@prestyj/agent 5.11.0 → 5.13.0

package/dist/index.d.cts

@@ -55,6 +55,13 @@ interface AgentToolCallEndEvent {
     details?: unknown;
     isError: boolean;
     durationMs: number;
+    /**
+     * Set only when the call failed schema validation: how many consecutive
+     * times this tool produced this same validation error. 1 means the model
+     * still has room to self-correct; 3 is the threshold that ends the turn.
+     * Logged so a retry loop shows up as a count instead of identical lines.
+     */
+    invalidArgAttempt?: number;
 }
 interface AgentTurnTiming {
     /** Logical turn start, before context transforms or provider retries. Unix epoch milliseconds. */
@@ -133,7 +140,7 @@ interface AgentTurnBudgetExtendedEvent {
  */
 interface AgentTruncatedEvent {
     type: "truncated";
-    reason: "max_tokens" | "refusal" | "provider_error";
+    reason: "max_tokens" | "refusal" | "provider_error" | "empty_response";
     /** True when the loop injected a continuation and will keep going. */
     continued: boolean;
 }
@@ -215,6 +222,22 @@ interface AgentOptions {
     temperature?: number;
     thinking?: StreamOptions["thinking"];
     apiKey?: string;
+    /**
+     * Re-resolve the credential at the start of every turn. A run can span many
+     * minutes, and an OAuth grant refreshed by any process (another app window, a
+     * CLI session, the usage poller) invalidates the access token captured when
+     * the run began — so a pinned `apiKey` goes dead mid-run and every remaining
+     * turn fails with an authentication error. Returning the current credential
+     * here keeps a long run alive across rotations.
+     *
+     * Falls back to `apiKey`/`accountId`/`projectId` when omitted or when the
+     * resolver throws (the provider call then surfaces the real auth error).
+     */
+    resolveCredentials?: () => Promise<{
+        apiKey: string;
+        accountId?: string;
+        projectId?: string;
+    }>;
     baseUrl?: string;
     signal?: AbortSignal;
     accountId?: string;
@@ -374,7 +397,7 @@ declare function isBillingError(err: unknown): boolean;
  * plan running out of usage). Unlike a transient per-minute 429, this does NOT
  * clear with a quick retry — the user must wait for the window to reset — so the
  * loop surfaces it immediately instead of retrying for minutes. Matches the
- * canonical message gg-ai stamps onto the provider error.
+ * canonical message @prestyj/ai stamps onto the provider error.
  */
 declare function isUsageLimitError(err: unknown): boolean;
 declare function agentLoop(messages: Message[], options: AgentOptions): AsyncGenerator<AgentEvent, AgentResult>;

package/dist/index.d.ts

@@ -55,6 +55,13 @@ interface AgentToolCallEndEvent {
     details?: unknown;
     isError: boolean;
     durationMs: number;
+    /**
+     * Set only when the call failed schema validation: how many consecutive
+     * times this tool produced this same validation error. 1 means the model
+     * still has room to self-correct; 3 is the threshold that ends the turn.
+     * Logged so a retry loop shows up as a count instead of identical lines.
+     */
+    invalidArgAttempt?: number;
 }
 interface AgentTurnTiming {
     /** Logical turn start, before context transforms or provider retries. Unix epoch milliseconds. */
@@ -133,7 +140,7 @@ interface AgentTurnBudgetExtendedEvent {
  */
 interface AgentTruncatedEvent {
     type: "truncated";
-    reason: "max_tokens" | "refusal" | "provider_error";
+    reason: "max_tokens" | "refusal" | "provider_error" | "empty_response";
     /** True when the loop injected a continuation and will keep going. */
     continued: boolean;
 }
@@ -215,6 +222,22 @@ interface AgentOptions {
     temperature?: number;
     thinking?: StreamOptions["thinking"];
     apiKey?: string;
+    /**
+     * Re-resolve the credential at the start of every turn. A run can span many
+     * minutes, and an OAuth grant refreshed by any process (another app window, a
+     * CLI session, the usage poller) invalidates the access token captured when
+     * the run began — so a pinned `apiKey` goes dead mid-run and every remaining
+     * turn fails with an authentication error. Returning the current credential
+     * here keeps a long run alive across rotations.
+     *
+     * Falls back to `apiKey`/`accountId`/`projectId` when omitted or when the
+     * resolver throws (the provider call then surfaces the real auth error).
+     */
+    resolveCredentials?: () => Promise<{
+        apiKey: string;
+        accountId?: string;
+        projectId?: string;
+    }>;
     baseUrl?: string;
     signal?: AbortSignal;
     accountId?: string;
@@ -374,7 +397,7 @@ declare function isBillingError(err: unknown): boolean;
  * plan running out of usage). Unlike a transient per-minute 429, this does NOT
  * clear with a quick retry — the user must wait for the window to reset — so the
  * loop surfaces it immediately instead of retrying for minutes. Matches the
- * canonical message gg-ai stamps onto the provider error.
+ * canonical message @prestyj/ai stamps onto the provider error.
  */
 declare function isUsageLimitError(err: unknown): boolean;
 declare function agentLoop(messages: Message[], options: AgentOptions): AsyncGenerator<AgentEvent, AgentResult>;

