@dudko.dev/agent-web 0.0.20 → 0.0.22

package/dist/index.cjs

@@ -30,6 +30,364 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
 ));
 var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
 
+// src/llm/util.ts
+var clip, promptOf, normalizeUsage, timeoutSignal;
+var init_util = __esm({
+  "src/llm/util.ts"() {
+    "use strict";
+    clip = (s, max) => s.length > max ? `${s.slice(0, max - 1)}\u2026` : s;
+    promptOf = (opts) => opts.messages ? { messages: opts.messages } : { prompt: opts.prompt ?? "" };
+    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
+      };
+    };
+    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 import_ai, callSettings, generate, stream, generateStructured;
+var init_generate = __esm({
+  "src/llm/generate.ts"() {
+    "use strict";
+    import_ai = require("ai");
+    init_util();
+    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)
+    });
+    generate = (model, opts = {}) => (0, import_ai.generateText)({
+      model,
+      ...callSettings(opts),
+      ...promptOf(opts),
+      tools: opts.tools
+    });
+    stream = (model, opts = {}) => (0, import_ai.streamText)({
+      model,
+      ...callSettings(opts),
+      ...promptOf(opts),
+      tools: opts.tools
+    });
+    generateStructured = (model, schema, opts = {}) => (0, import_ai.generateObject)({
+      model,
+      schema,
+      ...callSettings(opts),
+      ...promptOf(opts)
+    });
+  }
+});
+
+// src/memory/compress.ts
+var resolveCompaction, estimateTokens, messageTokens, SUMMARY_SYSTEM, compressHistory, compactHistory, TRACE_SYSTEM, compactSteps;
+var init_compress = __esm({
+  "src/memory/compress.ts"() {
+    "use strict";
+    init_generate();
+    init_util();
+    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
+      };
+    };
+    estimateTokens = (text) => Math.ceil((text ?? "").length / 4);
+    messageTokens = (msgs) => msgs.reduce((n, m) => n + estimateTokens(`${m.role}: ${m.content}
+`), 0);
+    SUMMARY_SYSTEM = "Summarize the following conversation into a compact set of durable facts, decisions, identifiers, open tasks and user preferences. Keep names, numbers and IDs verbatim. Output plain text only.";
+    compressHistory = async (messages, model, opts = {}) => {
+      const keepRecent = opts.keepRecent ?? 6;
+      if (messages.length <= keepRecent + 1) return messages;
+      if (!opts.force) {
+        if (opts.thresholdTokens !== void 0) {
+          if (opts.thresholdTokens <= 0 || messageTokens(messages) <= opts.thresholdTokens) {
+            return messages;
+          }
+        } else {
+          const maxChars = opts.maxChars ?? 12e3;
+          const chars = messages.reduce((n, m) => n + m.content.length, 0);
+          if (maxChars <= 0 || chars <= maxChars) return messages;
+        }
+      }
+      const head = messages.slice(0, messages.length - keepRecent);
+      const tail = messages.slice(messages.length - keepRecent);
+      const transcript = head.map((m) => `${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;
+      }
+    };
+    compactHistory = compressHistory;
+    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.";
+    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): ${clean}`, ...done.slice(head.length)];
+      } catch {
+        return done;
+      }
+    };
+  }
+});
+
+// src/images.ts
+var DATA_URL_RE, HTTP_RE, mediaTypeOf, attachmentKind, KIND_LABEL, KIND_HINT, AttachmentsNotSupportedError, ImagesNotSupportedError, unsupportedAttachment, dataOf, toFilePart, toImagePart, modelLabel, REFUSAL_RE, isAttachmentRefusal, isImageRefusal;
+var init_images = __esm({
+  "src/images.ts"() {
+    "use strict";
+    DATA_URL_RE = /^data:([^;,]+)?;base64,(.*)$/s;
+    HTTP_RE = /^https?:\/\//i;
+    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/${ext === "jpg" ? "jpeg" : ext}`;
+        }
+      }
+      return void 0;
+    };
+    attachmentKind = (f) => {
+      const mt = mediaTypeOf(f) ?? "";
+      if (mt.startsWith("image/")) return "image";
+      if (mt === "application/pdf") return "pdf";
+      return "file";
+    };
+    KIND_LABEL = {
+      image: "images",
+      pdf: "PDF files",
+      file: "file attachments"
+    };
+    KIND_HINT = {
+      image: "Remove the image, or switch to a vision-capable model such as Gemini, Claude or GPT-4o/GPT-5.",
+      pdf: "Convert the PDF to text first (e.g. PDF \u2192 Markdown), or switch to a model that reads PDFs such as Gemini, Claude or GPT-4o/GPT-5.",
+      file: "Paste its text instead, or switch to a model that accepts this file type."
+    };
+    AttachmentsNotSupportedError = class extends Error {
+      kind;
+      constructor(model, kind, detail) {
+        super(
+          `The model "${model}" can't take ${KIND_LABEL[kind]}${detail ? ` (${detail})` : ""}. ${KIND_HINT[kind]}`
+        );
+        this.name = "AttachmentsNotSupportedError";
+        this.kind = kind;
+      }
+      toJSON() {
+        return this.message;
+      }
+    };
+    ImagesNotSupportedError = class extends AttachmentsNotSupportedError {
+      constructor(model, detail) {
+        super(model, "image", detail);
+        this.name = "ImagesNotSupportedError";
+      }
+    };
+    unsupportedAttachment = (model, kind, detail) => kind === "image" ? new ImagesNotSupportedError(model, detail) : new AttachmentsNotSupportedError(model, kind, detail);
+    dataOf = (f) => {
+      if (typeof f.data === "string") {
+        const m = DATA_URL_RE.exec(f.data);
+        if (m) return m[2];
+        if (HTTP_RE.test(f.data)) return new URL(f.data);
+      }
+      return f.data;
+    };
+    toFilePart = (f) => {
+      const kind = attachmentKind(f);
+      const mediaType = mediaTypeOf(f) ?? (kind === "image" ? "image" : "application/octet-stream");
+      return {
+        type: "file",
+        data: dataOf(f),
+        mediaType,
+        ...f.name ? { filename: f.name } : {}
+      };
+    };
+    toImagePart = toFilePart;
+    modelLabel = (model) => {
+      if (typeof model === "string") return model;
+      const m = model;
+      return m?.modelId ? `${m.modelId}${m.provider ? ` (${m.provider})` : ""}` : "this model";
+    };
+    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;
+    isAttachmentRefusal = (err) => REFUSAL_RE.test(err instanceof Error ? `${err.name} ${err.message}` : String(err));
+    isImageRefusal = isAttachmentRefusal;
+  }
+});
+
+// src/context-window.ts
