@librechat/agents 3.1.98 → 3.1.99

package/dist/esm/messages/prune.mjs

@@ -364,7 +364,7 @@ function addThinkingBlock(message, thinkingBlock) {
             },
         ];
     /** Edge case, the message already has the thinking block */
-    if (content[0].type === thinkingBlock.type) {
+    if (content[0]?.type === thinkingBlock.type) {
         return message;
     }
     content.unshift(thinkingBlock);
@@ -397,6 +397,26 @@ function calculateTotalTokens(usage) {
         total_tokens: totalInputTokens + totalOutputTokens,
     };
 }
+/**
+ * Locates a reasoning block in assistant content. Reasoning blocks carry
+ * provider-specific `type` tags: Anthropic emits `thinking`, while Bedrock and
+ * OpenAI-compatible reasoning providers (DeepSeek-R1, DashScope/Qwen-thinking)
+ * emit `reasoning_content`. DeepSeek/Qwen route through the `THINKING` default
+ * even though their blocks are `reasoning_content` and aren't normalized
+ * upstream, so for the `THINKING` case we also accept `reasoning_content` — this
+ * is what fixes issue #191.
+ *
+ * The broadening is intentionally one-directional. A Bedrock run
+ * (`REASONING_CONTENT`) must NOT match an Anthropic `thinking` block: the
+ * Bedrock input converter rejects `thinking` blocks outright
+ * (`src/llm/bedrock/utils/message_inputs.ts`), so reattaching one to a
+ * surviving message would make the request fail before it is sent.
+ */
+function findReasoningBlock(content, reasoningType) {
+    return content.find((part) => part.type === reasoningType ||
+        (reasoningType === ContentTypes.THINKING &&
+            part.type === ContentTypes.REASONING_CONTENT));
+}
 /**
  * Processes an array of messages and returns a context of messages that fit within a specified token limit.
  * It iterates over the messages from newest to oldest, adding them to the context until the token limit is reached.
@@ -429,7 +449,7 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
     if (_thinkingStartIndex > -1) {
         const thinkingMessageContent = messages[_thinkingStartIndex]?.content;
         if (Array.isArray(thinkingMessageContent)) {
-            thinkingBlock = thinkingMessageContent.find((content) => content.type === reasoningType);
+            thinkingBlock = findReasoningBlock(thinkingMessageContent, reasoningType);
         }
     }
     if (currentTokenCount < remainingContextTokens) {
@@ -456,7 +476,7 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
                 thinkingStartIndex < 0 &&
                 messageType === 'ai' &&
                 Array.isArray(poppedMessage.content)) {
-                thinkingBlock = poppedMessage.content.find((content) => content.type === reasoningType);
+                thinkingBlock = findReasoningBlock(poppedMessage.content, reasoningType);
                 thinkingStartIndex = thinkingBlock != null ? currentIndex : -1;
             }
             /**
@@ -540,11 +560,28 @@ function getMessagesWithinTokenLimit({ messages: _messages, maxContextTokens, in
         result.context = context.reverse();
         return result;
     }
-    if (thinkingEndIndex > -1 && thinkingStartIndex < 0) {
-        throw new Error('The payload is malformed. There is a thinking sequence but no "AI" messages with thinking blocks.');
-    }
-    if (!thinkingBlock) {
-        throw new Error('The payload is malformed. There is a thinking sequence but no thinking block found.');
+    /**
+     * A trailing reasoning sequence was detected but its block could not be
+     * located in the surviving context. Rather than throw — which permanently
+     * bricks the conversation, re-firing on every retry of the same thread (see
+     * issue #191) — return the partially-pruned context and let the provider
+     * surface a real, recoverable error if the payload is genuinely malformed.
+     * Strict providers (Anthropic) reject it cleanly; lenient ones (DeepSeek,
+     * Qwen) proceed. The pruner cannot know which applies, so it must not be the
+     * one to make the failure fatal.
+     */
+    if ((thinkingEndIndex > -1 && thinkingStartIndex < 0) || !thinkingBlock) {
+        /**
+         * No block was located, so any `thinkingStartIndex` set above came from a
+         * stale carried-over index pointing at a block-less message. Drop it:
+         * `createPruneMessages` persists the returned index as
+         * `runThinkingStartIndex`, and a stale value would suppress the trailing
+         * scan (`thinkingStartIndex < 0`) on later turns, causing a real reasoning
+         * block to be missed and never reattached.
+         */
+        delete result.thinkingStartIndex;
+        result.context = context.reverse();
+        return result;
     }
     let assistantIndex = -1;
     for (let i = 0; i < context.length; i++) {