Core concepts
Project context
The relevant slice of shared knowledge, close to the work.
A project is an application directory connected to one or more homebases. Each connection makes selected Markdown available to the project and its coding agents.
The goal is useful context, not the largest possible pile of documentation.
Connect an application
From the application directory:
mindslice init .Review the detected role, project location, selected homebases, and context rules. Existing valid metadata is reused; conflicting or incomplete metadata should be investigated rather than overwritten.
If the directory is inside another repository, setup may ask whether integration belongs in this directory or the nearest Git root. Choose the location where the agent actually works.
Select for relevance
Project include/exclude rules choose which homebase Markdown belongs in a context view. For example, a frontend may need product flows and UI conventions while an API needs data contracts and architecture decisions.
Rules live in the homebase manifest and are reviewed during project setup. Don’t broaden another project’s selection just to make a missing document appear. Review the intended scope first.
Mindslice also filters agent instruction files, hidden files, and template directories from application-facing knowledge surfaces. A broad include rule does not override those filters.
Selection is not access control. Use appropriate filesystem and repository permissions for confidential knowledge.
Understand the generated view
Project metadata is kept in .mindslice/project.json. Generated context appears under .mindslice/context/, using a linked view or a copy fallback depending on the environment.
The homebase remains canonical. Generated context is not something to edit independently or check into the application repository.
If you use Git, review and share the descriptor and owned integration files as appropriate. Keep generated context ignored.
Inspect what the agent can see
From the connected project:
mindslice status
mindslice knowledge tree
mindslice knowledge search "architecture"Use the tree to find a page, then read its actual path:
mindslice knowledge read architecture/system.mdThe final path is an example. Use a selected path from your own tree. Read-only knowledge commands expose the current project’s selected Markdown, not unrestricted homebase contents.
With multiple connected homebases, specify one explicitly:
mindslice knowledge tree --hb <homebase-id>Keep context current
The local service watches for local changes and maintains connected views. You can request a refresh explicitly:
mindslice refresh <project-connection-id>Use the connection ID reported by mindslice status, not the project ID from the descriptor. A refresh updates the generated view; it does not change selection rules or fetch remote Git changes.
If a view is unhealthy, inspect the reported issue. Don’t delete unexpected files or reset project metadata just to clear a warning.