+var import_ai2, WEBLLM_DEFAULT_CONTEXT_WINDOW, withWebLLMContextWindow, webLLMContextWindow, contextWindowOf, fitCompactionToWindow, toolDefinitionTokens, TOOL_SEARCH_WINDOW_SHARE, ContextWindowExceededError, OVERFLOW_RE, WEBLLM_SIZES_RE, contextOverflowOf;
+var init_context_window = __esm({
+  "src/context-window.ts"() {
+    "use strict";
+    import_ai2 = require("ai");
+    init_images();
+    init_compress();
+    WEBLLM_DEFAULT_CONTEXT_WINDOW = 4096;
+    withWebLLMContextWindow = (appConfig, modelId, tokens) => ({
+      ...appConfig,
+      model_list: appConfig.model_list.map(
+        (r) => r.model_id === modelId ? { ...r, overrides: { ...r.overrides, context_window_size: tokens } } : r
+      )
+    });
+    webLLMContextWindow = (model) => {
+      const m = model;
+      if (!m || typeof m !== "object" || m.provider !== "web-llm") return void 0;
+      const id = String(m.modelId ?? "");
+      try {
+        const live = m.engine?.loadedModelIdToPipeline?.get?.(id)?.contextWindowSize;
+        if (typeof live === "number" && live > 0) return live;
+      } catch {
+      }
+      const record = m.config?.options?.engineConfig?.appConfig?.model_list?.find(
+        (r) => r.model_id === id
+      );
+      const configured = record?.overrides?.context_window_size;
+      if (typeof configured === "number" && configured > 0) return configured;
+      return /-1k$/i.test(id) ? 1024 : WEBLLM_DEFAULT_CONTEXT_WINDOW;
+    };
+    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;
+    };
+    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)))
+      };
+    };
+    toolDefinitionTokens = async (tools) => {
+      const sizes = await Promise.all(
+        Object.entries(tools).map(async ([name, t]) => {
+          let parameters;
+          try {
+            parameters = await (0, import_ai2.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);
+    };
+    TOOL_SEARCH_WINDOW_SHARE = 0.25;
+    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}"${sizes}. Start a new chat or compact this one, connect fewer tools, or switch to a model with a larger window.`
+        );
+        this.name = "ContextWindowExceededError";
+        this.promptTokens = promptTokens;
+        this.contextWindowTokens = contextWindowTokens;
+      }
+      toJSON() {
+        return this.message;
+      }
+    };
+    OVERFLOW_RE = /prompt tokens exceed context window|context window size|context_length_exceeded|maximum context length|prompt is too long|input token count.*exceeds|exceeds the maximum number of tokens|exceeds the available context size|context length exceeded|too many tokens in (the )?(prompt|input)/i;
+    WEBLLM_SIZES_RE = /number of prompt tokens:\s*(\d+);\s*context window size:\s*(\d+)/i;
+    contextOverflowOf = (err, model) => {
+      if (err instanceof ContextWindowExceededError) return err;
+      const text = err instanceof Error ? `${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 webllm_exports = {};
 __export(webllm_exports, {
@@ -38,14 +396,15 @@ __export(webllm_exports, {
   preloadWebLLMModel: () => preloadWebLLMModel,
   unloadWebLLMModel: () => unloadWebLLMModel
 });
-var import_ai2, isWebGPUAvailable, preloadWebLLMModel, unloadWebLLMModel, createWebLLMModel;
+var import_ai3, isWebGPUAvailable, preloadWebLLMModel, unloadWebLLMModel, createWebLLMModel;
 var init_webllm = __esm({
   "src/providers/webllm.ts"() {
     "use strict";
-    import_ai2 = require("ai");
+    import_ai3 = require("ai");
+    init_context_window();
     isWebGPUAvailable = () => typeof navigator !== "undefined" && "gpu" in navigator;
     preloadWebLLMModel = async (model) => {
-      await (0, import_ai2.generateText)({ model, prompt: "ok", maxOutputTokens: 1 });
+      await (0, import_ai3.generateText)({ model, prompt: "ok", maxOutputTokens: 1 });
     };
     unloadWebLLMModel = async (model) => {
       const engine = model.engine;
@@ -63,7 +422,29 @@ var init_webllm = __esm({
           '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, ...settings } = options ?? {};
+      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
@@ -78,6 +459,7 @@ var init_webllm = __esm({
 var src_exports = {};
 __export(src_exports, {
   AttachmentsNotSupportedError: () => AttachmentsNotSupportedError,
+  ContextWindowExceededError: () => ContextWindowExceededError,
   FILES_STORE: () => FILES_STORE,
   FIND_TOOLS_NAME: () => FIND_TOOLS_NAME,
   ImagesNotSupportedError: () => ImagesNotSupportedError,
@@ -93,10 +475,12 @@ __export(src_exports, {
   SESSIONS_STORE: () => SESSIONS_STORE,
   SKILL_TOOL_NAMES: () => SKILL_TOOL_NAMES,
   TOOL_APPROVAL_MODES: () => TOOL_APPROVAL_MODES,
+  TOOL_SEARCH_WINDOW_SHARE: () => TOOL_SEARCH_WINDOW_SHARE,
   ToolBudgetError: () => ToolBudgetError,
   ToolDeniedError: () => ToolDeniedError,
   VaultCredentialStore: () => VaultCredentialStore,
   VirtualFileSystem: () => VirtualFileSystem,
+  WEBLLM_DEFAULT_CONTEXT_WINDOW: () => WEBLLM_DEFAULT_CONTEXT_WINDOW,
   attachmentKind: () => attachmentKind,
   buildModelFromStage: () => buildModelFromStage,
   bytesToBase64: () => bytesToBase64,
@@ -104,6 +488,8 @@ __export(src_exports, {
   compactHistory: () => compactHistory,
   compactSteps: () => compactSteps,
   compressHistory: () => compressHistory,
+  contextOverflowOf: () => contextOverflowOf,
+  contextWindowOf: () => contextWindowOf,
   createAgent: () => createAgent,
   createFileTools: () => createFileTools,
   createFindToolsTool: () => createFindToolsTool,
@@ -125,6 +511,7 @@ __export(src_exports, {
   estimateMessagesTokens: () => messageTokens,
   estimateTokens: () => estimateTokens,
   executeStep: () => executeStep,
+  fitCompactionToWindow: () => fitCompactionToWindow,
   generate: () => generate,
   generateStructured: () => generateStructured,
   getOrCreateVaultKey: () => getOrCreateVaultKey,
@@ -183,10 +570,13 @@ __export(src_exports, {
   toImagePart: () => toImagePart,
   tokenize: () => tokenize,
   toolCatalogOf: () => toolCatalogOf,
+  toolDefinitionTokens: () => toolDefinitionTokens,
   toolRunContextOf: () => toolRunContextOf,
   unloadWebLLMModel: () => unloadWebLLMModel,
+  webLLMContextWindow: () => webLLMContextWindow,
   withRollingBreakpoint: () => withRollingBreakpoint,
-  withSystem: () => withSystem
+  withSystem: () => withSystem,
+  withWebLLMContextWindow: () => withWebLLMContextWindow
 });
 module.exports = __toCommonJS(src_exports);
 
