@kenkaiiii/gg-agent 5.22.2 → 5.22.4

package/dist/index.cjs

@@ -227,6 +227,7 @@ async function* agentLoop(messages, options) {
   let overloadRetries = 0;
   let emptyResponseRetries = 0;
   let stallRetries = 0;
+  let runawayToolcallRetries = 0;
   let overflowCompactionAttempts = 0;
   let toolResultTruncationAttempted = false;
   const invalidToolArgumentCounts = /* @__PURE__ */ new Map();
@@ -235,6 +236,8 @@ async function* agentLoop(messages, options) {
   const MAX_OVERLOAD_RETRIES = 10;
   const MAX_EMPTY_RESPONSE_RETRIES = 2;
   const MAX_STALL_RETRIES = 5;
+  const MAX_RUNAWAY_TOOLCALL_RETRIES = 2;
+  const RUNAWAY_TOOLCALL_RETRY_DELAY_MS = 1e3;
   const MAX_OVERFLOW_COMPACTIONS = 2;
   const STALL_RETRIES_BEFORE_NON_STREAMING = 2;
   const STALL_DELAY_MS = 1e3;
@@ -252,9 +255,9 @@ async function* agentLoop(messages, options) {
   const STREAM_THINKING_IDLE_TIMEOUT_MS = 3e5;
   const STREAM_THINKING_HARD_TIMEOUT_MS = 6e5;
   const NON_STREAMING_HARD_TIMEOUT_MS = 3e5;
-  const isSakana = options.provider === "sakana";
-  const firstEventTimeoutMs = isSakana ? STREAM_THINKING_IDLE_TIMEOUT_MS : STREAM_FIRST_EVENT_TIMEOUT_MS;
-  const initialHardTimeoutMs = isSakana ? STREAM_THINKING_HARD_TIMEOUT_MS : STREAM_HARD_TIMEOUT_MS;
+  const usesSilentReasoningBudget = options.provider === "sakana" || options.provider === "openai" && options.thinking != null;
+  const firstEventTimeoutMs = usesSilentReasoningBudget ? STREAM_THINKING_IDLE_TIMEOUT_MS : STREAM_FIRST_EVENT_TIMEOUT_MS;
+  const initialHardTimeoutMs = usesSilentReasoningBudget ? STREAM_THINKING_HARD_TIMEOUT_MS : STREAM_HARD_TIMEOUT_MS;
   const MAX_TOOLCALL_DELTA_CHARS = 1e6;
   const MAX_TOOLCALL_DELTA_EVENTS = 2e4;
   let logicalTurnStartedAt = 0;
@@ -282,7 +285,10 @@ async function* agentLoop(messages, options) {
           messages: messages.length,
           chars: msgChars,
           provider: options.provider,
-          model: options.model
+          model: options.model,
+          thinking: options.thinking ?? "off",
+          firstEventTimeoutMs,
+          initialHardTimeoutMs
         });
       }
       if (firstTurn && options.getSteeringMessages) {
@@ -639,11 +645,33 @@ async function* agentLoop(messages, options) {
             provider: options.provider,
             model: options.model
           });
+          if (runawayToolcallRetries < MAX_RUNAWAY_TOOLCALL_RETRIES) {
+            runawayToolcallRetries++;
+            const delayMs = RUNAWAY_TOOLCALL_RETRY_DELAY_MS * runawayToolcallRetries;
+            diag("retry", {
+              reason: "runaway_toolcall",
+              attempt: runawayToolcallRetries,
+              maxAttempts: MAX_RUNAWAY_TOOLCALL_RETRIES,
+              delayMs,
+              ...runawayDetected
+            });
+            yield {
+              type: "retry",
+              reason: "runaway_toolcall",
+              attempt: runawayToolcallRetries,
+              maxAttempts: MAX_RUNAWAY_TOOLCALL_RETRIES,
+              delayMs,
+              silent: true
+            };
+            await abortableSleep(delayMs, options.signal);
+            turn--;
+            continue;
+          }
           const detail = runawayDetected.kind === "chars" ? `${(runawayDetected.chars / 1024).toFixed(0)} KB of tool-call arguments` : `${runawayDetected.events} tool-call delta events`;
           yield {
             type: "error",
             error: new Error(
-              `The model glitched mid-tool-call and produced ${detail} without closing the call. This is usually an upstream model bug \u2014 try the same request again or switch models. Your conversation is preserved.`
+              `The model repeatedly failed to close a tool call after ${MAX_RUNAWAY_TOOLCALL_RETRIES} automatic retries (${detail}). Switch models and retry; your conversation is preserved.`
             )
           };
           break;
@@ -696,15 +724,29 @@ async function* agentLoop(messages, options) {
           continue;
         }
         if (transportFailure) {
+          const cause = malformed ? "malformed_stream" : socketDrop ? "socket_drop" : "stream_stall";
           diag("stall_exhausted", {
             stallRetries: MAX_STALL_RETRIES,
             provider: options.provider,
-            model: options.model
+            model: options.model,
+            cause,
+            nonStreaming: useNonStreamingFallback,
+            events: streamEventCount,
+            eventTypes: eventTypeCounts,
+            lastEventType,
+            sinceLastEventMs: Date.now() - lastEventTime,
+            attemptDurationMs: Date.now() - streamCallStart,
+            maxConsumerLagMs
           });
           yield {
             type: "error",
-            error: new Error(
-              `The API provider's stream stalled ${MAX_STALL_RETRIES} times \u2014 the provider may be experiencing capacity issues. Your conversation is preserved. Send another message to retry.`
+            error: new import_gg_ai.GGAIError(
+              `The connection to the API provider stopped responding after ${MAX_STALL_RETRIES} automatic retries. Your conversation is preserved.`,
+              {
+                source: "network",
+                hint: "Retry once. If it keeps happening on this device, disable any VPN or proxy and allow GG Coder through firewall or antivirus web protection.",
+                cause: err
+              }
             )
           };
           break;
@@ -744,6 +786,7 @@ async function* agentLoop(messages, options) {
       }
       overloadRetries = 0;
       stallRetries = 0;
+      runawayToolcallRetries = 0;
       const contentArr = Array.isArray(response.message.content) ? response.message.content : null;
       const hasActionableContent = response.message.content !== "" && contentArr !== null && contentArr.some(
         (p) => p.type === "text" || p.type === "tool_call" || p.type === "server_tool_call"