Bug: Sub-agents created outside qwen-code /agent create are not detected on Windows

[Aayushyaash]

What happened?

On Windows, sub-agent markdown files created or edited with external editors (e.g., Notepad) does not appear in the agent list and is silently ignored by qwen-code.

Steps to reproduce:

1. On Windows, create/edit a sub-agent markdown file using Notepad / VS Code (same format as a working agent).
2. Save it in the local agents directory.
3. Start or reload qwen-code.
4. Run the command that lists agents.

Result: the sub-agent is not detected.

What did you expect to happen?

Sub-agents created or edited outside qwen-code (for example, with Notepad or other editors) should be detected and usable by qwen-code just like agents created with /agent create.

Client information

Client Information

Run qwen to enter the interactive CLI, then run the /about command.

Status   
                                                      
Qwen Code      0.10.4 (509a88a4)                      
Runtime        Node.js v20.20.0 / npm 11.10.0          
OS             win32 x64 (10.0.26100)                 
                                                      
Auth           qwen-oauth                             
Model          coder-model                            
Session ID     1e483c60-b9c2-4348-aaf5-935f23e6306d   
Sandbox        no sandbox                             
Proxy          no proxy                               
Memory Usage   165.0 MB                               

Login information

I log in via the qwen-oauth. No API key or special config is used.

Anything else we need to know?

I’d like to work on this issue.

[rizalord]

I had the same issue on Windows. Turns out the problem is CRLF line endings — Windows editors save files with \r\n but the parser only expects \n, so the file silently fails to load.

Workaround: change the line ending to LF before saving (in VS Code, click CRLF at the bottom right → select LF). After that the agent is detected immediately.

[Aayushyaash]

Yeah, I noticed this while going through the codebase. The sub-agent parser only accepts \n line endings, whereas the skills parser doesn’t have this issue—SkillManager normalizes CRLF to LF before parsing skills. Since, I hadn’t come across CRLF vs. LF line endings before, so I had to dig into the codebase to understand that this was even a factor.

I’m thinking of creating a unified normalizer for both, as long as it doesn’t break anything.