Update DeepResearcherAgent report handling and citation verification

frontends/ui/src/adapters/api/websocket-client.ts

@@ -56,8 +56,6 @@ export interface NATWebSocketClientOptions {
   reconnectAttempts?: number
   /** Delay between reconnection attempts in ms (default: 1000) */
   reconnectDelay?: number
-  /** Auth token for backend authentication */
-  authToken?: string
   /** Override WebSocket URL (uses same-origin by default, proxied through UI server) */
   websocketUrl?: string
 }
@@ -105,13 +103,7 @@ export class NATWebSocketClient {
     this.options.callbacks.onConnectionChange?.('connecting')
 
     try {
-      const baseWsUrl = this.options.websocketUrl || (await getWebSocketUrl())
-      let wsUrl = baseWsUrl
-      if (this.options.authToken) {
-        const separator = wsUrl.includes('?') ? '&' : '?'
-        wsUrl = `${wsUrl}${separator}token=${encodeURIComponent(this.options.authToken)}`
-      }
-
+      const wsUrl = this.options.websocketUrl || (await getWebSocketUrl())
       this.ws = new WebSocket(wsUrl)
       this.setupEventHandlers()
     } catch {
@@ -199,13 +191,6 @@ export class NATWebSocketClient {
     return this.ws?.readyState === WebSocket.OPEN
   }
 
-  /**
-   * Update the auth token (e.g., after refresh)
-   */
-  updateAuthToken = (authToken: string): void => {
-    this.options.authToken = authToken
-  }
-
   /**
    * Update conversation ID (e.g., when switching conversations)
    */

frontends/ui/src/features/chat/hooks/use-websocket-chat.spec.ts

@@ -151,7 +151,6 @@ const mockWsClient = {
   sendMessage: vi.fn(),
   sendInteractionResponse: vi.fn(),
   isConnected: vi.fn(() => false),
-  updateAuthToken: vi.fn(),
   updateConversationId: vi.fn(),
 }
 

frontends/ui/src/features/chat/hooks/use-websocket-chat.ts

@@ -34,7 +34,6 @@ import { checkBackendHealthCached, invalidateHealthCache } from '@/shared/hooks/
 import { useChatStore } from '../store'
 import { useConnectionRecovery } from './use-connection-recovery'
 import { useLayoutStore } from '@/features/layout/store'
-import { WEB_SEARCH_SOURCE_ID } from '@/features/layout/data-sources'
 import { useDocumentsStore } from '@/features/documents/store'
 import { useAuth } from '@/adapters/auth'
 import type {
@@ -153,9 +152,7 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
   // Ref to track the current status for detecting status changes
   const currentStatusRef = useRef<StatusType | null>(null)
 
-  // Auth hook for getting user and token
-  // Note: idToken is used for backend auth, not accessToken
-  const { user, idToken } = useAuth()
+  const { user } = useAuth()
 
   // Chat store
   const {
@@ -562,7 +559,6 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
       wsClientRef.current = createNATWebSocketClient({
         conversationId: currentConversation.id,
         callbacks: createCallbacks(),
-        authToken: idToken,
       })
       wsClientRef.current.connect()
     } else {
@@ -577,16 +573,7 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
         wsClientRef.current = null
       }
     }
-  }, [currentConversation?.id, autoConnect, idToken, createCallbacks])
-
-  /**
-   * Update auth token when it changes
-   */
-  useEffect(() => {
-    if (wsClientRef.current && idToken) {
-      wsClientRef.current.updateAuthToken(idToken)
-    }
-  }, [idToken])
+  }, [currentConversation?.id, autoConnect, createCallbacks])
 
   /**
    * Send a message via WebSocket
@@ -597,14 +584,7 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
 
       // Collect metadata about data sources and files before adding user message
       const layoutState = useLayoutStore.getState()
-      let enabledDataSources = layoutState.enabledDataSourceIds
-
-      // Filter out authenticated sources if user doesn't have a valid idToken
-      if (!idToken) {
-        enabledDataSources = enabledDataSources.filter(
-          (sourceId) => sourceId === WEB_SEARCH_SOURCE_ID
-        )
-      }
+      const enabledDataSources = layoutState.enabledDataSourceIds
 
