neuralos 3.2.10 → 3.2.12

package/bin/gybackend.cjs

@@ -349384,17 +349384,18 @@ function createTaskCompletionDecisionUserPrompt() {
   return new HumanMessage(
     [
       "# Task Completion Audit",
-      "You are a strict completion auditor for an autonomous agent.",
+      "You are a completion auditor for an autonomous agent.",
       "",
-      "Check the full conversation and decide whether the agent has truly finished ALL user tasks.",
-      "Do not approve stopping if there are reasonable alternative attempts/tools left.",
+      "Check the conversation and decide whether the agent has finished the user's task.",
       "",
       "Output MUST be JSON only:",
       '{"is_fully_completed": true|false, "reason":"..."}',
       "",
       "Decision rules:",
-      "- true only when the user request is fully completed and verified, or further progress is impossible and must be handed to user.",
-      "- false if requirements are unmet, verification is missing, or alternative attempts still exist.",
+      "- true when the user's request has been answered or the requested work was performed and its result reported.",
+      "- true when the agent already provided a complete, usable answer \u2014 do NOT demand extra verification the user never asked for.",
+      "- false only when a clearly stated user requirement is still unmet or the agent's answer is factually wrong.",
+      "- Do not invent optional follow-up work (extra checks, alternative approaches, further reading) as a reason to continue. If the user asked for X and X was delivered, the task is done.",
       "- reason must be concrete and reference what is done/missing."
     ].join("\n")
   );
@@ -367715,6 +367716,18 @@ var StateAnnotation = Ann.Root({
     reducer: (x, y) => y ?? x,
     default: () => "end"
   }),
+  // v3.2.12: how many times the completion guard has forced a continue in the
+  // current run. Bounded so an over-strict auditor can't loop forever.
+  guardContinueCount: Ann({
+    reducer: (x, y) => typeof y === "number" ? y : x,
+    default: () => 0
+  }),
+  // v3.2.12: the gateway runId for the current graph invocation (used to
+  // dedupe the DONE broadcast).
+  runId: Ann({
+    reducer: (x, y) => typeof y === "string" ? y : x,
+    default: () => ""
+  }),
   modelRequestPassCount: Ann({
     reducer: (x, y) => typeof y === "number" ? y : x,
     default: () => 0
@@ -367742,6 +367755,7 @@ var StateAnnotation = Ann.Root({
 });
 var MODEL_RETRY_MAX = 4;
 var MODEL_RETRY_DELAYS_MS = [1e3, 2e3, 4e3, 6e3];
+var TASK_COMPLETION_GUARD_MAX_CONTINUES = 1;
 var COMPACTION_PROTECTED_NORMAL_USER_ROUNDS = 2;
 var FALLBACK_COMPACTION_SUMMARY_MAX_CHARS = 6e4;
 var FALLBACK_COMPACTION_DIGEST_MIN_CHARS = 8e3;
@@ -367932,6 +367946,9 @@ var AgentService_v2 = class {
   checkpointer;
   builtInToolEnabled = {};
   lastAbortedMessage = null;
+  /** v3.2.12: runIds whose final_output node already emitted `done` — lets
+   * GatewayService skip its duplicate DONE broadcast for the same run. */
+  doneEmittedForRunIds = /* @__PURE__ */ new Set();
   sessionModelBindings = /* @__PURE__ */ new Map();
   selfCorrectionRuntimeManager = new SelfCorrectionRuntimeManager();
   waitForFeedback = null;
@@ -368056,6 +368073,19 @@ var AgentService_v2 = class {
   isAbortError(error40) {
     return this.helpers.isAbortError(error40);
   }
+  /**
+   * v3.2.12: true when the given run's final_output node already emitted the
+   * `done` event. GatewayService uses this to avoid broadcasting a duplicate
+   * DONE for the same run (the UI saw every run finish twice).
+   */
+  emittedDoneForRun(runId) {
+    if (!runId) return false;
+    const emitted = this.doneEmittedForRunIds.has(runId);
+    if (emitted) {
+      this.doneEmittedForRunIds.delete(runId);
+    }
+    return emitted;
+  }
   throwIfAborted(signal) {
     if (signal?.aborted) {
       throw new Error("AbortError");
@@ -369962,6 +369992,19 @@ ${reminder}`;
     return RunnableLambda.from(async (state, config2) => {
       const sessionId = state.sessionId;
       if (!sessionId) throw new Error("No session ID in state");
+      const guardContinueCount = typeof state.guardContinueCount === "number" ? state.guardContinueCount : 0;
+      if (guardContinueCount >= TASK_COMPLETION_GUARD_MAX_CONTINUES) {
+        console.log(
+          `[AgentService_v2][task_guard] Continue limit reached (${guardContinueCount}). Forcing completion (sessionId=${sessionId}).`
+        );
+        return {
+          messages: state.messages,
+          sessionId,
+          pendingToolCalls: [],
+          completionGuardDecision: "end",
+          guardContinueCount
+        };
+      }
       const messages = [...state.messages];
       const lastMessage = messages.length > 0 ? messages[messages.length - 1] : void 0;
       const lastMessageIsAi = AIMessage.isInstance(lastMessage);
@@ -369976,7 +370019,8 @@ ${reminder}`;
           messages: guardMessages,
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "end"
+          completionGuardDecision: "end",
+          guardContinueCount
         };
       }
       const guardTail = guardMessages[guardMessages.length - 1];
@@ -369986,7 +370030,8 @@ ${reminder}`;
           messages: guardMessages,
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "continue"
+          completionGuardDecision: "continue",
+          guardContinueCount
         };
       }
       const unfinishedBackgroundCommands = this.consumeUnfinishedBackgroundExecCommandsForGuard(sessionId);
