You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/architecture/four-gb-vm-design.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -23,6 +23,6 @@ Five hundred logical agents are feasible because most are parked in SQLite as co
23
23
24
24
Book-writing and file-writing agents can run at higher active counts than coding agents because they do not all need browsers, sandboxes, test runners, or heavy model contexts. Expensive work should be scarce, leased, and scheduled.
25
25
26
-
Resource leases let SparseKernel answer which task owned which expensive resource and when it was released or expired.
26
+
Resource leases let SparseKernel answer which task owned which expensive resource and when it was released or expired. Trust-zone budgets are enforced at lease creation for sandbox work: `max_processes` caps active sandbox leases, and `max_runtime_seconds` clamps lease runtime and expiry. This keeps a 4 GB machine from materializing more heavy execution work than its configured trust zone allows.
27
27
28
28
Browser targets and observations are compact ledger rows, not retained screenshots or traces. This lets small machines keep enough browser provenance to answer which target made a request, emitted console output, or produced an artifact while still pruning old observations with `openclaw runtime prune`.
Copy file name to clipboardExpand all lines: docs/architecture/local-agent-kernel.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -108,7 +108,7 @@ The broker applies configured trust-zone network policy to explicit allowed orig
108
108
109
109
## Sandbox broker
110
110
111
-
The sandbox broker records allocations and leases behind a backend abstraction. Current v0 includes a local/no-isolation backend that can run trusted local commands behind an active sandbox lease for scheduling, timeout, output, usage, and audit accounting. Lease metadata persists the selected backend, isolation description, and trust-zone policy snapshot in SQLite so a restarted broker instance can recover the allocation backend without relying on an in-memory map. The broker can also build command spawn plans for requested bwrap and minijail backends when those binaries are available. Docker is now a policy-backed command backend: it requires a locally available `docker` CLI plus an explicit `dockerImage` or `OPENCLAW_SPARSEKERNEL_DOCKER_IMAGE`, uses `--pull never`, drops capabilities, sets `no-new-privileges`, applies read-only root/tmpfs defaults, maps trust-zone memory/process limits to Docker flags, and keeps networking disabled unless a policy proxy is configured. Set `OPENCLAW_RUNTIME_SANDBOX_REQUIRE_PROXY=1` to fail closed when a network-allowing trust zone lacks a valid loopback `network_policies.proxy_ref`; Docker command plans translate that host loopback proxy to `host.docker.internal` for container workers. This is proxy configuration and allocation gating, not a kernel firewall. SSH, OpenShell, and VM-backed execution remain future wrappers and are not silently executed on the host. In daemon broker mode, embedded runs grant sandbox allocation capability and allocate/release the `code_execution` sandbox lease through the SparseKernel daemon API before falling back to local accounting.
111
+
The sandbox broker records allocations and leases behind a backend abstraction. Current v0 includes a local/no-isolation backend that can run trusted local commands behind an active sandbox lease for scheduling, timeout, output, usage, and audit accounting. Lease metadata persists the selected backend, isolation description, and trust-zone policy snapshot in SQLite so a restarted broker instance can recover the allocation backend without relying on an in-memory map. The broker can also build command spawn plans for requested bwrap and minijail backends when those binaries are available. Docker is now a policy-backed command backend: it requires a locally available `docker` CLI plus an explicit `dockerImage` or `OPENCLAW_SPARSEKERNEL_DOCKER_IMAGE`, uses `--pull never`, drops capabilities, sets `no-new-privileges`, applies read-only root/tmpfs defaults, maps trust-zone memory/process limits to Docker flags, and keeps networking disabled unless a policy proxy is configured. Trust-zone `max_processes` now caps active sandbox leases, and `max_runtime_seconds` clamps resource lease runtime and expiry. Set `OPENCLAW_RUNTIME_SANDBOX_REQUIRE_PROXY=1` to fail closed when a network-allowing trust zone lacks a valid loopback `network_policies.proxy_ref`; Docker command plans translate that host loopback proxy to `host.docker.internal` for container workers. This is proxy configuration and allocation gating, not a kernel firewall. SSH, OpenShell, and VM-backed execution remain future wrappers and are not silently executed on the host. In daemon broker mode, embedded runs grant sandbox allocation capability and allocate/release the `code_execution` sandbox lease through the SparseKernel daemon API before falling back to local accounting.
112
112
113
113
Important boundary: `local/no_isolation` means accounting only. It does not provide process, filesystem, network, kernel, or VM isolation. Docker, bwrap, minijail, gVisor, or VM backends must be described by their actual guarantees when implemented.
0 commit comments