The agent starts every thread knowing Creator and Deluge: Zenith ships with a knowledge pack written for the model, and a Deluge linter whose rules come from Creator’s real save errors. You add what is specific to your project.
What Zenith already knows
- Reference topics on Deluge syntax, criteria and fetches, inserts and updates, loops and limits,
invokeurland connections, workflows and schedules, pages, environments, error handling, performance and security. The agent loads a topic only when the task needs it. - Lint rules for the mistakes that fail on save or misbehave silently. Some are checked against your app’s own schema, such as a field name with the wrong case. Browse them in Deluge lint rules.
- Your app, from the mirror: every component, its fields and who calls what.
Bring your own rules
Point Zenith at a project folder that contains CLAUDE.md, AGENTS.md, .cursorrules or a docs/ folder. Before anything is kept, Import review splits each file into sections and sorts them:
| Classification | Default |
|---|---|
| Covered by the built-in pack | Off |
| Project-specific | Kept |
| Conflicts with a pack rule | Shown side by side; you decide |
| Generated or stale (for example schema dumps the mirror replaces) | Off |
A budget meter shows how many tokens are loaded into every turn and how many are loaded on demand. Secrets in imported text are redacted before it is stored.
Feature folders
If you keep feature specs in a repository (a features/ folder with ZENITH.md, IMPLEMENTATION.md and TESTING_PLAN.md), Zenith can import them, link the folder so it reads and writes in place, and check each imported feature against the mirror. Git stays yours.