rikrdo/arnes
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
3d41579ad31a215a87dbe4c5505ef96a7a4b27c6
arnes/harness/policies/model-routing.md
rikrdo 3ff9b70e4c refactor: complete bootstrap of ARNES agent harness framework 
- Add complete agent harness structure with 8 roles (leader, triager, architect, implementer, reviewer, security, qa, documenter)
- Implement strict workflow with 9 stages and mandatory gates
- Add comprehensive verification script and runtime status tracking
- Create artifact-based evidence system with contracts and schemas
- Add agent policy matrix with permissions and anti-cheat rules
- Include test suite (44 tests passing) and CI-ready structure
- Add documentation: README, HOWTO, CHECKPOINTS, templates
- Configure model routing policies and token-aware task assignment
- Add BDD/SDD specification guides and feature templates
- Include starter pack for quick project onboarding

All verification checks pass. Framework ready for production use.
2026-05-17 23:25:35 +02:00

638 B
Raw Blame History

Policy: Model routing

Use model by task complexity, not by habit.

Core rule

  • Start small.
  • Escalate only when blocked or quality poor.

Escalation triggers

  • Repeated failed attempts.
  • Ambiguous requirements.
  • Cross-module side effects.
  • Security-critical code paths.

De-escalation triggers

  • Routine CRUD edits.
  • Mechanical refactors.
  • Artifact writing.
  • Status/timeline updates.

Required behavior

  • Record chosen model class in artifact header when work is non-trivial.
  • Keep outputs concise to reduce token burn.
  • If harness/project.config.json has model_mode=lean, prefer tiny/small whenever possible.