Merge origin/master (v0.5.69) into gitea/new_feature

open-sse/translator/response/gemini-to-openai.js

@@ -6,6 +6,7 @@ import { toOpenAIUsage } from "../concerns/usage.js";
 import { reasoningDelta } from "../concerns/reasoning.js";
 import { encodeDataUri } from "../concerns/image.js";
 import { toOpenAIFinish } from "../concerns/finishReason.js";
+import { storeGeminiThoughtSignature } from "../../services/thoughtSignatureStore.js";
 
 // Build chunk meta for current gemini state
 function chunkMeta(state) {
@@ -13,14 +14,18 @@ function chunkMeta(state) {
 }
 
 // Build a tool_call chunk from a gemini functionCall part (shared by sig/non-sig branches)
-function emitFunctionCall(functionCall, state) {
+function emitFunctionCall(functionCall, state, signature = null) {
   const rawName = functionCall.name;
   // Restore original tool name from mapping (AG cloaking)
   const fcName = state.toolNameMap?.get(rawName) || rawName;
   const fcArgs = functionCall.args || {};
   const toolCallIndex = state.functionIndex++;
+  const callId = functionCall.id || `${fcName}-${Date.now()}-${toolCallIndex}`;
+  if (signature) {
+    storeGeminiThoughtSignature(callId, signature, state.sessionId);
+  }
   const toolCall = {
-    id: `${fcName}-${Date.now()}-${toolCallIndex}`,
+    id: callId,
     index: toolCallIndex,
     type: OPENAI_BLOCK.FUNCTION,
     function: { name: fcName, arguments: JSON.stringify(fcArgs) },
@@ -57,13 +62,21 @@ export function geminiToOpenAIResponse(chunk, state) {
   if (content?.parts) {
     for (const part of content.parts) {
       const hasThoughtSig = part.thoughtSignature || part.thought_signature;
+      if (hasThoughtSig && typeof hasThoughtSig === "string") {
+        state.pendingThoughtSignature = hasThoughtSig;
+      }
       const isThought = part.thought === true;
-      
+
       // Handle thought signature (thinking mode)
       if (hasThoughtSig) {
         const hasTextContent = part.text !== undefined && part.text !== "";
         const hasFunctionCall = !!part.functionCall;
-        
+
+        // Standalone thoughtSignature part (no text, no functionCall): keep pending for next functionCall
+        if (!hasTextContent && !hasFunctionCall) {
+          continue;
+        }
+
         if (hasTextContent) {
           results.push(buildChunk(
             chunkMeta(state),
@@ -71,9 +84,10 @@ export function geminiToOpenAIResponse(chunk, state) {
             null
           ));
         }
-        
+
         if (hasFunctionCall) {
-          results.push(emitFunctionCall(part.functionCall, state));
+          results.push(emitFunctionCall(part.functionCall, state, hasThoughtSig));
+          state.pendingThoughtSignature = null;
         }
         continue;
       }
@@ -92,7 +106,9 @@ export function geminiToOpenAIResponse(chunk, state) {
 
       // Function call
       if (part.functionCall) {
-        results.push(emitFunctionCall(part.functionCall, state));
+        const sig = state.pendingThoughtSignature || null;
+        results.push(emitFunctionCall(part.functionCall, state, sig));
+        state.pendingThoughtSignature = null;
       }
 
       // Inline data (images)

open-sse/translator/response/openai-responses.js

@@ -446,6 +446,13 @@ export function openaiResponsesToOpenAIResponse(chunk, state) {
     state.created = Math.floor(Date.now() / 1000);
     state.toolCallIndex = 0;
     state.currentToolCallId = null;
+    // item_id → chat tool_calls index. Deltas carry item_id; keying on it (not
+    // stream position) keeps parallel calls separate when upstream emits all
+    // output_item.added events before any done/delta. Lazily created so callers
+    // that build their own state object (stream.js) need no changes.
+    state.respToolChatIndex ??= new Map();
+    // Indices that already received argument deltas (guards done-with-args).
+    state.respToolArgsEmitted ??= new Set();
   }
 
