Bug for the last step among the five steps!

[xiaohai-AI]

i find one of big trouble is the last step among the five steps! chat with report agent or chat with any agent , if you start the conversation, seems only chat for one round! if you keep chat for more rounds, even you ask the different question! the answer is just repeat the same answer form the first round chat ! thats very bad ! hope the team can improve that !

[dosubot]

Hi @xiaohai-AI! I'm Dosu and I’m helping the MiroFish team.

Hi @xiaohai-AI — this issue has actually been reported several times before. The most important thing to check first: are you using the online demo or a local deployment?

The online demo is a static page that is not connected to any LLM. It uses pre-prepared template responses, so no matter what you ask, the answer will always be the same. This was confirmed by the maintainer. If this is your case, you'd need to deploy MiroFish locally with a real LLM API configured (e.g., deepseek-chat, gemini-2.5-pro, qwen-plus).

If you are running a local deployment, the issue could be related to how conversation history is managed. MiroFish uses a stateless backend — the frontend must send the complete chat_history with every request to /api/report/chat [1]. A new ReportAgent instance is created per request, and only the last 10 messages are retained [2]. If the frontend fails to properly accumulate and send the history, each message gets treated as a fresh conversation, which would explain the repeated answers.

Some additional factors that can cause repetitive responses on local deployments:

• Context length overflow — the LLM's context window gets exhausted, causing truncated/repeated output [3]
• API timeouts — 300-second response limits can cause report truncation
• LLM hallucination — previous bugs where agents fabricated repeated data have been fixed in recent PRs [4] [5], so make sure you're on the latest version

Could you clarify whether you're using the online demo or a local deployment?

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[rqd6f4g6zn-bit]

A defensive fix for this is up in #626 — in the backend chat() function, the chat_history is now filtered to drop any history item that matches the current user message exactly, which prevents the LLM from seeing a duplicate trailing user message and echoing back the previous answer. PR: #626