@dudko.dev/agent-web 0.0.21 → 0.0.23

package/dist/chunk-YA6TQQ3Q.js

@@ -0,0 +1,425 @@
+// src/providers/webllm.ts
+import { generateText as generateText2 } from "ai";
+
+// src/context-window.ts
+import { asSchema } from "ai";
+
+// src/images.ts
+var DATA_URL_RE = /^data:([^;,]+)?;base64,(.*)$/s;
+var HTTP_RE = /^https?:\/\//i;
+var mediaTypeOf = (f) => {
+  if (f.mediaType) return f.mediaType;
+  if (typeof f.data === "string") {
+    const m = DATA_URL_RE.exec(f.data);
+    if (m?.[1]) return m[1];
+  }
+  const url = f.data instanceof URL ? f.data.href : typeof f.data === "string" ? f.data : "";
+  if (HTTP_RE.test(url) || f.name) {
+    const ext = (f.name ?? new URL(url, "http://x").pathname).split(".").pop()?.toLowerCase();
+    if (ext === "pdf") return "application/pdf";
+    if (ext && ["png", "jpg", "jpeg", "gif", "webp"].includes(ext)) {
+      return `image/${ext === "jpg" ? "jpeg" : ext}`;
+    }
+  }
+  return void 0;
+};
+var attachmentKind = (f) => {
+  const mt = mediaTypeOf(f) ?? "";
+  if (mt.startsWith("image/")) return "image";
+  if (mt === "application/pdf") return "pdf";
+  return "file";
+};
+var KIND_LABEL = {
+  image: "images",
+  pdf: "PDF files",
+  file: "file attachments"
+};
+var KIND_HINT = {
+  image: "Remove the image, or switch to a vision-capable model such as Gemini, Claude or GPT-4o/GPT-5.",
+  pdf: "Convert the PDF to text first (e.g. PDF \u2192 Markdown), or switch to a model that reads PDFs such as Gemini, Claude or GPT-4o/GPT-5.",
+  file: "Paste its text instead, or switch to a model that accepts this file type."
+};
+var AttachmentsNotSupportedError = class extends Error {
+  kind;
+  constructor(model, kind, detail) {
+    super(
+      `The model "${model}" can't take ${KIND_LABEL[kind]}${detail ? ` (${detail})` : ""}. ${KIND_HINT[kind]}`
+    );
+    this.name = "AttachmentsNotSupportedError";
+    this.kind = kind;
+  }
+  toJSON() {
+    return this.message;
+  }
+};
+var ImagesNotSupportedError = class extends AttachmentsNotSupportedError {
+  constructor(model, detail) {
+    super(model, "image", detail);
+    this.name = "ImagesNotSupportedError";
+  }
+};
+var unsupportedAttachment = (model, kind, detail) => kind === "image" ? new ImagesNotSupportedError(model, detail) : new AttachmentsNotSupportedError(model, kind, detail);
+var dataOf = (f) => {
+  if (typeof f.data === "string") {
+    const m = DATA_URL_RE.exec(f.data);
+    if (m) return m[2];
+    if (HTTP_RE.test(f.data)) return new URL(f.data);
+  }
+  return f.data;
+};
+var toFilePart = (f) => {
+  const kind = attachmentKind(f);
+  const mediaType = mediaTypeOf(f) ?? (kind === "image" ? "image" : "application/octet-stream");
+  return {
+    type: "file",
+    data: dataOf(f),
+    mediaType,
+    ...f.name ? { filename: f.name } : {}
+  };
+};
+var toImagePart = toFilePart;
+var modelLabel = (model) => {
+  if (typeof model === "string") return model;
+  const m = model;
+  return m?.modelId ? `${m.modelId}${m.provider ? ` (${m.provider})` : ""}` : "this model";
+};
+var REFUSAL_RE = /image|vision|multimodal|multi-modal|media type|mediatype|pdf|document|unsupported (file|content|part|mime)|does not support (file|content)|invalid content type|image_url|inline_?data|file_?data/i;
+var isAttachmentRefusal = (err) => REFUSAL_RE.test(err instanceof Error ? `${err.name} ${err.message}` : String(err));
+var isImageRefusal = isAttachmentRefusal;
+
+// src/llm/generate.ts
+import {
+  generateObject,
+  generateText,
+  streamText
+} from "ai";
+
+// src/llm/util.ts
+var clip = (s, max) => s.length > max ? `${s.slice(0, max - 1)}\u2026` : s;
+var promptOf = (opts) => opts.messages ? { messages: opts.messages } : { prompt: opts.prompt ?? "" };
+var normalizeUsage = (u) => {
+  const inputTokens = u?.inputTokens ?? 0;
+  const outputTokens = u?.outputTokens ?? 0;
+  return {
+    inputTokens,
+    outputTokens,
+    totalTokens: u?.totalTokens ?? inputTokens + outputTokens,
+    reasoningTokens: u?.outputTokenDetails?.reasoningTokens ?? u?.reasoningTokens ?? 0,
+    cachedInputTokens: u?.inputTokenDetails?.cacheReadTokens ?? u?.cachedInputTokens ?? 0,
+    cacheWriteTokens: u?.inputTokenDetails?.cacheWriteTokens ?? 0
+  };
+};
+var timeoutSignal = (signal, ms) => {
+  if (!ms || ms <= 0) return signal;
+  const timeout = AbortSignal.timeout(ms);
+  if (!signal) return timeout;
+  return typeof AbortSignal.any === "function" ? AbortSignal.any([signal, timeout]) : signal;
+};
+
+// src/llm/generate.ts
+var callSettings = (opts) => ({
+  ...opts.system !== void 0 ? { instructions: opts.system } : {},
+  maxOutputTokens: opts.maxOutputTokens,
+  temperature: opts.temperature,
+  ...opts.reasoning ? { reasoning: opts.reasoning } : {},
+  ...opts.providerOptions ? { providerOptions: opts.providerOptions } : {},
+  abortSignal: timeoutSignal(opts.abortSignal, opts.timeoutMs)
+});
+var generate = (model, opts = {}) => generateText({
+  model,
+  ...callSettings(opts),
+  ...promptOf(opts),
+  tools: opts.tools
+});
+var stream = (model, opts = {}) => streamText({
+  model,
+  ...callSettings(opts),
+  ...promptOf(opts),
+  tools: opts.tools
+});
+var generateStructured = (model, schema, opts = {}) => generateObject({
+  model,
+  schema,
+  ...callSettings(opts),
+  ...promptOf(opts)
+});
+
+// src/memory/compress.ts
+var resolveCompaction = (c) => {
+  const contextWindowTokens = c?.contextWindowTokens ?? 128e3;
+  return {
+    auto: c?.auto ?? true,
+    contextWindowTokens,
+    thresholdTokens: c?.thresholdTokens ?? Math.floor(contextWindowTokens / 2),
+    keepRecentTurns: c?.keepRecentTurns ?? 4,
+    keepRecentSteps: c?.keepRecentSteps ?? 3,
+    summaryMaxTokens: c?.summaryMaxTokens ?? 1024,
+    maxToolOutputChars: c?.maxToolOutputChars ?? 2e4,
+    clearToolResultsAfterTokens: c?.clearToolResultsAfterTokens ?? Math.floor(contextWindowTokens / 4),
+    keepToolResults: c?.keepToolResults ?? 3
+  };
+};
+var estimateTokens = (text) => Math.ceil((text ?? "").length / 4);
+var messageTokens = (msgs) => msgs.reduce((n, m) => n + estimateTokens(`${m.role}: ${m.content}
+`), 0);
+var SUMMARY_SYSTEM = "Summarize the following conversation into a compact set of durable facts, decisions, identifiers, open tasks and user preferences. Keep names, numbers and IDs verbatim. Output plain text only.";
+var compressHistory = async (messages, model, opts = {}) => {
+  const keepRecent = opts.keepRecent ?? 6;
+  if (messages.length <= keepRecent + 1) return messages;
+  if (!opts.force) {
+    if (opts.thresholdTokens !== void 0) {
+      if (opts.thresholdTokens <= 0 || messageTokens(messages) <= opts.thresholdTokens) {
+        return messages;
+      }
+    } else {
+      const maxChars = opts.maxChars ?? 12e3;
+      const chars = messages.reduce((n, m) => n + m.content.length, 0);
+      if (maxChars <= 0 || chars <= maxChars) return messages;
+    }
+  }
+  const head = messages.slice(0, messages.length - keepRecent);
+  const tail = messages.slice(messages.length - keepRecent);
+  const transcript = head.map((m) => `${m.role}: ${m.content}`).join("\n").slice(-48e3);
+  try {
+    const result = await generate(model, {
+      system: SUMMARY_SYSTEM,
+      prompt: transcript,
+      maxOutputTokens: opts.summaryMaxTokens ?? 400,
+      timeoutMs: opts.timeoutMs,
+      abortSignal: opts.abortSignal
+    });