@@ -233,142 +623,8 @@ var withRollingBreakpoint = (messages, caching) => {
   return out;
 };
 
-// src/llm/generate.ts
-var import_ai = require("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 = {}) => (0, import_ai.generateText)({
-  model,
-  ...callSettings(opts),
-  ...promptOf(opts),
-  tools: opts.tools
-});
-var stream = (model, opts = {}) => (0, import_ai.streamText)({
-  model,
-  ...callSettings(opts),
-  ...promptOf(opts),
-  tools: opts.tools
-});
-var generateStructured = (model, schema, opts = {}) => (0, import_ai.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/config.ts
+init_compress();
 var resolveConfig = (c) => ({
   clientName: c.clientName ?? "agent-web",
   sessionId: c.sessionId ?? "default",
@@ -432,6 +688,11 @@ var ToolBudgetError = class extends Error {
   }
 };
 
+// src/agent/runner.ts
+init_context_window();
+init_util();
+init_compress();
+
 // src/providers/types.ts
 var isDirectModel = (m) => typeof m === "string" || typeof m === "object" && m !== null && "specificationVersion" in m;
 var isProviderSpec = (m) => typeof m === "object" && m !== null && !isDirectModel(m) && typeof m.providerType === "string";
@@ -620,6 +881,7 @@ var buildModelFromStage = async (resolved, credentials, clientName = "agent-web"
 var resolveModel = (input, credentials, clientName) => buildModelFromStage(resolveStage(input, void 0, "model"), credentials, clientName);
 
 // src/prompts.ts
+init_util();
 var numbered = (items, empty) => items.length ? items.map((s, i) => `${i + 1}. ${s}`).join("\n") : empty;
 var stateBlock = (state) => state && state.trim() ? `
 
@@ -744,7 +1006,8 @@ var withSystem = (base, extra) => extra && extra.trim() ? `${extra.trim()}
 ${base}` : base;
 
