a working copy of the caller's
+* `modelOptions`. The caller always wins: if they already set any recognised
+* token-limit key (flat or, for Ollama, nested `options.num_predict`), the
+* default is left untouched. Unknown/unrecognised adapter names fall back to
+* NOT setting a token key (the prompt hint still asks the model to stay under
+* `maxLength`) rather than writing a dead key no provider reads.
+*
+* Caveat (intentional): "caller wins" keys off ANY recognised spelling in
+* `KNOWN_MAX_TOKENS_KEYS`, but only the adapter's native key is read on the
+* wire. So a caller who sets a NON-native spelling for this provider — e.g.
+* `maxTokens`, or Anthropic's `max_tokens` against an OpenAI adapter — suppresses
+* the summarize default WITHOUT getting their own value applied either: neither
+* cap reaches the wire. This favours never clobbering a migration leftover over
+* guaranteeing a cap; the prompt-level hint still asks the model to stay under
+* `maxLength`. Rename the key to the provider-native spelling to forward it.
+*/
 function applyMaxLength(adapterName, maxLength, modelOptions) {
-  const merged = { ...modelOptions };
-  if (adapterName === "ollama") {
-    const callerSetFlatLimit = KNOWN_MAX_TOKENS_KEYS.some(
-      (k) => typeof merged[k] === "number"
-    );
-    const existing = merged.options && typeof merged.options === "object" ? merged.options : void 0;
-    if (callerSetFlatLimit || existing && typeof existing.num_predict === "number") {
-      return merged;
-    }
-    merged.options = { num_predict: maxLength, ...existing };
-    return merged;
-  }
-  const key = MAX_TOKENS_KEY_BY_ADAPTER[adapterName];
-  if (key === void 0) return merged;
-  const callerSetLimit = KNOWN_MAX_TOKENS_KEYS.some(
-    (k) => typeof merged[k] === "number"
-  );
-  if (callerSetLimit) return merged;
-  merged[key] = maxLength;
-  return merged;
-}
-class ChatStreamSummarizeAdapter extends BaseSummarizeAdapter {
-  name;
-  textAdapter;
-  constructor(textAdapter, model, name = "chat-stream-summarize") {
-    super({}, model);
-    this.name = name;
-    this.textAdapter = textAdapter;
-  }
-  async summarize(options) {
-    const systemPrompt = this.buildSummarizationPrompt(options);
-    let summary = "";
-    const id = this.generateId();
-    let model = options.model;
-    let usage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
-    options.logger.request(
-      `activity=summarize provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? "unset"}`,
-      { provider: this.name, model: options.model }
-    );
-    try {
-      for await (const chunk of this.textAdapter.chatStream(
-        this.buildTextOptions(options, systemPrompt)
-      )) {
-        if (chunk.type === "TEXT_MESSAGE_CONTENT") {
-          if (chunk.content) {
-            summary = chunk.content;
-          } else if (chunk.delta) {
-            summary += chunk.delta;
-          }
-          model = chunk.model || model;
-        }
-        if (chunk.type === "RUN_FINISHED") {
-          if (chunk.usage) {
-            usage = chunk.usage;
-          }
-        }
-        if (chunk.type === "RUN_ERROR") {
-          const message = (chunk.error && typeof chunk.error.message === "string" ? chunk.error.message : null) ?? "Summarization failed";
-          const code = chunk.error && typeof chunk.error.code === "string" ? chunk.error.code : void 0;
-          const err = new Error(message);
-          if (code) {
-            ;
-            err.code = code;
-          }
-          throw err;
-        }
-      }
-    } catch (error) {
-      options.logger.errors(`${this.name}.summarize fatal`, {
-        error: toRunErrorPayload(error, `${this.name}.summarize failed`),
-        source: `${this.name}.summarize`
-      });
-      throw error;
-    }
-    return { id, model, summary, usage };
-  }
-  async *summarizeStream(options) {
-    const systemPrompt = this.buildSummarizationPrompt(options);
-    options.logger.request(
-      `activity=summarizeStream provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? "unset"}`,
-      { provider: this.name, model: options.model }
-    );
-    const id = this.generateId();
-    let summary = "";
-    let model = options.model;
-    let usage = {
-      promptTokens: 0,
-      completionTokens: 0,
-      totalTokens: 0
-    };
-    try {
-      for await (const chunk of this.textAdapter.chatStream(
-        this.buildTextOptions(options, systemPrompt)
-      )) {
-        if (chunk.type === "TEXT_MESSAGE_CONTENT") {
-          if (chunk.content) {
-            summary = chunk.content;
-          } else if (chunk.delta) {
-            summary += chunk.delta;
-          }
-          if (chunk.model) model = chunk.model;
-        }
-        if (chunk.type === "RUN_FINISHED") {
-          if (chunk.usage) usage = chunk.usage;
-          if (chunk.model) model = chunk.model;
-          yield {
-            type: EventType.CUSTOM,
-            name: "generation:result",
-            value: { id, model, summary, usage },
-            model,
-            timestamp: Date.now()
-          };
-        }
-        yield chunk;
-      }
-    } catch (error) {
-      options.logger.errors(`${this.name}.summarizeStream fatal`, {
-        error: toRunErrorPayload(error, `${this.name}.summarizeStream failed`),
-        source: `${this.name}.summarizeStream`
-      });
-      throw error;
-    }
-  }
-  /**
-   * Build the TextOptions passed to the underlying chatStream. Provider
-   * `modelOptions` from the summarize call are forwarded as-is so knobs like
-   * Anthropic cache headers, Gemini safety settings, or Ollama tuning params
-   * still reach the wire layer.
-   */
-  buildTextOptions(options, systemPrompt) {
-    let working = {
-      ...options.modelOptions
-    };
-    working = applyDefaultTemperature(this.name, 0.3, working);
-    if (options.maxLength !== void 0) {
-      if (!isKnownMaxTokensAdapter(this.name)) {
-        options.logger.warn(
-          `summarize: maxLength=${options.maxLength} could not be mapped to a provider token key for adapter name "${this.name}" — it was dropped from modelOptions (the prompt still asks the model to stay under it). Construct ChatStreamSummarizeAdapter with a recognised provider name to forward the cap.`,
-          { provider: this.name }
-        );
-      }
-      working = applyMaxLength(this.name, options.maxLength, working);
-    }
-    const modelOptions = working;
-    return {
-      model: options.model,
-      messages: [{ role: "user", content: options.text }],
-      systemPrompts: [systemPrompt],
-      modelOptions,
-      logger: options.logger
-    };
-  }
-  buildSummarizationPrompt(options) {
-    let prompt = "You are a professional summarizer. ";
-    switch (options.style) {
-      case "bullet-points":
-        prompt += "Provide a summary in bullet point format. ";
-        break;
-      case "paragraph":
-        prompt += "Provide a summary in paragraph format. ";
-        break;
-      case "concise":
-        prompt += "Provide a very concise summary in 1-2 sentences. ";
-        break;
-      case void 0:
-        prompt += "Provide a clear and concise summary. ";
-        break;
-      default:
-        prompt += "Provide a clear and concise summary. ";
-    }
-    if (options.focus && options.focus.length > 0) {
-      prompt += `Focus on the following aspects: ${options.focus.join(", ")}. `;
-    }
-    if (options.maxLength) {
-      prompt += `Keep the summary under ${options.maxLength} tokens. `;
-    }
-    return prompt;
-  }
+	const merged = { ...modelOptions };
+	if (adapterName === "ollama") {
+		const callerSetFlatLimit = KNOWN_MAX_TOKENS_KEYS.some((k) => typeof merged[k] === "number");
+		const existing = merged.options && typeof merged.options === "object" ? merged.options : void 0;
+		if (callerSetFlatLimit || existing && typeof existing.num_predict === "number") return merged;
+		merged.options = {
+			num_predict: maxLength,
+			...existing
+		};
+		return merged;
+	}
+	const key = MAX_TOKENS_KEY_BY_ADAPTER[adapterName];
+	if (key === void 0) return merged;
+	if (KNOWN_MAX_TOKENS_KEYS.some((k) => typeof merged[k] === "number")) return merged;
+	merged[key] = maxLength;
+	return merged;
 }
-export {
-  ChatStreamSummarizeAdapter
+/**
+* Summarize adapter that wraps any `ChatStreamCapable` text adapter and
+* prompts it for summarization. Not tied to any wire format.