+    opts.onUsage?.(normalizeUsage(result.usage));
+    const clean = (result.text || "").trim();
+    if (!clean) return messages;
+    const summaryMsg = {
+      role: "system",
+      content: `Summary of earlier conversation:
+${clean}`,
+      ts: head[0]?.ts
+    };
+    return [summaryMsg, ...tail];
+  } catch {
+    return messages;
+  }
+};
+var compactHistory = compressHistory;
+var TRACE_SYSTEM = "Summarize these completed agent steps into a compact progress note: what was done, what was found (keep names, numbers and IDs verbatim) and what failed. Plain text only.";
+var compactSteps = async (done, model, opts) => {
+  const keepRecent = opts.keepRecent ?? 3;
+  if (done.length <= keepRecent + 1) return done;
+  if (!opts.force && estimateTokens(done.join("\n")) <= opts.thresholdTokens) return done;
+  const head = done.slice(0, done.length - keepRecent);
+  try {
+    const result = await generate(model, {
+      system: TRACE_SYSTEM,
+      prompt: head.map((d, i) => `${i + 1}. ${d}`).join("\n"),
+      maxOutputTokens: opts.summaryMaxTokens ?? 600,
+      timeoutMs: opts.timeoutMs,
+      abortSignal: opts.abortSignal
+    });
+    opts.onUsage?.(normalizeUsage(result.usage));
+    const clean = (result.text || "").trim();
+    if (!clean) return done;
+    return [`Summary of ${head.length} earlier step(s): ${clean}`, ...done.slice(head.length)];
+  } catch {
+    return done;
+  }
+};
+
+// src/context-window.ts
+var WEBLLM_DEFAULT_CONTEXT_WINDOW = 4096;
+var withWebLLMContextWindow = (appConfig, modelId, tokens) => ({
+  ...appConfig,
+  model_list: appConfig.model_list.map(
+    (r) => r.model_id === modelId ? { ...r, overrides: { ...r.overrides, context_window_size: tokens } } : r
+  )
+});
+var webLLMContextWindow = (model) => {
+  const m = model;
+  if (!m || typeof m !== "object" || m.provider !== "web-llm") return void 0;
+  const id = String(m.modelId ?? "");
+  try {
+    const live = m.engine?.loadedModelIdToPipeline?.get?.(id)?.contextWindowSize;
+    if (typeof live === "number" && live > 0) return live;
+  } catch {
+  }
+  const record = m.config?.options?.engineConfig?.appConfig?.model_list?.find(
+    (r) => r.model_id === id
+  );
+  const configured = record?.overrides?.context_window_size;
+  if (typeof configured === "number" && configured > 0) return configured;
+  return /-1k$/i.test(id) ? 1024 : WEBLLM_DEFAULT_CONTEXT_WINDOW;
+};
+var contextWindowOf = (model) => {
+  const webllm = webLLMContextWindow(model);
+  if (webllm !== void 0) return webllm;
+  const m = model;
+  if (m && typeof m === "object" && m.provider === "browser-ai") {
+    try {
+      const w = m.getContextWindow?.();
+      return typeof w === "number" && w > 0 ? w : void 0;
+    } catch {
+      return void 0;
+    }
+  }
+  return void 0;
+};
+var fitCompactionToWindow = (resolved, configured, modelWindow) => {
+  const window = Math.min(
+    configured?.contextWindowTokens ?? modelWindow ?? resolved.contextWindowTokens,
+    modelWindow ?? Infinity
+  );
+  if (window === resolved.contextWindowTokens && modelWindow === void 0) return resolved;
+  const half = Math.floor(window / 2);
+  const quarter = Math.floor(window / 4);
+  const clearAfter = configured?.clearToolResultsAfterTokens;
+  return {
+    ...resolved,
+    contextWindowTokens: window,
+    thresholdTokens: Math.min(configured?.thresholdTokens ?? half, half),
+    // 0 = never stays never.
+    clearToolResultsAfterTokens: clearAfter === 0 ? 0 : Math.min(clearAfter ?? quarter, quarter),
+    // One result may take a quarter of the window (~4 chars per token).
+    maxToolOutputChars: configured?.maxToolOutputChars === 0 ? 0 : Math.min(configured?.maxToolOutputChars ?? 2e4, window),
+    // A small window holds fewer verbatim results.
+    keepToolResults: configured?.keepToolResults ?? Math.max(1, Math.min(3, Math.floor(window / 8192)))
+  };
+};
+var toolDefinitionTokens = async (tools) => {
+  const sizes = await Promise.all(
+    Object.entries(tools).map(async ([name, t]) => {
+      let parameters;
+      try {
+        parameters = await asSchema(t.inputSchema).jsonSchema;
+      } catch {
+        parameters = void 0;
+      }
+      return estimateTokens(JSON.stringify({ name, description: t.description ?? "", parameters }));
+    })
+  );
+  return sizes.reduce((a, b) => a + b, 0);
+};
+var TOOL_SEARCH_WINDOW_SHARE = 0.25;
+var ContextWindowExceededError = class extends Error {
+  /** The prompt's size as the provider counted it, when it said. */
+  promptTokens;
+  /** The model's window as the provider reported it, when it said. */
+  contextWindowTokens;
+  constructor(model, promptTokens, contextWindowTokens) {
+    const sizes = promptTokens && contextWindowTokens ? ` (${promptTokens} tokens; the window is ${contextWindowTokens})` : "";
+    super(
+      `The conversation no longer fits the context window of "${model}"${sizes}. Start a new chat or compact this one, connect fewer tools, or switch to a model with a larger window.`
+    );
+    this.name = "ContextWindowExceededError";
+    this.promptTokens = promptTokens;
+    this.contextWindowTokens = contextWindowTokens;
+  }
+  toJSON() {
+    return this.message;
+  }
+};
+var OVERFLOW_RE = /prompt tokens exceed context window|context window size|context_length_exceeded|maximum context length|prompt is too long|input token count.*exceeds|exceeds the maximum number of tokens|exceeds the available context size|context length exceeded|too many tokens in (the )?(prompt|input)/i;
+var WEBLLM_SIZES_RE = /number of prompt tokens:\s*(\d+);\s*context window size:\s*(\d+)/i;
+var contextOverflowOf = (err, model) => {
+  if (err instanceof ContextWindowExceededError) return err;
+  const text = err instanceof Error ? `${err.name} ${err.message}` : String(err);
+  if (!OVERFLOW_RE.test(text)) return void 0;
+  const sizes = WEBLLM_SIZES_RE.exec(text);
+  return new ContextWindowExceededError(
+    modelLabel(model),
+    sizes ? Number(sizes[1]) : void 0,
+    sizes ? Number(sizes[2]) : webLLMContextWindow(model) ?? void 0
+  );
+};
+
+// src/providers/webllm.ts
+var isWebGPUAvailable = () => typeof navigator !== "undefined" && "gpu" in navigator;
+var preloadWebLLMModel = async (model) => {
+  await generateText2({ model, prompt: "ok", maxOutputTokens: 1 });
+};
+var unloadWebLLMModel = async (model) => {
+  const engine = model.engine;
+  try {
+    await engine?.unload?.();
+  } catch {
+  }
+};
+var createWebLLMModel = async (model, options) => {
+  let mod;
+  try {
+    mod = await import("@browser-ai/web-llm");
+  } catch {
+    throw new Error(
+      'Local models require "@browser-ai/web-llm" and its peer "@mlc-ai/web-llm". Install: npm install @browser-ai/web-llm @mlc-ai/web-llm'
+    );
+  }
+  const { preload = true, contextWindowTokens, ...settings } = options ?? {};
+  if (contextWindowTokens) {
+    const engineConfig = settings.engineConfig ?? {};
+    let base = engineConfig.appConfig;
+    if (!base) {
+      try {
+        const spec = "@mlc-ai/web-llm";
+        base = (await import(
+          /* webpackIgnore: true */
+          /* @vite-ignore */
+          spec
+        )).prebuiltAppConfig;
+      } catch {
+        throw new Error(
+          'contextWindowTokens needs "@mlc-ai/web-llm" (for its prebuilt model list), or pass engineConfig.appConfig yourself.'
