ADR 0026: Per-agent admission of new runs

Status: implemented in development.

Operators need to take a registered agent out of rotation without deleting its identity, key, history or currently running work. Reuse the existing disabled flag rather than introduce a second scheduling state. No metadata migration or VCS format change is required.

The local CLI and API share the same transactional state-change function. The CLI agent list/disable/enable --remote NAME uses the same public gRPC API, negotiates agents.state for changes and captures the observed state before submitting a guarded, retry-safe mutation. Starting a run reads the agent's state inside the run creation transaction, so a start and a disable are serialized by repository metadata writes. Once disable commits, later starts are refused. Runs already admitted remain valid; disabling is not credential revocation or process termination. Task cancellation and run reaping retain their separate meanings.

PATCH /agents/{name} and AgentService.SetState require agent.operate and the existing observed expected_disabled value. An agent principal may change only its own identity, consistent with the other agent operations. This is a current-state comparison, not a revision history check. An idempotency key records the response in the same transaction as the state change, audit record and agent.state outbox event. Retries replay that response; reusing a key for another actor or request fails. Repeating a matching no-op state does not add another audit event.

The agents.state feature controls the browser actions. The confirmation names the agent and explains that existing sessions continue. The UI captures the observed state before confirmation, preserves errors, and discards late results after repository/account changes. Older servers retain a read-only status column.