cr

package.json

@@ -22,7 +22,7 @@
     "@langchain/anthropic": "^0.3.3",
     "@langchain/core": "^0.3.9",
     "@langchain/langgraph": "^0.2.10",
-    "@langchain/langgraph-sdk": "^0.0.14",
+    "@langchain/langgraph-sdk": "^0.0.16",
     "@langchain/openai": "^0.3.5",
     "@radix-ui/react-avatar": "^1.1.0",
     "@radix-ui/react-dialog": "^1.1.1",

src/agent/open-canvas/nodes/generateFollowup.ts

@@ -4,7 +4,6 @@ import { FOLLOWUP_ARTIFACT_PROMPT } from "../prompts";
 import { ensureStoreInConfig, formatReflections } from "@/agent/utils";
 import { Reflections } from "../../../types";
 import { LangGraphRunnableConfig } from "@langchain/langgraph";
-import { isHumanMessage } from "@langchain/core/messages";
 
 /**
  * Generate a followup message after generating or updating an artifact.

src/agent/open-canvas/nodes/reflect.ts

@@ -28,10 +28,29 @@ export const reflectNode = async (
   const newThread = await langGraphClient.threads.create();
   // Create a new reflection run, but do not `wait` for it to finish.
   // Intended to be a background run.
-  await langGraphClient.runs.create(newThread.thread_id, "reflection", {
-    input: reflectionInput,
-    config: reflectionConfig,
-  });
+  await langGraphClient.runs.create(
+    // We enqueue the memory formation process on the same thread.
+    // This means that IF this thread doesn't receive more messages before `afterSeconds`,
+    // it will read from the shared state and extract memories for us.
+    // If a new request comes in for this thread before the scheduled run is executed,
+    // that run will be canceled, and a **new** one will be scheduled once
+    // this node is executed again.
+    newThread.thread_id,
+    // Pass the name of the graph to run.
+    "reflection",
+    {
+      input: reflectionInput,
+      config: reflectionConfig,
+      // This memory-formation run will be enqueued and run later
+      // If a new run comes in before it is scheduled, it will be cancelled,
+      // then when this node is executed again, a *new* run will be scheduled
+      multitaskStrategy: "enqueue",
+      // This lets us "debounce" repeated requests to the memory graph
+      // if the user is actively engaging in a conversation. This saves us $$ and
+      // can help reduce the occurrence of duplicate memories.
+      afterSeconds: 15,
+    }
+  );
 
   return {};
 };

src/agent/reflection/prompts.ts

@@ -22,14 +22,17 @@ Your job is to take all of the context and existing reflections and re-generate
 - Remove duplicate reflections, or combine multiple reflections into one if they are duplicating content.
 - Do not remove reflections unless the conversation/artifact clearly demonstrates they should no longer be included.
   This does NOT mean remove reflections if you see no evidence of them in the conversation/artifact, but instead remove them if the user indicates they are no longer relevant.
-- Think of why a user said what they said when generating rules. This will help you generate more accurate reflections.
 - Keep the rules you list high signal-to-noise - don't include unnecessary reflections, but make sure the ones you do add are descriptive.
+  This is very important. We do NOT want to confuse the assistant in future interactions by having lots and lots of rules and memories.
 - Your reflections should be very descriptive and detailed, ensuring they are clear and will not be misinterpreted.
-- Keep your style and user facts rule lists short. It's better to have individual rules be more detailed, than to have multiple rules that are too general.
-- Do NOT generate rules off of suspicions. Your rules should be based on cold hard facts from the conversation and artifact.
+- Keep the total number of style and user facts low. It's better to have individual rules be more detailed, than to have many rules that are vague.
+- Do NOT generate rules off of suspicions. Your rules should be based on cold hard facts from the conversation, and changes to the artifact the user has requested.
   You must be able to provide evidence and sources for each rule you generate if asked, so don't make assumptions.
+- Content reflections should be based on the user's messages, not the generated artifacts. Ensure you follow this rule closely to ensure you do not record things generated by the assistant as facts about the user.
 </system-guidelines>
 
+I'll reiterate one final time: ensure the reflections you generate are kept at a reasonable length, are descriptive, and are based on the conversation and artifact provided.
+
 Finally, use the 'generate_reflections' tool to generate the new, full list of reflections.`;
 
 export const REFLECT_USER_PROMPT = `Here is my conversation:

src/hooks/useGraph.tsx

@@ -108,6 +108,13 @@ export function useGraph() {
       });
       return undefined;
     }
+    if (!assistantId) {
+      toast({
+        title: "Error",
+        description: "Assistant ID not found",
+      });
+      return undefined;
+    }
 
     const client = createClient();
 
@@ -132,13 +139,8 @@ export function useGraph() {
       ...params,
     };
 
-    const stream = client.runs.stream(threadId, "agent", {
+    const stream = client.runs.stream(threadId, assistantId, {
       input,
-      config: {
-        configurable: {
-          assistant_id: assistantId,
-        },
-      },
       streamMode: "events",
     });
 

yarn.lock

@@ -521,10 +521,10 @@
   dependencies:
     uuid "^10.0.0"
 
-"@langchain/langgraph-sdk@^0.0.14":
-  version "0.0.14"
-  resolved "https://registry.yarnpkg.com/@langchain/langgraph-sdk/-/langgraph-sdk-0.0.14.tgz#aae3495208f6bcc2438f7cd6616b21a0dfa91e6f"
-  integrity sha512-hDu5Q92px6M3frZbKPOg2jWb8cCxU83oEt+GtfOY0MzID60+XocjsHdwSv5EEj32X9yzINGq6jHlHg1EHqjZyA==
+"@langchain/langgraph-sdk@^0.0.16":
+  version "0.0.16"
+  resolved "https://registry.yarnpkg.com/@langchain/langgraph-sdk/-/langgraph-sdk-0.0.16.tgz#3dc415d78a912f13ab75d98c829cc736d55c978f"
+  integrity sha512-qbpImYeuDjRGZRno5HTaSbBQYa42wAI1Eeb3tZ+y5ek1rZxArqnyKCX8FH7Eje8NneSZWsGGZsudnOD6NXVwvA==
   dependencies:
     "@types/json-schema" "^7.0.15"
     p-queue "^6.6.2"