+        );
+      }
+    }
+    settings.engineConfig = {
+      ...engineConfig,
+      appConfig: withWebLLMContextWindow(base, model, contextWindowTokens)
+    };
+  }
+  const languageModel = mod.webLLM(
+    model,
+    settings
+  );
+  if (preload) await preloadWebLLMModel(languageModel);
+  return languageModel;
+};
+
+export {
+  clip,
+  promptOf,
+  normalizeUsage,
+  timeoutSignal,
+  generate,
+  stream,
+  generateStructured,
+  resolveCompaction,
+  estimateTokens,
+  messageTokens,
+  compressHistory,
+  compactHistory,
+  compactSteps,
+  mediaTypeOf,
+  attachmentKind,
+  AttachmentsNotSupportedError,
+  ImagesNotSupportedError,
+  unsupportedAttachment,
+  toFilePart,
+  toImagePart,
+  modelLabel,
+  isAttachmentRefusal,
+  isImageRefusal,
+  WEBLLM_DEFAULT_CONTEXT_WINDOW,
+  withWebLLMContextWindow,
+  webLLMContextWindow,
+  contextWindowOf,
+  fitCompactionToWindow,
+  toolDefinitionTokens,
+  TOOL_SEARCH_WINDOW_SHARE,
+  ContextWindowExceededError,
+  contextOverflowOf,
+  isWebGPUAvailable,
+  preloadWebLLMModel,
+  unloadWebLLMModel,
+  createWebLLMModel
+};
+//# sourceMappingURL=chunk-YA6TQQ3Q.js.map

package/dist/chunk-YA6TQQ3Q.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/providers/webllm.ts","../src/context-window.ts","../src/images.ts","../src/llm/generate.ts","../src/llm/util.ts","../src/memory/compress.ts"],"sourcesContent":["import { generateText, type LanguageModel } from 'ai'\nimport { withWebLLMContextWindow, type WebLLMAppConfigLike } from '../context-window.js'\n\n/** WebGPU feature-detect — WebLLM requires it. Call before offering local models. */\nexport const isWebGPUAvailable = (): boolean =>\n  typeof navigator !== 'undefined' && 'gpu' in navigator\n\nexport interface WebLLMModelOptions {\n  /**\n   * Progress callback for weight download / engine init. `report.progress` is\n   * 0..1. Forwarded to WebLLM's `initProgressCallback`.\n   */\n  initProgressCallback?: (report: { progress: number; text: string }) => void\n  /**\n   * Download the weights and initialize the engine NOW (default true).\n   *\n   * The underlying provider initializes lazily on the first generation — with\n   * a multi-GB model that means the download would silently happen on the\n   * user's first message, with `initProgressCallback` never firing at a moment\n   * the host can show it. Preloading runs a 1-token warm-up so the returned\n   * model is truly ready and progress is reported during creation. Re-creating\n   * an already-downloaded model is fast (weights live in the browser cache).\n   *\n   * Set to `false` to keep the lazy behavior.\n   */\n  preload?: boolean\n  /**\n   * The context window to load the model with, in tokens. WebLLM loads its\n   * prebuilt models with 4096 (to save VRAM), far below what Qwen3 or Llama\n   * 3.x were trained on; a larger window costs KV-cache memory, not a new\n   * download. Applied through `engineConfig.appConfig` (WebLLM's prebuilt\n   * config with this model's `context_window_size` overridden) unless you pass\n   * an `appConfig` of your own; see `withWebLLMContextWindow`.\n   */\n  contextWindowTokens?: number\n  /** Any other @browser-ai/web-llm setting (temperature, worker handler, …). */\n  [k: string]: unknown\n}\n\n/**\n * Force a model to download its weights and initialize its engine by running\n * a 1-token generation. Exposed for hosts that create models with\n * `preload: false` and want to warm them up later (e.g. behind their own UI).\n */\nexport const preloadWebLLMModel = async (model: LanguageModel): Promise<void> => {\n  await generateText({ model, prompt: 'ok', maxOutputTokens: 1 })\n}\n\n/**\n * Best-effort release of a model's GPU memory (call when switching models —\n * an abandoned engine otherwise holds its buffers until the page reloads).\n *\n * `@browser-ai/web-llm` keeps its MLCEngine private and exposes no public\n * `unload()`, so this reaches into the instance defensively; it is a no-op on\n * an uninitialized model or if the provider's internals change shape.\n */\nexport const unloadWebLLMModel = async (model: LanguageModel): Promise<void> => {\n  const engine = (model as { engine?: { unload?: () => Promise<void> } }).engine\n  try {\n    await engine?.unload?.()\n  } catch {\n    /* best-effort */\n  }\n}\n\n/**\n * Build an AI SDK `LanguageModel` backed by WebLLM (WebGPU) via the community\n * provider `@browser-ai/web-llm` (an optional peer, dynamically imported so it\n * is code-split and never loads until a local model is actually used). Accepts\n * any WebLLM prebuilt model id, e.g. \"Llama-3.2-3B-Instruct-q4f16_1-MLC\".\n *\n * By default the weights are downloaded eagerly with progress (see\n * `WebLLMModelOptions.preload`).\n */\nexport const createWebLLMModel = async (\n  model: string,\n  options?: WebLLMModelOptions,\n): Promise<LanguageModel> => {\n  let mod: typeof import('@browser-ai/web-llm')\n  try {\n    mod = await import('@browser-ai/web-llm')\n  } catch {\n    throw new Error(\n      'Local models require \"@browser-ai/web-llm\" and its peer \"@mlc-ai/web-llm\". Install: npm install @browser-ai/web-llm @mlc-ai/web-llm',\n    )\n  }\n  const { preload = true, contextWindowTokens, ...settings } = options ?? {}\n  if (contextWindowTokens) {\n    const engineConfig = (settings.engineConfig ?? {}) as { appConfig?: WebLLMAppConfigLike }\n    let base = engineConfig.appConfig\n    if (!base) {\n      try {\n        const spec = '@mlc-ai/web-llm'\n        base = (\n          (await import(/* webpackIgnore: true */ /* @vite-ignore */ spec)) as {\n            prebuiltAppConfig: WebLLMAppConfigLike\n          }\n        ).prebuiltAppConfig\n      } catch {\n        throw new Error(\n          'contextWindowTokens needs \"@mlc-ai/web-llm\" (for its prebuilt model list), or pass engineConfig.appConfig yourself.',\n        )\n      }\n    }\n    settings.engineConfig = {\n      ...engineConfig,\n      appConfig: withWebLLMContextWindow(base, model, contextWindowTokens),\n    }\n  }\n  // webLLM(modelId, settings) returns a LanguageModel. The model id is a strict\n  // union in the provider's types; we accept any string and cast at the seam.\n  const languageModel = mod.webLLM(\n    model as Parameters<typeof mod.webLLM>[0],\n    settings as Parameters<typeof mod.webLLM>[1],\n  )\n  if (preload) await preloadWebLLMModel(languageModel)\n  return languageModel\n}\n","import { asSchema, type ToolSet } from 'ai'\nimport { modelLabel } from './images.js'\nimport {\n  estimateTokens,\n  type CompactionConfig,\n  type ResolvedCompaction,\n} from './memory/compress.js'\n\n/**\n * The model's context window is a hard limit: a prompt that does not fit is\n * refused outright (WebLLM: \"Prompt tokens exceed context window size\"). The\n * agent sizes everything that grows — compaction thresholds, tool-result\n * clearing, the cap on one tool result, whether tools are listed or searched —\n * from the window, so it has to be the model's real one.\n *\n * Cloud windows are large and the host states them (`compaction.