Nope, your presets are fine. Its how you are calling qwen. If you interact with it directly and let it muck about, it's just going to do too much to be useful.

So you need something else that interacts with qwen for you, one that ensures you get a clean slate for each task. One that reads the tasks you created and only passes in what's relevant to make the code change. I wrote this toml and the python script picks it up and only passes these tasks one at a time to qwen code CLI. each task starts with near 0 context so you arent dealing with compaction.

Then you have your unit tests so that it can fix its mistakes on code change.

The important part of this is you need a way to manage context because local LLM can't be trusted to manage it on its own.

tasks.toml

[[tasks]]
name = "fix-venue-priority"
prompt = "Fix _user_role() in offers.py so venue role takes priority over band role."
files = ["backend/app/routes/offers.py"]


[[tasks]]
name = "add-genre-filter"
prompt = "Add a 'genre' query parameter to the search endpoint that filters
bands by their genre field."
files = ["backend/app/routes/search.py"]

# execute every test in the scripts folder after every single task.
# python feeds stack trace back into LLM with minimal context
[[tests]]
command = "pytest"
args = ["scripts/", "-x", "--tb=native", "-q"]
timeout = 30
python = "./venv/bin/python"