 // src/skills.ts
-var import_ai3 = require("ai");
+var import_ai4 = require("ai");
+init_util();
 var SKILL_NAME_RE = /^[a-z0-9][a-z0-9-]{0,63}$/;
 var SKILL_TOOL_NAMES = ["load_skill", "read_skill_file"];
 var defineSkill = (skill) => {
@@ -845,9 +1108,9 @@ ${s.content}${files}`, budget);
 var createSkillTools = (skills, onLoad) => {
   const byName = new Map(skills.map((s) => [s.name, s]));
   const known = skills.map((s) => s.name).join(", ");
-  const loadSkill = (0, import_ai3.tool)({
+  const loadSkill = (0, import_ai4.tool)({
     description: "Load the full instructions of a skill from the SKILLS list before doing work it covers.",
-    inputSchema: (0, import_ai3.jsonSchema)({
+    inputSchema: (0, import_ai4.jsonSchema)({
       type: "object",
       properties: { name: { type: "string", description: "The skill name" } },
       required: ["name"],
@@ -864,9 +1127,9 @@ var createSkillTools = (skills, onLoad) => {
       };
     }
   });
-  const readFile = (0, import_ai3.tool)({
+  const readFile = (0, import_ai4.tool)({
     description: "Read one of a skill's bundled files (paths are listed by load_skill).",
-    inputSchema: (0, import_ai3.jsonSchema)({
+    inputSchema: (0, import_ai4.jsonSchema)({
       type: "object",
       properties: {
         name: { type: "string", description: "The skill name" },
@@ -1100,7 +1363,7 @@ var dispatch = async (action, tools, ctx = {}) => {
 };
 
 // src/tools/search.ts
-var import_ai4 = require("ai");
+var import_ai5 = require("ai");
 var FIND_TOOLS_NAME = "find_tools";
 var toolCatalogOf = (tools) => Object.entries(tools).map(([name, t]) => {
   const sep = name.indexOf("__");
@@ -1170,9 +1433,9 @@ var renderSearchCatalog = (catalog, budget = 12e3) => {
   return lines.join("\n");
 };
 var createFindToolsTool = (catalog, onFound, hintOf) => {
-  const t = (0, import_ai4.tool)({
+  const t = (0, import_ai5.tool)({
     description: "Search the full tool catalogue by keywords and make the matching tools callable from the next step on. Use it whenever the tool you need is not in your current tool list.",
-    inputSchema: (0, import_ai4.jsonSchema)({
+    inputSchema: (0, import_ai5.jsonSchema)({
       type: "object",
       properties: {
         query: { type: "string", description: "Keywords: what the tool should do" },
@@ -1206,7 +1469,7 @@ var createFindToolsTool = (catalog, onFound, hintOf) => {
 };
 
 // src/tools/wrap.ts
-var import_ai5 = require("ai");
+var import_ai6 = require("ai");
 var toolRunContextOf = (options) => options?.agentRun;
 var truncated = (s, max) => s.length > max ? `${s.slice(0, max)}\u2026 [truncated ${s.length - max} chars]` : s;
 var toText = (v) => {
@@ -1330,15 +1593,18 @@ var schemaHintOf = async (t) => {
   const own = promptHintOf(t);
   if (own) return own;
   try {
-    const schema = await (0, import_ai5.asSchema)(t.inputSchema).jsonSchema;
+    const schema = await (0, import_ai6.asSchema)(t.inputSchema).jsonSchema;
     return schema && typeof schema === "object" ? objectHint(schema) : void 0;
   } catch {
     return void 0;
   }
 };
 
+// src/agent/executor.ts
+init_context_window();
+
 // src/llm/tool-loop.ts
-var import_ai6 = require("ai");
+var import_ai7 = require("ai");
 
 // src/agent/loop-types.ts
 var BLOCKER = "[BLOCKER]";
@@ -1504,6 +1770,7 @@ var parsePlainText = (raw) => {
 };
 
 // src/llm/context-editing.ts
+init_compress();
 var STUB = (name) => `[${name} result cleared to save context \u2014 call the tool again if you still need it]`;
 var estimate = (messages) => {
   try {
@@ -1548,6 +1815,7 @@ var createToolResultClearer = (opts, onCleared) => {
 };
 
 // src/llm/tool-loop.ts
+init_util();
 var runToolLoop = (model, opts) => opts.mode === "prompted" ? runPrompted(model, opts) : runNative(model, opts);
 var activeOf = (opts) => typeof opts.activeTools === "function" ? opts.activeTools() : opts.activeTools;
 var callExtras = (opts) => ({
@@ -1561,7 +1829,7 @@ var runNative = async (model, opts) => {
   const toolCalls = [];
   const pending = /* @__PURE__ */ new Map();
   let stoppedByBudget = false;
-  const stopWhen = [(0, import_ai6.stepCountIs)(opts.maxSteps ?? 4)];
+  const stopWhen = [(0, import_ai7.stepCountIs)(opts.maxSteps ?? 4)];
   if (opts.shouldStop) {
     const shouldStop = opts.shouldStop;
     stopWhen.push(({ steps }) => {
@@ -1576,7 +1844,7 @@ var runNative = async (model, opts) => {
   const caching = opts.caching?.enabled ? opts.caching : void 0;
   const clear = opts.clearToolResults ? createToolResultClearer(opts.clearToolResults, opts.callbacks?.onToolResultsCleared) : void 0;
   const editMessages = caching || clear;
-  const result = (0, import_ai6.streamText)({
+  const result = (0, import_ai7.streamText)({
     model,
     tools: opts.tools,
     ...initialActive ? { activeTools: initialActive } : {},
@@ -1683,7 +1951,7 @@ var runPrompted = async (model, opts) => {
       break;
     }
     const before = new Set(Object.keys(callable()));
-    const result = await (0, import_ai6.generateText)({
+    const result = await (0, import_ai7.generateText)({
       model,
       ...callExtras(opts),
       messages: opts.caching ? withRollingBreakpoint(messages, opts.caching) : messages,
@@ -1721,6 +1989,9 @@ var runPrompted = async (model, opts) => {
   return { text, toolCalls, usage, stoppedByBudget };
 };
 