+*/
+var ChatStreamSummarizeAdapter = class extends BaseSummarizeAdapter {
+	name;
+	textAdapter;
+	constructor(textAdapter, model, name = "chat-stream-summarize") {
+		super({}, model);
+		this.name = name;
+		this.textAdapter = textAdapter;
+	}
+	async summarize(options) {
+		const systemPrompt = this.buildSummarizationPrompt(options);
+		let summary = "";
+		const id = this.generateId();
+		let model = options.model;
+		let usage = {
+			promptTokens: 0,
+			completionTokens: 0,
+			totalTokens: 0
+		};
+		options.logger.request(`activity=summarize provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? "unset"}`, {
+			provider: this.name,
+			model: options.model
+		});
+		try {
+			for await (const chunk of this.textAdapter.chatStream(this.buildTextOptions(options, systemPrompt))) {
+				if (chunk.type === "TEXT_MESSAGE_CONTENT") {
+					if (chunk.content) summary = chunk.content;
+					else if (chunk.delta) summary += chunk.delta;
+					model = chunk.model || model;
+				}
+				if (chunk.type === "RUN_FINISHED") {
+					if (chunk.usage) usage = chunk.usage;
+				}
+				if (chunk.type === "RUN_ERROR") {
+					const message = (chunk.error && typeof chunk.error.message === "string" ? chunk.error.message : null) ?? "Summarization failed";
+					const code = chunk.error && typeof chunk.error.code === "string" ? chunk.error.code : void 0;
+					const err = new Error(message);
+					if (code) err.code = code;
+					throw err;
+				}
+			}
+		} catch (error) {
+			options.logger.errors(`${this.name}.summarize fatal`, {
+				error: toRunErrorPayload(error, `${this.name}.summarize failed`),
+				source: `${this.name}.summarize`
+			});
+			throw error;
+		}
+		return {
+			id,
+			model,
+			summary,
+			usage
+		};
+	}
+	async *summarizeStream(options) {
+		const systemPrompt = this.buildSummarizationPrompt(options);
+		options.logger.request(`activity=summarizeStream provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? "unset"}`, {
+			provider: this.name,
+			model: options.model
+		});
+		const id = this.generateId();
+		let summary = "";
+		let model = options.model;
+		let usage = {
+			promptTokens: 0,
+			completionTokens: 0,
+			totalTokens: 0
+		};
+		try {
+			for await (const chunk of this.textAdapter.chatStream(this.buildTextOptions(options, systemPrompt))) {
+				if (chunk.type === "TEXT_MESSAGE_CONTENT") {
+					if (chunk.content) summary = chunk.content;
+					else if (chunk.delta) summary += chunk.delta;
+					if (chunk.model) model = chunk.model;
+				}
+				if (chunk.type === "RUN_FINISHED") {
+					if (chunk.usage) usage = chunk.usage;
+					if (chunk.model) model = chunk.model;
+					yield {
+						type: EventType.CUSTOM,
+						name: "generation:result",
+						value: {
+							id,
+							model,
+							summary,
+							usage
+						},
+						model,
+						timestamp: Date.now()
+					};
+				}
+				yield chunk;
+			}
+		} catch (error) {
+			options.logger.errors(`${this.name}.summarizeStream fatal`, {
+				error: toRunErrorPayload(error, `${this.name}.summarizeStream failed`),
+				source: `${this.name}.summarizeStream`
+			});
+			throw error;
+		}
+	}
+	/**
+	* Build the TextOptions passed to the underlying chatStream. Provider
+	* `modelOptions` from the summarize call are forwarded as-is so knobs like
+	* Anthropic cache headers, Gemini safety settings, or Ollama tuning params
+	* still reach the wire layer.
+	*/
+	buildTextOptions(options, systemPrompt) {
+		let working = { ...options.modelOptions };
+		working = applyDefaultTemperature(this.name, .3, working);
+		if (options.maxLength !== void 0) {
+			if (!isKnownMaxTokensAdapter(this.name)) options.logger.warn(`summarize: maxLength=${options.maxLength} could not be mapped to a provider token key for adapter name "${this.name}" — it was dropped from modelOptions (the prompt still asks the model to stay under it). Construct ChatStreamSummarizeAdapter with a recognised provider name to forward the cap.`, { provider: this.name });
+			working = applyMaxLength(this.name, options.maxLength, working);
+		}
+		const modelOptions = working;
+		return {
+			model: options.model,
+			messages: [{
+				role: "user",
+				content: options.text
+			}],
+			systemPrompts: [systemPrompt],
+			modelOptions,
+			logger: options.logger,
+			...options.runId !== void 0 ? { runId: options.runId } : {},
+			...options.threadId !== void 0 ? { threadId: options.threadId } : {}
+		};
+	}
+	buildSummarizationPrompt(options) {
+		let prompt = "You are a professional summarizer. ";
+		switch (options.style) {
+			case "bullet-points":
+				prompt += "Provide a summary in bullet point format. ";
+				break;
+			case "paragraph":
+				prompt += "Provide a summary in paragraph format. ";
+				break;
+			case "concise":
+				prompt += "Provide a very concise summary in 1-2 sentences. ";
+				break;
+			case void 0:
+				prompt += "Provide a clear and concise summary. ";
+				break;
+			default: prompt += "Provide a clear and concise summary. ";
+		}
+		if (options.focus && options.focus.length > 0) prompt += `Focus on the following aspects: ${options.focus.join(", ")}. `;
+		if (options.maxLength) prompt += `Keep the summary under ${options.maxLength} tokens. `;
+		return prompt;
+	}
 };
-//# sourceMappingURL=chat-stream-summarize.js.map
+//#endregion
+export { ChatStreamSummarizeAdapter };
+
+//# sourceMappingURL=chat-stream-summarize.js.map

package/dist/esm/activities/summarize/chat-stream-summarize.js.map

@@ -1 +1 @@
-{"version":3,"file":"chat-stream-summarize.js","sources":["../../../../src/activities/summarize/chat-stream-summarize.ts"],"sourcesContent":["import { EventType } from '@ag-ui/core'\nimport { toRunErrorPayload } from '../error-payload'\nimport { MAX_TOKENS_KEYS } from '../../utilities/sampling-keys'\nimport { BaseSummarizeAdapter } from './adapter'\nimport type {\n  StreamChunk,\n  SummarizationOptions,\n  SummarizationResult,\n  TextOptions,\n} from '../../types'\n\n/**\n * Minimal contract for a text adapter that supports `chatStream`. Lets\n * `ChatStreamSummarizeAdapter` work with any text adapter without coupling\n * to a specific implementation.\n *\n * The provider-options shape is intentionally `any` here — the wrapper only\n * forwards `modelOptions` straight through, so a text adapter with a richer\n * per-model options type (e.g. `ResolveProviderOptions<TModel>`) is still\n * acceptable. Summarize-level type safety is enforced via\n * `SummarizationOptions<TProviderOptions>` on the wrapper itself.