jeopi-agent-core 16.4.3 → 16.4.5

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [16.4.5] - 2026-08-03
+
+### Fixed
+
+- Fixed the agent stopping mid-task with raw `<invoke name="…">` markup printed in the reply. When a native-tool-calling model emitted its tool call as visible text instead of a `tool_use` block, the turn carried zero runnable calls, so the loop treated it as a finished answer and the announced work never ran. Such turns are now re-materialized into real tool calls and dispatched normally; the leaked markup is stripped from the visible text. Turns that merely quote tool syntax inside a fenced code block are left untouched, and owned/in-band dialect sessions are unaffected (their stream wrapper already handles this).
+
 ## [16.4.3] - 2026-07-22
 
 ### Fixed

package/dist/types/compaction/utils.d.ts

@@ -50,8 +50,11 @@ export declare function upsertFileOperations(summary: string, readFiles: string[
  */
 export declare function truncateToolResultForSummary(text: string): string;
 /**
- * Serialize LLM messages to text for summarization.
- * This prevents the model from treating it as a conversation to continue.
+ * Serialize LLM messages as plain summary input without provider control tokens.
+ */
+export declare function serializeConversationForSummary(messages: Message[], dialect?: Dialect): string;
+/**
+ * Serialize LLM messages to transcript text.
  * Call convertToLlm() first to handle custom message types.
  */
 export declare function serializeConversation(messages: Message[], dialect?: Dialect): string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "jeopi-agent-core",
-	"version": "16.4.3",
+	"version": "16.4.5",
 	"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": "16.4.3",
-		"jeopi-catalog": "16.4.3",
-		"jeopi-natives": "16.4.3",
-		"jeopi-utils": "16.4.3",
-		"jeopi-wire": "16.4.3",
-		"jeopi-snapcompact": "16.4.3",
+		"jeopi-ai": "16.4.5",
+		"jeopi-catalog": "16.4.5",
+		"jeopi-natives": "16.4.5",
+		"jeopi-utils": "16.4.5",
+		"jeopi-wire": "16.4.5",
+		"jeopi-snapcompact": "16.4.5",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/agent-loop.ts

@@ -24,6 +24,7 @@ import {
 	encodeInbandToolHistory,
 	renderInbandToolPrompt,
 	renderToolExamples,
+	salvageLeakedToolCalls,
 	wrapInbandToolStream,
 } from "jeopi-ai/dialect";
 import * as AIError from "jeopi-ai/error";
@@ -38,7 +39,7 @@ import {
 	signalListLabel,
 } from "jeopi-ai/utils/harmony-leak";
 import { preferredDialect } from "jeopi-catalog/identity";
-import { sanitizeText, structuredCloneJSON } from "jeopi-utils";
+import { logger, sanitizeText, structuredCloneJSON } from "jeopi-utils";
 import { INTENT_FIELD } from "jeopi-wire";
 import { type AgentRunCoverage, type AgentRunSummary, ToolCallBlockedError } from "./run-collector";
 import {
@@ -1233,6 +1234,12 @@ async function streamAssistantResponse(
 		};
 	}
 
+	// Tool catalog as sent natively this turn. Empty under an owned dialect (the
+	// wire carries no `tools`), which is exactly when leaked-markup salvage must
+	// NOT run — that path already re-materializes in-band calls via the stream
+	// wrapper, and double-parsing would duplicate every call.
+	const nativeWireTools: Context["tools"] = ownedDialect ? undefined : llmContext.tools;
+
 	const streamFunction = streamFn || streamSimple;
 
 	const dynamicReasoning = config.getReasoning?.();
@@ -1427,6 +1434,7 @@ async function streamAssistantResponse(
 								throw new HarmonyLeakInterruption(detection, removed, recovered);
 							}
 						}
+						finalMessage = applyLeakedToolCallSalvage(finalMessage, nativeWireTools, config);
 						finalMessage = snapshotAssistantMessage(finalMessage);
 						// Expand inline macros (and any other registered rewrite) on the
 						// finalized message before it reaches the context, the UI, or tool
@@ -1529,6 +1537,7 @@ async function streamAssistantResponse(
 					throw new HarmonyLeakInterruption(detection, removed, recovered);
 				}
 			}
+			trailing = applyLeakedToolCallSalvage(trailing, nativeWireTools, config);
 			trailing = snapshotAssistantMessage(trailing);
 			if (addedPartial) {
 				context.messages[context.messages.length - 1] = trailing;
@@ -1606,6 +1615,30 @@ function recoverTransientErrorToolTurn(
 	};
 }
 
