Middleware audit for CK 1.62 / MAF 1.10: retire ReconcileSnapshot, inject dropped AG-UI context

- ReconcileSnapshotMiddleware removed: MAF now reuses the streamed
  message id in MESSAGES_SNAPSHOT (verified in _agent_run.py at
  python-1.10.0 and live event logs); keeping the surgery duplicated
  user messages on a2ui follow-up runs.
- beforeRequestMiddleware now injects RunAgentInput.context as a system
  message: MAF declares the field but never feeds it to the model,
  which starves the model of CopilotKit's A2UI catalog/guidelines.
  Remove when MAF consumes context upstream.
- Workaround chain env-gated: CK_MIDDLEWARES=off runs bare for audits,
  CK_DEBUG_EVENTS=1 enables the event logger.
- Kept: SuppressRenderToolText, StripModelArtifacts,
  DeduplicateToolCall, pairOrphanToolCalls (behavior-verified).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

frontend/server.ts

@@ -191,70 +191,6 @@ class StripModelArtifactsMiddleware extends (Middleware as any) {
   }
 }
 
-// ── ReconcileSnapshotMiddleware ──────────────────────────────────────────────
-// MAF's `_build_messages_snapshot` (agent_framework_ag_ui/_agent_run.py:686)
-// mints a fresh UUID for the post-tool-call assistant text instead of reusing
-// the streamed TEXT_MESSAGE_START id. The resulting MESSAGES_SNAPSHOT then
-// contains TWO assistant entries: the streamed id (holding the toolCalls) and
-// a brand-new id (holding the duplicated text). ag-ui's snapshot merge replaces
-// by id then APPENDS unknown ids, so the browser ends up with two assistant
-// bubbles for the same answer. Dropping the snapshot entirely fixes the dupe
-// but breaks render_a2ui card persistence (cards rely on the snapshot to keep
-// the assistant-with-toolCalls message in state past the run). The right fix
-// is to drop just the orphan text-only assistant message that has no streamed
-// counterpart. Remove once `_agent_run.py:686` reuses `flow.message_id`.
-
-// eslint-disable-next-line @typescript-eslint/no-explicit-any
-class ReconcileSnapshotMiddleware extends (Middleware as any) {
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  run(input: any, next: any): Observable<any> {
-    const streamedTextIds = new Set<string>();
-
-    // eslint-disable-next-line @typescript-eslint/no-explicit-any
-    return new Observable<any>((observer) => {
-      // eslint-disable-next-line @typescript-eslint/no-explicit-any
-      const sub = (this as any).runNextWithState(input, next).subscribe({
-        // eslint-disable-next-line @typescript-eslint/no-explicit-any
-        next: (eventWithState: any) => {
-          const event = eventWithState.event;
-          const type: string = event?.type ?? "";
-
-          if (type === EventType.TEXT_MESSAGE_START && event?.messageId) {
-            streamedTextIds.add(String(event.messageId));
-          }
-
-          if (type === EventType.MESSAGES_SNAPSHOT) {
-            // eslint-disable-next-line @typescript-eslint/no-explicit-any
-            const msgs: any[] = Array.isArray(event?.messages) ? event.messages : [];
-            const filtered = msgs.filter((m) => {
-              if (m?.role !== "assistant") return true;
-              const id = String(m?.id ?? "");
-              const hasText = typeof m?.content === "string" && m.content.length > 0;
-              const hasToolCalls =
-                (Array.isArray(m?.toolCalls) && m.toolCalls.length > 0) ||
-                (Array.isArray(m?.tool_calls) && m.tool_calls.length > 0);
-              if (hasText && !hasToolCalls && !streamedTextIds.has(id)) {
-                return false; // drop orphan text-only assistant duplicate
-              }
-              return true;
-            });
-            if (filtered.length !== msgs.length) {
-              observer.next({ ...event, messages: filtered });
-              return;
-            }
-          }
-
-          observer.next(event);
-        },
-        // eslint-disable-next-line @typescript-eslint/no-explicit-any
-        error: (err: any) => observer.error(err),
-        complete: () => observer.complete(),
-      });
-      return () => sub.unsubscribe();
-    });
-  }
-}
-
 // ── SuppressRenderToolTextMiddleware ──────────────────────────────────────────
 // When the LLM emits text content in the same response as a render tool call
 // (render_a2ui or render_spending_summary), the text appears as a duplicate
@@ -356,16 +292,22 @@ app.all("/api/copilotkit/*", async (c) => {
     : {};
 