\n */\nexport interface ChatStreamCapable {\n  chatStream: (options: TextOptions<any>) => AsyncIterable<StreamChunk>\n}\n\n/**\n * Provider-native max-output-tokens key per summarize-adapter `name`. summarize\n * is provider-agnostic and forwards `modelOptions` opaquely to the wrapped text\n * adapter, so `maxLength` must be written under the exact key the underlying\n * provider reads — no adapter reads a generic `maxTokens`. Ollama is the one\n * exception: it nests sampling under `options`, so it has no entry here and is\n * handled as a special nested case in `applyMaxLength`/`applyDefaultTemperature`.\n *\n * Keep in sync with each adapter's wire mapping:\n * - OpenAI (Responses): `max_output_tokens`\n * - Anthropic / Grok: `max_tokens`\n * - Groq: `max_completion_tokens`\n * - Gemini: `maxOutputTokens`\n * - OpenRouter: `maxCompletionTokens`\n * - Ollama: nested `options.num_predict` (no entry — see `applyMaxLength`)\n */\nconst MAX_TOKENS_KEY_BY_ADAPTER: Record<string, string> = {\n  openai: 'max_output_tokens',\n  anthropic: 'max_tokens',\n  grok: 'max_tokens',\n  groq: 'max_completion_tokens',\n  gemini: 'maxOutputTokens',\n  openrouter: 'maxCompletionTokens',\n}\n\n/**\n * Every flat key any supported provider uses to cap output tokens (plus the\n * generic `maxTokens` spelling no adapter reads). Used to detect a\n * caller-supplied token limit so the summarize default never overrides an\n * explicit caller value. Shared with the OTel middleware via\n * `MAX_TOKENS_KEYS` so the two spelling sets cannot drift.\n */\nconst KNOWN_MAX_TOKENS_KEYS = MAX_TOKENS_KEYS\n\n/**\n * Whether `applyMaxLength` knows how to place a token limit for this adapter\n * `name` (either the nested Ollama shape or a flat provider-native key).\n * Used to surface a warning when `maxLength` would otherwise be silently\n * dropped for an unrecognised adapter name.\n */\nfunction isKnownMaxTokensAdapter(adapterName: string): boolean {\n  return (\n    adapterName === 'ollama' ||\n    MAX_TOKENS_KEY_BY_ADAPTER[adapterName] !== undefined\n  )\n}\n\n/**\n * Apply the low-temperature summarize default to a working copy of the\n * caller's `modelOptions`, placed where the wrapped provider actually reads\n * it (nested under `options` for Ollama, flat otherwise). The caller always\n * wins: if they already set `temperature` in that location, it is untouched.\n */\nfunction applyDefaultTemperature(\n  adapterName: string,\n  temperature: number,\n  modelOptions: Record<string, unknown>,\n): Record<string, unknown> {\n  const merged: Record<string, unknown> = { ...modelOptions }\n\n  if (adapterName === 'ollama') {\n    const existing =\n      merged.options && typeof merged.options === 'object'\n        ? (merged.options as Record<string, unknown>)\n        : undefined\n    if (existing && 'temperature' in existing) return merged\n    merged.options = { temperature, ...existing }\n    return merged\n  }\n\n  if ('temperature' in merged) return merged\n  merged.temperature = temperature\n  return merged\n}\n\n/**\n * Resolve `maxLength` to the provider-native max-output-tokens key for the\n * given summarize-adapter `name` (this wrapper's OWN `name`, not the wrapped\n * text adapter's) and merge it into a working copy of the caller's\n * `modelOptions`. The caller always wins: if they already set any recognised\n * token-limit key (flat or, for Ollama, nested `options.num_predict`), the\n * default is left untouched. Unknown/unrecognised adapter names fall back to\n * NOT setting a token key (the prompt hint still asks the model to stay under\n * `maxLength`) rather than writing a dead key no provider reads.\n *\n * Caveat (intentional): \"caller wins\" keys off ANY recognised spelling in\n * `KNOWN_MAX_TOKENS_KEYS`, but only the adapter's native key is read on the\n * wire. So a caller who sets a NON-native spelling for this provider — e.g.\n * `maxTokens`, or Anthropic's `max_tokens` against an OpenAI adapter — suppresses\n * the summarize default WITHOUT getting their own value applied either: neither\n * cap reaches the wire. This favours never clobbering a migration leftover over\n * guaranteeing a cap; the prompt-level hint still asks the model to stay under\n * `maxLength`. Rename the key to the provider-native spelling to forward it.\n */\nfunction applyMaxLength(\n  adapterName: string,\n  maxLength: number,\n  modelOptions: Record<string, unknown>,\n): Record<string, unknown> {\n  const merged: Record<string, unknown> = { ...modelOptions }\n\n  if (adapterName === 'ollama') {\n    // Honor a caller-set limit in either shape: a recognised flat key (e.g.\n    // left over from a migration) or the nested `options.num_predict`.\n    const callerSetFlatLimit = KNOWN_MAX_TOKENS_KEYS.some(\n      (k) => typeof merged[k] === 'number',\n    )\n    const existing =\n      merged.options && typeof merged.options === 'object'\n        ? (merged.options as Record<string, unknown>)\n        : undefined\n    if (\n      callerSetFlatLimit ||\n      (existing && typeof existing.num_predict === 'number')\n    ) {\n      return merged\n    }\n    merged.options = { num_predict: maxLength, ...existing }\n    return merged\n  }\n\n  const key = MAX_TOKENS_KEY_BY_ADAPTER[adapterName]\n  if (key === undefined) return merged\n\n  const callerSetLimit = KNOWN_MAX_TOKENS_KEYS.some(\n    (k) => typeof merged[k] === 'number',\n  )\n  if (callerSetLimit) return merged\n\n  merged[key] = maxLength\n  return merged\n}\n\n/**\n * Extract the per-model `modelOptions` type a text adapter accepts. Used by\n * provider summarize factories so their `modelOptions` IntelliSense matches\n * what the underlying text adapter actually understands.\n */\nexport type InferTextProviderOptions<TAdapter> = TAdapter extends {\n  '~types': { providerOptions: infer P }\n}\n  ? P extends object\n    ? P\n    : object\n  : object\n\n/**\n * Summarize adapter that wraps any `ChatStreamCapable` text adapter and\n * prompts it for summarization. Not tied to any wire format.\n */\nexport class ChatStreamSummarizeAdapter<\n  TModel extends string,\n  TProviderOptions extends object = Record<string, unknown>,\n> extends BaseSummarizeAdapter<TModel, TProviderOptions> {\n  readonly name: string\n\n  private readonly textAdapter: ChatStreamCapable\n\n  constructor(\n    textAdapter: ChatStreamCapable,\n    model: TModel,\n    name: string = 'chat-stream-summarize',\n  ) {\n    super({}, model)\n    this.name = name\n    this.textAdapter = textAdapter\n  }\n\n  async summarize(\n    options: SummarizationOptions<TProviderOptions>,\n  ): Promise<SummarizationResult> {\n    const systemPrompt = this.buildSummarizationPrompt(options)\n\n    let summary = ''\n    const id = this.generateId()\n    let model = options.model\n    let usage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 }\n\n    options.logger.request(\n      `activity=summarize provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? 