+// src/agent/executor.ts
+init_util();
+
 // src/thinking.ts
 var normalize = (setting) => {
   if (setting === void 0 || setting === false) return void 0;
@@ -1732,7 +2003,8 @@ var resolveThinking = (setting) => {
   const cfg = normalize(setting);
   if (!cfg) return {};
   const level = cfg.level ?? "medium";
-  if (level === "none") return { reasoning: "none" };
+  const webllm = (on) => ({ "web-llm": { extra_body: { enable_thinking: on } } });
+  if (level === "none") return { reasoning: "none", providerOptions: webllm(false) };
   const includeThoughts = cfg.includeThoughts !== false;
   const budget = cfg.budgetTokens;
   if (typeof budget === "number" && budget > 0) {
@@ -1740,16 +2012,18 @@ var resolveThinking = (setting) => {
       reasoning: level,
       providerOptions: {
         anthropic: { thinking: { type: "enabled", budgetTokens: budget } },
-        google: { thinkingConfig: { thinkingBudget: budget, includeThoughts } }
+        google: { thinkingConfig: { thinkingBudget: budget, includeThoughts } },
+        ...webllm(true)
       }
     };
   }
-  if (!includeThoughts) return { reasoning: level };
+  if (!includeThoughts) return { reasoning: level, providerOptions: webllm(true) };
   return {
     reasoning: level,
     providerOptions: {
       google: { thinkingConfig: { includeThoughts: true } },
-      openai: { reasoningSummary: "auto" }
+      openai: { reasoningSummary: "auto" },
+      ...webllm(true)
     }
   };
 };
@@ -1771,90 +2045,8 @@ var mergeProviderOptions = (...maps) => {
   return out;
 };
 
-// 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/agent/internal.ts
+init_images();
 var promptFor = (ctx, prompt) => ctx.images?.length ? {
   messages: [
     {
@@ -1977,7 +2169,7 @@ var executeStep = async (ctx, goal, step, index, total, done) => {
       }
     }
   }).catch((err) => {
-    throw imageRefusal(ctx, err, ctx.executorModel) ?? err;
+    throw imageRefusal(ctx, err, ctx.executorModel) ?? contextOverflowOf(err, ctx.executorModel) ?? err;
   });
   ctx.log.debug("executor text:", loop.text);
   const { summary, blocked } = splitBlocker(loop.text);
@@ -2022,6 +2214,11 @@ var replanWanted = async (trigger, result, opts = {}) => {
   ]);
 };
 
+// src/agent/planner.ts
+init_context_window();
+init_generate();
+init_util();
+
 // src/agent/schemas.ts
 var import_zod = require("zod");
 var PlanStepSchema = import_zod.z.object({
@@ -2101,12 +2298,14 @@ ${nudge}` : goal;
       if (ctx.signal?.aborted) throw err;
       const refused = imageRefusal(ctx, err, ctx.plannerModel);
       if (refused) throw refused;
+      const overflow = contextOverflowOf(err, ctx.plannerModel);
+      if (overflow) throw overflow;
       ctx.log.warn("planner: native structured output failed, salvaging:", asMessage(err));
       ctx.emit({ type: "retry", phase: "plan", attempt: 1, error: asMessage(err) });
     }
   }
   const result = await generate(ctx.plannerModel, commonFor("prompted")).catch((err) => {
-    throw imageRefusal(ctx, err, ctx.plannerModel) ?? err;
+    throw imageRefusal(ctx, err, ctx.plannerModel) ?? contextOverflowOf(err, ctx.plannerModel) ?? err;
   });
   const parsed = parsePlannerResponse(result.text);
   ctx.log.debug("planner (prompted) raw:", result.text);
@@ -2126,6 +2325,8 @@ ${nudge}` : goal;
 var asMessage = (err) => err instanceof Error ? err.message : String(err);
 
 // src/agent/replanner.ts
+init_generate();
+init_util();
 var decideReplan = async (ctx, goal, done, remaining) => {
   const state = await ctx.state();
   const active = ctx.activeSkills?.() ?? [];
@@ -2174,6 +2375,9 @@ var decideReplan = async (ctx, goal, done, remaining) => {
 };
 
 // src/agent/synthesizer.ts
+init_context_window();
+init_generate();
+init_util();
 var synthesizeAnswer = async (ctx, goal, done, findings = []) => {
   const state = await ctx.state();
   const active = ctx.activeSkills?.() ?? [];
@@ -2201,7 +2405,7 @@ var synthesizeAnswer = async (ctx, goal, done, findings = []) => {
     }
     if (part.type === "error") {
       const err = part.error instanceof Error ? part.error : new Error(String(part.error));
-      throw imageRefusal(ctx, err, ctx.synthesizerModel) ?? err;
+      throw imageRefusal(ctx, err, ctx.synthesizerModel) ?? contextOverflowOf(err, ctx.synthesizerModel) ?? err;
     }
     if (part.type !== "text-delta") continue;
     const delta = part.text;
@@ -2247,6 +2451,9 @@ var createLogger = (level = "warn", sink = console) => {
   };
 };
 