   // Text content delta
@@ -464,16 +471,29 @@ export function openaiResponsesToOpenAIResponse(chunk, state) {
     return null;
   }
 
-  // Function call started (standard function_call or custom_tool_call)
+  // Function call started (standard function_call or custom_tool_call).
+  // Index is assigned here (not on done): attributing deltas by stream position
+  // merges parallel calls into index 0 whenever upstream emits all addeds
+  // before dones — the client then concatenates N JSON payloads into one
+  // tool input and fails validation. The server item id is the correlator.
   if (eventType === "response.output_item.added" && (data.item?.type === RESPONSES_ITEM.FUNCTION_CALL || data.item?.type === "custom_tool_call")) {
     const item = data.item;
     state.currentToolCallId = item.call_id || fallbackToolCallId();
+    state.respToolChatIndex ??= new Map();
+    const key = item.id || data.item_id || state.currentToolCallId;
+    let idx;
+    if (key && state.respToolChatIndex.has(key)) {
+      idx = state.respToolChatIndex.get(key); // duplicate added (retry) — reuse
+    } else {
+      idx = state.toolCallIndex++;
+      if (key) state.respToolChatIndex.set(key, idx);
+    }
 
     return buildChunk(
       { id: state.chatId, created: state.created, model: state.model || MODEL_FALLBACK },
       {
         tool_calls: [{
-          index: state.toolCallIndex,
+          index: idx,
           id: state.currentToolCallId,
           type: OPENAI_BLOCK.FUNCTION,
           function: { name: item.name || "", arguments: "" }
@@ -482,20 +502,39 @@ export function openaiResponsesToOpenAIResponse(chunk, state) {
     );
   }
 
-  // Function call arguments delta (standard or custom_tool_call variant)
+  // Function call arguments delta (standard or custom_tool_call variant).
+  // Routed by item_id so interleaved parallel fragments stay on their own call.
   if (eventType === "response.function_call_arguments.delta" || eventType === "response.custom_tool_call_input.delta") {
     const argsDelta = data.delta || "";
     if (!argsDelta) return null;
 
+    const known = data.item_id ? state.respToolChatIndex?.get(data.item_id) : undefined;
+    const idx = known ?? Math.max(0, (state.toolCallIndex || 1) - 1);
+    state.respToolArgsEmitted ??= new Set();
+    state.respToolArgsEmitted.add(idx);
     return buildChunk(
       { id: state.chatId, created: state.created, model: state.model || MODEL_FALLBACK },
-      { tool_calls: [{ index: state.toolCallIndex, function: { arguments: argsDelta } }] }
+      { tool_calls: [{ index: idx, function: { arguments: argsDelta } }] }
     );
   }
 
-  // Function call done (standard or custom_tool_call variant)
+  // Function call done (standard or custom_tool_call variant).
+  // Index was assigned at added-time; nothing to advance. Some upstreams send
+  // complete arguments only here (no deltas) — emit them once in that case.
   if (eventType === "response.output_item.done" && (data.item?.type === RESPONSES_ITEM.FUNCTION_CALL || data.item?.type === "custom_tool_call")) {
-    state.toolCallIndex++;
+    const key = data.item?.id || data.item_id;
+    const idx = (key && state.respToolChatIndex?.get(key)) ?? Math.max(0, (state.toolCallIndex || 1) - 1);
+    const fullArgs = data.item?.arguments;
+    if (typeof fullArgs === "string" && fullArgs) {
+      state.respToolArgsEmitted ??= new Set();
+      if (!state.respToolArgsEmitted.has(idx)) {
+        state.respToolArgsEmitted.add(idx);
+        return buildChunk(
+          { id: state.chatId, created: state.created, model: state.model || MODEL_FALLBACK },
+          { tool_calls: [{ index: idx, function: { arguments: fullArgs } }] }
+        );
+      }
+    }
     return null;
   }