+/**
+ * Re-materialize tool calls a native-tool-calling model wrote as visible text.
+ *
+ * Without this the turn holds zero runnable tool calls, so the loop treats it
+ * as a finished answer: the raw `<invoke …>` markup renders to the user and the
+ * work the model announced never runs. `salvageLeakedToolCalls` only fires on
+ * unambiguous leaks (see its guards); anything else returns the message
+ * untouched.
+ */
+function applyLeakedToolCallSalvage(
+	message: AssistantMessage,
+	nativeWireTools: Context["tools"],
+	config: AgentLoopConfig,
+): AssistantMessage {
+	const salvaged = salvageLeakedToolCalls(message, nativeWireTools);
+	if (!salvaged) return message;
+	logger.warn("Recovered tool calls emitted as text", {
+		model: config.model.id,
+		provider: config.model.provider,
+		tools: salvaged.calls.map(call => call.name),
+	});
+	return salvaged.message;
+}
+
 function emitDiscardedHarmonyPartial(
 	partialMessage: AssistantMessage | null,
 	stream: EventStream<AgentEvent, AgentMessage[]>,

package/src/compaction/branch-summarization.ts

@@ -27,7 +27,7 @@ import {
 	extractFileOpsFromMessage,
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
-	serializeConversation,
+	serializeConversationForSummary,
 	stripReadSelector,
 	truncateToolResultForSummary,
 	upsertFileOperations,
@@ -320,7 +320,7 @@ export async function generateBranchSummary(
 	// Transform to LLM-compatible messages, then serialize to text
 	// Serialization prevents the model from treating it as a conversation to continue
 	const llmMessages = (options.convertToLlm ?? defaultConvertToLlm)(messages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	// Build prompt
 	const instructions = customInstructions || BRANCH_SUMMARY_PROMPT;

package/src/compaction/compaction.ts

@@ -63,7 +63,7 @@ import {
 	extractFileOpsFromMessage,
 	type FileOperations,
 	SUMMARIZATION_SYSTEM_PROMPT,
-	serializeConversation,
+	serializeConversationForSummary,
 	stripReadSelector,
 	upsertFileOperations,
 } from "./utils";
@@ -786,7 +786,7 @@ export async function generateSummary(
 	// Serialize conversation to text so model doesn't try to continue it
 	// Convert to LLM messages first (handles custom app messages when caller provides a transformer).
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(currentMessages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	// Build the prompt with conversation wrapped in tags
 	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
@@ -983,7 +983,7 @@ async function generateShortSummary(
 ): Promise<string> {
 	const maxTokens = Math.min(512, Math.floor(0.2 * reserveTokens));
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(recentMessages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 
 	let promptText = `<conversation>\n${conversationText}\n</conversation>\n\n`;
 	if (historySummary) {
@@ -1518,7 +1518,7 @@ async function generateTurnPrefixSummary(
 	const maxTokens = Math.floor(0.5 * reserveTokens); // Smaller budget for turn prefix
 
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(messages);
-	const conversationText = serializeConversation(llmMessages, preferredDialect(model.id));
+	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));
 	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
 	const summarizationMessages = [
 		{

package/src/compaction/utils.ts

@@ -207,9 +207,19 @@ export function truncateToolResultForSummary(text: string): string {
 	return `${text.slice(0, TOOL_RESULT_MAX_CHARS)}\n\n[... ${truncatedChars} more characters truncated]`;
 }
 
+const HARMONY_CONTROL_TOKEN_RE = /<\|(start|end|message|channel|constrain|return|call)\|>/g;
+
+/**
+ * Serialize LLM messages as plain summary input without provider control tokens.
+ */
+export function serializeConversationForSummary(messages: Message[], dialect?: Dialect): string {
+	const conversation = serializeConversation(messages, dialect);
+	if (dialect !== "harmony") return conversation;
+	return conversation.replace(HARMONY_CONTROL_TOKEN_RE, "<\\|$1\\|>");
+}
+
 /**
- * Serialize LLM messages to text for summarization.
- * This prevents the model from treating it as a conversation to continue.
+ * Serialize LLM messages to transcript text.
  * Call convertToLlm() first to handle custom message types.
  */
 export function serializeConversation(messages: Message[], dialect?: Dialect): string {