@genesislcap/ai-assistant 15.14.2 → 15.15.0

package/dist/chat-driver.mjs

@@ -1154,6 +1154,14 @@ function vendorOfModel(modelId) {
   return void 0;
 }
 
+// ../../foundation-ai/dist/esm/utils/image-mime.js
+function normalizeImageMime(mimeType) {
+  return mimeType === "image/jpg" ? "image/jpeg" : mimeType;
+}
+function stripImageDataUrl(data) {
+  return data.replace(/^data:[^,]*;base64,/, "");
+}
+
 // ../../foundation-ai/dist/esm/utils/temperature.js
 var DEFAULT_ANCHOR = 0.5;
 function scaleTemperature(normalized, { defaultTemp, maxTemp }) {
@@ -1762,8 +1770,28 @@ function anthropicThinking(model, policy) {
   if (policy === "auto") {
     return supportsAdaptiveThinking(model) ? adaptive : void 0;
   }
+  if (policy !== void 0) {
+    return supportsAdaptiveThinking(model) ? adaptive : void 0;
+  }
   return defaultsToThinking ? adaptive : void 0;
 }
+function supportsEffort(model) {
+  return model === "claude-fable-5" || model === "claude-opus-4-8" || model === "claude-sonnet-5";
+}
+var ANTHROPIC_EFFORT = {
+  minimal: "low",
+  low: "low",
+  medium: "medium",
+  high: "high",
+  max: "max"
+};
+function anthropicEffort(model, policy) {
+  if (policy === void 0 || policy === "auto" || policy === "off")
+    return void 0;
+  if (!supportsEffort(model) || !supportsAdaptiveThinking(model))
+    return void 0;
+  return ANTHROPIC_EFFORT[policy];
+}
 var ANTHROPIC_PROVIDER_KEY = "anthropic";
 var ResponseTruncatedError = class extends Error {
   constructor(model, maxTokens, outputTokens, toolNames) {
@@ -1788,11 +1816,42 @@ var AnthropicTransport = class _AnthropicTransport {
   warnIfThinkingUnclampable(policy) {
     if (this.warnedThinkingClamped)
       return;
-    const clamped = policy === "off" && thinkingIsMandatory(this.model) || policy === "auto" && !supportsAdaptiveThinking(this.model);
-    if (!clamped)
+    const message = this.clampWarning(policy);
+    if (!message)
       return;
     this.warnedThinkingClamped = true;
-    logger.warn(policy === "off" ? `AnthropicTransport: thinkingPolicy 'off' ignored \u2014 ${this.model} always thinks and rejects an explicit disable. Reasoning tokens are still billed as output; switch model if you need them gone.` : `AnthropicTransport: thinkingPolicy 'auto' ignored \u2014 ${this.model} does not support adaptive thinking, so this turn runs without reasoning. Use a Sonnet or Opus tier if the agent needs it.`);
+    logger.warn(message);
+  }
+  /**
+   * The warning for a policy this model cannot honour, or `undefined` when it can.
+   *
+   * An exhaustive `switch` with a `never` check rather than a condition chain: a new
+   * {@link ChatThinkingPolicy} member must be considered HERE or the build fails. The condition
+   * chain this replaced would have let every graded level clamp in silence, which is the one
+   * outcome the type's own docs promise callers never happens.
