Problem / Value
When the Inline Terminal is enabled, Zoo tells the model that its commands will run in the shell configured in VS Code, but the commands actually run in a different shell - on Windows, the plain Command Prompt. The model then writes commands in a dialect the terminal cannot understand, and has to guess and retry before anything works. Making the reported shell match the real one would remove failed commands, wasted turns, and needless confusion for both the model and the user.
Context
Affects Windows users whose VS Code default terminal is something other than Command Prompt (PowerShell, PowerShell 7, Git Bash, etc.) while the recommended "Use Inline Terminal (recommended)" setting is left on - which is the out-of-the-box configuration.
What happens today: the environment summary handed to the model names the VS Code shell (PowerShell in my case), regardless of the fact that, with the Inline Terminal active, commands the model runs are executed by Command Prompt. The model only discovers the mismatch by failing.
What should happen: the model should always be told the kind of shell it will actually get, so its very first command is written for the right shell.
Reproduction
- On Windows, set your VS Code default terminal to PowerShell
- In Zoo settings, leave "Use Inline Terminal (recommended)" enabled
- Ask Zoo to do anything that runs a command, and observe the shell Zoo reports to the model versus the shell the command actually runs in
- Expected result: Zoo tells the model the same shell its commands really run in
- Actual result: Zoo tells the model the shell is PowerShell, while the commands are executed by Command Prompt, so shell-specific commands fail
Constraints/Preferences
- Whatever execution mode is active, the shell named to the model must be the shell that actually runs its commands.
- Accompanying advice about shell-specific command syntax and chaining should come from that same, accurate source rather than the VS Code setting alone.
Related reports
Previously reported and closed: #82 (System prompt on Windows says shell is cmd.exe but PowerShell is used via the VS Code terminal) and #321 (Windows: shell mismatch between AI prompt generation and Inline Terminal execution). Still open, #508 (Inline terminal always uses system shell instead of what shell is configured in VS Code) covers the execution side only, not what the model is told. No open issue was found for the system prompt reporting the wrong shell.
Problem / Value
When the Inline Terminal is enabled, Zoo tells the model that its commands will run in the shell configured in VS Code, but the commands actually run in a different shell - on Windows, the plain Command Prompt. The model then writes commands in a dialect the terminal cannot understand, and has to guess and retry before anything works. Making the reported shell match the real one would remove failed commands, wasted turns, and needless confusion for both the model and the user.
Context
Affects Windows users whose VS Code default terminal is something other than Command Prompt (PowerShell, PowerShell 7, Git Bash, etc.) while the recommended "Use Inline Terminal (recommended)" setting is left on - which is the out-of-the-box configuration.
What happens today: the environment summary handed to the model names the VS Code shell (PowerShell in my case), regardless of the fact that, with the Inline Terminal active, commands the model runs are executed by Command Prompt. The model only discovers the mismatch by failing.
What should happen: the model should always be told the kind of shell it will actually get, so its very first command is written for the right shell.
Reproduction
Constraints/Preferences
Related reports
Previously reported and closed: #82 (System prompt on Windows says shell is cmd.exe but PowerShell is used via the VS Code terminal) and #321 (Windows: shell mismatch between AI prompt generation and Inline Terminal execution). Still open, #508 (Inline terminal always uses system shell instead of what shell is configured in VS Code) covers the execution side only, not what the model is told. No open issue was found for the system prompt reporting the wrong shell.