From 9d5d7996d4fa72a759e363d0e69ef07b8f65610b Mon Sep 17 00:00:00 2001
From: jfilipiuk <jakub.filipiuk@huawei.com>
Date: Tue, 16 Jun 2026 12:06:37 +0200
Subject: [PATCH] fix: render full thread history in chat (#8)

Long threads were truncated, hiding the original prompt
and early agent reasoning. Three layered causes, fixed together:

- Read messages from threads.get (persisted record) instead of
  threads.getState (graph checkpoint, which the backend windows for
  agent memory). Still call getState in parallel for run-status fields.
- Prefer the polled snapshot over stream.messages once the run has
  settled; stream length can be inflated by sub-agent noise from
  streamSubgraphs.
- Skip the subgraph-namespace filter for human messages; getMessagesMetadata
  can return stale subgraph tags for prompts from a prior run.
---
 src/app/components/ChatInterface.tsx | 20 +++++++++++------
 src/app/hooks/useChat.ts             | 33 ++++++++++++++++++++++------
 2 files changed, 39 insertions(+), 14 deletions(-)

diff --git a/src/app/components/ChatInterface.tsx b/src/app/components/ChatInterface.tsx
index 9142453..8feed45 100644
--- a/src/app/components/ChatInterface.tsx
+++ b/src/app/components/ChatInterface.tsx
@@ -686,6 +686,19 @@ export const ChatInterface = React.memo<ChatInterfaceProps>(
       // aren't in thread state anyway.)
       const seenAsyncUpdates = new Set<string>();
       const visibleMessages = messages.filter((message: Message) => {
+        // Humans are always user-typed (or our injected async-update pills) —
+        // never sub-agent noise. Run their checks first so a stale subgraph
+        // namespace on a previous-run human (left over in stream metadata)
+        // can't silently drop the original prompt.
+        if (message.type === "human") {
+          const key = asyncUpdateMessageKey(
+            extractStringFromMessageContent(message)
+          );
+          if (!key) return true;
+          if (seenAsyncUpdates.has(key)) return false;
+          seenAsyncUpdates.add(key);
+          return true;
+        }
         const meta = stream.getMessagesMetadata(message)?.streamMetadata;
         const ns = meta?.["langgraph_checkpoint_ns"];
         if (typeof ns === "string" && ns.includes("|")) return false;
@@ -696,13 +709,6 @@ export const ChatInterface = React.memo<ChatInterfaceProps>(
         // never persisted in `messages`. Drop it here; the stable summary is
         // surfaced from `_summarization_event` as a collapsible block instead.
         if (isSummarizationMessage(message)) return false;
-        if (message.type !== "human") return true;
-        const key = asyncUpdateMessageKey(
-          extractStringFromMessageContent(message)
-        );
-        if (!key) return true;
-        if (seenAsyncUpdates.has(key)) return false;
-        seenAsyncUpdates.add(key);
         return true;
       });
       const completedToolCallIds = new Set<string>();
diff --git a/src/app/hooks/useChat.ts b/src/app/hooks/useChat.ts
index 59f01ee..db05f91 100644
--- a/src/app/hooks/useChat.ts
+++ b/src/app/hooks/useChat.ts
@@ -231,13 +231,24 @@ export function useChat({
     const attempt = async () => {
       tries += 1;
       try {
-        const state = (await client.threads.getState(threadId)) as {
-          tasks?: Array<{ interrupts?: unknown[] }>;
-          next?: unknown[];
-          values?: { messages?: Message[] };
-        };
+        // `getState` returns the GRAPH CHECKPOINT state — which the backend
+        // windows/compacts for memory, so its `values.messages` is only the
+        // recent slice. `threads.get` returns the persisted THREAD RECORD with
+        // the full message history. We need both: state for run status
+        // (`next` / `tasks` / `interrupts`), record for the messages the UI
+        // displays. Done in parallel to keep the round trip tight.
+        const [state, threadRecord] = await Promise.all([
+          client.threads.getState(threadId) as Promise<{
+            tasks?: Array<{ interrupts?: unknown[] }>;
+            next?: unknown[];
+            values?: { messages?: Message[] };
+          }>,
+          client.threads.get(threadId) as Promise<{
+            values?: { messages?: Message[] };
+          }>,
+        ]);
         if (cancelled || recoveryRunRef.current !== recoveryRunId) return;
-        const msgs = state.values?.messages;
+        const msgs = threadRecord.values?.messages;
         const pending = latestTaskInterrupt(state.tasks);
         const stillPending = Array.isArray(state.next) && state.next.length > 0;
         const safePending = normalizePendingInterrupt(pending);
@@ -319,13 +330,21 @@ export function useChat({
   // the blank live version (the bug where the answer only appears after a manual
   // refresh). The equal-count/more-text rule is gated on `!isLoading` so a
   // mid-stream poll snapshot never flickers over the actively updating stream.
+  //
+  // Once the run has settled AND we have a snapshot, ALWAYS prefer the snapshot.
+  // `stream.messages` can carry subgraph noise (streamSubgraphs: true) plus
+  // stale per-message metadata from earlier runs, inflating its length above the
+  // persisted main-thread state. A pure `>` compare against that bloated count
+  // would keep us on the stream — which makes the downstream subgraph-namespace
+  // filter (ChatInterface.processedMessages) drop legitimate main-thread
+  // history that's only tagged subgraph in stale stream metadata.
   const messages = (() => {
     if (!fetchedMessages || fetchedThreadId !== threadId)
       return stream.messages;
     if (fetchedInterrupt) return fetchedMessages;
+    if (!stream.isLoading) return fetchedMessages;
     if (fetchedMessages.length > stream.messages.length) return fetchedMessages;
     if (
-      !stream.isLoading &&
       fetchedMessages.length === stream.messages.length &&
       totalTextLength(fetchedMessages) > totalTextLength(stream.messages)
     ) {