+   */
+  clampWarning(policy) {
+    switch (policy) {
+      case void 0:
+        return void 0;
+      case "off":
+        return thinkingIsMandatory(this.model) ? `AnthropicTransport: thinkingPolicy 'off' ignored \u2014 ${this.model} always thinks and rejects an explicit disable. Reasoning tokens are still billed as output; switch model if you need them gone.` : void 0;
+      case "auto":
+        return supportsAdaptiveThinking(this.model) ? void 0 : `AnthropicTransport: thinkingPolicy 'auto' ignored \u2014 ${this.model} does not support adaptive thinking, so this turn runs without reasoning. Use a Sonnet or Opus tier if the agent needs it.`;
+      case "minimal":
+      case "low":
+      case "medium":
+      case "high":
+      case "max":
+        if (!supportsAdaptiveThinking(this.model) || !supportsEffort(this.model)) {
+          return `AnthropicTransport: thinkingPolicy '${policy}' ignored \u2014 ${this.model} predates output_config.effort, so this turn runs at its default posture. Use Opus 4.8, Sonnet 5, or Fable 5 to set reasoning depth.`;
+        }
+        return void 0;
+      default: {
+        const exhaustive = policy;
+        return exhaustive;
+      }
+    }
   }
   constructor(config = {}) {
     var _a, _b, _c, _d;
@@ -1854,9 +1913,7 @@ var AnthropicTransport = class _AnthropicTransport {
       const useNative = responseSchema != null && supportsNativeStructuredOutput(this.model);
       const useForcedTool = responseSchema != null && !useNative;
       if (useNative) {
-        body.output_config = {
-          format: { type: "json_schema", schema: responseSchema }
-        };
+        body.output_config = Object.assign(Object.assign({}, body.output_config), { format: { type: "json_schema", schema: responseSchema } });
       } else if (useForcedTool) {
         body.tools = [
           {
@@ -1880,7 +1937,7 @@ var AnthropicTransport = class _AnthropicTransport {
   // ── ChatTransport (multi-turn chat) ────────────────────────────────────
   sendChatMessage(history, userMessage, options) {
     return __awaiter(this, void 0, void 0, function* () {
-      var _a, _b, _c, _d;
+      var _a, _b, _c, _d, _e;
       const reachableModels = /* @__PURE__ */ new Set([
         this.model,
         ...((_a = options === null || options === void 0 ? void 0 : options.fallbacks) !== null && _a !== void 0 ? _a : []).map((f) => f.model)
@@ -1888,15 +1945,16 @@ var AnthropicTransport = class _AnthropicTransport {
       const messages = this.toAnthropicMessages(history, userMessage, options === null || options === void 0 ? void 0 : options.attachments, reachableModels);
       const body = {
         model: this.model,
-        max_tokens: this.maxTokens,
+        // Per-turn override wins over the instance default; see ChatRequestOptions.maxTokens.
+        max_tokens: (_b = options === null || options === void 0 ? void 0 : options.maxTokens) !== null && _b !== void 0 ? _b : this.maxTokens,
         messages
       };
       if (options === null || options === void 0 ? void 0 : options.systemPrompt)
         body.system = options.systemPrompt;
-      if ((_b = options === null || options === void 0 ? void 0 : options.tools) === null || _b === void 0 ? void 0 : _b.length) {
+      if ((_c = options === null || options === void 0 ? void 0 : options.tools) === null || _c === void 0 ? void 0 : _c.length) {
         body.tools = options.tools.map((t) => Object.assign({ name: t.name, description: t.description, input_schema: t.enforceSchema ? enforceAnthropicToolSchema(t.parameters, t.name) : t.parameters }, t.enforceSchema ? { strict: true } : {}));
       }
-      if ((_c = body.tools) === null || _c === void 0 ? void 0 : _c.length) {
+      if ((_d = body.tools) === null || _d === void 0 ? void 0 : _d.length) {
         const toolChoice = toAnthropicToolChoice(options === null || options === void 0 ? void 0 : options.toolChoice);
         if (toolChoice)
           body.tool_choice = toolChoice;
@@ -1905,6 +1963,9 @@ var AnthropicTransport = class _AnthropicTransport {
       const thinking = anthropicThinking(this.model, options === null || options === void 0 ? void 0 : options.thinkingPolicy);
       if (thinking)
         body.thinking = thinking;
+      const effort = anthropicEffort(this.model, options === null || options === void 0 ? void 0 : options.thinkingPolicy);
+      if (effort)
+        body.output_config = Object.assign(Object.assign({}, body.output_config), { effort });
       const thinkingEnabled = (thinking === null || thinking === void 0 ? void 0 : thinking.type) === "adaptive";
       if ((options === null || options === void 0 ? void 0 : options.temperature) != null && !rejectsSamplingParams(this.model) && !thinkingEnabled) {
         body.temperature = scaleTemperature(options.temperature, {
@@ -1913,11 +1974,9 @@ var AnthropicTransport = class _AnthropicTransport {
         });
       }
       if ((options === null || options === void 0 ? void 0 : options.responseSchema) && supportsNativeStructuredOutput(this.model)) {
-        body.output_config = {
-          format: { type: "json_schema", schema: options.responseSchema }
-        };
+        body.output_config = Object.assign(Object.assign({}, body.output_config), { format: { type: "json_schema", schema: options.responseSchema } });
       }
-      if ((_d = options === null || options === void 0 ? void 0 : options.fallbacks) === null || _d === void 0 ? void 0 : _d.length) {
+      if ((_e = options === null || options === void 0 ? void 0 : options.fallbacks) === null || _e === void 0 ? void 0 : _e.length) {
         body.fallbacks = options.fallbacks.map((f) => f.maxTokens != null ? { model: f.model, max_tokens: f.maxTokens } : { model: f.model });
       }
       if (options === null || options === void 0 ? void 0 : options.cachePolicy) {
@@ -1927,7 +1986,7 @@ var AnthropicTransport = class _AnthropicTransport {
         this.appendTailContext(body, options.tailContext);
       }
       const response = yield this.post(body, options === null || options === void 0 ? void 0 : options.signal);
-      return this.fromAnthropicResponse(response);
+      return this.fromAnthropicResponse(response, body.max_tokens);
     });
   }
   /**
@@ -2020,7 +2079,7 @@ var AnthropicTransport = class _AnthropicTransport {
    * the payload tidy.
