@kenkaiiii/gg-agent 5.83.2 → 5.83.4

package/dist/index.cjs

@@ -638,6 +638,9 @@ async function* agentLoop(messages, options) {
   let overloadRetries = 0;
   let emptyResponseRetries = 0;
   let stallRetries = 0;
+  let splitToolCallHintAdded = false;
+  let stalledInToolCall = false;
+  let stalledAfterConnecting = false;
   let runawayToolcallRetries = 0;
   let overflowCompactionAttempts = 0;
   let toolResultTruncationAttempted = false;
@@ -655,6 +658,7 @@ async function* agentLoop(messages, options) {
   const STALL_DELAY_MS = 1e3;
   const MIN_PARTIAL_PRESERVE_CHARS = 200;
   const PARTIAL_CONTINUATION_PROMPT = "[Your previous response was cut off by a connection failure. The text above is what was already delivered to the user. Continue exactly from where it stopped \u2014 do not repeat or restart it.]";
+  const SPLIT_TOOL_CALL_PROMPT = "[Your previous response was cut off while you were still writing a tool call's arguments, after minutes without data from the provider. None of that call was applied. Do not resend it as one large call: split the work into smaller pieces (e.g. a shorter first version, then further edits, or one file per call).]";
   const MAX_OUTPUT_CONTINUATIONS = 2;
   const MAX_TOKENS_CONTINUATION_PROMPT = "[Your previous response hit the output-token limit and was cut off. The text above is what was already delivered to the user. Continue exactly from where it stopped \u2014 do not repeat or restart it.]";
   let maxTokensContinuations = 0;
@@ -680,7 +684,7 @@ async function* agentLoop(messages, options) {
   const NON_STREAMING_HARD_TIMEOUT_MS = 3e5;
   const nonStreamingHardTimeoutMs = options.thinking != null ? STREAM_THINKING_HARD_TIMEOUT_MS : NON_STREAMING_HARD_TIMEOUT_MS;
   const VISIBLE_STALL_AFTER_MS = 6e4;
-  const usesSilentReasoningBudget = options.provider === "sakana" || options.provider === "openai" && options.thinking != null;
+  const usesSilentReasoningBudget = options.provider === "sakana" || options.provider === "xai" || options.provider === "openai" && options.thinking != null;
   const localBackend = isLocalBackendUrl(options.baseUrl);
   const baseFirstEventTimeoutMs = localBackend ? Number.POSITIVE_INFINITY : usesSilentReasoningBudget ? STREAM_THINKING_IDLE_TIMEOUT_MS : STREAM_FIRST_EVENT_TIMEOUT_MS;
   const baseHardTimeoutMs = localBackend || usesSilentReasoningBudget ? STREAM_THINKING_HARD_TIMEOUT_MS : STREAM_HARD_TIMEOUT_MS;
@@ -902,7 +906,7 @@ async function* agentLoop(messages, options) {
               }, hardTimeoutMs);
             }
           }
-          if (event.type === "thinking_delta" && !hasReceivedThinking) {
+          if ((event.type === "thinking_delta" || event.type === "keepalive" && event.reasoning === true) && !hasReceivedThinking) {
             hasReceivedThinking = true;
             if (hardTimer) clearTimeout(hardTimer);
             hardTimeoutMs = STREAM_THINKING_HARD_TIMEOUT_MS;
@@ -1231,6 +1235,9 @@ async function* agentLoop(messages, options) {
         const malformed = isMalformedStream(err);
         const socketDrop = isTransportFailure(err);
         const transportFailure = (idleTimedOut || malformed || socketDrop) && !options.signal?.aborted;
+        const attemptMs = Date.now() - streamCallStart;
+        if (transportFailure && toolInputOpen) stalledInToolCall = true;
+        if (transportFailure && streamEventCount > 0) stalledAfterConnecting = true;
         if (transportFailure && stallRetries < MAX_STALL_RETRIES) {
           stallRetries++;
           const cause = malformed ? "malformed_stream" : socketDrop ? "socket_drop" : "stream_stall";
@@ -1261,6 +1268,20 @@ async function* agentLoop(messages, options) {
             });
             preservedChars = attemptText.length;
           }
+          let splitHint = false;
+          if (toolInputOpen && attemptMs >= VISIBLE_STALL_AFTER_MS && !splitToolCallHintAdded) {
+            splitToolCallHintAdded = true;
+            splitHint = true;
+            messages.push({
+              role: "user",
+              content: SPLIT_TOOL_CALL_PROMPT,
+              provenance: {
+                source: "runtime",
+                kind: "continuation",
+                visibility: "hidden"
+              }
+            });
+          }
           diag("retry", {
             reason: cause,
             attempt: stallRetries,
@@ -1268,7 +1289,8 @@ async function* agentLoop(messages, options) {
             delayMs,
             events: streamEventCount,
             nonStreaming: useNonStreamingFallback,
-            preservedChars
+            preservedChars,
+            splitHint
           });
           yield {
             type: "retry",
@@ -1295,9 +1317,25 @@ async function* agentLoop(messages, options) {
             eventTypes: eventTypeCounts,
             lastEventType,
             sinceLastEventMs: Date.now() - lastEventTime,
-            attemptDurationMs: Date.now() - streamCallStart,
-            maxConsumerLagMs
+            attemptDurationMs: attemptMs,
+            maxConsumerLagMs,
+            stalledAfterConnecting,
+            stalledInToolCall
           });
+          if (stalledAfterConnecting) {
+            yield {
+              type: "error",
+              error: new import_gg_ai2.ResponseStalledError(
+                `The provider stopped sending data before the response finished, ${MAX_STALL_RETRIES} automatic retries in a row. Your conversation is preserved.`,
+                {
+                  inToolCall: stalledInToolCall,
+                  hint: stalledInToolCall ? "The model was writing one very large tool call each time. Ask it to make the change in smaller pieces (shorter files or several edits), then retry." : "Retry once. If it keeps happening, choose another model or lower the thinking level, and disable any VPN or proxy that may buffer streamed responses.",
+                  cause: err
+                }
+              )
+            };
+            break;
+          }
           yield {
             type: "error",
             error: new import_gg_ai2.GGAIError(
@@ -1346,6 +1384,9 @@ async function* agentLoop(messages, options) {
       }
       overloadRetries = 0;
       stallRetries = 0;
+      splitToolCallHintAdded = false;
+      stalledInToolCall = false;
+      stalledAfterConnecting = false;
       runawayToolcallRetries = 0;
       const contentArr = Array.isArray(response.message.content) ? response.message.content : null;
       const hasActionableContent = response.message.content !== "" && contentArr !== null && contentArr.some(