Policy

This page settles what each policy knob changes and what a sane default looks like, for the four values most teams touch first.

review_tier

Which provider tier fills the charter reviewer and adversary seats for this team's runs. Raising it spends more per review in exchange for a stronger read on harder changes. A sane default is the same tier the builder runs on — spending less on review than on the work it's checking usually means the review is the weaker of the two.

plan_competition.min_tier

The minimum tier a candidate plan must have run on to be considered when more than one plan is generated for the same task. A sane default is whatever tier plan itself normally runs on — this knob exists for teams that want cheap draft plans filtered out before a human ever compares them, not to gate every team's planning step.

build_budget_usd_max

The ceiling budget_usd on a build node can't exceed, even for a team that wants to try raising it. This is the enforcement point for Budgets's tighten-only rule at the team level: a team can set a build's own budget_usd lower than this, never higher. A sane default is the base tier's own build ceiling — a team wanting more room should be splitting the task, not raising this.

dispatch.max_in_flight

The cap on how many runs the coxswain will dispatch at once — the in-flight bound described in Crew and seats. A sane default is small enough that a human reviewing pull requests isn't handed more at once than they can actually read; raising it trades review attention for throughput, and that trade should be a deliberate policy choice, not a default nobody looked at.