+// src/agent/runner.ts
+init_images();
+
 // src/providers/capabilities.ts
 var CLOUD = /* @__PURE__ */ new Set([
   "openai",
@@ -2259,10 +2466,11 @@ var CLOUD = /* @__PURE__ */ new Set([
 ]);
 var supportsNativeTools = (p) => CLOUD.has(p);
 var supportsStructuredOutput = (p) => CLOUD.has(p);
-var directBrowserOk = (p) => p === "google" || p === "gateway" || p === "openai-compatible" || p === "anthropic";
+var directBrowserOk = (p) => p !== "web-llm";
 var LOCAL_RE2 = /web-?llm|mlc|browser-ai|transformers|built-?in/i;
+var LOCAL_VISION_RE = /vision|[-_]vl\b|[-_]vl[-_]|llava|smolvlm|moondream|gemma-?3n/i;
 var TEXT_ONLY_MODEL_RE = /gpt-3\.5|o1-mini|o3-mini|deepseek|text-(davinci|embedding)|embedding|babbage|davinci|codestral/i;
-var VISION_PROVIDER_RE = /^(google|anthropic|openai|azure|xai|vertex|bedrock|gemini)/i;
+var VISION_PROVIDER_RE = /^(google|anthropic|openai|azure|xai|vertex|bedrock|gemini|moonshot|mistral)/i;
 var PDF_PROVIDER_RE = /^(google|anthropic|openai|azure|vertex|bedrock|gemini)/i;
 var supportsImages = (model) => {
   if (typeof model === "string") {
@@ -2273,14 +2481,16 @@ var supportsImages = (model) => {
   if (!model || typeof model !== "object") return void 0;
   const provider = String(model.provider ?? "");
   const id = String(model.modelId ?? "");
-  if (LOCAL_RE2.test(provider)) return false;
-  if (/deepseek/i.test(provider) || TEXT_ONLY_MODEL_RE.test(id)) return false;
+  if (LOCAL_RE2.test(provider)) return LOCAL_VISION_RE.test(id);
+  if (/deepseek/i.test(provider)) return /flash/i.test(id);
+  if (TEXT_ONLY_MODEL_RE.test(id)) return false;
   return VISION_PROVIDER_RE.test(provider) ? true : void 0;
 };
 var supportsPdf = (model) => {
   const vision = supportsImages(model);
   if (vision === false) return false;
   const provider = typeof model === "string" ? model.split("/")[0] ?? "" : String(model?.provider ?? "");
+  if (LOCAL_RE2.test(provider) || /moonshot/i.test(provider)) return false;
   return PDF_PROVIDER_RE.test(provider) ? true : void 0;
 };
 
