Knowledge source reference
Goal Loop and Skill Routing
Route a request into one primary Valdris skill and execute bounded proof-bearing checkpoints.
Start
Use each installed skill's SKILL.md YAML frontmatter for implicit Codex discovery. Read skills/codex-routing.yaml for the complete selection projection and skills/registry.json for the gate-authoritative source. Select exactly one primary skill and the smallest supporting set, never more than four.
Durable State
Write goal/goal.json with the objective, owner, authority boundary, budgets, checkpoints, and measurable stopping conditions. Runtime goal/loop features may mirror this state but never replace it.
Loop
At each checkpoint:
- Inspect evidence and remaining risk.
- Choose the smallest bounded action that can advance a stopping condition.
- Execute within tool, cost, attempt, token, and time budgets.
- Run the required deterministic, eval, trajectory, production, AI, and proof gates.
- Continue only with new evidence and remaining authority.
Finish
A completed goal requires all stopping conditions to pass with typed evidence, no trajectory violation, no exceeded budget, no unresolved Red Zone approval, and an evidence-backed handoff.