       // Get session files
       const sessionId = useChatStore.getState().currentConversation?.id
@@ -683,7 +663,6 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
         wsClientRef.current = createNATWebSocketClient({
           conversationId,
           callbacks,
-          authToken: idToken,
         })
         wsClientRef.current.connect()
       } else {
@@ -702,7 +681,6 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
       setCurrentStatus,
       setStreaming,
       setLoading,
-      idToken,
       createCallbacks,
     ]
   )
@@ -771,11 +749,10 @@ export const useWebSocketChat = (options: UseWebSocketChatOptions = {}): UseWebS
       wsClientRef.current = createNATWebSocketClient({
         conversationId: currentConversation.id,
         callbacks: createCallbacks(),
-        authToken: idToken,
       })
       wsClientRef.current.connect()
     }
-  }, [currentConversation, idToken, createCallbacks])
+  }, [currentConversation, createCallbacks])
 
   // Activate recovery polling when connection error cards are visible
   useConnectionRecovery(connect)

frontends/ui/src/lib/pdf/ReactPdfDocument.tsx

@@ -19,6 +19,10 @@ Font.register({
   fonts: [{ src: 'Helvetica' }, { src: 'Helvetica-Bold', fontWeight: 'bold' }],
 })
 
+// Disable hyphenation — react-pdf's default English hyphenation inserts
+// hyphens into long words (including URLs), making them unusable when copied.
+Font.registerHyphenationCallback((word) => [word])
+
 const styles = StyleSheet.create({
   page: {
     padding: 56.69,

src/aiq_agent/agents/deep_researcher/agent.py

@@ -5,6 +5,7 @@
 from __future__ import annotations
 
 import logging
+import re
 from collections.abc import Sequence
 from datetime import datetime
 from pathlib import Path
@@ -14,6 +15,7 @@
 from deepagents.backends import CompositeBackend
 from deepagents.backends import StateBackend
 from langchain.agents.middleware import ModelRetryMiddleware
+from langchain_core.messages import AIMessage
 from langchain_core.messages import HumanMessage
 from langchain_core.tools import BaseTool
 from langchain_core.tools import tool
@@ -36,6 +38,11 @@
 
 logger = logging.getLogger(__name__)
 
+# Minimum character count for a report to be considered substantive.
+# Used by both _extract_report_content (to decide if write_file fallback is needed)
+# and _is_report_complete (to reject too-short reports).
+_MIN_REPORT_LENGTH = 1500
+
 # Path to this agent's directory (for loading prompts)
 AGENT_DIR = Path(__file__).parent
 
@@ -255,6 +262,29 @@ def backend(runtime):
         )
         return agent.with_config({"recursion_limit": 1000})
 
+    @staticmethod
+    def _extract_report_content(messages: list) -> str:
+        """Extract report content from the last message, falling back to write_file tool calls if text is too short."""
+        if not messages:
+            return ""
+        last_msg = messages[-1]
+        raw = last_msg.content or ""
+        if isinstance(raw, list):
+            content = " ".join(p.get("text", "") for p in raw if isinstance(p, dict) and p.get("type") == "text")
+        else:
+            content = raw if isinstance(raw, str) else str(raw)
+        if len(content) >= _MIN_REPORT_LENGTH:
+            return content
+        # If the last message is an AIMessage with a write_file tool call,
+        # the LLM may have written the report via tool instead of text output.
+        if isinstance(last_msg, AIMessage) and getattr(last_msg, "tool_calls", None):
+            for tc in last_msg.tool_calls:
+                if tc.get("name") == "write_file":
+                    file_content = tc.get("args", {}).get("content", "")
+                    if isinstance(file_content, str) and len(file_content) > len(content):
+                        content = file_content
+        return content
+
     def _is_report_complete(self, result: dict | Any) -> tuple[bool, str]:
         """
         Check if the agent produced a complete report using tool calls or heuristics.
@@ -266,10 +296,9 @@ def _is_report_complete(self, result: dict | Any) -> tuple[bool, str]:
         if not messages:
             return False, "no_messages"
 
