@kenkaiiii/gg-ai 5.67.1 → 5.68.1

package/dist/index.d.cts

@@ -186,8 +186,24 @@ interface Usage {
         webFetchRequests?: number;
     };
 }
+interface PreparedContext {
+    /** Read-only observation after generic sanitization/image limiting, before provider encoding.
+     * Do not log or retain these messages: they may contain secrets and image data. */
+    messages: readonly Message[];
+    tools: readonly {
+        name: string;
+        description: string;
+        parameters: Record<string, unknown>;
+    }[];
+    imagesBefore: number;
+    imagesAfter: number;
+    /** First message whose images were removed for this request, or null. */
+    firstImageDropMessage: number | null;
+}
 interface StreamOptions {
     provider: Provider;
+    /** Optional, synchronous diagnostics observer. Exceptions cannot fail a model request. */
+    onContextPrepared?: (context: PreparedContext) => void;
     model: string;
     messages: Message[];
     tools?: Tool[];
@@ -245,6 +261,14 @@ interface StreamOptions {
      *  stalls — broken SSE connections (transient CDN / proxy issues) often
      *  recover when the same request is issued over a plain HTTP request/response. */
     streaming?: boolean;
+    /** Cache prewarm (Anthropic only; other providers ignore it). Sends the exact
+     *  same request prefix (system, tools, messages, thinking, betas) with
+     *  `max_tokens: 1` over a non-streaming request so the prompt cache is written
+     *  before the user's next real turn. If the configured thinking mode cannot be
+     *  kept identical at `max_tokens: 1` (budget thinking requires
+     *  budget_tokens < max_tokens), no request is sent and the response has
+     *  zero usage with stopReason "end_turn". */
+    prewarm?: boolean;
     /** Override the User-Agent sent with OAuth-authenticated Anthropic requests.
      *  Anthropic's OAuth edge rejects requests whose claude-cli version lags too
      *  far behind the real Claude Code release; callers that track the live
@@ -574,7 +598,8 @@ type JsonSchema = Record<string, unknown>;
 declare function resolveToolSchema(tool: Tool): JsonSchema;
 
 /**
- * Cap historical images before provider dispatch, removing the oldest first.
+ * Cap historical images before provider dispatch, removing the oldest first, in batches so the
+ * cached conversation prefix stays byte-identical between requests.
  * The persisted/live conversation is never mutated; only modified messages and
  * tool results are cloned for the outgoing request.
  */
@@ -685,4 +710,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.d.ts

@@ -186,8 +186,24 @@ interface Usage {
         webFetchRequests?: number;
     };
 }
+interface PreparedContext {
+    /** Read-only observation after generic sanitization/image limiting, before provider encoding.
+     * Do not log or retain these messages: they may contain secrets and image data. */
+    messages: readonly Message[];
+    tools: readonly {
+        name: string;
+        description: string;
+        parameters: Record<string, unknown>;
+    }[];
+    imagesBefore: number;
+    imagesAfter: number;
+    /** First message whose images were removed for this request, or null. */
+    firstImageDropMessage: number | null;
+}
 interface StreamOptions {
     provider: Provider;
+    /** Optional, synchronous diagnostics observer. Exceptions cannot fail a model request. */
+    onContextPrepared?: (context: PreparedContext) => void;
     model: string;
     messages: Message[];
     tools?: Tool[];
@@ -245,6 +261,14 @@ interface StreamOptions {
      *  stalls — broken SSE connections (transient CDN / proxy issues) often
      *  recover when the same request is issued over a plain HTTP request/response. */
     streaming?: boolean;
+    /** Cache prewarm (Anthropic only; other providers ignore it). Sends the exact
+     *  same request prefix (system, tools, messages, thinking, betas) with
+     *  `max_tokens: 1` over a non-streaming request so the prompt cache is written
+     *  before the user's next real turn. If the configured thinking mode cannot be
+     *  kept identical at `max_tokens: 1` (budget thinking requires
+     *  budget_tokens < max_tokens), no request is sent and the response has
+     *  zero usage with stopReason "end_turn". */
+    prewarm?: boolean;
     /** Override the User-Agent sent with OAuth-authenticated Anthropic requests.
      *  Anthropic's OAuth edge rejects requests whose claude-cli version lags too
      *  far behind the real Claude Code release; callers that track the live
@@ -574,7 +598,8 @@ type JsonSchema = Record<string, unknown>;
 declare function resolveToolSchema(tool: Tool): JsonSchema;
 
 /**
- * Cap historical images before provider dispatch, removing the oldest first.
+ * Cap historical images before provider dispatch, removing the oldest first, in batches so the
+ * cached conversation prefix stays byte-identical between requests.
  * The persisted/live conversation is never mutated; only modified messages and
  * tool results are cloned for the outgoing request.
  */
@@ -685,4 +710,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, type ImageContent, type Message, type MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type PreparedContext, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.js

@@ -743,10 +743,17 @@ 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)) {
@@ -1376,7 +1383,7 @@ 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);
@@ -1400,6 +1407,17 @@ 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,
@@ -3830,6 +3848,41 @@ 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"}`;
@@ -3977,6 +4030,14 @@ 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) {