    */
   toAnthropicMessages(history, userMessage, attachments, reachableModels = /* @__PURE__ */ new Set([this.model])) {
-    var _a, _b, _c;
+    var _a, _b, _c, _d;
     const messages = [];
     const pushBlock = (role, block) => {
       const last = messages[messages.length - 1];
@@ -2030,18 +2089,54 @@ var AnthropicTransport = class _AnthropicTransport {
       }
       messages.push({ role, content: [block] });
     };
+    const splitAttachments = (atts) => {
+      const images = [];
+      const texts = [];
+      for (const att of atts) {
+        if (att.kind === "image") {
+          images.push({
+            type: "image",
+            source: {
+              type: "base64",
+              media_type: normalizeImageMime(att.mimeType),
+              data: stripImageDataUrl(att.data)
+            }
+          });
+        } else {
+          texts.push({ type: "text", text: `[File: ${att.name}]
+${att.content}` });
+        }
+      }
+      return { images, texts };
+    };
     for (const msg of history) {
       if (msg.role === "system" || msg.role === "system-event" || msg.role === "synthetic-user" || msg.role === "compacted-summary" || msg.category === "reasoning" || msg.category === "narration" || msg.thinking)
         continue;
       if (msg.toolResult) {
+        const resultImages = (_a = msg.toolResult.attachments) !== null && _a !== void 0 ? _a : [];
         pushBlock("user", {
           type: "tool_result",
           tool_use_id: msg.toolResult.toolCallId,
-          content: msg.toolResult.content
+          // Bare string when there are no images — the historical shape, byte for byte.
+          content: resultImages.length ? [
+            // Text block only when there IS text. A handler can return an image with no
+            // caption (`{ content: '', attachments: [img] }` — the mainline vision-agent
+            // "just look at this" shape), and Anthropic 400s an empty text block, live and on
+            // replay. Guarded like every user-turn path; the array is then image-only.
+            ...msg.toolResult.content ? [{ type: "text", text: msg.toolResult.content }] : [],
+            ...resultImages.map((att) => ({
+              type: "image",
+              source: {
+                type: "base64",
+                media_type: normalizeImageMime(att.mimeType),
+                data: stripImageDataUrl(att.data)
+              }
+            }))
+          ] : msg.toolResult.content
         });
         continue;
       }
-      if ((_a = msg.toolCalls) === null || _a === void 0 ? void 0 : _a.length) {
+      if ((_b = msg.toolCalls) === null || _b === void 0 ? void 0 : _b.length) {
         for (const block of this.reasoningToReplay(msg.toolCalls[0], reachableModels)) {
           pushBlock("assistant", block);
         }
@@ -2053,30 +2148,33 @@ var AnthropicTransport = class _AnthropicTransport {
             type: "tool_use",
             id: tc.id,
             name: tc.name,
-            input: (_b = tc.args) !== null && _b !== void 0 ? _b : {}
+            input: (_c = tc.args) !== null && _c !== void 0 ? _c : {}
           });
         }
         continue;
       }
       const role = msg.role === "user" ? "user" : "assistant";
-      if (role === "user" && ((_c = msg.attachments) === null || _c === void 0 ? void 0 : _c.length)) {
-        pushBlock(role, { type: "text", text: msg.content });
-        for (const att of msg.attachments) {
-          pushBlock(role, { type: "text", text: `[File: ${att.name}]
-${att.content}` });
-        }
+      if (role === "user" && ((_d = msg.attachments) === null || _d === void 0 ? void 0 : _d.length)) {
+        const { images, texts } = splitAttachments(msg.attachments);
+        for (const image of images)
+          pushBlock(role, image);
+        if (msg.content)
+          pushBlock(role, { type: "text", text: msg.content });
+        for (const text of texts)
+          pushBlock(role, text);
       } else if (msg.content) {
         pushBlock(role, { type: "text", text: msg.content });
       }
     }
     if (userMessage || (attachments === null || attachments === void 0 ? void 0 : attachments.length)) {
+      const { images, texts } = splitAttachments(attachments !== null && attachments !== void 0 ? attachments : []);