'unset'}`,\n      { provider: this.name, model: options.model },\n    )\n\n    try {\n      for await (const chunk of this.textAdapter.chatStream(\n        this.buildTextOptions(options, systemPrompt),\n      )) {\n        if (chunk.type === 'TEXT_MESSAGE_CONTENT') {\n          if (chunk.content) {\n            summary = chunk.content\n          } else if (chunk.delta) {\n            // Append delta only when present — a content-less chunk with no\n            // delta would otherwise concat literal `'undefined'`.\n            summary += chunk.delta\n          }\n          model = chunk.model || model\n        }\n        if (chunk.type === 'RUN_FINISHED') {\n          if (chunk.usage) {\n            usage = chunk.usage\n          }\n        }\n        // Surface failures: the underlying chatStream emits RUN_ERROR instead\n        // of throwing, so without this branch summarize() would return an\n        // empty summary and pretend a failed run succeeded.\n        if (chunk.type === 'RUN_ERROR') {\n          const message =\n            (chunk.error && typeof chunk.error.message === 'string'\n              ? chunk.error.message\n              : null) ?? 'Summarization failed'\n          const code =\n            chunk.error && typeof chunk.error.code === 'string'\n              ? chunk.error.code\n              : undefined\n          const err = new Error(message)\n          if (code) {\n            ;(err as Error & { code?: string }).code = code\n          }\n          throw err\n        }\n      }\n    } catch (error: unknown) {\n      // Narrow before logging: raw SDK errors can carry request metadata\n      // (including auth headers) which we must never surface to user loggers.\n      options.logger.errors(`${this.name}.summarize fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.summarize failed`),\n        source: `${this.name}.summarize`,\n      })\n      throw error\n    }\n\n    return { id, model, summary, usage }\n  }\n\n  override async *summarizeStream(\n    options: SummarizationOptions<TProviderOptions>,\n  ): AsyncIterable<StreamChunk> {\n    const systemPrompt = this.buildSummarizationPrompt(options)\n\n    options.logger.request(\n      `activity=summarizeStream provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? 'unset'}`,\n      { provider: this.name, model: options.model },\n    )\n\n    const id = this.generateId()\n    let summary = ''\n    let model = options.model\n    let usage: SummarizationResult['usage'] = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n    }\n\n    try {\n      for await (const chunk of this.textAdapter.chatStream(\n        this.buildTextOptions(options, systemPrompt),\n      )) {\n        // Accumulate the same way `summarize()` does so consumers see deltas\n        // AND the terminal `generation:result` event below carries the same\n        // final summary that non-streaming returns.\n        if (chunk.type === 'TEXT_MESSAGE_CONTENT') {\n          if (chunk.content) {\n            summary = chunk.content\n          } else if (chunk.delta) {\n            summary += chunk.delta\n          }\n          if (chunk.model) model = chunk.model\n        }\n\n        // Emit the GenerationClient-shaped result event just before the\n        // terminal RUN_FINISHED so subscribers (useSummarize) populate\n        // `result` before flipping `status` to success.\n        if (chunk.type === 'RUN_FINISHED') {\n          if (chunk.usage) usage = chunk.usage\n          if (chunk.model) model = chunk.model\n          yield {\n            type: EventType.CUSTOM,\n            name: 'generation:result',\n            value: { id, model, summary, usage } satisfies SummarizationResult,\n            model,\n            timestamp: Date.now(),\n          }\n        }\n\n        yield chunk\n      }\n    } catch (error: unknown) {\n      options.logger.errors(`${this.name}.summarizeStream fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.summarizeStream failed`),\n        source: `${this.name}.summarizeStream`,\n      })\n      throw error\n    }\n  }\n\n  /**\n   * Build the TextOptions passed to the underlying chatStream. Provider\n   * `modelOptions` from the summarize call are forwarded as-is so knobs like\n   * Anthropic cache headers, Gemini safety settings, or Ollama tuning params\n   * still reach the wire layer.\n   */\n  protected buildTextOptions(\n    options: SummarizationOptions<TProviderOptions>,\n    systemPrompt: string,\n  ): TextOptions<TProviderOptions> {\n    // Sampling knobs now live in provider-native `modelOptions`. Apply the\n    // low-temperature default where the wrapped provider actually reads it\n    // (nested under `options` for Ollama, flat otherwise) so callers can still\n    // override it. Resolving the placement from this summarize adapter's OWN\n    // `name` keeps the default off the wire correctly per provider — a flat\n    // `temperature` would be silently dropped by Ollama while still showing up\n    // in OTel.\n    let working: Record<string, unknown> = {\n      ...(options.modelOptions as Record<string, unknown> | undefined),\n    }\n    working = applyDefaultTemperature(this.name, 0.3, working)\n    // `maxLength` must reach the wire under the provider-native token key (it\n    // differs per provider, and no adapter reads a generic `maxTokens`).\n    // Resolve it from this summarize adapter's `name` (the constructor arg,\n    // not the wrapped text adapter's name), never overriding a caller-supplied\n    // token limit.\n    if (options.maxLength !== undefined) {\n      if (!isKnownMaxTokensAdapter(this.name)) {\n        options.logger.warn(\n          `summarize: maxLength=${options.maxLength} could not be mapped to a provider token key for adapter name \"${this.name}\" — it was dropped from modelOptions (the prompt still asks the model to stay under it). Construct ChatStreamSummarizeAdapter with a recognised provider name to forward the cap.