@kenkaiiii/gg-agent 5.23.2 → 5.24.0

package/dist/index.d.cts

@@ -324,4 +324,13 @@ declare function isBillingError(err: unknown): boolean;
 declare function isUsageLimitError(err: unknown): boolean;
 declare function agentLoop(messages: Message[], options: AgentOptions): AsyncGenerator<AgentEvent, AgentResult>;
 
-export { Agent, type AgentDoneEvent, type AgentErrorEvent, type AgentEvent, type AgentFollowUpMessageEvent, type AgentOptions, type AgentResult, type AgentRetryEvent, type AgentServerToolCallEvent, type AgentServerToolResultEvent, type AgentSteeringMessageEvent, AgentStream, type AgentTextDeltaEvent, type AgentThinkingDeltaEvent, type AgentTool, type AgentToolCallDeltaEvent, type AgentToolCallEndEvent, type AgentToolCallStartEvent, type AgentToolCallUpdateEvent, type AgentTurnEndEvent, type AgentTurnTiming, type StreamDiagnosticFn, type StructuredToolResult, type ToolContext, type ToolExecuteResult, type ToolExecutionMode, type TransformContextOptions, agentLoop, isAbortError, isBillingError, isContextOverflow, isUsageLimitError, setStreamDiagnostic };
+/**
+ * A local inference server (llama.cpp, vLLM, Ollama, LM Studio) can spend
+ * minutes prefilling a large prompt before it emits its first token. The
+ * first-event watchdog that protects hosted streams turns that into an abort
+ * → retry → cold-prefill loop that never converges, so it is disabled for
+ * loopback backends.
+ */
+declare function isLocalBackendUrl(baseUrl?: string): boolean;
+
+export { Agent, type AgentDoneEvent, type AgentErrorEvent, type AgentEvent, type AgentFollowUpMessageEvent, type AgentOptions, type AgentResult, type AgentRetryEvent, type AgentServerToolCallEvent, type AgentServerToolResultEvent, type AgentSteeringMessageEvent, AgentStream, type AgentTextDeltaEvent, type AgentThinkingDeltaEvent, type AgentTool, type AgentToolCallDeltaEvent, type AgentToolCallEndEvent, type AgentToolCallStartEvent, type AgentToolCallUpdateEvent, type AgentTurnEndEvent, type AgentTurnTiming, type StreamDiagnosticFn, type StructuredToolResult, type ToolContext, type ToolExecuteResult, type ToolExecutionMode, type TransformContextOptions, agentLoop, isAbortError, isBillingError, isContextOverflow, isLocalBackendUrl, isUsageLimitError, setStreamDiagnostic };

package/dist/index.d.ts

@@ -324,4 +324,13 @@ declare function isBillingError(err: unknown): boolean;
 declare function isUsageLimitError(err: unknown): boolean;
 declare function agentLoop(messages: Message[], options: AgentOptions): AsyncGenerator<AgentEvent, AgentResult>;
 