+      for (const image of images)
+        pushBlock("user", image);
       if (userMessage) {
         pushBlock("user", { type: "text", text: userMessage });
       }
-      for (const att of attachments !== null && attachments !== void 0 ? attachments : []) {
-        pushBlock("user", { type: "text", text: `[File: ${att.name}]
-${att.content}` });
-      }
+      for (const text of texts)
+        pushBlock("user", text);
     }
     return messages;
   }
@@ -2183,7 +2281,7 @@ ${att.content}` });
     const cacheCreation = (_c = usage.cache_creation_input_tokens) !== null && _c !== void 0 ? _c : breakdown ? ((_d = breakdown.ephemeral_5m_input_tokens) !== null && _d !== void 0 ? _d : 0) + cacheCreation1h : 0;
     return this.logTokenUsage(model, (_e = usage.input_tokens) !== null && _e !== void 0 ? _e : 0, (_f = usage.output_tokens) !== null && _f !== void 0 ? _f : 0, cacheRead, cacheCreation, cacheCreation1h);
   }
-  fromAnthropicResponse(response) {
+  fromAnthropicResponse(response, effectiveMaxTokens = this.maxTokens) {
     var _a, _b, _c, _d, _e, _f, _g;
     let inputTokens;
     let outputTokens;
@@ -2260,7 +2358,7 @@ ${att.content}` });
       toolCalls[0].providerMetadata = Object.assign(Object.assign({}, toolCalls[0].providerMetadata), { [ANTHROPIC_PROVIDER_KEY]: state });
     }
     if (response.stop_reason === "max_tokens" && toolCalls.length > 0) {
-      throw new ResponseTruncatedError(this.model, this.maxTokens, outputTokens, toolCalls.map((tc) => tc.name));
+      throw new ResponseTruncatedError(this.model, effectiveMaxTokens, outputTokens, toolCalls.map((tc) => tc.name));
     }
     const reasoning = thoughtParts.join("") || void 0;
     const base = toolCalls.length > 0 ? { role: "assistant", content: textParts.join(""), reasoning, toolCalls } : { role: "assistant", content: textParts.join(""), reasoning };
@@ -2586,7 +2684,43 @@ var GEMINI_THINKING_DISABLEABLE = [
   "gemini-2.5-flash",
   "gemini-2.5-flash-lite"
 ];
+var GEMINI_MEDIA_RESOLUTION_TIERS = [
+  "gemini-3.5-flash",
+  "gemini-3.1-flash-lite",
+  "gemini-3.1-pro-preview"
+];
+var GEMINI_THINKING_LEVEL_TIERS = [
+  "gemini-3.5-flash",
+  "gemini-3.1-flash-lite",
+  "gemini-3.1-pro-preview"
+];
+var GEMINI_THINKING_LEVEL = {
+  minimal: "minimal",
+  low: "low",
+  medium: "medium",
+  high: "high",
+  max: "high"
+};
+var GEMINI_COARSE_LEVEL_TIERS = [
+  "gemini-3.5-flash",
+  "gemini-3.1-flash-lite",
+  "gemini-3.1-pro-preview"
+];
+var GEMINI_COARSE_LEVEL = {
+  minimal: "low",
+  low: "low",
+  medium: "high",
+  high: "high"
+};
+var isGradedLevel = (p) => p !== void 0 && p !== "auto" && p !== "off";
 function geminiThinkingConfig(model, policy) {
+  if (isGradedLevel(policy) && GEMINI_THINKING_LEVEL_TIERS.includes(model)) {
+    const level = GEMINI_THINKING_LEVEL[policy];
+    return {
+      includeThoughts: true,
+      thinkingLevel: GEMINI_COARSE_LEVEL_TIERS.includes(model) ? GEMINI_COARSE_LEVEL[level] : level
+    };
+  }
   if (policy === "off" && GEMINI_THINKING_DISABLEABLE.includes(model)) {
     return { includeThoughts: false, thinkingBudget: 0 };
   }
@@ -2597,6 +2731,24 @@ function geminiThinkingConfig(model, policy) {
 }
 var SKIP_SIGNATURE = "skip_thought_signature_validator";
 var GEMINI_PROVIDER_KEY = "gemini";
+function splitAttachmentParts(atts) {
+  const imageParts = [];
+  const textParts = [];
+  for (const att of atts) {
+    if (att.kind === "image") {
+      imageParts.push({
+        inlineData: {
+          mimeType: normalizeImageMime(att.mimeType),
+          data: stripImageDataUrl(att.data)
+        }
+      });
+    } else {
+      textParts.push({ text: `[File: ${att.name}]
+${att.content}` });
+    }
+  }
+  return { imageParts, textParts };
+}
 var MalformedFunctionCallError = class extends Error {
   constructor(finishMessage) {
     super("Gemini returned MALFORMED_FUNCTION_CALL");
@@ -2606,21 +2758,32 @@ var MalformedFunctionCallError = class extends Error {
 };
 var GeminiTransport = class _GeminiTransport {
   /**
-   * Warn once when a requested policy is silently clamped. Turning thinking off is a *cost*
-   * decision, so a caller who asked for it and kept paying for reasoning tokens needs to hear
-   * about it — but only once, not per turn.
