fix(claude): normalize trailing assistant turn

open-sse/translator/request/openai-to-claude.js

@@ -96,6 +96,40 @@ export function openaiToClaudeRequest(model, body, stream) {
 
     flushCurrentMessage();
 
+    // GUARD: some Claude auth channels (OAuth/Claude Code) reject requests
+    // that end on an assistant turn ("assistant message prefill" 400).
+    // Agentic loops (e.g. Hermes) sometimes resend their own last output as
+    // the new final message to request a continuation, with no new user
+    // turn in between. Normalize by appending a synthetic turn so the
+    // request always ends on `user`, regardless of auth channel or model.
+    // If the trailing assistant message has unresolved tool_use blocks,
+    // Anthropic separately requires a matching tool_result for each one
+    // (not just any user turn), so synthesize those instead of plain text.
+    {
+      const lastMsg = result.messages[result.messages.length - 1];
+      if (lastMsg && lastMsg.role === ROLE.ASSISTANT) {
+        const unresolvedToolUseIds = Array.isArray(lastMsg.content)
+          ? lastMsg.content.filter(b => b.type === CLAUDE_BLOCK.TOOL_USE).map(b => b.id)
+          : [];
+
+        if (unresolvedToolUseIds.length > 0) {
+          result.messages.push({
+            role: ROLE.USER,
+            content: unresolvedToolUseIds.map(id => ({
+              type: CLAUDE_BLOCK.TOOL_RESULT,
+              tool_use_id: id,
+              content: "Continuing."
+            }))
+          });
+        } else {
+          result.messages.push({
+            role: ROLE.USER,
+            content: [{ type: CLAUDE_BLOCK.TEXT, text: "Continue." }]
+          });
+        }
+      }
+    }
+
     // Add cache_control to last assistant message
     for (let i = result.messages.length - 1; i >= 0; i--) {
       const message = result.messages[i];