Operations
Keep outcomes inspectable.
A decision you cannot reconstruct is a decision you cannot defend. Evidence, audit history, recovery, and emergency stop exist so that the record survives the incident — and so that stopping is always available.
- Concepts
- Evidence and recovery
- Read before going to production
Signed evidence
Every governed request leaves an event, and every event is chained to the one before it with an HMAC-SHA-256 digest. The chain head is signed and anchored, so a record cannot be edited, reordered, or quietly removed without the chain failing verification.
What the chain gives you
- Tamper evidence. Verification covers the full chain, not a sample. A broken link is reported, not smoothed over.
- Fail-secure writes. If an event cannot be written, the request does not proceed. The system will decline an action before it will perform one it cannot record.
- Redaction on export. Sensitive input is masked when a record leaves the site, while the fingerprint still identifies the exact request.
Evidence is stored in your own database in versioned custom tables. There is no requirement to ship decision records to a third party for the record to be complete.
Audit history
The activity timeline is the operational view of the chain: agents, events, outcomes, dates, and sessions, written in human-readable summaries with the technical detail one click away. It answers “what did this agent do last Tuesday” without requiring a database query.
Retention is a decision, not a default
Retention is opt-in and bounded: 7 to 365 days, and 1,000 to 1,000,000 events. Extending retention is routine. Shortening it is gated behind explicit confirmation, because shortening retention destroys evidence, and that should never happen as a side effect of adjusting a setting.
What is kept
- Agent lifecycle
- Creation, activation, restriction, pause, revocation, and who performed each one.
- Request decisions
- Identity, Ability, classification, gate reached, outcome, and linked evidence.
- Approvals
- Request details, the administrator’s decision, and the binding fingerprint.
- Connections and sessions
- Health tests, rotation, suspension, revocation, and session termination.
Recovery
Recovery starts from the record. Because each outcome links to what actually changed, the first question after an unwanted action is answerable without guesswork: what ran, what it touched, and whether the effect is reversible.
Work in this order
- Stop the source. Pause or restrict the agent before undoing anything, so the same request cannot repeat while you work.
- Read the decision. Confirm the gate that allowed it. An action that should have paused points at a permission to change, not only an edit to revert.
- Reverse what is reversible. Use WordPress’ own recovery paths — revisions, trash, restores — against the specific objects the evidence names.
- Narrow the permission. Move the Ability from allow to approval, or to block, before returning the agent to service.
- Verify. Confirm both the content state and the decision record before closing the issue.
Classification tells you in advance which Abilities cannot be undone. Treat an irreversible, public-facing, bulk-capable Ability as an approval candidate from the start rather than a recovery problem later.
Emergency stop
Three site modes are always available, and they take effect immediately across every agent and every connection.
| Mode | Effect | Use when |
|---|---|---|
| Normal | Configured policy applies as written. | Routine operation. |
| Read‑only | Anything that writes is refused. Reads continue, so diagnostics still work. | You suspect a problem and need to investigate without further change. |
| Paused | All governed requests are refused. | An incident is in progress, or you are handing the site to someone else. |
Emergency mode overrides allow decisions; it never overrides a refusal. Returning to normal restores exactly the policy you had configured — the modes do not rewrite your permission matrix, and every mode change is recorded as an administrative event.
Cutting a connection
When the concern is a credential rather than a policy, terminate the connection instead. Sessions are ended connection-wide and credentials are invalidated before the connection state changes, so there is no window in which an in-flight session outlives the revocation.
Emergency controls are identical in every edition. Scale and optional features differ between plans; the ability to stop does not.
Where to go next
Keep the record, keep the control
Run agents you can account for.
Signed evidence, readable history, and an emergency stop that is always one click away.