Keep agents inside the current reality.
An agent should not just validate context before it reasons. It should validate that the world it reasoned about is still the world it is about to act on.
Where context comes from
Context leases
Pre-action gate
Immediately before a consequential action, re-read every watched source and run four independent checks plus concurrency. The first failure decides.
Human decision
In live mode, decisions happen in Linear: comment on a watched issue or edit the task description, then sync.
Event trail
Audit
Who knew what, decided what, acted when. Each decision carries the lease revision and every fact version behind it.
| Time | Agent | Action | Rev | Decision | Sources |
|---|
Pause the process, not just the PR
Wrap a long-running agent with the session guard. On invalidation the process receives SIGSTOP; a clean re-plan sends SIGCONT. It reports here so you can watch it. Wrap single consequential commands with guarded-action.
Import from Linear
Reads the issue, its parent, issues that block it (dependencies), and related issues (constraints). Consequential gate calls re-sync before deciding.