What we're thinking about

[DOsinga]

What we're thinking - mid august

⚡ Finish up (auto) compaction

Deliver the left over bits. Merge the better token count approach, refresh the UI so it is all in one place, get the todo tool merged, agent tool calling loop compaction and having a provider specific model for compaction (cheaper, faster)
See: #3944

🔄 Refresh our MCP server selection

Go through all the MCP servers we suggest on Goose and see if they work on a clean install. Remove any MCP server that doesn’t work or doesn’t seem maintained. Also go through the most popular new ones and add them. See #3849

🚀 Make the installing of MCP servers faster and report errors better

We should improve our whole error stack and make it more transparent. There's more low-hanging fruit to tackle: supporting all features in the latest spec (roots, sampling, elicitation, improved oauth, etc). For the slowness, much of it stems from blocking on each server's startup - this is especially noticeable on the CLI.
See Discord, #3910, #4016, #3886, #3945, #4006, #3990, ✅: #3872

⚙️ Settings and Providers

After we merge #3824 we have per (custom) provider jsons. We can expand this to cover also the non custom providers so all settings for all providers are then done the same. Then by bringing in elements of #3975: unify the naming, add types to configs, remove the reliance of the desktop app on environment variables, make all access to what are now environment variables go through the settings module and pick a better way to store settings (use something that doesn’t rewrite the file, just updates so comments are maintained, possibly switch to a INI file like format)
See: #3857, #4046, #4036, #4029, #3868

🍳 Make recipes universally work

Enable subagents to run with specified model/settings and extensions. Refactor subagent implementation to reuse agent implementation as much as possible, including retry and structured output. Move recipe logic to agent core and merge UI for scheduling and recipes.
See: #4040, #3977, #3972, #3938

🪟 One window client

This is a big one and will require multiple steps. Roughly, first make the server the source of truth for agent interactions, then make agent.reply stateless, then use only one goosed for all windows, then switch to one window.
See: #4044

[towry]

The issues listed have not updated over a month.