Agent documentation
August 19, 2026 ยท View on GitHub
This directory indexes the repository context used by coding agents and human contributors. The content is portable: agents that discover repository skills can load them automatically, and other agents can follow the same links from AGENTS.md.
Choose the right layer
| Layer | Purpose | Location |
|---|---|---|
| Persistent guidance | Rules, commands, source routing, and review expectations that apply to every task | AGENTS.md |
| Architecture | Current runtime boundaries, modules, persistence, calculations, file flows, and invariants | ARCHITECTURE.md |
| Unified UI/UX migration | v0/v1 boundaries, shared-contract protection, and just-in-time workflow notes | unified-ux-migration.md |
| Data workflow | Practical examples for workspace reads, selections, persistence, transactions, and committed refreshes | data-access-and-workspace.md |
| Task context | Minimal entrypoints, document sections, modes, skills, risks, and initial validation | context-index.md |
| Task modes | Implementer, Investigator, Reviewer, Test Engineer, and Designer checklists | personas.md |
| Skills | Repeatable task workflows loaded only when relevant | .agents/skills |
Executable configuration and source code remain authoritative for volatile facts. These documents explain how to find and safely change that source.
Start a task
Use the task context index to select entrypoints, exact sections, mode, skill, cross-boundary triggers, and initial validation.
Maintenance
- Keep
AGENTS.mdsmall and limited to durable repository-wide guidance. - Put descriptive system context in
ARCHITECTURE.md. - Put workflow steps in the matching skill instead of duplicating them here.
- Keep
context-index.mdlimited to durable routing and starting points, not implementation inventories. - Update
personas.mdonly when a task mode's responsibilities change. - Prefer source links over copied inventories or version numbers.
- Validate relative links, commands, and paths in the same pull request.
When an agent makes a recurring wrong assumption, identify the narrowest layer that would have prevented it and update that source.