Governance

Breaker

This is one policy kind in the ADR-014 governance engine. Policies resolve most-specific-wins with deny-wins at a tie, across scopes that run agent → user/role → workspace → team → org. Rules of the same kind combine rather than replace, so what matters is which rules exist, not which policy row holds them.

What it does

A breaker names a signal, a window, a threshold and an action. Signals are counted per agent as runs happen. When the count crosses the threshold inside the window the breaker trips, once per cooldown, and either notifies the agent's author or pauses the agent. Until now every stop in Trellaris was a person pulling a switch; a breaker pulls it before anyone is at a keyboard.

Signals

Signal	Counts
`injection_blocks`	Each tool result the injection screen refused.
`dlp_redactions`	Runs whose output needed DLP redaction (once per run, however many spans).
`perimeter_blocks`	Each outbound call the data perimeter refused.
`tool_errors`	Each tool call that returned an error.
`run_failures`	Runs that ended failed (once per run).
`thumbs_down`	Runs the person rated 👎 (once per run).
`approval_denials`	Each approval a person denied.
`cost_velocity_usd`	Dollars spent; the threshold is an amount, not a count.

Actions

Action	On trip
`notify`	The agent's author is told once per cooldown. Nothing else changes.
`pause`	The agent is paused through the same kill switch an admin uses: new runs are refused, in-flight runs finish. The author is notified, an incident is opened, and the agent's page shows Paused by a breaker until an admin resumes it. Resuming re-arms the agent's breakers.

The Orchestrator is never paused by a breaker A breaker scoped to the router with pause is refused at save; on the router every breaker is notify-only. Pausing the router would stop every conversation in the organisation, and the brake for that is the emergency stop, which an admin pulls deliberately.

The defaults

Two notify-only breakers are seeded per organisation: injection blocks ≥ 5 in 60 min and perimeter blocks ≥ 5 in 60 min, each with a four-hour cooldown. They change nothing an agent does; they tell an author that something is repeatedly being refused. Switching one to pause, or adding a thumbs_down or cost_velocity_usd breaker, is the admin's decision.

Admin → Safety → Breakers: the two seeded rules in a table with signal, window, threshold, action and cooldown, and a per-agent panel of count/threshold chips. — Admin → Safety → Breakers: the rules in force and, for every agent with recent signals, how close it is to tripping.

Floors

Under ADR-030 a floored breaker is a minimum: a narrower scope may lower the threshold, shorten neither the window nor the cooldown, and may move notify to pause but not back.