How to deal with Raptor Mini?

[davrot]

Select Topic Area

Question

Copilot Feature Area

Copilot Agent Mode

Body

I am testing Raptor Mini in Agent Mode under VS Code Copilot. I noticed that the Raptor Mini agent acts rather autistic.

I talked to it normally, very polite and even insulting it. It tells me "I hear you" and it does exactly what I told it not to do.
If I can get it to work on the problem it produces progress. But...

1. I tell it how to build the sources and start the produced docker container. I even let it write the procedure down. No, instead of using a procedure that is working, it used his own approach that will never ever work.
2. It is working on tests that will never work. It ignores me if I try to explain that.
3. I press the button for stop. It ignores me. It just continues to start commands while I am mashing the stop button.
4. If I tell it to generate E2E playwright test which are in this project the only way that tests work then it builds some none E2E tests & runs them.
5. It builds some tests, runs them and tells me that everything is okay. I check it myself and the feature that was worked on was not even activated.
6. It tells me that problems have been solved, without changing any file at all. Shocker: Nothing was fixed.

And this is not just random behavior. I am letting it develop a plugin for the second time beginning from the same starting point but with different prompts. (Why? Because I need to explain my students how to use that thing in a productive way; Thanks Copilot Education! otherwise I couldn't do that research).

How do you get Raptor mini agent to do what it should do / is asked to do?

[ThanhDang-Vn]

Plan the steps first. Do not mark the task as done until you verify the specific file modification time on disk. I usually create markdown file and write exactly what it need to do step by step.

[davrot]

Yeah, tried that. I wrote a plan.md with Claude and a skillset.md with the commands. I told Raptor to read it but it ignores the info in that constantly, except the goals in the plan.md.

[davrot]

Or is the problem that I am extending an extended existing project (i.e. Overleaf)? Maybe this is confusing it.

[gayathra45]

Raptor Mini agent mode is experiencing known issues with following instructions, ignoring stop commands, and making unauthorized changes. Here's how to improve its behavior:​

Best Practices
Set thinking level to high: Raptor Mini is specifically designed for agent mode and performs better with high thinking settings​

Use custom instructions files: Create .copilot-instructions.md or AGENTS.md files at workspace root with clear, specific rules​

Be extremely specific in prompts: Clearly define goals, expected output, and boundaries​

• Use "Mandatory" or "Must Follow" sections in instructions​
• Avoid phrases like "stop doing..." - state what TO do instead​
• Set explicit constraints (e.g., "only plan, don't make changes")​

Provide documented procedures: Reference existing working procedures in your prompts, though agents may still ignore them​

Known Limitations

• Agents sometimes take unauthorized actions even when explicitly told not to​
• Stop button may not work immediately during active operations​
• Agent may claim work is complete without actual changes​

These issues reflect current model limitations - the behavior is being actively addressed but remains inconsistent

[VeVarunSharma]

Here's a tl;dr article I wrote about the mini-raptor, what it can do, limitations, and what it is. I hope it at least sheds light a bit more for you on how to best use it!

https://dev.to/vevarunsharma/so-what-is-github-copilots-raptor-miniand-why-should-devs-care-3n30