@prestyj/agent 5.12.0 → 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;
 }

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;
 }

package/dist/index.js

@@ -925,6 +925,7 @@ async function* agentLoop(messages, options) {
           continue;
         }
       }
+      const emptyExhausted = !hasActionableContent;
       emptyResponseRetries = 0;
       useNonStreamingFallback = false;
       totalUsage.inputTokens += response.usage.inputTokens;
@@ -938,9 +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 = {
@@ -972,8 +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", {
@@ -1195,6 +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: ${toolCall.name}`;
@@ -1232,6 +1243,7 @@ async function executeSingleToolCall(toolCall, options, pushEvent) {
         const failureKey = `${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) {
@@ -1262,7 +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 };
 }