@@ -2317,6 +2527,10 @@ var createAgent = async (config) => {
   ]);
   const plannerMode = selectToolMode(plannerModel, cfg.toolMode);
   const executorMode = selectToolMode(executorModel, cfg.toolMode);
+  const modelWindows = [executorModel, plannerModel, synthesizerModel].map(contextWindowOf).filter((w) => w !== void 0);
+  const modelWindow = modelWindows.length ? Math.min(...modelWindows) : void 0;
+  cfg.compaction = fitCompactionToWindow(cfg.compaction, config.compaction, modelWindow);
+  const byTokens = Boolean(config.compaction) || modelWindow !== void 0;
   const baseTools = filterTools(
     config.tools ?? {},
     config.availableTools,
@@ -2329,7 +2543,8 @@ var createAgent = async (config) => {
     skillNames.add(s.name);
   }
   const catalog = toolCatalogOf(baseTools);
-  const strategy = cfg.toolSelectionStrategy === "auto" ? catalog.length > cfg.toolSearchThreshold ? "search" : "all" : cfg.toolSelectionStrategy;
+  const toolTokens = cfg.toolSelectionStrategy === "auto" ? await toolDefinitionTokens(baseTools) : 0;
+  const strategy = cfg.toolSelectionStrategy === "auto" ? catalog.length > cfg.toolSearchThreshold || toolTokens > cfg.compaction.contextWindowTokens * TOOL_SEARCH_WINDOW_SHARE ? "search" : "all" : cfg.toolSelectionStrategy;
   const reserved = [
     ...skills.length ? SKILL_TOOL_NAMES : [],
     ...strategy === "search" ? [FIND_TOOLS_NAME] : []
@@ -2355,14 +2570,15 @@ var createAgent = async (config) => {
   const caching = resolveCaching(config.promptCaching);
   const textOnly = executorMode === "prompted";
   const accepts = {
-    image: config.inputs?.images ?? config.vision ?? (textOnly ? false : supportsImages(executorModel)),
+    // A local vision model (the prompted path carries images too) says so by its id.
+    image: config.inputs?.images ?? config.vision ?? (textOnly ? supportsImages(executorModel) === true : supportsImages(executorModel)),
     pdf: config.inputs?.pdf ?? (textOnly ? false : supportsPdf(executorModel)),
     file: config.inputs?.files ?? (textOnly ? false : void 0)
   };
   const approval = createApprovalState(config.toolApproval);
   const log = createLogger(cfg.logLevel, config.logger);
   log.info(
-    `agent ready \u2014 ${catalog.length} tool(s) (${strategy}), skills: [${[...skillNames].join(", ") || "none"}], planner mode: ${plannerMode}, executor mode: ${executorMode}, approval: ${approval.mode}`
+    `agent ready \u2014 ${catalog.length} tool(s) (${strategy}, ~${toolTokens || "?"} tokens), window ${cfg.compaction.contextWindowTokens}${modelWindow ? " (the model\u2019s)" : ""}, skills: [${[...skillNames].join(", ") || "none"}], planner mode: ${plannerMode}, executor mode: ${executorMode}, approval: ${approval.mode}`
   );
   const run = async (goal, opts = {}) => {
     const emit = (event) => {
@@ -2473,7 +2689,7 @@ var createAgent = async (config) => {
       } catch {
       }
     }
-    if (config.memory && config.compaction && cfg.compaction.auto && history.length > 0) {
+    if (config.memory && byTokens && cfg.compaction.auto && history.length > 0) {
       const before = messageTokens(history);
       if (before > cfg.compaction.thresholdTokens) {
         const compacted = await compressHistory(history, synthesizerModel, {
@@ -2685,7 +2901,8 @@ var createAgent = async (config) => {
         }
       }
       if (isAborted()) return stop();
-      let summary = "Done \u2014 the changes have been applied.";
+      const stepWords = result.trace.map((r) => r.summary.trim()).filter((t) => t && !/^(Executed \d+ tool call\(s\)\.|Step produced no output\.)$/.test(t));
+      let summary = stepWords.length ? stepWords.join("\n\n") : "Done \u2014 the changes have been applied.";
       if (cfg.synthesize) {
         try {
           const synth = await synthesizeAnswer(ctx, text, done, findings);
@@ -2700,12 +2917,12 @@ var createAgent = async (config) => {
       emit({ type: "final", text: summary });
       await remember({ role: "assistant", content: summary });
       result.final = summary;
-      if (config.memory && (config.compaction ? cfg.compaction.auto : cfg.compressAfterChars > 0)) {
+      if (config.memory && (byTokens ? cfg.compaction.auto : cfg.compressAfterChars > 0)) {
         try {
           const transcript = await config.memory.load(sessionId);
           const before = messageTokens(transcript);
           const compacted = await compressHistory(transcript, synthesizerModel, {
-            ...config.compaction ? {
+            ...byTokens ? {
               thresholdTokens: cfg.compaction.thresholdTokens,
               keepRecent: cfg.compaction.keepRecentTurns,
               summaryMaxTokens: cfg.budgets.compaction
@@ -2777,6 +2994,7 @@ var createAgent = async (config) => {
     },
     listTools: () => catalog.map((e) => ({ ...e })),
     toolStrategy: strategy,
+    contextWindowTokens: cfg.compaction.contextWindowTokens,
     skills: skills.map((s) => ({ name: s.name, description: s.description })),
     capabilities: { images: accepts.image, pdf: accepts.pdf, files: accepts.file },
     models: { planner: plannerModel, executor: executorModel, synthesizer: synthesizerModel }
@@ -2784,7 +3002,8 @@ var createAgent = async (config) => {
 };
 
 // src/subagent/tool.ts
-var import_ai7 = require("ai");
+var import_ai8 = require("ai");
+init_util();
 
 // src/subagent/protocol.ts
 var toCloneable = (value) => {
@@ -2886,7 +3105,7 @@ var createSubagentTool = (opts) => {
     Object.entries(opts.tools ?? {}).map(async ([name, t2]) => ({
       name,
       description: typeof t2.description === "string" ? t2.description : "",
-      inputSchema: toCloneable(await (0, import_ai7.asSchema)(t2.inputSchema).jsonSchema),
+      inputSchema: toCloneable(await (0, import_ai8.asSchema)(t2.inputSchema).jsonSchema),
       readOnly: isReadOnlyTool(t2)
     }))
   );
@@ -2959,9 +3178,9 @@ var createSubagentTool = (opts) => {
       safePost(worker, { type: "run", task, config, proxied });
     });
   };
-  const t = (0, import_ai7.tool)({
+  const t = (0, import_ai8.tool)({
     description: opts.description,
-    inputSchema: (0, import_ai7.jsonSchema)({
+    inputSchema: (0, import_ai8.jsonSchema)({
       type: "object",
       properties: {
         task: {
@@ -3021,7 +3240,7 @@ var createSubagentTool = (opts) => {
 };
 
 // src/subagent/worker.ts
-var import_ai8 = require("ai");
+var import_ai9 = require("ai");
 
 // src/storage/db.ts
 var import_idb = require("idb");
@@ -3188,9 +3407,9 @@ var serveSubagentWorker = (opts = {}) => {
   const pending = /* @__PURE__ */ new Map();
   let callSeq = 0;
   const proxy = (spec) => {
-    const t = (0, import_ai8.dynamicTool)({
+    const t = (0, import_ai9.dynamicTool)({
       description: spec.description,
-      inputSchema: (0, import_ai8.jsonSchema)(spec.inputSchema),
+      inputSchema: (0, import_ai9.jsonSchema)(spec.inputSchema),
       execute: (input, options) => new Promise((resolve, reject) => {
         const callId = `call-${callSeq += 1}`;
         const signal = options?.abortSignal;
@@ -3268,6 +3487,9 @@ var serveSubagentWorker = (opts = {}) => {
   });
 };
 
