You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
bug? Denying a tool call will continue generation #31
Not sure if this is expected behavior but when I deny a tool call I normally like to tell the llm the reasoning as to why, but I have noticed that if I deny a tool call it will just continue the loop and the LLM will start generating again right away, (not giving me a chance to instruct it better).
I wasn't sure if this was intentional behavior or if it was a bug. I was kinda used to the previous paradigm so I figured I would open issue to see if it was intentional, if it is then I can close it.
this sounds like a bug. i hadn't ran into this but my expectation is the same as yours, it should stop. i also want to add real-time steering and there's a current bug where after interrupting we don't always send a valid message structure back to the model on subsequent calls. those are separate but related issues.
Since Dax is doing his big refactor I will leave this be until he is done, if the issue persists I will open a pr
yeah, good shout for everyone that's been active here @PhantomReactor@ezynda3@Adictya@l0gicgate, probably best to hold off on any changes that don't live in the internal/tui dir, dax is rewriting the backend in JS (mullet stack!)
Reacted by Aiden Cline, Adictya and phantomreactor
Interesting. Do we plan to basically spawn the TS server in the background? Also I generally stay on the bleeding edge by using go install direct from the github repo but I suppose that won't work now with there being two separate binaries. Could be cool if the CLI downloads the server on first start maybe.
Activity
adamdotdevin commentedon May 17, 2025
this sounds like a bug. i hadn't ran into this but my expectation is the same as yours, it should stop. i also want to add real-time steering and there's a current bug where after interrupting we don't always send a valid message structure back to the model on subsequent calls. those are separate but related issues.
ezynda3 commentedon May 18, 2025
I have run into this as well. I totally forgot about it and already conditioned myself to press
ESCas soon as I deny a tool call lol.rekram1-node commentedon May 18, 2025
I think Ik the issue, I will prolly make a pr for the fix today or tmr
rekram1-node commentedon May 19, 2025
Since Dax is doing his big refactor I will leave this be until he is done, if the issue persists I will open a pr
adamdotdevin commentedon May 19, 2025
yeah, good shout for everyone that's been active here @PhantomReactor @ezynda3 @Adictya @l0gicgate, probably best to hold off on any changes that don't live in the internal/tui dir, dax is rewriting the backend in JS (mullet stack!)
adamdotdevin commentedon May 19, 2025
this will reduce scope a ton (we can leverage the AI SDK to support all providers without owning any of the code) among other things
ezynda3 commentedon May 19, 2025
Interesting. Do we plan to basically spawn the TS server in the background? Also I generally stay on the bleeding edge by using
go installdirect from the github repo but I suppose that won't work now with there being two separate binaries. Could be cool if the CLI downloads the server on first start maybe.adamdotdevin commentedon May 19, 2025
i know distribution is the tricky part of this, i'll let @thdxr chime in on that when he's ready
rekram1-node commentedon Jun 14, 2025
closing, this is no longer relevant in latest opencode release.
feat(session-start): flexible project structure discovery (#31)
Onboarding 1.0: complete workspace-first flow (anomalyco#31)
feat(rust): third LLM protocol — gemini (decode + lowering)