\n * contextWindowTokens`); a local WebLLM model's window is small (4096 tokens\n * by default — WebLLM's choice to save VRAM, not the model's limit) and is read\n * from the model itself.\n */\n\ninterface WebLLMRecordLike {\n  model_id: string\n  overrides?: { context_window_size?: number } & Record<string, unknown>\n}\n\n/** The shape of a WebLLM `AppConfig` this module touches. */\nexport interface WebLLMAppConfigLike {\n  model_list: WebLLMRecordLike[]\n}\n\n/** WebLLM's default window for its prebuilt models. */\nexport const WEBLLM_DEFAULT_CONTEXT_WINDOW = 4096\n\n/**\n * A copy of a WebLLM app config (e.g. `prebuiltAppConfig` from\n * `@mlc-ai/web-llm`) with one model's context window set. Pass it as\n * `engineConfig.appConfig` when building the model: the weights are the same,\n * only the KV cache (VRAM) grows with the window. Models trained on long\n * contexts (Qwen3, Llama 3.x) take far more than the default 4096; models\n * trained on 4k (Llama 2) don't.\n */\nexport const withWebLLMContextWindow = <T extends WebLLMAppConfigLike>(\n  appConfig: T,\n  modelId: string,\n  tokens: number,\n): T => ({\n  ...appConfig,\n  model_list: appConfig.model_list.map((r) =>\n    r.model_id === modelId\n      ? { ...r, overrides: { ...r.overrides, context_window_size: tokens } }\n      : r,\n  ),\n})\n\ninterface WebLLMModelLike {\n  provider?: string\n  modelId?: string\n  engine?: {\n    loadedModelIdToPipeline?: { get?: (id: string) => { contextWindowSize?: number } | undefined }\n  }\n  config?: { options?: { engineConfig?: { appConfig?: WebLLMAppConfigLike } } }\n}\n\n/**\n * The context window of a WebLLM model (`@browser-ai/web-llm`), or undefined\n * for any other model. Read from the loaded engine when it runs in this\n * thread, else from the app config it was built with, else WebLLM's defaults\n * (1024 for its \"-1k\" builds, 4096 otherwise). The provider keeps these\n * private, so this reaches in defensively.\n */\nexport const webLLMContextWindow = (model: unknown): number | undefined => {\n  const m = model as WebLLMModelLike | undefined\n  if (!m || typeof m !== 'object' || m.provider !== 'web-llm') return undefined\n  const id = String(m.modelId ?? '')\n  try {\n    const live = m.engine?.loadedModelIdToPipeline?.get?.(id)?.contextWindowSize\n    if (typeof live === 'number' && live > 0) return live\n  } catch {\n    /* the engine lives in a worker, or changed shape */\n  }\n  const record = m.config?.options?.engineConfig?.appConfig?.model_list?.find(\n    (r) => r.model_id === id,\n  )\n  const configured = record?.overrides?.context_window_size\n  if (typeof configured === 'number' && configured > 0) return configured\n  return /-1k$/i.test(id) ? 1024 : WEBLLM_DEFAULT_CONTEXT_WINDOW\n}\n\n/**\n * The model's own context window when it can be known (local models), else\n * undefined: WebLLM's (see above), or the browser's built-in model's\n * (`@browser-ai/core` reports it once its session exists).\n */\nexport const contextWindowOf = (model: unknown): number | undefined => {\n  const webllm = webLLMContextWindow(model)\n  if (webllm !== undefined) return webllm\n  const m = model as { provider?: string; getContextWindow?: () => number | undefined } | undefined\n  if (m && typeof m === 'object' && m.provider === 'browser-ai') {\n    try {\n      const w = m.getContextWindow?.()\n      return typeof w === 'number' && w > 0 ? w : undefined\n    } catch {\n      return undefined\n    }\n  }\n  return undefined\n}\n\n/**\n * Fit the compaction settings to the model's window: the window in use is the\n * configured one, never larger than the model's own, and every size derived\n * from it shrinks with it — also a size the host set for a larger window.\n */\nexport const fitCompactionToWindow = (\n  resolved: ResolvedCompaction,\n  configured: CompactionConfig | undefined,\n  modelWindow: number | undefined,\n): ResolvedCompaction => {\n  const window = Math.min(\n    configured?.contextWindowTokens ?? modelWindow ?? resolved.contextWindowTokens,\n    modelWindow ?? Infinity,\n  )\n  if (window === resolved.contextWindowTokens && modelWindow === undefined) return resolved\n  const half = Math.floor(window / 2)\n  const quarter = Math.floor(window / 4)\n  const clearAfter = configured?.clearToolResultsAfterTokens\n  return {\n    ...resolved,\n    contextWindowTokens: window,\n    thresholdTokens: Math.min(configured?.thresholdTokens ?? half, half),\n    // 0 = never stays never.\n    clearToolResultsAfterTokens: clearAfter === 0 ? 0 : Math.min(clearAfter ?? quarter, quarter),\n    // One result may take a quarter of the window (~4 chars per token).\n    maxToolOutputChars:\n      configured?.maxToolOutputChars === 0\n        ? 0\n        : Math.min(configured?.maxToolOutputChars ?? 20_000, window),\n    // A small window holds fewer verbatim results.\n    keepToolResults:\n      configured?.keepToolResults ?? Math.max(1, Math.min(3, Math.floor(window / 8192))),\n  }\n}\n\n/**\n * Estimated tokens of the tool definitions a model is sent (name, description,\n * JSON schema of the input) — for tools natively, or rendered into the prompt.\n */\nexport const toolDefinitionTokens = async (tools: ToolSet): Promise<number> => {\n  const sizes = await Promise.all(\n    Object.entries(tools).map(async ([name, t]) => {\n      let parameters: unknown\n      try {\n        parameters = await asSchema(t.inputSchema as never).jsonSchema\n      } catch {\n        parameters = undefined\n      }\n      return estimateTokens(JSON.stringify({ name, description: t.description ?? '', parameters }))\n    }),\n  )\n  return sizes.reduce((a, b) => a + b, 0)\n}\n\n/** 'auto' switches to tool search once the definitions would take this share of the window. */\nexport const TOOL_SEARCH_WINDOW_SHARE = 0.25\n\n/** Raised when a prompt does not fit the model's context window. */\nexport class ContextWindowExceededError extends Error {\n  /** The prompt's size as the provider counted it, when it said. */\n  readonly promptTokens?: number\n  /** The model's window as the provider reported it, when it said. */\n  readonly contextWindowTokens?: number\n  constructor(model: string, promptTokens?: number, contextWindowTokens?: number) {\n    const sizes =\n      promptTokens && contextWindowTokens\n        ? ` (${promptTokens} tokens; the window is ${contextWindowTokens})`\n        : ''\n    super(\n      `The conversation no longer fits the context window of \"${model}\"${sizes}. Start a new chat or compact this one, connect fewer tools, or switch to a model with a larger window.`,\n    )\n    this.name = 'ContextWindowExceededError'\n    this.promptTokens = promptTokens\n    this.contextWindowTokens = contextWindowTokens\n  }\n\n  toJSON(): string {\n    return this.message\n  }\n}\n\n// What providers say when the prompt is too long: WebLLM, OpenAI\n// (context_length_exceeded), Anthropic, Gemini, llama.cpp, vLLM / OpenRouter.\nconst OVERFLOW_RE =\n  /prompt tokens exceed context window|context window size|context_length_exceeded|maximum context length|prompt is too long|input token count.*exceeds|exceeds the maximum number of tokens|exceeds the available context size|context length exceeded|too many tokens in (the )?