@prestyj/ai 5.8.0 → 5.9.0

package/dist/index.d.cts

@@ -41,6 +41,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";

package/dist/index.d.ts

@@ -41,6 +41,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";

package/dist/index.js

@@ -264,6 +264,9 @@ function providerGuidance(provider, message, statusCode) {
   if (lower.includes("context_length_exceeded") || lower.includes("prompt is too long")) {
     return `Context window for this ${name} model is full. Compact the conversation to shrink history, or start a new session.`;
   }
+  if (lower.includes("many-image request") || lower.includes("image dimensions") && lower.includes("max allowed size")) {
+    return `An image in conversation history exceeds ${name}'s many-image limit. Restart EZ Coder so restored images are resized, then retry; if it persists, start a new session.`;
+  }
   if (statusCode === 413 || lower.includes("request_too_large") || lower.includes("request exceeds the maximum size")) {
     return `The request to ${name} is too large. Compact the conversation to shrink history, or start a new session.`;
   }
@@ -760,9 +763,14 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "user") {
+      if (typeof msg.content === "string") {
+        if (msg.content === "") continue;
+      } else if (!msg.content.some((p) => !(p.type === "text" && p.text === ""))) {
+        continue;
+      }
       out.push({
         role: "user",
-        content: typeof msg.content === "string" ? msg.content : msg.content.map((part) => {
+        content: typeof msg.content === "string" ? msg.content : msg.content.filter((part) => !(part.type === "text" && part.text === "")).map((part) => {
           if (part.type === "text") return { type: "text", text: part.text };
           if (part.type === "video") {
             return {
@@ -787,6 +795,7 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "assistant") {
+      if (typeof msg.content === "string" && msg.content === "") continue;
       const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(msg.content, msgIdx > trajectoryStartIdx, idMap);
       if (Array.isArray(content) && content.length === 0) continue;
       out.push({ role: "assistant", content });
@@ -907,7 +916,7 @@ function remapToolCallId(id, idMap) {
   if (!id.startsWith("toolu_")) return id;
   const existing = idMap.get(id);
   if (existing) return existing;
-  const mapped = `call_${id.slice(5)}`;
+  const mapped = `call_${id.slice(6)}`;
   idMap.set(id, mapped);
   return mapped;
 }
@@ -1518,6 +1527,12 @@ async function* runStream(options) {
       statusCode: 504
     });
   }
+  if (stopReason === null) {
+    throw new ProviderError("anthropic", "Stream ended before completion (no stop_reason).", {
+      statusCode: 504,
+      cause: { partialContent: contentParts, outputTokens }
+    });
+  }
   const normalizedStop = normalizeAnthropicStopReason(stopReason);
   const response = {
     message: {
@@ -1784,6 +1799,17 @@ function getEnvironment() {
 }
 
 // src/providers/openai.ts
+function toKimiK3Effort(level) {
+  switch (level) {
+    case "low":
+      return "low";
+    case "medium":
+    case "high":
+      return "high";
+    default:
+      return "max";
+  }
+}
 function extractOpenAIUsage(usage) {
   let cacheRead = 0;
   let cacheWrite = 0;
@@ -1840,6 +1866,7 @@ async function* runStream2(options) {
   const client = createClient2(options);
   const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
   const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").endsWith("/coding/v1") === true;
+  const k3Effort = options.thinking ? toKimiK3Effort(options.thinking) : void 0;
   const isKimiK27 = options.provider === "moonshot" && options.model.startsWith("kimi-k2.7-code");
   const hasFixedKimiSampling = isKimiK3 || isKimiK27;
   const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" && !isKimiK3 && !isKimiK27 || options.provider === "xiaomi";
@@ -1854,9 +1881,11 @@ async function* runStream2(options) {
   }
   const messages = toOpenAIMessages(downgradedMessages, {
     provider: options.provider,
-    // K3 and K2.7 preserve reasoning even when the user hides thinking in the
-    // UI; keep assistant tool-call history wire-valid in that display mode.
-    thinking: isKimiK3 || isKimiK27 || !!options.thinking,
+    // K2.7 preserves reasoning even when the user hides thinking in the UI;
+    // keep assistant tool-call history wire-valid in that display mode. A
+    // disabled K3 must NOT carry placeholder reasoning_content (mirrors the
+    // official CLI: reasoning is preserved only while thinking is enabled).
+    thinking: isKimiK27 || !!options.thinking,
     supportsImages: options.supportsImages
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
@@ -1889,9 +1918,11 @@ async function* runStream2(options) {
   if (isKimiK3) {
     const paramsAny = params;
     if (isManagedKimiK3) {
-      paramsAny.thinking = { type: "enabled", effort: "max", keep: "all" };
+      paramsAny.thinking = k3Effort ? { type: "enabled", effort: k3Effort, keep: "all" } : { type: "disabled" };
+    } else if (k3Effort) {
+      paramsAny.reasoning_effort = k3Effort;
     } else {
-      paramsAny.reasoning_effort = "max";
+      paramsAny.thinking = { type: "disabled" };
     }
   }
   if (usesThinkingParam) {
@@ -1997,6 +2028,12 @@ async function* runStream2(options) {
       statusCode: 504
     });
   }
+  if (finishReason === null) {
+    throw new ProviderError(providerName, "Stream ended before completion (no finish_reason).", {
+      statusCode: 504,
+      cause: { partialText: textAccum, outputTokens }
+    });
+  }
   if (thinkingAccum) {
     contentParts.push({ type: "thinking", text: thinkingAccum });
   }