+   * Warn once when a requested policy cannot be honoured on this model — a *cost* decision the
+   * caller made that the wire silently reversed, so they hear about it, but only once, not per
+   * turn. Two shapes reach here (the Anthropic twin warns on the same two):
    *
-   * Only `'off'` can actually be denied. `'auto'` is already what an omitted budget produces on
-   * every model here — dynamic thinking is the documented default — so warning that it was
-   * "ignored" would be false, and latching on it would spend the one warning the genuinely
-   * unhonourable `'off'` needs.
+   * - `'off'` on a tier that can't disable thinking — only the flash tiers accept a zero thinking
+   *   budget; everything else keeps paying for reasoning it was told to stop. (`'auto'` is NOT this
+   *   case: it is already what an omitted budget produces on every model here, so warning it was
+   *   "ignored" would be false and would spend the one warning the genuine cases need.)
+   * - A graded depth (`minimal`…`max`) on a 2.5 tier — those speak the numeric-budget dialect,
+   *   not the Gemini 3 `thinkingLevel` enum, so the level is dropped and the turn runs at the
+   *   default posture. Without this, a caller asking for `high` on `gemini-2.5-pro` and quietly
+   *   getting the default is left to wonder why.
    */
   warnIfThinkingUnclampable(policy) {
-    if (policy !== "off" || GEMINI_THINKING_DISABLEABLE.includes(this.model) || this.warnedThinkingClamped) {
+    if (this.warnedThinkingClamped) {
       return;
     }
-    this.warnedThinkingClamped = true;
-    logger.warn(`GeminiTransport: thinkingPolicy 'off' ignored \u2014 ${this.model} runs its default thinking posture and takes no budget we can safely set. Reasoning tokens are still billed at the candidate rate; use a flash tier if you need them gone.`);
+    if (policy === "off" && !GEMINI_THINKING_DISABLEABLE.includes(this.model)) {
+      this.warnedThinkingClamped = true;
+      logger.warn(`GeminiTransport: thinkingPolicy 'off' ignored \u2014 ${this.model} runs its default thinking posture and takes no budget we can safely set. Reasoning tokens are still billed at the candidate rate; use a flash tier if you need them gone.`);
+      return;
+    }
+    if (isGradedLevel(policy) && !GEMINI_THINKING_LEVEL_TIERS.includes(this.model)) {
+      this.warnedThinkingClamped = true;
+      logger.warn(`GeminiTransport: thinkingPolicy '${policy}' ignored \u2014 ${this.model} takes a numeric thinking budget, not the graded 'thinkingLevel' the Gemini 3 tiers use, so this turn runs at its default reasoning posture. Use a Gemini 3 tier to set reasoning depth.`);
+    }
   }
   constructor(config = {}) {
     var _a, _b, _c;
@@ -2677,7 +2840,7 @@ var GeminiTransport = class _GeminiTransport {
   // ── ChatTransport (multi-turn chat) ────────────────────────────────────
   sendChatMessage(history, userMessage, options) {
     return __awaiter(this, void 0, void 0, function* () {
-      var _a, _b;
+      var _a, _b, _c;
       const contents = this.toGeminiContents(history, userMessage, options === null || options === void 0 ? void 0 : options.attachments);
       if (options === null || options === void 0 ? void 0 : options.tailContext) {
         contents.push({ role: "user", parts: [{ text: options.tailContext }] });
@@ -2695,13 +2858,33 @@ var GeminiTransport = class _GeminiTransport {
       const toolConfig = tools ? toGeminiToolConfig(options === null || options === void 0 ? void 0 : options.toolChoice) : void 0;
       this.warnIfThinkingUnclampable(options === null || options === void 0 ? void 0 : options.thinkingPolicy);
       const generationConfig = { thinkingConfig: geminiThinkingConfig(this.model, options === null || options === void 0 ? void 0 : options.thinkingPolicy) };
+      if ((options === null || options === void 0 ? void 0 : options.maxTokens) != null) {
+        generationConfig.maxOutputTokens = options.maxTokens;
+      }
+      const resolutionHint = [
+        ...(_b = options === null || options === void 0 ? void 0 : options.attachments) !== null && _b !== void 0 ? _b : [],
+        ...history.flatMap((m) => {
+          var _a2;
+          return (_a2 = m.attachments) !== null && _a2 !== void 0 ? _a2 : [];
+        }),
+        // Tool-result images too: `ChatToolResult.attachments` is typed ChatImageAttachment[], so
+        // `detail` is offerable there, and a render handed back by a tool is the whole point of
+        // the tool-result image path. Omitting this source made the field silently inert there.
+        ...history.flatMap((m) => {
+          var _a2, _b2;
+          return (_b2 = (_a2 = m.toolResult) === null || _a2 === void 0 ? void 0 : _a2.attachments) !== null && _b2 !== void 0 ? _b2 : [];
+        })
+      ].find((att) => att.kind === "image" && att.detail);
+      if ((resolutionHint === null || resolutionHint === void 0 ? void 0 : resolutionHint.kind) === "image" && resolutionHint.detail && GEMINI_MEDIA_RESOLUTION_TIERS.includes(this.model)) {
+        generationConfig.mediaResolution = `MEDIA_RESOLUTION_${resolutionHint.detail.toUpperCase()}`;
+      }
       if ((options === null || options === void 0 ? void 0 : options.temperature) != null) {
         generationConfig.temperature = scaleTemperature(options.temperature, {
           defaultTemp: GEMINI_DEFAULT_TEMPERATURE,
           maxTemp: GEMINI_MAX_TEMPERATURE
         });
       }
-      const offeredToolNames = new Set(((_b = options === null || options === void 0 ? void 0 : options.tools) !== null && _b !== void 0 ? _b : []).map((t) => t.name));
+      const offeredToolNames = new Set(((_c = options === null || options === void 0 ? void 0 : options.tools) !== null && _c !== void 0 ? _c : []).map((t) => t.name));
       const schemaCoexistsWithTools = this.model.startsWith("gemini-3");
       const applyResponseSchema = (options === null || options === void 0 ? void 0 : options.responseSchema) != null && (!tools || schemaCoexistsWithTools);
       const response = yield this.post(Object.assign({
@@ -2712,7 +2895,7 @@ var GeminiTransport = class _GeminiTransport {
         toolConfig,
         generationConfig
       }, applyResponseSchema ? { responseSchema: toGeminiSchema(options.responseSchema) } : {}), options === null || options === void 0 ? void 0 : options.signal);
-      return this.fromGeminiResponse(response, offeredToolNames);
+      return this.fromGeminiResponse(response, offeredToolNames, generationConfig.maxOutputTokens);
     });
   }
   /**
@@ -2740,7 +2923,7 @@ var GeminiTransport = class _GeminiTransport {
     return costUsd;
   }
   toGeminiContents(history, userMessage, attachments) {
-    var _a, _b, _c;
+    var _a, _b, _c, _d;
     const contents = [];
     const toolCallNameById = /* @__PURE__ */ new Map();
     for (const msg of history) {
@@ -2760,7 +2943,16 @@ var GeminiTransport = class _GeminiTransport {
             }
           ]
         });
-      } else if ((_b = msg.toolCalls) === null || _b === void 0 ? void 0 : _b.length) {
+        const resultImages = (_b = msg.toolResult.attachments) !== null && _b !== void 0 ? _b : [];
+        if (resultImages.length) {
+          contents.push({
+            role: "user",
+            parts: resultImages.map((att) => ({
+              inlineData: { mimeType: normalizeImageMime(att.mimeType), data: att.data }
+            }))
+          });
+        }
+      } else if ((_c = msg.toolCalls) === null || _c === void 0 ? void 0 : _c.length) {
         for (const tc of msg.toolCalls) {
           toolCallNameById.set(tc.id, tc.name);
         }
@@ -2779,22 +2971,22 @@ var GeminiTransport = class _GeminiTransport {
             return part;
           })
         });
-      } else if (role === "user" && ((_c = msg.attachments) === null || _c === void 0 ? void 0 : _c.length)) {
-        const msgAttachmentParts = msg.attachments.map((att) => ({
-          text: `[File: ${att.name}]
-${att.content}`
-        }));
-        contents.push({ role: "user", parts: [{ text: msg.content }, ...msgAttachmentParts] });
+      } else if (role === "user" && ((_d = msg.attachments) === null || _d === void 0 ? void 0 : _d.length)) {
+        const { imageParts: imageParts2, textParts: textParts2 } = splitAttachmentParts(msg.attachments);
+        contents.push({
+          role: "user",
+          parts: [...imageParts2, ...msg.content ? [{ text: msg.content }] : [], ...textParts2]
+        });
       } else {