package/dist/index.js

@@ -57,7 +57,13 @@ function isContextOverflow(err) {
   if (overflowStatus === 402) return false;
   if (isBillingError(err)) return false;
   const msg = err.message.toLowerCase();
-  return msg.includes("prompt is too long") || msg.includes("prompt too long") || msg.includes("input is too long") || msg.includes("context_length_exceeded") || msg.includes("context_window_exceeded") || msg.includes("maximum context length") || msg.includes("exceeds model context window") || msg.includes("exceeds the context window") || msg.includes("content_too_large") || msg.includes("request_too_large") || msg.includes("reduce the length") || msg.includes("please shorten") || msg.includes("token") && msg.includes("exceed");
+  if (msg.includes("prompt is too long") || msg.includes("prompt too long") || msg.includes("input is too long") || msg.includes("context_length_exceeded") || msg.includes("context_window_exceeded") || msg.includes("maximum context length") || msg.includes("exceeds model context window") || msg.includes("exceeds the context window") || msg.includes("content_too_large") || msg.includes("request_too_large") || msg.includes("reduce the length") || msg.includes("please shorten")) {
+    return true;
+  }
+  const rateLimited = overflowStatus === 429 || msg.includes("rate limit") || msg.includes("rate_limit") || msg.includes("too many requests");
+  const perUnitTime = msg.includes("per min") || msg.includes("/min") || msg.includes("per minute") || msg.includes("per hour") || msg.includes("per day") || msg.includes("tpm") || msg.includes("rpm");
+  if (rateLimited && perUnitTime) return false;
+  return msg.includes("token") && msg.includes("exceed");
 }
 function parseOverflowNumber(value) {
   return Number(value.replace(/[,_\s]/g, ""));
@@ -170,6 +176,17 @@ function isMalformedStream(err) {
   const msg = err.message;
   return /\bin JSON at position \d+/i.test(msg);
 }
+var TIMEOUT_NAMES = /* @__PURE__ */ new Set(["TimeoutError", "ConnectTimeoutError", "HeadersTimeoutError"]);
+var TIMEOUT_MESSAGES = [
+  /^request timed out\.?$/i,
+  /\brequest to [\w .-]+ timed out\b/i,
+  /\b(?:connection|socket|headers|stream) timed out\b/i
+];
+function isBareTimeout(e) {
+  if (typeof e.name === "string" && TIMEOUT_NAMES.has(e.name)) return true;
+  if (typeof e.message !== "string") return false;
+  return TIMEOUT_MESSAGES.some((re) => re.test(e.message));
+}
 function isTransportFailure(err) {
   const codes = /* @__PURE__ */ new Set([
     "ECONNRESET",
@@ -206,6 +223,8 @@ function isTransportFailure(err) {
     if (typeof e.message === "string") {
       for (const re of messages) if (re.test(e.message)) return true;
     }
+    const clientError = typeof e.status === "number" && e.status >= 400 && e.status < 500;
+    if (!clientError && isBareTimeout(e)) return true;
     cur = e.cause;
   }
   return false;
@@ -431,6 +450,21 @@ async function* agentLoop(messages, options) {
         diag("stream_call", { nonStreaming: useNonStreamingFallback });
         streamCallStart = Date.now();
         providerAttemptStartedAt = streamCallStart;
+        let liveApiKey = options.apiKey;
+        let liveAccountId = options.accountId;
+        let liveProjectId = options.projectId;
+        if (options.resolveCredentials) {
+          try {
+            const fresh = await options.resolveCredentials();
+            liveApiKey = fresh.apiKey;
+            if (fresh.accountId !== void 0) liveAccountId = fresh.accountId;
+            if (fresh.projectId !== void 0) liveProjectId = fresh.projectId;
+          } catch (credErr) {
+            diag("credential_refresh_failed", {
+              error: (credErr instanceof Error ? credErr.message : String(credErr)).slice(0, 200)
+            });
+          }
+        }
         const result = stream({
           provider: options.provider,
