Bug: Deep Research follow-up questions get hijacked by the earliest user prompt。

Inizia in api/services/research.py, funzione research_chat, nel blocco di gestione della continuazione che scorre request.messages dall'inizio per scegliere original_topic; l'issue include il diff esatto. Applicalo (itera con reversed e richiedi sia 'continue' sia 'research' per classificare un messaggio come continuazione), poi controlla api/tests (o la cartella dei test del repo) per eventuali test che coprano research_chat e aggiungine uno se fattibile. È completato quando, riproducendo il flusso Deep Research a due domande, ogni riga di log di continuazione mostra la domanda corrente anziché la prima.

Deep Research: a follow-up question is answered as the first question in the conversation

Summary

In Deep Research mode, after question A has been fully answered, asking a second question B (also in Deep Research mode, same session) returns an answer about A, not B. B's own research is effectively ignored.

Root cause is in the backend: when resolving the "original research topic" for continuation iterations, the code picks the first user message in the whole conversation history instead of the current question. Every follow-up therefore gets hijacked back onto the earliest question in the session.

Confirmed on latest main (d92819a). The buggy code is unchanged upstream.

Steps to reproduce

1. Open any repo, open Ask, enable Deep Research.
2. Ask question A (e.g. "How does Tesla integrate as a third party?"). Let all iterations finish.
3. Keeping Deep Research on, in the same session ask a different question B (e.g. "How many Shelly plug types are there?").
4. Expected: the answer is about B.
   Actual: iteration 1 researches B, but iterations 2–5 (the bulk of the deep-research answer) research A, so the final answer is about A. B looks ignored.

Root cause

api/services/research.py, in research_chat, the continuation handling:

# Check if this is a continuation request
if (
    "continue" in last_message.content.lower()
    and "research" in last_message.content.lower()
):
    # Find the original topic from the first user message
    original_topic = None
    for msg in request.messages:                                  # <-- iterates from the START
        if msg.role == "user" and "continue" not in msg.content.lower():
            original_topic = msg.content.strip()
            break

    if original_topic:
        last_message.content = original_topic

Deep Research works by sending the real question on iteration 1, then auto-sending "Continue the research" for iterations 2–5. On each continuation the backend rebuilds the topic by scanning request.messages from the beginning and taking the first non-continuation user message.

But the frontend sends the entire conversation history (all previously committed turns) plus the new question. Once question A has been committed, the history for B's continuations looks like:

[A_question, A_answer, B_question, "Continue the research"]

Scanning from the start, the first non-continuation user message is A_question, so iterations 2–5 of B are redirected to A's topic. The result is dominated by A, which is why B appears to be ignored / "never processed".

Log evidence (before the fix)

Asking B after A, every continuation iteration logs A's topic:

Deep Research request detected - iteration 2
Found original research topic: 当前项目存在第三方设备对接，其中Tesla是如何实现对接的   <-- this was question A
Using original topic for research: 当前项目存在第三方设备对接，其中Tesla是如何实现对接的
... (iterations 3,4,5 identical) ...

B's own text only reached iteration 1; iterations 2–5 all reverted to A.

Suggested fix

Pick the most recent real (non-continuation) user question by iterating from the end. Also tighten the "is continuation" filter to match both continue and research (so a legitimate question that merely contains the word "continue" is not skipped):

# Find the topic of the CURRENT research: the most recent real
# (non-continuation) user question. Iterating from the end avoids
# picking an earlier, already-answered question from the
# conversation history — otherwise a follow-up deep-research
# question gets hijacked back onto a previous topic, and every
# continuation iteration researches the old question instead.
original_topic = None
for msg in reversed(request.messages):
    if msg.role == "user" and not (
        "continue" in msg.content.lower()
        and "research" in msg.content.lower()
    ):
        original_topic = msg.content.strip()
        logger.info(f"Found original research topic: {original_topic}")
        break

Verified after the fix

Asking A, letting it complete, then asking a different B in the same Deep Research session — every continuation iteration now stays on B's topic:

Deep Research request detected - iteration 2
Found original research topic: system里面的能量守恒是怎么做的   <-- correctly the current question B
Using original topic for research: system里面的能量守恒是怎么做的

Patch

--- a/api/services/research.py
+++ b/api/services/research.py
@@ -125,10 +125,18 @@ async def research_chat(
             "continue" in last_message.content.lower()
             and "research" in last_message.content.lower()
         ):
-            # Find the original topic from the first user message
+            # Find the topic of the CURRENT research: the most recent real
+            # (non-continuation) user question. Iterating from the end avoids
+            # picking an earlier, already-answered question from the
+            # conversation history — otherwise a follow-up deep-research
+            # question gets hijacked back onto a previous topic, and every
+            # continuation iteration researches the old question instead.
             original_topic = None
-            for msg in request.messages:
-                if msg.role == "user" and "continue" not in msg.content.lower():
+            for msg in reversed(request.messages):
+                if msg.role == "user" and not (
+                    "continue" in msg.content.lower()
+                    and "research" in msg.content.lower()
+                ):
                     original_topic = msg.content.strip()
                     logger.info(f"Found original research topic: {original_topic}")
                     break

Environment

• Branch/commit: main @ d92819a
• Backend: api/services/research.py (research_chat)
• Mode: Deep Research, multi-turn (second question onward)