         contents.push({ role, parts: [{ text: msg.content }] });
       }
     }
-    const attachmentParts = (attachments !== null && attachments !== void 0 ? attachments : []).map((att) => ({
-      text: `[File: ${att.name}]
-${att.content}`
-    }));
-    if (userMessage || attachmentParts.length) {
-      contents.push({ role: "user", parts: [{ text: userMessage }, ...attachmentParts] });
+    const { imageParts, textParts } = splitAttachmentParts(attachments !== null && attachments !== void 0 ? attachments : []);
+    if (userMessage || imageParts.length || textParts.length) {
+      contents.push({
+        role: "user",
+        parts: [...imageParts, ...userMessage ? [{ text: userMessage }] : [], ...textParts]
+      });
     }
     return contents;
   }
@@ -2831,7 +3023,7 @@ ${att.content}`
       cost
     };
   }
-  fromGeminiResponse(response, offeredToolNames = /* @__PURE__ */ new Set()) {
+  fromGeminiResponse(response, offeredToolNames = /* @__PURE__ */ new Set(), effectiveMaxTokens) {
     var _a, _b, _c, _d;
     const { inputTokens, outputTokens, thoughtsTokens, cacheReadTokens, cost } = this.usageFromGemini(response.usageMetadata);
     const candidates = response === null || response === void 0 ? void 0 : response.candidates;
@@ -2878,7 +3070,7 @@ ${att.content}`
         textParts: textParts.length
       });
     }
-    this.guardMaxTokensTruncation(finishReason, toolCalls, narration, outputTokens);
+    this.guardMaxTokensTruncation(finishReason, toolCalls, narration, outputTokens, effectiveMaxTokens);
     if (inputTokens != null)
       base.inputTokens = inputTokens;
     if (outputTokens != null)
@@ -2916,12 +3108,12 @@ ${att.content}`
    * `responseMeta.finishReason`, for the caller to decide. Usage/cost is logged
    * before this check runs, so the spent tokens stay accounted for.
    */
-  guardMaxTokensTruncation(finishReason, toolCalls, narration, outputTokens) {
+  guardMaxTokensTruncation(finishReason, toolCalls, narration, outputTokens, effectiveMaxTokens) {
     if (finishReason !== "MAX_TOKENS")
       return;
     if (toolCalls.length === 0 && narration.trim() !== "")
       return;
-    throw new ResponseTruncatedError(this.model, void 0, outputTokens, toolCalls.map((tc) => tc.name));
+    throw new ResponseTruncatedError(this.model, effectiveMaxTokens, outputTokens, toolCalls.map((tc) => tc.name));
   }
   /**
    * Log the full shape of a blank or non-STOP response so its cause is legible
@@ -3482,11 +3674,12 @@ function triggerReason(trigger) {
       return "superseded";
   }
 }
-function condenseStub(target, tool, trigger, origLen, restorable) {
+function condenseStub(target, tool, trigger, origLen, restorable, imageCount = 0) {
   const what = target === "args" ? "args" : "result";
   const key = trigger.kind === "superseded" ? ` ${trigger.by}` : "";
   const restore = restorable ? "; re-call to restore" : "";
-  return `[${tool}${key} \u2014 ${what} elided, ~${origLen} chars (${triggerReason(trigger)})${restore}]`;
+  const images = imageCount > 0 ? ` + ${imageCount} image${imageCount === 1 ? "" : "s"}` : "";
+  return `[${tool}${key} \u2014 ${what} elided, ~${origLen} chars${images} (${triggerReason(trigger)})${restore}]`;
 }
 function triggerLabel(trigger) {
   switch (trigger.kind) {
@@ -3585,12 +3778,14 @@ function applyCondensation(history, policies, ctx, onCondensed) {
     }
     if (msg.toolResult) {
       const entry = policies.get(msg.toolResult.toolCallId);
-      const origLen = msg.toolResult.content?.length ?? 0;
+      const textLen = msg.toolResult.content?.length ?? 0;
+      const images = msg.toolResult.attachments ?? [];
+      const origLen = textLen + images.reduce((n, att) => n + (att.data?.length ?? 0), 0);
       const fired = entry?.policy.response ? firstFired(msg.toolResult.toolCallId, entry) : void 0;
       if (entry && fired && origLen >= CONDENSE_MIN_CHARS) {
         const tool = nameById.get(msg.toolResult.toolCallId) ?? msg.toolResult.toolCallId;
