jeopi-agent-core 16.2.28 → 16.2.30

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 ## [Unreleased]
 
+## [16.2.30] - 2026-07-10
+
+### Fixed
+
+- The agent loop now auto-continues a text response truncated by the output-token limit (`stopReason: "length"` with no tool calls, e.g. Sonnet 5 hitting `max_tokens` mid-answer): a synthetic "continue from where you were cut off" user turn is injected and the loop re-samples instead of silently stopping with a cut-off message. Capped at 8 consecutive continuations (`MAX_LENGTH_TRUNCATE_CONTINUATIONS`); any tool call resets the counter. Truncated turns that carry tool calls keep the existing abandoned-toolcall path.
+- Steering/aside dequeues at the turn boundary no longer overwrite already-queued pending messages (the injected length-continuation turn was previously dropped when steering arrived in the same window); they now append.
+
 ## [16.2.26] - 2026-07-05
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "jeopi-agent-core",
-	"version": "16.2.28",
+	"version": "16.2.30",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://github.com/akillness/jeopi",
 	"author": "Can Boluk",
@@ -35,12 +35,12 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"jeopi-ai": "16.2.28",
-		"jeopi-catalog": "16.2.28",
-		"jeopi-natives": "16.2.28",
-		"jeopi-utils": "16.2.28",
-		"jeopi-wire": "16.2.28",
-		"jeopi-snapcompact": "16.2.28",
+		"jeopi-ai": "16.2.30",
+		"jeopi-catalog": "16.2.30",
+		"jeopi-natives": "16.2.30",
+		"jeopi-utils": "16.2.30",
+		"jeopi-wire": "16.2.30",
+		"jeopi-snapcompact": "16.2.30",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/agent-loop.ts

@@ -16,6 +16,7 @@ import {
 	type ToolResultMessage,
 	type TSchema,
 	toolWireSchema,
+	type UserMessage,
 	validateToolArguments,
 } from "jeopi-ai";
 import {
@@ -84,6 +85,12 @@ const ABORTED: unique symbol = Symbol("agent-loop-aborted");
  */
 const MAX_PAUSED_TURN_CONTINUATIONS = 8;
 
+/**
+ * Cap on consecutive re-samples triggered by output truncation stop
+ * (`stopReason === "length"`) without an intervening tool call.
+ */
+const MAX_LENGTH_TRUNCATE_CONTINUATIONS = 8;
+
 /**
  * Cap on consecutive forced escalations for a single soft tool requirement.
  * A forced `toolChoice` guarantees the call, so this is purely defensive: if a
@@ -758,6 +765,7 @@ async function runLoopBody(
 		let harmonyRetryAttempt = 0;
 		let harmonyTruncateResumeCount = 0;
 		let pausedTurnContinuations = 0;
+		let lengthTruncateContinuations = 0;
 
 		// Soft tool requirement lifecycle (reminder → escalate; see SoftToolRequirement).
 		// `forcedToolChoice` carries a one-turn escalation into the next model call. It
@@ -1042,6 +1050,7 @@ async function runLoopBody(
 
 				if (toolCalls.length > 0) {
 					pausedTurnContinuations = 0;
+					lengthTruncateContinuations = 0;
 				} else if (
 					!hasMoreToolCalls &&
 					message.stopReason === "stop" &&
@@ -1055,6 +1064,22 @@ async function runLoopBody(
 					// mid-work turn.
 					pausedTurnContinuations++;
 					hasMoreToolCalls = true;
+				} else if (
+					!hasMoreToolCalls &&
+					message.stopReason === "length" &&
+					lengthTruncateContinuations < MAX_LENGTH_TRUNCATE_CONTINUATIONS &&
+					!deadlinePassed
+				) {
+					// Prompt the model to continue the truncated response
+					lengthTruncateContinuations++;
+					const continuationMessage: UserMessage = {
+						role: "user",
+						content: "Continue your response exactly from where it was cut off.",
+						synthetic: true,
+						timestamp: Date.now(),
+					};
+					pendingMessages = [continuationMessage];
+					hasMoreToolCalls = true;
 				}
 
 				await emitTurnEnd(stream, currentContext, message, toolResults, config, signal, {
@@ -1075,12 +1100,13 @@ async function runLoopBody(
 				if (hasMoreToolCalls) {
 					// Mid-work: fold any non-interrupting asides into the next turn alongside steering.
 					const asides = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.());
-					pendingMessages = asides.length > 0 ? [...steering, ...asides] : steering;
+					const basePending = asides.length > 0 ? [...steering, ...asides] : steering;
+					pendingMessages = [...pendingMessages, ...basePending];
 				} else {
 					// Stop boundary: only steering (live user input) forces another turn here. Leave
 					// asides for the outer drain below so a passive aside can't trigger an extra model
 					// turn ahead of a queued follow-up — the outer drain batches asides + follow-ups together.
-					pendingMessages = steering;
+					pendingMessages = [...pendingMessages, ...steering];
 				}
 			}