   const agent = new HttpAgent({ url: AGENT_URL, headers: agentHeaders });
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+  // CK_MIDDLEWARES=off runs the stack bare — used to audit which of these
-  agent.use(new SuppressRenderToolTextMiddleware() as any);
+  // workarounds are still needed after CopilotKit/MAF upgrades.
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+  /* eslint-disable @typescript-eslint/no-explicit-any */
-  agent.use(new StripModelArtifactsMiddleware() as any);
+  if (process.env.CK_MIDDLEWARES !== "off") {
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    agent.use(new SuppressRenderToolTextMiddleware() as any);
-  agent.use(new ReconcileSnapshotMiddleware() as any);
+    agent.use(new StripModelArtifactsMiddleware() as any);
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    // ReconcileSnapshotMiddleware retired 2026-07-04: MAF 1.9+ reuses the
-  agent.use(new DeduplicateToolCallMiddleware() as any);
+    // streamed message id in MESSAGES_SNAPSHOT (verified in _agent_run.py at
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    // python-1.10.0 and in live event logs), and keeping the surgery caused
-  agent.use(new DebugLogMiddleware() as any);
+    // duplicated user messages on a2ui follow-up runs.
+    agent.use(new DeduplicateToolCallMiddleware() as any);
+  }
+  if (process.env.CK_DEBUG_EVENTS === "1" || process.env.CK_MIDDLEWARES === "off") {
+    agent.use(new DebugLogMiddleware() as any);
+  }
+  /* eslint-enable @typescript-eslint/no-explicit-any */
 
   const runtime = new CopilotRuntime({
     agents: { wealthysmart: agent },
@@ -375,19 +317,44 @@ app.all("/api/copilotkit/*", async (c) => {
       const ct = outbound.headers.get("content-type") ?? "";
       if (!ct.includes("application/json")) return;
       try {
-        const body = (await outbound.clone().json()) as { messages?: AGUIMessage[] };
+        const body = (await outbound.clone().json()) as {
+          messages?: AGUIMessage[];
+          context?: { description?: string; value?: string }[];
+        };
         if (!Array.isArray(body.messages)) return;
-        const paired = pairOrphanToolCalls(body.messages);
+        let messages = pairOrphanToolCalls(body.messages);
-        if (paired.length === body.messages.length) return;
+
+        // MAF's AG-UI endpoint declares `context` on its request model but
+        // never feeds it to the model (agent_framework_ag_ui, checked at
+        // python-1.10.0). CopilotKit 1.60+ ships the A2UI catalog schema and
+        // guidelines in that context — without them the model emits ops the
+        // A2UI validator rejects and the surface never paints. Inject the
+        // context as a system message ourselves. Remove once MAF consumes
+        // RunAgentInput.context.
+        if (Array.isArray(body.context) && body.context.length > 0) {
+          const contextText = body.context
+            .map((c) => `## ${c.description ?? "Context"}\n${c.value ?? ""}`)
+            .join("\n\n");
+          messages = [
+            {
+              id: `ctx-${Date.now()}`,
+              role: "system",
+              content: `Additional run context:\n\n${contextText}`,
+            } as AGUIMessage,
+            ...messages,
+          ];
+        }
+
+        if (messages === body.messages) return;
         return new Request(outbound.url, {
           method: outbound.method,
           headers: outbound.headers,
-          body: JSON.stringify({ ...body, messages: paired }),
+          body: JSON.stringify({ ...body, messages }),
         });
       } catch (err) {
         // Skipping the repair is survivable; doing it silently is not —
         // orphan tool calls then fail downstream with no trace of why.
-        console.error("[copilotkit] orphan-tool-call repair skipped:", err);
+        console.error("[copilotkit] outbound request repair skipped:", err);
         return;
       }
     },