Skip to main content

Approach and Rationale

This page explains why Livon uses a small control plane for agents instead of full-load prompt files.

Goals​

  1. Lower token usage without dropping critical rules.
  2. Improve consistency by routing only relevant context.
  3. Keep one documentation truth for humans and agents.
  4. Support both Codex and Copilot with the same governance model.

Design decisions​

  1. Keep canonical engineering and product rules in docs and package prompts.
  2. Keep AGENTS.md small and operational (routing, scope lock, gates).
  3. Use scoped loading over hard full-load links.
  4. Use active rules plus gate review as enforcement.
  5. Use skills for deep, task-specific behavior.
  6. Keep routing machine-readable for regression checks and metrics.

Why this model​

Full-load context increases token cost and reduces instruction clarity in large monorepos. Scoped routing and progressive disclosure keep high-signal context active while preserving full rule coverage through conditional loading.

Operational checks​

  • pnpm check:policies validates:
    • required AI control files,
    • routing fixtures from configs/ai/context-routing.json,
    • root load-budget limits,
    • naming consistency for AI markdown.
  • pnpm metrics:ai-control reports routing/load metrics for trend tracking.

Multi-agent policy​

Use multi-agent mode for parallelizable tasks (exploration, review shards, migration decomposition). Keep single-agent mode for small or linear tasks. Use the role council contract from Multi-Agent Council when product intent, architecture, and performance decisions must be coordinated.

Official references​

Codex and AGENTS:

Copilot custom instructions and skills:

Open standards used by this approach: