Project model-use policy

Optimize total usage per correctly completed task while preserving the project’s correctness, validation requirements, and intended scope.

MODEL ROLES

Sol (gpt-5.6-sol) is the preferred lead. Own requirements, task breakdown, important decisions, integration, and final verification. Handle small tasks directly when delegation would add more overhead than value.
Explicitly delegate independent, well-defined implementation, investigation, and testing tasks to Terra (gpt-5.6-terra).
Use Luna (gpt-5.6-luna) for narrow searches, factual extraction, log summaries, and mechanical edits with clear acceptance criteria.
Use Astra (gpt-6-astra) for genuinely difficult reasoning: ambiguous architecture, subtle correctness problems, complex reverse engineering, or unresolved contradictions. Delegate the smallest complete problem that needs its expertise. Escalate immediately when difficulty is clear; do not require cheaper models to fail first.

This policy explicitly authorizes subagents and the model choices above. Use actual supported model-selection tools; never merely roleplay a model.
Delegate only when a bounded task can run independently alongside useful lead work and is likely to save total effort or improve the result.
Start with at most two workers at once. Workers should return to the lead for further delegation rather than spawning more workers.
Give each worker the goal, relevant files and evidence, constraints, acceptance criteria, and clear ownership of any files it may edit.
Supply a concise, self-contained brief and required project instructions instead of copying the full conversation when supported.
Prevent overlapping edits. Do not duplicate a worker’s assignment.
Request concise results: findings or changes, evidence and file references, checks performed, and remaining uncertainty.

Start routine worker tasks at low or medium reasoning; use high effort for difficult reasoning. Do not default every task to maximum effort.
If an attempt fails, use the evidence to refine or escalate it. Do not repeat essentially the same attempt or cycle through models without new information.
The lead must assess worker output and verify integrated behavior with appropriate tests or direct evidence. Agreement between agents is not proof. Preserve all required project checks.
Prefer small, evidence-backed changes and existing tools. Avoid speculative scaffolding, unnecessary dependencies, and redundant testing.
If a requested model or delegation capability is unavailable, state that briefly and use the best available approach without claiming a switch.
Finish with the outcome, validation, and unresolved limitations. Mention model escalation briefly when relevant; do not invent usage savings.