@prestyj/agent 5.9.0 → 5.10.0

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 "@prestyj/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) {
@@ -276,9 +300,10 @@ 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 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;
@@ -306,7 +331,11 @@ 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,
+          localBackend
         });
       }
       if (firstTurn && options.getSteeringMessages) {
@@ -364,6 +393,7 @@ 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,
@@ -748,15 +778,29 @@ 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 EZCoderAIError(
+              `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 EZ Coder through firewall or antivirus web protection.",
+                cause: err
+              }
             )
           };
           break;
@@ -1575,6 +1619,7 @@ export {
   isAbortError,
   isBillingError,
   isContextOverflow,
+  isLocalBackendUrl,
   isUsageLimitError,
   setStreamDiagnostic
 };