# Leo AI destructive error handling
**Summary**
Leo fails to handle errors gracefully, destroying partial answers and restarting work. It mis-reports conversation length as the issue.
**Error message**
Conversation length exceeded. Please shorten your request or start a new conversation.
**Observed behavior**
- During generation, the rendered answer is removed and replaced with “thinking” status. This repeats multiple times within a single turn.
- The system re-performs web searches already completed in that turn, repeating all prior work.
- Pressing Stop mid-render does not predictably preserve the partial answer.
- Page reload reproduces the same error message without recovery.
- No feedback entry point exists in the Leo UI
**Evidence this is not a context window limit**
- A subsequent turn requesting “output your last answer” succeeds and returns content. Context grows monotonically per turn, so a longer context succeeding after a shorter one failed contradicts the stated cause.
- The failure occurs mid-turn during generation, not at submission time when context size would be evaluated.
- The system re-executes tool calls (web searches) rather than truncating history, indicating an error-handling path rather than token exhaustion.
**Reproduction steps**
1. Go to brave://leo-ai/
2. Submit a research-heavy prompt requiring multiple web lookups (e.g., provide a professional consultant homepage URL and ask which services are performed).
3. Observe the answer begin rendering normally.
4. Watch as rendered text is cleared and replaced with “thinking”; searches re-execute. This may happen 1-4 times.
5. On completion or timeout, the error dialog appears and no answer is retained.
6. Reload the page: same error persists.
7. Send a follow-up turn asking for the previous answer: a response is returned, confirming the context was never actually over-limit.
**Expected behavior**
On error, do not discard rendered output or silently re-run completed searches. Retain and display the partial answer. May include a notice such as: “An error occurred. The answer may be incomplete.”
**Environment**
Brave Browser desktop, current stable. Leo AI chat interface.
**Additional steps**
1. New turn: “Last turn, you rendered an answer multiple times and deleted it each time. After deletion, it seemed you re-performed steps, such as search. Please output the response.”
2. A complete, thorough answer was output, but the turn was still running. Thinking reported: “Executing code in the sandbox to help with your query”. (Conversation did not include any request for code generation; it was non-technical)
3. Click Stop
4. Answer remained displayed