Win/Pwsh Roo can't see output from follow-up commands

[kfsone]

Which version of the app are you using?

3.7.12

Which API Provider are you using?

Anthropic

Which Model are you using?

Claude 3.7 Sonnet

What happened?

I'm uncertain whether the issue is that I'm using a custom prompt or just an issue seeing output from follow-on commands, but unless I routinely delete the Roo terminal before allowing the execution of a second command, Roo only ever sees output from a first command, e.g.

cd cmd/tool && go run .
... sees output ...

cd cmd/tool && go run .
... cannot see output

Steps to reproduce

1. Windows 11,
2. Powershell core as default terminal/shell,
3. Use a Prompt() function in the $profile script, (Mine is multi-line)
4. Go or Python project,
5. Have Roo run a command, e.g. "cd subdir && dir"
6. Confirm roo sees output
7. Have Roo run a second command
   -> Roo cannot see output

Relevant API REQUEST output

Additional context

No response

[KJ7LNW]

try this and tell me what happens:

1. close all VSCE terminals (or open a completely different project directory in another VSCE instance)
2. tell the model to "run echo test"
3. change directories to somewhere else in the terminal window that it opened, eg: cd \
4. tell it to run echo test again
5. does it fail ?

[kfsone]

Sure:

(extra detail for my own checklist, not pedantry)

• closed all vs windows and terminal windows,
• opened an old rust project folder (never used with vs before),
• check config: "Code", w/default profile (Anthropic, claude-3-7-sonnet-20250219)
• first command:
  
• second command, after I ran "cd src"
  

[KJ7LNW]

@hannesrudolph : I know exactly where that line of code is, I was considering deleting it yesterday as part of #1365, but I think I need to check with you and developers who have been part of this for longer than I:

Is it really that important that a terminal changes directories to keep it always running in the project working directory? I mean, we are all developers here, and if somebody changes directories and something fails, it's their own fault...

What do you think?

https://github.com/RooVetGit/Roo-Code/blob/70a88aee4d7d31a1521c976e691aa5a89130e176/src/integrations/terminal/TerminalManager.ts#L336

and this is where it has been moved in #1365:

https://github.com/RooVetGit/Roo-Code/blob/f9cd2a96fd04903a2764e9fd7126996d398e3a52/src/integrations/terminal/TerminalRegistry.ts#L245

[hannesrudolph]

@mrubens is the pro on this one

[kfsone]

A note: Working with Go and submodules, claude-via-Roo [edit :)] very often likes to issue commands like:

cd cmd/thatthing && go test .

rather than

go test ./cmd/thatthing

[KJ7LNW]

cd cmd/thatthing && go test .

yes and I wish it would not. In fact, I have instructions in my custom code profile that tells it not to do that (yet it still does sometimes):

- NEVER change directory (cd) when executing commands
 - NEVER do this: cd some/path && some-command some/file
- DO NOT cd dir/ && run_program

[kfsone]

LLMs are bad at negatives like that if there is no strong correlation with cases in their training data, in-fact the model typically ends up applying attention to the very thing you were trying to tell it not to do.

(demonstrated by asking ChatGPT: implement std::move but do not use std::remove_reference)

[kfsone]

if there is no strong correlation with cases in their training data,

aka if that's not the way people wrote their questions on Stack Overflow when OpenAI trawled it :)

[KJ7LNW]

I will keep that in mind. so perhaps a better prompt would be something like "only run programs from the project's directory" would be better ...