Agent system
Veyra is a hierarchy of agents — four primaries that decompose work and delegate to 47 specialised sub-agents.
The four primary agents
| Agent | Responsibility | Typical sub-agents |
|---|---|---|
| Planner | Decompose intent into a verifiable plan | Intent parser, Spec writer, Risk scorer |
| Coder | Edit, refactor, generate code and tests | Type-fixer, Test author, Migration writer |
| Reviewer | Verify diffs, run static analysis | Lint, Sec-scan, Diff-critic |
| Operator | Execute commands, run tests, deploy | Sandbox, Self-heal, Rollback |
Delegation
Primary agents never call models directly for execution work. They emit delegations — typed requests that the runtime routes to a sub-agent with a narrow toolset and a strict success contract.
Self-healing
If a step fails — a test breaks, a type fails, an env var is missing — the Operator triggers a Self-heal sub-agent. Self-heal can rerun, patch, or escalate up to the Planner. Most recoveries complete without notifying the user.
Recovery budget
- Default: 3 self-heal attempts per failed step.
- If exhausted, the run halts and is surfaced for review.
- Workspace policy can raise or lower this per project.
Tracing delegations
Every delegation is visible in the trace as a nested span. Open Run history and expand any node to see which sub-agent ran, with what tools, and what it returned.