feat: lead/worker model #2719
feat: lead/worker model #2719
Conversation
|
@angiejones interesting - I would think that you would configure that ahead of time - if you don't want it to do it, dont' have that extra model (as by default wont have one)? |
|
baxen
left a comment
baxen
left a comment
There was a problem hiding this comment.
Looks good! love the idea and good call to make it a new provider type here
i think there are some issues with the failure handling still? but it mostly worked for me!
|
okay @michaelneale so making sure I understand this:
Am I understanding this correctly and if so, could we add a little visual indicator? Besides that, this PR is cool with me. I'm team "switch the model without me knowing..do the heavy decision making for me" Sidenote: apologies for my rogue issue and comment that was added to this PR. I asked Goose to help me catch up on the Discord conversations related to this PR..and then it was like I will leave a comment on and open an issue based on the Discord conversation. Claude 4 is too powerful and fast 😆 |
|
@blackgirlbytes no worries - meant to ask for your feedback, yes you have it right, will be a subtle indicator it is switching models (we can improve it over time too) of course few will use this feature out of the gate as it won't be a default feature of course. It isn't just about cost too, it could apply to local models where very very slow ones are used as the lead vs the worker which is faster and more efficient |
oh cool! I like it! |
* main: (82 commits) feat: lead/worker model (#2719) fix: pass ref in pr comment workflow (#2777) feat: goose web for local terminal alternative (#2718) chore: run CI on merge_group (#2786) fix: Don't break from consuming subprocess output in shell tool until both streams are done (#2771) Add retries w/ exponential backoff for databricks provider (#2764) Fix paths in google drive mcp documentation (#2775) testing windows build (#2770) docs: Add Context7 YouTube Video (#2779) cli(command): Add `export` command to CLI for markdown export of sessions (#2533) fix(copilot): gh copilot auth token conflicts w/ gh mcp env var (#2743) feat(providers): Add support for Gemini 2.5 Flash Preview and Pro Preview models (#2780) fix: pr comment build cli workflow (#2774) hotfix: don't always run prompt (#2773) Lifei/test workflow (#2772) chore: use hermit to install node, rust and protoc (#2766) Feat: Refined the documentation for Goose (#2751) mcp(developer): add fallback on .gitignore if no .gooseignore is present (#2661) cli(ux): Show active context length in CLI (#2315) cli(config): Add GOOSE_CONTEXT_STRATEGY setting (#2666) ...
* main: feat: lead/worker model (#2719) fix: pass ref in pr comment workflow (#2777) feat: goose web for local terminal alternative (#2718) chore: run CI on merge_group (#2786) fix: Don't break from consuming subprocess output in shell tool until both streams are done (#2771) Add retries w/ exponential backoff for databricks provider (#2764) Fix paths in google drive mcp documentation (#2775) testing windows build (#2770) docs: Add Context7 YouTube Video (#2779) cli(command): Add `export` command to CLI for markdown export of sessions (#2533) fix(copilot): gh copilot auth token conflicts w/ gh mcp env var (#2743) feat(providers): Add support for Gemini 2.5 Flash Preview and Pro Preview models (#2780)
* upstream/main: Iand/blog goosehints metadata update (block#2800) Iand/blog goosehints (block#2798) blog post about goosehints and persistent context (block#2796) [goose-llm] system prompt override (block#2791) chore: small bit of a cleanup - removing unused dir (block#2761) feat: lead/worker model (block#2719) fix: pass ref in pr comment workflow (block#2777) feat: goose web for local terminal alternative (block#2718) chore: run CI on merge_group (block#2786)
I have been using with good effect an expensive model like opus claude4 for a bit, and switching to claude4 sonnet mid stream.
This feature will automatically switch from large to small as session goes on so you can make more efficient use if you set a lead model. this is similar to the
/planmode but you don't have to remember to call it.How it works:
set the GOOSE_LEAD_MODEL and optionally GOOSE_LEAD_PROVIDER
this works with the plan feature, and similar to it but is automatic for tool calling/execution, vs plan which is more for planning/chat.
Also added docs and connected it with the plan feature (existing)