           model: options.model,
@@ -442,12 +476,12 @@ async function* agentLoop(messages, options) {
           maxTokens: options.maxTokens,
           temperature: options.temperature,
           thinking: options.thinking,
-          apiKey: options.apiKey,
+          apiKey: liveApiKey,
           baseUrl: options.baseUrl,
           signal: streamController.signal,
-          accountId: options.accountId,
+          accountId: liveAccountId,
           transportSessionId: options.transportSessionId,
-          projectId: options.projectId,
+          projectId: liveProjectId,
           cacheRetention: options.cacheRetention,
           promptCacheKey: options.promptCacheKey,
           serviceTier: options.serviceTier,
@@ -891,6 +925,7 @@ async function* agentLoop(messages, options) {
           continue;
         }
       }
+      const emptyExhausted = !hasActionableContent;
       emptyResponseRetries = 0;
       useNonStreamingFallback = false;
       totalUsage.inputTokens += response.usage.inputTokens;
@@ -904,9 +939,11 @@ async function* agentLoop(messages, options) {
       if (response.usage.cacheWrite) {
         totalUsage.cacheWrite = (totalUsage.cacheWrite ?? 0) + response.usage.cacheWrite;
       }
-      messages.push(response.message);
-      latestProviderUsage = response.usage;
-      usageAnchorIndex = messages.length - 1;
+      if (!emptyExhausted) {
+        messages.push(response.message);
+        latestProviderUsage = response.usage;
+        usageAnchorIndex = messages.length - 1;
+      }
       const completedAt = Date.now();
       const outputTokensPerSecond = providerDurationMs > 0 && response.usage.outputTokens > 0 ? response.usage.outputTokens / (providerDurationMs / 1e3) : void 0;
       const timing = {
@@ -938,8 +975,15 @@ async function* agentLoop(messages, options) {
       }
       consecutivePauses = 0;
       const allToolCalls = extractToolCalls(response.message.content);
-      if (response.stopReason !== "tool_use" && allToolCalls.length === 0) {
-        if (response.stopReason === "max_tokens") {
+      if (emptyExhausted || response.stopReason !== "tool_use" && allToolCalls.length === 0) {
+        if (emptyExhausted) {
+          diag("empty_response_exhausted", {
+            provider: options.provider,
+            model: options.model,
+            stopReason: response.stopReason
+          });
+          yield { type: "truncated", reason: "empty_response", continued: false };
+        } else if (response.stopReason === "max_tokens") {
           if (maxTokensContinuations < MAX_OUTPUT_CONTINUATIONS) {
             maxTokensContinuations++;
             diag("max_tokens_continuation", {
@@ -1161,6 +1205,7 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
   let resultContent;
   let details;
   let isError = false;
+  let invalidArgAttempt;
   const tool = options.toolMap.get(toolCall.name);
   if (!tool) {
     resultContent = `Unknown tool: ${toolCall.name}`;
@@ -1198,6 +1243,7 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
         const failureKey = `${toolCall.name}:${prettyError}`;
         const failureCount = (options.invalidToolArgumentCounts.get(failureKey) ?? 0) + 1;
         options.invalidToolArgumentCounts.set(failureKey, failureCount);
+        invalidArgAttempt = failureCount;
         resultContent = `Invalid arguments for tool \`${toolCall.name}\`:
 ` + prettyError + "\nRe-issue the call with each field as the correct type.";
         if (failureCount >= 3) {
@@ -1228,7 +1274,8 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
     result: toolResultPreview(resultContent),
     details,
     isError,
-    durationMs
+    durationMs,
+    ...invalidArgAttempt === void 0 ? {} : { invalidArgAttempt }
   });
   return { toolCallId: toolCall.id, content: resultContent, isError };
 }