         const result = entry.policy.response;
-        const content = result === "drop" || result === "pointer" ? condenseStub("response", tool, fired, origLen, result === "pointer") : result.replaceWith;
+        const content = result === "drop" || result === "pointer" ? condenseStub("response", tool, fired, textLen, result === "pointer", images.length) : result.replaceWith;
         if (!entry.reportedResponse) {
           entry.reportedResponse = true;
           onCondensed({
@@ -3603,7 +3798,8 @@ function applyCondensation(history, policies, ctx, onCondensed) {
             tokensSaved: estimateTokensSaved(origLen, content.length)
           });
         }
-        return { ...msg, toolResult: { ...msg.toolResult, content } };
+        const { attachments: _elided, ...restOfResult } = msg.toolResult;
+        return { ...msg, toolResult: { ...restOfResult, content } };
       }
     }
     return msg;
@@ -3616,9 +3812,10 @@ function maskToolPayload(msg) {
     return { ...msg, toolCalls: msg.toolCalls.map((tc) => ({ ...tc, args: {} })) };
   }
   if (msg.toolResult) {
+    const { attachments: _masked, ...restOfResult } = msg.toolResult;
     return {
       ...msg,
-      toolResult: { ...msg.toolResult, content: "[other agent's tool result omitted]" }
+      toolResult: { ...restOfResult, content: "[other agent's tool result omitted]" }
     };
   }
   return msg;
@@ -3772,6 +3969,15 @@ function accumulate(messages, into) {
 var TOOL_FOLD_SYMBOL = Symbol("toolFold");
 
 // src/components/chat-driver/chat-driver.ts
+var toolResultImages = (result) => {
+  if (!result || typeof result !== "object") return void 0;
+  const candidate = result.attachments;
+  if (!Array.isArray(candidate) || candidate.length === 0) return void 0;
+  const allImages = candidate.every(
+    (att) => !!att && typeof att === "object" && att.kind === "image"
+  );
+  return allImages ? candidate : void 0;
+};
 var providerRefusedDetailOf = (e) => ({
   vendorLabel: e.vendorLabel,
   kind: e.kind,
@@ -5902,10 +6108,12 @@ ${tailBody}
             const capturedTrace = () => traceCapture.traces.length ? traceCapture.traces.flat() : void 0;
             try {
               const result = await handler(tc.args, this.buildHandlerContext(tc.id, traceCapture));
-              const content = typeof result === "string" ? result : JSON.stringify(result);
+              const resultImages = toolResultImages(result);
+              const content = resultImages ? typeof result.content === "string" ? result.content : JSON.stringify({ ...result, attachments: void 0 }) : typeof result === "string" ? result : JSON.stringify(result);
               executedById.set(tc.id, {
                 toolCallId: tc.id,
                 content,
+                ...resultImages ? { attachments: resultImages } : {},
                 // Concatenated when a handler invoked several children, so none is lost.
                 // Stays `undefined` when nothing was captured — readers key off presence,
                 // and an empty array is a different claim from "no sub-agent ran".
@@ -5954,7 +6162,14 @@ RECOVERY: this tool failed once \u2014 you may retry it, or take a different val
             this.appendToHistory({
               role: "tool",
               content: "",
-              toolResult: { toolCallId: r.toolCallId, content: r.content }
+              toolResult: {
+                toolCallId: r.toolCallId,
+                content: r.content,
+                // Images DO persist, unlike the sub-agent trace above: they are the payload
+                // the model must see on every replay of this turn, not a one-off UI artefact.
+                // `condenseWhen` is what sheds them once they are spent — see condense-history.
+                ...r.attachments ? { attachments: r.attachments } : {}
+              }
             });
           }
         }