+// src/index.ts
+init_images();
+
 // src/files/vfs.ts
 var MIME = {
   txt: "text/plain",
@@ -3417,15 +3639,15 @@ var VirtualFileSystem = class {
 };
 
 // src/files/tools.ts
-var import_ai9 = require("ai");
+var import_ai10 = require("ai");
 var createFileTools = (vfs, opts = {}) => {
   const p = opts.prefix ?? "fs_";
   const maxRead = opts.maxReadChars ?? 5e4;
   const tools = {
     [`${p}list`]: markReadOnly(
-      (0, import_ai9.tool)({
+      (0, import_ai10.tool)({
         description: "List the files in the workspace (attachments the user added, files you wrote). Optionally only under a directory prefix.",
-        inputSchema: (0, import_ai9.jsonSchema)({
+        inputSchema: (0, import_ai10.jsonSchema)({
           type: "object",
           properties: { prefix: { type: "string", description: 'Directory prefix, e.g. "/docs"' } },
           additionalProperties: false
@@ -3438,9 +3660,9 @@ var createFileTools = (vfs, opts = {}) => {
       })
     ),
     [`${p}read`]: markReadOnly(
-      (0, import_ai9.tool)({
+      (0, import_ai10.tool)({
         description: "Read a text file from the workspace by its path.",
-        inputSchema: (0, import_ai9.jsonSchema)({
+        inputSchema: (0, import_ai10.jsonSchema)({
           type: "object",
           properties: { path: { type: "string", description: 'Absolute path, e.g. "/notes.md"' } },
           required: ["path"],
@@ -3462,9 +3684,9 @@ var createFileTools = (vfs, opts = {}) => {
     )
   };
   if (opts.readOnly) return tools;
-  tools[`${p}write`] = (0, import_ai9.tool)({
+  tools[`${p}write`] = (0, import_ai10.tool)({
     description: "Write a text file to the workspace (creates or replaces it). Use it for reports, drafts, data the user should be able to open.",
-    inputSchema: (0, import_ai9.jsonSchema)({
+    inputSchema: (0, import_ai10.jsonSchema)({
       type: "object",
       properties: {
         path: { type: "string", description: 'Absolute path, e.g. "/reports/summary.md"' },
@@ -3479,9 +3701,9 @@ var createFileTools = (vfs, opts = {}) => {
       return { written: info.path, size: info.size };
     }
   });
-  tools[`${p}delete`] = (0, import_ai9.tool)({
+  tools[`${p}delete`] = (0, import_ai10.tool)({
     description: "Delete a file from the workspace.",
-    inputSchema: (0, import_ai9.jsonSchema)({
+    inputSchema: (0, import_ai10.jsonSchema)({
       type: "object",
       properties: { path: { type: "string" } },
       required: ["path"],
@@ -3494,12 +3716,15 @@ var createFileTools = (vfs, opts = {}) => {
 
 // src/index.ts
 init_webllm();
+init_context_window();
+init_generate();
+init_util();
 
 // src/tools/define.ts
-var import_ai10 = require("ai");
+var import_ai11 = require("ai");
 var defineTool = (config) => {
   const { promptHint, readOnly, ...rest } = config;
-  const t = (0, import_ai10.tool)(rest);
+  const t = (0, import_ai11.tool)(rest);
   if (promptHint) t.promptHint = promptHint;
   if (readOnly) t.readOnly = true;
   return t;
@@ -3554,9 +3779,13 @@ var IndexedDBStore = class {
     return (await db.getAllKeys(SESSIONS_STORE)).map(String);
   }
 };
+
+// src/index.ts
+init_compress();
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
   AttachmentsNotSupportedError,
+  ContextWindowExceededError,
   FILES_STORE,
   FIND_TOOLS_NAME,
   ImagesNotSupportedError,
@@ -3572,10 +3801,12 @@ var IndexedDBStore = class {
   SESSIONS_STORE,
   SKILL_TOOL_NAMES,
   TOOL_APPROVAL_MODES,
+  TOOL_SEARCH_WINDOW_SHARE,
   ToolBudgetError,
   ToolDeniedError,
   VaultCredentialStore,
   VirtualFileSystem,
+  WEBLLM_DEFAULT_CONTEXT_WINDOW,
   attachmentKind,
   buildModelFromStage,
   bytesToBase64,
@@ -3583,6 +3814,8 @@ var IndexedDBStore = class {
   compactHistory,
   compactSteps,
   compressHistory,
+  contextOverflowOf,
+  contextWindowOf,
   createAgent,
   createFileTools,
   createFindToolsTool,
@@ -3604,6 +3837,7 @@ var IndexedDBStore = class {
   estimateMessagesTokens,
   estimateTokens,
   executeStep,
+  fitCompactionToWindow,
   generate,
   generateStructured,
   getOrCreateVaultKey,
@@ -3662,9 +3896,12 @@ var IndexedDBStore = class {
   toImagePart,
   tokenize,
   toolCatalogOf,
+  toolDefinitionTokens,
   toolRunContextOf,
   unloadWebLLMModel,
+  webLLMContextWindow,
   withRollingBreakpoint,
-  withSystem
+  withSystem,
+  withWebLLMContextWindow
 });
 //# sourceMappingURL=index.cjs.map