`,\n          { provider: this.name },\n        )\n      }\n      working = applyMaxLength(this.name, options.maxLength, working)\n    }\n    const modelOptions = working as TProviderOptions\n\n    return {\n      model: options.model,\n      messages: [{ role: 'user', content: options.text }],\n      systemPrompts: [systemPrompt],\n      modelOptions,\n      logger: options.logger,\n    }\n  }\n\n  protected buildSummarizationPrompt(\n    options: SummarizationOptions<TProviderOptions>,\n  ): string {\n    let prompt = 'You are a professional summarizer. '\n\n    switch (options.style) {\n      case 'bullet-points':\n        prompt += 'Provide a summary in bullet point format. '\n        break\n      case 'paragraph':\n        prompt += 'Provide a summary in paragraph format. '\n        break\n      case 'concise':\n        prompt += 'Provide a very concise summary in 1-2 sentences. '\n        break\n      case undefined:\n        prompt += 'Provide a clear and concise summary. '\n        break\n      default:\n        prompt += 'Provide a clear and concise summary. '\n    }\n\n    if (options.focus && options.focus.length > 0) {\n      prompt += `Focus on the following aspects: ${options.focus.join(', ')}. `\n    }\n\n    if (options.maxLength) {\n      prompt += `Keep the summary under ${options.maxLength} tokens. `\n    }\n\n    return prompt\n  }\n}\n"],"names":[],"mappings":";;;;AA0CA,MAAM,4BAAoD;AAAA,EACxD,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,MAAM;AAAA,EACN,MAAM;AAAA,EACN,QAAQ;AAAA,EACR,YAAY;AACd;AASA,MAAM,wBAAwB;AAQ9B,SAAS,wBAAwB,aAA8B;AAC7D,SACE,gBAAgB,YAChB,0BAA0B,WAAW,MAAM;AAE/C;AAQA,SAAS,wBACP,aACA,aACA,cACyB;AACzB,QAAM,SAAkC,EAAE,GAAG,aAAA;AAE7C,MAAI,gBAAgB,UAAU;AAC5B,UAAM,WACJ,OAAO,WAAW,OAAO,OAAO,YAAY,WACvC,OAAO,UACR;AACN,QAAI,YAAY,iBAAiB,SAAU,QAAO;AAClD,WAAO,UAAU,EAAE,aAAa,GAAG,SAAA;AACnC,WAAO;AAAA,EACT;AAEA,MAAI,iBAAiB,OAAQ,QAAO;AACpC,SAAO,cAAc;AACrB,SAAO;AACT;AAqBA,SAAS,eACP,aACA,WACA,cACyB;AACzB,QAAM,SAAkC,EAAE,GAAG,aAAA;AAE7C,MAAI,gBAAgB,UAAU;AAG5B,UAAM,qBAAqB,sBAAsB;AAAA,MAC/C,CAAC,MAAM,OAAO,OAAO,CAAC,MAAM;AAAA,IAAA;AAE9B,UAAM,WACJ,OAAO,WAAW,OAAO,OAAO,YAAY,WACvC,OAAO,UACR;AACN,QACE,sBACC,YAAY,OAAO,SAAS,gBAAgB,UAC7C;AACA,aAAO;AAAA,IACT;AACA,WAAO,UAAU,EAAE,aAAa,WAAW,GAAG,SAAA;AAC9C,WAAO;AAAA,EACT;AAEA,QAAM,MAAM,0BAA0B,WAAW;AACjD,MAAI,QAAQ,OAAW,QAAO;AAE9B,QAAM,iBAAiB,sBAAsB;AAAA,IAC3C,CAAC,MAAM,OAAO,OAAO,CAAC,MAAM;AAAA,EAAA;AAE9B,MAAI,eAAgB,QAAO;AAE3B,SAAO,GAAG,IAAI;AACd,SAAO;AACT;AAmBO,MAAM,mCAGH,qBAA+C;AAAA,EAC9C;AAAA,EAEQ;AAAA,EAEjB,YACE,aACA,OACA,OAAe,yBACf;AACA,UAAM,CAAA,GAAI,KAAK;AACf,SAAK,OAAO;AACZ,SAAK,cAAc;AAAA,EACrB;AAAA,EAEA,MAAM,UACJ,SAC8B;AAC9B,UAAM,eAAe,KAAK,yBAAyB,OAAO;AAE1D,QAAI,UAAU;AACd,UAAM,KAAK,KAAK,WAAA;AAChB,QAAI,QAAQ,QAAQ;AACpB,QAAI,QAAQ,EAAE,cAAc,GAAG,kBAAkB,GAAG,aAAa,EAAA;AAEjE,YAAQ,OAAO;AAAA,MACb,+BAA+B,KAAK,IAAI,UAAU,QAAQ,KAAK,gBAAgB,QAAQ,KAAK,MAAM,cAAc,QAAQ,aAAa,OAAO;AAAA,MAC5I,EAAE,UAAU,KAAK,MAAM,OAAO,QAAQ,MAAA;AAAA,IAAM;AAG9C,QAAI;AACF,uBAAiB,SAAS,KAAK,YAAY;AAAA,QACzC,KAAK,iBAAiB,SAAS,YAAY;AAAA,MAAA,GAC1C;AACD,YAAI,MAAM,SAAS,wBAAwB;AACzC,cAAI,MAAM,SAAS;AACjB,sBAAU,MAAM;AAAA,UAClB,WAAW,MAAM,OAAO;AAGtB,uBAAW,MAAM;AAAA,UACnB;AACA,kBAAQ,MAAM,SAAS;AAAA,QACzB;AACA,YAAI,MAAM,SAAS,gBAAgB;AACjC,cAAI,MAAM,OAAO;AACf,oBAAQ,MAAM;AAAA,UAChB;AAAA,QACF;AAIA,YAAI,MAAM,SAAS,aAAa;AAC9B,gBAAM,WACH,MAAM,SAAS,OAAO,MAAM,MAAM,YAAY,WAC3C,MAAM,MAAM,UACZ,SAAS;AACf,gBAAM,OACJ,MAAM,SAAS,OAAO,MAAM,MAAM,SAAS,WACvC,MAAM,MAAM,OACZ;AACN,gBAAM,MAAM,IAAI,MAAM,OAAO;AAC7B,cAAI,MAAM;AACR;AAAE,gBAAkC,OAAO;AAAA,UAC7C;AACA,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF,SAAS,OAAgB;AAGvB,cAAQ,OAAO,OAAO,GAAG,KAAK,IAAI,oBAAoB;AAAA,QACpD,OAAO,kBAAkB,OAAO,GAAG,KAAK,IAAI,mBAAmB;AAAA,QAC/D,QAAQ,GAAG,KAAK,IAAI;AAAA,MAAA,CACrB;AACD,YAAM;AAAA,IACR;AAEA,WAAO,EAAE,IAAI,OAAO,SAAS,MAAA;AAAA,EAC/B;AAAA,EAEA,OAAgB,gBACd,SAC4B;AAC5B,UAAM,eAAe,KAAK,yBAAyB,OAAO;AAE1D,YAAQ,OAAO;AAAA,MACb,qCAAqC,KAAK,IAAI,UAAU,QAAQ,KAAK,gBAAgB,QAAQ,KAAK,MAAM,cAAc,QAAQ,aAAa,OAAO;AAAA,MAClJ,EAAE,UAAU,KAAK,MAAM,OAAO,QAAQ,MAAA;AAAA,IAAM;AAG9C,UAAM,KAAK,KAAK,WAAA;AAChB,QAAI,UAAU;AACd,QAAI,QAAQ,QAAQ;AACpB,QAAI,QAAsC;AAAA,MACxC,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,IAAA;AAGf,QAAI;AACF,uBAAiB,SAAS,KAAK,YAAY;AAAA,QACzC,KAAK,iBAAiB,SAAS,YAAY;AAAA,MAAA,GAC1C;AAID,YAAI,MAAM,SAAS,wBAAwB;AACzC,cAAI,MAAM,SAAS;AACjB,sBAAU,MAAM;AAAA,UAClB,WAAW,MAAM,OAAO;AACtB,uBAAW,MAAM;AAAA,UACnB;AACA,cAAI,MAAM,MAAO,SAAQ,MAAM;AAAA,QACjC;AAKA,YAAI,MAAM,SAAS,gBAAgB;AACjC,cAAI,MAAM,MAAO,SAAQ,MAAM;AAC/B,cAAI,MAAM,MAAO,SAAQ,MAAM;AAC/B,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB,MAAM;AAAA,YACN,OAAO,EAAE,IAAI,OAAO,SAAS,MAAA;AAAA,YAC7B;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,UAAI;AAAA,QAExB;AAEA,cAAM;AAAA,MACR;AAAA,IACF,SAAS,OAAgB;AACvB,cAAQ,OAAO,OAAO,GAAG,KAAK,IAAI,0BAA0B;AAAA,QAC1D,OAAO,kBAAkB,OAAO,GAAG,KAAK,IAAI,yBAAyB;AAAA,QACrE,QAAQ,GAAG,KAAK,IAAI;AAAA,MAAA,CACrB;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQU,iBACR,SACA,cAC+B;AAQ/B,QAAI,UAAmC;AAAA,MACrC,GAAI,QAAQ;AAAA,IAAA;AAEd,cAAU,wBAAwB,KAAK,MAAM,KAAK,OAAO;AAMzD,QAAI,QAAQ,cAAc,QAAW;AACnC,UAAI,CAAC,wBAAwB,KAAK,IAAI,GAAG;AACvC,gBAAQ,OAAO;AAAA,UACb,wBAAwB,QAAQ,SAAS,kEAAkE,KAAK,IAAI;AAAA,UACpH,EAAE,UAAU,KAAK,KAAA;AAAA,QAAK;AAAA,MAE1B;AACA,gBAAU,eAAe,KAAK,MAAM,QAAQ,WAAW,OAAO;AAAA,IAChE;AACA,UAAM,eAAe;AAErB,WAAO;AAAA,MACL,OAAO,QAAQ;AAAA,MACf,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,QAAQ,MAAM;AAAA,MAClD,eAAe,CAAC,YAAY;AAAA,MAC5B;AAAA,MACA,QAAQ,QAAQ;AAAA,IAAA;AAAA,EAEpB;AAAA,EAEU,yBACR,SACQ;AACR,QAAI,SAAS;AAEb,YAAQ,QAAQ,OAAA;AAAA,MACd,KAAK;AACH,kBAAU;AACV;AAAA,MACF,KAAK;AACH,kBAAU;AACV;AAAA,MACF,KAAK;AACH,kBAAU;AACV;AAAA,MACF,KAAK;AACH,kBAAU;AACV;AAAA,MACF;AACE,kBAAU;AAAA,IAAA;AAGd,QAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAAG;AAC7C,gBAAU,mCAAmC,QAAQ,MAAM,KAAK,IAAI,CAAC;AAAA,IACvE;AAEA,QAAI,QAAQ,WAAW;AACrB,gBAAU,0BAA0B,QAAQ,SAAS;AAAA,IACvD;AAEA,WAAO;AAAA,EACT;AACF;"}
+{"version":3,"file":"chat-stream-summarize.js","names":[],"sources":["../../../../src/activities/summarize/chat-stream-summarize.ts"],"sourcesContent":["import { EventType } from '@ag-ui/core'\nimport { toRunErrorPayload } from '../error-payload'\nimport { MAX_TOKENS_KEYS } from '../../utilities/sampling-keys'\nimport { BaseSummarizeAdapter } from './adapter'\nimport type {\n  StreamChunk,\n  SummarizationOptions,\n  SummarizationResult,\n  TextOptions,\n} from '../../types'\n\n/**\n * Minimal contract for a text adapter that supports `chatStream`. Lets\n * `ChatStreamSummarizeAdapter` work with any text adapter without coupling\n * to a specific implementation.\n *\n * The provider-options shape is intentionally `any` here — the wrapper only\n * forwards `modelOptions` straight through, so a text adapter with a richer\n * per-model options type (e.g. `ResolveProviderOptions<TModel>`) is still\n * acceptable. Summarize-level type safety is enforced via\n * `SummarizationOptions<TProviderOptions>` on the wrapper itself.\n */\nexport interface ChatStreamCapable {\n  chatStream: (options: TextOptions<any>) => AsyncIterable<StreamChunk>\n}\n\n/**\n * Provider-native max-output-tokens key per summarize-adapter `name`. summarize\n * is provider-agnostic and forwards `modelOptions` opaquely to the wrapped text\n * adapter, so `maxLength` must be written under the exact key the underlying\n * provider reads — no adapter reads a generic `maxTokens`. Ollama is the one\n * exception: it nests sampling under `options`, so it has no entry here and is\n * handled as a special nested case in `applyMaxLength`/`applyDefaultTemperature`.