Conversation
|
We could totally do this, but I could see some downsides. The model might expect to be using bash specifically, unless we update system prompts/tool descriptions. So if it isn't instructed to do so specifically, it could generate invalid shell commands thinking it's bash |
|
@jamadeo oh... hrm... that is a great point. zsh is kind of a superset of bash so I expect would work the same. Conversely, bash on macos is getting old (bash 3.2 when linux is > 4) and is lagging modern bash so it could go the other way where it will try to do things that won't work (especially if it ends up deprecated), not that I Have seen that. so if zsh is a reasonable updated shell, then probably better to do this (but then also is a change). If it is not, probably better to stay with bash. Zsh is a posix shell which I would think LLMs are mostly trained with (due to linux). We don't prompt either way I believe, so the LLM will just see "shell" and think whatever is most idiomatic on that platform, which could go either way I guess (and probably drift over time). |
|
I looked at how claude code works, and while its tool is always called bash, it does use zsh on my system. Presumably it's using $SHELL then? So, yeah, let's go for it. Maybe though, should we tell the model in the tool description which shell executable it will use? I suppose if it wants to it can always specify by running |
|
I think it is enough to say it says to use the system SHELL - it should be able to work that out. |
|
we could still inject it into the system prompt, yeah? |
|
yeah - not sure if it is necessary, as it can work it out itself as it goes if it needs to? |
Co-authored-by: Jack Amadeo <jackamadeo@squareup.com>
…ovements * 'main' of github.com:block/goose: (23 commits) blog post on subagents vs subrecipes (#4829) fix chat button alignment and spacing for attachments (#4794) fix: remove nested double quotes in windows automation_script tool description (#4824) fix: a few things with the mcp snapshot test (#4818) Revert "fix(compaction): try to catch more context limit exceeded erors and compact" (#4820) test: add test coverage for Tools Inspector (#4700) feat: Parse and use retryDelay from Google API RateLimitExceeded errors (#4124) cleanup: remove unused link preview and goose response form components (#4795) fix build: latest bedrock version (#4812) prefer users SHELL (#4702) feat: update aws-sdk-bedrockruntime to enable AWS_BEARER_TOKEN_BEDROCK auth (#4327) correct the tests from an odd merge (#4804) docs: import yaml recipe (#4799) docs: Add openmetadata extension to goose mcp docs (#4547) Add elapsed time to the CLI output. (#4609) fix: Fix cell coordinate ordering in XlsxTool and add unit tests (#4551) Use gemini flash for summarization on open router (#4290) chore(deps): bump xcb from 1.5.0 to 1.6.0 (#4289) feat(shell): throw errors on interactive commands (#4788) feat: AgentManager - foundation for unified execution (#4389) (#4684) ...
Co-authored-by: Jack Amadeo <jackamadeo@squareup.com> Signed-off-by: HikaruEgashira <hikaru-egashira@c-fo.com>
3 participants
This is one option - use the $SHELL value and delegate to that. That allows it to load appropriately if zsh etc, fall back to bash