> NEW TASK
> TASKS
⋯
> MODELS
DISCOVERING...

Every model once. Its chips are the harnesses that run it, with the router's pass rate (~ is an estimate, no runs yet). Tap one to dispatch with that pair.

COORDINATOR
…

Who coordinates the fleet: writes the task specs, reviews the results, answers approvals. Recorded with every task and shown in both dashboards. It does not pick the lane for a task (the router does), but its own account stays in reserve by default (Budgets), so the fleet leaves it its quota.

ROUTING STRATEGY: Router: measured pass rate per category over cost, failover with one repair pass
SANDBOX ISOLATION: checking...
CODE INTEGRITY: Scope, syntax, repository gate, task gate
BUDGETS
…

How much of each provider's quota the fleet may use. Normal: the router decides by measured quality. Reserve: only when everything else failed. Manual: never automatic, only when you pick the lane. Pick a policy, add an optional reason, then Save on that row.

> SETTINGS
Configure fleet dispatch autonomy governance and automated execution gating.
LimitsHow long one agent run may take
How long one agent run (daemon executor or fleet run) may take; then its whole process tree is stopped and the work so far stays on the task's branch. With the grid backend the runner has its own timeout.
Quality gatesWhat has to pass before a task is done
ALWAYS ON
Before a task counts as done: the change stays inside its allowed paths, Python still parses, the repository gate passes (projects.<name>.gate, verify.sh or pytest), the task's own test passes and the files it claimed changed. The gates run in the agent's sandbox without credentials. They cannot be switched off here: a task that skipped them would not be verified.
Keys and sandboxWhere provider keys come from; the kernel sandbox
LANDLOCK + SECCOMP SANDBOX
Agent CLIs write only to their worktree and their own state. Not switchable from the web: only FLEET_SANDBOX=0 in the daemon's environment turns it off.
checking...