\n *\n * Keep in sync with each adapter's wire mapping:\n * - OpenAI (Responses): `max_output_tokens`\n * - Anthropic / Grok: `max_tokens`\n * - Groq: `max_completion_tokens`\n * - Gemini: `maxOutputTokens`\n * - OpenRouter: `maxCompletionTokens`\n * - Ollama: nested `options.num_predict` (no entry — see `applyMaxLength`)\n */\nconst MAX_TOKENS_KEY_BY_ADAPTER: Record<string, string> = {\n  openai: 'max_output_tokens',\n  anthropic: 'max_tokens',\n  grok: 'max_tokens',\n  groq: 'max_completion_tokens',\n  gemini: 'maxOutputTokens',\n  openrouter: 'maxCompletionTokens',\n}\n\n/**\n * Every flat key any supported provider uses to cap output tokens (plus the\n * generic `maxTokens` spelling no adapter reads). Used to detect a\n * caller-supplied token limit so the summarize default never overrides an\n * explicit caller value. Shared with the OTel middleware via\n * `MAX_TOKENS_KEYS` so the two spelling sets cannot drift.\n */\nconst KNOWN_MAX_TOKENS_KEYS = MAX_TOKENS_KEYS\n\n/**\n * Whether `applyMaxLength` knows how to place a token limit for this adapter\n * `name` (either the nested Ollama shape or a flat provider-native key).\n * Used to surface a warning when `maxLength` would otherwise be silently\n * dropped for an unrecognised adapter name.\n */\nfunction isKnownMaxTokensAdapter(adapterName: string): boolean {\n  return (\n    adapterName === 'ollama' ||\n    MAX_TOKENS_KEY_BY_ADAPTER[adapterName] !== undefined\n  )\n}\n\n/**\n * Apply the low-temperature summarize default to a working copy of the\n * caller's `modelOptions`, placed where the wrapped provider actually reads\n * it (nested under `options` for Ollama, flat otherwise). The caller always\n * wins: if they already set `temperature` in that location, it is untouched.\n */\nfunction applyDefaultTemperature(\n  adapterName: string,\n  temperature: number,\n  modelOptions: Record<string, unknown>,\n): Record<string, unknown> {\n  const merged: Record<string, unknown> = { ...modelOptions }\n\n  if (adapterName === 'ollama') {\n    const existing =\n      merged.options && typeof merged.options === 'object'\n        ? (merged.options as Record<string, unknown>)\n        : undefined\n    if (existing && 'temperature' in existing) return merged\n    merged.options = { temperature, ...existing }\n    return merged\n  }\n\n  if ('temperature' in merged) return merged\n  merged.temperature = temperature\n  return merged\n}\n\n/**\n * Resolve `maxLength` to the provider-native max-output-tokens key for the\n * given summarize-adapter `name` (this wrapper's OWN `name`, not the wrapped\n * text adapter's) and merge it into a working copy of the caller's\n * `modelOptions`. The caller always wins: if they already set any recognised\n * token-limit key (flat or, for Ollama, nested `options.num_predict`), the\n * default is left untouched. Unknown/unrecognised adapter names fall back to\n * NOT setting a token key (the prompt hint still asks the model to stay under\n * `maxLength`) rather than writing a dead key no provider reads.\n *\n * Caveat (intentional): \"caller wins\" keys off ANY recognised spelling in\n * `KNOWN_MAX_TOKENS_KEYS`, but only the adapter's native key is read on the\n * wire. So a caller who sets a NON-native spelling for this provider — e.g.\n * `maxTokens`, or Anthropic's `max_tokens` against an OpenAI adapter — suppresses\n * the summarize default WITHOUT getting their own value applied either: neither\n * cap reaches the wire. This favours never clobbering a migration leftover over\n * guaranteeing a cap; the prompt-level hint still asks the model to stay under\n * `maxLength`. Rename the key to the provider-native spelling to forward it.\n */\nfunction applyMaxLength(\n  adapterName: string,\n  maxLength: number,\n  modelOptions: Record<string, unknown>,\n): Record<string, unknown> {\n  const merged: Record<string, unknown> = { ...modelOptions }\n\n  if (adapterName === 'ollama') {\n    // Honor a caller-set limit in either shape: a recognised flat key (e.g.\n    // left over from a migration) or the nested `options.num_predict`.\n    const callerSetFlatLimit = KNOWN_MAX_TOKENS_KEYS.some(\n      (k) => typeof merged[k] === 'number',\n    )\n    const existing =\n      merged.options && typeof merged.options === 'object'\n        ? (merged.options as Record<string, unknown>)\n        : undefined\n    if (\n      callerSetFlatLimit ||\n      (existing && typeof existing.num_predict === 'number')\n    ) {\n      return merged\n    }\n    merged.options = { num_predict: maxLength, ...existing }\n    return merged\n  }\n\n  const key = MAX_TOKENS_KEY_BY_ADAPTER[adapterName]\n  if (key === undefined) return merged\n\n  const callerSetLimit = KNOWN_MAX_TOKENS_KEYS.some(\n    (k) => typeof merged[k] === 'number',\n  )\n  if (callerSetLimit) return merged\n\n  merged[key] = maxLength\n  return merged\n}\n\n/**\n * Extract the per-model `modelOptions` type a text adapter accepts. Used by\n * provider summarize factories so their `modelOptions` IntelliSense matches\n * what the underlying text adapter actually understands.\n */\nexport type InferTextProviderOptions<TAdapter> = TAdapter extends {\n  '~types': { providerOptions: infer P }\n}\n  ? P extends object\n    ? P\n    : object\n  : object\n\n/**\n * Summarize adapter that wraps any `ChatStreamCapable` text adapter and\n * prompts it for summarization. Not tied to any wire format.\n */\nexport class ChatStreamSummarizeAdapter<\n  TModel extends string,\n  TProviderOptions extends object = Record<string, unknown>,\n> extends BaseSummarizeAdapter<TModel, TProviderOptions> {\n  readonly name: string\n\n  private readonly textAdapter: ChatStreamCapable\n\n  constructor(\n    textAdapter: ChatStreamCapable,\n    model: TModel,\n    name: string = 'chat-stream-summarize',\n  ) {\n    super({}, model)\n    this.name = name\n    this.textAdapter = textAdapter\n  }\n\n  async summarize(\n    options: SummarizationOptions<TProviderOptions>,\n  ): Promise<SummarizationResult> {\n    const systemPrompt = this.buildSummarizationPrompt(options)\n\n    let summary = ''\n    const id = this.generateId()\n    let model = options.model\n    let usage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 }\n\n    options.logger.request(\n      `activity=summarize provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? 'unset'}`,\n      { provider: this.name, model: options.model },\n    )\n\n    try {\n      for await (const chunk of this.textAdapter.chatStream(\n        this.buildTextOptions(options, systemPrompt),\n      )) {\n        if (chunk.type === 'TEXT_MESSAGE_CONTENT') {\n          if (chunk.content) {\n            summary = chunk.content\n          } else if (chunk.delta) {\n            // Append delta only when present — a content-less chunk with no\n            // delta would otherwise concat literal `'undefined'`.