@kenkaiiii/gg-ai 5.67.1 → 5.68.1

package/dist/index.cjs

@@ -810,10 +810,17 @@ function countContextImages(messages) {
   }
   return count;
 }
+var IMAGE_DROP_BATCH = 30;
+function providerImageDropCount(imageCount2, budget) {
+  const overflow = imageCount2 - budget;
+  if (overflow <= 0) return 0;
+  const batch = Math.max(1, Math.min(IMAGE_DROP_BATCH, Math.floor(budget / 3)));
+  return Math.min(imageCount2, Math.ceil(overflow / batch) * batch);
+}
 function clampProviderContextImages(messages, provider, supportsImages) {
   if (supportsImages === false) return messages;
   const budget = PROVIDER_IMAGE_BUDGETS[provider] ?? 5;
-  let remainingToRemove = countContextImages(messages) - budget;
+  let remainingToRemove = providerImageDropCount(countContextImages(messages), budget);
   if (remainingToRemove <= 0) return messages;
   return messages.map((message) => {
     if (message.role === "user" && Array.isArray(message.content)) {
@@ -1443,7 +1450,7 @@ function streamAnthropic(options) {
 async function* runStream(options) {
   const client = createClient(options);
   const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
-  const useStreaming = options.streaming !== false;
+  const useStreaming = options.streaming !== false && !options.prewarm;
   const cacheControl = toAnthropicCacheControl(options.cacheRetention, options.baseUrl);
   const supportsFirstPartyToolExtras = !options.baseUrl || options.baseUrl.includes("api.anthropic.com");
   const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
@@ -1467,6 +1474,17 @@ async function* runStream(options) {
       outputConfig = t.outputConfig;
     }
   }
+  if (options.prewarm) {
+    const budget = thinking?.budget_tokens;
+    if (budget != null && budget >= 1) {
+      return {
+        message: { role: "assistant", content: [] },
+        stopReason: "end_turn",
+        usage: { inputTokens: 0, outputTokens: 0 }
+      };
+    }
+    maxTokens = 1;
+  }
   const params = {
     model: options.model,
     max_tokens: maxTokens,
@@ -3897,6 +3915,41 @@ function sanitizeMessagesForWire(messages) {
   return sanitized ?? messages;
 }
 
+// src/utils/context-observation.ts
+function imageCount(message) {
+  if (!message || !Array.isArray(message.content)) return 0;
+  if (message.role === "user")
+    return message.content.filter((part) => part.type === "image").length;
+  if (message.role !== "tool") return 0;
+  return message.content.reduce(
+    (count, result) => count + (Array.isArray(result.content) ? result.content.filter((part) => part.type === "image").length : 0),
+    0
+  );
+}
+function observePreparedContext(before, after, tools) {
+  let imagesBefore = 0;
+  let imagesAfter = 0;
+  let firstImageDropMessage = null;
+  for (let index = 0; index < before.length; index++) {
+    const oldCount = imageCount(before[index]);
+    const newCount = imageCount(after[index]);
+    imagesBefore += oldCount;
+    imagesAfter += newCount;
+    if (firstImageDropMessage === null && newCount < oldCount) firstImageDropMessage = index;
+  }
+  return {
+    messages: after,
+    tools: tools.map((tool) => ({
+      name: tool.name,
+      description: tool.description,
+      parameters: resolveToolSchema(tool)
+    })),
+    imagesBefore,
+    imagesAfter,
+    firstImageDropMessage
+  };
+}
+
 // src/stream.ts
 var GLM_CODING_BASE_URL = "https://api.z.ai/api/coding/paas/v4";
 var KIMI_CODE_USER_AGENT = `kimi-code-cli/${process.env.KIMI_CODE_VERSION ?? "1.0.11"}`;
@@ -4044,6 +4097,14 @@ function stream(options) {
     options.provider,
     options.supportsImages
   );
+  if (options.onContextPrepared) {
+    try {
+      options.onContextPrepared(
+        observePreparedContext(wireMessages, messages, options.tools ?? [])
+      );
+    } catch {
+    }
+  }
   return entry.stream(messages === options.messages ? options : { ...options, messages });
 }
 function stripMessageProvenance(messages) {