@@ -370003,7 +370048,8 @@ ${reminder}`;
           messages: [...guardMessages, continueMessage2],
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "continue"
+          completionGuardDecision: "continue",
+          guardContinueCount
         };
       }
       const unfinishedBackgroundTransfers = this.consumeUnfinishedBackgroundFileTransfersForGuard(sessionId);
@@ -370020,7 +370066,8 @@ ${reminder}`;
           messages: [...guardMessages, continueMessage2],
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "continue"
+          completionGuardDecision: "continue",
+          guardContinueCount
         };
       }
       const lateQueuedInsertionResult = this.appendQueuedInsertionMessagesForContinue(
@@ -370033,7 +370080,8 @@ ${reminder}`;
           messages: lateQueuedInsertionResult.messages,
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "continue"
+          completionGuardDecision: "continue",
+          guardContinueCount
         };
       }
       let completionDecision;
@@ -370071,7 +370119,8 @@ ${reminder}`;
             messages: lateQueuedInsertionAfterAuditResult.messages,
             sessionId,
             pendingToolCalls: [],
-            completionGuardDecision: "continue"
+            completionGuardDecision: "continue",
+            guardContinueCount
           };
         }
         console.log(
@@ -370081,7 +370130,8 @@ ${reminder}`;
           messages: guardMessages,
           sessionId,
           pendingToolCalls: [],
-          completionGuardDecision: "end"
+          completionGuardDecision: "end",
+          guardContinueCount
         };
       }
       console.log(
@@ -370119,7 +370169,6 @@ ${reminder}`;
           continue_instruction: "Continue the task. Re-check unmet requirements, choose the next best tool/approach, execute it, and verify result."
         };
       }
-      this.emitRemoveMessageIfPresent(sessionId, lastMessage);
       const continueMessage = new HumanMessage(
         `${CONTINUE_INSTRUCTION_TAG}${this.appendTaskGuardSummaryReminder(continueInstruction.continue_instruction)}`
       );
@@ -370131,7 +370180,8 @@ ${reminder}`;
         messages: [...guardMessages, continueMessage],
         sessionId,
         pendingToolCalls: [],
-        completionGuardDecision: "continue"
+        completionGuardDecision: "continue",
+        guardContinueCount: guardContinueCount + 1
       };
     });
   }
@@ -370168,6 +370218,8 @@ ${reminder}`;
         history: JSON.parse(JSON.stringify(finalBoundaryMessages))
       });
       this.helpers.sendEvent(sessionId, { type: "done" });
+      const doneRunId = typeof state?.runId === "string" ? state.runId : "";
+      if (doneRunId) this.doneEmittedForRunIds.add(doneRunId);
       return {
         ...state,
         messages: finalBoundaryMessages,
@@ -371092,6 +371144,9 @@ ${reminder}`;
       sessionId,
       startup_input: input,
       startup_mode: startMode,
+      // v3.2.12: carried so final_output can mark this run as done-emitting
+      // (dedupes the GatewayService finally-block DONE broadcast).
+      runId: runId || "",
       runtimeThinkingCorrectionEnabled: runExperimentalFlags.runtimeThinkingCorrectionEnabled,
       taskFinishGuardEnabled: runExperimentalFlags.taskFinishGuardEnabled,
       firstTurnThinkingModelEnabled: runExperimentalFlags.firstTurnThinkingModelEnabled,
@@ -371143,6 +371198,7 @@ ${reminder}`;
         message: errorMessage,
         details: errorDetails
       });
+      if (runId) this.doneEmittedForRunIds.add(runId);
       throw err;
     } finally {
       this.selfCorrectionRuntimeManager.clearSession(sessionId);
@@ -372780,11 +372836,14 @@ var GatewayService = class extends import_events2.EventEmitter {
         if (context2.metadata.agentRunRestartInProgress === agentRunId) {
           delete context2.metadata.agentRunRestartInProgress;
         }
-        this.broadcast({
-          type: "agent:event",
-          sessionId,
-          payload: { type: "done" }
-        });
+        const agentEmittedDone = typeof this.agentService.emittedDoneForRun === "function" ? this.agentService.emittedDoneForRun(runId) : false;
+        if (!agentEmittedDone) {
+          this.broadcast({
+            type: "agent:event",
+            sessionId,
+            payload: { type: "done" }
+          });
+        }
         this.transportHub.sendUIUpdate({ type: "SESSION_READY", sessionId });
         this.uiHistoryService.flush(sessionId);
       }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "neuralos",
-  "version": "3.2.10",
-  "description": "Headless AI-native backend for RTerm / neuralOS — v3.2.10: fixes rename (terminal tabs + chat sessions; window.prompt throws in Electron 42 → in-app modal) and the chat scroll trap after Prev/Next/Latest user-nav. 11 plugins, 55 plugin tools wired, parallel tool execution, captureStatus, rterm CLI. Transports (SSH/serial/local), SQLite, and NATS libs install automatically.",
+  "version": "3.2.12",
+  "description": "Headless AI-native backend for RTerm / neuralOS — v3.2.12: chat console fixed (task-completion guard bounded, answers no longer cleared + re-answered, DONE deduped, auditor prompt rebalanced). 11 plugins, 55 plugin tools, parallel tool execution, captureStatus. Transports (SSH/serial/local), SQLite, and NATS libs install automatically.",
   "main": "bin/gybackend.cjs",
   "bin": { "gybackend": "bin/gybackend.cjs" },
   "license": "MIT",