-        last_msg = messages[-1]
-        content = last_msg.content or ""
+        content = self._extract_report_content(messages)
 
-        if len(content) < 1500:
+        if len(content) < _MIN_REPORT_LENGTH:
             return False, f"too_short ({len(content)} chars)"
 
         if content.count("## ") < 2:
@@ -299,7 +328,7 @@ def _is_report_complete(self, result: dict | Any) -> tuple[bool, str]:
                     url_match = _URL_IN_LINE_RE.search(ref_text)
                     if url_match:
                         url = url_match.group(0).rstrip(".,;)")
-                        if registry.has_url(url):
+                        if registry.resolve_url(url):
                             has_any_valid = True
                             break
                         continue
@@ -391,7 +420,12 @@ async def run(self, state: DeepResearchAgentState) -> DeepResearchAgentState:
                     if source_list:
                         feedback_msg += "\n\n" + source_list
 
-                feedback_msg += " Please fix this immediately and call 'submit_final_report' when done."
+                feedback_msg += (
+                    " IMPORTANT: Do NOT restart the research from scratch."
+                    " First check if /report.md already exists using read_file."
+                    " If it does, use that content as your report — just fix the specific issue above"
+                    " and return the corrected report in your final message."
+                )
 
                 if isinstance(result, dict):
                     next_state = {**result}
@@ -400,27 +434,52 @@ async def run(self, state: DeepResearchAgentState) -> DeepResearchAgentState:
                     next_state = result.model_dump() if hasattr(result, "model_dump") else dict(result)
                     messages = getattr(result, "messages", next_state.get("messages", []))
                 next_state["messages"] = list(messages) + [HumanMessage(content=feedback_msg)]
-                result = await agent.ainvoke(
-                    next_state,
-                    config={"callbacks": self.callbacks} if self.callbacks else None,
-                )
+
+                try:
+                    result = await agent.ainvoke(
+                        next_state,
+                        config={"callbacks": self.callbacks} if self.callbacks else None,
+                    )
+                    last_error = None
+                except Exception as ex:
+                    logger.error("Deep Research feedback retry %d failed: %s", attempt + 1, ex, exc_info=True)
+                    last_error = ex
+                    if "recursion" in str(ex).lower() or "reuse already awaited" in str(ex):
+                        raise ex
+                    # Non-fatal: ainvoke raised before producing a result, so
+                    # `result` still holds the previous iteration's value.
+                    # The next loop iteration will rebuild next_state from it.
+                    continue
+
+                # Evaluate the feedback-retry result before the next iteration
+                is_complete, reason = self._is_report_complete(result)
+                if is_complete:
+                    logger.info(f"Report completed after feedback retry. Reason: {reason}")
+                    break
+
+                # Update state so next iteration builds on progress, not the original state
+                state = result
 
             if result is None and last_error is not None:
                 raise last_error
 
             final_message = "Research failed to produce a report."
             if result and result.get("messages"):
-                final_content = result["messages"][-1].content
-                final_message = final_content if isinstance(final_content, str) else str(final_content)
+                final_message = self._extract_report_content(result["messages"])
 
             # Post-process: verify citations against source registry
             if self.source_registry_middleware._get_registry().all_sources():
                 verification = verify_citations(final_message, self.source_registry_middleware._get_registry())
                 if verification.removed_citations:
+                    removed_details = []
+                    for c in verification.removed_citations:
+                        url_match = re.search(r"https?://\S+", c.get("line", ""))
+                        url_str = url_match.group(0).rstrip(".,;)") if url_match else "(no url)"
+                        removed_details.append(f"[{c['number']}] {c['reason']}: {url_str}")
                     logger.info(
-                        "Citation verification removed %d invalid citations: %s",
+                        "Citation verification removed %d invalid citation(s):\n  %s",
                         len(verification.removed_citations),
-                        [c["reason"] for c in verification.removed_citations],
+                        "\n  ".join(removed_details),
                     )
                 final_message = verification.verified_report
             else: