XERXES SI / ARCHITECTURE / ALPHA

How XMAN Governs Agents Through Its API

XMAN is a management API with separate authority lanes—not a root shell controlled by a model.

Reviewed against XMAN 0.4.76a1 source · November 1, 2026 · No independent certification claimed

EXISTING ALPHA CONTRACT

Who is authorized to request, decide and execute?

The design is not a chatbot with a Podman password. The trusted OpenClaw Gateway bridge identifies a registered requesting agent, whose token scopes access to that requester. All agents—including the coordinating master—request work directly from the management daemon. The master can advise but cannot silently grant operator authority.

Requester lane

Agent → trusted Gateway bridge → authenticated requester API → XMAN policy and durable ledger. Authenticated access is scoped to the registered requester.

Operator lane

Human operator → separate local operator API. Operator-only operations include sensitive reviews and approvals. A requester token cannot become an operator credential.

Execution lane

Validated policy → limited typed Podman adapter for eligible work, or reviewed foreground host execution for special.script. The agent does not receive the raw Podman socket.

IMPLEMENTED INTERFACE DESIGN

A private API boundary, not a public dashboard endpoint

XMAN API v3 uses local Unix sockets with one newline-terminated JSON request per connection. It does not expose a TCP listener, HTTP endpoint or CORS surface in this release. Both endpoints check the Linux peer UID; the requester socket additionally validates per-agent tokens. The trusted bridge itself is host software that must be qualified against the installed Gateway version.

Neither a master role nor model-generated text confers operator credentials. The daemon records request revisions and outcomes; events may be delivered more than once until acknowledged without replaying the associated operation.

API / POLICY

Two different routes for two different classes of work

Routine resource adjustments

A requester calls resource.prepare and then resource.commit for its own enrollment. The daemon resolves registered identity, evaluates policy, reserves resources and returns proposal and request receipts. Queued does not mean applied.

Sensitive host requests

Agents submit request.submit with operation special.script or another special type. Operator review remains mandatory; approvals are content-bound and sensitive host scripts execute in the operator foreground.

Recorded uncertainty

Receipt, acknowledgement, attempted execution and verified outcome are distinct. Timeouts or a failed notification must not be assumed to represent completed work.

WHAT IS NOT PROVEN

The boundary is real, and its limits are documented

Rootless Podman and a typed adapter reduce unintended host authority. They do not make a compromised host owner, root account or kernel harmless. The request ledger is not tamper-proof merely because it uses SHA-256, and local logs are not independent offsite evidence. XMAN’s own architecture document lists integration, reboot, restore, audit-signing and operational qualification work that remains outstanding.

Deployment qualification is required before promising that a particular host, Gateway version, plugin set or security posture is protected.

PROPOSED / NOT SHIPPED

Where future daemons fit

The proposed SysOp rootless daemon would handle a narrowly bounded set of system-operations and health-observation tasks. The future Security Officer rootless daemon would have a different policy, credentials, audit trail and scope for plugin preflight analysis and runtime security observations. Neither exists as a shipped v0.4.76a1 feature, and neither would inherit XMAN’s operator authority or gain its raw Podman engine socket.

Both would interact with XMAN through documented, permissioned interfaces when appropriate. Resource allocation should have one authoritative owner; a second daemon should not independently resize a container behind XMAN’s ledger.