@kenkaiiii/gg-agent 5.19.6 → 5.20.0

package/dist/index.d.cts

@@ -87,6 +87,19 @@ interface AgentMaxTurnsEvent {
     totalTurns: number;
     maxTurns: number;
 }
+/**
+ * Warning signal emitted when a turn ended on a non-clean stop reason —
+ * `max_tokens` (output clipped at the model's output-token limit), `refusal`,
+ * or a provider-reported `error` stop. Distinguishes a truncated/degraded
+ * completion from a clean one so hosts can warn the user instead of silently
+ * presenting incomplete output as done.
+ */
+interface AgentTruncatedEvent {
+    type: "truncated";
+    reason: "max_tokens" | "refusal" | "provider_error";
+    /** True when the loop injected a continuation and will keep going. */
+    continued: boolean;
+}
 interface AgentRetryEvent {
     type: "retry";
     reason: "overloaded" | "rate_limit" | "provider_error" | "empty_response" | "stream_stall" | "overflow_compact" | "tool_argument_glitch";
@@ -135,7 +148,7 @@ interface AgentFollowUpMessageEvent {
     type: "follow_up_message";
     content: Message["content"];
 }
-type AgentEvent = AgentTextDeltaEvent | AgentThinkingDeltaEvent | AgentToolCallStartEvent | AgentToolCallUpdateEvent | AgentToolCallEndEvent | AgentToolCallDeltaEvent | AgentServerToolCallEvent | AgentServerToolResultEvent | AgentSteeringMessageEvent | AgentFollowUpMessageEvent | AgentRetryEvent | AgentTurnEndEvent | AgentDoneEvent | AgentMaxTurnsEvent | AgentErrorEvent;
+type AgentEvent = AgentTextDeltaEvent | AgentThinkingDeltaEvent | AgentToolCallStartEvent | AgentToolCallUpdateEvent | AgentToolCallEndEvent | AgentToolCallDeltaEvent | AgentServerToolCallEvent | AgentServerToolResultEvent | AgentSteeringMessageEvent | AgentFollowUpMessageEvent | AgentRetryEvent | AgentTurnEndEvent | AgentDoneEvent | AgentMaxTurnsEvent | AgentTruncatedEvent | AgentErrorEvent;
 interface TransformContextOptions {
     /** Force a transform after the provider reports context overflow. */
     force?: boolean;

package/dist/index.d.ts

@@ -87,6 +87,19 @@ interface AgentMaxTurnsEvent {
     totalTurns: number;
     maxTurns: number;
 }
+/**
+ * Warning signal emitted when a turn ended on a non-clean stop reason —
+ * `max_tokens` (output clipped at the model's output-token limit), `refusal`,
+ * or a provider-reported `error` stop. Distinguishes a truncated/degraded
+ * completion from a clean one so hosts can warn the user instead of silently
+ * presenting incomplete output as done.
+ */
+interface AgentTruncatedEvent {
+    type: "truncated";
+    reason: "max_tokens" | "refusal" | "provider_error";
+    /** True when the loop injected a continuation and will keep going. */
+    continued: boolean;
+}
 interface AgentRetryEvent {
     type: "retry";
     reason: "overloaded" | "rate_limit" | "provider_error" | "empty_response" | "stream_stall" | "overflow_compact" | "tool_argument_glitch";
@@ -135,7 +148,7 @@ interface AgentFollowUpMessageEvent {
     type: "follow_up_message";
     content: Message["content"];
 }
-type AgentEvent = AgentTextDeltaEvent | AgentThinkingDeltaEvent | AgentToolCallStartEvent | AgentToolCallUpdateEvent | AgentToolCallEndEvent | AgentToolCallDeltaEvent | AgentServerToolCallEvent | AgentServerToolResultEvent | AgentSteeringMessageEvent | AgentFollowUpMessageEvent | AgentRetryEvent | AgentTurnEndEvent | AgentDoneEvent | AgentMaxTurnsEvent | AgentErrorEvent;
+type AgentEvent = AgentTextDeltaEvent | AgentThinkingDeltaEvent | AgentToolCallStartEvent | AgentToolCallUpdateEvent | AgentToolCallEndEvent | AgentToolCallDeltaEvent | AgentServerToolCallEvent | AgentServerToolResultEvent | AgentSteeringMessageEvent | AgentFollowUpMessageEvent | AgentRetryEvent | AgentTurnEndEvent | AgentDoneEvent | AgentMaxTurnsEvent | AgentTruncatedEvent | AgentErrorEvent;
 interface TransformContextOptions {
     /** Force a transform after the provider reports context overflow. */
     force?: boolean;

package/dist/index.js

@@ -213,6 +213,9 @@ async function* agentLoop(messages, options) {
   const STALL_DELAY_MS = 1e3;
   const MIN_PARTIAL_PRESERVE_CHARS = 200;
   const PARTIAL_CONTINUATION_PROMPT = "[Your previous response was cut off by a connection failure. The text above is what was already delivered to the user. Continue exactly from where it stopped \u2014 do not repeat or restart it.]";
+  const MAX_OUTPUT_CONTINUATIONS = 2;
+  const MAX_TOKENS_CONTINUATION_PROMPT = "[Your previous response hit the output-token limit and was cut off. The text above is what was already delivered to the user. Continue exactly from where it stopped \u2014 do not repeat or restart it.]";
+  let maxTokensContinuations = 0;
   const OVERLOAD_BASE_DELAY_MS = 2e3;
   const OVERLOAD_MAX_DELAY_MS = 3e4;
   const STREAM_FIRST_EVENT_TIMEOUT_MS = 45e3;
@@ -785,6 +788,27 @@ 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 (maxTokensContinuations < MAX_OUTPUT_CONTINUATIONS) {
+            maxTokensContinuations++;
+            diag("max_tokens_continuation", {
+              attempt: maxTokensContinuations,
+              maxAttempts: MAX_OUTPUT_CONTINUATIONS,
+              provider: options.provider,
+              model: options.model
+            });
+            yield { type: "truncated", reason: "max_tokens", continued: true };
+            messages.push({ role: "user", content: MAX_TOKENS_CONTINUATION_PROMPT });
+            continue;
+          }
+          yield { type: "truncated", reason: "max_tokens", continued: false };
+        } else if (response.stopReason === "refusal" || response.stopReason === "error") {
+          yield {
+            type: "truncated",
+            reason: response.stopReason === "refusal" ? "refusal" : "provider_error",
+            continued: false
+          };
+        }
         if (options.getSteeringMessages) {
           const steering = await options.getSteeringMessages();
           if (steering && steering.length > 0) {