(prompt|input)/i\nconst WEBLLM_SIZES_RE = /number of prompt tokens:\\s*(\\d+);\\s*context window size:\\s*(\\d+)/i\n\n/**\n * The clear error to raise when a provider refused a prompt for its length\n * (else undefined) — instead of the provider's raw text.\n */\nexport const contextOverflowOf = (\n  err: unknown,\n  model: unknown,\n): ContextWindowExceededError | undefined => {\n  if (err instanceof ContextWindowExceededError) return err\n  const text = err instanceof Error ? `${err.name} ${err.message}` : String(err)\n  if (!OVERFLOW_RE.test(text)) return undefined\n  const sizes = WEBLLM_SIZES_RE.exec(text)\n  return new ContextWindowExceededError(\n    modelLabel(model),\n    sizes ? Number(sizes[1]) : undefined,\n    sizes ? Number(sizes[2]) : (webLLMContextWindow(model) ?? undefined),\n  )\n}\n","import type { FilePart } from 'ai'\n\n/**\n * Attachments sent with a run: images, PDFs and other files the provider can\n * read, given as bytes / base64 / a data URL — or as an http(s) URL, which\n * providers that accept links (Gemini, Claude, OpenAI) fetch themselves and the\n * AI SDK downloads for the rest (from the browser: subject to CORS).\n */\nexport interface RunFile {\n  /** Base64 (no prefix), a data URL (\"data:image/png;base64,…\"), raw bytes, or an http(s) URL. */\n  data: string | Uint8Array | URL\n  /** e.g. \"image/png\", \"application/pdf\" — read from a data URL when omitted. */\n  mediaType?: string\n  /** Shown in the transcript, in memory and in errors. */\n  name?: string\n}\n\n/** An image attachment (same shape; kept as its own name for clarity). */\nexport type RunImage = RunFile\n\n/** The kinds of attachment the agent reasons about separately. */\nexport type AttachmentKind = 'image' | 'pdf' | 'file'\n\nconst DATA_URL_RE = /^data:([^;,]+)?;base64,(.*)$/s\nconst HTTP_RE = /^https?:\\/\\//i\n\n/** The media type of an attachment (declared, from a data URL, or from a URL's extension). */\nexport const mediaTypeOf = (f: RunFile): string | undefined => {\n  if (f.mediaType) return f.mediaType\n  if (typeof f.data === 'string') {\n    const m = DATA_URL_RE.exec(f.data)\n    if (m?.[1]) return m[1]\n  }\n  const url = f.data instanceof URL ? f.data.href : typeof f.data === 'string' ? f.data : ''\n  if (HTTP_RE.test(url) || f.name) {\n    const ext = (f.name ?? new URL(url, 'http://x').pathname).split('.').pop()?.toLowerCase()\n    if (ext === 'pdf') return 'application/pdf'\n    if (ext && ['png', 'jpg', 'jpeg', 'gif', 'webp'].includes(ext)) {\n      return `image/${ext === 'jpg' ? 'jpeg' : ext}`\n    }\n  }\n  return undefined\n}\n\nexport const attachmentKind = (f: RunFile): AttachmentKind => {\n  const mt = mediaTypeOf(f) ?? ''\n  if (mt.startsWith('image/')) return 'image'\n  if (mt === 'application/pdf') return 'pdf'\n  return 'file'\n}\n\nconst KIND_LABEL: Record<AttachmentKind, string> = {\n  image: 'images',\n  pdf: 'PDF files',\n  file: 'file attachments',\n}\n\nconst KIND_HINT: Record<AttachmentKind, string> = {\n  image:\n    'Remove the image, or switch to a vision-capable model such as Gemini, Claude or GPT-4o/GPT-5.',\n  pdf: 'Convert the PDF to text first (e.g. PDF → Markdown), or switch to a model that reads PDFs such as Gemini, Claude or GPT-4o/GPT-5.',\n  file: 'Paste its text instead, or switch to a model that accepts this file type.',\n}\n\n/** Raised when attachments are sent to a model that cannot take them. */\nexport class AttachmentsNotSupportedError extends Error {\n  readonly kind: AttachmentKind\n  constructor(model: string, kind: AttachmentKind, detail?: string) {\n    super(\n      `The model \"${model}\" can't take ${KIND_LABEL[kind]}${detail ? ` (${detail})` : ''}. ${KIND_HINT[kind]}`,\n    )\n    this.name = 'AttachmentsNotSupportedError'\n    this.kind = kind\n  }\n\n  toJSON(): string {\n    return this.message\n  }\n}\n\n/** The image flavour of {@link AttachmentsNotSupportedError}. */\nexport class ImagesNotSupportedError extends AttachmentsNotSupportedError {\n  constructor(model: string, detail?: string) {\n    super(model, 'image', detail)\n    this.name = 'ImagesNotSupportedError'\n  }\n}\n\n/** The error for a kind: images get the dedicated class. */\nexport const unsupportedAttachment = (\n  model: string,\n  kind: AttachmentKind,\n  detail?: string,\n): AttachmentsNotSupportedError =>\n  kind === 'image'\n    ? new ImagesNotSupportedError(model, detail)\n    : new AttachmentsNotSupportedError(model, kind, detail)\n\nconst dataOf = (f: RunFile): string | Uint8Array | URL => {\n  if (typeof f.data === 'string') {\n    const m = DATA_URL_RE.exec(f.data)\n    if (m) return m[2]\n    if (HTTP_RE.test(f.data)) return new URL(f.data)\n  }\n  return f.data\n}\n\n/**\n * An AI SDK file part (images included — the v7 SDK deprecates the separate\n * image part). An http(s) URL is passed as a URL: providers that accept links\n * fetch it themselves, and the SDK downloads it for the rest.\n */\nexport const toFilePart = (f: RunFile): FilePart => {\n  const kind = attachmentKind(f)\n  const mediaType = mediaTypeOf(f) ?? (kind === 'image' ? 'image' : 'application/octet-stream')\n  return {\n    type: 'file',\n    data: dataOf(f),\n    mediaType,\n    ...(f.name ? { filename: f.name } : {}),\n  }\n}\n\n/** Back-compat name. */\nexport const toImagePart = toFilePart\n\n/** A short, human label of a model for messages. */\nexport const modelLabel = (model: unknown): string => {\n  if (typeof model === 'string') return model\n  const m = model as { provider?: string; modelId?: string } | undefined\n  return m?.modelId ? `${m.modelId}${m.provider ? ` (${m.provider})` : ''}` : 'this model'\n}\n\nconst REFUSAL_RE =\n  /image|vision|multimodal|multi-modal|media type|mediatype|pdf|document|unsupported (file|content|part|mime)|does not support (file|content)|invalid content type|image_url|inline_?data|file_?data/i\n\n/** Does a provider error look like a refusal of attachment input? */\nexport const isAttachmentRefusal = (err: unknown): boolean =>\n  REFUSAL_RE.test(err instanceof Error ? `${err.name} ${err.message}` : String(err))\n\n/** Back-compat name. */\nexport const isImageRefusal = isAttachmentRefusal\n","import {\n  generateObject,\n  generateText,\n  streamText,\n  type LanguageModel,\n  type ModelMessage,\n  type SystemModelMessage,\n  type ToolSet,\n} from 'ai'\nimport type { ZodType } from 'zod'\nimport type { ProviderOptionsMap, ThinkingLevel } from '../thinking.js'\nimport { promptOf, timeoutSignal } from './util.js'\n\n/** Options shared by the low-level generation helpers. Provide `prompt` OR `messages`. */\nexport interface GenerateOptions {\n  /** The system prompt — a string, or a SystemModelMessage (e.g. with a cache breakpoint). */\n  system?: string | SystemModelMessage\n  prompt?: string\n  messages?: ModelMessage[]\n  tools?: ToolSet\n  maxOutputTokens?: number\n  temperature?: number\n  /** Portable thinking level (see `resolveThinking`). */\n  reasoning?: ThinkingLevel\n  /** Provider-keyed options (thinking budgets, cache keys, …). */\n  providerOptions?: ProviderOptionsMap\n  abortSignal?: AbortSignal\n  /** Time-box the call; combined with `abortSignal` (0/undefined = no timeout). */\n  timeoutMs?: number\n}\n\n/** The settings every helper forwards; omits undefined keys the SDK would reject. */\nconst callSettings = (opts: GenerateOptions) => ({\n  ...