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
This includes brainstorming ideas (long-term vision). A lot is out-of-scope for the MVP: we can only implement the MVP-relevant parts.
Sidebars:
First sidebar (on the left):
Three sections: Overview, Projects, Queue
Projects: shows the list of all projects where The Framework is installed
Queue: essentially all the TODO.md entries of all projects
Overview: some kind of nice overview of current prompts AI is working on, the queue, and recent projects
Second sidebar (on the left): shows the latest AI loops/prompts of the currently selected project (the most recently active project if none is selected)
In other words: clicking a project => essentially shows .the-framework/LOGS.md in the second sidebar
Third sidebar (on the right)
Show arbitrary views (e.g. when AI calls the function showMarkdown(), <Choices>, ...)
Sticky top nav to jump between views (e.g. for jumping between PLAN.md and TODO.md)
Main view: shows the last loop/prompt of the selected project (similar to https://claude.ai/code).
Installing projects
A repo is activated when it has a .the-framework/ directory
Installation: the user can click on a button "Add project(s)" => user can add a directory of repos, or a single repo
When adding a directory: add all child directories that are Git repos
Commits an empty .the-framework/LOGS.md with commit message [The Framework] install The Framework
If the Git repo is dirty, then commit the uncommitted changes with a commit message [The Framework] uncommitted changes
<Choices>
When AI needs the user to choose between a list of options to proceed
Ideally, choices always have a recommended choice (if a recommendation is missing, then pick first choice)
Click green button Accept (or shortcut Ctrl + Enter) => accepts (picks the recommended choices)
If [x] autopilot => auto-accepts after 10 seconds (while the UI says "move mouse to abort countdown")
All sets of choices (Choices[], not Choice[]) are shown at once in the right sidebar (no pagination, long vertical scroll instead, with sticky top nav to quick-jump between all the sets of choices)
Global options
boolean Autopilot (whether to auto-accept)
boolean Technical control (expose technical details, e.g. for tech stack)
boolean Eco (fine-grain control over system prompt to save resources)
boolean Auto planning
boolean Auto research
boolean Auto maintenance
boolean Vanilla (remove all system prompts, fully transparent, same as directly using Claude Code)
Context selector — multi-select dropdown (to select list of directories)
This just adds one line to the system prompt: Context: [list-of-dirs]
AI always has access to all repos (but it doesn't mean it should read all repos, we therefore need a "context selector")
When adding repos: "Do you trust this repository? ("You must trust it, otherwise high risk of [prompt injection](link-that-explains-the-problem).")
Note
This includes brainstorming ideas (long-term vision). A lot is out-of-scope for the MVP: we can only implement the MVP-relevant parts.
Overview,Projects,QueueProjects: shows the list of all projects where The Framework is installedQueue: essentially all theTODO.mdentries of all projectsOverview: some kind of nice overview of current prompts AI is working on, the queue, and recent projects.the-framework/LOGS.mdin the second sidebarshowMarkdown(),<Choices>, ...)PLAN.mdandTODO.md).the-framework/directory.the-framework/LOGS.mdwith commit message[The Framework] install The Framework[The Framework] uncommitted changes<Choices>Accept(or shortcutCtrl + Enter) => accepts (picks the recommended choices)[x] autopilot=> auto-accepts after 10 seconds (while the UI says "move mouse to abort countdown")Choices[], notChoice[]) are shown at once in the right sidebar (no pagination, long vertical scroll instead, with sticky top nav to quick-jump between all the sets of choices)booleanAutopilot (whether to auto-accept)booleanTechnical control (expose technical details, e.g. for tech stack)booleanEco (fine-grain control over system prompt to save resources)booleanAuto planningbooleanAuto researchbooleanAuto maintenancebooleanVanilla (remove all system prompts, fully transparent, same as directly using Claude Code)Context: [list-of-dirs][prompt injection](link-that-explains-the-problem).")