\n            summary += chunk.delta\n          }\n          model = chunk.model || model\n        }\n        if (chunk.type === 'RUN_FINISHED') {\n          if (chunk.usage) {\n            usage = chunk.usage\n          }\n        }\n        // Surface failures: the underlying chatStream emits RUN_ERROR instead\n        // of throwing, so without this branch summarize() would return an\n        // empty summary and pretend a failed run succeeded.\n        if (chunk.type === 'RUN_ERROR') {\n          const message =\n            (chunk.error && typeof chunk.error.message === 'string'\n              ? chunk.error.message\n              : null) ?? 'Summarization failed'\n          const code =\n            chunk.error && typeof chunk.error.code === 'string'\n              ? chunk.error.code\n              : undefined\n          const err = new Error(message)\n          if (code) {\n            ;(err as Error & { code?: string }).code = code\n          }\n          throw err\n        }\n      }\n    } catch (error: unknown) {\n      // Narrow before logging: raw SDK errors can carry request metadata\n      // (including auth headers) which we must never surface to user loggers.\n      options.logger.errors(`${this.name}.summarize fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.summarize failed`),\n        source: `${this.name}.summarize`,\n      })\n      throw error\n    }\n\n    return { id, model, summary, usage }\n  }\n\n  override async *summarizeStream(\n    options: SummarizationOptions<TProviderOptions>,\n  ): AsyncIterable<StreamChunk> {\n    const systemPrompt = this.buildSummarizationPrompt(options)\n\n    options.logger.request(\n      `activity=summarizeStream provider=${this.name} model=${options.model} text-length=${options.text.length} maxLength=${options.maxLength ?? 'unset'}`,\n      { provider: this.name, model: options.model },\n    )\n\n    const id = this.generateId()\n    let summary = ''\n    let model = options.model\n    let usage: SummarizationResult['usage'] = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n    }\n\n    try {\n      for await (const chunk of this.textAdapter.chatStream(\n        this.buildTextOptions(options, systemPrompt),\n      )) {\n        // Accumulate the same way `summarize()` does so consumers see deltas\n        // AND the terminal `generation:result` event below carries the same\n        // final summary that non-streaming returns.\n        if (chunk.type === 'TEXT_MESSAGE_CONTENT') {\n          if (chunk.content) {\n            summary = chunk.content\n          } else if (chunk.delta) {\n            summary += chunk.delta\n          }\n          if (chunk.model) model = chunk.model\n        }\n\n        // Emit the GenerationClient-shaped result event just before the\n        // terminal RUN_FINISHED so subscribers (useSummarize) populate\n        // `result` before flipping `status` to success.\n        if (chunk.type === 'RUN_FINISHED') {\n          if (chunk.usage) usage = chunk.usage\n          if (chunk.model) model = chunk.model\n          yield {\n            type: EventType.CUSTOM,\n            name: 'generation:result',\n            value: { id, model, summary, usage } satisfies SummarizationResult,\n            model,\n            timestamp: Date.now(),\n          }\n        }\n\n        yield chunk\n      }\n    } catch (error: unknown) {\n      options.logger.errors(`${this.name}.summarizeStream fatal`, {\n        error: toRunErrorPayload(error, `${this.name}.summarizeStream failed`),\n        source: `${this.name}.summarizeStream`,\n      })\n      throw error\n    }\n  }\n\n  /**\n   * Build the TextOptions passed to the underlying chatStream. Provider\n   * `modelOptions` from the summarize call are forwarded as-is so knobs like\n   * Anthropic cache headers, Gemini safety settings, or Ollama tuning params\n   * still reach the wire layer.\n   */\n  protected buildTextOptions(\n    options: SummarizationOptions<TProviderOptions>,\n    systemPrompt: string,\n  ): TextOptions<TProviderOptions> {\n    // Sampling knobs now live in provider-native `modelOptions`. Apply the\n    // low-temperature default where the wrapped provider actually reads it\n    // (nested under `options` for Ollama, flat otherwise) so callers can still\n    // override it. Resolving the placement from this summarize adapter's OWN\n    // `name` keeps the default off the wire correctly per provider — a flat\n    // `temperature` would be silently dropped by Ollama while still showing up\n    // in OTel.\n    let working: Record<string, unknown> = {\n      ...(options.modelOptions as Record<string, unknown> | undefined),\n    }\n    working = applyDefaultTemperature(this.name, 0.3, working)\n    // `maxLength` must reach the wire under the provider-native token key (it\n    // differs per provider, and no adapter reads a generic `maxTokens`).\n    // Resolve it from this summarize adapter's `name` (the constructor arg,\n    // not the wrapped text adapter's name), never overriding a caller-supplied\n    // token limit.\n    if (options.maxLength !== undefined) {\n      if (!isKnownMaxTokensAdapter(this.name)) {\n        options.logger.warn(\n          `summarize: maxLength=${options.maxLength} could not be mapped to a provider token key for adapter name \"${this.name}\" — it was dropped from modelOptions (the prompt still asks the model to stay under it). Construct ChatStreamSummarizeAdapter with a recognised provider name to forward the cap.`,\n          { provider: this.name },\n        )\n      }\n      working = applyMaxLength(this.name, options.maxLength, working)\n    }\n    const modelOptions = working as TProviderOptions\n\n    return {\n      model: options.model,\n      messages: [{ role: 'user', content: options.text }],\n      systemPrompts: [systemPrompt],\n      modelOptions,\n      logger: options.logger,\n      // Forward the run identity so the wrapped chat stamps it onto RUN_STARTED\n      // (chat uses `runId` as its `runIdOverride`). Conditional spreads keep the\n      // fields absent when unset, under `exactOptionalPropertyTypes`.\n      ...(options.runId !== undefined ? { runId: options.runId } : {}),\n      ...(options.threadId !