(opts.system !== undefined ? { instructions: opts.system } : {}),\n  maxOutputTokens: opts.maxOutputTokens,\n  temperature: opts.temperature,\n  ...(opts.reasoning ? { reasoning: opts.reasoning } : {}),\n  ...(opts.providerOptions ? { providerOptions: opts.providerOptions as never } : {}),\n  abortSignal: timeoutSignal(opts.abortSignal, opts.timeoutMs),\n})\n\n/**\n * One-shot text generation. Thin, provider-agnostic wrapper over the AI SDK's\n * `generateText` — the same call works against a cloud model, a local WebLLM\n * model, or Chrome/Edge built-in AI. Returns the full AI SDK result (`.text`,\n * `.usage`, `.toolCalls`, `.steps`, …).\n */\nexport const generate = (\n  model: LanguageModel,\n  opts: GenerateOptions = {},\n): ReturnType<typeof generateText> =>\n  generateText({\n    model,\n    ...callSettings(opts),\n    ...promptOf(opts),\n    tools: opts.tools,\n  })\n\n/** Streaming text generation. Returns the AI SDK `streamText` result (`.textStream`, `.fullStream`). */\nexport const stream = (\n  model: LanguageModel,\n  opts: GenerateOptions = {},\n): ReturnType<typeof streamText> =>\n  streamText({\n    model,\n    ...callSettings(opts),\n    ...promptOf(opts),\n    tools: opts.tools,\n  })\n\n/**\n * Structured output constrained to a Zod schema, via the AI SDK's\n * `generateObject`. Cloud models and WebLLM both support this; tiny local\n * models can be unreliable — the agent's prompted mode is the fallback.\n * Returns the AI SDK result (`.object`, `.usage`).\n */\nexport const generateStructured = <OBJECT>(\n  model: LanguageModel,\n  schema: ZodType<OBJECT>,\n  opts: GenerateOptions = {},\n) =>\n  generateObject({\n    model,\n    schema,\n    ...callSettings(opts),\n    ...promptOf(opts),\n  })\n","import type { ModelMessage } from 'ai'\nimport type { IUsage } from '../agent/loop-types.js'\n\n/** Truncate to `max` characters, ellipsising the overflow. */\nexport const clip = (s: string, max: number): string =>\n  s.length > max ? `${s.slice(0, max - 1)}…` : s\n\n/**\n * The AI SDK prompt is `{ prompt } XOR { messages }` — passing both keys (even\n * with one undefined) breaks the discriminated union. Pick exactly one.\n */\nexport const promptOf = (opts: {\n  prompt?: string\n  messages?: ModelMessage[]\n}): { prompt: string } | { messages: ModelMessage[] } =>\n  opts.messages ? { messages: opts.messages } : { prompt: opts.prompt ?? '' }\n\n/** The AI SDK usage shape, every field optional (providers omit what they don't report). */\nexport interface SdkUsageLike {\n  inputTokens?: number\n  outputTokens?: number\n  totalTokens?: number\n  inputTokenDetails?: { cacheReadTokens?: number; cacheWriteTokens?: number }\n  outputTokenDetails?: { reasoningTokens?: number }\n  /** Pre-v6 flat fields, still emitted by some providers. */\n  reasoningTokens?: number\n  cachedInputTokens?: number\n}\n\n/** Normalise the AI SDK usage shape (fields can be undefined) into IUsage. */\nexport const normalizeUsage = (u?: SdkUsageLike): IUsage => {\n  const inputTokens = u?.inputTokens ?? 0\n  const outputTokens = u?.outputTokens ?? 0\n  return {\n    inputTokens,\n    outputTokens,\n    totalTokens: u?.totalTokens ?? inputTokens + outputTokens,\n    reasoningTokens: u?.outputTokenDetails?.reasoningTokens ?? u?.reasoningTokens ?? 0,\n    cachedInputTokens: u?.inputTokenDetails?.cacheReadTokens ?? u?.cachedInputTokens ?? 0,\n    cacheWriteTokens: u?.inputTokenDetails?.cacheWriteTokens ?? 0,\n  }\n}\n\n/**\n * Combine the caller's AbortSignal with a timeout so a hung generation can't\n * freeze a run. Returns the original signal when no timeout is set. Uses the\n * standard `AbortSignal.timeout`/`AbortSignal.any` (Node ≥ 20, modern browsers).\n */\nexport const timeoutSignal = (signal?: AbortSignal, ms?: number): AbortSignal | undefined => {\n  if (!ms || ms <= 0) return signal\n  const timeout = AbortSignal.timeout(ms)\n  if (!signal) return timeout\n  return typeof AbortSignal.any === 'function' ? AbortSignal.any([signal, timeout]) : signal\n}\n","import type { LanguageModel } from 'ai'\nimport { generate } from '../llm/generate.js'\nimport { normalizeUsage } from '../llm/util.js'\nimport type { IUsage } from '../agent/loop-types.js'\nimport type { StoredMessage } from './store.js'\n\n/**\n * Context compaction. A model's window is finite and every token in it is paid\n * for on every call, so long transcripts and long runs are folded into a\n * summary: the oldest material is summarised by a model, the most recent is\n * kept verbatim.\n */\nexport interface CompactionConfig {\n  /** Compact automatically before planning and between steps (default true). */\n  auto?: boolean\n  /** The model's context window, in tokens (default 128 000). */\n  contextWindowTokens?: number\n  /** Compact once the estimated context exceeds this (default: half the window). */\n  thresholdTokens?: number\n  /** Transcript messages kept verbatim (default 4). */\n  keepRecentTurns?: number\n  /** Executed steps kept verbatim in the run's trace (default 3). */\n  keepRecentSteps?: number\n  /** Output cap of a summary call (default 1024). */\n  summaryMaxTokens?: number\n  /** Cap on ONE tool result as the model sees it, in chars (default 20 000; 0 = no cap). */\n  maxToolOutputChars?: number\n  /**\n   * Inside one step's tool loop: once its context passes this many tokens,\n   * replace the oldest tool results with one-line stubs (context editing;\n   * default: a quarter of the window; 0 = never).\n   */\n  clearToolResultsAfterTokens?: number\n  /** Most recent tool results kept verbatim when clearing (default 3). */\n  keepToolResults?: number\n}\n\nexport interface ResolvedCompaction {\n  auto: boolean\n  contextWindowTokens: number\n  thresholdTokens: number\n  keepRecentTurns: number\n  keepRecentSteps: number\n  summaryMaxTokens: number\n  maxToolOutputChars: number\n  clearToolResultsAfterTokens: number\n  keepToolResults: number\n}\n\nexport const resolveCompaction = (c: CompactionConfig | undefined): ResolvedCompaction => {\n  const contextWindowTokens = c?.contextWindowTokens ?? 128_000\n  return {\n    auto: c?.auto ?? true,\n    contextWindowTokens,\n    thresholdTokens: c?.thresholdTokens ?? Math.floor(contextWindowTokens / 2),\n    keepRecentTurns: c?.keepRecentTurns ?? 4,\n    keepRecentSteps: c?.keepRecentSteps ?? 3,\n    summaryMaxTokens: c?.summaryMaxTokens ?? 1024,\n    maxToolOutputChars: c?.maxToolOutputChars ?? 20_000,\n    clearToolResultsAfterTokens:\n      c?.clearToolResultsAfterTokens ?? Math.floor(contextWindowTokens / 4),\n    keepToolResults: c?.keepToolResults ?? 