Flow

Fleet, evidence, and operations

Signed inventory, desired state, typed control, exact exchange evidence, audit, monitoring, and bounded automation.

Inventory and health

The native management dashboard summarizes connected nodes, policy coverage, decision outcomes, pending approvals, evaluator health, command state, and recent activity. Fleet views support server-side filters and pagination for node status, group, tag, policy drift, connector capability, runtime version, and last report time.

Each node detail joins its signed identity, harness and connector versions, uptime, counters, connectivity, current and desired policy revisions, groups, tags, quarantine posture, recent actions, receipts, commands, state root, and acknowledgements. A stale or missing report is visible; the UI does not silently treat configured state as observed state.

Desired state and drift

Authority-signed desired state can set group, tags, desired runtime version, desired policy revision, and quarantine posture. The node verifies the command key and monotonic floors, persists before applying, reconciles locally, then returns a signed applied, partially_applied, or rejected acknowledgement. Drift remains visible until the reported state converges.

Runtime version is a desired target, not permission for the control plane to replace arbitrary executables. The deployment supervisor remains responsible for changing installed software.

Remote commands

Remote operations are typed, targeted, authority-signed, reason-bound, expiring, cancellable, and auditable. Supported commands are:

Pilot is not a remote shell. Operators cannot type arbitrary commands into the fleet console. Nodes return bounded signed result codes, and protected local authority cannot be widened through command parameters.

.pilot state visibility

An adopted node can report a signed manifest of its managed state directory: relative paths, file or directory kind, size, mode, modification time, content hash, revision, and root hash. Safe text files may expose sanitized or bounded content. Keys, seeds, credentials, identity, policy internals, trust material, and rollback floors are fingerprint-only and immutable through this channel.

Supported state changes use optimistic concurrency. A mutation names the expected tree revision and existing file hash, contains an auditable reason, expires, and is signed by the authority. The node independently validates safe paths and modes, applies the change locally, and reports the resulting signed revision. History records additions, changes, removals, and root hashes.

Exchanges and tracing

The Exchanges workspace stores tenant-scoped request and response evidence with server-side search and pagination. A detail view correlates:

Browser previews are bounded. Complete retained objects require explicit evidence permission and an audited download. Audit metadata and exact exchange content are separate stores with different access and retention concerns.

Audit and investigations

Management mutations write an append-only attempt event and a terminal result event. Records include actor, roles, authentication method, operation, reason, correlation ID, status, duration, and request/response hashes without copying secrets. Operators can filter, paginate, inspect, export NDJSON, monitor continuous delivery, and retry dead-letter exports.

Investigations have assignee, severity, status, resolution, timestamps, and links to action, exchange, receipt, command, and audit evidence. Legal hold and retention policy preserve required objects. Signed action receipts are verified at ingest and may be mirrored into compliance-retention object storage.

Procedures

Procedures are immutable, revisioned, post-action response policies. Triggers match action, node, risk, decision, or observed result. The closed response vocabulary can create an alert, open an investigation, quarantine a node, request policy refresh, request receipt export, or notify an active integration.

A run is durable and idempotent for the activity and procedure revision. Procedures cannot grant or retry the original action, inspect arbitrary content, execute shell commands, or invent remote methods.

External integrations

KindPurposeLifecycle
OIDC / KeycloakOptional workforce identity federation.Stage, inspect, probe, activate, rotate, disable.
OpenBao / KMSExternal signing-key custody.Reference-only secrets and signature verification.
S3-compatible Object LockImmutable receipt and evidence mirror.Retention and legal-hold evidence.
PostgreSQLReplica-safe shared authority state.TLS, migrations, backup, restore, and health.
Prometheus-compatible storageMetrics retention.Pilot's native UI remains the operator dashboard.
Usage sinkAsynchronous idempotent commercial usage delivery.Tenant-scoped credentials and delivery health.
Audit/SIEM and webhookManagement export, approval, and procedure notifications.Retry, dead letter, health, and rotation.

Connectors store secret references, not rendered secret values. Hosted semantic evaluation is Pilot platform infrastructure and is not configured as a customer-owned integration.

Object lifecycle

ObjectCreateChangeTerminal behavior
NodeOne-time adopted identity and signed report.Desired state, tags, groups, quarantine, commands.Retire or revoke; reactivate only where policy permits.
Policy statementInactive source revision.Compile, review, approve, deploy.Reject or retire; signed policy removal is a higher revision.
Policy changeDraft candidate.Diff, simulate, approve, publish, canary, promote.Withdraw, terminate, or monotonic rollback.
Approval transactionExact intent and payload binding.Vote, escalate, cancel.Execute once, reject, cancel, or expire.
ProcedureInactive draft revision.New revision, activate, disable.Retire while retaining run evidence.
IntegrationStaged revision.Probe, activate, rotate, disable.Retain revision and audit history.
High-volume behavior. Fleet, exchanges, approvals, receipts, audit, procedures, and activity lists use tenant-scoped server-side filters and pagination. Production sizing still requires load, retention, backup, and failover evidence for the intended fleet and event rate.