Engineering framework and approvals

The repository has three instruction layers:

Layer	Source	Purpose
Task framing	prompts/catalog.yaml and selected prompt	Define intent and analysis focus
Repository policy	Root and nested AGENTS.md	Describe constraints and engineering expectations
Workflow	playbooks/catalog.yaml and selected playbook	Sequence implementation, validation, and review

Load only the context relevant to a task. A catalog entry is a routing index, not a permission grant.

Execution and review levels

The historical “four-tier” document numbered levels 0 through 4. This site uses five execution and review levels to remove that ambiguity:

1. Level 0 — Deterministic checks: linters, tests, and repeatable automation.
2. Level 1 — Local inference: explain, retrieve, and draft using local models.
3. Level 2 — Engineering agents: inspect and modify files within authorized scope.
4. Level 3 — Deeper analysis: review complex tradeoffs with suitable models and evidence.
5. Level 4 — Human authority: approve consequential actions and accept outcomes.

These levels describe responsibilities, not a ranking that makes higher model output inherently correct. Human authorization can be required at any stage.

Practical workflow

State the goal and permissions → inspect repository policy → choose a prompt/playbook → plan → implement within scope → verify → review → separately authorize publication.

The backend reads selected YAML metadata for model routing. It does not enforce the complete playbook lifecycle or an agent’s filesystem/network permissions. The agent harness and operator must enforce those boundaries.

Handoffs

Use the planner contract to make proposed execution concrete. Include evidence, assumptions, acceptance criteria, and rollback implications rather than merely asking another model to “finish the task.”

Next: Contribution workflow.