Skip to content

Conversation

cwlbraa
Copy link
Contributor

@cwlbraa cwlbraa commented Jul 15, 2025

based on #226, this PR reworks the prompt engineering to be less reliant on a rule file.

it still sometimes ignores the "don't guess environment IDs" instruction (1/5 times in my testing, admittedly that's a low N).

it does behave better when it needs to use git (common for manually testing container-use)


const (
repositoryToolSuffix = "You MUST tell the user how to view environment changes using \"container-use log <environment_id>\", \"container-use diff <environment_id>\", AND \"container-use checkout <env_id>\". Failure to do so will make your work completely inaccessible."
environmentToolSuffix = "You must call `environment_create` or `environment_open` to obtain a valid environment_id value. LLM-generated environment IDs WILL cause task failure."
Copy link
Contributor Author

@cwlbraa cwlbraa Jul 16, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

right now i'm using both a tool description suffix AND an argument description, and models still ignore both.

maybe this is a place to apply dynamic tools 🤔... while there's no strong guarantee that 1 server process = 1 session, we could maybe still hide env tools prior to open/create and list env ids in tool descriptions...

that said, i've been told that when you have a hammer, everything looks like a nail. (also i still think maybe only vscode and goose have that hammer)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants