feat(kubernetes): add sidecar supervisor topology (#2076)

* feat(kubernetes): add sidecar supervisor topology

Add the Kubernetes sidecar supervisor topology, its Helm/Skaffold configuration, topology documentation, and sidecar e2e matrix coverage. Skip root-only sandbox identity rewriting when process enforcement is network-only so the low-permission sidecar process container can start successfully.

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(supervisor): avoid similar process id names

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(supervisor): avoid similar process id names

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(sandbox): avoid similar proxy id names

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* docs(kubernetes): clarify sidecar topology limits

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): keep sidecar process leaf capless

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): refresh sidecar provider env snapshots

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* test(supervisor): align hot-swap identity regression

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): stage sidecar mtls files before proxy chown

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): simplify sidecar supervisor topology

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* chore(helm): reuse sidecar skaffold values

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(supervisor): avoid similar iptables helper names

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(e2e): harden kube gateway wrapper setup

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(supervisor): avoid nft batch rollback on OCP

Run nftables setup as individual commands so optional conntrack and log expressions can fail without rolling back required table, chain, and reject rules.

Signed-off-by: Seth Jennings <sjenning@redhat.com>
Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): preserve process identity in sidecar topology

Render sidecar pods with a shared process namespace, keep binary-aware network policy enabled, and move Kubernetes sidecar settings under the nested sidecar config table.

Also apply unprivileged Landlock/seccomp setup in NetworkOnly supervisor mode so sidecar topology keeps sandbox child hardening without privileged process setup.

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* refactor(kubernetes): replace sidecar snapshots with control socket

Coordinate sidecar policy and provider bootstrap over a local Unix socket so the process leaf no longer reads policy/provider snapshot files.

Report entrypoint startup through the control channel and keep gateway credentials confined to the network sidecar.

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* feat(kubernetes): support relaxed sidecar network identity

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(sandbox): satisfy sidecar clippy lint

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* refactor(kubernetes): standardize topology naming

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(sandbox): satisfy linux clippy timeout import

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): support kata sidecar on ipv4 pods

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): satisfy linux clippy for sidecar fallback

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* chore(kubernetes): remove stale supervisor topology references

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): enable sidecar binary policy inspection

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): harden sidecar control boundary

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

* fix(kubernetes): couple sidecar supervisor lifecycles

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>

---------

Signed-off-by: Taylor Mutch <taylormutch@gmail.com>
Signed-off-by: Seth Jennings <sjenning@redhat.com>
Co-authored-by: Seth Jennings <sjenning@redhat.com>

.agents/skills/debug-openshell-cluster/SKILL.md

@@ -268,6 +268,57 @@ kubectl -n openshell get configmap openshell-config -o jsonpath='{.data.gateway\
 kubectl -n <sandbox-namespace> get sandbox <sandbox-name> -o jsonpath='{.spec.template.spec.serviceAccountName}{"\n"}'
 ```
 
+If `topology = "sidecar"` is rendered under `[openshell.drivers.kubernetes]`,
+sandbox pods should have an `openshell-network-init` init container running
+`--mode=network-init`, an `agent` container running
+`openshell-sandbox --mode=process`, and an `openshell-supervisor-network`
+container running `--mode=network`. The init container owns nftables setup and
+should be the only sidecar topology container with `NET_ADMIN`. It also needs
+`CHOWN`/`FOWNER` to hand shared emptyDir state to the effective sidecar UID. The
+default binary-aware network sidecar runs as UID 0 with primary GID
+`sandbox_gid` and adds `SYS_PTRACE` plus `DAC_READ_SEARCH`. When
+`process_binary_aware_network_policy = false`, it runs as the configured
+non-root `proxy_uid` without those inspection capabilities. The pod `fsGroup`
+is set to `sandbox_gid` in both modes.
+
+In sidecar topology only the network sidecar should mount the gateway bootstrap
+credentials (`openshell-sa-token` and `openshell-client-tls`). The process
+container should not receive `OPENSHELL_ENDPOINT`, gateway TLS env vars, the
+sandbox token file, or those credential mounts. Instead, the network sidecar
+serves policy and provider environment state over the Unix control socket from
+`OPENSHELL_SIDECAR_CONTROL_SOCKET` (`/run/openshell-sidecar/control.sock` by
+default). The process supervisor must be the first and only client. After
+validating its peer UID, GID, and PID, the sidecar unlinks the listener. If the
+connection later closes, the network sidecar exits non-zero so Kubernetes can
+restart it with a fresh listener. If the process supervisor fails before
+launching the workload,
+inspect both containers for control-socket bind, connect, bootstrap, or update
+errors. If new SSH/exec sessions do not pick up refreshed provider environment,
+inspect the network sidecar settings-poll logs and the process container logs
+for provider environment update handling; the process container should consume
+newer provider-env revisions without receiving gateway credentials.
+
+The process container reports the workload entrypoint PID over the same control
+socket, and the network sidecar uses that PID for binary-scoped policy
+decisions through `/proc`. If rules with `policy.binaries` are unexpectedly
+denied, inspect the sidecar control logs and confirm the pod has
+`shareProcessNamespace: true`.
+The shared state directory should preserve `sandbox_gid` inheritance
+(`02775`). Sidecar SSH uses the Linux abstract socket
+`@openshell-sidecar-ssh`; the network sidecar verifies its peer PID before
+bridging gateway relay requests. No `ssh.sock` file should appear in the shared
+state directory.
+Inspect all three when sandbox registration or egress enforcement fails:
+
+```bash
+kubectl -n openshell get configmap openshell-config -o jsonpath='{.data.gateway\.toml}' | grep -E '^\[openshell\.drivers\.kubernetes\]|^topology\s*='
+kubectl -n <sandbox-namespace> get pod <sandbox-pod> -o jsonpath='{range .spec.initContainers[*]}{.name}{" "}{.command}{"\n"}{end}'
+kubectl -n <sandbox-namespace> get pod <sandbox-pod> -o jsonpath='{range .spec.containers[*]}{.name}{" "}{.command}{"\n"}{end}'
+kubectl -n <sandbox-namespace> logs <sandbox-pod> -c openshell-network-init --tail=200
+kubectl -n <sandbox-namespace> logs <sandbox-pod> -c openshell-supervisor-network --tail=200
+kubectl -n <sandbox-namespace> logs <sandbox-pod> -c agent --tail=200
+```
+
 ### Step 6: Check VM-Backed Gateways
 
 Use the VM driver logs and host diagnostics available in the user's environment. Verify:

.agents/skills/helm-dev-environment/SKILL.md

@@ -60,9 +60,25 @@ mise run helm:skaffold:dev
 mise run helm:skaffold:run
 ```
 
+**Supervisor sidecar topology** (build once and leave running):
+```bash
+mise run helm:skaffold:run:sidecar
+```
+
+**Supervisor sidecar topology with TLS/mTLS enabled** (build once and leave running):
+```bash
+mise run helm:skaffold:run:sidecar-mtls
+```
+
 Both commands build the `gateway` and `supervisor` images and deploy the OpenShell Helm
-chart. The `pkiInitJob` hook (a pre-install Job that runs `openshell-gateway generate-certs`)
-generates mTLS secrets on first install. Envoy Gateway opt-in; see the Optional Add-ons section below.
+chart. The sidecar profile renders an `openshell-network-init` init container for
+nftables setup and an `openshell-supervisor-network` runtime sidecar for proxying.
+Binary-aware policy mode runs that sidecar as UID 0 with `SYS_PTRACE` and
+`DAC_READ_SEARCH`; relaxed mode can run it as the configured proxy UID. The
+sidecar-mTLS profile reuses `ci/values-sidecar.yaml` and restores
+`server.disableTls=false` inline for Skaffold. The `pkiInitJob` hook (a pre-install
+Job that runs `openshell-gateway generate-certs`) generates mTLS secrets on first
+install. Envoy Gateway opt-in; see the Optional Add-ons section below.
 
 The gateway Service uses ClusterIP. Access is via Envoy Gateway (port `8080`) or `kubectl port-forward`.
 
@@ -74,7 +90,8 @@ create the Secret named `openshell-ha-pg` with a `uri` key, then run
 ### TLS behaviour
 
 `ci/values-skaffold.yaml` sets `server.disableTls: true`, so Skaffold-based deploys run
-plaintext by default. To test with TLS enabled, comment out that line and redeploy.
+plaintext by default. To test sidecar topology with TLS enabled, use
+`mise run helm:skaffold:run:sidecar-mtls`.
 
 | Mode | `server.disableTls` | Gateway scheme |
 |------|---------------------|----------------|
@@ -126,6 +143,12 @@ openshell sandbox list --gateway-endpoint https://localhost:8090
 mise run helm:skaffold:delete
 ```
 
+For a sidecar-profile deployment:
+
+```bash
+mise run helm:skaffold:delete:sidecar
+```
+
 ### Delete the cluster entirely
 
 ```bash
@@ -250,6 +273,7 @@ for dependencies still declared in `Chart.yaml`.
 | `deploy/helm/openshell/ci/values-gateway.yaml` | Envoy Gateway GRPCRoute + Gateway overlay |
 | `deploy/helm/openshell/ci/values-high-availability.yaml` | HA test overlay (`replicaCount: 2` with external PostgreSQL Secret) |
 | `deploy/helm/openshell/ci/values-keycloak.yaml` | Keycloak OIDC overlay |
+| `deploy/helm/openshell/ci/values-sidecar.yaml` | Supervisor sidecar topology overlay for Kubernetes e2e/dev |
 | `deploy/helm/openshell/ci/values-spire.yaml` | SPIFFE/SPIRE provider token grant overlay |
 | `deploy/helm/openshell/ci/values-spire-stack.yaml` | SPIRE hardened chart values for local dev |
 | `deploy/helm/openshell/ci/values-tls-disabled.yaml` | Lint-only: TLS + auth disabled (reverse-proxy edge termination) |

.github/workflows/branch-e2e.yml

@@ -121,16 +121,25 @@ jobs:
         include:
           - agent_sandbox_api: v1beta1
             agent_sandbox_version: v0.5.0
+            topology: combined
+            extra_helm_values: ""
           - agent_sandbox_api: v1alpha1
             agent_sandbox_version: v0.4.6
+            topology: combined
+            extra_helm_values: ""
+          - agent_sandbox_api: v1beta1
+            agent_sandbox_version: v0.5.0
+            topology: sidecar
+            extra_helm_values: deploy/helm/openshell/ci/values-sidecar.yaml
     permissions:
       contents: read
       packages: read
     uses: ./.github/workflows/e2e-kubernetes-test.yml
     with:
       image-tag: ${{ github.sha }}
-      job-name: Kubernetes E2E (Rust smoke, Agent Sandbox ${{ matrix.agent_sandbox_api }})
+      job-name: Kubernetes E2E (Rust smoke, ${{ matrix.topology }}, Agent Sandbox ${{ matrix.agent_sandbox_api }})
       agent-sandbox-version: ${{ matrix.agent_sandbox_version }}
+      extra-helm-values: ${{ matrix.extra_helm_values }}
 
   kubernetes-ha-e2e:
     needs: [pr_metadata, build-gateway, build-supervisor]

Cargo.lock

@@ -3827,12 +3827,15 @@ dependencies = [
  "clap",
  "futures",
  "miette",
+ "nix",
  "openshell-core",
  "openshell-ocsf",
  "openshell-policy",
  "openshell-supervisor-network",
  "openshell-supervisor-process",
+ "prost",
  "rustls",
+ "serde",
  "serde_json",
  "temp-env",
  "tempfile",

architecture/build.md

@@ -91,10 +91,11 @@ Runtime layout:
   as a release artifact. Linux GNU VM driver binaries must not reference
   `GLIBC_*` symbols newer than `GLIBC_2.28`; release workflows verify this
   before publishing artifacts.
-- **Supervisor**: `scratch` base, static musl binary at `/openshell-sandbox`.
-  Static linkage is required because the image is mounted/extracted into
-  sandbox environments (Docker extraction, Podman image volumes, Kubernetes
-  init-container copy-self) and cannot rely on a dynamic loader.
+- **Supervisor**: Alpine base with `nftables`, static musl binary at
+  `/openshell-sandbox`. Static linkage keeps the binary usable when the image
+  is mounted/extracted into sandbox environments (Docker extraction, Podman
+  image volumes, Kubernetes init-container copy-self), while `nftables` supports
+  Kubernetes supervisor sidecar egress enforcement.
 
 Gateway image builds bake the corresponding supervisor image tag into the
 gateway binary so Docker sandboxes do not depend on `:latest` by default.

architecture/compute-runtimes.md

@@ -81,7 +81,7 @@ The supervisor must be available inside each sandbox workload:
 |---|---|
 | Docker | Bind-mounted local supervisor binary, or a binary extracted from the configured supervisor image. |
 | Podman | Read-only OCI image volume containing the supervisor binary. |
-| Kubernetes | Sandbox pod image or pod template configuration. |
+| Kubernetes | Supervisor image side-loaded into the sandbox pod by image volume or init container. |
 | VM | Embedded in the guest rootfs bundle. |
 | Extension | Defined by the out-of-tree driver. |
 
@@ -89,6 +89,44 @@ Driver-controlled environment variables must override sandbox image or template
 values for sandbox ID, sandbox name, gateway endpoint, relay socket path, TLS
 paths, and command metadata.
 
+Kubernetes can run the supervisor in the default combined topology or in a
+sidecar topology. Combined mode keeps network and process supervision in the
+agent container. Sidecar mode runs network enforcement, the proxy, and gateway
+session in a dedicated sidecar, while the agent container runs only the
+process-supervision leaf and launches the user workload after the sidecar
+serves bootstrap state over a local control socket. The network sidecar owns
+gateway credentials and sends policy plus workload-facing provider environment
+state to the process leaf over that socket. It also streams provider
+environment updates after settings polls so future process sessions see
+updated provider env without giving the process leaf gateway access. The
+pre-workload process supervisor is the only accepted control client: the
+network sidecar verifies its UID, GID, and PID with peer credentials, removes
+the listener after accepting it, and ignores workload-supplied relay targets.
+SSH relays use a Linux abstract socket and verify its peer PID against that
+authenticated process-supervisor connection, so workload filesystem access
+cannot replace the relay endpoint. Either supervisor exits when this control
+connection closes. This couples their restart lifecycle and prevents a workload
+that survives an isolated network-sidecar restart from becoming the next
+authoritative control client. In sidecar mode, an init container performs the
+privileged pod-network nftables setup with
+`NET_ADMIN`. The default binary-aware network sidecar runs as UID 0 without
+`NET_ADMIN` and adds `SYS_PTRACE` plus `DAC_READ_SEARCH` so it can resolve
+cross-UID workload process/binary identity through shared `/proc`. Operators
+can set the sidecar `process_binary_aware_network_policy` flag false to run the
+sidecar as the configured non-root proxy UID, omit both inspection capabilities,
+and downgrade network policy to endpoint/L7 matching without `policy.binaries`.
+The init path applies nftables as individual commands so optional conntrack and
+log expressions can fail without rolling back the required table, chain, and
+reject rules.
+The agent container runs as the resolved sandbox UID/GID with no added Linux
+capabilities. Sidecar mode preserves gateway session and SSH behavior, but
+treats the process leaf as network-only: Landlock filesystem policy and child
+seccomp still apply where supported, while process privilege dropping and
+supervisor identity mount isolation do not run because the agent container is
+already unprivileged. Sidecar pods use a shared process namespace so the
+network sidecar can resolve workload process and binary identity through
+`/proc/<entrypoint-pid>`.
+
 ## Images
 
 The gateway image and Helm chart are built from this repository. Sandbox images

crates/openshell-core/src/grpc_client.rs

@@ -167,9 +167,14 @@ async fn build_plain_channel(endpoint: &str) -> Result<Channel> {
             .into_diagnostic()
             .wrap_err_with(|| format!("failed to read client key from {key_path}"))?;
 
-        let tls_config = ClientTlsConfig::new()
+        let mut tls_config = ClientTlsConfig::new()
             .ca_certificate(Certificate::from_pem(ca_pem))
             .identity(Identity::from_pem(cert_pem, key_pem));
+        if let Ok(server_name) = std::env::var(sandbox_env::GATEWAY_TLS_SERVER_NAME)
+            && !server_name.is_empty()
+        {
+            tls_config = tls_config.domain_name(server_name);
+        }
 
         ep = ep
             .tls_config(tls_config)

crates/openshell-core/src/provider_credentials.rs

@@ -63,6 +63,63 @@ impl ProviderCredentialState {
         }
     }
 
+    /// Build a static provider state from an already-prepared child
+    /// environment snapshot.
+    ///
+    /// Kubernetes sidecar topology uses this in the process-only supervisor:
+    /// the network sidecar owns provider credential resolvers and sends the
+    /// workload-facing env map over a local control channel. The process leaf
+    /// must inject that map into child processes without re-placeholderizing it
+    /// or holding the gateway-side resolver material.
+    pub fn from_child_env_snapshot(revision: u64, child_env: HashMap<String, String>) -> Self {
+        let snapshot = Arc::new(ProviderCredentialSnapshot {
+            revision,
+            child_env,
+            dynamic_credentials: HashMap::new(),
+        });
+
+        Self {
+            inner: Arc::new(RwLock::new(ProviderCredentialStateInner {
+                current: snapshot,
+                generations: VecDeque::new(),
+                current_resolver: None,
+                combined_resolver: None,
+                suppressed_keys: HashSet::new(),
+            })),
+        }
+    }
+
+    /// Install an already-prepared child environment snapshot.
+    ///
+    /// This is intentionally narrower than [`Self::install_environment`]: it
+    /// updates only the workload-facing env map and clears resolver state so a
+    /// process that does not own gateway/provider resolver material can still
+    /// pick up refreshed provider env for future child processes.
+    pub fn install_child_env_snapshot(
+        &self,
+        revision: u64,
+        mut child_env: HashMap<String, String>,
+    ) -> usize {
+        let mut inner = self
+            .inner
+            .write()
+            .expect("provider credential state poisoned");
+
+        for key in &inner.suppressed_keys {
+            child_env.remove(key);
+        }
+
+        inner.current = Arc::new(ProviderCredentialSnapshot {
+            revision,
+            child_env,
+            dynamic_credentials: HashMap::new(),
+        });
+        inner.generations.clear();
+        inner.current_resolver = None;
+        inner.combined_resolver = None;
+        inner.current.child_env.len()
+    }
+
     pub fn snapshot(&self) -> Arc<ProviderCredentialSnapshot> {
         self.inner
             .read()
@@ -594,6 +651,39 @@ mod tests {
         );
     }
 
+    #[test]
+    fn child_env_snapshot_install_updates_env_without_resolver_material() {
+        let state = ProviderCredentialState::from_child_env_snapshot(
+            1,
+            HashMap::from([
+                ("GITHUB_TOKEN".to_string(), "old".to_string()),
+                ("GCE_METADATA_HOST".to_string(), "marker".to_string()),
+            ]),
+        );
+        state.remove_env_key("GCE_METADATA_HOST");
+
+        let env_count = state.install_child_env_snapshot(
+            2,
+            HashMap::from([
+                ("GITHUB_TOKEN".to_string(), "new".to_string()),
+                ("GCE_METADATA_HOST".to_string(), "marker".to_string()),
+            ]),
+        );
+
+        let snapshot = state.snapshot();
+        assert_eq!(snapshot.revision, 2);
+        assert_eq!(env_count, 1);
+        assert_eq!(
+            snapshot.child_env.get("GITHUB_TOKEN").map(String::as_str),
+            Some("new")
+        );
+        assert!(!snapshot.child_env.contains_key("GCE_METADATA_HOST"));
+        assert!(
+            state.resolver().is_none(),
+            "child-env snapshots must not install provider resolver material"
+        );
+    }
+
     #[test]
     fn stale_generation_falls_back_to_current_credential_after_retention_window() {
         let state = ProviderCredentialState::from_environment(

crates/openshell-core/src/sandbox_env.rs

@@ -29,6 +29,34 @@ pub const SANDBOX_COMMAND: &str = "OPENSHELL_SANDBOX_COMMAND";
 /// Deployment-controlled telemetry toggle propagated to the sandbox supervisor.
 pub const TELEMETRY_ENABLED: &str = "OPENSHELL_TELEMETRY_ENABLED";
 
+/// Supervisor pod/runtime topology. Kubernetes sidecar mode sets this to
+/// `"sidecar"`; the default combined supervisor path omits it.
+pub const SUPERVISOR_TOPOLOGY: &str = "OPENSHELL_SUPERVISOR_TOPOLOGY";
+
+/// Network enforcement backend selected by the compute driver.
+pub const NETWORK_ENFORCEMENT_MODE: &str = "OPENSHELL_NETWORK_ENFORCEMENT_MODE";
+
+/// Whether network policy evaluation must bind requests to the peer binary.
+///
+/// The default when unset is `"required"`. Kubernetes sidecar experiments may
+/// set this to `"relaxed"` to enforce endpoint and L7 policy without per-binary
+/// `/proc` identity binding.
+pub const NETWORK_BINARY_IDENTITY: &str = "OPENSHELL_NETWORK_BINARY_IDENTITY";
+
+/// Unix socket used by Kubernetes sidecar topology for local coordination.
+///
+/// The network sidecar owns gateway credentials and serves policy/provider
+/// state over this socket instead of exposing gateway credentials to the agent
+/// container.
+pub const SIDECAR_CONTROL_SOCKET: &str = "OPENSHELL_SIDECAR_CONTROL_SOCKET";
+
+/// Optional TLS server name override used when connecting to the gateway.
+pub const GATEWAY_TLS_SERVER_NAME: &str = "OPENSHELL_GATEWAY_TLS_SERVER_NAME";
+
+/// Directory where the network supervisor writes the proxy CA files consumed
+/// by workload child processes.
+pub const PROXY_TLS_DIR: &str = "OPENSHELL_PROXY_TLS_DIR";
+
 /// Path to the CA certificate for mTLS communication with the gateway.
 pub const TLS_CA: &str = "OPENSHELL_TLS_CA";
 

crates/openshell-driver-kubernetes/README.md

@@ -53,9 +53,43 @@ pods do not need direct external ingress for SSH.
 
 ## Container Security Context
 
-The driver grants the sandbox agent container the Linux capabilities the
-supervisor needs for namespace setup and policy enforcement. It can also request
-a Kubernetes AppArmor profile through `app_armor_profile`.
+The default `combined` supervisor topology grants the sandbox agent container
+the Linux capabilities the supervisor needs for namespace setup and process,
+filesystem, and network policy enforcement.
+
+The `sidecar` supervisor topology moves pod-level network setup into a root init
+container. In the default process/binary-aware mode, the long-lived network
+sidecar runs as UID 0 with `allowPrivilegeEscalation: false`, drops default
+Linux capabilities, and adds only `SYS_PTRACE` plus `DAC_READ_SEARCH` for
+cross-UID workload `/proc` inspection. The agent container also runs as the
+resolved sandbox UID/GID with `allowPrivilegeEscalation: false` and
+`capabilities.drop: ["ALL"]`.
+Set `sidecar.process_binary_aware_network_policy = false` to run the network
+sidecar as the configured non-root `sidecar.proxy_uid`, omit the extra `/proc`
+inspection capabilities, and enforce endpoint/L7 network policy without
+matching `policy.binaries`.
+In this mode OpenShell preserves gateway session and SSH behavior, but the
+process supervisor does not perform root-to-sandbox privilege dropping or
+supervisor identity mount isolation. It still applies Landlock filesystem policy
+and child seccomp filters where the kernel/runtime supports them. Network
+endpoint and L7 policy remain enforced by the network sidecar, and
+sidecar pods use a shared process namespace so the network sidecar can resolve
+process/binary identity through `/proc/<entrypoint-pid>`.
+
+Sidecar mode keeps gateway credentials in the network sidecar. The agent
+container does not mount the projected service-account token used for sandbox
+token bootstrap, does not mount the sandbox client TLS secret, and does not get
+gateway callback environment variables. The process supervisor receives policy
+and provider environment state from the sidecar over a local control socket in
+the shared sidecar state volume. The sidecar accepts only the pre-workload
+process-supervisor connection, authenticates its UID/GID/PID with peer
+credentials, and removes the listener afterward. SSH relays use a Linux
+abstract socket whose peer PID must match that authenticated supervisor. Both
+supervisors exit if the control connection closes, coupling their container
+restart lifecycle before a new authoritative client can be established.
+
+The driver can request a Kubernetes AppArmor profile through
+`app_armor_profile`.
 
 Supported values are `Unconfined`, `RuntimeDefault`, and
 `Localhost/<profile-name>`. An empty or unset value omits

crates/openshell-driver-kubernetes/src/config.rs

@@ -15,6 +15,9 @@ pub const DEFAULT_SANDBOX_SERVICE_ACCOUNT_NAME: &str = "default";
 /// Default storage size for the workspace PVC.
 pub const DEFAULT_WORKSPACE_STORAGE_SIZE: &str = "2Gi";
 
+/// Default non-root UID for relaxed Kubernetes network supervisor sidecars.
+pub const DEFAULT_PROXY_UID: u32 = 1337;
+
 /// How the supervisor binary is delivered into sandbox pods.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
 #[serde(rename_all = "kebab-case")]
@@ -59,12 +62,16 @@ pub enum SupervisorTopology {
     /// Run networking and process supervision in the agent container.
     #[default]
     Combined,
+    /// Run network supervision in a privileged sidecar and process supervision
+    /// as a low-capability wrapper in the agent container.
+    Sidecar,
 }
 
 impl std::fmt::Display for SupervisorTopology {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         match self {
             Self::Combined => f.write_str("combined"),
+            Self::Sidecar => f.write_str("sidecar"),
         }
     }
 }
@@ -75,11 +82,49 @@ impl FromStr for SupervisorTopology {
     fn from_str(s: &str) -> Result<Self, Self::Err> {
         match s {
             "combined" => Ok(Self::Combined),
-            other => Err(format!("unknown supervisor topology '{other}'")),
+            "sidecar" => Ok(Self::Sidecar),
+            other => Err(format!("unknown topology '{other}'")),
         }
     }
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+pub struct KubernetesSidecarConfig {
+    /// UID used by relaxed long-running network sidecars in `sidecar`
+    /// topology. The network init container installs nftables rules that
+    /// exempt this UID, so it must not match the sandbox workload UID.
+    /// Strict process/binary-aware sidecars run as UID 0 so Kubernetes grants
+    /// the requested `/proc` inspection capabilities into the effective set.
+    pub proxy_uid: u32,
+    /// Require process/binary-aware network policy enforcement in sidecar
+    /// topology. When disabled, the network sidecar runs as `proxy_uid`,
+    /// drops the extra `/proc` inspection permissions, and evaluates
+    /// endpoint/L7 policy without matching `policy.binaries`.
+    pub process_binary_aware_network_policy: bool,
+}
+
+impl Default for KubernetesSidecarConfig {
+    fn default() -> Self {
+        Self {
+            proxy_uid: DEFAULT_PROXY_UID,
+            process_binary_aware_network_policy: true,
+        }
+    }
+}
+
+impl KubernetesSidecarConfig {
+    pub fn validate_proxy_uid(&self) -> Result<(), String> {
+        if self.proxy_uid < openshell_policy::MIN_SANDBOX_UID {
+            return Err(format!(
+                "sidecar.proxy_uid must be at least {}",
+                openshell_policy::MIN_SANDBOX_UID
+            ));
+        }
+        Ok(())
+    }
+}
+
 /// Kubernetes `AppArmor` profile requested for the sandbox agent container.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub enum AppArmorProfile {
@@ -205,7 +250,9 @@ pub struct KubernetesComputeConfig {
     /// How the supervisor binary is delivered into sandbox pods.
     pub supervisor_sideload_method: SupervisorSideloadMethod,
     /// How the supervisor is arranged for Kubernetes sandbox pods.
-    pub supervisor_topology: SupervisorTopology,
+    pub topology: SupervisorTopology,
+    /// Sidecar-only settings used when `topology = "sidecar"`.
+    pub sidecar: KubernetesSidecarConfig,
     pub grpc_endpoint: String,
     pub ssh_socket_path: String,
     pub client_tls_secret_name: String,
@@ -291,7 +338,8 @@ impl Default for KubernetesComputeConfig {
             supervisor_image: config::default_supervisor_image(),
             supervisor_image_pull_policy: String::new(),
             supervisor_sideload_method: SupervisorSideloadMethod::default(),
-            supervisor_topology: SupervisorTopology::default(),
+            topology: SupervisorTopology::default(),
+            sidecar: KubernetesSidecarConfig::default(),
             grpc_endpoint: String::new(),
             ssh_socket_path: "/run/openshell/ssh.sock".to_string(),
             client_tls_secret_name: String::new(),
@@ -336,6 +384,10 @@ impl KubernetesComputeConfig {
         )
     }
 
+    pub fn validate_proxy_uid(&self) -> Result<(), String> {
+        self.sidecar.validate_proxy_uid()
+    }
+
     /// Resolve the sandbox UID/GID pair.
     ///
     /// Resolution order:
@@ -351,6 +403,7 @@ impl KubernetesComputeConfig {
         if let Some(uid) = self.sandbox_uid {
             return uid;
         }
+        // Try OpenShift SCC annotation.
         if let Some(anns) = namespace_annotations
             && let Some(range) = anns.get(ANNOTATION_SCC_UID_RANGE)
             && let Some(uid) = Self::from_open_shift_uid_range(range)
@@ -461,6 +514,132 @@ mod tests {
         );
     }
 
+    #[test]
+    fn default_topology_is_combined() {
+        let cfg = KubernetesComputeConfig::default();
+        assert_eq!(cfg.topology, SupervisorTopology::Combined);
+        assert_eq!(cfg.topology.to_string(), "combined");
+    }
+
+    #[test]
+    fn default_proxy_uid_is_dedicated_non_root_uid() {
+        let cfg = KubernetesComputeConfig::default();
+        assert_eq!(cfg.sidecar.proxy_uid, DEFAULT_PROXY_UID);
+    }
+
+    #[test]
+    fn default_sidecar_requires_process_binary_aware_network_policy() {
+        let cfg = KubernetesComputeConfig::default();
+        assert!(cfg.sidecar.process_binary_aware_network_policy);
+    }
+
+    #[test]
+    fn serde_override_topology_sidecar() {
+        let json = serde_json::json!({
+            "topology": "sidecar"
+        });
+        let cfg: KubernetesComputeConfig = serde_json::from_value(json).unwrap();
+        assert_eq!(cfg.topology, SupervisorTopology::Sidecar);
+    }
+
+    #[test]
+    fn serde_override_topology_combined() {
+        let json = serde_json::json!({
+            "topology": "combined"
+        });
+        let cfg: KubernetesComputeConfig = serde_json::from_value(json).unwrap();
+        assert_eq!(cfg.topology, SupervisorTopology::Combined);
+    }
+
+    #[test]
+    fn serde_rejects_sidecar_binary_identity_field() {
+        let json = serde_json::json!({
+            "sidecar": {
+                "binary_identity": "shared-pid"
+            }
+        });
+        let err = serde_json::from_value::<KubernetesComputeConfig>(json).unwrap_err();
+        assert!(err.to_string().contains("unknown field"));
+    }
+
+    #[test]
+    fn serde_override_sidecar_process_binary_aware_network_policy_nested() {
+        let json = serde_json::json!({
+            "sidecar": {
+                "process_binary_aware_network_policy": false
+            }
+        });
+        let cfg: KubernetesComputeConfig = serde_json::from_value(json).unwrap();
+        assert!(!cfg.sidecar.process_binary_aware_network_policy);
+    }
+
+    #[test]
+    fn serde_override_sidecar_proxy_uid_nested() {
+        let json = serde_json::json!({
+            "sidecar": {
+                "proxy_uid": 2000
+            }
+        });
+        let cfg: KubernetesComputeConfig = serde_json::from_value(json).unwrap();
+        assert_eq!(cfg.sidecar.proxy_uid, 2000);
+        cfg.validate_proxy_uid().unwrap();
+    }
+
+    #[test]
+    fn validate_proxy_uid_rejects_privileged_uid() {
+        let cfg = KubernetesComputeConfig {
+            sidecar: KubernetesSidecarConfig {
+                proxy_uid: 999,
+                ..KubernetesSidecarConfig::default()
+            },
+            ..KubernetesComputeConfig::default()
+        };
+        let err = cfg.validate_proxy_uid().unwrap_err();
+        assert!(err.contains("proxy_uid"));
+    }
+
+    #[test]
+    fn serde_rejects_invalid_topology() {
+        let json = serde_json::json!({
+            "topology": "unsupported"
+        });
+        let err = serde_json::from_value::<KubernetesComputeConfig>(json).unwrap_err();
+        assert!(err.to_string().contains("unknown variant"));
+    }
+
+    #[test]
+    fn serde_rejects_removed_topology_alias_field() {
+        let mut json = serde_json::Map::new();
+        json.insert(
+            ["supervisor", "topology"].join("_"),
+            serde_json::json!("sidecar"),
+        );
+        let err =
+            serde_json::from_value::<KubernetesComputeConfig>(serde_json::Value::Object(json))
+                .unwrap_err();
+        assert!(err.to_string().contains("unknown field"));
+    }
+
+    #[test]
+    fn serde_rejects_removed_flat_sidecar_fields() {
+        for json in [
+            serde_json::json!({ "sidecar_binary_identity": "shared-pid" }),
+            serde_json::json!({ "proxy_uid": 2000 }),
+        ] {
+            let err = serde_json::from_value::<KubernetesComputeConfig>(json).unwrap_err();
+            assert!(err.to_string().contains("unknown field"));
+        }
+    }
+
+    #[test]
+    fn serde_rejects_removed_process_enforcement_field() {
+        let json = serde_json::json!({
+            "process_enforcement": "network-only"
+        });
+        let err = serde_json::from_value::<KubernetesComputeConfig>(json).unwrap_err();
+        assert!(err.to_string().contains("unknown field"));
+    }
+
     #[test]
     fn default_service_account_name_is_default() {
         let cfg = KubernetesComputeConfig::default();
@@ -470,31 +649,6 @@ mod tests {
         );
     }
 
-    #[test]
-    fn default_supervisor_topology_is_combined() {
-        let cfg = KubernetesComputeConfig::default();
-        assert_eq!(cfg.supervisor_topology, SupervisorTopology::Combined);
-        assert_eq!(cfg.supervisor_topology.to_string(), "combined");
-    }
-
-    #[test]
-    fn serde_override_supervisor_topology_combined() {
-        let json = serde_json::json!({
-            "supervisor_topology": "combined"
-        });
-        let cfg: KubernetesComputeConfig = serde_json::from_value(json).unwrap();
-        assert_eq!(cfg.supervisor_topology, SupervisorTopology::Combined);
-    }
-
-    #[test]
-    fn serde_rejects_invalid_supervisor_topology() {
-        let json = serde_json::json!({
-            "supervisor_topology": "unsupported"
-        });
-        let err = serde_json::from_value::<KubernetesComputeConfig>(json).unwrap_err();
-        assert!(err.to_string().contains("unknown variant"));
-    }
-
     #[test]
     fn serde_override_workspace_storage_size() {
         let json = serde_json::json!({

crates/openshell-driver-kubernetes/src/lib.rs

@@ -6,8 +6,9 @@ pub mod driver;
 pub mod grpc;
 
 pub use config::{
-    AppArmorProfile, DEFAULT_SANDBOX_SERVICE_ACCOUNT_NAME, DEFAULT_WORKSPACE_STORAGE_SIZE,
-    KubernetesComputeConfig, SupervisorSideloadMethod, SupervisorTopology,
+    AppArmorProfile, DEFAULT_PROXY_UID, DEFAULT_SANDBOX_SERVICE_ACCOUNT_NAME,
+    DEFAULT_WORKSPACE_STORAGE_SIZE, KubernetesComputeConfig, KubernetesSidecarConfig,
+    SupervisorSideloadMethod, SupervisorTopology,
 };
 pub use driver::{KubernetesComputeDriver, KubernetesDriverError};
 pub use grpc::ComputeDriverService;

crates/openshell-driver-kubernetes/src/main.rs

@@ -1,7 +1,7 @@
 // SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-use clap::Parser;
+use clap::{ArgAction, Parser};
 use miette::{IntoDiagnostic, Result};
 use std::net::SocketAddr;
 use tracing::info;
@@ -10,8 +10,9 @@ use tracing_subscriber::EnvFilter;
 use openshell_core::VERSION;
 use openshell_core::proto::compute::v1::compute_driver_server::ComputeDriverServer;
 use openshell_driver_kubernetes::{
-    AppArmorProfile, ComputeDriverService, DEFAULT_SANDBOX_SERVICE_ACCOUNT_NAME,
-    KubernetesComputeConfig, KubernetesComputeDriver, SupervisorSideloadMethod, SupervisorTopology,
+    AppArmorProfile, ComputeDriverService, DEFAULT_PROXY_UID, DEFAULT_SANDBOX_SERVICE_ACCOUNT_NAME,
+    KubernetesComputeConfig, KubernetesComputeDriver, KubernetesSidecarConfig,
+    SupervisorSideloadMethod, SupervisorTopology,
 };
 
 #[derive(Parser, Debug)]
@@ -80,12 +81,24 @@ struct Args {
     )]
     supervisor_sideload_method: SupervisorSideloadMethod,
 
+    #[arg(long, env = "OPENSHELL_K8S_TOPOLOGY", default_value = "combined")]
+    topology: SupervisorTopology,
+
     #[arg(
-        long,
-        env = "OPENSHELL_SUPERVISOR_TOPOLOGY",
-        default_value = "combined"
+        long = "sidecar-proxy-uid",
+        alias = "proxy-uid",
+        env = "OPENSHELL_K8S_SIDECAR_PROXY_UID",
+        default_value_t = DEFAULT_PROXY_UID
     )]
-    supervisor_topology: SupervisorTopology,
+    sidecar_proxy_uid: u32,
+
+    #[arg(
+        long = "sidecar-process-binary-aware-network-policy",
+        env = "OPENSHELL_K8S_SIDECAR_PROCESS_BINARY_AWARE_NETWORK_POLICY",
+        default_value_t = true,
+        action = ArgAction::Set
+    )]
+    sidecar_process_binary_aware_network_policy: bool,
 
     #[arg(long, env = "OPENSHELL_ENABLE_USER_NAMESPACES")]
     enable_user_namespaces: bool,
@@ -130,7 +143,11 @@ async fn main() -> Result<()> {
             .unwrap_or_else(openshell_core::config::default_supervisor_image),
         supervisor_image_pull_policy: args.supervisor_image_pull_policy.unwrap_or_default(),
         supervisor_sideload_method: args.supervisor_sideload_method,
-        supervisor_topology: args.supervisor_topology,
+        topology: args.topology,
+        sidecar: KubernetesSidecarConfig {
+            proxy_uid: args.sidecar_proxy_uid,
+            process_binary_aware_network_policy: args.sidecar_process_binary_aware_network_policy,
+        },
         grpc_endpoint: args.grpc_endpoint.unwrap_or_default(),
         ssh_socket_path: args.sandbox_ssh_socket_path,
         client_tls_secret_name: args.client_tls_secret_name.unwrap_or_default(),

crates/openshell-driver-podman/README.md

@@ -130,8 +130,8 @@ sequenceDiagram
     C->>C: entrypoint: /opt/openshell/bin/openshell-sandbox
 ```
 
-The supervisor image from `deploy/docker/Dockerfile.supervisor` copies the static
-`openshell-sandbox` binary to `/openshell-sandbox`.
+The supervisor image from `deploy/docker/Dockerfile.supervisor` provides the
+static `openshell-sandbox` binary at `/openshell-sandbox`.
 Mounting that image at `/opt/openshell/bin` makes the binary available as
 `/opt/openshell/bin/openshell-sandbox`.
 

crates/openshell-driver-podman/src/container.rs

@@ -888,9 +888,8 @@ pub fn build_container_spec_with_token_and_gpu_devices(
         // Side-load the supervisor binary from a standalone OCI image.
         // Podman resolves image_volumes at the libpod layer, mounting the
         // image's filesystem at the destination path without starting a
-        // container from it. The supervisor image is FROM scratch with just
-        // the binary at /openshell-sandbox, so it appears at
-        // /opt/openshell/bin/openshell-sandbox.
+        // container from it. The supervisor image exposes the binary at
+        // /openshell-sandbox, so it appears at /opt/openshell/bin/openshell-sandbox.
         image_volumes,
         hostname: format!("sandbox-{}", sandbox.name),
         // Override the image's ENTRYPOINT so the supervisor binary runs

crates/openshell-sandbox/Cargo.toml

@@ -33,11 +33,16 @@ clap = { workspace = true }
 # Error handling
 miette = { workspace = true }
 
+# Unix ownership for Kubernetes sidecar init setup
+nix = { workspace = true }
+
 # TLS crypto provider install (main.rs)
 rustls = { workspace = true }
 
 # Serialization (serde_json::json! for OCSF unmapped fields)
+serde = { workspace = true }
 serde_json = { workspace = true }
+prost = { workspace = true }
 
 # Logging
 tracing = { workspace = true }

crates/openshell-sandbox/src/lib.rs

@@ -12,8 +12,9 @@ mod google_cloud_metadata;
 mod mechanistic_mapper;
 #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
 mod metadata_server;
+mod sidecar_control;
 
-use miette::Result;
+use miette::{IntoDiagnostic, Result, WrapErr};
 use std::future::Future;
 use std::sync::Arc;
 use std::sync::atomic::AtomicU32;
@@ -64,12 +65,20 @@ use openshell_core::denial::DenialEvent;
 use openshell_core::policy::{NetworkMode, NetworkPolicy, ProxyPolicy, SandboxPolicy};
 use openshell_core::provider_credentials::ProviderCredentialState;
 use openshell_supervisor_network::opa::OpaEngine;
+use openshell_supervisor_process::process::ProcessEnforcementMode;
 pub use openshell_supervisor_process::process::{ProcessHandle, ProcessStatus};
 use openshell_supervisor_process::skills;
 use tokio::sync::mpsc::UnboundedSender;
-#[cfg(target_os = "linux")]
+#[cfg(any(test, target_os = "linux"))]
 use tokio::time::timeout;
 
+const SIDECAR_NETWORK_ENFORCEMENT_MODE: &str = "sidecar-nftables";
+const SIDECAR_TLS_DIR: &str = "/etc/openshell-tls/proxy";
+const SIDECAR_CA_CERT: &str = "openshell-ca.pem";
+const SIDECAR_CA_BUNDLE: &str = "ca-bundle.pem";
+const SIDECAR_PROCESS_PROXY_ADDR: &str = "127.0.0.1:3128";
+const SIDECAR_READY_TIMEOUT_SECS: u64 = 120;
+
 /// Run a command in the sandbox.
 ///
 /// # Errors
@@ -125,17 +134,45 @@ pub async fn run_sandbox(
         }
     }
 
+    let sidecar_network_enforcement = sidecar_network_enforcement_enabled();
+    let process_enforcement_mode = process_enforcement_mode();
+    let process_uses_sidecar_control =
+        process_enabled && !network_enabled && sidecar_network_enforcement;
+    let mut process_control_connection = None;
+    let sidecar_bootstrap = if process_uses_sidecar_control {
+        let socket = sidecar_control_socket().ok_or_else(|| {
+            miette::miette!(
+                "{} is required for process-only sidecar topology",
+                openshell_core::sandbox_env::SIDECAR_CONTROL_SOCKET
+            )
+        })?;
+        let (bootstrap, connection) = sidecar_control::connect_process_client(
+            &socket,
+            Duration::from_secs(SIDECAR_READY_TIMEOUT_SECS),
+        )
+        .await?;
+        process_control_connection = Some(connection);
+        Some(bootstrap)
+    } else {
+        None
+    };
+
     // Load policy and initialize OPA engine
     let openshell_endpoint_for_proxy = openshell_endpoint.clone();
     let sandbox_name_for_agg = sandbox.clone();
-    let (mut policy, opa_engine, retained_proto, loaded_policy_origin) = load_policy(
-        sandbox_id.clone(),
-        sandbox,
-        openshell_endpoint.clone(),
-        policy_rules,
-        policy_data,
-    )
-    .await?;
+    let (mut policy, opa_engine, retained_proto, loaded_policy_origin) =
+        if let Some(bootstrap) = sidecar_bootstrap.as_ref() {
+            load_policy_from_sidecar_bootstrap(bootstrap)?
+        } else {
+            load_policy(
+                sandbox_id.clone(),
+                sandbox,
+                openshell_endpoint.clone(),
+                policy_rules,
+                policy_data,
+            )
+            .await?
+        };
 
     // Override the policy's process identity with the driver-resolved UID/GID
     // from the pod environment. The policy defaults to the name "sandbox" which
@@ -170,71 +207,89 @@ pub async fn run_sandbox(
         policy.process.run_as_group = Some(gid);
     }
 
-    // Fetch provider environment variables from the server.
-    // This is done after loading the policy so the sandbox can still start
-    // even if provider env fetch fails (graceful degradation).
-    let (
-        provider_env_revision,
-        provider_env,
-        provider_credential_expires_at_ms,
-        dynamic_credentials,
-    ) = if let (Some(id), Some(endpoint)) = (&sandbox_id, &openshell_endpoint) {
-        match openshell_core::grpc_client::fetch_provider_environment(endpoint, id).await {
-            Ok(result) => {
-                ocsf_emit!(
-                    ConfigStateChangeBuilder::new(ocsf_ctx())
-                        .severity(SeverityId::Informational)
-                        .status(StatusId::Success)
-                        .state(StateId::Enabled, "loaded")
-                        .message(format!(
-                            "Fetched provider environment [env_count:{}]",
-                            result.environment.len()
-                        ))
-                        .build()
-                );
-                (
-                    result.provider_env_revision,
-                    result.environment,
-                    result.credential_expires_at_ms,
-                    result.dynamic_credentials,
-                )
-            }
-            Err(e) => {
-                ocsf_emit!(
-                    ConfigStateChangeBuilder::new(ocsf_ctx())
-                        .severity(SeverityId::Medium)
-                        .status(StatusId::Failure)
-                        .state(StateId::Other, "degraded")
-                        .message(format!(
-                            "Failed to fetch provider environment, continuing without: {e}"
-                        ))
-                        .build()
-                );
+    #[cfg_attr(not(target_os = "linux"), allow(unused_mut))]
+    let (provider_credentials, mut provider_env) =
+        if let Some(bootstrap) = sidecar_bootstrap.as_ref() {
+            let provider_credentials = ProviderCredentialState::from_child_env_snapshot(
+                bootstrap.provider_env_revision,
+                bootstrap.provider_child_env.clone(),
+            );
+            (provider_credentials, bootstrap.provider_child_env.clone())
+        } else {
+            // Fetch provider environment variables from the server.
+            // This is done after loading the policy so the sandbox can still start
+            // even if provider env fetch fails (graceful degradation).
+            let (
+                provider_env_revision,
+                provider_env,
+                provider_credential_expires_at_ms,
+                dynamic_credentials,
+            ) = if let (Some(id), Some(endpoint)) = (&sandbox_id, &openshell_endpoint) {
+                match openshell_core::grpc_client::fetch_provider_environment(endpoint, id).await {
+                    Ok(result) => {
+                        ocsf_emit!(
+                            ConfigStateChangeBuilder::new(ocsf_ctx())
+                                .severity(SeverityId::Informational)
+                                .status(StatusId::Success)
+                                .state(StateId::Enabled, "loaded")
+                                .message(format!(
+                                    "Fetched provider environment [env_count:{}]",
+                                    result.environment.len()
+                                ))
+                                .build()
+                        );
+                        (
+                            result.provider_env_revision,
+                            result.environment,
+                            result.credential_expires_at_ms,
+                            result.dynamic_credentials,
+                        )
+                    }
+                    Err(e) => {
+                        ocsf_emit!(
+                            ConfigStateChangeBuilder::new(ocsf_ctx())
+                                .severity(SeverityId::Medium)
+                                .status(StatusId::Failure)
+                                .state(StateId::Other, "degraded")
+                                .message(format!(
+                                    "Failed to fetch provider environment, continuing without: {e}"
+                                ))
+                                .build()
+                        );
+                        (
+                            0,
+                            std::collections::HashMap::new(),
+                            std::collections::HashMap::new(),
+                            std::collections::HashMap::new(),
+                        )
+                    }
+                }
+            } else {
                 (
                     0,
                     std::collections::HashMap::new(),
                     std::collections::HashMap::new(),
                     std::collections::HashMap::new(),
                 )
-            }
-        }
-    } else {
-        (
-            0,
-            std::collections::HashMap::new(),
-            std::collections::HashMap::new(),
-            std::collections::HashMap::new(),
-        )
-    };
+            };
 
-    let provider_credentials = ProviderCredentialState::from_environment(
-        provider_env_revision,
-        provider_env,
-        provider_credential_expires_at_ms,
-        dynamic_credentials,
-    );
-    #[cfg_attr(not(target_os = "linux"), allow(unused_mut))]
-    let mut provider_env = provider_credentials.child_env_with_gcp_resolved();
+            let provider_credentials = ProviderCredentialState::from_environment(
+                provider_env_revision,
+                provider_env,
+                provider_credential_expires_at_ms,
+                dynamic_credentials,
+            );
+            let provider_env = provider_credentials.child_env_with_gcp_resolved();
+            (provider_credentials, provider_env)
+        };
+    let process_control_writer = process_control_connection
+        .as_ref()
+        .map(|connection| connection.writer.clone());
+    let mut process_control_closed = None;
+    if let Some(connection) = process_control_connection {
+        process_control_closed = Some(connection.closed);
+        spawn_sidecar_control_update_watcher(connection.updates, provider_credentials.clone());
+    }
 
     // Initialize the agent-proposals feature flag. Default false until the
     // initial settings fetch (or the poll loop) tells us otherwise. The flag
@@ -258,7 +313,7 @@ pub async fn run_sandbox(
     // it via setns(). The RAII handle lives in this frame for the duration
     // of the sandbox.
     #[cfg(target_os = "linux")]
-    let netns = if network_enabled {
+    let netns = if network_enabled && !sidecar_network_enforcement {
         openshell_supervisor_process::netns::create_netns_for_proxy(&policy)?
     } else {
         None
@@ -328,6 +383,80 @@ pub async fn run_sandbox(
         None
     };
 
+    #[cfg(target_os = "linux")]
+    let sidecar_control_server = if network_enabled && sidecar_network_enforcement {
+        if !matches!(policy.network.mode, NetworkMode::Proxy) {
+            return Err(miette::miette!(
+                "sidecar network enforcement requires proxy network mode"
+            ));
+        }
+        let socket = sidecar_control_socket().ok_or_else(|| {
+            miette::miette!(
+                "{} is required for sidecar topology",
+                openshell_core::sandbox_env::SIDECAR_CONTROL_SOCKET
+            )
+        })?;
+        let proto = retained_proto.as_ref().ok_or_else(|| {
+            miette::miette!(
+                "sidecar topology requires gateway policy data for the process supervisor"
+            )
+        })?;
+        let ca_paths = networking.as_ref().and_then(|n| n.ca_file_paths.clone());
+        Some(sidecar_control::spawn_server(
+            &socket,
+            sidecar_control::BootstrapData {
+                policy_proto: proto.clone(),
+                provider_env_revision: provider_credentials.snapshot().revision,
+                provider_child_env: provider_env.clone(),
+                proxy_ca_cert_path: ca_paths.as_ref().map(|paths| paths.0.clone()),
+                proxy_ca_bundle_path: ca_paths.as_ref().map(|paths| paths.1.clone()),
+            },
+            sidecar_expected_peer()?,
+        )?)
+    } else {
+        None
+    };
+    #[cfg(not(target_os = "linux"))]
+    let sidecar_control_server: Option<sidecar_control::ServerHandle> = None;
+
+    let sidecar_control_publisher = sidecar_control_server
+        .as_ref()
+        .map(sidecar_control::ServerHandle::publisher);
+
+    #[cfg(target_os = "linux")]
+    let mut sidecar_control_task = None;
+
+    #[cfg(target_os = "linux")]
+    if network_enabled
+        && sidecar_network_enforcement
+        && let Some(server) = sidecar_control_server
+    {
+        let trusted_ssh_socket_path = ssh_socket_path.clone().ok_or_else(|| {
+            miette::miette!(
+                "{} is required for sidecar network topology",
+                openshell_core::sandbox_env::SSH_SOCKET_PATH
+            )
+        })?;
+        let (entrypoint_rx, connection_task) = server.into_runtime_parts();
+        sidecar_control_task = Some(connection_task);
+        spawn_sidecar_entrypoint_handler(
+            entrypoint_rx,
+            entrypoint_pid.clone(),
+            opa_engine.clone(),
+            retained_proto.clone(),
+            openshell_endpoint.clone(),
+            sandbox_id.clone(),
+            std::path::PathBuf::from(trusted_ssh_socket_path),
+        );
+    }
+
+    #[cfg(not(target_os = "linux"))]
+    if network_enabled && sidecar_network_enforcement {
+        return Err(miette::miette!(
+            "sidecar network enforcement is only supported on Linux"
+        ));
+    }
+
     // Spawn the denial-aggregator flush task. The aggregator drains denial
     // events from the proxy + bypass monitor, batches them, and ships
     // summaries to the gateway via `SubmitPolicyAnalysis`.
@@ -398,11 +527,13 @@ pub async fn run_sandbox(
     }
 
     // Spawn background policy poll task (gRPC mode only).
-    if let (Some(id), Some(endpoint), Some(engine)) = (
-        sandbox_id.as_deref(),
-        openshell_endpoint.as_deref(),
-        opa_engine.as_ref(),
-    ) {
+    if !process_uses_sidecar_control
+        && let (Some(id), Some(endpoint), Some(engine)) = (
+            sandbox_id.as_deref(),
+            openshell_endpoint.as_deref(),
+            opa_engine.as_ref(),
+        )
+    {
         let poll_id = id.to_string();
         let poll_endpoint = endpoint.to_string();
         let poll_engine = engine.clone();
@@ -424,6 +555,7 @@ pub async fn run_sandbox(
             ocsf_enabled: poll_ocsf_enabled,
             provider_credentials: poll_provider_credentials,
             policy_local_ctx: poll_policy_local,
+            sidecar_control_publisher: sidecar_control_publisher.clone(),
         };
 
         tokio::spawn(async move {
@@ -479,10 +611,51 @@ pub async fn run_sandbox(
         }
     }
 
-    let exit_code = if process_enabled {
-        let ca_file_paths = networking.as_ref().and_then(|n| n.ca_file_paths.clone());
+    let process_policy = process_policy_for_topology(&policy, sidecar_network_enforcement)?;
+    let sidecar_bootstrap_ca_file_paths = sidecar_bootstrap.as_ref().and_then(|bootstrap| {
+        bootstrap
+            .proxy_ca_cert_path
+            .clone()
+            .zip(bootstrap.proxy_ca_bundle_path.clone())
+    });
 
-        openshell_supervisor_process::run::run_process(
+    let exit_code = if process_enabled {
+        let ca_file_paths = networking
+            .as_ref()
+            .and_then(|n| n.ca_file_paths.clone())
+            .or_else(|| {
+                if sidecar_network_enforcement {
+                    sidecar_bootstrap_ca_file_paths
+                        .clone()
+                        .or_else(sidecar_ca_file_paths)
+                } else {
+                    None
+                }
+            });
+
+        let entrypoint_started_tx =
+            if process_uses_sidecar_control && let Some(writer) = process_control_writer.clone() {
+                let (tx, rx) = tokio::sync::oneshot::channel();
+                tokio::spawn(async move {
+                    match rx.await {
+                        Ok(pid) => {
+                            if let Err(err) =
+                                sidecar_control::send_entrypoint_started(&writer, pid).await
+                            {
+                                warn!(error = %err, "Failed to send sidecar entrypoint event");
+                            }
+                        }
+                        Err(_closed) => {
+                            debug!("Entrypoint exited before sidecar entrypoint event was sent");
+                        }
+                    }
+                });
+                Some(tx)
+            } else {
+                None
+            };
+
+        let process = openshell_supervisor_process::run::run_process(
             program,
             args,
             workdir.as_deref(),
@@ -491,8 +664,11 @@ pub async fn run_sandbox(
             sandbox_id.as_deref(),
             openshell_endpoint.as_deref(),
             ssh_socket_path,
-            &policy,
+            sidecar_network_enforcement,
+            &process_policy,
+            process_enforcement_mode,
             entrypoint_pid,
+            entrypoint_started_tx,
             provider_credentials,
             provider_env,
             ca_file_paths,
@@ -502,14 +678,54 @@ pub async fn run_sandbox(
             bypass_denial_tx,
             #[cfg(target_os = "linux")]
             bypass_activity_tx,
-        )
-        .await?
+        );
+
+        if let Some(control_closed) = process_control_closed.as_mut() {
+            tokio::select! {
+                result = process => result?,
+                _ = control_closed => {
+                    ocsf_emit!(
+                        AppLifecycleBuilder::new(ocsf_ctx())
+                            .activity(ActivityId::Fail)
+                            .severity(SeverityId::High)
+                            .status(StatusId::Failure)
+                            .message(
+                                "Authoritative network-sidecar control channel closed; terminating process container"
+                            )
+                            .build()
+                    );
+                    return Err(miette::miette!(
+                        "authoritative network-sidecar control channel closed"
+                    ));
+                }
+            }
+        } else {
+            process.await?
+        }
     } else {
         // Network-only sidecar mode: keep the proxy and its background
-        // tasks alive (held via the `networking` value) until SIGINT or
-        // SIGTERM. Exit 0 on clean shutdown.
-        wait_for_shutdown_signal().await;
-        0
+        // tasks alive (held via the `networking` value) until shutdown. If the
+        // sole authenticated process-supervisor control connection closes,
+        // exit non-zero so Kubernetes restarts the network sidecar and creates
+        // a fresh one-client bootstrap listener for the restarted agent.
+        #[cfg(target_os = "linux")]
+        if let Some(control_task) = sidecar_control_task {
+            tokio::select! {
+                () = wait_for_shutdown_signal() => 0,
+                result = control_task => {
+                    warn!(?result, "Authoritative sidecar control channel exited; restarting sidecar");
+                    1
+                }
+            }
+        } else {
+            wait_for_shutdown_signal().await;
+            0
+        }
+        #[cfg(not(target_os = "linux"))]
+        {
+            wait_for_shutdown_signal().await;
+            0
+        }
     };
 
     // Drop networking explicitly so the proxy + bypass monitor RAII
@@ -552,6 +768,206 @@ async fn wait_for_shutdown_signal() {
     }
 }
 
+fn sidecar_network_enforcement_enabled() -> bool {
+    std::env::var(openshell_core::sandbox_env::NETWORK_ENFORCEMENT_MODE)
+        .is_ok_and(|value| value == SIDECAR_NETWORK_ENFORCEMENT_MODE)
+}
+
+fn process_enforcement_mode() -> ProcessEnforcementMode {
+    match std::env::var(openshell_core::sandbox_env::SUPERVISOR_TOPOLOGY)
+        .ok()
+        .as_deref()
+    {
+        Some("sidecar") => ProcessEnforcementMode::NetworkOnly,
+        _ => ProcessEnforcementMode::Full,
+    }
+}
+
+fn sidecar_control_socket() -> Option<std::path::PathBuf> {
+    std::env::var(openshell_core::sandbox_env::SIDECAR_CONTROL_SOCKET)
+        .ok()
+        .filter(|path| !path.is_empty())
+        .map(std::path::PathBuf::from)
+}
+
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+fn sidecar_expected_peer() -> Result<sidecar_control::ExpectedPeer> {
+    fn required_numeric_env(name: &str) -> Result<u32> {
+        let value = std::env::var(name)
+            .into_diagnostic()
+            .wrap_err_with(|| format!("{name} is required for sidecar control authentication"))?;
+        value.parse::<u32>().into_diagnostic().wrap_err_with(|| {
+            format!("{name} must be a numeric ID for sidecar control authentication")
+        })
+    }
+
+    Ok(sidecar_control::ExpectedPeer {
+        uid: required_numeric_env(openshell_core::sandbox_env::SANDBOX_UID)?,
+        gid: required_numeric_env(openshell_core::sandbox_env::SANDBOX_GID)?,
+    })
+}
+
+type LoadedPolicyBundle = (
+    SandboxPolicy,
+    Option<Arc<OpaEngine>>,
+    Option<openshell_core::proto::SandboxPolicy>,
+    LoadedPolicyOrigin,
+);
+
+fn load_policy_from_sidecar_bootstrap(
+    bootstrap: &sidecar_control::BootstrapData,
+) -> Result<LoadedPolicyBundle> {
+    let proto = bootstrap.policy_proto.clone();
+    let opa_engine = Some(Arc::new(OpaEngine::from_proto(&proto)?));
+    let policy = SandboxPolicy::try_from(proto.clone())?;
+    info!("Loaded sidecar policy from control socket bootstrap");
+    Ok((
+        policy,
+        opa_engine,
+        Some(proto),
+        LoadedPolicyOrigin::Gateway { revision: None },
+    ))
+}
+
+fn spawn_sidecar_control_update_watcher(
+    mut updates: tokio::sync::mpsc::UnboundedReceiver<sidecar_control::ControlUpdate>,
+    provider_credentials: ProviderCredentialState,
+) -> tokio::task::JoinHandle<()> {
+    tokio::spawn(async move {
+        while let Some(update) = updates.recv().await {
+            match update {
+                sidecar_control::ControlUpdate::ProviderEnvUpdated {
+                    revision,
+                    provider_child_env,
+                } => {
+                    if revision <= provider_credentials.snapshot().revision {
+                        continue;
+                    }
+                    let env_count = provider_credentials
+                        .install_child_env_snapshot(revision, provider_child_env);
+                    ocsf_emit!(
+                        ConfigStateChangeBuilder::new(ocsf_ctx())
+                            .severity(SeverityId::Informational)
+                            .status(StatusId::Success)
+                            .state(StateId::Enabled, "loaded")
+                            .unmapped("provider_env_revision", serde_json::json!(revision))
+                            .message(format!(
+                                "Sidecar provider environment refreshed [revision:{revision} env_count:{env_count}]"
+                            ))
+                            .build()
+                    );
+                }
+                sidecar_control::ControlUpdate::PolicyUpdated {
+                    policy_proto,
+                    policy_hash,
+                    config_revision,
+                } => {
+                    debug!(
+                        version = policy_proto.version,
+                        policy_hash,
+                        config_revision,
+                        "Received sidecar policy update for process supervisor"
+                    );
+                }
+            }
+        }
+    })
+}
+
+#[cfg(target_os = "linux")]
+fn spawn_sidecar_entrypoint_handler(
+    mut entrypoint_rx: tokio::sync::mpsc::Receiver<sidecar_control::EntrypointStarted>,
+    entrypoint_pid: Arc<AtomicU32>,
+    opa_engine: Option<Arc<OpaEngine>>,
+    retained_proto: Option<openshell_core::proto::SandboxPolicy>,
+    openshell_endpoint: Option<String>,
+    sandbox_id: Option<String>,
+    trusted_ssh_socket_path: std::path::PathBuf,
+) {
+    tokio::spawn(async move {
+        let mut session_started = false;
+        let mut trusted_supervisor_pid = None;
+        while let Some(started) = entrypoint_rx.recv().await {
+            entrypoint_pid.store(started.pid, std::sync::atomic::Ordering::Release);
+            if started.start_session {
+                info!(
+                    pid = started.pid,
+                    ssh_socket = %trusted_ssh_socket_path.display(),
+                    "Sidecar process supervisor reported entrypoint start"
+                );
+            } else {
+                trusted_supervisor_pid = Some(started.pid);
+                info!(
+                    pid = started.pid,
+                    "Sidecar process supervisor reported initial process anchor"
+                );
+            }
+
+            if let (Some(engine), Some(proto)) = (opa_engine.as_ref(), retained_proto.as_ref()) {
+                match engine.reload_from_proto_with_pid(proto, started.pid) {
+                    Ok(()) => info!(
+                        pid = started.pid,
+                        "Policy binary symlink resolution complete for sidecar process anchor"
+                    ),
+                    Err(err) => warn!(
+                        error = %err,
+                        pid = started.pid,
+                        "Failed to rebuild OPA engine with sidecar process anchor PID"
+                    ),
+                }
+            }
+
+            if started.start_session
+                && !session_started
+                && let (Some(endpoint), Some(id)) =
+                    (openshell_endpoint.as_ref(), sandbox_id.as_ref())
+            {
+                let Some(supervisor_pid) = trusted_supervisor_pid else {
+                    warn!(
+                        pid = started.pid,
+                        "Ignoring sidecar entrypoint event before authenticated supervisor anchor"
+                    );
+                    continue;
+                };
+                openshell_supervisor_process::supervisor_session::spawn(
+                    endpoint.clone(),
+                    id.clone(),
+                    trusted_ssh_socket_path.clone(),
+                    None,
+                    Some(supervisor_pid),
+                );
+                session_started = true;
+                info!("sidecar supervisor session task spawned");
+            }
+        }
+    });
+}
+
+fn sidecar_ca_file_paths() -> Option<(std::path::PathBuf, std::path::PathBuf)> {
+    let tls_dir = std::env::var(openshell_core::sandbox_env::PROXY_TLS_DIR)
+        .unwrap_or_else(|_| SIDECAR_TLS_DIR.to_string());
+    let cert = std::path::Path::new(&tls_dir).join(SIDECAR_CA_CERT);
+    let bundle = std::path::Path::new(&tls_dir).join(SIDECAR_CA_BUNDLE);
+    (cert.exists() && bundle.exists()).then_some((cert, bundle))
+}
+
+fn process_policy_for_topology(
+    policy: &SandboxPolicy,
+    sidecar_network_enforcement: bool,
+) -> Result<SandboxPolicy> {
+    let mut process_policy = policy.clone();
+    if sidecar_network_enforcement && matches!(process_policy.network.mode, NetworkMode::Proxy) {
+        let proxy = process_policy
+            .network
+            .proxy
+            .get_or_insert(ProxyPolicy { http_addr: None });
+        if proxy.http_addr.is_none() {
+            proxy.http_addr = Some(SIDECAR_PROCESS_PROXY_ADDR.parse().into_diagnostic()?);
+        }
+    }
+    Ok(process_policy)
+}
+
 /// Flush aggregated denial summaries to the gateway via `SubmitPolicyAnalysis`.
 async fn flush_proposals_to_gateway(
     endpoint: &str,
@@ -1947,6 +2363,7 @@ struct PolicyPollLoopContext {
     ocsf_enabled: Arc<std::sync::atomic::AtomicBool>,
     provider_credentials: ProviderCredentialState,
     policy_local_ctx: Option<Arc<openshell_supervisor_network::policy_local::PolicyLocalContext>>,
+    sidecar_control_publisher: Option<sidecar_control::Publisher>,
 }
 
 async fn run_policy_poll_loop(ctx: PolicyPollLoopContext) -> Result<()> {
@@ -2058,12 +2475,20 @@ async fn run_policy_poll_loop(ctx: PolicyPollLoopContext) -> Result<()> {
             .await
             {
                 Ok(env_result) => {
-                    let env_count = ctx.provider_credentials.install_environment(
+                    ctx.provider_credentials.install_environment(
                         env_result.provider_env_revision,
                         env_result.environment,
                         env_result.credential_expires_at_ms,
                         env_result.dynamic_credentials,
                     );
+                    let child_env = ctx.provider_credentials.child_env_with_gcp_resolved();
+                    let env_count = child_env.len();
+                    if let Some(publisher) = ctx.sidecar_control_publisher.as_ref() {
+                        publisher.publish_provider_env(
+                            env_result.provider_env_revision,
+                            child_env.clone(),
+                        );
+                    }
                     current_provider_env_revision = env_result.provider_env_revision;
                     ocsf_emit!(
                         ConfigStateChangeBuilder::new(ocsf_ctx())
@@ -2072,10 +2497,11 @@ async fn run_policy_poll_loop(ctx: PolicyPollLoopContext) -> Result<()> {
                             .state(StateId::Enabled, "loaded")
                             .unmapped(
                                 "provider_env_revision",
-                                serde_json::json!(current_provider_env_revision)
+                                serde_json::json!(env_result.provider_env_revision)
                             )
                             .message(format!(
-                                "Provider environment refreshed [revision:{current_provider_env_revision} env_count:{env_count}]"
+                                "Provider environment refreshed [revision:{} env_count:{env_count}]",
+                                env_result.provider_env_revision
                             ))
                             .build()
                     );
@@ -2111,6 +2537,13 @@ async fn run_policy_poll_loop(ctx: PolicyPollLoopContext) -> Result<()> {
                     if let Some(policy_local_ctx) = ctx.policy_local_ctx.as_ref() {
                         policy_local_ctx.set_current_policy(policy.clone()).await;
                     }
+                    if let Some(publisher) = ctx.sidecar_control_publisher.as_ref() {
+                        publisher.publish_policy(
+                            policy.clone(),
+                            result.policy_hash.clone(),
+                            result.config_revision,
+                        );
+                    }
                     if result.global_policy_version > 0 {
                         ocsf_emit!(ConfigStateChangeBuilder::new(ocsf_ctx())
                             .severity(SeverityId::Informational)
@@ -2317,8 +2750,24 @@ fn format_setting_value(es: &openshell_core::proto::EffectiveSetting) -> String
 )]
 mod tests {
     use super::*;
+    use openshell_core::policy::{
+        FilesystemPolicy, LandlockPolicy, NetworkMode, NetworkPolicy, ProcessPolicy, ProxyPolicy,
+    };
     use std::sync::atomic::{AtomicBool, Ordering};
 
+    fn proxy_policy(http_addr: Option<std::net::SocketAddr>) -> SandboxPolicy {
+        SandboxPolicy {
+            version: 1,
+            filesystem: FilesystemPolicy::default(),
+            network: NetworkPolicy {
+                mode: NetworkMode::Proxy,
+                proxy: Some(ProxyPolicy { http_addr }),
+            },
+            landlock: LandlockPolicy::default(),
+            process: ProcessPolicy::default(),
+        }
+    }
+
     fn effective_bool(value: bool) -> openshell_core::proto::EffectiveSetting {
         openshell_core::proto::EffectiveSetting {
             value: Some(openshell_core::proto::SettingValue {
@@ -2330,6 +2779,100 @@ mod tests {
         }
     }
 
+    #[test]
+    fn sidecar_process_policy_sets_loopback_proxy_addr() {
+        let policy = proxy_policy(None);
+
+        let process_policy = process_policy_for_topology(&policy, true).unwrap();
+
+        let http_addr = process_policy
+            .network
+            .proxy
+            .and_then(|proxy| proxy.http_addr)
+            .expect("sidecar process policy should set proxy address");
+        assert_eq!(http_addr.to_string(), SIDECAR_PROCESS_PROXY_ADDR);
+        assert!(
+            policy
+                .network
+                .proxy
+                .as_ref()
+                .expect("original policy should keep proxy config")
+                .http_addr
+                .is_none(),
+            "process policy normalization must not mutate the network policy"
+        );
+    }
+
+    #[test]
+    fn non_sidecar_process_policy_preserves_proxy_addr() {
+        let policy = proxy_policy(None);
+
+        let process_policy = process_policy_for_topology(&policy, false).unwrap();
+
+        assert!(
+            process_policy
+                .network
+                .proxy
+                .and_then(|proxy| proxy.http_addr)
+                .is_none()
+        );
+    }
+
+    #[tokio::test]
+    async fn sidecar_control_provider_env_update_installs_newer_revision() {
+        let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
+        let provider_credentials = ProviderCredentialState::from_child_env_snapshot(
+            1,
+            std::collections::HashMap::from([("TOKEN".to_string(), "old".to_string())]),
+        );
+        let handle = spawn_sidecar_control_update_watcher(rx, provider_credentials.clone());
+
+        tx.send(sidecar_control::ControlUpdate::ProviderEnvUpdated {
+            revision: 2,
+            provider_child_env: std::collections::HashMap::from([(
+                "TOKEN".to_string(),
+                "new".to_string(),
+            )]),
+        })
+        .unwrap();
+
+        timeout(Duration::from_secs(1), async {
+            loop {
+                if provider_credentials.snapshot().revision == 2 {
+                    break;
+                }
+                tokio::time::sleep(Duration::from_millis(10)).await;
+            }
+        })
+        .await
+        .unwrap();
+        let snapshot = provider_credentials.snapshot();
+        assert_eq!(snapshot.revision, 2);
+        assert_eq!(
+            snapshot.child_env.get("TOKEN").map(String::as_str),
+            Some("new")
+        );
+
+        tx.send(sidecar_control::ControlUpdate::ProviderEnvUpdated {
+            revision: 1,
+            provider_child_env: std::collections::HashMap::from([(
+                "TOKEN".to_string(),
+                "stale".to_string(),
+            )]),
+        })
+        .unwrap();
+        tokio::time::sleep(Duration::from_millis(20)).await;
+        assert_eq!(
+            provider_credentials
+                .snapshot()
+                .child_env
+                .get("TOKEN")
+                .map(String::as_str),
+            Some("new")
+        );
+        handle.abort();
+    }
+
     #[test]
     fn apply_ocsf_json_setting_enables_from_initial_settings_snapshot() {
         let enabled = AtomicBool::new(false);

crates/openshell-sandbox/src/main.rs

@@ -35,15 +35,36 @@ const DEBUG_RPC_SUBCOMMAND: &str = "debug-rpc";
 
 /// Default `--mode` value: run both supervisor leaves in a single binary.
 const DEFAULT_MODE: &str = "network,process";
+const SIDECAR_STATE_DIR: &str = "/run/openshell-sidecar";
+const SIDECAR_TLS_DIR: &str = "/etc/openshell-tls/proxy";
+#[cfg(target_os = "linux")]
+const CLIENT_TLS_DIR: &str = "/etc/openshell-tls/client";
+#[cfg(target_os = "linux")]
+const SIDECAR_CLIENT_TLS_SUBDIR: &str = "client";
+#[cfg(target_os = "linux")]
+const CLIENT_TLS_FILES: [&str; 3] = ["ca.crt", "tls.crt", "tls.key"];
+#[cfg(target_os = "linux")]
+const SIDECAR_STATE_DIR_MODE: u32 = 0o2775;
+#[cfg(target_os = "linux")]
+const SIDECAR_TLS_DIR_MODE: u32 = 0o755;
+#[cfg(target_os = "linux")]
+const SIDECAR_TLS_STAGING_DIR_MODE: u32 = 0o700;
+#[cfg(target_os = "linux")]
+const SIDECAR_CLIENT_TLS_DIR_MODE: u32 = 0o750;
+#[cfg(target_os = "linux")]
+const SIDECAR_CLIENT_TLS_FILE_MODE: u32 = 0o400;
 
 /// Which supervisor leaves are enabled in this process.
 ///
 /// Parsed from a comma-separated `--mode` value, e.g. `network`,
-/// `process`, or `network,process`. At least one must be set.
+/// `process`, or `network,process`. `network-init` is a one-shot setup mode
+/// used by the Kubernetes sidecar topology and cannot be combined with other
+/// mode components. At least one must be set.
 #[derive(Clone, Copy, Debug)]
 struct Mode {
     network: bool,
     process: bool,
+    network_init: bool,
 }
 
 impl std::str::FromStr for Mode {
@@ -53,20 +74,27 @@ impl std::str::FromStr for Mode {
         let mut mode = Self {
             network: false,
             process: false,
+            network_init: false,
         };
         for part in s.split(',').map(str::trim).filter(|p| !p.is_empty()) {
             match part {
                 "network" => mode.network = true,
                 "process" => mode.process = true,
+                "network-init" => mode.network_init = true,
                 other => {
                     return Err(format!(
-                        "unknown mode component '{other}' (expected 'network' and/or 'process')"
+                        "unknown mode component '{other}' (expected 'network', 'process', or 'network-init')"
                     ));
                 }
             }
         }
-        if !mode.network && !mode.process {
-            return Err("--mode must enable at least one of: network, process".into());
+        if mode.network_init && (mode.network || mode.process) {
+            return Err("--mode=network-init cannot be combined with other components".into());
+        }
+        if !mode.network && !mode.process && !mode.network_init {
+            return Err(
+                "--mode must enable at least one of: network, process, network-init".into(),
+            );
         }
         Ok(mode)
     }
@@ -125,7 +153,8 @@ struct Args {
     #[arg(long, default_value = "warn", env = openshell_core::sandbox_env::LOG_LEVEL)]
     log_level: String,
 
-    /// Filesystem path to the Unix socket the embedded SSH daemon binds.
+    /// Unix socket the embedded SSH daemon binds. On Linux, a value beginning
+    /// with `@` selects an abstract socket in the network namespace.
     /// The supervisor bridges `RelayStream` traffic from the gateway onto
     /// this socket; nothing else should connect to it.
     #[arg(long, env = openshell_core::sandbox_env::SSH_SOCKET_PATH)]
@@ -149,9 +178,28 @@ struct Args {
     /// "network" and/or "process". Defaults to both (single-binary
     /// topology). Use --mode=network for a network-only sidecar, or
     /// --mode=process for a process-only supervisor when network
-    /// enforcement runs in another pod.
+    /// enforcement runs in another pod. Use --mode=network-init only in
+    /// the Kubernetes init container that prepares sidecar nftables.
     #[arg(long, default_value = DEFAULT_MODE)]
     mode: Mode,
+
+    /// UID that the long-running Kubernetes network sidecar will run as.
+    /// `--mode=network-init` installs nftables rules that exempt this UID.
+    #[arg(long, env = "OPENSHELL_PROXY_UID", default_value_t = 1337)]
+    proxy_uid: u32,
+
+    /// GID assigned to shared sidecar state directories. Defaults to
+    /// `--proxy-uid` when omitted.
+    #[arg(long, env = "OPENSHELL_PROXY_GID")]
+    proxy_gid: Option<u32>,
+
+    /// Shared state directory between the network init container and sidecar.
+    #[arg(long, env = "OPENSHELL_SIDECAR_STATE_DIR", default_value = SIDECAR_STATE_DIR)]
+    sidecar_state_dir: String,
+
+    /// Shared TLS work directory between the network init container and sidecar.
+    #[arg(long, env = "OPENSHELL_PROXY_TLS_DIR", default_value = SIDECAR_TLS_DIR)]
+    sidecar_tls_dir: String,
 }
 
 /// Copy the running executable to `dest`, creating parent directories as
@@ -194,6 +242,189 @@ fn copy_self(dest: &str) -> Result<()> {
     Ok(())
 }
 
+#[cfg(target_os = "linux")]
+fn prepare_sidecar_directory(path: &Path, uid: u32, gid: u32, mode: u32) -> Result<()> {
+    use miette::Context as _;
+    use nix::unistd::{Gid, Uid, chown};
+    use std::os::unix::fs::PermissionsExt;
+
+    std::fs::create_dir_all(path)
+        .into_diagnostic()
+        .wrap_err_with(|| format!("failed to create sidecar directory {}", path.display()))?;
+    let mut perms = std::fs::metadata(path).into_diagnostic()?.permissions();
+    perms.set_mode(mode);
+    std::fs::set_permissions(path, perms)
+        .into_diagnostic()
+        .wrap_err_with(|| format!("failed to chmod sidecar directory {}", path.display()))?;
+    chown(path, Some(Uid::from_raw(uid)), Some(Gid::from_raw(gid)))
+        .into_diagnostic()
+        .wrap_err_with(|| {
+            format!(
+                "failed to chown sidecar directory {} to {uid}:{gid}",
+                path.display()
+            )
+        })?;
+    Ok(())
+}
+
+#[cfg(target_os = "linux")]
+fn prepare_sidecar_directory_for_current_user(path: &Path, mode: u32) -> Result<()> {
+    use miette::Context as _;
+    use nix::unistd::{Gid, Uid, chown};
+    use std::os::unix::fs::PermissionsExt;
+
+    let uid = Uid::current();
+    let gid = Gid::current();
+    std::fs::create_dir_all(path)
+        .into_diagnostic()
+        .wrap_err_with(|| format!("failed to create sidecar directory {}", path.display()))?;
+    chown(path, Some(uid), Some(gid))
+        .into_diagnostic()
+        .wrap_err_with(|| {
+            format!(
+                "failed to chown sidecar directory {} to {}:{}",
+                path.display(),
+                uid.as_raw(),
+                gid.as_raw()
+            )
+        })?;
+    let mut perms = std::fs::metadata(path).into_diagnostic()?.permissions();
+    perms.set_mode(mode);
+    std::fs::set_permissions(path, perms)
+        .into_diagnostic()
+        .wrap_err_with(|| format!("failed to chmod sidecar directory {}", path.display()))?;
+    Ok(())
+}
+
+#[cfg(target_os = "linux")]
+fn copy_sidecar_client_tls_if_present(
+    source_dir: &Path,
+    sidecar_tls_dir: &Path,
+    uid: u32,
+    gid: u32,
+) -> Result<()> {
+    use miette::Context as _;
+    use nix::unistd::{Gid, Uid, chown};
+    use std::os::unix::fs::PermissionsExt;
+
+    if !source_dir.exists() {
+        return Ok(());
+    }
+
+    let dest_dir = sidecar_tls_dir.join(SIDECAR_CLIENT_TLS_SUBDIR);
+    prepare_sidecar_directory_for_current_user(&dest_dir, SIDECAR_TLS_STAGING_DIR_MODE)?;
+    for file_name in CLIENT_TLS_FILES {
+        let source = source_dir.join(file_name);
+        if !source.exists() {
+            return Err(miette::miette!(
+                "client TLS source file is missing: {}",
+                source.display()
+            ));
+        }
+        let dest = dest_dir.join(file_name);
+        if dest.exists() {
+            std::fs::remove_file(&dest)
+                .into_diagnostic()
+                .wrap_err_with(|| {
+                    format!("failed to remove stale client TLS file {}", dest.display())
+                })?;
+        }
+        std::fs::copy(&source, &dest)
+            .into_diagnostic()
+            .wrap_err_with(|| {
+                format!(
+                    "failed to copy client TLS file {} to {}",
+                    source.display(),
+                    dest.display()
+                )
+            })?;
+        let mut perms = std::fs::metadata(&dest).into_diagnostic()?.permissions();
+        perms.set_mode(SIDECAR_CLIENT_TLS_FILE_MODE);
+        std::fs::set_permissions(&dest, perms)
+            .into_diagnostic()
+            .wrap_err_with(|| {
+                format!("failed to chmod copied client TLS file {}", dest.display())
+            })?;
+        chown(&dest, Some(Uid::from_raw(uid)), Some(Gid::from_raw(gid)))
+            .into_diagnostic()
+            .wrap_err_with(|| {
+                format!(
+                    "failed to chown copied client TLS file {} to {uid}:{gid}",
+                    dest.display()
+                )
+            })?;
+    }
+
+    prepare_sidecar_directory(&dest_dir, uid, gid, SIDECAR_CLIENT_TLS_DIR_MODE)?;
+
+    Ok(())
+}
+
+#[cfg(target_os = "linux")]
+fn run_network_init(
+    proxy_user_id: u32,
+    proxy_primary_group_id: u32,
+    sidecar_state_dir: &str,
+    sidecar_tls_dir: &str,
+) -> Result<()> {
+    validate_network_init_ids(proxy_user_id, proxy_primary_group_id)?;
+
+    let sidecar_state_dir = Path::new(sidecar_state_dir);
+    let sidecar_tls_dir = Path::new(sidecar_tls_dir);
+    prepare_sidecar_directory(
+        sidecar_state_dir,
+        proxy_user_id,
+        proxy_primary_group_id,
+        SIDECAR_STATE_DIR_MODE,
+    )?;
+    // The init container runs as uid 0 with CAP_DAC_OVERRIDE dropped. Keep the
+    // TLS work directory owned by the init user until the client cert copy is
+    // complete, then hand it to the long-running proxy UID.
+    prepare_sidecar_directory_for_current_user(sidecar_tls_dir, SIDECAR_TLS_DIR_MODE)?;
+    copy_sidecar_client_tls_if_present(
+        Path::new(CLIENT_TLS_DIR),
+        sidecar_tls_dir,
+        proxy_user_id,
+        proxy_primary_group_id,
+    )?;
+    prepare_sidecar_directory(
+        sidecar_tls_dir,
+        proxy_user_id,
+        proxy_primary_group_id,
+        SIDECAR_TLS_DIR_MODE,
+    )?;
+    openshell_supervisor_process::netns::install_sidecar_bypass_rules(proxy_user_id)
+}
+
+#[cfg(target_os = "linux")]
+fn validate_network_init_ids(proxy_user_id: u32, proxy_primary_group_id: u32) -> Result<()> {
+    if proxy_user_id != 0 && proxy_user_id < openshell_policy::MIN_SANDBOX_UID {
+        return Err(miette::miette!(
+            "--proxy-uid must be 0 or at least {}",
+            openshell_policy::MIN_SANDBOX_UID
+        ));
+    }
+    if proxy_primary_group_id < openshell_policy::MIN_SANDBOX_UID {
+        return Err(miette::miette!(
+            "--proxy-gid must be at least {}",
+            openshell_policy::MIN_SANDBOX_UID
+        ));
+    }
+    Ok(())
+}
+
+#[cfg(not(target_os = "linux"))]
+fn run_network_init(
+    _proxy_uid: u32,
+    _proxy_gid: u32,
+    _sidecar_state_dir: &str,
+    _sidecar_tls_dir: &str,
+) -> Result<()> {
+    Err(miette::miette!(
+        "--mode=network-init is only supported on Linux"
+    ))
+}
+
 fn main() -> Result<()> {
     // Handle `copy-self <DEST>` before clap so it works without any of the
     // sandbox flags. Kubernetes init containers invoke this path to seed an
@@ -222,6 +453,16 @@ fn main() -> Result<()> {
 
     let args = Args::parse();
 
+    if args.mode.network_init {
+        let proxy_gid = args.proxy_gid.unwrap_or(args.proxy_uid);
+        return run_network_init(
+            args.proxy_uid,
+            proxy_gid,
+            &args.sidecar_state_dir,
+            &args.sidecar_tls_dir,
+        );
+    }
+
     // Try to open a rolling log file; fall back to stderr-only logging if it fails
     // (e.g., /var/log is not writable in custom workload images).
     // Rotates daily, keeps the 3 most recent files to bound disk usage.
@@ -421,4 +662,50 @@ mod tests {
         let final_path = dest_dir.join("openshell-sandbox");
         assert!(final_path.exists(), "binary should land inside dest dir");
     }
+
+    #[test]
+    fn mode_parses_network_init_standalone() {
+        let mode = "network-init".parse::<Mode>().unwrap();
+        assert!(mode.network_init);
+        assert!(!mode.network);
+        assert!(!mode.process);
+    }
+
+    #[test]
+    fn mode_rejects_combined_network_init() {
+        let err = "network-init,network".parse::<Mode>().unwrap_err();
+        assert!(err.contains("cannot be combined"));
+    }
+
+    #[test]
+    fn mode_rejects_empty_value() {
+        let err = "".parse::<Mode>().unwrap_err();
+        assert!(err.contains("at least one"));
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn sidecar_tls_modes_preserve_proxy_owned_parent_and_private_client_dir() {
+        assert_eq!(SIDECAR_TLS_DIR_MODE, 0o755);
+        assert_eq!(SIDECAR_TLS_STAGING_DIR_MODE, 0o700);
+        assert_eq!(SIDECAR_CLIENT_TLS_DIR_MODE, 0o750);
+        assert_eq!(SIDECAR_CLIENT_TLS_FILE_MODE, 0o400);
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn network_init_accepts_root_proxy_uid_for_binary_aware_sidecar() {
+        validate_network_init_ids(0, openshell_policy::MIN_SANDBOX_UID).unwrap();
+    }
+
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn network_init_still_rejects_low_non_root_proxy_ids() {
+        let uid_err =
+            validate_network_init_ids(999, openshell_policy::MIN_SANDBOX_UID).unwrap_err();
+        assert!(uid_err.to_string().contains("--proxy-uid"));
+
+        let gid_err = validate_network_init_ids(0, 999).unwrap_err();
+        assert!(gid_err.to_string().contains("--proxy-gid"));
+    }
 }

crates/openshell-sandbox/src/sidecar_control.rs

@@ -0,0 +1,783 @@
+// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+//! Local control channel for Kubernetes sidecar topology.
+//!
+//! The network sidecar owns gateway credentials. The process supervisor in the
+//! agent container connects over this Unix socket to receive policy/provider
+//! state without mounting gateway credentials into the agent container.
+
+use miette::{IntoDiagnostic, Result, WrapErr};
+use prost::Message;
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::{Arc, RwLock};
+use std::time::Duration;
+use tokio::io::{AsyncBufReadExt, AsyncWrite, AsyncWriteExt, BufReader};
+use tokio::net::UnixListener;
+use tokio::net::unix::OwnedWriteHalf;
+use tokio::sync::{Mutex, broadcast, mpsc};
+use tracing::{debug, info, warn};
+
+#[derive(Debug, Clone)]
+pub struct BootstrapData {
+    pub policy_proto: openshell_core::proto::SandboxPolicy,
+    pub provider_env_revision: u64,
+    pub provider_child_env: HashMap<String, String>,
+    pub proxy_ca_cert_path: Option<PathBuf>,
+    pub proxy_ca_bundle_path: Option<PathBuf>,
+}
+
+#[derive(Debug, Clone)]
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+pub struct EntrypointStarted {
+    pub pid: u32,
+    pub start_session: bool,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct ExpectedPeer {
+    pub uid: u32,
+    pub gid: u32,
+}
+
+#[derive(Debug, Clone)]
+pub enum ControlUpdate {
+    ProviderEnvUpdated {
+        revision: u64,
+        provider_child_env: HashMap<String, String>,
+    },
+    PolicyUpdated {
+        policy_proto: openshell_core::proto::SandboxPolicy,
+        policy_hash: String,
+        config_revision: u64,
+    },
+}
+
+#[derive(Clone)]
+pub struct Publisher {
+    state: Arc<RwLock<BootstrapData>>,
+    updates: broadcast::Sender<WireServerMessage>,
+}
+
+impl Publisher {
+    pub fn publish_provider_env(&self, revision: u64, provider_child_env: HashMap<String, String>) {
+        {
+            let mut state = self.state.write().expect("sidecar control state poisoned");
+            if revision <= state.provider_env_revision {
+                return;
+            }
+            state.provider_env_revision = revision;
+            state.provider_child_env.clone_from(&provider_child_env);
+        }
+
+        let _ = self.updates.send(WireServerMessage::ProviderEnvUpdated {
+            revision,
+            provider_child_env,
+        });
+    }
+
+    pub fn publish_policy(
+        &self,
+        policy_proto: openshell_core::proto::SandboxPolicy,
+        policy_hash: String,
+        config_revision: u64,
+    ) {
+        {
+            let mut state = self.state.write().expect("sidecar control state poisoned");
+            state.policy_proto = policy_proto.clone();
+        }
+
+        let _ = self.updates.send(WireServerMessage::PolicyUpdated {
+            policy_proto: policy_proto.encode_to_vec(),
+            policy_hash,
+            config_revision,
+        });
+    }
+}
+
+pub struct ServerHandle {
+    publisher: Publisher,
+    #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+    entrypoint_rx: mpsc::Receiver<EntrypointStarted>,
+    connection_task: tokio::task::JoinHandle<()>,
+}
+
+impl ServerHandle {
+    pub fn publisher(&self) -> Publisher {
+        self.publisher.clone()
+    }
+
+    #[cfg(test)]
+    pub fn into_entrypoint_receiver(self) -> mpsc::Receiver<EntrypointStarted> {
+        self.entrypoint_rx
+    }
+
+    #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+    pub fn into_runtime_parts(
+        self,
+    ) -> (
+        mpsc::Receiver<EntrypointStarted>,
+        tokio::task::JoinHandle<()>,
+    ) {
+        (self.entrypoint_rx, self.connection_task)
+    }
+}
+
+pub struct ProcessConnection {
+    pub writer: Arc<Mutex<OwnedWriteHalf>>,
+    pub updates: mpsc::UnboundedReceiver<ControlUpdate>,
+    pub closed: tokio::sync::oneshot::Receiver<()>,
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+enum WireClientMessage {
+    BootstrapRequest { supervisor_pid: u32 },
+    EntrypointStarted { pid: u32 },
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+enum WireServerMessage {
+    BootstrapResponse {
+        policy_proto: Vec<u8>,
+        provider_env_revision: u64,
+        provider_child_env: HashMap<String, String>,
+        proxy_ca_cert_path: Option<String>,
+        proxy_ca_bundle_path: Option<String>,
+    },
+    ProviderEnvUpdated {
+        revision: u64,
+        provider_child_env: HashMap<String, String>,
+    },
+    PolicyUpdated {
+        policy_proto: Vec<u8>,
+        policy_hash: String,
+        config_revision: u64,
+    },
+}
+
+impl BootstrapData {
+    #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+    fn to_wire(&self) -> WireServerMessage {
+        WireServerMessage::BootstrapResponse {
+            policy_proto: self.policy_proto.encode_to_vec(),
+            provider_env_revision: self.provider_env_revision,
+            provider_child_env: self.provider_child_env.clone(),
+            proxy_ca_cert_path: self
+                .proxy_ca_cert_path
+                .as_ref()
+                .map(|path| path.display().to_string()),
+            proxy_ca_bundle_path: self
+                .proxy_ca_bundle_path
+                .as_ref()
+                .map(|path| path.display().to_string()),
+        }
+    }
+}
+
+impl TryFrom<WireServerMessage> for BootstrapData {
+    type Error = miette::Report;
+
+    fn try_from(message: WireServerMessage) -> Result<Self> {
+        let WireServerMessage::BootstrapResponse {
+            policy_proto,
+            provider_env_revision,
+            provider_child_env,
+            proxy_ca_cert_path,
+            proxy_ca_bundle_path,
+        } = message
+        else {
+            return Err(miette::miette!(
+                "expected sidecar bootstrap response, received update message"
+            ));
+        };
+
+        let policy_proto = openshell_core::proto::SandboxPolicy::decode(policy_proto.as_slice())
+            .into_diagnostic()
+            .wrap_err("failed to decode sidecar bootstrap policy")?;
+
+        Ok(Self {
+            policy_proto,
+            provider_env_revision,
+            provider_child_env,
+            proxy_ca_cert_path: proxy_ca_cert_path.map(PathBuf::from),
+            proxy_ca_bundle_path: proxy_ca_bundle_path.map(PathBuf::from),
+        })
+    }
+}
+
+impl TryFrom<WireServerMessage> for ControlUpdate {
+    type Error = miette::Report;
+
+    fn try_from(message: WireServerMessage) -> Result<Self> {
+        match message {
+            WireServerMessage::ProviderEnvUpdated {
+                revision,
+                provider_child_env,
+            } => Ok(Self::ProviderEnvUpdated {
+                revision,
+                provider_child_env,
+            }),
+            WireServerMessage::PolicyUpdated {
+                policy_proto,
+                policy_hash,
+                config_revision,
+            } => {
+                let policy_proto =
+                    openshell_core::proto::SandboxPolicy::decode(policy_proto.as_slice())
+                        .into_diagnostic()
+                        .wrap_err("failed to decode sidecar policy update")?;
+                Ok(Self::PolicyUpdated {
+                    policy_proto,
+                    policy_hash,
+                    config_revision,
+                })
+            }
+            WireServerMessage::BootstrapResponse { .. } => Err(miette::miette!(
+                "unexpected sidecar bootstrap response after initial handshake"
+            )),
+        }
+    }
+}
+
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+pub fn spawn_server(
+    path: &Path,
+    bootstrap: BootstrapData,
+    expected_peer: ExpectedPeer,
+) -> Result<ServerHandle> {
+    if let Some(parent) = path.parent() {
+        std::fs::create_dir_all(parent)
+            .into_diagnostic()
+            .wrap_err_with(|| {
+                format!(
+                    "failed to create sidecar control socket dir {}",
+                    parent.display()
+                )
+            })?;
+    }
+    match std::fs::remove_file(path) {
+        Ok(()) => {}
+        Err(err) if err.kind() == std::io::ErrorKind::NotFound => {}
+        Err(err) => {
+            return Err(err).into_diagnostic().wrap_err_with(|| {
+                format!(
+                    "failed to remove stale sidecar control socket {}",
+                    path.display()
+                )
+            });
+        }
+    }
+
+    let listener = UnixListener::bind(path)
+        .into_diagnostic()
+        .wrap_err_with(|| format!("failed to bind sidecar control socket {}", path.display()))?;
+
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::PermissionsExt;
+        std::fs::set_permissions(path, std::fs::Permissions::from_mode(0o660))
+            .into_diagnostic()
+            .wrap_err_with(|| {
+                format!(
+                    "failed to set permissions on sidecar control socket {}",
+                    path.display()
+                )
+            })?;
+    }
+
+    let state = Arc::new(RwLock::new(bootstrap));
+    let (updates, _) = broadcast::channel(32);
+    let (entrypoint_tx, entrypoint_rx) = mpsc::channel(8);
+    let publisher = Publisher {
+        state: state.clone(),
+        updates: updates.clone(),
+    };
+
+    let connection_task = tokio::spawn(accept_authoritative_connection(
+        listener,
+        path.to_path_buf(),
+        expected_peer,
+        state,
+        updates,
+        entrypoint_tx,
+    ));
+    info!(path = %path.display(), "Sidecar control socket listening");
+
+    Ok(ServerHandle {
+        publisher,
+        entrypoint_rx,
+        connection_task,
+    })
+}
+
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+async fn accept_authoritative_connection(
+    listener: UnixListener,
+    socket_path: PathBuf,
+    expected_peer: ExpectedPeer,
+    state: Arc<RwLock<BootstrapData>>,
+    updates: broadcast::Sender<WireServerMessage>,
+    entrypoint_tx: mpsc::Sender<EntrypointStarted>,
+) {
+    let stream = match listener.accept().await {
+        Ok((stream, _addr)) => stream,
+        Err(err) => {
+            warn!(error = %err, "Failed to accept authoritative sidecar control connection");
+            return;
+        }
+    };
+
+    // The process supervisor connects before it launches the workload. Drop
+    // the listener and unlink its pathname after that first accept so workload
+    // processes can neither open a second control channel nor impersonate a
+    // restarted server at the trusted path.
+    drop(listener);
+    if let Err(err) = std::fs::remove_file(&socket_path)
+        && err.kind() != std::io::ErrorKind::NotFound
+    {
+        warn!(
+            path = %socket_path.display(),
+            error = %err,
+            "Failed to unlink accepted sidecar control socket"
+        );
+    }
+
+    if let Err(err) = handle_connection(stream, expected_peer, state, updates, entrypoint_tx).await
+    {
+        warn!(error = %err, "Authoritative sidecar control connection closed");
+    }
+}
+
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+async fn handle_connection(
+    stream: tokio::net::UnixStream,
+    expected_peer: ExpectedPeer,
+    state: Arc<RwLock<BootstrapData>>,
+    updates: broadcast::Sender<WireServerMessage>,
+    entrypoint_tx: mpsc::Sender<EntrypointStarted>,
+) -> Result<()> {
+    let credentials = stream
+        .peer_cred()
+        .into_diagnostic()
+        .wrap_err("failed to read sidecar control peer credentials")?;
+    if credentials.uid() != expected_peer.uid || credentials.gid() != expected_peer.gid {
+        return Err(miette::miette!(
+            "sidecar control peer identity mismatch: expected uid:gid {}:{}, got {}:{}",
+            expected_peer.uid,
+            expected_peer.gid,
+            credentials.uid(),
+            credentials.gid(),
+        ));
+    }
+    let peer_pid = credentials
+        .pid()
+        .and_then(|pid| u32::try_from(pid).ok())
+        .ok_or_else(|| miette::miette!("sidecar control peer PID is unavailable"))?;
+
+    let (reader, mut writer) = stream.into_split();
+    let mut lines = BufReader::new(reader).lines();
+
+    let first_line =
+        lines.next_line().await.into_diagnostic()?.ok_or_else(|| {
+            miette::miette!("sidecar control client disconnected before bootstrap")
+        })?;
+    match decode_client_message(&first_line)? {
+        WireClientMessage::BootstrapRequest { supervisor_pid } => {
+            if supervisor_pid == 0 || supervisor_pid != peer_pid {
+                return Err(miette::miette!(
+                    "sidecar bootstrap PID mismatch: peer PID {peer_pid}, claimed PID {supervisor_pid}"
+                ));
+            }
+            entrypoint_tx
+                .send(EntrypointStarted {
+                    pid: supervisor_pid,
+                    start_session: false,
+                })
+                .await
+                .map_err(|_| miette::miette!("sidecar entrypoint receiver closed"))?;
+        }
+        WireClientMessage::EntrypointStarted { .. } => {
+            return Err(miette::miette!(
+                "sidecar control client sent entrypoint event before bootstrap"
+            ));
+        }
+    }
+
+    let bootstrap = {
+        let state = state.read().expect("sidecar control state poisoned");
+        state.to_wire()
+    };
+    write_json_line(&mut writer, &bootstrap).await?;
+
+    let mut update_rx = updates.subscribe();
+    loop {
+        tokio::select! {
+            line = lines.next_line() => {
+                let Some(line) = line.into_diagnostic()? else {
+                    return Ok(());
+                };
+                match decode_client_message(&line)? {
+                    WireClientMessage::BootstrapRequest { .. } => {
+                        debug!("Ignoring duplicate sidecar bootstrap request");
+                    }
+                    WireClientMessage::EntrypointStarted { pid } => {
+                        if pid == 0 {
+                            warn!("Ignoring sidecar entrypoint event with pid=0");
+                            continue;
+                        }
+                        entrypoint_tx
+                            .send(EntrypointStarted {
+                                pid,
+                                start_session: true,
+                            })
+                            .await
+                            .map_err(|_| miette::miette!("sidecar entrypoint receiver closed"))?;
+                    }
+                }
+            }
+            update = update_rx.recv() => {
+                match update {
+                    Ok(message) => write_json_line(&mut writer, &message).await?,
+                    Err(broadcast::error::RecvError::Lagged(skipped)) => {
+                        warn!(skipped, "Sidecar control client lagged behind updates");
+                    }
+                    Err(broadcast::error::RecvError::Closed) => return Ok(()),
+                }
+            }
+        }
+    }
+}
+
+pub async fn connect_process_client(
+    path: &Path,
+    timeout: Duration,
+) -> Result<(BootstrapData, ProcessConnection)> {
+    let stream = connect_with_retry(path, timeout).await?;
+    let (reader, mut writer) = stream.into_split();
+    write_json_line(
+        &mut writer,
+        &WireClientMessage::BootstrapRequest {
+            supervisor_pid: std::process::id(),
+        },
+    )
+    .await?;
+
+    let mut lines = BufReader::new(reader).lines();
+    let first_line = lines
+        .next_line()
+        .await
+        .into_diagnostic()?
+        .ok_or_else(|| miette::miette!("sidecar control closed before bootstrap response"))?;
+    let bootstrap = BootstrapData::try_from(decode_server_message(&first_line)?)?;
+
+    let (update_tx, updates) = mpsc::unbounded_channel();
+    let (closed_tx, closed) = tokio::sync::oneshot::channel();
+    tokio::spawn(async move {
+        while let Ok(Some(line)) = lines.next_line().await {
+            match decode_server_message(&line).and_then(ControlUpdate::try_from) {
+                Ok(update) => {
+                    if update_tx.send(update).is_err() {
+                        break;
+                    }
+                }
+                Err(err) => {
+                    warn!(error = %err, "Ignoring invalid sidecar control update");
+                }
+            }
+        }
+        let _ = closed_tx.send(());
+    });
+
+    Ok((
+        bootstrap,
+        ProcessConnection {
+            writer: Arc::new(Mutex::new(writer)),
+            updates,
+            closed,
+        },
+    ))
+}
+
+async fn connect_with_retry(path: &Path, timeout: Duration) -> Result<tokio::net::UnixStream> {
+    let deadline = tokio::time::Instant::now() + timeout;
+    loop {
+        match tokio::net::UnixStream::connect(path).await {
+            Ok(stream) => return Ok(stream),
+            Err(err) if tokio::time::Instant::now() < deadline => {
+                debug!(
+                    path = %path.display(),
+                    error = %err,
+                    "Waiting for sidecar control socket"
+                );
+                tokio::time::sleep(Duration::from_millis(100)).await;
+            }
+            Err(err) => {
+                return Err(err).into_diagnostic().wrap_err_with(|| {
+                    format!(
+                        "timed out waiting for sidecar control socket {}",
+                        path.display()
+                    )
+                });
+            }
+        }
+    }
+}
+
+pub async fn send_entrypoint_started(writer: &Arc<Mutex<OwnedWriteHalf>>, pid: u32) -> Result<()> {
+    let message = WireClientMessage::EntrypointStarted { pid };
+    let mut writer = writer.lock().await;
+    write_json_line(&mut *writer, &message).await
+}
+
+async fn write_json_line<W, T>(writer: &mut W, value: &T) -> Result<()>
+where
+    W: AsyncWrite + Unpin + Send,
+    T: Serialize + Sync,
+{
+    let bytes = serde_json::to_vec(value).into_diagnostic()?;
+    writer.write_all(&bytes).await.into_diagnostic()?;
+    writer.write_all(b"\n").await.into_diagnostic()?;
+    writer.flush().await.into_diagnostic()?;
+    Ok(())
+}
+
+#[cfg_attr(not(target_os = "linux"), allow(dead_code))]
+fn decode_client_message(line: &str) -> Result<WireClientMessage> {
+    serde_json::from_str(line)
+        .into_diagnostic()
+        .wrap_err("failed to decode sidecar client message")
+}
+
+fn decode_server_message(line: &str) -> Result<WireServerMessage> {
+    serde_json::from_str(line)
+        .into_diagnostic()
+        .wrap_err("failed to decode sidecar server message")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openshell_core::proto::SandboxPolicy;
+
+    fn current_peer() -> ExpectedPeer {
+        ExpectedPeer {
+            uid: nix::unistd::Uid::current().as_raw(),
+            gid: nix::unistd::Gid::current().as_raw(),
+        }
+    }
+
+    #[tokio::test]
+    async fn bootstrap_round_trips_policy_and_provider_env() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let mut env = HashMap::new();
+        env.insert("GITHUB_TOKEN".to_string(), "secret".to_string());
+        let bootstrap = BootstrapData {
+            policy_proto: SandboxPolicy {
+                version: 7,
+                ..SandboxPolicy::default()
+            },
+            provider_env_revision: 3,
+            provider_child_env: env.clone(),
+            proxy_ca_cert_path: Some(PathBuf::from("/tmp/ca.pem")),
+            proxy_ca_bundle_path: Some(PathBuf::from("/tmp/bundle.pem")),
+        };
+
+        let _server = spawn_server(&socket, bootstrap, current_peer()).unwrap();
+        let (received, _connection) = connect_process_client(&socket, Duration::from_secs(1))
+            .await
+            .unwrap();
+
+        assert_eq!(received.policy_proto.version, 7);
+        assert_eq!(received.provider_env_revision, 3);
+        assert_eq!(received.provider_child_env, env);
+        assert_eq!(
+            received.proxy_ca_cert_path,
+            Some(PathBuf::from("/tmp/ca.pem"))
+        );
+        assert_eq!(
+            received.proxy_ca_bundle_path,
+            Some(PathBuf::from("/tmp/bundle.pem"))
+        );
+    }
+
+    #[tokio::test]
+    async fn entrypoint_started_is_delivered_to_server() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let server = spawn_server(
+            &socket,
+            BootstrapData {
+                policy_proto: SandboxPolicy::default(),
+                provider_env_revision: 0,
+                provider_child_env: HashMap::new(),
+                proxy_ca_cert_path: None,
+                proxy_ca_bundle_path: None,
+            },
+            current_peer(),
+        )
+        .unwrap();
+        let mut entrypoint_rx = server.into_entrypoint_receiver();
+        let (_bootstrap, connection) = connect_process_client(&socket, Duration::from_secs(1))
+            .await
+            .unwrap();
+
+        let anchor = tokio::time::timeout(Duration::from_secs(1), entrypoint_rx.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(anchor.pid, std::process::id());
+        assert!(!anchor.start_session);
+
+        send_entrypoint_started(&connection.writer, 4242)
+            .await
+            .unwrap();
+
+        let started = tokio::time::timeout(Duration::from_secs(1), entrypoint_rx.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(started.pid, 4242);
+        assert!(started.start_session);
+    }
+
+    #[tokio::test]
+    async fn second_control_client_is_rejected_after_authoritative_bootstrap() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let _server = spawn_server(
+            &socket,
+            BootstrapData {
+                policy_proto: SandboxPolicy::default(),
+                provider_env_revision: 0,
+                provider_child_env: HashMap::new(),
+                proxy_ca_cert_path: None,
+                proxy_ca_bundle_path: None,
+            },
+            current_peer(),
+        )
+        .unwrap();
+
+        let (_bootstrap, _connection) = connect_process_client(&socket, Duration::from_secs(1))
+            .await
+            .unwrap();
+
+        let err = tokio::net::UnixStream::connect(&socket)
+            .await
+            .expect_err("control listener must be removed after the first bootstrap");
+        assert!(
+            matches!(
+                err.kind(),
+                std::io::ErrorKind::NotFound | std::io::ErrorKind::ConnectionRefused
+            ),
+            "unexpected second-client error: {err}"
+        );
+    }
+
+    #[tokio::test]
+    async fn authoritative_connection_task_ends_when_process_supervisor_disconnects() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let server = spawn_server(
+            &socket,
+            BootstrapData {
+                policy_proto: SandboxPolicy::default(),
+                provider_env_revision: 0,
+                provider_child_env: HashMap::new(),
+                proxy_ca_cert_path: None,
+                proxy_ca_bundle_path: None,
+            },
+            current_peer(),
+        )
+        .unwrap();
+        let (_entrypoint_rx, connection_task) = server.into_runtime_parts();
+        let (_bootstrap, connection) = connect_process_client(&socket, Duration::from_secs(1))
+            .await
+            .unwrap();
+
+        drop(connection);
+        tokio::time::timeout(Duration::from_secs(1), connection_task)
+            .await
+            .expect("server must observe authoritative client disconnect")
+            .expect("control task must not panic");
+    }
+
+    #[tokio::test]
+    async fn process_client_reports_network_sidecar_restart() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let server = spawn_server(
+            &socket,
+            BootstrapData {
+                policy_proto: SandboxPolicy::default(),
+                provider_env_revision: 0,
+                provider_child_env: HashMap::new(),
+                proxy_ca_cert_path: None,
+                proxy_ca_bundle_path: None,
+            },
+            current_peer(),
+        )
+        .unwrap();
+        let (_entrypoint_rx, connection_task) = server.into_runtime_parts();
+        let (_bootstrap, connection) = connect_process_client(&socket, Duration::from_secs(1))
+            .await
+            .unwrap();
+
+        connection_task.abort();
+        let _ = connection_task.await;
+        tokio::time::timeout(Duration::from_secs(1), connection.closed)
+            .await
+            .expect("process supervisor must observe network sidecar disconnect")
+            .expect("disconnect notifier must remain live");
+    }
+
+    #[tokio::test]
+    async fn bootstrap_rejects_claimed_pid_that_does_not_match_peer_credentials() {
+        let dir = tempfile::tempdir().unwrap();
+        let socket = dir.path().join("control.sock");
+        let server = spawn_server(
+            &socket,
+            BootstrapData {
+                policy_proto: SandboxPolicy::default(),
+                provider_env_revision: 0,
+                provider_child_env: HashMap::new(),
+                proxy_ca_cert_path: None,
+                proxy_ca_bundle_path: None,
+            },
+            current_peer(),
+        )
+        .unwrap();
+        let mut entrypoint_rx = server.into_entrypoint_receiver();
+
+        let mut stream = tokio::net::UnixStream::connect(&socket).await.unwrap();
+        write_json_line(
+            &mut stream,
+            &WireClientMessage::BootstrapRequest {
+                supervisor_pid: std::process::id().saturating_add(1),
+            },
+        )
+        .await
+        .unwrap();
+
+        assert!(
+            tokio::time::timeout(Duration::from_secs(1), entrypoint_rx.recv())
+                .await
+                .unwrap()
+                .is_none(),
+            "mismatched bootstrap must not publish a process anchor"
+        );
+    }
+
+    #[test]
+    fn malformed_client_message_is_rejected() {
+        let err = decode_client_message("not-json").unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("failed to decode sidecar client message")
+        );
+    }
+}

crates/openshell-supervisor-network/data/sandbox-policy.rego

@@ -19,6 +19,10 @@ allow_network if {
 	network_policy_for_request
 }
 
+binary_identity_required if {
+	object.get(object.get(data, "runtime", {}), "require_binary_identity", true)
+}
+
 # --- Deny reasons (specific diagnostics for debugging policy denials) ---
 
 deny_reason := "missing input.network" if {
@@ -131,6 +135,12 @@ endpoint_allowed(policy, network) if {
 	endpoint.ports[_] == network.port
 }
 
+# Binary matching can be relaxed by trusted runtime configuration. In that
+# mode, network policies are endpoint/L7 scoped and ignore policy.binaries.
+binary_allowed(_, _) if {
+	not binary_identity_required
+}
+
 # Binary matching: exact path.
 # SHA256 integrity is enforced in Rust via trust-on-first-use (TOFU) cache,
 # not in Rego. The proxy computes and caches binary hashes at runtime.
@@ -161,6 +171,10 @@ binary_allowed(policy, exec) if {
 	glob.match(b.path, ["/"], p)
 }
 
+user_declared_binary_allowed(_, _) if {
+	not binary_identity_required
+}
+
 user_declared_binary_allowed(policy, exec) if {
 	some b
 	b := policy.binaries[_]

crates/openshell-supervisor-network/src/identity.rs

@@ -100,23 +100,34 @@ impl BinaryIdentityCache {
     ///   Returns `Ok(hash)` if it matches, `Err` if the hash changed (binary tampered).
     #[cfg_attr(not(target_os = "linux"), allow(dead_code))]
     pub fn verify_or_cache(&self, path: &Path) -> Result<String> {
-        self.verify_or_cache_with_hasher(path, procfs::file_sha256)
+        self.verify_or_cache_with_paths(path, path, procfs::file_sha256)
     }
 
-    fn verify_or_cache_with_hasher<F>(&self, path: &Path, mut hash_file: F) -> Result<String>
+    #[cfg(target_os = "linux")]
+    pub fn verify_or_cache_process_exe(&self, display_path: &Path, pid: u32) -> Result<String> {
+        let proc_exe = PathBuf::from(format!("/proc/{pid}/exe"));
+        self.verify_or_cache_with_paths(display_path, &proc_exe, procfs::file_sha256)
+    }
+
+    fn verify_or_cache_with_paths<F>(
+        &self,
+        cache_path: &Path,
+        access_path: &Path,
+        mut hash_file: F,
+    ) -> Result<String>
     where
         F: FnMut(&Path) -> Result<String>,
     {
         let start = std::time::Instant::now();
-        let metadata = std::fs::metadata(path)
-            .map_err(|error| miette::miette!("Failed to stat {}: {error}", path.display()))?;
+        let metadata = std::fs::metadata(access_path)
+            .map_err(|error| miette::miette!("Failed to stat {}: {error}", cache_path.display()))?;
         let fingerprint = FileFingerprint::from_metadata(&metadata);
 
         let cached = self
             .hashes
             .lock()
             .map_err(|_| miette::miette!("Binary identity cache lock poisoned"))?
-            .get(path)
+            .get(cache_path)
             .cloned();
 
         if let Some(cached_binary) = &cached
@@ -125,7 +136,7 @@ impl BinaryIdentityCache {
             debug!(
                 "      verify_or_cache: {}ms CACHE HIT path={}",
                 start.elapsed().as_millis(),
-                path.display()
+                cache_path.display()
             );
             return Ok(cached_binary.hash.clone());
         }
@@ -133,29 +144,29 @@ impl BinaryIdentityCache {
         debug!(
             "      verify_or_cache: CACHE MISS size={} path={}",
             metadata.len(),
-            path.display()
+            cache_path.display()
         );
 
-        let current_hash = hash_file(path)?;
+        let current_hash = hash_file(access_path)?;
 
         let mut hashes = self
             .hashes
             .lock()
             .map_err(|_| miette::miette!("Binary identity cache lock poisoned"))?;
 
-        if let Some(existing) = hashes.get(path)
+        if let Some(existing) = hashes.get(cache_path)
             && existing.hash != current_hash
         {
             return Err(miette::miette!(
                 "Binary integrity violation: {} hash changed (cached: {}, current: {})",
-                path.display(),
+                cache_path.display(),
                 existing.hash,
                 current_hash
             ));
         }
 
         hashes.insert(
-            path.to_path_buf(),
+            cache_path.to_path_buf(),
             CachedBinary {
                 hash: current_hash.clone(),
                 fingerprint,
@@ -165,7 +176,7 @@ impl BinaryIdentityCache {
         debug!(
             "      verify_or_cache TOTAL (cold): {}ms path={}",
             start.elapsed().as_millis(),
-            path.display()
+            cache_path.display()
         );
 
         Ok(current_hash)
@@ -212,13 +223,13 @@ mod tests {
         let mut hash_calls = 0;
 
         let hash1 = cache
-            .verify_or_cache_with_hasher(tmp.path(), |path| {
+            .verify_or_cache_with_paths(tmp.path(), tmp.path(), |path| {
                 hash_calls += 1;
                 procfs::file_sha256(path)
             })
             .unwrap();
         let hash2 = cache
-            .verify_or_cache_with_hasher(tmp.path(), |path| {
+            .verify_or_cache_with_paths(tmp.path(), tmp.path(), |path| {
                 hash_calls += 1;
                 procfs::file_sha256(path)
             })
@@ -238,7 +249,7 @@ mod tests {
         let mut hash_calls = 0;
 
         let hash1 = cache
-            .verify_or_cache_with_hasher(tmp.path(), |path| {
+            .verify_or_cache_with_paths(tmp.path(), tmp.path(), |path| {
                 hash_calls += 1;
                 procfs::file_sha256(path)
             })
@@ -254,7 +265,7 @@ mod tests {
             .unwrap();
 
         let hash2 = cache
-            .verify_or_cache_with_hasher(tmp.path(), |path| {
+            .verify_or_cache_with_paths(tmp.path(), tmp.path(), |path| {
                 hash_calls += 1;
                 procfs::file_sha256(path)
             })
@@ -275,7 +286,7 @@ mod tests {
         let mut hash_calls = 0;
 
         cache
-            .verify_or_cache_with_hasher(&path, |path| {
+            .verify_or_cache_with_paths(&path, &path, |path| {
                 hash_calls += 1;
                 procfs::file_sha256(path)
             })
@@ -292,7 +303,7 @@ mod tests {
             .set_modified(original_mtime)
             .unwrap();
 
-        let result = cache.verify_or_cache_with_hasher(&path, |path| {
+        let result = cache.verify_or_cache_with_paths(&path, &path, |path| {
             hash_calls += 1;
             procfs::file_sha256(path)
         });
@@ -301,6 +312,28 @@ mod tests {
         assert_eq!(hash_calls, 2);
     }
 
+    #[test]
+    fn display_path_can_differ_from_access_path() {
+        let mut tmp = tempfile::NamedTempFile::new().unwrap();
+        tmp.write_all(b"binary content").unwrap();
+        tmp.flush().unwrap();
+        let display_path = Path::new("/usr/bin/python3");
+
+        let cache = BinaryIdentityCache::new();
+        let hash = cache
+            .verify_or_cache_with_paths(display_path, tmp.path(), procfs::file_sha256)
+            .unwrap();
+
+        assert!(!hash.is_empty());
+        assert!(
+            cache
+                .hashes
+                .lock()
+                .unwrap()
+                .contains_key(Path::new("/usr/bin/python3"))
+        );
+    }
+
     #[test]
     fn hash_mismatch_returns_error() {
         let dir = tempfile::tempdir().unwrap();

crates/openshell-supervisor-network/src/opa.rs

@@ -18,6 +18,7 @@ use std::sync::{
     Arc, Mutex,
     atomic::{AtomicU64, Ordering},
 };
+use tracing::info;
 
 /// Baked-in rego rules for OPA policy evaluation.
 /// These rules define the network access decision logic and static config
@@ -55,6 +56,49 @@ pub struct NetworkInput {
     pub cmdline_paths: Vec<PathBuf>,
 }
 
+pub(crate) fn network_binary_identity_required() -> bool {
+    std::env::var(openshell_core::sandbox_env::NETWORK_BINARY_IDENTITY).map_or(true, |value| {
+        !matches!(
+            value.as_str(),
+            "relaxed" | "disabled" | "endpoint-only" | "false" | "0"
+        )
+    })
+}
+
+fn inject_runtime_policy_data(data: &mut serde_json::Value, require_binary_identity: bool) {
+    let Some(obj) = data.as_object_mut() else {
+        return;
+    };
+    obj.insert(
+        "runtime".to_string(),
+        serde_json::json!({
+            "require_binary_identity": require_binary_identity,
+        }),
+    );
+}
+
+fn emit_binary_identity_mode(require_binary_identity: bool, source: &str) {
+    info!(
+        require_binary_identity,
+        source, "Configured OPA runtime binary identity mode"
+    );
+    openshell_ocsf::ocsf_emit!(
+        openshell_ocsf::ConfigStateChangeBuilder::new(openshell_ocsf::ctx::ctx())
+            .severity(openshell_ocsf::SeverityId::Informational)
+            .status(openshell_ocsf::StatusId::Success)
+            .state(openshell_ocsf::StateId::Enabled, "configured")
+            .unmapped(
+                "require_binary_identity",
+                serde_json::json!(require_binary_identity)
+            )
+            .unmapped("source", serde_json::json!(source))
+            .message(format!(
+                "OPA runtime binary identity mode configured [source:{source} require_binary_identity:{require_binary_identity}]"
+            ))
+            .build()
+    );
+}
+
 /// Sandbox configuration extracted from OPA data at startup.
 pub struct SandboxConfig {
     pub filesystem: FilesystemPolicy,
@@ -146,7 +190,9 @@ impl OpaEngine {
         engine
             .add_policy_from_file(policy_path)
             .map_err(|e| miette::miette!("{e}"))?;
-        let data_json = preprocess_yaml_data(&yaml_str)?;
+        let require_binary_identity = network_binary_identity_required();
+        emit_binary_identity_mode(require_binary_identity, "files");
+        let data_json = preprocess_yaml_data(&yaml_str, require_binary_identity)?;
         engine
             .add_data_json(&data_json)
             .map_err(|e| miette::miette!("{e}"))?;
@@ -160,11 +206,24 @@ impl OpaEngine {
     ///
     /// Preprocesses the YAML data to expand access presets and validate L7 config.
     pub fn from_strings(policy: &str, data_yaml: &str) -> Result<Self> {
+        Self::from_strings_with_binary_identity_required(
+            policy,
+            data_yaml,
+            network_binary_identity_required(),
+        )
+    }
+
+    pub(crate) fn from_strings_with_binary_identity_required(
+        policy: &str,
+        data_yaml: &str,
+        require_binary_identity: bool,
+    ) -> Result<Self> {
         let mut engine = regorus::Engine::new();
         engine
             .add_policy("policy.rego".into(), policy.into())
             .map_err(|e| miette::miette!("{e}"))?;
-        let data_json = preprocess_yaml_data(data_yaml)?;
+        emit_binary_identity_mode(require_binary_identity, "strings");
+        let data_json = preprocess_yaml_data(data_yaml, require_binary_identity)?;
         engine
             .add_data_json(&data_json)
             .map_err(|e| miette::miette!("{e}"))?;
@@ -193,11 +252,25 @@ impl OpaEngine {
     /// gap between user-specified symlink paths (e.g., `/usr/bin/python3`) and
     /// kernel-resolved canonical paths (e.g., `/usr/bin/python3.11`).
     pub fn from_proto_with_pid(proto: &ProtoSandboxPolicy, entrypoint_pid: u32) -> Result<Self> {
+        Self::from_proto_with_pid_and_binary_identity_required(
+            proto,
+            entrypoint_pid,
+            network_binary_identity_required(),
+        )
+    }
+
+    fn from_proto_with_pid_and_binary_identity_required(
+        proto: &ProtoSandboxPolicy,
+        entrypoint_pid: u32,
+        require_binary_identity: bool,
+    ) -> Result<Self> {
+        emit_binary_identity_mode(require_binary_identity, "proto");
         let data_json_str = proto_to_opa_data_json(proto, entrypoint_pid);
 
         // Parse back to Value for preprocessing, then re-serialize
         let mut data: serde_json::Value = serde_json::from_str(&data_json_str)
             .map_err(|e| miette::miette!("internal: failed to parse proto JSON: {e}"))?;
+        inject_runtime_policy_data(&mut data, require_binary_identity);
 
         // Validate BEFORE expanding presets
         let (errors, warnings) = crate::l7::validate_l7_policies(&data);
@@ -720,9 +793,10 @@ fn parse_process_policy(val: &regorus::Value) -> ProcessPolicy {
 }
 
 /// Preprocess YAML policy data: parse, normalize, validate, expand access presets, return JSON.
-fn preprocess_yaml_data(yaml_str: &str) -> Result<String> {
+fn preprocess_yaml_data(yaml_str: &str, require_binary_identity: bool) -> Result<String> {
     let mut data: serde_json::Value = serde_yml::from_str(yaml_str)
         .map_err(|e| miette::miette!("failed to parse YAML data: {e}"))?;
+    inject_runtime_policy_data(&mut data, require_binary_identity);
 
     // Normalize port → ports for all endpoints so Rego always sees "ports" array.
     normalize_endpoint_ports(&mut data);
@@ -2298,6 +2372,88 @@ process:
         assert!(eval_l7(&engine, &input));
     }
 
+    #[test]
+    fn l7_get_allowed_by_rules_when_binary_identity_relaxed() {
+        let engine =
+            OpaEngine::from_strings_with_binary_identity_required(TEST_POLICY, L7_TEST_DATA, false)
+                .expect("Failed to load relaxed L7 test data");
+        let mut input = l7_input("api.example.com", 8080, "GET", "/repos/myorg/foo");
+        input["exec"]["path"] = "".into();
+        assert!(eval_l7(&engine, &input));
+    }
+
+    #[test]
+    fn relaxed_binary_identity_preserves_matched_policy_and_l7_for_proto() {
+        let mut network_policies = std::collections::HashMap::new();
+        network_policies.insert(
+            "test_l7".to_string(),
+            NetworkPolicyRule {
+                name: "test_l7".to_string(),
+                endpoints: vec![NetworkEndpoint {
+                    host: "host.k3d.internal".to_string(),
+                    port: 56123,
+                    protocol: "rest".to_string(),
+                    enforcement: "enforce".to_string(),
+                    rules: vec![L7Rule {
+                        allow: Some(L7Allow {
+                            method: "GET".to_string(),
+                            path: "/allowed".to_string(),
+                            command: String::new(),
+                            query: std::collections::HashMap::new(),
+                            operation_type: String::new(),
+                            operation_name: String::new(),
+                            fields: Vec::new(),
+                            params: std::collections::HashMap::new(),
+                        }),
+                    }],
+                    allowed_ips: vec!["192.168.0.0/16".to_string()],
+                    ..Default::default()
+                }],
+                binaries: vec![NetworkBinary {
+                    path: "/usr/bin/curl".to_string(),
+                    ..Default::default()
+                }],
+            },
+        );
+        let proto = ProtoSandboxPolicy {
+            version: 1,
+            filesystem: Some(ProtoFs {
+                include_workdir: true,
+                read_only: vec![],
+                read_write: vec![],
+            }),
+            landlock: Some(openshell_core::proto::LandlockPolicy {
+                compatibility: "best_effort".to_string(),
+            }),
+            process: Some(ProtoProc {
+                run_as_user: "sandbox".to_string(),
+                run_as_group: "sandbox".to_string(),
+            }),
+            network_policies,
+        };
+        let engine = OpaEngine::from_proto_with_pid_and_binary_identity_required(&proto, 0, false)
+            .expect("engine from relaxed proto");
+        let network_input = NetworkInput {
+            host: "host.k3d.internal".into(),
+            port: 56123,
+            binary_path: PathBuf::new(),
+            binary_sha256: String::new(),
+            ancestors: vec![],
+            cmdline_paths: vec![],
+        };
+        let action = engine.evaluate_network_action(&network_input).unwrap();
+        assert_eq!(
+            action,
+            NetworkAction::Allow {
+                matched_policy: Some("test_l7".to_string())
+            }
+        );
+
+        let mut input = l7_input("host.k3d.internal", 56123, "GET", "/allowed");
+        input["exec"]["path"] = "".into();
+        assert!(eval_l7(&engine, &input));
+    }
+
     #[test]
     fn l7_post_allowed_by_rules() {
         let engine = l7_engine();
@@ -4626,6 +4782,46 @@ process:
         );
     }
 
+    #[test]
+    fn relaxed_binary_identity_allows_declared_endpoint_without_binary_match() {
+        let engine = OpaEngine::from_strings_with_binary_identity_required(
+            TEST_POLICY,
+            INFERENCE_TEST_DATA,
+            false,
+        )
+        .expect("Failed to load relaxed binary identity test data");
+        let input = NetworkInput {
+            host: "api.anthropic.com".into(),
+            port: 443,
+            binary_path: PathBuf::from("/tmp/unlisted-agent"),
+            binary_sha256: "unused".into(),
+            ancestors: vec![],
+            cmdline_paths: vec![],
+        };
+
+        let action = engine.evaluate_network_action(&input).unwrap();
+        assert_eq!(
+            action,
+            NetworkAction::Allow {
+                matched_policy: Some("claude_code".to_string())
+            },
+        );
+        assert!(
+            engine.query_exact_declared_endpoint_host(&input).unwrap(),
+            "relaxed identity should preserve exact declared endpoint handling"
+        );
+
+        let undeclared = NetworkInput {
+            host: "api.openai.com".into(),
+            ..input
+        };
+        let action = engine.evaluate_network_action(&undeclared).unwrap();
+        assert!(
+            matches!(action, NetworkAction::Deny { .. }),
+            "relaxed identity must not allow undeclared endpoints"
+        );
+    }
+
     #[test]
     fn unknown_endpoint_returns_deny() {
         let engine = inference_engine();

crates/openshell-supervisor-network/src/proxy.rs

@@ -42,6 +42,8 @@ const TUNNEL_PROTOCOL_PEEK_POLL: std::time::Duration = std::time::Duration::from
 const TUNNEL_PROTOCOL_PEEK_POLL: std::time::Duration = std::time::Duration::from_millis(1);
 const INFERENCE_LOCAL_HOST: &str = "inference.local";
 const INFERENCE_LOCAL_PORT: u16 = 443;
+#[cfg(target_os = "linux")]
+const SIDECAR_SUPERVISOR_TOPOLOGY: &str = "sidecar";
 
 /// Hostnames injected by compute drivers as `/etc/hosts` aliases for the host
 /// machine. Traffic to these names is eligible for the trusted-gateway SSRF
@@ -1426,7 +1428,7 @@ fn resolve_owner_identity(
         })?;
 
     let bin_hash = identity_cache
-        .verify_or_cache(&bin_path)
+        .verify_or_cache_process_exe(&bin_path, owner_pid)
         .map_err(|e| IdentityError {
             reason: format!("binary integrity check failed: {e}"),
             binary: Some(bin_path.clone()),
@@ -1434,11 +1436,15 @@ fn resolve_owner_identity(
             ancestors: vec![],
         })?;
 
-    let ancestors = crate::procfs::collect_ancestor_binaries(owner_pid, entrypoint_pid);
+    let ancestor_identities = collect_ancestor_identities(owner_pid, entrypoint_pid);
+    let ancestors: Vec<PathBuf> = ancestor_identities
+        .iter()
+        .map(|(_, path)| path.clone())
+        .collect();
 
-    for ancestor in &ancestors {
+    for (ancestor_pid, ancestor) in &ancestor_identities {
         identity_cache
-            .verify_or_cache(ancestor)
+            .verify_or_cache_process_exe(ancestor, *ancestor_pid)
             .map_err(|e| IdentityError {
                 reason: format!(
                     "ancestor integrity check failed for {}: {e}",
@@ -1463,6 +1469,31 @@ fn resolve_owner_identity(
     })
 }
 
+#[cfg(target_os = "linux")]
+fn collect_ancestor_identities(start_pid: u32, stop_pid: u32) -> Vec<(u32, PathBuf)> {
+    const MAX_DEPTH: usize = 64;
+    let mut ancestors = Vec::new();
+    let mut current = start_pid;
+
+    for _ in 0..MAX_DEPTH {
+        let parent_pid = match crate::procfs::read_ppid(current) {
+            Some(parent) if parent > 0 && parent != current => parent,
+            _ => break,
+        };
+
+        if let Ok(path) = crate::procfs::binary_path(parent_pid.cast_signed()) {
+            ancestors.push((parent_pid, path));
+        }
+
+        if parent_pid == stop_pid || parent_pid == 1 {
+            break;
+        }
+        current = parent_pid;
+    }
+
+    ancestors
+}
+
 /// Resolve the identity of the process owning a TCP peer connection.
 ///
 /// Walks `/proc/<entrypoint_pid>/net/tcp` to find the socket inode, locates
@@ -1472,10 +1503,10 @@ fn resolve_owner_identity(
 ///
 /// This is the identity-resolution block of [`evaluate_opa_tcp`] extracted
 /// into a standalone helper so it can be exercised by Linux-only regression
-/// tests without a full OPA engine. The key invariant under test is that on
-/// a hot-swap of the peer binary, the failure mode is
-/// `"Binary integrity violation"` (from the identity cache) rather than
-/// `"Failed to stat ... (deleted)"` (from the kernel-tainted path).
+/// tests without a full OPA engine. The key hot-swap invariant under test is
+/// that display paths are stripped for policy/logging, while integrity hashing
+/// reads the live executable via `/proc/<pid>/exe` instead of the replacement
+/// file that now exists at the display path.
 #[cfg(target_os = "linux")]
 fn resolve_process_identity(
     entrypoint_pid: u32,
@@ -1573,8 +1604,17 @@ fn evaluate_opa_tcp(
         }
     };
 
-    let pid = entrypoint_pid.load(Ordering::Acquire);
-    if pid == 0 {
+    if !crate::opa::network_binary_identity_required() {
+        let result = evaluate_endpoint_only_opa(engine, host, port);
+        debug!(
+            "evaluate_opa_tcp endpoint-only: host={host} port={port} action={:?}",
+            result.action
+        );
+        return result;
+    }
+
+    let entrypoint_pid = entrypoint_pid.load(Ordering::Acquire);
+    let Some(proc_net_anchor_pid) = proc_net_anchor_pid(entrypoint_pid) else {
         return deny(
             "entrypoint process not yet spawned".into(),
             None,
@@ -1582,12 +1622,12 @@ fn evaluate_opa_tcp(
             vec![],
             vec![],
         );
-    }
+    };
 
     let total_start = std::time::Instant::now();
     let peer_port = peer_addr.port();
 
-    let identity = match resolve_process_identity(pid, peer_port, identity_cache) {
+    let identity = match resolve_process_identity(proc_net_anchor_pid, peer_port, identity_cache) {
         Ok(id) => id,
         Err(err) => {
             return deny(
@@ -1641,6 +1681,52 @@ fn evaluate_opa_tcp(
     result
 }
 
+#[cfg(target_os = "linux")]
+fn proc_net_anchor_pid(entrypoint_pid: u32) -> Option<u32> {
+    if entrypoint_pid != 0 {
+        return Some(entrypoint_pid);
+    }
+    sidecar_topology_enabled().then(std::process::id)
+}
+
+#[cfg(target_os = "linux")]
+fn sidecar_topology_enabled() -> bool {
+    std::env::var(openshell_core::sandbox_env::SUPERVISOR_TOPOLOGY)
+        .is_ok_and(|value| value == SIDECAR_SUPERVISOR_TOPOLOGY)
+}
+
+fn evaluate_endpoint_only_opa(engine: &OpaEngine, host: &str, port: u16) -> ConnectDecision {
+    let input = crate::opa::NetworkInput {
+        host: host.to_string(),
+        port,
+        binary_path: PathBuf::new(),
+        binary_sha256: String::new(),
+        ancestors: vec![],
+        cmdline_paths: vec![],
+    };
+
+    match engine.evaluate_network_action_with_generation(&input) {
+        Ok((action, generation)) => ConnectDecision {
+            action,
+            generation,
+            binary: None,
+            binary_pid: None,
+            ancestors: vec![],
+            cmdline_paths: vec![],
+        },
+        Err(e) => ConnectDecision {
+            action: NetworkAction::Deny {
+                reason: format!("policy evaluation error: {e}"),
+            },
+            generation: engine.current_generation(),
+            binary: None,
+            binary_pid: None,
+            ancestors: vec![],
+            cmdline_paths: vec![],
+        },
+    }
+}
+
 /// Non-Linux stub: OPA identity binding requires /proc.
 #[cfg(not(target_os = "linux"))]
 fn evaluate_opa_tcp(
@@ -1648,9 +1734,13 @@ fn evaluate_opa_tcp(
     engine: &OpaEngine,
     _identity_cache: &BinaryIdentityCache,
     _entrypoint_pid: &AtomicU32,
-    _host: &str,
-    _port: u16,
+    host: &str,
+    port: u16,
 ) -> ConnectDecision {
+    if !crate::opa::network_binary_identity_required() {
+        return evaluate_endpoint_only_opa(engine, host, port);
+    }
+
     ConnectDecision {
         action: NetworkAction::Deny {
             reason: "identity binding unavailable on this platform".into(),
@@ -2152,14 +2242,24 @@ fn query_l7_route_snapshot(
     };
 
     match engine.query_endpoint_configs_with_generation(&input) {
-        Ok((vals, generation)) => Some(L7RouteSnapshot {
-            configs: vals
+        Ok((vals, generation)) => {
+            let configs: Vec<_> = vals
                 .into_iter()
                 .filter_map(|val| crate::l7::parse_l7_config(&val))
                 .map(|config| L7ConfigSnapshot { config })
-                .collect(),
-            generation,
-        }),
+                .collect();
+            debug!(
+                host,
+                port,
+                generation,
+                config_count = configs.len(),
+                "Forward proxy L7 route lookup complete"
+            );
+            Some(L7RouteSnapshot {
+                configs,
+                generation,
+            })
+        }
         Err(e) => {
             let event = NetworkActivityBuilder::new(openshell_ocsf::ctx::ctx())
                 .activity(ActivityId::Fail)
@@ -3337,10 +3437,29 @@ async fn handle_forward_proxy(
         }
     };
     let policy_str = matched_policy.as_deref().unwrap_or("-");
+    debug!(
+        host = %host_lc,
+        port,
+        binary = %binary_str,
+        binary_pid = %pid_str,
+        matched_policy = %policy_str,
+        decision_generation = decision.generation,
+        current_generation = opa_engine.current_generation(),
+        action = ?decision.action,
+        "Forward proxy L4 policy decision"
+    );
     let sandbox_entrypoint_pid = entrypoint_pid.load(Ordering::Acquire);
     let forward_generation_guard = match opa_engine.generation_guard(decision.generation) {
         Ok(guard) => guard,
         Err(e) => {
+            warn!(
+                host = %host_lc,
+                port,
+                decision_generation = decision.generation,
+                current_generation = opa_engine.current_generation(),
+                error = %e,
+                "Forward proxy rejected request because policy generation changed after L4 decision"
+            );
             emit_l7_tunnel_close_after_policy_change(&host_lc, port, e);
             emit_activity_simple(activity_tx, true, "policy_stale");
             respond(
@@ -3401,6 +3520,15 @@ async fn handle_forward_proxy(
         && !route.configs.is_empty()
     {
         if route.generation != forward_generation_guard.captured_generation() {
+            warn!(
+                host = %host_lc,
+                port,
+                decision_generation = decision.generation,
+                guard_generation = forward_generation_guard.captured_generation(),
+                route_generation = route.generation,
+                current_generation = opa_engine.current_generation(),
+                "Forward proxy rejected request because L7 route lookup used a different policy generation"
+            );
             emit_l7_tunnel_close_after_policy_change(
                 &host_lc,
                 port,
@@ -3426,6 +3554,14 @@ async fn handle_forward_proxy(
         let tunnel_engine = match opa_engine.clone_engine_for_tunnel(route.generation) {
             Ok(engine) => engine,
             Err(e) => {
+                warn!(
+                    host = %host_lc,
+                    port,
+                    route_generation = route.generation,
+                    current_generation = opa_engine.current_generation(),
+                    error = %e,
+                    "Forward proxy rejected request because L7 tunnel engine could not be cloned"
+                );
                 emit_l7_tunnel_close_after_policy_change(&host_lc, port, e);
                 emit_activity_simple(activity_tx, true, "policy_stale");
                 respond(
@@ -4105,6 +4241,14 @@ async fn handle_forward_proxy(
     };
 
     if let Err(e) = forward_generation_guard.ensure_current() {
+        warn!(
+            host = %host_lc,
+            port,
+            captured_generation = forward_generation_guard.captured_generation(),
+            current_generation = forward_generation_guard.current_generation(),
+            error = %e,
+            "Forward proxy rejected request because policy changed before upstream connect"
+        );
         emit_l7_tunnel_close_after_policy_change(&host_lc, port, e);
         emit_activity_simple(activity_tx, true, "policy_stale");
         respond(
@@ -4243,6 +4387,14 @@ async fn handle_forward_proxy(
     };
 
     if let Err(e) = forward_generation_guard.ensure_current() {
+        warn!(
+            host = %host_lc,
+            port,
+            captured_generation = forward_generation_guard.captured_generation(),
+            current_generation = forward_generation_guard.current_generation(),
+            error = %e,
+            "Forward proxy rejected request because policy changed before relay"
+        );
         emit_l7_tunnel_close_after_policy_change(&host_lc, port, e);
         respond(
             client,
@@ -4379,6 +4531,46 @@ mod tests {
     use tokio::io::{AsyncRead, AsyncReadExt, AsyncWriteExt};
     use tokio::net::{TcpListener, TcpStream};
 
+    #[test]
+    fn endpoint_only_opa_allows_declared_endpoint_without_process_identity() {
+        let policy = include_str!("../data/sandbox-policy.rego");
+        let data = r#"
+version: 1
+network_policies:
+  test_l7:
+    name: test_l7
+    endpoints:
+      - host: host.k3d.internal
+        port: 56123
+        protocol: rest
+        enforcement: enforce
+        rules:
+          - allow:
+              method: GET
+              path: /allowed
+    binaries:
+      - path: /usr/bin/curl
+"#;
+        let engine = OpaEngine::from_strings_with_binary_identity_required(policy, data, false)
+            .expect("relaxed engine");
+
+        let decision = evaluate_endpoint_only_opa(&engine, "host.k3d.internal", 56123);
+        assert_eq!(
+            decision.action,
+            NetworkAction::Allow {
+                matched_policy: Some("test_l7".to_string()),
+            }
+        );
+        assert!(decision.binary.is_none());
+        assert!(decision.ancestors.is_empty());
+
+        let denied = evaluate_endpoint_only_opa(&engine, "api.example.com", 443);
+        assert!(
+            matches!(denied.action, NetworkAction::Deny { .. }),
+            "endpoint-only mode must still deny undeclared endpoints"
+        );
+    }
+
     fn websocket_l7_config(
         protocol: crate::l7::L7Protocol,
         websocket_credential_rewrite: bool,
@@ -7992,27 +8184,23 @@ network_policies:
         assert_eq!(resp_str[body_start..].len(), cl);
     }
 
-    /// End-to-end regression for the `docker cp` hot-swap hazard that
-    /// motivated `binary_path()` stripping the kernel's `" (deleted)"`
-    /// suffix (PR #844).
+    /// End-to-end regression for the `docker cp` hot-swap hazard around
+    /// unlinked process executables.
     ///
-    /// Before the strip, the identity-resolution chain inside
-    /// `evaluate_opa_tcp` failed with `"Failed to stat
-    /// /opt/openshell/bin/openshell-sandbox (deleted)"` because
-    /// `BinaryIdentityCache::verify_or_cache()` tried to `metadata()` the
-    /// tainted path. That masked the real security signal: a live process
-    /// was now bound to a *different* binary on disk than the one that was
-    /// TOFU-cached. After the strip, `binary_path()` returns a path that
-    /// stats fine, the cache rehashes the new bytes, and the hash mismatch
-    /// surfaces as a `Binary integrity violation` error — the contract this
-    /// PR is trying to establish.
+    /// `binary_path()` strips the kernel's `" (deleted)"` suffix so policy
+    /// identity and logs use a clean display path. Integrity verification must
+    /// not hash that display path after a hot-swap, because it may now point to
+    /// unrelated replacement bytes. It hashes `/proc/<pid>/exe` instead, which
+    /// resolves to the live executable inode even after the original path was
+    /// unlinked.
     ///
     /// Test shape (from the review comment on the initial PR):
     /// 1. Start a `TcpListener` in the test process.
     /// 2. Copy `/bin/bash` to a temp path we control.
     /// 3. Prime `BinaryIdentityCache` with that temp binary's hash.
     /// 4. Spawn the temp bash as a child with a `/dev/tcp` one-liner that
-    ///    opens a real TCP connection to the listener and holds it open.
+    ///    opens a real TCP connection to the listener and holds it open
+    ///    inside the bash process.
     /// 5. Accept the connection on the listener side and capture the peer's
     ///    ephemeral port — that's what `resolve_process_identity` uses to
     ///    walk `/proc/net/tcp` back to the child PID.
@@ -8022,13 +8210,12 @@ network_policies:
     ///    now readlink to `" (deleted)"` OR the overwritten file, depending
     ///    on whether the filesystem reused the inode.
     /// 7. Call `resolve_process_identity` and assert:
-    ///    - the error reason contains `"Binary integrity violation"` (the
-    ///      cache detected the tampered on-disk bytes), and
-    ///    - the error reason does NOT contain `"Failed to stat"` or
-    ///      `"(deleted)"` (the old pre-strip failure mode).
+    ///    - identity resolution succeeds using the live executable hash, and
+    ///    - the returned display path does not contain the kernel-added
+    ///      `"(deleted)"` suffix.
     #[cfg(target_os = "linux")]
     #[test]
-    fn resolve_process_identity_surfaces_binary_integrity_violation_on_hot_swap() {
+    fn resolve_process_identity_hashes_live_exe_after_hot_swap() {
         use crate::identity::BinaryIdentityCache;
         use std::io::Read;
         use std::net::TcpListener;
@@ -8060,9 +8247,12 @@ network_policies:
         assert!(!v1_hash.is_empty());
 
         // 4. Spawn the temp bash with a /dev/tcp one-liner that opens a real
-        //    connection to the listener and sleeps to keep it open. The
-        //    `read -t` blocks on stdin so the shell stays resident.
-        let script = format!("exec 3<>/dev/tcp/127.0.0.1/{listener_port}; sleep 30 <&3");
+        //    connection to the listener and blocks in bash's `read` builtin
+        //    to keep it open. Do not use an external command like `sleep`:
+        //    it inherits the socket fd and intentionally trips the shared
+        //    socket ambiguity guard instead of exercising the hot-swap path.
+        let script =
+            format!("exec 3<>/dev/tcp/127.0.0.1/{listener_port}; read -r -t 30 _ <&3 || true");
         let mut child = Command::new(&bash_v1)
             .arg("-c")
             .arg(&script)
@@ -8106,10 +8296,11 @@ network_policies:
         std::fs::write(&bash_v1, tampered_bytes).expect("write replacement bytes");
 
         // 7. Resolve identity through the real helper and assert the
-        //    contract: we want "Binary integrity violation", not
-        //    "Failed to stat ... (deleted)".
+        //    contract: hash the live executable via /proc/<pid>/exe while
+        //    returning a clean display path for policy/logging.
         let test_pid = std::process::id();
         let result = resolve_process_identity(test_pid, peer_port, &cache);
+        let child_pid = child.id();
 
         // Always clean up the child before asserting so a failure doesn't
         // leak a sleeping process across test runs.
@@ -8117,40 +8308,29 @@ network_policies:
         let _ = child.wait();
 
         match result {
-            Ok(_) => panic!(
-                "resolve_process_identity unexpectedly succeeded after hot-swap; \
-                 the cache should have detected the tampered on-disk bytes"
-            ),
-            Err(err) => {
-                assert!(
-                    err.reason.contains("Binary integrity violation"),
-                    "expected 'Binary integrity violation' error, got: {}",
-                    err.reason
+            Ok(identity) => {
+                assert_eq!(
+                    identity.binary_pid, child_pid,
+                    "expected the hot-swapped bash child to own the socket"
+                );
+                assert_eq!(
+                    identity.bin_path, bash_v1,
+                    "expected stripped display path to remain the original binary path"
                 );
                 assert!(
-                    !err.reason.contains("Failed to stat"),
-                    "pre-PR-#844 failure mode leaked: {}",
-                    err.reason
+                    !identity.bin_path.to_string_lossy().contains("(deleted)"),
+                    "resolved binary path still tainted: {}",
+                    identity.bin_path.display()
                 );
-                assert!(
-                    !err.reason.contains("(deleted)"),
-                    "resolved path still contains '(deleted)' suffix: {}",
-                    err.reason
+                assert_eq!(
+                    identity.bin_hash, v1_hash,
+                    "expected integrity hash from the live executable, not replacement bytes"
                 );
-                // The binary field should be populated — we did resolve a
-                // path before failing.
-                assert!(
-                    err.binary.is_some(),
-                    "expected resolved binary path on integrity failure"
-                );
-                if let Some(path) = &err.binary {
-                    assert!(
-                        !path.to_string_lossy().contains("(deleted)"),
-                        "resolved binary path still tainted: {}",
-                        path.display()
-                    );
-                }
             }
+            Err(err) => panic!(
+                "resolve_process_identity failed after hot-swap; expected live-exe identity: {}",
+                err.reason
+            ),
         }
     }
 

crates/openshell-supervisor-network/src/run.rs

@@ -201,7 +201,9 @@ pub async fn run_networking(
     let (tls_state, ca_file_paths) = if matches!(policy.network.mode, NetworkMode::Proxy) {
         match SandboxCa::generate() {
             Ok(ca) => {
-                let tls_dir = std::path::Path::new("/etc/openshell-tls");
+                let tls_dir = std::env::var(openshell_core::sandbox_env::PROXY_TLS_DIR)
+                    .unwrap_or_else(|_| "/etc/openshell-tls".to_string());
+                let tls_dir = std::path::Path::new(&tls_dir);
                 let system_ca_bundle = read_system_ca_bundle();
                 match write_ca_files(&ca, tls_dir, &system_ca_bundle) {
                     Ok(paths) => {

crates/openshell-supervisor-process/src/lib.rs

@@ -19,6 +19,8 @@ pub mod skills;
 pub mod ssh;
 pub mod supervisor_session;
 
+mod unix_socket;
+
 #[cfg(target_os = "linux")]
 pub mod bypass_monitor;
 #[cfg(target_os = "linux")]

crates/openshell-supervisor-process/src/netns/mod.rs

@@ -285,43 +285,22 @@ impl NetworkNamespace {
         // monitor can see log entries from the sandbox namespace.
         enable_nf_log_all_netns();
 
-        // Try combined ruleset with log rules first. Log rules must appear
-        // before reject rules in the chain so packets are logged before being
-        // rejected. If the kernel lacks nft_log support, fall back to the
-        // reject-only ruleset.
-        let ruleset_with_log =
-            nft_ruleset::generate_bypass_ruleset(&host_ip_str, proxy_port, Some(&log_prefix));
+        let commands =
+            nft_ruleset::generate_bypass_commands(&host_ip_str, proxy_port, Some(&log_prefix));
 
-        if let Err(e) = run_nft_netns(&self.name, &nft_path, &ruleset_with_log) {
+        if let Err(e) = run_nft_commands_netns(&self.name, &nft_path, &commands) {
             openshell_ocsf::ocsf_emit!(
                 openshell_ocsf::ConfigStateChangeBuilder::new(openshell_ocsf::ctx::ctx())
-                    .severity(openshell_ocsf::SeverityId::Low)
+                    .severity(openshell_ocsf::SeverityId::Medium)
                     .status(openshell_ocsf::StatusId::Failure)
-                    .state(openshell_ocsf::StateId::Other, "degraded")
+                    .state(openshell_ocsf::StateId::Disabled, "failed")
                     .message(format!(
-                        "Failed to install bypass log rules (non-fatal), falling back to reject-only [ns:{}]: {e}",
+                        "Failed to install bypass detection rules [ns:{}]: {e}",
                         self.name
                     ))
                     .build()
             );
-
-            let ruleset_no_log =
-                nft_ruleset::generate_bypass_ruleset(&host_ip_str, proxy_port, None);
-
-            if let Err(e) = run_nft_netns(&self.name, &nft_path, &ruleset_no_log) {
-                openshell_ocsf::ocsf_emit!(
-                    openshell_ocsf::ConfigStateChangeBuilder::new(openshell_ocsf::ctx::ctx())
-                        .severity(openshell_ocsf::SeverityId::Medium)
-                        .status(openshell_ocsf::StatusId::Failure)
-                        .state(openshell_ocsf::StateId::Disabled, "failed")
-                        .message(format!(
-                            "Failed to install bypass detection rules [ns:{}]: {e}",
-                            self.name
-                        ))
-                        .build()
-                );
-                return Err(e);
-            }
+            return Err(e);
         }
 
         openshell_ocsf::ocsf_emit!(
@@ -467,6 +446,193 @@ pub fn create_netns_for_proxy(
     }
 }
 
+/// Install pod-network bypass enforcement for Kubernetes sidecar topology.
+///
+/// This runs in the current network namespace, not in a per-workload netns.
+/// The rules allow loopback and the sidecar proxy UID, then reject direct
+/// TCP/UDP egress from other UIDs so traffic must use the sidecar's local
+/// proxy.
+///
+/// # Errors
+///
+/// Returns an error when `nft` is unavailable or the ruleset cannot be loaded.
+pub fn install_sidecar_bypass_rules(proxy_uid: u32) -> Result<()> {
+    match install_sidecar_nft_bypass_rules(proxy_uid) {
+        Ok(()) => Ok(()),
+        Err(nft_error) => {
+            warn!(
+                error = %nft_error,
+                "Failed to install nftables sidecar rules; trying iptables-legacy fallback"
+            );
+            install_sidecar_iptables_legacy_bypass_rules(proxy_uid).map_err(|iptables_error| {
+                miette::miette!(
+                    "sidecar nft ruleset load failed: {nft_error}; sidecar iptables-legacy fallback failed: {iptables_error}"
+                )
+            })
+        }
+    }
+}
+
+fn install_sidecar_nft_bypass_rules(proxy_uid: u32) -> Result<()> {
+    let nft_cmd = find_nft().ok_or_else(|| {
+        miette::miette!(
+            "trusted nft helper not found; sidecar network enforcement requires nftables"
+        )
+    })?;
+    let log_prefix = Some("openshell:sidecar-bypass:");
+    let commands = nft_ruleset::generate_sidecar_bypass_commands(proxy_uid, log_prefix);
+    run_nft_commands_current_namespace(&nft_cmd, &commands)
+}
+
+const SIDECAR_IPTABLES_CHAIN: &str = "OPENSHELL_SIDECAR_BYPASS";
+const PROC_NET_IF_INET6_PATH: &str = "/proc/net/if_inet6";
+
+fn install_sidecar_iptables_legacy_bypass_rules(proxy_uid: u32) -> Result<()> {
+    let ipv4_filter_tool = find_iptables_legacy().ok_or_else(|| {
+        miette::miette!(
+            "trusted iptables-legacy helper not found; sidecar network enforcement fallback unavailable"
+        )
+    })?;
+
+    let ipv6_fence_tool = if current_namespace_has_non_loopback_ipv6()? {
+        Some(find_ip6tables_legacy().ok_or_else(|| {
+            miette::miette!(
+                "trusted ip6tables-legacy helper not found; sidecar network enforcement fallback cannot fence IPv6"
+            )
+        })?)
+    } else {
+        warn!(
+            "Skipping IPv6 sidecar iptables-legacy fallback because the current namespace has no non-loopback IPv6 interface"
+        );
+        None
+    };
+
+    cleanup_sidecar_iptables_legacy_rule_families(&ipv4_filter_tool, ipv6_fence_tool.as_deref());
+
+    if let Err(e) = install_sidecar_iptables_legacy_family_rules(
+        &ipv4_filter_tool,
+        proxy_uid,
+        "icmp-port-unreachable",
+    ) {
+        cleanup_sidecar_iptables_legacy_rule_families(
+            &ipv4_filter_tool,
+            ipv6_fence_tool.as_deref(),
+        );
+        return Err(e);
+    }
+
+    if let Some(ipv6_fence_tool) = ipv6_fence_tool
+        && let Err(e) = install_sidecar_iptables_legacy_family_rules(
+            &ipv6_fence_tool,
+            proxy_uid,
+            "icmp6-port-unreachable",
+        )
+    {
+        cleanup_sidecar_iptables_legacy_rule_families(&ipv4_filter_tool, Some(&ipv6_fence_tool));
+        return Err(e);
+    }
+
+    Ok(())
+}
+
+fn current_namespace_has_non_loopback_ipv6() -> Result<bool> {
+    match std::fs::read_to_string(PROC_NET_IF_INET6_PATH) {
+        Ok(content) => Ok(has_non_loopback_ipv6_interface(&content)),
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok(false),
+        Err(e) => Err(miette::miette!(
+            "failed to inspect {PROC_NET_IF_INET6_PATH} before installing sidecar IPv6 fence: {e}"
+        )),
+    }
+}
+
+fn has_non_loopback_ipv6_interface(content: &str) -> bool {
+    content.lines().any(|line| {
+        line.split_whitespace()
+            .nth(5)
+            .is_some_and(|iface| iface != "lo")
+    })
+}
+
+fn install_sidecar_iptables_legacy_family_rules(
+    cmd: &str,
+    proxy_uid: u32,
+    udp_reject_with: &str,
+) -> Result<()> {
+    let proxy_uid_arg = proxy_uid.to_string();
+    let commands: Vec<Vec<&str>> = vec![
+        vec!["-N", SIDECAR_IPTABLES_CHAIN],
+        vec!["-A", SIDECAR_IPTABLES_CHAIN, "-o", "lo", "-j", "ACCEPT"],
+        vec![
+            "-A",
+            SIDECAR_IPTABLES_CHAIN,
+            "-m",
+            "conntrack",
+            "--ctstate",
+            "ESTABLISHED,RELATED",
+            "-j",
+            "ACCEPT",
+        ],
+        vec![
+            "-A",
+            SIDECAR_IPTABLES_CHAIN,
+            "-m",
+            "owner",
+            "--uid-owner",
+            &proxy_uid_arg,
+            "-j",
+            "ACCEPT",
+        ],
+        vec![
+            "-A",
+            SIDECAR_IPTABLES_CHAIN,
+            "-p",
+            "tcp",
+            "-j",
+            "REJECT",
+            "--reject-with",
+            "tcp-reset",
+        ],
+        vec![
+            "-A",
+            SIDECAR_IPTABLES_CHAIN,
+            "-p",
+            "udp",
+            "-j",
+            "REJECT",
+            "--reject-with",
+            udp_reject_with,
+        ],
+        vec!["-A", "OUTPUT", "-j", SIDECAR_IPTABLES_CHAIN],
+    ];
+
+    for args in commands {
+        if let Err(e) = run_iptables_legacy_current_namespace(cmd, &args) {
+            cleanup_sidecar_iptables_legacy_rules(cmd);
+            return Err(e);
+        }
+    }
+
+    Ok(())
+}
+
+fn cleanup_sidecar_iptables_legacy_rules(iptables_cmd: &str) {
+    while run_iptables_legacy_current_namespace(
+        iptables_cmd,
+        &["-D", "OUTPUT", "-j", SIDECAR_IPTABLES_CHAIN],
+    )
+    .is_ok()
+    {}
+    let _ = run_iptables_legacy_current_namespace(iptables_cmd, &["-F", SIDECAR_IPTABLES_CHAIN]);
+    let _ = run_iptables_legacy_current_namespace(iptables_cmd, &["-X", SIDECAR_IPTABLES_CHAIN]);
+}
+
+fn cleanup_sidecar_iptables_legacy_rule_families(ipv4_cmd: &str, ipv6_cmd: Option<&str>) {
+    cleanup_sidecar_iptables_legacy_rules(ipv4_cmd);
+    if let Some(ipv6_cmd) = ipv6_cmd {
+        cleanup_sidecar_iptables_legacy_rules(ipv6_cmd);
+    }
+}
+
 /// Run an `ip` command on the host.
 fn run_ip(args: &[&str]) -> Result<()> {
     let ip_path = find_trusted_binary("ip", IP_SEARCH_PATHS)?;
@@ -490,6 +656,68 @@ fn run_ip(args: &[&str]) -> Result<()> {
     Ok(())
 }
 
+fn run_iptables_legacy_current_namespace(iptables_cmd: &str, args: &[&str]) -> Result<()> {
+    debug!(
+        command = %format!("{iptables_cmd} {}", args.join(" ")),
+        "Running iptables-legacy sidecar command"
+    );
+
+    let output = Command::new(iptables_cmd)
+        .args(args)
+        .output()
+        .into_diagnostic()?;
+
+    if !output.status.success() {
+        let stderr = String::from_utf8_lossy(&output.stderr);
+        return Err(miette::miette!(
+            "{iptables_cmd} {} failed: {}",
+            args.join(" "),
+            stderr.trim()
+        ));
+    }
+
+    Ok(())
+}
+
+/// Run a sequence of nft commands in the current network namespace.
+///
+/// Each command is executed as a separate `nft` invocation to avoid atomic
+/// batch rollback (where one unsupported expression like `ct state` or `log`
+/// causes the entire transaction, including table creation, to fail).
+///
+/// Commands marked as non-required are allowed to fail with a warning.
+/// Required commands that fail abort the sequence immediately.
+fn run_nft_commands_current_namespace(
+    nft_cmd: &str,
+    commands: &[nft_ruleset::NftCommand],
+) -> Result<()> {
+    for cmd in commands {
+        let args_str = cmd.args.join(" ");
+        debug!(command = %format!("{nft_cmd} {args_str}"), "Running nft command");
+
+        let output = Command::new(nft_cmd)
+            .args(&cmd.args)
+            .output()
+            .into_diagnostic()?;
+
+        if !output.status.success() {
+            let stderr = String::from_utf8_lossy(&output.stderr);
+            if cmd.required {
+                return Err(miette::miette!(
+                    "{nft_cmd} {args_str} failed: {}",
+                    stderr.trim()
+                ));
+            }
+            warn!(
+                command = %args_str,
+                error = %stderr.trim(),
+                "non-required nft command failed (continuing)"
+            );
+        }
+    }
+    Ok(())
+}
+
 /// Run an `ip` command inside a network namespace via `nsenter --net=`.
 ///
 /// We use `nsenter` instead of `ip netns exec` because `ip netns exec`
@@ -532,47 +760,51 @@ fn run_ip_netns(netns: &str, args: &[&str]) -> Result<()> {
     Ok(())
 }
 
-/// Load an nftables ruleset inside a network namespace via `nsenter --net=`.
+/// Run a sequence of nft commands inside a network namespace via `nsenter --net=`.
 ///
-/// Writes the ruleset to a temp file and loads it with `nft -f <path>`.
-/// A temp file is used instead of piping to stdin (`nft -f -`) because
-/// `nft` resolves `-` to `/dev/stdin`, which may not exist in minimal
-/// VM guest environments (e.g. virtiofs rootfs without /proc mounted
-/// at nft invocation time).
-fn run_nft_netns(netns: &str, nft_cmd: &str, ruleset: &str) -> Result<()> {
-    use std::io::Write;
-    let mut tmp = tempfile::Builder::new()
-        .prefix("openshell-nft-")
-        .suffix(".conf")
-        .tempfile()
-        .into_diagnostic()?;
-    tmp.write_all(ruleset.as_bytes()).into_diagnostic()?;
-    let ruleset_path = tmp.path().to_string_lossy().to_string();
-
+/// Each command is executed as a separate invocation to avoid atomic batch
+/// rollback. See [`run_nft_commands_current_namespace`] for rationale.
+fn run_nft_commands_netns(
+    netns: &str,
+    nft_cmd: &str,
+    commands: &[nft_ruleset::NftCommand],
+) -> Result<()> {
     let nsenter_path = find_trusted_binary("nsenter", NSENTER_SEARCH_PATHS)?;
     let ns_path = format!("/var/run/netns/{netns}");
     let net_flag = format!("--net={ns_path}");
 
-    debug!(
-        command = %format!("{nsenter_path} {net_flag} -- {nft_cmd} -f {ruleset_path}"),
-        "Loading nftables ruleset in namespace"
-    );
+    for cmd in commands {
+        let args_str = cmd.args.join(" ");
+        debug!(
+            command = %format!("{nsenter_path} {net_flag} -- {nft_cmd} {args_str}"),
+            "Running nft command in namespace"
+        );
 
-    let output = Command::new(nsenter_path)
-        .args([net_flag.as_str(), "--", nft_cmd, "-f", &ruleset_path])
-        .output()
-        .into_diagnostic()?;
+        let mut full_args = vec![net_flag.as_str(), "--", nft_cmd];
+        let arg_refs: Vec<&str> = cmd.args.iter().map(String::as_str).collect();
+        full_args.extend(&arg_refs);
 
-    drop(tmp);
+        let output = Command::new(nsenter_path)
+            .args(&full_args)
+            .output()
+            .into_diagnostic()?;
 
-    if !output.status.success() {
-        let stderr = String::from_utf8_lossy(&output.stderr);
-        return Err(miette::miette!(
-            "nft ruleset load failed in netns {netns}: {}",
-            stderr.trim()
-        ));
+        if !output.status.success() {
+            let stderr = String::from_utf8_lossy(&output.stderr);
+            if cmd.required {
+                return Err(miette::miette!(
+                    "nft {args_str} failed in netns {netns}: {}",
+                    stderr.trim()
+                ));
+            }
+            warn!(
+                command = %args_str,
+                error = %stderr.trim(),
+                netns = %netns,
+                "non-required nft command failed in namespace (continuing)"
+            );
+        }
     }
-
     Ok(())
 }
 
@@ -605,6 +837,16 @@ fn enable_nf_log_all_netns() {
 
 /// Well-known paths where nft may be installed.
 const NFT_SEARCH_PATHS: &[&str] = &["/usr/sbin/nft", "/sbin/nft", "/usr/bin/nft"];
+const IPTABLES_LEGACY_SEARCH_PATHS: &[&str] = &[
+    "/usr/sbin/iptables-legacy",
+    "/sbin/iptables-legacy",
+    "/usr/bin/iptables-legacy",
+];
+const IP6TABLES_LEGACY_SEARCH_PATHS: &[&str] = &[
+    "/usr/sbin/ip6tables-legacy",
+    "/sbin/ip6tables-legacy",
+    "/usr/bin/ip6tables-legacy",
+];
 
 fn find_trusted_binary<'a>(name: &str, paths: &'a [&str]) -> Result<&'a str> {
     paths
@@ -629,6 +871,18 @@ fn find_nft() -> Option<String> {
         .map(String::from)
 }
 
+fn find_iptables_legacy() -> Option<String> {
+    find_trusted_binary("iptables-legacy", IPTABLES_LEGACY_SEARCH_PATHS)
+        .ok()
+        .map(String::from)
+}
+
+fn find_ip6tables_legacy() -> Option<String> {
+    find_trusted_binary("ip6tables-legacy", IP6TABLES_LEGACY_SEARCH_PATHS)
+        .ok()
+        .map(String::from)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -668,6 +922,49 @@ mod tests {
         }
     }
 
+    #[test]
+    fn iptables_legacy_search_paths_are_absolute() {
+        for path in IPTABLES_LEGACY_SEARCH_PATHS {
+            assert!(
+                path.starts_with('/'),
+                "IPTABLES_LEGACY_SEARCH_PATHS entry must be absolute: {path}"
+            );
+        }
+    }
+
+    #[test]
+    fn ip6tables_legacy_search_paths_are_absolute() {
+        for path in IP6TABLES_LEGACY_SEARCH_PATHS {
+            assert!(
+                path.starts_with('/'),
+                "IP6TABLES_LEGACY_SEARCH_PATHS entry must be absolute: {path}"
+            );
+        }
+    }
+
+    #[test]
+    fn non_loopback_ipv6_detector_ignores_empty_input() {
+        assert!(!has_non_loopback_ipv6_interface(""));
+        assert!(!has_non_loopback_ipv6_interface("\n\n"));
+    }
+
+    #[test]
+    fn non_loopback_ipv6_detector_ignores_loopback() {
+        let content = "00000000000000000000000000000001 01 80 10 80 lo\n";
+
+        assert!(!has_non_loopback_ipv6_interface(content));
+    }
+
+    #[test]
+    fn non_loopback_ipv6_detector_detects_pod_interface() {
+        let content = "\
+00000000000000000000000000000001 01 80 10 80 lo
+fe800000000000000000000000000001 02 40 20 80 eth0
+";
+
+        assert!(has_non_loopback_ipv6_interface(content));
+    }
+
     #[test]
     #[ignore = "requires root privileges"]
     fn test_create_and_drop_namespace() {

crates/openshell-supervisor-process/src/netns/nft_ruleset.rs

@@ -6,85 +6,417 @@
 //! This module provides pure functions to generate nftables rulesets that enforce
 //! the sandbox network policy: all traffic must go through the proxy, with bypass
 //! attempts logged and rejected.
+//!
+//! Rulesets are returned as a sequence of individual nft commands rather than a
+//! monolithic file. Running each command as a separate `nft` invocation avoids
+//! `nft -f` atomic batch semantics, where a single unsupported expression (e.g.
+//! `ct state` without `nf_conntrack`, `log` without `nf_log`) rolls back the
+//! entire transaction including table/chain creation.
 
-/// Generate a complete nftables ruleset for sandbox network bypass enforcement.
+/// A single nft command with metadata about whether it is required.
+pub struct NftCommand {
+    /// The nft command arguments (e.g. `["add", "table", "inet", "openshell_bypass"]`).
+    pub args: Vec<String>,
+    /// When false, failure of this command is non-fatal; the caller should
+    /// log a warning and continue with the remaining commands.
+    pub required: bool,
+}
+
+/// Generate nft commands for sandbox network bypass enforcement.
 ///
 /// Creates an `inet` family table (handles both IPv4 and IPv6) with rules that:
 /// 1. Accept traffic to the proxy (IPv4 only)
 /// 2. Accept loopback traffic
-/// 3. Accept established/related connections
+/// 3. Accept established/related connections (optional; requires `nf_conntrack`)
 /// 4. Reject TCP and UDP bypass attempts (both IPv4 and IPv6)
 ///
 /// If `log_prefix` is provided, log rules are inserted before each reject rule
 /// so that bypass attempts are recorded in the kernel ring buffer before being
-/// rejected. The `log` expression requires kernel `nft_log` module support;
-/// pass `None` for `log_prefix` as a fallback when that module is unavailable.
-pub fn generate_bypass_ruleset(host_ip: &str, proxy_port: u16, log_prefix: Option<&str>) -> String {
-    let log_tcp = log_prefix
-        .map(|p| {
-            format!(
-                "\n        tcp flags syn limit rate 5/second burst 10 packets log prefix \"{p}\" flags skuid"
-            )
-        })
-        .unwrap_or_default();
-    let log_udp = log_prefix
-        .map(|p| {
-            format!(
-                "\n        meta l4proto udp limit rate 5/second burst 10 packets log prefix \"{p}\" flags skuid"
-            )
-        })
-        .unwrap_or_default();
+/// rejected. Log rules are always non-required since they need `nf_log` support.
+pub fn generate_bypass_commands(
+    host_ip: &str,
+    proxy_port: u16,
+    log_prefix: Option<&str>,
+) -> Vec<NftCommand> {
+    let table = "openshell_bypass";
+    let mut cmds = vec![
+        nft_cmd(true, &["add", "table", "inet", table]),
+        nft_cmd(true, &["flush", "table", "inet", table]),
+        nft_cmd(
+            true,
+            &[
+                "add",
+                "chain",
+                "inet",
+                table,
+                "output",
+                "{ type filter hook output priority 0; policy accept; }",
+            ],
+        ),
+        nft_cmd(
+            true,
+            &[
+                "add",
+                "rule",
+                "inet",
+                table,
+                "output",
+                "ip",
+                "daddr",
+                host_ip,
+                "tcp",
+                "dport",
+                &proxy_port.to_string(),
+                "accept",
+            ],
+        ),
+        nft_cmd(
+            true,
+            &[
+                "add", "rule", "inet", table, "output", "oifname", "lo", "accept",
+            ],
+        ),
+        nft_cmd(
+            false,
+            &[
+                "add",
+                "rule",
+                "inet",
+                table,
+                "output",
+                "ct",
+                "state",
+                "established,related",
+                "accept",
+            ],
+        ),
+    ];
 
-    format!(
-        r#"table inet openshell_bypass {{
-    chain output {{
-        type filter hook output priority 0; policy accept;
+    if let Some(prefix) = log_prefix {
+        cmds.push(nft_cmd(
+            false,
+            &[
+                "add", "rule", "inet", table, "output", "tcp", "flags", "syn", "limit", "rate",
+                "5/second", "burst", "10", "packets", "log", "prefix", prefix, "flags", "skuid",
+            ],
+        ));
+    }
 
-        ip daddr {host_ip} tcp dport {proxy_port} accept
-        oifname "lo" accept
-        ct state established,related accept{log_tcp}
-        meta nfproto ipv4 meta l4proto tcp reject with icmp type port-unreachable
-        meta nfproto ipv6 meta l4proto tcp reject with icmpv6 type port-unreachable{log_udp}
-        meta nfproto ipv4 meta l4proto udp reject with icmp type port-unreachable
-        meta nfproto ipv6 meta l4proto udp reject with icmpv6 type port-unreachable
-    }}
-}}
-"#
-    )
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv4",
+            "meta",
+            "l4proto",
+            "tcp",
+            "reject",
+            "with",
+            "icmp",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv6",
+            "meta",
+            "l4proto",
+            "tcp",
+            "reject",
+            "with",
+            "icmpv6",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+
+    if let Some(prefix) = log_prefix {
+        cmds.push(nft_cmd(
+            false,
+            &[
+                "add", "rule", "inet", table, "output", "meta", "l4proto", "udp", "limit", "rate",
+                "5/second", "burst", "10", "packets", "log", "prefix", prefix, "flags", "skuid",
+            ],
+        ));
+    }
+
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv4",
+            "meta",
+            "l4proto",
+            "udp",
+            "reject",
+            "with",
+            "icmp",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv6",
+            "meta",
+            "l4proto",
+            "udp",
+            "reject",
+            "with",
+            "icmpv6",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+
+    cmds
+}
+
+/// Generate nft commands for Kubernetes sidecar enforcement.
+///
+/// The network sidecar and the process supervisor share a pod network
+/// namespace. The sidecar runs as `proxy_uid` and owns external egress;
+/// sandbox traffic must use loopback services hosted by that sidecar
+/// (gateway forward and HTTP CONNECT proxy). The generated fence rejects
+/// TCP/UDP bypass attempts from non-proxy UIDs; other L4 protocols are outside
+/// the sidecar policy fence.
+pub fn generate_sidecar_bypass_commands(
+    proxy_uid: u32,
+    log_prefix: Option<&str>,
+) -> Vec<NftCommand> {
+    let table = "openshell_sidecar_bypass";
+    let uid_str = proxy_uid.to_string();
+    let mut cmds = vec![
+        nft_cmd(true, &["add", "table", "inet", table]),
+        nft_cmd(true, &["flush", "table", "inet", table]),
+        nft_cmd(
+            true,
+            &[
+                "add",
+                "chain",
+                "inet",
+                table,
+                "output",
+                "{ type filter hook output priority 0; policy accept; }",
+            ],
+        ),
+        nft_cmd(
+            true,
+            &[
+                "add", "rule", "inet", table, "output", "oifname", "lo", "accept",
+            ],
+        ),
+        nft_cmd(
+            false,
+            &[
+                "add",
+                "rule",
+                "inet",
+                table,
+                "output",
+                "ct",
+                "state",
+                "established,related",
+                "accept",
+            ],
+        ),
+        nft_cmd(
+            true,
+            &[
+                "add", "rule", "inet", table, "output", "meta", "skuid", &uid_str, "accept",
+            ],
+        ),
+    ];
+
+    if let Some(prefix) = log_prefix {
+        cmds.push(nft_cmd(
+            false,
+            &[
+                "add", "rule", "inet", table, "output", "tcp", "flags", "syn", "limit", "rate",
+                "5/second", "burst", "10", "packets", "log", "prefix", prefix, "flags", "skuid",
+            ],
+        ));
+    }
+
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv4",
+            "meta",
+            "l4proto",
+            "tcp",
+            "reject",
+            "with",
+            "icmp",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv6",
+            "meta",
+            "l4proto",
+            "tcp",
+            "reject",
+            "with",
+            "icmpv6",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+
+    if let Some(prefix) = log_prefix {
+        cmds.push(nft_cmd(
+            false,
+            &[
+                "add", "rule", "inet", table, "output", "meta", "l4proto", "udp", "limit", "rate",
+                "5/second", "burst", "10", "packets", "log", "prefix", prefix, "flags", "skuid",
+            ],
+        ));
+    }
+
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv4",
+            "meta",
+            "l4proto",
+            "udp",
+            "reject",
+            "with",
+            "icmp",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+    cmds.push(nft_cmd(
+        true,
+        &[
+            "add",
+            "rule",
+            "inet",
+            table,
+            "output",
+            "meta",
+            "nfproto",
+            "ipv6",
+            "meta",
+            "l4proto",
+            "udp",
+            "reject",
+            "with",
+            "icmpv6",
+            "type",
+            "port-unreachable",
+        ],
+    ));
+
+    cmds
+}
+
+fn nft_cmd(required: bool, args: &[&str]) -> NftCommand {
+    NftCommand {
+        args: args.iter().map(|s| (*s).to_string()).collect(),
+        required,
+    }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
 
-    #[test]
-    fn generates_bypass_ruleset_with_proxy_rule() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, None);
-        assert!(ruleset.contains("table inet openshell_bypass"));
-        assert!(ruleset.contains("chain output"));
-        assert!(ruleset.contains("ip daddr 10.0.2.2 tcp dport 8080 accept"));
+    fn cmd_str(cmd: &NftCommand) -> String {
+        cmd.args.join(" ")
+    }
+
+    fn all_strs(cmds: &[NftCommand]) -> String {
+        cmds.iter().map(cmd_str).collect::<Vec<_>>().join("\n")
     }
 
     #[test]
-    fn ruleset_has_inet_family_table_and_output_chain() {
-        let ruleset = generate_bypass_ruleset("192.168.1.1", 3128, None);
-        assert!(ruleset.contains("table inet openshell_bypass"));
-        assert!(ruleset.contains("type filter hook output priority 0; policy accept;"));
+    fn generates_bypass_commands_with_proxy_rule() {
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, None);
+        let text = all_strs(&cmds);
+        assert!(text.contains("add table inet openshell_bypass"));
+        assert!(text.contains("add chain inet openshell_bypass output"));
+        assert!(text.contains("ip daddr 10.0.2.2 tcp dport 8080 accept"));
+    }
+
+    #[test]
+    fn bypass_commands_have_table_and_chain() {
+        let cmds = generate_bypass_commands("192.168.1.1", 3128, None);
+        let text = all_strs(&cmds);
+        assert!(text.contains("add table inet openshell_bypass"));
+        assert!(text.contains("type filter hook output priority 0; policy accept;"));
     }
 
     #[test]
     fn proxy_accept_rule_uses_provided_ip_and_port() {
-        let ruleset = generate_bypass_ruleset("172.16.0.1", 9999, None);
-        assert!(ruleset.contains("ip daddr 172.16.0.1 tcp dport 9999 accept"));
+        let cmds = generate_bypass_commands("172.16.0.1", 9999, None);
+        let text = all_strs(&cmds);
+        assert!(text.contains("ip daddr 172.16.0.1 tcp dport 9999 accept"));
     }
 
     #[test]
     fn rules_are_ordered_accept_then_reject() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, None);
-        let proxy_pos = ruleset.find("ip daddr").unwrap();
-        let lo_pos = ruleset.find("oifname \"lo\"").unwrap();
-        let ct_pos = ruleset.find("ct state established,related").unwrap();
-        let reject_pos = ruleset.find("reject with icmp type").unwrap();
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, None);
+        let text = all_strs(&cmds);
+        let proxy_pos = text.find("ip daddr").unwrap();
+        let lo_pos = text.find("oifname lo").unwrap();
+        let ct_pos = text.find("ct state established,related").unwrap();
+        let reject_pos = text.find("reject with icmp type").unwrap();
 
         assert!(proxy_pos < lo_pos);
         assert!(lo_pos < ct_pos);
@@ -93,11 +425,12 @@ mod tests {
 
     #[test]
     fn both_ipv4_and_ipv6_reject_types_are_present() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, None);
-        let icmp_count = ruleset
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, None);
+        let text = all_strs(&cmds);
+        let icmp_count = text
             .matches("reject with icmp type port-unreachable")
             .count();
-        let icmpv6_count = ruleset
+        let icmpv6_count = text
             .matches("reject with icmpv6 type port-unreachable")
             .count();
         assert_eq!(icmp_count, 2, "need IPv4 ICMP rejects for TCP + UDP");
@@ -105,34 +438,35 @@ mod tests {
     }
 
     #[test]
-    fn no_log_ruleset_omits_log_rules() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, None);
+    fn no_log_commands_omit_log_rules() {
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, None);
+        let text = all_strs(&cmds);
         assert!(
-            !ruleset.contains("log prefix"),
-            "no-log ruleset must not contain log rules"
+            !text.contains("log prefix"),
+            "no-log commands must not contain log rules"
         );
     }
 
     #[test]
-    fn log_ruleset_contains_prefix_for_tcp_and_udp() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, Some("openshell:bypass:test:"));
-        let count = ruleset
-            .matches("log prefix \"openshell:bypass:test:\"")
-            .count();
+    fn log_commands_contain_prefix_for_tcp_and_udp() {
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, Some("openshell:bypass:test:"));
+        let text = all_strs(&cmds);
+        let count = text.matches("log prefix openshell:bypass:test:").count();
         assert_eq!(count, 2, "need log rules for both TCP and UDP");
-        assert!(ruleset.contains("tcp flags syn limit rate 5/second burst 10 packets"));
-        assert!(ruleset.contains("meta l4proto udp limit rate 5/second burst 10 packets"));
+        assert!(text.contains("tcp flags syn limit rate 5/second burst 10 packets"));
+        assert!(text.contains("meta l4proto udp limit rate 5/second burst 10 packets"));
     }
 
     #[test]
     fn log_rules_appear_before_reject_rules() {
-        let ruleset = generate_bypass_ruleset("10.0.2.2", 8080, Some("openshell:bypass:test:"));
-        let tcp_log_pos = ruleset.find("tcp flags syn").unwrap();
-        let tcp_reject_pos = ruleset
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, Some("openshell:bypass:test:"));
+        let text = all_strs(&cmds);
+        let tcp_log_pos = text.find("tcp flags syn").unwrap();
+        let tcp_reject_pos = text
             .find("meta nfproto ipv4 meta l4proto tcp reject")
             .unwrap();
-        let udp_log_pos = ruleset.find("meta l4proto udp limit rate").unwrap();
-        let udp_reject_pos = ruleset
+        let udp_log_pos = text.find("meta l4proto udp limit rate").unwrap();
+        let udp_reject_pos = text
             .find("meta nfproto ipv4 meta l4proto udp reject")
             .unwrap();
 
@@ -145,4 +479,53 @@ mod tests {
             "UDP log rule must come before UDP reject rule"
         );
     }
+
+    #[test]
+    fn ct_state_rule_is_not_required() {
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, None);
+        let ct_cmd = cmds
+            .iter()
+            .find(|c| cmd_str(c).contains("ct state"))
+            .unwrap();
+        assert!(
+            !ct_cmd.required,
+            "ct state rule should be non-required (needs nf_conntrack)"
+        );
+    }
+
+    #[test]
+    fn log_rules_are_not_required() {
+        let cmds = generate_bypass_commands("10.0.2.2", 8080, Some("openshell:bypass:test:"));
+        for cmd in &cmds {
+            if cmd_str(cmd).contains("log prefix") {
+                assert!(
+                    !cmd.required,
+                    "log rules should be non-required (needs nf_log)"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn sidecar_commands_allow_supervisor_uid_and_loopback() {
+        let cmds = generate_sidecar_bypass_commands(1337, None);
+        let text = all_strs(&cmds);
+        assert!(text.contains("add table inet openshell_sidecar_bypass"));
+        assert!(text.contains("oifname lo accept"));
+        assert!(text.contains("meta skuid 1337 accept"));
+    }
+
+    #[test]
+    fn sidecar_commands_reject_tcp_and_udp_egress() {
+        let cmds = generate_sidecar_bypass_commands(0, Some("openshell:sidecar:test:"));
+        let text = all_strs(&cmds);
+        assert!(text.contains("meta nfproto ipv4 meta l4proto tcp reject"));
+        assert!(text.contains("meta nfproto ipv6 meta l4proto tcp reject"));
+        assert!(text.contains("meta nfproto ipv4 meta l4proto udp reject"));
+        assert!(text.contains("meta nfproto ipv6 meta l4proto udp reject"));
+        assert_eq!(
+            text.matches("log prefix openshell:sidecar:test:").count(),
+            2
+        );
+    }
 }

crates/openshell-supervisor-process/src/process.rs

@@ -28,10 +28,55 @@ use std::sync::OnceLock;
 use tokio::process::{Child, Command};
 use tracing::{debug, info};
 
+/// Process/filesystem enforcement performed by the process supervisor.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ProcessEnforcementMode {
+    /// Preserve the existing supervisor behavior: prepare filesystem policy,
+    /// drop privileges, and apply Landlock/seccomp to workload processes.
+    Full,
+    /// Preserve process launch and SSH/session behavior, but skip controls
+    /// that require root or extra Linux capabilities. Kubernetes sidecar mode
+    /// uses this when network policy is enforced by the network sidecar.
+    NetworkOnly,
+}
+
+impl ProcessEnforcementMode {
+    #[must_use]
+    pub const fn uses_privileged_process_setup(self) -> bool {
+        matches!(self, Self::Full)
+    }
+
+    #[must_use]
+    pub const fn enforces_child_sandbox(self) -> bool {
+        matches!(self, Self::Full | Self::NetworkOnly)
+    }
+}
+
+#[cfg(target_os = "linux")]
+pub(crate) fn prepare_child_sandbox(
+    policy: &SandboxPolicy,
+    workdir: Option<&str>,
+    enforcement_mode: ProcessEnforcementMode,
+) -> Result<Option<sandbox::linux::PreparedSandbox>> {
+    if !enforcement_mode.enforces_child_sandbox() {
+        return Ok(None);
+    }
+
+    let prepared = if enforcement_mode.uses_privileged_process_setup() {
+        sandbox::linux::prepare(policy, workdir)
+    } else {
+        sandbox::linux::prepare_current_user(policy, workdir)
+    }?;
+    Ok(Some(prepared))
+}
+
 const SUPERVISOR_ONLY_ENV_VARS: &[&str] = &[
     openshell_core::sandbox_env::SANDBOX_TOKEN,
     openshell_core::sandbox_env::SANDBOX_TOKEN_FILE,
     openshell_core::sandbox_env::K8S_SA_TOKEN_FILE,
+    openshell_core::sandbox_env::TLS_CA,
+    openshell_core::sandbox_env::TLS_CERT,
+    openshell_core::sandbox_env::TLS_KEY,
     openshell_core::sandbox_env::PROVIDER_SPIFFE_WORKLOAD_API_SOCKET,
 ];
 
@@ -443,6 +488,7 @@ impl ProcessHandle {
         workdir: Option<&str>,
         interactive: bool,
         policy: &SandboxPolicy,
+        enforcement_mode: ProcessEnforcementMode,
         netns: Option<&NetworkNamespace>,
         ca_paths: Option<&(PathBuf, PathBuf)>,
         provider_env: &HashMap<String, String>,
@@ -453,6 +499,7 @@ impl ProcessHandle {
             workdir,
             interactive,
             policy,
+            enforcement_mode,
             netns.and_then(NetworkNamespace::ns_fd),
             ca_paths,
             provider_env,
@@ -465,12 +512,14 @@ impl ProcessHandle {
     ///
     /// Returns an error if the process fails to start.
     #[cfg(not(target_os = "linux"))]
+    #[allow(clippy::too_many_arguments)]
     pub fn spawn(
         program: &str,
         args: &[String],
         workdir: Option<&str>,
         interactive: bool,
         policy: &SandboxPolicy,
+        enforcement_mode: ProcessEnforcementMode,
         ca_paths: Option<&(PathBuf, PathBuf)>,
         provider_env: &HashMap<String, String>,
     ) -> Result<Self> {
@@ -480,6 +529,7 @@ impl ProcessHandle {
             workdir,
             interactive,
             policy,
+            enforcement_mode,
             ca_paths,
             provider_env,
         )
@@ -493,6 +543,7 @@ impl ProcessHandle {
         workdir: Option<&str>,
         interactive: bool,
         policy: &SandboxPolicy,
+        enforcement_mode: ProcessEnforcementMode,
         netns_fd: Option<RawFd>,
         ca_paths: Option<&(PathBuf, PathBuf)>,
         provider_env: &HashMap<String, String>,
@@ -552,18 +603,26 @@ impl ProcessHandle {
         // process where the tracing subscriber is functional. The child's
         // pre_exec context cannot reliably emit structured logs.
         #[cfg(target_os = "linux")]
-        sandbox::linux::log_sandbox_readiness(policy, workdir);
+        if enforcement_mode.enforces_child_sandbox() {
+            sandbox::linux::log_sandbox_readiness(policy, workdir);
+        }
 
-        // Phase 1 (as root): Prepare Landlock ruleset by opening PathFds.
-        // This MUST happen before drop_privileges() so that root-only paths
-        // (e.g. mode 700 directories) can be opened. See issue #803.
+        // Phase 1: Prepare Landlock ruleset by opening PathFds.
+        // In full mode this runs before drop_privileges() so root-only paths
+        // can be opened. In sidecar network-only mode the container already
+        // runs as the sandbox UID, so inaccessible paths are unavailable to
+        // the workload and best-effort compatibility skips them.
         #[cfg(target_os = "linux")]
-        let prepared_sandbox = sandbox::linux::prepare(policy, workdir)
+        let prepared_sandbox = prepare_child_sandbox(policy, workdir, enforcement_mode)
             .map_err(|err| miette::miette!("Failed to prepare sandbox: {err}"))?;
         #[cfg(target_os = "linux")]
-        let supervisor_identity_mount = supervisor_identity_mount_from_env().map_err(|err| {
-            miette::miette!("Failed to prepare supervisor identity isolation: {err}")
-        })?;
+        let supervisor_identity_mount = if enforcement_mode.uses_privileged_process_setup() {
+            supervisor_identity_mount_from_env().map_err(|err| {
+                miette::miette!("Failed to prepare supervisor identity isolation: {err}")
+            })?
+        } else {
+            None
+        };
 
         // Set up process group for signal handling (non-interactive mode only).
         // In interactive mode, we inherit the parent's process group to maintain
@@ -575,7 +634,7 @@ impl ProcessHandle {
             // Wrap in Option so we can .take() it out of the FnMut closure.
             // pre_exec is only called once (after fork, before exec).
             #[cfg(target_os = "linux")]
-            let mut prepared_sandbox = Some(prepared_sandbox);
+            let mut prepared_sandbox = prepared_sandbox;
             #[allow(unsafe_code)]
             unsafe {
                 cmd.pre_exec(move || {
@@ -600,8 +659,10 @@ impl ProcessHandle {
                     // Drop privileges. initgroups/setgid/setuid need access to
                     // /etc/group and /etc/passwd which would be blocked if
                     // Landlock were already enforced.
-                    drop_privileges(&policy)
-                        .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    if enforcement_mode.uses_privileged_process_setup() {
+                        drop_privileges(&policy)
+                            .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    }
 
                     harden_child_process().map_err(|err| std::io::Error::other(err.to_string()))?;
 
@@ -629,12 +690,14 @@ impl ProcessHandle {
     }
 
     #[cfg(not(target_os = "linux"))]
+    #[allow(clippy::too_many_arguments)]
     fn spawn_impl(
         program: &str,
         args: &[String],
         workdir: Option<&str>,
         interactive: bool,
         policy: &SandboxPolicy,
+        enforcement_mode: ProcessEnforcementMode,
         ca_paths: Option<&(PathBuf, PathBuf)>,
         provider_env: &HashMap<String, String>,
     ) -> Result<Self> {
@@ -697,13 +760,17 @@ impl ProcessHandle {
                     // Drop privileges before applying sandbox restrictions.
                     // initgroups/setgid/setuid need access to /etc/group and /etc/passwd
                     // which may be blocked by Landlock.
-                    drop_privileges(&policy)
-                        .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    if enforcement_mode.uses_privileged_process_setup() {
+                        drop_privileges(&policy)
+                            .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    }
 
                     harden_child_process().map_err(|err| std::io::Error::other(err.to_string()))?;
 
-                    sandbox::apply(&policy, workdir.as_deref())
-                        .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    if enforcement_mode.enforces_child_sandbox() {
+                        sandbox::apply(&policy, workdir.as_deref())
+                            .map_err(|err| std::io::Error::other(err.to_string()))?;
+                    }
 
                     Ok(())
                 });
@@ -1456,6 +1523,18 @@ mod tests {
         );
     }
 
+    #[test]
+    fn full_enforcement_uses_privileged_setup_and_child_sandbox() {
+        assert!(ProcessEnforcementMode::Full.uses_privileged_process_setup());
+        assert!(ProcessEnforcementMode::Full.enforces_child_sandbox());
+    }
+
+    #[test]
+    fn network_only_enforcement_keeps_child_sandbox_without_privileged_setup() {
+        assert!(!ProcessEnforcementMode::NetworkOnly.uses_privileged_process_setup());
+        assert!(ProcessEnforcementMode::NetworkOnly.enforces_child_sandbox());
+    }
+
     #[cfg(target_os = "linux")]
     fn capability_bounding_set_clear_available() -> bool {
         capctl::caps::CapState::get_current()

crates/openshell-supervisor-process/src/run.rs

@@ -33,7 +33,7 @@ use openshell_core::denial::DenialEvent;
 
 #[cfg(target_os = "linux")]
 use crate::managed_children;
-use crate::process::ProcessHandle;
+use crate::process::{ProcessEnforcementMode, ProcessHandle};
 
 fn ocsf_ctx() -> &'static openshell_ocsf::SandboxContext {
     openshell_ocsf::ctx::ctx()
@@ -56,8 +56,11 @@ pub async fn run_process(
     sandbox_id: Option<&str>,
     openshell_endpoint: Option<&str>,
     ssh_socket_path: Option<String>,
+    shared_ssh_socket: bool,
     policy: &SandboxPolicy,
+    enforcement_mode: ProcessEnforcementMode,
     entrypoint_pid: Arc<AtomicU32>,
+    entrypoint_started_tx: Option<tokio::sync::oneshot::Sender<u32>>,
     provider_credentials: ProviderCredentialState,
     provider_env: std::collections::HashMap<String, String>,
     ca_file_paths: Option<(std::path::PathBuf, std::path::PathBuf)>,
@@ -71,21 +74,26 @@ pub async fn run_process(
     // /etc/group so the "sandbox" entry matches. Must run before
     // validate_sandbox_user so passwd lookups see the correct identity.
     #[cfg(unix)]
-    crate::process::update_sandbox_passwd_entries()?;
+    if enforcement_mode.uses_privileged_process_setup() {
+        crate::process::update_sandbox_passwd_entries()?;
+    }
 
     // Validate that the sandbox user exists in the image. All sandbox images
     // must include a "sandbox" user for privilege dropping; failing fast here
     // beats silently running children as root.
     #[cfg(unix)]
-    crate::process::validate_sandbox_user(policy)?;
-    #[cfg(unix)]
-    crate::process::validate_sandbox_group(policy)?;
+    if enforcement_mode.uses_privileged_process_setup() {
+        crate::process::validate_sandbox_user(policy)?;
+        crate::process::validate_sandbox_group(policy)?;
+    }
 
     // Create read_write directories and chown newly-created ones to the
     // sandbox user/group. Runs as the supervisor (root) before the child
     // is forked so the workload sees writable paths it owns.
     #[cfg(unix)]
-    crate::process::prepare_filesystem(policy)?;
+    if enforcement_mode.uses_privileged_process_setup() {
+        crate::process::prepare_filesystem(policy)?;
+    }
 
     // Eagerly fetch initial settings and install the agent skill if the
     // proposals flag is on at startup, rather than waiting for the policy
@@ -206,31 +214,10 @@ pub async fn run_process(
     // their env so cooperative tools (curl, npm, Node) route through the
     // CONNECT proxy. Linux uses the netns host_ip; on other targets fall back
     // to the policy-declared http_addr directly.
-    let ssh_proxy_url = if matches!(policy.network.mode, NetworkMode::Proxy) {
-        #[cfg(target_os = "linux")]
-        {
-            netns.map(|ns| {
-                let port = policy
-                    .network
-                    .proxy
-                    .as_ref()
-                    .and_then(|p| p.http_addr)
-                    .map_or(3128, |addr| addr.port());
-                format!("http://{}:{port}", ns.host_ip())
-            })
-        }
-        #[cfg(not(target_os = "linux"))]
-        {
-            policy
-                .network
-                .proxy
-                .as_ref()
-                .and_then(|p| p.http_addr)
-                .map(|addr| format!("http://{addr}"))
-        }
-    } else {
-        None
-    };
+    #[cfg(target_os = "linux")]
+    let ssh_proxy_url = ssh_proxy_url_for_policy(policy, netns.map(NetworkNamespace::host_ip));
+    #[cfg(not(target_os = "linux"))]
+    let ssh_proxy_url = ssh_proxy_url_for_policy(policy, None);
 
     let ssh_socket_path: Option<std::path::PathBuf> = ssh_socket_path.map(std::path::PathBuf::from);
     if let Some(listen_path) = ssh_socket_path.clone() {
@@ -259,6 +246,8 @@ pub async fn run_process(
                 ca_paths,
                 provider_credentials_clone,
                 user_env_clone,
+                enforcement_mode,
+                shared_ssh_socket,
             )
             .await
             {
@@ -314,6 +303,7 @@ pub async fn run_process(
             id.to_string(),
             socket.clone(),
             ssh_netns_fd,
+            None,
         );
         info!("supervisor session task spawned");
     }
@@ -325,6 +315,7 @@ pub async fn run_process(
         workdir,
         interactive,
         policy,
+        enforcement_mode,
         netns,
         ca_file_paths.as_ref(),
         &provider_env,
@@ -337,12 +328,16 @@ pub async fn run_process(
         workdir,
         interactive,
         policy,
+        enforcement_mode,
         ca_file_paths.as_ref(),
         &provider_env,
     )?;
 
     // Store the entrypoint PID so the proxy can resolve TCP peer identity
     entrypoint_pid.store(handle.pid(), Ordering::Release);
+    if let Some(tx) = entrypoint_started_tx {
+        let _ = tx.send(handle.pid());
+    }
     ocsf_emit!(
         ProcessActivityBuilder::new(ocsf_ctx())
             .activity(ActivityId::Open)
@@ -395,6 +390,23 @@ pub async fn run_process(
     Ok(status.code())
 }
 
+fn ssh_proxy_url_for_policy(
+    policy: &SandboxPolicy,
+    netns_proxy_host: Option<std::net::IpAddr>,
+) -> Option<String> {
+    if !matches!(policy.network.mode, NetworkMode::Proxy) {
+        return None;
+    }
+
+    let proxy = policy.network.proxy.as_ref()?;
+    if let Some(host) = netns_proxy_host {
+        let port = proxy.http_addr.map_or(3128, |addr| addr.port());
+        return Some(format!("http://{host}:{port}"));
+    }
+
+    proxy.http_addr.map(|addr| format!("http://{addr}"))
+}
+
 /// Eagerly fetch initial settings and install the agent-driven policy
 /// proposal skill if the flag is on at startup.
 ///
@@ -451,3 +463,53 @@ async fn install_initial_agent_skill(sandbox_id: Option<&str>, openshell_endpoin
         );
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openshell_core::policy::{
+        FilesystemPolicy, LandlockPolicy, NetworkMode, NetworkPolicy, ProcessPolicy, ProxyPolicy,
+    };
+
+    fn policy(mode: NetworkMode, http_addr: Option<std::net::SocketAddr>) -> SandboxPolicy {
+        SandboxPolicy {
+            version: 1,
+            filesystem: FilesystemPolicy::default(),
+            network: NetworkPolicy {
+                mode,
+                proxy: http_addr.map(|http_addr| ProxyPolicy {
+                    http_addr: Some(http_addr),
+                }),
+            },
+            landlock: LandlockPolicy::default(),
+            process: ProcessPolicy::default(),
+        }
+    }
+
+    #[test]
+    fn ssh_proxy_url_uses_policy_addr_without_netns() {
+        let policy = policy(NetworkMode::Proxy, Some(([127, 0, 0, 1], 3128).into()));
+
+        assert_eq!(
+            ssh_proxy_url_for_policy(&policy, None).as_deref(),
+            Some("http://127.0.0.1:3128")
+        );
+    }
+
+    #[test]
+    fn ssh_proxy_url_prefers_netns_host_with_policy_port() {
+        let policy = policy(NetworkMode::Proxy, Some(([127, 0, 0, 1], 8080).into()));
+
+        assert_eq!(
+            ssh_proxy_url_for_policy(&policy, Some([10, 200, 0, 1].into())).as_deref(),
+            Some("http://10.200.0.1:8080")
+        );
+    }
+
+    #[test]
+    fn ssh_proxy_url_skips_non_proxy_mode() {
+        let policy = policy(NetworkMode::Allow, Some(([127, 0, 0, 1], 3128).into()));
+
+        assert_eq!(ssh_proxy_url_for_policy(&policy, None), None);
+    }
+}

crates/openshell-supervisor-process/src/sandbox/linux/landlock.rs

@@ -95,6 +95,12 @@ pub struct PreparedRuleset {
     compatibility: LandlockCompatibility,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum PathOpenMode {
+    Privileged,
+    CurrentUser,
+}
+
 /// Phase 1: Open `PathFds` and build the Landlock ruleset **as root**.
 ///
 /// This must run before `drop_privileges()` so that `PathFd::new()` can open
@@ -103,6 +109,27 @@ pub struct PreparedRuleset {
 /// Returns `None` if there are no filesystem paths to restrict (no-op).
 /// Returns `Some(PreparedRuleset)` on success, or an error.
 pub fn prepare(policy: &SandboxPolicy, workdir: Option<&str>) -> Result<Option<PreparedRuleset>> {
+    prepare_with_path_open_mode(policy, workdir, PathOpenMode::Privileged)
+}
+
+/// Phase 1 for already-unprivileged workloads.
+///
+/// Kubernetes sidecar mode starts the process supervisor as the sandbox UID, so
+/// Landlock path FDs are opened as the same UID that will run the workload.
+/// Paths this UID cannot open are already unavailable to the workload; omit
+/// them from the allowlist and let the resulting ruleset deny everything else.
+pub fn prepare_current_user(
+    policy: &SandboxPolicy,
+    workdir: Option<&str>,
+) -> Result<Option<PreparedRuleset>> {
+    prepare_with_path_open_mode(policy, workdir, PathOpenMode::CurrentUser)
+}
+
+fn prepare_with_path_open_mode(
+    policy: &SandboxPolicy,
+    workdir: Option<&str>,
+    path_open_mode: PathOpenMode,
+) -> Result<Option<PreparedRuleset>> {
     let read_only = policy.filesystem.read_only.clone();
     let mut read_write = policy.filesystem.read_write.clone();
 
@@ -188,7 +215,7 @@ pub fn prepare(policy: &SandboxPolicy, workdir: Option<&str>) -> Result<Option<P
         let mut rules_applied: usize = 0;
 
         for path in &read_only {
-            if let Some(path_fd) = try_open_path(path, compatibility)? {
+            if let Some(path_fd) = try_open_path(path, compatibility, path_open_mode)? {
                 debug!(path = %path.display(), "Landlock allow read-only");
                 ruleset = ruleset
                     .add_rule(PathBeneath::new(path_fd, access_read))
@@ -198,7 +225,7 @@ pub fn prepare(policy: &SandboxPolicy, workdir: Option<&str>) -> Result<Option<P
         }
 
         for path in &read_write {
-            if let Some(path_fd) = try_open_path(path, compatibility)? {
+            if let Some(path_fd) = try_open_path(path, compatibility, path_open_mode)? {
                 debug!(path = %path.display(), "Landlock allow read-write");
                 ruleset = ruleset
                     .add_rule(PathBeneath::new(path_fd, access_all))
@@ -325,7 +352,11 @@ pub fn apply(policy: &SandboxPolicy, workdir: Option<&str>) -> Result<()> {
 ///
 /// In `HardRequirement` mode, any failure is fatal — the caller propagates the
 /// error, which ultimately aborts sandbox startup.
-fn try_open_path(path: &Path, compatibility: &LandlockCompatibility) -> Result<Option<PathFd>> {
+fn try_open_path(
+    path: &Path,
+    compatibility: &LandlockCompatibility,
+    path_open_mode: PathOpenMode,
+) -> Result<Option<PathFd>> {
     match PathFd::new(path) {
         Ok(fd) => Ok(Some(fd)),
         Err(err) => {
@@ -335,6 +366,28 @@ fn try_open_path(path: &Path, compatibility: &LandlockCompatibility) -> Result<O
                 PathFdError::OpenCall { source, .. }
                     if source.kind() == std::io::ErrorKind::NotFound
             );
+            if matches!(path_open_mode, PathOpenMode::CurrentUser) {
+                if is_not_found {
+                    debug!(
+                        path = %path.display(),
+                        reason,
+                        "Skipping non-existent Landlock path for current user"
+                    );
+                } else {
+                    openshell_ocsf::ocsf_emit!(
+                        openshell_ocsf::ConfigStateChangeBuilder::new(openshell_ocsf::ctx::ctx())
+                            .severity(openshell_ocsf::SeverityId::Informational)
+                            .status(openshell_ocsf::StatusId::Success)
+                            .state(openshell_ocsf::StateId::Other, "already-denied")
+                            .message(format!(
+                                "Skipping inaccessible Landlock path for current user [path:{} error:{err}]",
+                                path.display()
+                            ))
+                            .build()
+                    );
+                }
+                return Ok(None);
+            }
             match compatibility {
                 LandlockCompatibility::BestEffort => {
                     // NotFound is expected for stale baseline paths (e.g.
@@ -421,6 +474,7 @@ mod tests {
         let result = try_open_path(
             &PathBuf::from("/nonexistent/openshell/test/path"),
             &LandlockCompatibility::BestEffort,
+            PathOpenMode::Privileged,
         );
         assert!(result.is_ok());
         assert!(result.unwrap().is_none());
@@ -431,6 +485,7 @@ mod tests {
         let result = try_open_path(
             &PathBuf::from("/nonexistent/openshell/test/path"),
             &LandlockCompatibility::HardRequirement,
+            PathOpenMode::Privileged,
         );
         assert!(result.is_err());
         let err_msg = result.unwrap_err().to_string();
@@ -447,11 +502,26 @@ mod tests {
     #[test]
     fn try_open_path_succeeds_for_existing_path() {
         let dir = tempfile::tempdir().unwrap();
-        let result = try_open_path(dir.path(), &LandlockCompatibility::BestEffort);
+        let result = try_open_path(
+            dir.path(),
+            &LandlockCompatibility::BestEffort,
+            PathOpenMode::Privileged,
+        );
         assert!(result.is_ok());
         assert!(result.unwrap().is_some());
     }
 
+    #[test]
+    fn try_open_path_current_user_skips_missing_path_in_hard_requirement() {
+        let result = try_open_path(
+            &PathBuf::from("/nonexistent/openshell/test/path"),
+            &LandlockCompatibility::HardRequirement,
+            PathOpenMode::CurrentUser,
+        );
+        assert!(result.is_ok());
+        assert!(result.unwrap().is_none());
+    }
+
     #[test]
     fn classify_not_found() {
         let err = std::io::Error::from_raw_os_error(libc::ENOENT);

crates/openshell-supervisor-process/src/sandbox/linux/mod.rs

@@ -12,7 +12,7 @@ use std::path::PathBuf;
 use std::sync::Once;
 
 /// Opaque handle to a prepared-but-not-yet-enforced sandbox.
-/// Holds the Landlock ruleset with `PathFds` opened as root.
+/// Holds the Landlock ruleset with `PathFds` opened before child exec.
 pub struct PreparedSandbox {
     landlock: Option<landlock::PreparedRuleset>,
     policy: SandboxPolicy,
@@ -30,6 +30,21 @@ pub fn prepare(policy: &SandboxPolicy, workdir: Option<&str>) -> Result<Prepared
     })
 }
 
+/// Phase 1 for already-unprivileged workloads.
+///
+/// Opens Landlock `PathFds` as the current UID. This is used by Kubernetes
+/// sidecar mode, where the agent container already runs as the sandbox user.
+pub fn prepare_current_user(
+    policy: &SandboxPolicy,
+    workdir: Option<&str>,
+) -> Result<PreparedSandbox> {
+    let landlock = landlock::prepare_current_user(policy, workdir)?;
+    Ok(PreparedSandbox {
+        landlock,
+        policy: policy.clone(),
+    })
+}
+
 /// Phase 2: Enforce prepared sandbox restrictions (after `drop_privileges`).
 ///
 /// Calls `restrict_self()` for Landlock and applies seccomp filters.

crates/openshell-supervisor-process/src/ssh.rs

@@ -6,7 +6,7 @@
 use crate::child_env;
 #[cfg(target_os = "linux")]
 use crate::managed_children;
-use crate::process::{drop_privileges, is_supervisor_only_env_var};
+use crate::process::{ProcessEnforcementMode, drop_privileges, is_supervisor_only_env_var};
 use crate::sandbox;
 use miette::{IntoDiagnostic, Result};
 use nix::pty::{Winsize, openpty};
@@ -42,6 +42,8 @@ type SshServerInit = (
 fn ssh_server_init(
     listen_path: &Path,
     ca_file_paths: &Option<(PathBuf, PathBuf)>,
+    enforcement_mode: ProcessEnforcementMode,
+    shared_socket: bool,
 ) -> Result<SshServerInit> {
     let mut rng = OsRng;
     let host_key = PrivateKey::random(&mut rng, Algorithm::Ed25519).into_diagnostic()?;
@@ -55,13 +57,17 @@ fn ssh_server_init(
     let config = Arc::new(config);
     let ca_paths = ca_file_paths.as_ref().map(|p| Arc::new(p.clone()));
 
-    // Ensure the parent directory exists and is root-owned with 0700
-    // permissions. The sandbox entrypoint runs as an unprivileged user; it
-    // must not be able to enter this directory and connect to the socket.
-    if let Some(parent) = listen_path.parent() {
+    // In full enforcement mode the supervisor normally starts as root and can
+    // isolate the SSH socket in a root-only directory before spawning
+    // unprivileged children. Sidecar topology is different: the gateway relay
+    // runs in the network sidecar as a different UID, so the shared sidecar
+    // state directory must stay group-accessible. Sidecar mode uses a Linux
+    // abstract socket instead, so the workload cannot unlink the relay target.
+    let abstract_socket = crate::unix_socket::is_abstract(listen_path);
+    if !abstract_socket && let Some(parent) = listen_path.parent() {
         std::fs::create_dir_all(parent).into_diagnostic()?;
         #[cfg(unix)]
-        {
+        if enforcement_mode.uses_privileged_process_setup() && !shared_socket {
             use std::os::unix::fs::PermissionsExt;
             let perms = std::fs::Permissions::from_mode(0o700);
             std::fs::set_permissions(parent, perms).into_diagnostic()?;
@@ -69,19 +75,19 @@ fn ssh_server_init(
     }
 
     // Remove any stale socket from a previous run before binding.
-    if listen_path.exists() {
+    if !abstract_socket && listen_path.exists() {
         std::fs::remove_file(listen_path).into_diagnostic()?;
     }
-    let listener = UnixListener::bind(listen_path).into_diagnostic()?;
+    let runtime_path = crate::unix_socket::runtime_path(listen_path);
+    let listener = UnixListener::bind(runtime_path.as_ref()).into_diagnostic()?;
 
-    // Tighten permissions so only the supervisor (root) can connect. The
-    // sandbox entrypoint runs as an unprivileged user and must not be able to
-    // dial the SSH daemon directly — all access goes through the relay from
-    // the gateway.
+    // Tighten filesystem-socket permissions. Abstract sockets have no inode;
+    // sidecar relay connections authenticate the listener with SO_PEERCRED.
     #[cfg(unix)]
-    {
+    if !abstract_socket {
         use std::os::unix::fs::PermissionsExt;
-        let perms = std::fs::Permissions::from_mode(0o600);
+        let mode = if shared_socket { 0o660 } else { 0o600 };
+        let perms = std::fs::Permissions::from_mode(mode);
         std::fs::set_permissions(listen_path, perms).into_diagnostic()?;
     }
 
@@ -108,8 +114,15 @@ pub async fn run_ssh_server(
     ca_file_paths: Option<(PathBuf, PathBuf)>,
     provider_credentials: ProviderCredentialState,
     user_environment: HashMap<String, String>,
+    enforcement_mode: ProcessEnforcementMode,
+    shared_socket: bool,
 ) -> Result<()> {
-    let (listener, config, ca_paths) = match ssh_server_init(&listen_path, &ca_file_paths) {
+    let (listener, config, ca_paths) = match ssh_server_init(
+        &listen_path,
+        &ca_file_paths,
+        enforcement_mode,
+        shared_socket,
+    ) {
         Ok(v) => {
             // Signal that the SSH server has bound the socket and is ready to
             // accept connections. The parent task awaits this before spawning
@@ -145,6 +158,7 @@ pub async fn run_ssh_server(
                 ca_paths,
                 provider_credentials,
                 user_environment,
+                enforcement_mode,
             )
             .await
             {
@@ -172,6 +186,7 @@ async fn handle_connection(
     ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
     provider_credentials: ProviderCredentialState,
     user_environment: HashMap<String, String>,
+    enforcement_mode: ProcessEnforcementMode,
 ) -> Result<()> {
     // Access is gated by the Unix-socket filesystem permissions (root-only),
     // not by an application-level preface. The supervisor bridges the
@@ -195,6 +210,7 @@ async fn handle_connection(
         ca_file_paths,
         provider_credentials,
         user_environment,
+        enforcement_mode,
     );
     russh::server::run_stream(config, stream, handler)
         .await
@@ -223,6 +239,7 @@ struct SshHandler {
     ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
     provider_credentials: ProviderCredentialState,
     user_environment: HashMap<String, String>,
+    enforcement_mode: ProcessEnforcementMode,
     channels: HashMap<ChannelId, ChannelState>,
 }
 
@@ -236,6 +253,7 @@ impl SshHandler {
         ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
         provider_credentials: ProviderCredentialState,
         user_environment: HashMap<String, String>,
+        enforcement_mode: ProcessEnforcementMode,
     ) -> Self {
         Self {
             policy,
@@ -245,6 +263,7 @@ impl SshHandler {
             ca_file_paths,
             provider_credentials,
             user_environment,
+            enforcement_mode,
             channels: HashMap::new(),
         }
     }
@@ -468,6 +487,7 @@ impl russh::server::Handler for SshHandler {
                 self.ca_file_paths.clone(),
                 &self.provider_credentials.child_env_with_gcp_resolved(),
                 &self.user_environment,
+                self.enforcement_mode,
             )?;
             let state = self.channels.get_mut(&channel).ok_or_else(|| {
                 anyhow::anyhow!("subsystem_request on unknown channel {channel:?}")
@@ -564,6 +584,7 @@ impl SshHandler {
                 self.ca_file_paths.clone(),
                 &provider_env,
                 &self.user_environment,
+                self.enforcement_mode,
             )?;
             state.pty_master = Some(pty_master);
             state.input_sender = Some(input_sender);
@@ -582,6 +603,7 @@ impl SshHandler {
                 self.ca_file_paths.clone(),
                 &provider_env,
                 &self.user_environment,
+                self.enforcement_mode,
             )?;
             state.input_sender = Some(input_sender);
         }
@@ -748,6 +770,7 @@ fn spawn_pty_shell(
     ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
     provider_env: &HashMap<String, String>,
     user_environment: &HashMap<String, String>,
+    enforcement_mode: ProcessEnforcementMode,
 ) -> anyhow::Result<(std::fs::File, mpsc::Sender<Vec<u8>>)> {
     let winsize = Winsize {
         ws_row: to_u16(pty.row_height.max(1)),
@@ -806,12 +829,15 @@ fn spawn_pty_shell(
 
     // Probe Landlock availability from the parent process where tracing works.
     #[cfg(target_os = "linux")]
-    sandbox::linux::log_sandbox_readiness(policy, workdir.as_deref());
+    if enforcement_mode.enforces_child_sandbox() {
+        sandbox::linux::log_sandbox_readiness(policy, workdir.as_deref());
+    }
 
-    // Phase 1 (as root): Prepare Landlock ruleset before drop_privileges.
+    // Phase 1: Prepare Landlock ruleset before the child applies it.
     #[cfg(target_os = "linux")]
-    let prepared_sandbox = sandbox::linux::prepare(policy, workdir.as_deref())
-        .map_err(|err| anyhow::anyhow!("Failed to prepare sandbox: {err}"))?;
+    let prepared_sandbox =
+        crate::process::prepare_child_sandbox(policy, workdir.as_deref(), enforcement_mode)
+            .map_err(|err| anyhow::anyhow!("Failed to prepare sandbox: {err}"))?;
 
     #[cfg(unix)]
     {
@@ -821,6 +847,7 @@ fn spawn_pty_shell(
             workdir.clone(),
             slave_fd,
             netns_fd,
+            enforcement_mode,
             #[cfg(target_os = "linux")]
             prepared_sandbox,
         )?;
@@ -913,6 +940,7 @@ fn spawn_pipe_exec(
     ca_file_paths: Option<Arc<(PathBuf, PathBuf)>>,
     provider_env: &HashMap<String, String>,
     user_environment: &HashMap<String, String>,
+    enforcement_mode: ProcessEnforcementMode,
 ) -> anyhow::Result<mpsc::Sender<Vec<u8>>> {
     let mut cmd = command.map_or_else(
         || {
@@ -955,12 +983,15 @@ fn spawn_pipe_exec(
 
     // Probe Landlock availability from the parent process where tracing works.
     #[cfg(target_os = "linux")]
-    sandbox::linux::log_sandbox_readiness(policy, workdir.as_deref());
+    if enforcement_mode.enforces_child_sandbox() {
+        sandbox::linux::log_sandbox_readiness(policy, workdir.as_deref());
+    }
 
-    // Phase 1 (as root): Prepare Landlock ruleset before drop_privileges.
+    // Phase 1: Prepare Landlock ruleset before the child applies it.
     #[cfg(target_os = "linux")]
-    let prepared_sandbox = sandbox::linux::prepare(policy, workdir.as_deref())
-        .map_err(|err| anyhow::anyhow!("Failed to prepare sandbox: {err}"))?;
+    let prepared_sandbox =
+        crate::process::prepare_child_sandbox(policy, workdir.as_deref(), enforcement_mode)
+            .map_err(|err| anyhow::anyhow!("Failed to prepare sandbox: {err}"))?;
 
     #[cfg(unix)]
     {
@@ -969,6 +1000,7 @@ fn spawn_pipe_exec(
             policy.clone(),
             workdir.clone(),
             netns_fd,
+            enforcement_mode,
             #[cfg(target_os = "linux")]
             prepared_sandbox,
         )?;
@@ -1068,7 +1100,9 @@ fn spawn_pipe_exec(
 mod unsafe_pty {
     #[cfg(not(target_os = "linux"))]
     use super::sandbox;
-    use super::{Command, RawFd, SandboxPolicy, Winsize, drop_privileges, setsid};
+    use super::{
+        Command, ProcessEnforcementMode, RawFd, SandboxPolicy, Winsize, drop_privileges, setsid,
+    };
     #[cfg(unix)]
     use std::os::unix::process::CommandExt;
 
@@ -1107,17 +1141,21 @@ mod unsafe_pty {
         _workdir: Option<String>,
         slave_fd: RawFd,
         netns_fd: Option<RawFd>,
-        #[cfg(target_os = "linux")] prepared: crate::sandbox::linux::PreparedSandbox,
+        enforcement_mode: ProcessEnforcementMode,
+        #[cfg(target_os = "linux")] prepared: Option<crate::sandbox::linux::PreparedSandbox>,
     ) -> anyhow::Result<()> {
         // Wrap in Option so we can .take() it out of the FnMut closure.
         // pre_exec is only called once (after fork, before exec).
         #[cfg(target_os = "linux")]
-        let mut prepared = Some(prepared);
+        let mut prepared = prepared;
         #[cfg(target_os = "linux")]
-        let supervisor_identity_mount = crate::process::supervisor_identity_mount_from_env()
-            .map_err(|err| {
+        let supervisor_identity_mount = if enforcement_mode.uses_privileged_process_setup() {
+            crate::process::supervisor_identity_mount_from_env().map_err(|err| {
                 anyhow::anyhow!("failed to prepare supervisor identity isolation: {err}")
-            })?;
+            })?
+        } else {
+            None
+        };
         unsafe {
             cmd.pre_exec(move || {
                 setsid().map_err(|err| std::io::Error::other(err.to_string()))?;
@@ -1126,6 +1164,7 @@ mod unsafe_pty {
                 enter_netns_and_sandbox(
                     netns_fd,
                     &policy,
+                    enforcement_mode,
                     #[cfg(target_os = "linux")]
                     supervisor_identity_mount,
                     #[cfg(target_os = "linux")]
@@ -1152,20 +1191,25 @@ mod unsafe_pty {
         policy: SandboxPolicy,
         _workdir: Option<String>,
         netns_fd: Option<RawFd>,
-        #[cfg(target_os = "linux")] prepared: crate::sandbox::linux::PreparedSandbox,
+        enforcement_mode: ProcessEnforcementMode,
+        #[cfg(target_os = "linux")] prepared: Option<crate::sandbox::linux::PreparedSandbox>,
     ) -> anyhow::Result<()> {
         #[cfg(target_os = "linux")]
-        let mut prepared = Some(prepared);
+        let mut prepared = prepared;
         #[cfg(target_os = "linux")]
-        let supervisor_identity_mount = crate::process::supervisor_identity_mount_from_env()
-            .map_err(|err| {
+        let supervisor_identity_mount = if enforcement_mode.uses_privileged_process_setup() {
+            crate::process::supervisor_identity_mount_from_env().map_err(|err| {
                 anyhow::anyhow!("failed to prepare supervisor identity isolation: {err}")
-            })?;
+            })?
+        } else {
+            None
+        };
         unsafe {
             cmd.pre_exec(move || {
                 enter_netns_and_sandbox(
                     netns_fd,
                     &policy,
+                    enforcement_mode,
                     #[cfg(target_os = "linux")]
                     supervisor_identity_mount,
                     #[cfg(target_os = "linux")]
@@ -1179,6 +1223,7 @@ mod unsafe_pty {
     fn enter_netns_and_sandbox(
         netns_fd: Option<RawFd>,
         policy: &SandboxPolicy,
+        enforcement_mode: ProcessEnforcementMode,
         #[cfg(target_os = "linux")] supervisor_identity_mount: Option<
             &crate::process::SupervisorIdentityMountNamespace,
         >,
@@ -1207,7 +1252,9 @@ mod unsafe_pty {
 
         // Drop privileges. initgroups/setgid/setuid need /etc/group and
         // /etc/passwd which would be blocked if Landlock were already enforced.
-        drop_privileges(policy).map_err(|err| std::io::Error::other(err.to_string()))?;
+        if enforcement_mode.uses_privileged_process_setup() {
+            drop_privileges(policy).map_err(|err| std::io::Error::other(err.to_string()))?;
+        }
         crate::process::harden_child_process()
             .map_err(|err| std::io::Error::other(err.to_string()))?;
 
@@ -1220,7 +1267,9 @@ mod unsafe_pty {
         }
 
         #[cfg(not(target_os = "linux"))]
-        sandbox::apply(policy, None).map_err(|err| std::io::Error::other(err.to_string()))?;
+        if enforcement_mode.enforces_child_sandbox() {
+            sandbox::apply(policy, None).map_err(|err| std::io::Error::other(err.to_string()))?;
+        }
 
         Ok(())
     }
@@ -1275,6 +1324,71 @@ mod tests {
     use super::*;
     use std::process::Stdio;
 
+    #[cfg(unix)]
+    fn file_mode(path: &Path) -> u32 {
+        use std::os::unix::fs::PermissionsExt;
+        std::fs::metadata(path).unwrap().permissions().mode() & 0o7777
+    }
+
+    #[cfg(unix)]
+    fn set_file_mode(path: &Path, mode: u32) {
+        use std::os::unix::fs::PermissionsExt;
+        std::fs::set_permissions(path, std::fs::Permissions::from_mode(mode)).unwrap();
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn ssh_server_init_full_enforcement_keeps_private_socket() {
+        let temp = tempfile::tempdir().unwrap();
+        let parent = temp.path().join("ssh");
+        std::fs::create_dir_all(&parent).unwrap();
+        set_file_mode(&parent, 0o775);
+        let socket = parent.join("ssh.sock");
+
+        let (listener, _, _) =
+            ssh_server_init(&socket, &None, ProcessEnforcementMode::Full, false).unwrap();
+        drop(listener);
+
+        assert_eq!(file_mode(&parent), 0o700);
+        assert_eq!(file_mode(&socket), 0o600);
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn ssh_server_init_shared_socket_keeps_group_access() {
+        let temp = tempfile::tempdir().unwrap();
+        let parent = temp.path().join("ssh");
+        std::fs::create_dir_all(&parent).unwrap();
+        set_file_mode(&parent, 0o775);
+        let socket = parent.join("ssh.sock");
+
+        let (listener, _, _) =
+            ssh_server_init(&socket, &None, ProcessEnforcementMode::Full, true).unwrap();
+        drop(listener);
+
+        assert_eq!(file_mode(&parent), 0o775);
+        assert_eq!(file_mode(&socket), 0o660);
+    }
+
+    #[cfg(target_os = "linux")]
+    #[tokio::test]
+    async fn ssh_server_abstract_socket_cannot_be_replaced_while_bound() {
+        let socket = PathBuf::from(format!("@openshell-ssh-test-{}", uuid::Uuid::new_v4()));
+        let (listener, _, _) =
+            ssh_server_init(&socket, &None, ProcessEnforcementMode::NetworkOnly, true).unwrap();
+
+        assert!(
+            !socket.exists(),
+            "abstract socket must not create a filesystem inode"
+        );
+        let runtime_path = crate::unix_socket::runtime_path(&socket);
+        let err = UnixListener::bind(runtime_path.as_ref())
+            .expect_err("a workload must not be able to replace the bound abstract socket");
+        assert_eq!(err.kind(), std::io::ErrorKind::AddrInUse);
+
+        drop(listener);
+    }
+
     /// Verify that dropping the input sender (the operation `channel_eof`
     /// performs) causes the stdin writer loop to exit and close the child's
     /// stdin pipe.  Without this, commands like `cat | tar xf -` used by
@@ -1681,21 +1795,24 @@ mod tests {
             policy,
             None,
             None, // no netns fd
+            ProcessEnforcementMode::Full,
             #[cfg(target_os = "linux")]
-            sandbox::linux::prepare(
-                &SandboxPolicy {
-                    version: 0,
-                    filesystem: FilesystemPolicy::default(),
-                    network: NetworkPolicy::default(),
-                    landlock: LandlockPolicy::default(),
-                    process: ProcessPolicy {
-                        run_as_user: None,
-                        run_as_group: None,
+            Some(
+                sandbox::linux::prepare(
+                    &SandboxPolicy {
+                        version: 0,
+                        filesystem: FilesystemPolicy::default(),
+                        network: NetworkPolicy::default(),
+                        landlock: LandlockPolicy::default(),
+                        process: ProcessPolicy {
+                            run_as_user: None,
+                            run_as_group: None,
+                        },
                     },
-                },
-                None,
-            )
-            .expect("prepare should succeed in test environment"),
+                    None,
+                )
+                .expect("prepare should succeed in test environment"),
+            ),
         )
         .expect("install pre_exec should succeed");
 

crates/openshell-supervisor-process/src/supervisor_session.rs

@@ -235,12 +235,14 @@ pub fn spawn(
     sandbox_id: String,
     ssh_socket_path: std::path::PathBuf,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
 ) -> tokio::task::JoinHandle<()> {
     tokio::spawn(run_session_loop(
         endpoint,
         sandbox_id,
         ssh_socket_path,
         netns_fd,
+        expected_ssh_peer_pid,
     ))
 }
 
@@ -249,6 +251,7 @@ async fn run_session_loop(
     sandbox_id: String,
     ssh_socket_path: std::path::PathBuf,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
 ) {
     let mut backoff = INITIAL_BACKOFF;
     let mut attempt: u64 = 0;
@@ -256,7 +259,15 @@ async fn run_session_loop(
     loop {
         attempt += 1;
 
-        match run_single_session(&endpoint, &sandbox_id, &ssh_socket_path, netns_fd).await {
+        match run_single_session(
+            &endpoint,
+            &sandbox_id,
+            &ssh_socket_path,
+            netns_fd,
+            expected_ssh_peer_pid,
+        )
+        .await
+        {
             Ok(()) => {
                 let event =
                     session_closed_event(openshell_ocsf::ctx::ctx(), &endpoint, &sandbox_id);
@@ -283,6 +294,7 @@ async fn run_single_session(
     sandbox_id: &str,
     ssh_socket_path: &std::path::Path,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
 ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     // Connect to the gateway. The same `Channel` is used for both the
     // long-lived control stream and all data-plane `RelayStream` calls, so
@@ -352,6 +364,7 @@ async fn run_single_session(
                     sandbox_id,
                     ssh_socket_path,
                     netns_fd,
+                    expected_ssh_peer_pid,
                     &channel,
                     &tx,
                 );
@@ -375,6 +388,7 @@ fn handle_gateway_message(
     sandbox_id: &str,
     ssh_socket_path: &std::path::Path,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
     channel: &grpc_client::AuthedChannel,
     tx: &mpsc::Sender<SupervisorMessage>,
 ) {
@@ -395,7 +409,16 @@ fn handle_gateway_message(
 
             tokio::spawn(async move {
                 let event_open = relay_open.clone();
-                match handle_relay_open(relay_open, &ssh_socket_path, netns_fd, channel, tx).await {
+                match handle_relay_open(
+                    relay_open,
+                    &ssh_socket_path,
+                    netns_fd,
+                    expected_ssh_peer_pid,
+                    channel,
+                    tx,
+                )
+                .await
+                {
                     Ok(()) => {
                         let event = relay_closed_event(
                             openshell_ocsf::ctx::ctx(),
@@ -446,11 +469,19 @@ async fn handle_relay_open(
     relay_open: RelayOpen,
     ssh_socket_path: &std::path::Path,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
     channel: grpc_client::AuthedChannel,
     tx: mpsc::Sender<SupervisorMessage>,
 ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
     let channel_id = relay_open.channel_id.clone();
-    let target = match open_target(&relay_open, ssh_socket_path, netns_fd).await {
+    let target = match open_target(
+        &relay_open,
+        ssh_socket_path,
+        netns_fd,
+        expected_ssh_peer_pid,
+    )
+    .await
+    {
         Ok(target) => target,
         Err(err) => {
             send_relay_open_result(&tx, &channel_id, false, err.to_string()).await;
@@ -576,11 +607,23 @@ async fn open_target(
     relay_open: &RelayOpen,
     ssh_socket_path: &std::path::Path,
     netns_fd: Option<i32>,
+    expected_ssh_peer_pid: Option<u32>,
 ) -> Result<Box<dyn TargetStream>, Box<dyn std::error::Error + Send + Sync>> {
     match relay_open.target.as_ref() {
         Some(relay_open::Target::Tcp(target)) => open_tcp_target(target, netns_fd).await,
         Some(relay_open::Target::Ssh(_)) | None => {
-            let stream = tokio::net::UnixStream::connect(ssh_socket_path).await?;
+            let runtime_path = crate::unix_socket::runtime_path(ssh_socket_path);
+            let stream = tokio::net::UnixStream::connect(runtime_path.as_ref()).await?;
+            if let Some(expected_pid) = expected_ssh_peer_pid {
+                let credentials = stream.peer_cred()?;
+                let actual_pid = credentials.pid().and_then(|pid| u32::try_from(pid).ok());
+                if actual_pid != Some(expected_pid) {
+                    return Err(format!(
+                        "SSH relay peer PID mismatch: expected {expected_pid}, got {actual_pid:?}"
+                    )
+                    .into());
+                }
+            }
             Ok(Box::new(stream))
         }
     }
@@ -895,4 +938,37 @@ mod ocsf_event_tests {
             .expect_err("eof should force reconnect");
         assert_eq!(err.to_string(), "gateway closed stream");
     }
+
+    #[cfg(target_os = "linux")]
+    #[tokio::test]
+    async fn ssh_target_requires_authenticated_supervisor_peer_pid() {
+        let socket =
+            std::path::PathBuf::from(format!("@openshell-relay-test-{}", uuid::Uuid::new_v4()));
+        let runtime_path = crate::unix_socket::runtime_path(&socket);
+        let listener = tokio::net::UnixListener::bind(runtime_path.as_ref()).unwrap();
+        let accept_task = tokio::spawn(async move {
+            for _ in 0..2 {
+                let (_stream, _) = listener.accept().await.unwrap();
+            }
+        });
+        let relay = ssh_relay_open("peer-check");
+
+        let trusted = open_target(&relay, &socket, None, Some(std::process::id()))
+            .await
+            .expect("matching peer PID should be accepted");
+        drop(trusted);
+
+        let Err(err) = open_target(
+            &relay,
+            &socket,
+            None,
+            Some(std::process::id().saturating_add(1)),
+        )
+        .await
+        else {
+            panic!("mismatched peer PID must be rejected");
+        };
+        assert!(err.to_string().contains("peer PID mismatch"));
+        accept_task.await.unwrap();
+    }
 }

crates/openshell-supervisor-process/src/unix_socket.rs

@@ -0,0 +1,60 @@
+// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+//! Unix-socket address helpers shared by the SSH server and relay client.
+
+use std::borrow::Cow;
+use std::path::Path;
+#[cfg(target_os = "linux")]
+use std::path::PathBuf;
+
+/// Return whether a configured socket name denotes a Linux abstract socket.
+///
+/// Environment variables cannot contain the leading NUL byte used by the
+/// kernel ABI, so configuration uses the conventional `@name` spelling.
+pub fn is_abstract(path: &Path) -> bool {
+    #[cfg(target_os = "linux")]
+    {
+        use std::os::unix::ffi::OsStrExt;
+        path.as_os_str().as_bytes().starts_with(b"@")
+    }
+    #[cfg(not(target_os = "linux"))]
+    {
+        let _ = path;
+        false
+    }
+}
+
+/// Translate the configured `@name` spelling into Tokio's NUL-prefixed Linux
+/// abstract-socket path representation.
+pub fn runtime_path(path: &Path) -> Cow<'_, Path> {
+    #[cfg(target_os = "linux")]
+    {
+        use std::ffi::OsString;
+        use std::os::unix::ffi::{OsStrExt, OsStringExt};
+
+        let bytes = path.as_os_str().as_bytes();
+        if let Some(name) = bytes.strip_prefix(b"@") {
+            let mut abstract_name = Vec::with_capacity(name.len() + 1);
+            abstract_name.push(0);
+            abstract_name.extend_from_slice(name);
+            return Cow::Owned(PathBuf::from(OsString::from_vec(abstract_name)));
+        }
+    }
+    Cow::Borrowed(path)
+}
+
+#[cfg(all(test, target_os = "linux"))]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn at_name_maps_to_abstract_socket_path() {
+        use std::os::unix::ffi::OsStrExt;
+
+        let configured = Path::new("@openshell-test");
+        let runtime = runtime_path(configured);
+        assert!(is_abstract(configured));
+        assert_eq!(runtime.as_os_str().as_bytes(), b"\0openshell-test");
+    }
+}

deploy/docker/Dockerfile.supervisor

@@ -5,10 +5,10 @@
 
 # Supervisor image build.
 #
-# The final image is `scratch`: it only carries the static `openshell-sandbox`
-# binary used by Docker extraction, Podman image volumes, and the Kubernetes
-# init container copy-self path. A static musl binary lets the image stay
-# `scratch` while still being executable as an init container.
+# The final image carries the static `openshell-sandbox` binary used by Docker
+# extraction, Podman image volumes, and the Kubernetes init container copy-self
+# path. It also includes nftables so the Kubernetes supervisor sidecar can
+# install pod-namespace egress enforcement rules.
 #
 # The Rust binary is built natively before this image build runs and staged at:
 #   deploy/docker/.build/prebuilt-binaries/<arch>/openshell-sandbox
@@ -19,17 +19,16 @@
 # target) and uploads it as an artifact, which is downloaded into the same
 # staging directory before the image build job runs.
 
-FROM scratch AS supervisor
+FROM alpine:3.22 AS supervisor
 
 ARG TARGETARCH
 
-# --chmod=0550 drops world-execute and survives the actions/upload-artifact
-# + download-artifact roundtrip (which strips exec perms). Ownership is left
-# at root (0:0) deliberately: the Podman driver mounts this image as a
-# read-only image volume into the sandbox container and drops DAC_OVERRIDE,
-# so the container's UID 0 must own the binary to read+exec it. Mode 0550
-# (r-xr-x---) is the security win; the chown to a non-root UID was breaking
-# Podman without buying anything since the container is always UID 0.
-COPY --chmod=0550 deploy/docker/.build/prebuilt-binaries/${TARGETARCH}/openshell-sandbox /openshell-sandbox
+RUN apk add --no-cache nftables iptables iptables-legacy
+
+# --chmod=0555 restores execute bits after the actions/upload-artifact +
+# download-artifact roundtrip strips them. Ownership stays root (0:0) for
+# Podman image-volume mounts, while world-execute lets the Kubernetes
+# network sidecar run this binary as the dedicated non-root proxy UID.
+COPY --chmod=0555 deploy/docker/.build/prebuilt-binaries/${TARGETARCH}/openshell-sandbox /openshell-sandbox
 
 ENTRYPOINT ["/openshell-sandbox"]

deploy/helm/openshell/README.md

@@ -239,8 +239,10 @@ add `ci/values-spire.yaml` to the OpenShell release values files.
 | supervisor.image.pullPolicy | string | `""` | Supervisor image pull policy. Defaults to the gateway image pull policy when empty. |
 | supervisor.image.repository | string | `"ghcr.io/nvidia/openshell/supervisor"` | Supervisor image repository. Changing it uses the effective gateway image tag unless tag is also set. |
 | supervisor.image.tag | string | `""` | Supervisor image tag override. Empty uses the version pinned into the gateway unless repository is changed. |
+| supervisor.sidecar.processBinaryAwareNetworkPolicy | bool | `true` | Keep process/binary-aware network policy enabled in sidecar topology. When false, the network sidecar runs as proxyUid, drops the extra /proc inspection capabilities, and enforces endpoint/L7 policy without matching policy.binaries. |
+| supervisor.sidecar.proxyUid | int | `1337` | UID for relaxed long-running network sidecars in sidecar topology. Strict process/binary-aware sidecars run as UID 0 so Kubernetes grants the required /proc inspection capabilities into the effective set. The network init container installs nftables rules that exempt the effective sidecar UID. |
 | supervisor.sideloadMethod | string | `""` | How the supervisor binary is delivered into sandbox pods. Empty (default) = auto-detect from cluster version:   K8s >= v1.35 -> "image-volume" (ImageVolume enabled by default; GA in v1.36)   K8s < v1.35 -> "init-container" (copies via init container + emptyDir) On K8s v1.33-v1.34 with the ImageVolume feature gate manually enabled, set this to "image-volume" explicitly. |
-| supervisor.topology | string | `"combined"` | Supervisor pod topology for Kubernetes sandboxes. "combined" runs networking and process supervision in the agent container. |
+| supervisor.topology | string | `"combined"` | Supervisor pod topology for Kubernetes sandboxes. "combined" runs the current single supervisor container in the agent pod. "sidecar" runs network enforcement in a dedicated sidecar and the process supervisor as a low-capability wrapper in the agent container. |
 | tolerations | list | `[]` | Tolerations for the gateway pod. |
 | workload.allowMultiReplicaStatefulSet | bool | `false` | Allow replicaCount > 1 while rendering a StatefulSet. Prefer workload.kind=deployment for external database-backed multi-replica gateways; this override exists for operators who explicitly require StatefulSet identity or storage semantics. |
 | workload.kind | string | `"statefulset"` | Gateway workload controller kind. Use `statefulset` for the default SQLite database, or `deployment` when server.externalDbSecret points at an external database. |

deploy/helm/openshell/ci/values-sidecar-kata.yaml

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# CI/dev overlay for exercising the Kubernetes supervisor sidecar topology under
+# a Kata RuntimeClass.
+#
+# Use with e2e/with-kube-gateway.sh by setting:
+#   OPENSHELL_E2E_KUBE_EXTRA_VALUES=deploy/helm/openshell/ci/values-sidecar-kata.yaml
+# The e2e wrapper supplies the image repository and tag through OPENSHELL_REGISTRY
+# and IMAGE_TAG for existing-cluster runs.
+
+supervisor:
+  # Use the sidecar topology under Kata so network enforcement runs in the
+  # sidecar and the sandbox agent container stays low-privilege.
+  topology: sidecar
+  sidecar:
+    # Keep strict process/binary-aware network policy enabled for the Kata
+    # validation path. Set this false only when intentionally validating the
+    # documented endpoint/L7-only downgrade mode.
+    processBinaryAwareNetworkPolicy: true
+
+# Kata validation clusters normally install this RuntimeClass.
+server:
+  defaultRuntimeClassName: kata-qemu

deploy/helm/openshell/ci/values-sidecar.yaml

@@ -0,0 +1,18 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# CI/dev overlay for exercising the Kubernetes supervisor sidecar topology.
+#
+# Merge after values.yaml and ci/values-skaffold.yaml:
+#   helm install ... -f values.yaml -f ci/values-skaffold.yaml -f ci/values-sidecar.yaml
+#
+# Or set:
+#   OPENSHELL_E2E_KUBE_EXTRA_VALUES=deploy/helm/openshell/ci/values-sidecar.yaml
+# before running `mise run e2e:kubernetes`.
+supervisor:
+  topology: sidecar
+  sidecar:
+    # The strict sidecar default requires cross-container /proc identity access.
+    # CI/dev e2e uses the explicit downgraded mode so endpoint and L7 policy
+    # coverage remains runnable on local k3d while that path is hardened.
+    processBinaryAwareNetworkPolicy: false

deploy/helm/openshell/skaffold.yaml

@@ -119,6 +119,8 @@ deploy:
           # To enable SPIFFE/SPIRE provider token grants (requires the
           # spire-crds and spire releases above):
           #- ci/values-spire.yaml
+          # To exercise the Kubernetes supervisor sidecar topology:
+          #- ci/values-sidecar.yaml
           # To test multi-replica external PostgreSQL behavior:
           #- ci/values-high-availability.yaml
         setValueTemplates:
@@ -126,3 +128,18 @@ deploy:
           image.tag: '{{.IMAGE_TAG_openshell_gateway}}'
           supervisor.image.repository: '{{.IMAGE_REPO_openshell_supervisor}}'
           supervisor.image.tag: '{{.IMAGE_TAG_openshell_supervisor}}'
+profiles:
+  - name: sidecar
+    patches:
+      - op: add
+        path: /deploy/helm/releases/0/valuesFiles/-
+        value: ci/values-sidecar.yaml
+  - name: sidecar-mtls
+    patches:
+      - op: add
+        path: /deploy/helm/releases/0/valuesFiles/-
+        value: ci/values-sidecar.yaml
+      - op: add
+        path: /deploy/helm/releases/0/setValues
+        value:
+          server.disableTls: "false"

deploy/helm/openshell/templates/gateway-config.yaml

@@ -115,7 +115,7 @@ data:
     grpc_endpoint                = {{ include "openshell.grpcEndpoint" . | quote }}
     service_account_name         = {{ include "openshell.sandboxServiceAccountName" . | quote }}
     supervisor_sideload_method   = {{ include "openshell.supervisorSideloadMethod" . | quote }}
-    supervisor_topology          = {{ .Values.supervisor.topology | default "combined" | quote }}
+    topology                     = {{ .Values.supervisor.topology | default "combined" | quote }}
     sa_token_ttl_secs            = {{ .Values.server.sandboxJwt.k8sSaTokenTtlSecs | default 3600 }}
     {{- if .Values.server.providerTokenGrants.spiffe.enabled }}
     provider_spiffe_workload_api_socket_path = {{ .Values.server.providerTokenGrants.spiffe.workloadApiSocketPath | quote }}
@@ -144,3 +144,7 @@ data:
     {{- if .Values.supervisor.image.pullPolicy }}
     supervisor_image_pull_policy = {{ .Values.supervisor.image.pullPolicy | quote }}
     {{- end }}
+
+    [openshell.drivers.kubernetes.sidecar]
+    proxy_uid                    = {{ .Values.supervisor.sidecar.proxyUid | default 1337 }}
+    process_binary_aware_network_policy = {{ .Values.supervisor.sidecar.processBinaryAwareNetworkPolicy }}

deploy/helm/openshell/tests/gateway_config_test.yaml

@@ -88,7 +88,10 @@ tests:
     asserts:
       - matchRegex:
           path: data["gateway.toml"]
-          pattern: '(?ms)\[openshell\.drivers\.kubernetes\].*?supervisor_topology\s*=\s*"combined"'
+          pattern: '(?ms)\[openshell\.drivers\.kubernetes\].*?topology\s*=\s*"combined"'
+      - notMatchRegex:
+          path: data["gateway.toml"]
+          pattern: 'supervisor[_]topology\s*='
 
   - it: uses the gateway built-in supervisor image by default
     template: templates/gateway-config.yaml
@@ -126,14 +129,35 @@ tests:
           path: data["gateway.toml"]
           pattern: 'supervisor_image\s*=\s*"registry\.example\.com/openshell/supervisor:supervisor-build"'
 
-  - it: renders explicit combined supervisor topology under [openshell.drivers.kubernetes]
+  - it: renders sidecar supervisor topology under [openshell.drivers.kubernetes]
     template: templates/gateway-config.yaml
     set:
-      supervisor.topology: combined
+      supervisor.topology: sidecar
     asserts:
       - matchRegex:
           path: data["gateway.toml"]
-          pattern: '(?ms)\[openshell\.drivers\.kubernetes\].*?supervisor_topology\s*=\s*"combined"'
+          pattern: '(?ms)\[openshell\.drivers\.kubernetes\].*?topology\s*=\s*"sidecar"'
+      - notMatchRegex:
+          path: data["gateway.toml"]
+          pattern: 'supervisor[_]topology\s*='
+
+  - it: renders proxy uid under [openshell.drivers.kubernetes.sidecar]
+    template: templates/gateway-config.yaml
+    set:
+      supervisor.sidecar.proxyUid: 2200
+    asserts:
+      - matchRegex:
+          path: data["gateway.toml"]
+          pattern: '(?ms)\[openshell\.drivers\.kubernetes\.sidecar\].*?proxy_uid\s*=\s*2200'
+
+  - it: renders process binary aware network policy under [openshell.drivers.kubernetes.sidecar]
+    template: templates/gateway-config.yaml
+    set:
+      supervisor.sidecar.processBinaryAwareNetworkPolicy: false
+    asserts:
+      - matchRegex:
+          path: data["gateway.toml"]
+          pattern: '(?ms)\[openshell\.drivers\.kubernetes\.sidecar\].*?process_binary_aware_network_policy\s*=\s*false'
 
   - it: renders sandbox image pull secrets under [openshell.drivers.kubernetes]
     template: templates/gateway-config.yaml

deploy/helm/openshell/values.yaml

@@ -45,8 +45,22 @@ supervisor:
   # set this to "image-volume" explicitly.
   sideloadMethod: ""
   # -- Supervisor pod topology for Kubernetes sandboxes.
-  # "combined" runs networking and process supervision in the agent container.
+  # "combined" runs the current single supervisor container in the agent pod.
+  # "sidecar" runs network enforcement in a dedicated sidecar and the process
+  # supervisor as a low-capability wrapper in the agent container.
   topology: "combined"
+  sidecar:
+    # -- UID for relaxed long-running network sidecars in sidecar topology.
+    # Strict process/binary-aware sidecars run as UID 0 so Kubernetes grants
+    # the required /proc inspection capabilities into the effective set. The
+    # network init container installs nftables rules that exempt the effective
+    # sidecar UID.
+    proxyUid: 1337
+    # -- Keep process/binary-aware network policy enabled in sidecar topology.
+    # When false, the network sidecar runs as proxyUid, drops the extra /proc
+    # inspection capabilities, and enforces endpoint/L7 policy without matching
+    # policy.binaries.
+    processBinaryAwareNetworkPolicy: true
 
 # -- Image pull secrets attached to gateway and helper pods.
 imagePullSecrets: []

docs/kubernetes/access-control.mdx

@@ -5,7 +5,7 @@ title: "Access Control"
 sidebar-title: "Access Control"
 description: "Configure OIDC user authentication or reverse-proxy auth termination for a Kubernetes-deployed OpenShell gateway."
 keywords: "Generative AI, Cybersecurity, Kubernetes, Authentication, mTLS, OIDC, Keycloak, Entra ID, Okta, Gateway Auth"
-position: 4
+position: 5
 ---
 
 The OpenShell gateway supports two access-control models for human callers on Kubernetes:

docs/kubernetes/ingress.mdx

@@ -5,7 +5,7 @@ title: "Ingress"
 sidebar-title: "Ingress"
 description: "Expose the OpenShell gateway externally using the Kubernetes Gateway API and a GRPCRoute."
 keywords: "Generative AI, Cybersecurity, Kubernetes, Gateway API, Envoy Gateway, GRPCRoute, Ingress, External Access"
-position: 3
+position: 4
 ---
 
 By default, the OpenShell gateway is only reachable inside the cluster. To let CLI clients connect without a `kubectl port-forward`, expose the gateway through an ingress.

docs/kubernetes/managing-certificates.mdx

@@ -5,7 +5,7 @@ title: "Managing Certificates"
 sidebar-title: "Managing Certificates"
 description: "Configure the OpenShell Helm chart to use cert-manager for mTLS certificate issuance and automatic renewal."
 keywords: "Generative AI, Cybersecurity, Kubernetes, cert-manager, PKI, TLS, mTLS, Certificates"
-position: 2
+position: 3
 ---
 
 The OpenShell gateway uses mTLS certificates for transport between the gateway and sandbox supervisors. These certificates are not Kubernetes user authentication; configure OIDC or a trusted access proxy for user access. The Helm chart supports two ways to provision and manage the certificate bundle:

docs/kubernetes/openshift.mdx

@@ -5,7 +5,7 @@ title: "OpenShift"
 sidebar-title: "OpenShift"
 description: "Install the OpenShell Helm chart on OpenShift, including the SCC binding and chart overrides required by OpenShift's Security Context Constraints."
 keywords: "Generative AI, Cybersecurity, Kubernetes, OpenShift, SCC, Security Context Constraints, Helm, Gateway, Installation"
-position: 5
+position: 6
 ---
 
 <Warning>
@@ -14,6 +14,11 @@ The OpenShift install path is experimental. It currently requires running sandbo
 
 OpenShift's [Security Context Constraints](https://docs.openshift.com/container-platform/latest/authentication/managing-security-context-constraints.html) reject the chart's default pod security settings. Installing on OpenShift requires precreating the namespace, granting the `privileged` SCC to the sandbox service account, and overriding a few chart values so the cluster admission controller can assign UIDs and FS groups itself.
 
+OpenShell installs sandbox nftables rules as individual commands. On OpenShift
+nodes where optional conntrack or packet log expressions are unavailable, those
+optional rules can fail without rolling back the required proxy bypass reject
+rules.
+
 ## Prerequisites
 
 - OpenShift 4.x cluster with `oc` configured

docs/kubernetes/setup.mdx

@@ -161,6 +161,7 @@ The most commonly changed values are:
 | `pkiInitJob.serverDnsNames` / `certManager.serverDnsNames` | Additional gateway server DNS SANs. Wildcard SANs also enable sandbox service URLs under that domain. |
 | `supervisor.sideloadMethod` | How the supervisor binary is delivered into sandbox pods. Leave empty to auto-detect based on cluster version: clusters running Kubernetes 1.35 or later use `image-volume` (ImageVolume GA in 1.36); older clusters use `init-container`. Set explicitly to `image-volume` on Kubernetes 1.33 or 1.34 with the ImageVolume feature gate enabled, or to `init-container` to force the legacy path on any version. |
 | `supervisor.topology` | Sandbox pod topology. Refer to [Topology](/kubernetes/topology). |
+| `supervisor.sidecar.proxyUid` | Non-root UID used when sidecar process/binary-aware network policy is disabled. The default binary-aware sidecar runs as UID 0 instead. The configured UID must not match the sandbox UID. |
 
 Use a values file for repeatable deployments:
 
@@ -244,7 +245,7 @@ The gateway exposes `/healthz` for process liveness and `/readyz` for dependency
 
 ## Next Steps
 
-- To review Kubernetes sandbox topology, refer to [Topology](/kubernetes/topology).
+- To choose between combined and sidecar sandbox pods, refer to [Topology](/kubernetes/topology).
 - To enable automatic certificate rotation with cert-manager, refer to [Managing Certificates](/kubernetes/managing-certificates).
 - To expose the gateway externally without port-forwarding, refer to [Ingress](/kubernetes/ingress).
 - To configure OIDC or reverse-proxy authentication, refer to [Access Control](/kubernetes/access-control).

docs/kubernetes/topology.mdx

@@ -3,37 +3,36 @@
 # SPDX-License-Identifier: Apache-2.0
 title: "Kubernetes Sandbox Topology"
 sidebar-title: "Topology"
-description: "Review the default combined supervisor topology for Kubernetes sandbox pods."
-keywords: "Generative AI, Cybersecurity, Kubernetes, Sandboxing, RuntimeClass"
+description: "Choose between combined and sidecar supervisor topology for Kubernetes sandbox pods."
+keywords: "Generative AI, Cybersecurity, Kubernetes, Sandboxing, Sidecar, Network Policy, RuntimeClass"
 position: 2
 ---
 
-Kubernetes sandbox pods run the OpenShell supervisor in `combined` topology by
-default. Combined topology keeps network, filesystem, and process controls in
-the agent pod so the supervisor can enforce the complete OpenShell sandbox
-contract before launching the workload.
+Kubernetes sandbox pods can run the OpenShell supervisor in `combined` or
+`sidecar` topology. Choose the topology based on which controls you need inside
+the pod and how much privilege your cluster allows on the agent container.
 
 ## Choose a Topology
 
 The default `combined` topology preserves the full OpenShell enforcement model.
-Use it when you need OpenShell to apply all sandbox controls inside the workload
-pod and your cluster policy permits the required Linux capabilities.
+Use `sidecar` only when you accept network-focused enforcement in exchange for a
+lower-privilege agent container.
 
 | Topology | Use when | Main tradeoff |
 |---|---|---|
 | `combined` | You need OpenShell network, filesystem, and process controls in the sandbox workload. | The agent container carries the Linux capabilities the supervisor needs. |
-
-Additional Kubernetes sandbox topologies are still being designed. Until they
-are documented as supported configuration values, `combined` is the only
-supported value for `supervisor.topology`.
+| `sidecar` | You need the agent container to run as non-root without added Linux capabilities, and network policy is the primary control. | Privilege-dropping and supervisor mount isolation do not run in the agent container. |
 
 ## Privilege Model
 
-The long-running container permissions for `combined` topology are:
+The long-running container permissions differ by topology:
 
 | Topology | Pod or container | UID/GID | Privilege escalation | Capabilities | Result |
 |---|---|---|---|---|---|
 | `combined` | Agent container, which also runs the supervisor | Not forced by topology | Not explicitly disabled by the driver | Adds `SYS_ADMIN`, `NET_ADMIN`, `SYS_PTRACE`, and `SYSLOG`; adds `SETUID`, `SETGID`, and `DAC_READ_SEARCH` when user namespaces are enabled | Full supervisor controls run in the agent container. |
+| `sidecar` | Agent container, process-only supervisor (`network-only`) | `sandbox_uid:sandbox_gid` | `false` | Drops `ALL` | Agent and workload run without added Linux capabilities. |
+| `sidecar` | Network supervisor sidecar, binary-aware mode (default) | `0:sandbox_gid` | `false` | Drops `ALL`; adds `SYS_PTRACE` and `DAC_READ_SEARCH` | Root sidecar inspects cross-UID workload `/proc` entries. The nftables fence exempts UID 0, so do not inject other root containers into these pods. |
+| `sidecar` | Network supervisor sidecar, endpoint/L7-only mode | `proxyUid:sandbox_gid` | `false` | Drops `ALL` | Non-root sidecar enforces endpoint and L7 policy without matching `policy.binaries`. |
 
 Short-lived setup containers still have the permissions needed to prepare the
 pod:
@@ -41,6 +40,7 @@ pod:
 | Topology | Setup container | UID/GID | Privilege escalation | Capabilities | Purpose |
 |---|---|---|---|---|---|
 | `combined` | Supervisor install init container | `0` | Not set | Not set | Copies the supervisor binary into the agent container volume. |
+| `sidecar` | Network init container | `0` | `false` | Drops `ALL`; adds `NET_ADMIN`, `NET_RAW`, `CHOWN`, and `FOWNER` | Installs pod-local nftables rules and prepares shared sidecar state. |
 
 ## Combined Topology
 
@@ -48,6 +48,26 @@ Combined topology is the original Kubernetes mode and remains the default. The
 agent container starts the OpenShell supervisor, and the supervisor launches the
 workload after applying sandbox setup.
 
+```mermaid
+flowchart TB
+  Sandbox["agents.x-k8s.io Sandbox"]
+
+  subgraph Pod["Sandbox pod"]
+    subgraph Agent["agent container"]
+      Supervisor["OpenShell supervisor<br/>network + process + filesystem"]
+      Workload["Agent workload"]
+    end
+  end
+
+  Gateway["OpenShell Gateway"]
+  External["External services"]
+
+  Sandbox --> Pod
+  Supervisor --> Workload
+  Supervisor -->|"gateway callback / SSH relay"| Gateway
+  Supervisor -->|"policy-enforced egress"| External
+```
+
 Combined topology keeps these controls in one supervisor path:
 
 - Network endpoint and L7 policy enforcement.
@@ -61,12 +81,119 @@ controls from the agent container, Kubernetes grants that container elevated
 Linux capabilities. Use this mode when you need the complete OpenShell sandbox
 contract and your cluster policy permits those capabilities.
 
+## Sidecar Topology
+
+Sidecar topology splits the supervisor into a network sidecar and a
+low-privilege process supervisor in the agent container.
+
+```mermaid
+flowchart TB
+  Sandbox["agents.x-k8s.io Sandbox"]
+
+  subgraph Pod["Sandbox pod"]
+    Init["network init container<br/>root setup capabilities"]
+    State["shared state + TLS volumes"]
+    NetNS["pod network namespace"]
+
+    subgraph Agent["agent container"]
+      ProcessSupervisor["process supervisor<br/>network-only"]
+      Workload["Agent workload"]
+    end
+
+    NetworkSidecar["network supervisor sidecar<br/>UID 0 by default"]
+    SshEndpoint["abstract SSH relay socket<br/>peer-PID authenticated"]
+  end
+
+  Gateway["OpenShell Gateway"]
+  External["External services"]
+
+  Sandbox --> Pod
+  Init -->|"installs nftables rules"| NetNS
+  ProcessSupervisor --> Workload
+  Workload -->|"egress redirected on loopback"| NetworkSidecar
+  NetworkSidecar -->|"gateway session + relays"| Gateway
+  NetworkSidecar -->|"policy-enforced egress"| External
+  NetworkSidecar -->|"control socket + proxy TLS"| State
+  ProcessSupervisor -->|"bootstrap + updates"| State
+  ProcessSupervisor --> SshEndpoint
+  NetworkSidecar -->|"SSH relay"| SshEndpoint
+  NetworkSidecar --- State
+```
+
+The pod contains these OpenShell-managed pieces:
+
+| Component | Runs as | Purpose |
+|---|---|---|
+| Network init container | Root with setup capabilities | Installs pod-level nftables rules and prepares shared sidecar state. |
+| Network sidecar | UID 0 by default; `supervisor.sidecar.proxyUid` when binary-aware policy is disabled | Runs the proxy, enforces network policy, owns gateway authentication and the gateway session, and serves local policy/provider state over the sidecar control socket. |
+| Agent container | Resolved sandbox UID/GID | Runs the process supervisor and launches the user workload. |
+
+In this topology, the agent container defaults to `runAsNonRoot: true`,
+`allowPrivilegeEscalation: false`, and `capabilities.drop: ["ALL"]`. The
+default binary-aware network sidecar runs as UID 0, drops default Linux
+capabilities, and adds `SYS_PTRACE` plus `DAC_READ_SEARCH` for cross-UID workload
+process identity resolution. Setting
+`supervisor.sidecar.processBinaryAwareNetworkPolicy=false` runs the sidecar as
+the configured non-root `proxyUid`, omits both capabilities, and downgrades
+network policy to endpoint/L7 enforcement without binary matching. The root
+init container keeps the setup capabilities needed to configure pod networking.
+
+Sidecar mode preserves gateway session behavior, including SSH connectivity,
+because the network sidecar owns the gateway session and bridges relay requests
+to a Linux abstract SSH socket owned by the process supervisor. The relay
+verifies the socket peer PID against the authenticated control connection, so
+the workload cannot replace the relay endpoint. The agent container does not get a
+gateway endpoint, gateway TLS material, or the sandbox bootstrap token in the
+default sidecar path.
+
+<Warning>
+Sidecar mode runs the process supervisor in `network-only` mode. OpenShell still
+enforces network endpoint and L7 policy through the sidecar, and the process
+supervisor applies Landlock filesystem policy and child seccomp filters where
+the kernel/runtime supports them. The process supervisor does not perform
+root-to-sandbox privilege dropping because Kubernetes starts the container as
+the sandbox UID/GID, and it does not perform supervisor identity mount
+isolation because gateway credentials are not mounted into the agent container.
+Sidecar pods use `shareProcessNamespace: true` so the network sidecar can
+resolve workload process and binary identity through `/proc/<entrypoint-pid>`.
+</Warning>
+
+## Credential Exposure
+
+Sidecar topology keeps gateway credentials in the network sidecar. The agent
+container does not mount the projected ServiceAccount token used for sandbox
+token bootstrap, does not mount the sandbox client TLS secret, and does not get
+gateway callback environment variables.
+
+The network sidecar serves the policy and workload-facing provider environment
+over a Unix control socket in the shared sidecar state volume. Before launching
+the workload, the process supervisor establishes the only accepted connection.
+The sidecar validates its UID, GID, and PID with peer credentials, unlinks the
+listener, derives the SSH target from trusted configuration, and rejects later
+clients. The connection receives bootstrap state and provider-environment
+updates after settings polls. If it closes, the network sidecar exits so
+Kubernetes recreates the one-client bootstrap listener, and the process
+supervisor exits so Kubernetes terminates the workload and restarts the agent
+container. This symmetric failure behavior prevents a surviving workload from
+claiming the new control listener after an isolated sidecar restart. Future
+child processes can see refreshed provider env without giving the agent
+container gateway authentication material. This does not mutate the environment
+of the already-running workload entrypoint. Use `combined` topology when you
+need the full single-supervisor enforcement path; use additional runtime
+isolation when you need a stronger container boundary around sidecar workloads.
+
 ## RuntimeClass Isolation
 
-RuntimeClass isolation can add a stronger container boundary for the sandbox
-workload when the cluster supports it. Runtime classes do not replace the
-combined topology's supervisor controls; they add another isolation boundary
-around the same supervised workload.
+Sidecar topology has been validated with Kata Containers. It does not currently
+support gVisor because sidecar mode requires pod-local nftables setup, which
+gVisor does not provide to the init container. A supported sandboxed runtime
+strengthens the container boundary while OpenShell focuses on network policy
+enforcement from the sidecar.
+
+Runtime classes do not re-enable the OpenShell privilege-drop or supervisor
+mount-isolation controls that sidecar mode relaxes. Use them as an additional
+workload boundary, not as a replacement for the combined topology's full
+supervisor controls.
 
 You can set a default runtime class in the Kubernetes driver configuration or
 override it per sandbox with driver config:
@@ -77,24 +204,44 @@ openshell sandbox create \
   -- claude
 ```
 
-## Configure Combined Mode
+## Enable Sidecar Mode
 
-For direct gateway TOML configuration, leave `supervisor_topology` unset, or
-set it to `combined`, to use the default single-container supervisor path:
+For direct gateway TOML configuration, set the Kubernetes driver fields:
 
 ```toml
 [openshell.drivers.kubernetes]
-supervisor_topology = "combined"
+topology = "sidecar"
+
+[openshell.drivers.kubernetes.sidecar]
+proxy_uid = 1337
 ```
 
-When the Helm chart renders `gateway.toml`, leave `supervisor.topology` unset,
-or set it to `combined`, to produce the same driver configuration:
+`proxy_uid` configures only the relaxed endpoint/L7-only sidecar. It must be a
+non-root UID and must not match the sandbox UID. The default binary-aware mode
+runs the sidecar as UID 0 instead. The network init container exempts the
+effective sidecar UID from proxy redirection so the sidecar can reach the
+gateway.
+
+When the Helm chart renders `gateway.toml`, set the equivalent chart values:
 
 ```yaml
 supervisor:
-  topology: combined
+  topology: sidecar
+  sidecar:
+    proxyUid: 1337
+    processBinaryAwareNetworkPolicy: true
 ```
 
+Leave `topology` unset, or set it to `combined`, to keep the original
+single-container supervisor path. For Helm installs, leave
+`supervisor.topology` unset or set it to `combined`.
+
+Set `supervisor.sidecar.processBinaryAwareNetworkPolicy=false` only when you
+accept downgrading sidecar network policy to endpoint/L7 enforcement without
+matching `policy.binaries`. This changes the sidecar from UID 0 to `proxyUid`
+and removes its `SYS_PTRACE` and `DAC_READ_SEARCH` capabilities, which are used
+for cross-UID `/proc` inspection.
+
 ## Next Steps
 
 - To install OpenShell on Kubernetes, refer to [Setup](/kubernetes/setup).

docs/reference/gateway-config.mdx

@@ -179,9 +179,10 @@ supervisor_image_pull_policy   = "IfNotPresent"
 # Use the image volume on Kubernetes >= 1.35 (GA in 1.36); switch to "init-container"
 # on older clusters or where the ImageVolume feature gate is off.
 supervisor_sideload_method     = "image-volume"
-# "combined" runs networking and process supervision in the sandbox agent
-# container and preserves the existing Kubernetes sandbox behavior.
-supervisor_topology            = "combined"
+# "combined" runs the existing single supervisor container with full process,
+# filesystem, and network enforcement in the agent container. "sidecar" moves
+# pod-level network enforcement and gateway session handling into a network sidecar.
+topology                       = "combined"
 grpc_endpoint                  = "https://openshell-gateway.agents.svc:8080"
 ssh_socket_path                = "/run/openshell/ssh.sock"
 client_tls_secret_name         = "openshell-client-tls"
@@ -205,6 +206,18 @@ provider_spiffe_workload_api_socket_path = "/spiffe-workload-api/spire-agent.soc
 # back to 1000 on non-OpenShift clusters.
 # sandbox_uid = 1500
 # sandbox_gid = 1500
+
+[openshell.drivers.kubernetes.sidecar]
+# UID used by relaxed long-running network sidecars. Strict process/binary-aware
+# sidecars run as UID 0 so Kubernetes grants the required /proc inspection
+# capabilities into the effective set. In sidecar topology the network init
+# container installs nftables rules that exempt the effective sidecar UID.
+proxy_uid                      = 1337
+# Keep process/binary-aware network policy enabled in sidecar topology. Set
+# false to run the sidecar as proxy_uid, drop the sidecar's extra /proc
+# inspection capabilities, and enforce endpoint/L7 policy without matching
+# policy.binaries.
+process_binary_aware_network_policy = true
 ```
 
 ### Docker

docs/reference/sandbox-compute-drivers.mdx

@@ -306,10 +306,38 @@ For maintainer-level implementation details, refer to the [Kubernetes driver REA
 | `supervisor_image` | `supervisor.image.repository` / `supervisor.image.tag` | Override the supervisor image that provides the `openshell-sandbox` binary. The default repository with an empty tag uses the version-pinned image built into the gateway. Changing the repository uses the effective gateway image tag, while setting a tag pins that version explicitly. |
 | `supervisor_image_pull_policy` | `supervisor.image.pullPolicy` | Set the Kubernetes image pull policy for the supervisor image. |
 | `supervisor_sideload_method` | `supervisor.sideloadMethod` | How the supervisor binary is delivered into sandbox pods. Leave empty to auto-detect from cluster version. Set to `image-volume` to mount the supervisor OCI image directly as a volume (requires Kubernetes 1.33+ with the ImageVolume feature gate; GA in 1.36), or `init-container` to copy it through an init container on older clusters. |
+| `topology` | `supervisor.topology` | Set `combined` for the default single supervisor path, or `sidecar` to move pod-level network enforcement and the gateway session into a dedicated sidecar. |
+| `sidecar.proxy_uid` | `supervisor.sidecar.proxyUid` | Non-root UID used by the relaxed sidecar when process/binary-aware network policy is disabled. The default binary-aware sidecar runs as UID 0. The network init container exempts the effective sidecar UID from proxy redirection. |
+| `sidecar.process_binary_aware_network_policy` | `supervisor.sidecar.processBinaryAwareNetworkPolicy` | Keep process/binary-aware network policy enabled in `sidecar` topology. The default runs the sidecar as UID 0 with `SYS_PTRACE` and `DAC_READ_SEARCH`. Set false to run as `proxy_uid`, drop both capabilities, and enforce endpoint/L7 policy without matching `policy.binaries`. |
 | `app_armor_profile` | `server.appArmorProfile` | Set the sandbox agent container's AppArmor profile. Helm defaults this to `Unconfined` so AppArmor-enabled nodes do not block supervisor network namespace setup. Set the Helm value to an empty string to omit the field, or use `RuntimeDefault` or `Localhost/<profile-name>` for operator-managed profiles. |
 | `workspace_default_storage_size` | `server.workspaceDefaultStorageSize` | Set the default workspace PVC size for new sandboxes. |
 | `sa_token_ttl_secs` | `server.sandboxJwt.k8sSaTokenTtlSecs` | Set the projected ServiceAccount token TTL used for the bootstrap token exchange. |
 
+In `combined` topology, the agent container carries the Linux capabilities
+needed by the supervisor for network namespace setup, Landlock filesystem
+policy, process privilege changes, and network policy enforcement. In `sidecar`
+topology, the agent container runs as the resolved sandbox UID/GID with no added
+Linux capabilities. A root init container performs the nftables setup, and the
+long-running binary-aware sidecar runs as UID 0, drops default capabilities,
+and adds `SYS_PTRACE` plus `DAC_READ_SEARCH` for workload process identity
+resolution through shared `/proc`. The
+`sidecar.process_binary_aware_network_policy = false` setting runs it as the
+configured non-root `proxy_uid`, removes both capabilities, and relaxes network
+policy to endpoint/L7 matching only. The
+network sidecar owns gateway authentication and writes local policy/provider
+state to the process supervisor over a local control socket, so the agent
+container does not mount the sandbox bootstrap token or client TLS secret in
+the default sidecar path. The provider environment is refreshed by the network
+sidecar after settings polls and streamed to the process supervisor so future
+child processes can see updated provider env without gateway access in the
+agent container.
+Sidecar mode keeps gateway session and SSH behavior. The process supervisor
+applies Landlock filesystem policy and child seccomp filters where supported,
+but it does not perform root-to-sandbox privilege dropping or supervisor
+identity mount isolation. Network policy still runs in the sidecar, and sidecar
+pods set `shareProcessNamespace: true` so the network sidecar can resolve
+process/binary identity through `/proc/<entrypoint-pid>`.
+
 The Kubernetes driver creates namespaced `agents.x-k8s.io` `Sandbox` resources from the Kubernetes SIG Apps [agent-sandbox](https://github.com/kubernetes-sigs/agent-sandbox) project. It detects the served Sandbox API at runtime, caches the selected API version for the gateway process, and uses `v1beta1` when available before falling back to `v1alpha1`, so supported Agent Sandbox installations work without version-specific operator configuration. The Agent Sandbox controller turns those resources into sandbox pods and related storage.
 
 If Agent Sandbox is upgraded in place, restart the OpenShell gateway after the controller and CRD rollout completes so the gateway can detect the served API versions again.

e2e/with-kube-gateway.sh

@@ -393,6 +393,21 @@ require_cmd() {
   fi
 }
 
+configure_fixture_container_engine() {
+  [ -n "${CONTAINER_ENGINE:-}" ] || return 0
+  local selected_engine
+  selected_engine="$(printf '%s' "${CONTAINER_ENGINE}" | tr '[:upper:]' '[:lower:]')"
+  case "${selected_engine}" in
+    docker|podman)
+      ;;
+    *)
+      echo "ERROR: CONTAINER_ENGINE=${CONTAINER_ENGINE} is invalid; expected docker or podman" >&2
+      exit 2
+      ;;
+  esac
+  export CONTAINER_ENGINE="${selected_engine}"
+}
+
 require_cmd helm
 require_cmd kubectl
 require_cmd curl
@@ -423,6 +438,8 @@ else
   KUBE_CONTEXT="k3d-${CLUSTER_NAME}"
 fi
 
+configure_fixture_container_engine
+
 if [ -z "${OPENSHELL_E2E_KUBE_BUILD_IMAGES+x}" ]; then
   if [ "${CLUSTER_CREATED_BY_US}" = "1" ]; then
     OPENSHELL_E2E_KUBE_BUILD_IMAGES=1
@@ -501,7 +518,7 @@ if [ -z "${HOST_GATEWAY_IP}" ] \
     # is unreachable for the typical test-host listener (0.0.0.0 bind).
     detected="$(docker network inspect "${net}" \
       -f '{{range .IPAM.Config}}{{.Gateway}}{{"\n"}}{{end}}' 2>/dev/null \
-      | awk '/^[0-9.]+$/ { print; exit }')"
+      | awk '/^[0-9.]+$/ { print; exit }' || true)"
     if [ -n "${detected}" ]; then
       HOST_GATEWAY_IP="${detected}"
       echo "Detected host gateway IP ${HOST_GATEWAY_IP} from docker network '${net}'."

tasks/helm.toml

@@ -55,16 +55,46 @@ description = "Run skaffold dev for deploy/helm/openshell (iterative deploy)"
 dir = "deploy/helm/openshell"
 run = "skaffold dev"
 
+["helm:skaffold:dev:sidecar"]
+description = "Run skaffold dev with the Kubernetes supervisor sidecar topology"
+dir = "deploy/helm/openshell"
+run = "skaffold dev -p sidecar"
+
+["helm:skaffold:dev:sidecar-mtls"]
+description = "Run skaffold dev with the Kubernetes supervisor sidecar topology and TLS/mTLS enabled"
+dir = "deploy/helm/openshell"
+run = "skaffold dev -p sidecar-mtls"
+
 ["helm:skaffold:run"]
 description = "Run skaffold run for deploy/helm/openshell (one-shot deploy)"
 dir = "deploy/helm/openshell"
 run = "skaffold run"
 
+["helm:skaffold:run:sidecar"]
+description = "Run skaffold run with the Kubernetes supervisor sidecar topology"
+dir = "deploy/helm/openshell"
+run = "skaffold run -p sidecar"
+
+["helm:skaffold:run:sidecar-mtls"]
+description = "Run skaffold run with the Kubernetes supervisor sidecar topology and TLS/mTLS enabled"
+dir = "deploy/helm/openshell"
+run = "skaffold run -p sidecar-mtls"
+
 ["helm:skaffold:delete"]
 description = "Run skaffold delete for deploy/helm/openshell"
 dir = "deploy/helm/openshell"
 run = "skaffold delete"
 
+["helm:skaffold:delete:sidecar"]
+description = "Run skaffold delete for the Kubernetes supervisor sidecar topology"
+dir = "deploy/helm/openshell"
+run = "skaffold delete -p sidecar"
+
+["helm:skaffold:delete:sidecar-mtls"]
+description = "Run skaffold delete for the Kubernetes supervisor sidecar topology with TLS/mTLS enabled"
+dir = "deploy/helm/openshell"
+run = "skaffold delete -p sidecar-mtls"
+
 ["helm:skaffold:diagnose"]
 description = "Run skaffold diagnose for deploy/helm/openshell"
 dir = "deploy/helm/openshell"

tasks/test.toml

@@ -114,6 +114,11 @@ run = [
   "AGENT_SANDBOX_VERSION=v0.4.6 e2e/rust/e2e-kubernetes.sh",
 ]
 
+["e2e:kubernetes:sidecar"]
+description = "Run Kubernetes e2e with the supervisor sidecar topology overlay"
+env = { OPENSHELL_E2E_KUBE_EXTRA_VALUES = "deploy/helm/openshell/ci/values-sidecar.yaml" }
+run = "e2e/rust/e2e-kubernetes.sh"
+
 ["e2e:kubernetes:db"]
 description = "Run Kubernetes e2e with all database backend scenarios (SQLite and external PostgreSQL with existingSecret)"
 env = { OPENSHELL_E2E_KUBE_DB_SCENARIOS = "1" }