== undefined ? { threadId: options.threadId } : {}),\n    }\n  }\n\n  protected buildSummarizationPrompt(\n    options: SummarizationOptions<TProviderOptions>,\n  ): string {\n    let prompt = 'You are a professional summarizer. '\n\n    switch (options.style) {\n      case 'bullet-points':\n        prompt += 'Provide a summary in bullet point format. '\n        break\n      case 'paragraph':\n        prompt += 'Provide a summary in paragraph format. '\n        break\n      case 'concise':\n        prompt += 'Provide a very concise summary in 1-2 sentences. '\n        break\n      case undefined:\n        prompt += 'Provide a clear and concise summary. '\n        break\n      default:\n        prompt += 'Provide a clear and concise summary. '\n    }\n\n    if (options.focus && options.focus.length > 0) {\n      prompt += `Focus on the following aspects: ${options.focus.join(', ')}. `\n    }\n\n    if (options.maxLength) {\n      prompt += `Keep the summary under ${options.maxLength} tokens. `\n    }\n\n    return prompt\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AA0CA,IAAM,4BAAoD;CACxD,QAAQ;CACR,WAAW;CACX,MAAM;CACN,MAAM;CACN,QAAQ;CACR,YAAY;AACd;;;;;;;;AASA,IAAM,wBAAwB;;;;;;;AAQ9B,SAAS,wBAAwB,aAA8B;CAC7D,OACE,gBAAgB,YAChB,0BAA0B,iBAAiB,KAAA;AAE/C;;;;;;;AAQA,SAAS,wBACP,aACA,aACA,cACyB;CACzB,MAAM,SAAkC,EAAE,GAAG,aAAa;CAE1D,IAAI,gBAAgB,UAAU;EAC5B,MAAM,WACJ,OAAO,WAAW,OAAO,OAAO,YAAY,WACvC,OAAO,UACR,KAAA;EACN,IAAI,YAAY,iBAAiB,UAAU,OAAO;EAClD,OAAO,UAAU;GAAE;GAAa,GAAG;EAAS;EAC5C,OAAO;CACT;CAEA,IAAI,iBAAiB,QAAQ,OAAO;CACpC,OAAO,cAAc;CACrB,OAAO;AACT;;;;;;;;;;;;;;;;;;;;AAqBA,SAAS,eACP,aACA,WACA,cACyB;CACzB,MAAM,SAAkC,EAAE,GAAG,aAAa;CAE1D,IAAI,gBAAgB,UAAU;EAG5B,MAAM,qBAAqB,sBAAsB,MAC9C,MAAM,OAAO,OAAO,OAAO,QAC9B;EACA,MAAM,WACJ,OAAO,WAAW,OAAO,OAAO,YAAY,WACvC,OAAO,UACR,KAAA;EACN,IACE,sBACC,YAAY,OAAO,SAAS,gBAAgB,UAE7C,OAAO;EAET,OAAO,UAAU;GAAE,aAAa;GAAW,GAAG;EAAS;EACvD,OAAO;CACT;CAEA,MAAM,MAAM,0BAA0B;CACtC,IAAI,QAAQ,KAAA,GAAW,OAAO;CAK9B,IAHuB,sBAAsB,MAC1C,MAAM,OAAO,OAAO,OAAO,QAE1B,GAAgB,OAAO;CAE3B,OAAO,OAAO;CACd,OAAO;AACT;;;;;AAmBA,IAAa,6BAAb,cAGU,qBAA+C;CACvD;CAEA;CAEA,YACE,aACA,OACA,OAAe,yBACf;EACA,MAAM,CAAC,GAAG,KAAK;EACf,KAAK,OAAO;EACZ,KAAK,cAAc;CACrB;CAEA,MAAM,UACJ,SAC8B;EAC9B,MAAM,eAAe,KAAK,yBAAyB,OAAO;EAE1D,IAAI,UAAU;EACd,MAAM,KAAK,KAAK,WAAW;EAC3B,IAAI,QAAQ,QAAQ;EACpB,IAAI,QAAQ;GAAE,cAAc;GAAG,kBAAkB;GAAG,aAAa;EAAE;EAEnE,QAAQ,OAAO,QACb,+BAA+B,KAAK,KAAK,SAAS,QAAQ,MAAM,eAAe,QAAQ,KAAK,OAAO,aAAa,QAAQ,aAAa,WACrI;GAAE,UAAU,KAAK;GAAM,OAAO,QAAQ;EAAM,CAC9C;EAEA,IAAI;GACF,WAAW,MAAM,SAAS,KAAK,YAAY,WACzC,KAAK,iBAAiB,SAAS,YAAY,CAC7C,GAAG;IACD,IAAI,MAAM,SAAS,wBAAwB;KACzC,IAAI,MAAM,SACR,UAAU,MAAM;UACX,IAAI,MAAM,OAGf,WAAW,MAAM;KAEnB,QAAQ,MAAM,SAAS;IACzB;IACA,IAAI,MAAM,SAAS;SACb,MAAM,OACR,QAAQ,MAAM;IAAA;IAMlB,IAAI,MAAM,SAAS,aAAa;KAC9B,MAAM,WACH,MAAM,SAAS,OAAO,MAAM,MAAM,YAAY,WAC3C,MAAM,MAAM,UACZ,SAAS;KACf,MAAM,OACJ,MAAM,SAAS,OAAO,MAAM,MAAM,SAAS,WACvC,MAAM,MAAM,OACZ,KAAA;KACN,MAAM,MAAM,IAAI,MAAM,OAAO;KAC7B,IAAI,MACD,IAAmC,OAAO;KAE7C,MAAM;IACR;GACF;EACF,SAAS,OAAgB;GAGvB,QAAQ,OAAO,OAAO,GAAG,KAAK,KAAK,mBAAmB;IACpD,OAAO,kBAAkB,OAAO,GAAG,KAAK,KAAK,kBAAkB;IAC/D,QAAQ,GAAG,KAAK,KAAK;GACvB,CAAC;GACD,MAAM;EACR;EAEA,OAAO;GAAE;GAAI;GAAO;GAAS;EAAM;CACrC;CAEA,OAAgB,gBACd,SAC4B;EAC5B,MAAM,eAAe,KAAK,yBAAyB,OAAO;EAE1D,QAAQ,OAAO,QACb,qCAAqC,KAAK,KAAK,SAAS,QAAQ,MAAM,eAAe,QAAQ,KAAK,OAAO,aAAa,QAAQ,aAAa,WAC3I;GAAE,UAAU,KAAK;GAAM,OAAO,QAAQ;EAAM,CAC9C;EAEA,MAAM,KAAK,KAAK,WAAW;EAC3B,IAAI,UAAU;EACd,IAAI,QAAQ,QAAQ;EACpB,IAAI,QAAsC;GACxC,cAAc;GACd,kBAAkB;GAClB,aAAa;EACf;EAEA,IAAI;GACF,WAAW,MAAM,SAAS,KAAK,YAAY,WACzC,KAAK,iBAAiB,SAAS,YAAY,CAC7C,GAAG;IAID,IAAI,MAAM,SAAS,wBAAwB;KACzC,IAAI,MAAM,SACR,UAAU,MAAM;UACX,IAAI,MAAM,OACf,WAAW,MAAM;KAEnB,IAAI,MAAM,OAAO,QAAQ,MAAM;IACjC;IAKA,IAAI,MAAM,SAAS,gBAAgB;KACjC,IAAI,MAAM,OAAO,QAAQ,MAAM;KAC/B,IAAI,MAAM,OAAO,QAAQ,MAAM;KAC/B,MAAM;MACJ,MAAM,UAAU;MAChB,MAAM;MACN,OAAO;OAAE;OAAI;OAAO;OAAS;MAAM;MACnC;MACA,WAAW,KAAK,IAAI;KACtB;IACF;IAEA,MAAM;GACR;EACF,SAAS,OAAgB;GACvB,QAAQ,OAAO,OAAO,GAAG,KAAK,KAAK,yBAAyB;IAC1D,OAAO,kBAAkB,OAAO,GAAG,KAAK,KAAK,wBAAwB;IACrE,QAAQ,GAAG,KAAK,KAAK;GACvB,CAAC;GACD,MAAM;EACR;CACF;;;;;;;CAQA,iBACE,SACA,cAC+B;EAQ/B,IAAI,UAAmC,EACrC,GAAI,QAAQ,aACd;EACA,UAAU,wBAAwB,KAAK,MAAM,IAAK,OAAO;EAMzD,IAAI,QAAQ,cAAc,KAAA,GAAW;GACnC,IAAI,CAAC,wBAAwB,KAAK,IAAI,GACpC,QAAQ,OAAO,KACb,wBAAwB,QAAQ,UAAU,iEAAiE,KAAK,KAAK,oLACrH,EAAE,UAAU,KAAK,KAAK,CACxB;GAEF,UAAU,eAAe,KAAK,MAAM,QAAQ,WAAW,OAAO;EAChE;EACA,MAAM,eAAe;EAErB,OAAO;GACL,OAAO,QAAQ;GACf,UAAU,CAAC;IAAE,MAAM;IAAQ,SAAS,QAAQ;GAAK,CAAC;GAClD,eAAe,CAAC,YAAY;GAC5B;GACA,QAAQ,QAAQ;GAIhB,GAAI,QAAQ,UAAU,KAAA,IAAY,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;GAC9D,GAAI,QAAQ,aAAa,KAAA,IAAY,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;EACzE;CACF;CAEA,yBACE,SACQ;EACR,IAAI,SAAS;EAEb,QAAQ,QAAQ,OAAhB;GACE,KAAK;IACH,UAAU;IACV;GACF,KAAK;IACH,UAAU;IACV;GACF,KAAK;IACH,UAAU;IACV;GACF,KAAK,KAAA;IACH,UAAU;IACV;GACF,SACE,UAAU;EACd;EAEA,IAAI,QAAQ,SAAS,QAAQ,MAAM,SAAS,GAC1C,UAAU,mCAAmC,QAAQ,MAAM,KAAK,IAAI,EAAE;EAGxE,IAAI,QAAQ,WACV,UAAU,0BAA0B,QAAQ,UAAU;EAGxD,OAAO;CACT;AACF"}

package/dist/esm/activities/summarize/index.d.ts

@@ -1,4 +1,5 @@
 import { DebugOption } from '../../logger/types.js';
+import { GenerationMiddleware } from '../middleware/types.js';
 import { SummarizeAdapter } from './adapter.js';
 import { StreamChunk, SummarizationResult } from '../../types.js';
 /** The adapter kind this activity handles */
@@ -27,6 +28,32 @@ export interface SummarizeActivityOptions<TAdapter extends SummarizeAdapter<stri
     focus?: Array<string>;
     /** Provider-specific options */
     modelOptions?: SummarizeProviderOptions<TAdapter>;
+    /**
+     * Optional run identity. When set on a streaming summarize, it is stamped
+     * onto the emitted `RUN_STARTED` so a delivery-durable route keys the run's
+     * log by the same id the client rejoins with — making a mid-run reload
+     * resumable. Filed under `threadId` when persistence is wired.
+     */
+    runId?: string;
+    /**
+     * Stable conversation/thread id for correlating this run when persisted — the
+     * slot a reloading client hydrates the last summary by. Pass it whenever
+     * persistence is on; `withGenerationPersistence` refuses a run without one.
+     */
+    threadId?: string;
+    /**
+     * Observe-only middleware notified on start, usage, success, and error. Pass
+     * `otelMiddleware()` for OpenTelemetry, `withGenerationPersistence()` to
+     * record the run (summaries are text, so the run record holds the result and
+     * there are no artifacts to store), or implement the `GenerationMiddleware`
+     * contract for a custom backend.
+     *
+     * Streaming and non-streaming behave the same way: one `onStart`, then a
+     * terminal `onFinish` / `onError`, with the result transforms applied to the
+     * `SummarizationResult` in between. A streaming consumer that disconnects
+     * mid-summary fires `onAbort`.
+     */
+    middleware?: Array<GenerationMiddleware>;
     /**
      * Whether to stream the summarization result.
      * When true, returns an AsyncIterable<StreamChunk> for streaming output.