3,\n  }\n}\n\n/** Cheap, provider-independent token estimate (~4 chars per token). */\nexport const estimateTokens = (text: string): number => Math.ceil((text ?? '').length / 4)\n\nconst messageTokens = (msgs: StoredMessage[]): number =>\n  msgs.reduce((n, m) => n + estimateTokens(`${m.role}: ${m.content}\\n`), 0)\n\nexport interface CompressOptions {\n  /** Compress once the transcript exceeds this many characters (0 = never). */\n  maxChars?: number\n  /** Compress once the transcript's estimated tokens exceed this (wins over maxChars). */\n  thresholdTokens?: number\n  /** How many most-recent messages to keep verbatim. */\n  keepRecent?: number\n  /** Output cap of the summary call. */\n  summaryMaxTokens?: number\n  /** Compact even when below the threshold (manual \"compact now\"). */\n  force?: boolean\n  /** Time-box the summarisation call so it can never hang the caller. */\n  timeoutMs?: number\n  abortSignal?: AbortSignal\n  /** Receives the summary call's token usage. */\n  onUsage?: (usage: IUsage) => void\n}\n\nconst SUMMARY_SYSTEM =\n  'Summarize the following conversation into a compact set of durable facts, decisions, ' +\n  'identifiers, open tasks and user preferences. Keep names, numbers and IDs verbatim. ' +\n  'Output plain text only.'\n\n/**\n * When the transcript grows too long, summarise everything except the last\n * `keepRecent` messages into a single system message (via the model), so the\n * session stays within the model's context window. Returns the same array\n * unchanged when no compression is needed; never throws.\n */\nexport const compressHistory = async (\n  messages: StoredMessage[],\n  model: LanguageModel,\n  opts: CompressOptions = {},\n): Promise<StoredMessage[]> => {\n  const keepRecent = opts.keepRecent ?? 6\n  if (messages.length <= keepRecent + 1) return messages\n  if (!opts.force) {\n    if (opts.thresholdTokens !== undefined) {\n      if (opts.thresholdTokens <= 0 || messageTokens(messages) <= opts.thresholdTokens) {\n        return messages\n      }\n    } else {\n      const maxChars = opts.maxChars ?? 12_000\n      const chars = messages.reduce((n, m) => n + m.content.length, 0)\n      if (maxChars <= 0 || chars <= maxChars) return messages\n    }\n  }\n  const head = messages.slice(0, messages.length - keepRecent)\n  const tail = messages.slice(messages.length - keepRecent)\n  // The newest part of the head matters most; keep the tail end of it when it\n  // has to be cut to fit the summariser's own window.\n  const transcript = head\n    .map((m) => `${m.role}: ${m.content}`)\n    .join('\\n')\n    .slice(-48_000)\n  try {\n    const result = await generate(model, {\n      system: SUMMARY_SYSTEM,\n      prompt: transcript,\n      maxOutputTokens: opts.summaryMaxTokens ?? 400,\n      timeoutMs: opts.timeoutMs,\n      abortSignal: opts.abortSignal,\n    })\n    opts.onUsage?.(normalizeUsage(result.usage))\n    const clean = (result.text || '').trim()\n    if (!clean) return messages\n    const summaryMsg: StoredMessage = {\n      role: 'system',\n      content: `Summary of earlier conversation:\\n${clean}`,\n      ts: head[0]?.ts,\n    }\n    return [summaryMsg, ...tail]\n  } catch {\n    return messages\n  }\n}\n\n/** Alias with the sibling package's name; same behaviour as compressHistory. */\nexport const compactHistory = compressHistory\n\nconst TRACE_SYSTEM =\n  'Summarize these completed agent steps into a compact progress note: what was done, ' +\n  'what was found (keep names, numbers and IDs verbatim) and what failed. Plain text only.'\n\n/**\n * Fold the oldest entries of a run's step log into one summary entry, keeping\n * the last `keepRecent` verbatim. Returns the input unchanged when it fits the\n * threshold or on any failure; never throws.\n */\nexport const compactSteps = async (\n  done: string[],\n  model: LanguageModel,\n  opts: CompressOptions & { thresholdTokens: number },\n): Promise<string[]> => {\n  const keepRecent = opts.keepRecent ?? 3\n  if (done.length <= keepRecent + 1) return done\n  if (!opts.force && estimateTokens(done.join('\\n')) <= opts.thresholdTokens) return done\n  const head = done.slice(0, done.length - keepRecent)\n  try {\n    const result = await generate(model, {\n      system: TRACE_SYSTEM,\n      prompt: head.map((d, i) => `${i + 1}. ${d}`).join('\\n'),\n      maxOutputTokens: opts.summaryMaxTokens ?? 600,\n      timeoutMs: opts.timeoutMs,\n      abortSignal: opts.abortSignal,\n    })\n    opts.onUsage?.(normalizeUsage(result.usage))\n    const clean = (result.text || '').trim()\n    if (!clean) return done\n    return [`Summary of ${head.length} earlier step(s): ${clean}`, ...done.slice(head.length)]\n  } catch {\n    return done\n  }\n}\n\nexport { messageTokens as estimateMessagesTokens }\n"],"mappings":";AAAA,SAAS,gBAAAA,qBAAwC;;;ACAjD,SAAS,gBAA8B;;;ACuBvC,IAAM,cAAc;AACpB,IAAM,UAAU;AAGT,IAAM,cAAc,CAAC,MAAmC;AAC7D,MAAI,EAAE,UAAW,QAAO,EAAE;AAC1B,MAAI,OAAO,EAAE,SAAS,UAAU;AAC9B,UAAM,IAAI,YAAY,KAAK,EAAE,IAAI;AACjC,QAAI,IAAI,CAAC,EAAG,QAAO,EAAE,CAAC;AAAA,EACxB;AACA,QAAM,MAAM,EAAE,gBAAgB,MAAM,EAAE,KAAK,OAAO,OAAO,EAAE,SAAS,WAAW,EAAE,OAAO;AACxF,MAAI,QAAQ,KAAK,GAAG,KAAK,EAAE,MAAM;AAC/B,UAAM,OAAO,EAAE,QAAQ,IAAI,IAAI,KAAK,UAAU,EAAE,UAAU,MAAM,GAAG,EAAE,IAAI,GAAG,YAAY;AACxF,QAAI,QAAQ,MAAO,QAAO;AAC1B,QAAI,OAAO,CAAC,OAAO,OAAO,QAAQ,OAAO,MAAM,EAAE,SAAS,GAAG,GAAG;AAC9D,aAAO,SAAS,QAAQ,QAAQ,SAAS,GAAG;AAAA,IAC9C;AAAA,EACF;AACA,SAAO;AACT;AAEO,IAAM,iBAAiB,CAAC,MAA+B;AAC5D,QAAM,KAAK,YAAY,CAAC,KAAK;AAC7B,MAAI,GAAG,WAAW,QAAQ,EAAG,QAAO;AACpC,MAAI,OAAO,kBAAmB,QAAO;AACrC,SAAO;AACT;AAEA,IAAM,aAA6C;AAAA,EACjD,OAAO;AAAA,EACP,KAAK;AAAA,EACL,MAAM;AACR;AAEA,IAAM,YAA4C;AAAA,EAChD,OACE;AAAA,EACF,KAAK;AAAA,EACL,MAAM;AACR;AAGO,IAAM,+BAAN,cAA2C,MAAM;AAAA,EAC7C;AAAA,EACT,YAAY,OAAe,MAAsB,QAAiB;AAChE;AAAA,MACE,cAAc,KAAK,gBAAgB,WAAW,IAAI,CAAC,GAAG,SAAS,KAAK,MAAM,MAAM,EAAE,KAAK,UAAU,IAAI,CAAC;AAAA,IACxG;AACA,SAAK,OAAO;AACZ,SAAK,OAAO;AAAA,EACd;AAAA,EAEA,SAAiB;AACf,WAAO,KAAK;AAAA,EACd;AACF;AAGO,IAAM,0BAAN,cAAsC,6BAA6B;AAAA,EACxE,YAAY,OAAe,QAAiB;AAC1C,UAAM,OAAO,SAAS,MAAM;AAC5B,SAAK,OAAO;AAAA,EACd;AACF;AAGO,IAAM,wBAAwB,CACnC,OACA,MACA,WAEA,SAAS,UACL,IAAI,wBAAwB,OAAO,MAAM,IACzC,IAAI,6BAA6B,OAAO,MAAM,MAAM;AAE1D,IAAM,SAAS,CAAC,MAA0C;AACxD,MAAI,OAAO,EAAE,SAAS,UAAU;AAC9B,UAAM,IAAI,YAAY,KAAK,EAAE,IAAI;AACjC,QAAI,EAAG,QAAO,EAAE,CAAC;AACjB,QAAI,QAAQ,KAAK,EAAE,IAAI,EAAG,QAAO,IAAI,IAAI,EAAE,IAAI;AAAA,EACjD;AACA,SAAO,EAAE;AACX;AAOO,IAAM,aAAa,CAAC,MAAyB;AAClD,QAAM,OAAO,eAAe,CAAC;AAC7B,QAAM,YAAY,YAAY,CAAC,MAAM,SAAS,UAAU,UAAU;AAClE,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,OAAO,CAAC;AAAA,IACd;AAAA,IACA,GAAI,EAAE,OAAO,EAAE,UAAU,EAAE,KAAK,IAAI,CAAC;AAAA,EACvC;AACF;AAGO,IAAM,cAAc;AAGpB,IAAM,aAAa,CAAC,UAA2B;AACpD,MAAI,OAAO,UAAU,SAAU,QAAO;AACtC,QAAM,IAAI;AACV,SAAO,GAAG,UAAU,GAAG,EAAE,OAAO,GAAG,EAAE,WAAW,KAAK,EAAE,QAAQ,MAAM,EAAE,KAAK;AAC9E;AAEA,IAAM,aACJ;AAGK,IAAM,sBAAsB,CAAC,QAClC,WAAW,KAAK,eAAe,QAAQ,GAAG,IAAI,IAAI,IAAI,IAAI,OAAO,KAAK,OAAO,GAAG,CAAC;AAG5E,IAAM,iBAAiB;;;AC7I9B;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OAKK;;;ACJA,IAAM,OAAO,CAAC,GAAW,QAC9B,EAAE,SAAS,MAAM,GAAG,EAAE,MAAM,GAAG,MAAM,CAAC,CAAC,WAAM;AAMxC,IAAM,WAAW,CAAC,SAIvB,KAAK,WAAW,EAAE,UAAU,KAAK,SAAS,IAAI,EAAE,QAAQ,KAAK,UAAU,GAAG;AAerE,IAAM,iBAAiB,CAAC,MAA6B;AAC1D,QAAM,cAAc,GAAG,eAAe;AACtC,QAAM,eAAe,GAAG,gBAAgB;AACxC,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,aAAa,GAAG,eAAe,cAAc;AAAA,IAC7C,iBAAiB,GAAG,oBAAoB,mBAAmB,GAAG,mBAAmB;AAAA,IACjF,mBAAmB,GAAG,mBAAmB,mBAAmB,GAAG,qBAAqB;AAAA,IACpF,kBAAkB,GAAG,mBAAmB,oBAAoB;AAAA,EAC9D;AACF;AAOO,IAAM,gBAAgB,CAAC,QAAsB,OAAyC;AAC3F,MAAI,CAAC,MAAM,MAAM,EAAG,QAAO;AAC3B,QAAM,UAAU,YAAY,QAAQ,EAAE;AACtC,MAAI,CAAC,OAAQ,QAAO;AACpB,SAAO,OAAO,YAAY,QAAQ,aAAa,YAAY,IAAI,CAAC,QAAQ,OAAO,CAAC,IAAI;AACtF;;;ADrBA,IAAM,eAAe,CAAC,UAA2B;AAAA,EAC/C,GAAI,KAAK,WAAW,SAAY,EAAE,cAAc,KAAK,OAAO,IAAI,CAAC;AAAA,EACjE,iBAAiB,KAAK;AAAA,EACtB,aAAa,KAAK;AAAA,EAClB,GAAI,KAAK,YAAY,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,EACtD,GAAI,KAAK,kBAAkB,EAAE,iBAAiB,KAAK,gBAAyB,IAAI,CAAC;AAAA,EACjF,aAAa,cAAc,KAAK,aAAa,KAAK,SAAS;AAC7D;AAQO,IAAM,WAAW,CACtB,OACA,OAAwB,CAAC,MAEzB,aAAa;AAAA,EACX;AAAA,EACA,GAAG,aAAa,IAAI;AAAA,EACpB,GAAG,SAAS,IAAI;AAAA,EAChB,OAAO,KAAK;AACd,CAAC;AAGI,IAAM,SAAS,CACpB,OACA,OAAwB,CAAC,MAEzB,WAAW;AAAA,EACT;AAAA,EACA,GAAG,aAAa,IAAI;AAAA,EACpB,GAAG,SAAS,IAAI;AAAA,EAChB,OAAO,KAAK;AACd,CAAC;AAQI,IAAM,qBAAqB,CAChC,OACA,QACA,OAAwB,CAAC,MAEzB,eAAe;AAAA,EACb;AAAA,EACA;AAAA,EACA,GAAG,aAAa,IAAI;AAAA,EACpB,GAAG,SAAS,IAAI;AAClB,CAAC;;;AErCI,IAAM,oBAAoB,CAAC,MAAwD;AACxF,QAAM,sBAAsB,GAAG,uBAAuB;AACtD,SAAO;AAAA,IACL,MAAM,GAAG,QAAQ;AAAA,IACjB;AAAA,IACA,iBAAiB,GAAG,mBAAmB,KAAK,MAAM,sBAAsB,CAAC;AAAA,IACzE,iBAAiB,GAAG,mBAAmB;AAAA,IACvC,iBAAiB,GAAG,mBAAmB;AAAA,IACvC,kBAAkB,GAAG,oBAAoB;AAAA,IACzC,oBAAoB,GAAG,sBAAsB;AAAA,IAC7C,6BACE,GAAG,+BAA+B,KAAK,MAAM,sBAAsB,CAAC;AAAA,IACtE,iBAAiB,GAAG,mBAAmB;AAAA,EACzC;AACF;AAGO,IAAM,iBAAiB,CAAC,SAAyB,KAAK,MAAM,QAAQ,IAAI,SAAS,CAAC;AAEzF,IAAM,gBAAgB,CAAC,SACrB,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,eAAe,GAAG,EAAE,IAAI,KAAK,EAAE,OAAO;AAAA,CAAI,GAAG,CAAC;AAoB1E,IAAM,iBACJ;AAUK,IAAM,kBAAkB,OAC7B,UACA,OACA,OAAwB,CAAC,MACI;AAC7B,QAAM,aAAa,KAAK,cAAc;AACtC,MAAI,SAAS,UAAU,aAAa,EAAG,QAAO;AAC9C,MAAI,CAAC,KAAK,OAAO;AACf,QAAI,KAAK,oBAAoB,QAAW;AACtC,UAAI,KAAK,mBAAmB,KAAK,cAAc,QAAQ,KAAK,KAAK,iBAAiB;AAChF,eAAO;AAAA,MACT;AAAA,IACF,OAAO;AACL,YAAM,WAAW,KAAK,YAAY;AAClC,YAAM,QAAQ,SAAS,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,QAAQ,QAAQ,CAAC;AAC/D,UAAI,YAAY,KAAK,SAAS,SAAU,QAAO;AAAA,IACjD;AAAA,EACF;AACA,QAAM,OAAO,SAAS,MAAM,GAAG,SAAS,SAAS,UAAU;AAC3D,QAAM,OAAO,SAAS,MAAM,SAAS,SAAS,UAAU;AAGxD,QAAM,aAAa,KAChB,IAAI,CAAC,MAAM,GAAG,EAAE,IAAI,KAAK,EAAE,OAAO,EAAE,EACpC,KAAK,IAAI,EACT,MAAM,KAAO;AAChB,MAAI;AACF,UAAM,SAAS,MAAM,SAAS,OAAO;AAAA,MACnC,QAAQ;AAAA,MACR,QAAQ;AAAA,MACR,iBAAiB,KAAK,oBAAoB;AAAA,MAC1C,WAAW,KAAK;AAAA,MAChB,aAAa,KAAK;AAAA,IACpB,CAAC;AACD,SAAK,UAAU,eAAe,OAAO,KAAK,CAAC;AAC3C,UAAM,SAAS,OAAO,QAAQ,IAAI,KAAK;AACvC,QAAI,CAAC,MAAO,QAAO;AACnB,UAAM,aAA4B;AAAA,MAChC,MAAM;AAAA,MACN,SAAS;AAAA,EAAqC,KAAK;AAAA,MACnD,IAAI,KAAK,CAAC,GAAG;AAAA,IACf;AACA,WAAO,CAAC,YAAY,GAAG,IAAI;AAAA,EAC7B,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAGO,IAAM,iBAAiB;AAE9B,IAAM,eACJ;AAQK,IAAM,eAAe,OAC1B,MACA,OACA,SACsB;AACtB,QAAM,aAAa,KAAK,cAAc;AACtC,MAAI,KAAK,UAAU,aAAa,EAAG,QAAO;AAC1C,MAAI,CAAC,KAAK,SAAS,eAAe,KAAK,KAAK,IAAI,CAAC,KAAK,KAAK,gBAAiB,QAAO;AACnF,QAAM,OAAO,KAAK,MAAM,GAAG,KAAK,SAAS,UAAU;AACnD,MAAI;AACF,UAAM,SAAS,MAAM,SAAS,OAAO;AAAA,MACnC,QAAQ;AAAA,MACR,QAAQ,KAAK,IAAI,CAAC,GAAG,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,EAAE,EAAE,KAAK,IAAI;AAAA,MACtD,iBAAiB,KAAK,oBAAoB;AAAA,MAC1C,WAAW,KAAK;AAAA,MAChB,aAAa,KAAK;AAAA,IACpB,CAAC;AACD,SAAK,UAAU,eAAe,OAAO,KAAK,CAAC;AAC3C,UAAM,SAAS,OAAO,QAAQ,IAAI,KAAK;AACvC,QAAI,CAAC,MAAO,QAAO;AACnB,WAAO,CAAC,cAAc,KAAK,MAAM,qBAAqB,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,MAAM,CAAC;AAAA,EAC3F,QAAQ;AACN,WAAO;AAAA,EACT;AACF;;;AJxJO,IAAM,gCAAgC;AAUtC,IAAM,0BAA0B,CACrC,WACA,SACA,YACO;AAAA,EACP,GAAG;AAAA,EACH,YAAY,UAAU,WAAW;AAAA,IAAI,CAAC,MACpC,EAAE,aAAa,UACX,EAAE,GAAG,GAAG,WAAW,EAAE,GAAG,EAAE,WAAW,qBAAqB,OAAO,EAAE,IACnE;AAAA,EACN;AACF;AAkBO,IAAM,sBAAsB,CAAC,UAAuC;AACzE,QAAM,IAAI;AACV,MAAI,CAAC,KAAK,OAAO,MAAM,YAAY,EAAE,aAAa,UAAW,QAAO;AACpE,QAAM,KAAK,OAAO,EAAE,WAAW,EAAE;AACjC,MAAI;AACF,UAAM,OAAO,EAAE,QAAQ,yBAAyB,MAAM,EAAE,GAAG;AAC3D,QAAI,OAAO,SAAS,YAAY,OAAO,EAAG,QAAO;AAAA,EACnD,QAAQ;AAAA,EAER;AACA,QAAM,SAAS,EAAE,QAAQ,SAAS,cAAc,WAAW,YAAY;AAAA,IACrE,CAAC,MAAM,EAAE,aAAa;AAAA,EACxB;AACA,QAAM,aAAa,QAAQ,WAAW;AACtC,MAAI,OAAO,eAAe,YAAY,aAAa,EAAG,QAAO;AAC7D,SAAO,QAAQ,KAAK,EAAE,IAAI,OAAO;AACnC;AAOO,IAAM,kBAAkB,CAAC,UAAuC;AACrE,QAAM,SAAS,oBAAoB,KAAK;AACxC,MAAI,WAAW,OAAW,QAAO;AACjC,QAAM,IAAI;AACV,MAAI,KAAK,OAAO,MAAM,YAAY,EAAE,aAAa,cAAc;AAC7D,QAAI;AACF,YAAM,IAAI,EAAE,mBAAmB;AAC/B,aAAO,OAAO,MAAM,YAAY,IAAI,IAAI,IAAI;AAAA,IAC9C,QAAQ;AACN,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AAOO,IAAM,wBAAwB,CACnC,UACA,YACA,gBACuB;AACvB,QAAM,SAAS,KAAK;AAAA,IAClB,YAAY,uBAAuB,eAAe,SAAS;AAAA,IAC3D,eAAe;AAAA,EACjB;AACA,MAAI,WAAW,SAAS,uBAAuB,gBAAgB,OAAW,QAAO;AACjF,QAAM,OAAO,KAAK,MAAM,SAAS,CAAC;AAClC,QAAM,UAAU,KAAK,MAAM,SAAS,CAAC;AACrC,QAAM,aAAa,YAAY;AAC/B,SAAO;AAAA,IACL,GAAG;AAAA,IACH,qBAAqB;AAAA,IACrB,iBAAiB,KAAK,IAAI,YAAY,mBAAmB,MAAM,IAAI;AAAA;AAAA,IAEnE,6BAA6B,eAAe,IAAI,IAAI,KAAK,IAAI,cAAc,SAAS,OAAO;AAAA;AAAA,IAE3F,oBACE,YAAY,uBAAuB,IAC/B,IACA,KAAK,IAAI,YAAY,sBAAsB,KAAQ,MAAM;AAAA;AAAA,IAE/D,iBACE,YAAY,mBAAmB,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,MAAM,SAAS,IAAI,CAAC,CAAC;AAAA,EACrF;AACF;AAMO,IAAM,uBAAuB,OAAO,UAAoC;AAC7E,QAAM,QAAQ,MAAM,QAAQ;AAAA,IAC1B,OAAO,QAAQ,KAAK,EAAE,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM;AAC7C,UAAI;AACJ,UAAI;AACF,qBAAa,MAAM,SAAS,EAAE,WAAoB,EAAE;AAAA,MACtD,QAAQ;AACN,qBAAa;AAAA,MACf;AACA,aAAO,eAAe,KAAK,UAAU,EAAE,MAAM,aAAa,EAAE,eAAe,IAAI,WAAW,CAAC,CAAC;AAAA,IAC9F,CAAC;AAAA,EACH;AACA,SAAO,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AACxC;AAGO,IAAM,2BAA2B;AAGjC,IAAM,6BAAN,cAAyC,MAAM;AAAA;AAAA,EAE3C;AAAA;AAAA,EAEA;AAAA,EACT,YAAY,OAAe,cAAuB,qBAA8B;AAC9E,UAAM,QACJ,gBAAgB,sBACZ,KAAK,YAAY,0BAA0B,mBAAmB,MAC9D;AACN;AAAA,MACE,0DAA0D,KAAK,IAAI,KAAK;AAAA,IAC1E;AACA,SAAK,OAAO;AACZ,SAAK,eAAe;AACpB,SAAK,sBAAsB;AAAA,EAC7B;AAAA,EAEA,SAAiB;AACf,WAAO,KAAK;AAAA,EACd;AACF;AAIA,IAAM,cACJ;AACF,IAAM,kBAAkB;AAMjB,IAAM,oBAAoB,CAC/B,KACA,UAC2C;AAC3C,MAAI,eAAe,2BAA4B,QAAO;AACtD,QAAM,OAAO,eAAe,QAAQ,GAAG,IAAI,IAAI,IAAI,IAAI,OAAO,KAAK,OAAO,GAAG;AAC7E,MAAI,CAAC,YAAY,KAAK,IAAI,EAAG,QAAO;AACpC,QAAM,QAAQ,gBAAgB,KAAK,IAAI;AACvC,SAAO,IAAI;AAAA,IACT,WAAW,KAAK;AAAA,IAChB,QAAQ,OAAO,MAAM,CAAC,CAAC,IAAI;AAAA,IAC3B,QAAQ,OAAO,MAAM,CAAC,CAAC,IAAK,oBAAoB,KAAK,KAAK;AAAA,EAC5D;AACF;;;ADjNO,IAAM,oBAAoB,MAC/B,OAAO,cAAc,eAAe,SAAS;AAuCxC,IAAM,qBAAqB,OAAO,UAAwC;AAC/E,QAAMC,cAAa,EAAE,OAAO,QAAQ,MAAM,iBAAiB,EAAE,CAAC;AAChE;AAUO,IAAM,oBAAoB,OAAO,UAAwC;AAC9E,QAAM,SAAU,MAAwD;AACxE,MAAI;AACF,UAAM,QAAQ,SAAS;AAAA,EACzB,QAAQ;AAAA,EAER;AACF;AAWO,IAAM,oBAAoB,OAC/B,OACA,YAC2B;AAC3B,MAAI;AACJ,MAAI;AACF,UAAM,MAAM,OAAO,qBAAqB;AAAA,EAC1C,QAAQ;AACN,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,EAAE,UAAU,MAAM,qBAAqB,GAAG,SAAS,IAAI,WAAW,CAAC;AACzE,MAAI,qBAAqB;AACvB,UAAM,eAAgB,SAAS,gBAAgB,CAAC;AAChD,QAAI,OAAO,aAAa;AACxB,QAAI,CAAC,MAAM;AACT,UAAI;AACF,cAAM,OAAO;AACb,gBACG,MAAM;AAAA;AAAA;AAAA,UAAoD;AAAA,WAG3D;AAAA,MACJ,QAAQ;AACN,cAAM,IAAI;AAAA,UACR;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,aAAS,eAAe;AAAA,MACtB,GAAG;AAAA,MACH,WAAW,wBAAwB,MAAM,OAAO,mBAAmB;AAAA,IACrE;AAAA,EACF;AAGA,QAAM,gBAAgB,IAAI;AAAA,IACxB;AAAA,IACA;AAAA,EACF;AACA,MAAI,QAAS,OAAM,mBAAmB,aAAa;AACnD,SAAO;AACT;","names":["generateText","generateText"]}