-export { Agent, type AgentDoneEvent, type AgentErrorEvent, type AgentEvent, type AgentFollowUpMessageEvent, type AgentOptions, type AgentResult, type AgentRetryEvent, type AgentServerToolCallEvent, type AgentServerToolResultEvent, type AgentSteeringMessageEvent, AgentStream, type AgentTextDeltaEvent, type AgentThinkingDeltaEvent, type AgentTool, type AgentToolCallDeltaEvent, type AgentToolCallEndEvent, type AgentToolCallStartEvent, type AgentToolCallUpdateEvent, type AgentTurnEndEvent, type AgentTurnTiming, type StreamDiagnosticFn, type StructuredToolResult, type ToolContext, type ToolExecuteResult, type ToolExecutionMode, type TransformContextOptions, agentLoop, isAbortError, isBillingError, isContextOverflow, isUsageLimitError, setStreamDiagnostic };
+/**
+ * A local inference server (llama.cpp, vLLM, Ollama, LM Studio) can spend
+ * minutes prefilling a large prompt before it emits its first token. The
+ * first-event watchdog that protects hosted streams turns that into an abort
+ * → retry → cold-prefill loop that never converges, so it is disabled for
+ * loopback backends.
+ */
+declare function isLocalBackendUrl(baseUrl?: string): boolean;
+
+export { Agent, type AgentDoneEvent, type AgentErrorEvent, type AgentEvent, type AgentFollowUpMessageEvent, type AgentOptions, type AgentResult, type AgentRetryEvent, type AgentServerToolCallEvent, type AgentServerToolResultEvent, type AgentSteeringMessageEvent, AgentStream, type AgentTextDeltaEvent, type AgentThinkingDeltaEvent, type AgentTool, type AgentToolCallDeltaEvent, type AgentToolCallEndEvent, type AgentToolCallStartEvent, type AgentToolCallUpdateEvent, type AgentTurnEndEvent, type AgentTurnTiming, type StreamDiagnosticFn, type StructuredToolResult, type ToolContext, type ToolExecuteResult, type ToolExecutionMode, type TransformContextOptions, agentLoop, isAbortError, isBillingError, isContextOverflow, isLocalBackendUrl, isUsageLimitError, setStreamDiagnostic };

package/dist/index.js

@@ -10,6 +10,30 @@ import {
   isHardBillingMessage,
   redactValue
 } from "@kenkaiiii/gg-ai";
+
+// src/local-backend.ts
+function isLocalBackendUrl(baseUrl) {
+  if (!baseUrl) return false;
+  let host;
+  try {
+    host = new URL(baseUrl).hostname.toLowerCase();
+  } catch {
+    return false;
+  }
+  if (host === "[::1]" || host === "::1") return true;
+  if (host === "localhost" || host.endsWith(".localhost")) return true;
+  if (host === "0.0.0.0") return true;
+  if (host.endsWith(".local")) return true;
+  const ipv4 = /^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/.exec(host);
+  if (ipv4) {
+    const octets = ipv4.slice(1).map(Number);
+    if (octets.some((n) => n > 255)) return false;
+    return octets[0] === 127;
+  }
+  return false;
+}
+
+// src/agent-loop.ts
 var DEFAULT_MAX_TURNS = 300;
 var _diagFn = null;
 function setStreamDiagnostic(fn) {
@@ -229,8 +253,9 @@ async function* agentLoop(messages, options) {
   const STREAM_THINKING_HARD_TIMEOUT_MS = 6e5;
   const NON_STREAMING_HARD_TIMEOUT_MS = 3e5;
   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 localBackend = isLocalBackendUrl(options.baseUrl);
+  const firstEventTimeoutMs = localBackend ? Number.POSITIVE_INFINITY : usesSilentReasoningBudget ? STREAM_THINKING_IDLE_TIMEOUT_MS : STREAM_FIRST_EVENT_TIMEOUT_MS;
+  const initialHardTimeoutMs = localBackend || 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;
@@ -261,7 +286,8 @@ async function* agentLoop(messages, options) {
           model: options.model,
           thinking: options.thinking ?? "off",
           firstEventTimeoutMs,
-          initialHardTimeoutMs
+          initialHardTimeoutMs,
+          localBackend
         });
       }
       if (firstTurn && options.getSteeringMessages) {
@@ -319,6 +345,7 @@ async function* agentLoop(messages, options) {
         if (useNonStreamingFallback) return;
         if (idleTimer) clearTimeout(idleTimer);
         const timeoutMs = hasReceivedEvent ? STREAM_IDLE_TIMEOUT_MS : hasReceivedThinking ? STREAM_THINKING_IDLE_TIMEOUT_MS : firstEventTimeoutMs;
+        if (!Number.isFinite(timeoutMs)) return;
         idleTimer = setTimeout(() => {
           diag("idle_timeout_fired", {
             events: streamEventCount,
@@ -1535,6 +1562,7 @@ export {
   isAbortError,
   isBillingError,
   isContextOverflow,
+  isLocalBackendUrl,
   isUsageLimitError,
   setStreamDiagnostic
 };