Probe Architecture v0
Scope
Defines the v0 architecture of the Probe runtime in AIOS, including control-plane boundaries, internal modules, session supervision loops, adapter roles, and bounded self-heal authority.Terminology
- Control plane: Gateway-owned policy, routing, desired-state, approval, and audit authority.
- Data plane: Probe-owned local execution and forwarding runtime.
- Worker Sandbox: isolated execution unit hosting coding tools, SSH/PTY runtime, mounts, and workspace dependencies.
- Adapter: protocol translation boundary for channel or agent integration.
Normative Spec
System Layering
AIOS v0 MUST use this layered model:- Gateway as the only control plane.
- Probe as the host-side data-plane runtime.
- Worker Sandbox as the execution plane for session and task runtime.
Gateway Responsibilities
Gateway MUST own:- DevBox and session specification
- scheduling and lease issuance
- authentication, authorization, approval, and sandbox policy decisions
- desired-state writes
- metadata-only audit and trace
- capability registry and global recovery/archive policy
Probe Responsibilities
Probe MUST own:- execution of authorized Gateway intent
- Worker Sandbox lifecycle on the local host
- maintenance of
ssh_remoteanddialogchannels - supervision of channel adapters and agent adapters
- health, heartbeat, event, and observed-phase reporting
- bounded local self-heal only
- global routing authority
- global approval policy
- desired-state authorship
- cross-host placement changes
Worker Sandbox Responsibilities
Worker Sandbox MUST contain the execution environment for:codexclaude_codeopencode- SSH/PTY runtime
- workspace filesystem, dependencies, tools, and local helper contracts
Internal Probe Modules
The Probe architecture SHOULD include these modules:- Host Runtime Manager
- Lease Guard
- Session Supervisor
- Workspace Bridge
- SSH Remote Broker
- Dialog Forwarder
- Adapter Host
- Health Reporter
- Snapshot Client
- Local Self-Heal Executor
- Event Emitter
- Capability Reporter
Channel Model
- Top-level channels MUST remain
ssh_remoteanddialogonly. ssh_remoteMUST supportide_primaryandterminal_fallback.dialogMUST use forward-only delivery with metadata-only persistence.
Supervision Loops
The architecture SHOULD expose these supervision loops:- Gateway reconcile loop.
- Gateway lease loop.
- Gateway scheduler loop.
- Probe session supervision loop.
- Probe channel supervision loop.
- Probe self-heal loop.
Self-heal Boundary
Probe local self-heal MAY perform only bounded local actions such as reconnect, restart, remount, rebind, and local endpoint refresh. Probe MUST NOT:- change host placement
- override policy
- bypass platform approval
- rewrite desired state
- silently terminate or archive runtime
Examples
Responsibility split
| Layer | Owns | Does not own |
|---|---|---|
| Gateway | desired state, routing, approval, audit | local runtime process control |
| Probe | local runtime execution, channel forwarding, heartbeat | global policy and scheduling authority |
| Worker Sandbox | coding tools and isolated workspace runtime | control-plane decisions |
Probe internal composition
Error & Recovery
- Lease expiry MUST immediately fence local state-changing operations.
- Adapter faults MUST remain locally isolated and MUST NOT collapse unrelated sessions.
- Repeated local recovery failure MUST escalate back to Gateway.
Security & Audit
- Gateway-to-Probe transport SHOULD be authenticated and traceable.
- Metadata-only audit MUST capture approval references, reasons, and recovery trails.
- DevBox-internal coding-tool commands remain outside AIOS secondary approval.
Status
Current status:draft
Source discussions:
docs/maintainers/conversation-records/agent-platform/2026-03-24-probe-agent-architecture-complete-proposal.mddocs/maintainers/conversation-records/agent-platform/2026-03-24-p0-freeze-result-accepted.mddocs/maintainers/conversation-records/agent-platform/2026-03-24-p1-freeze-result-accepted.md
Changelog
- 2026-03-24: Initial formal Probe architecture specification created from frozen proposal.