fix(policy): allow npm child process and Vertex AI egress for opencode

The opencode network policy allowlists registry.npmjs.org as an endpoint
but not the npm binary, so opencode's background dependency install
(arborist) connects as /usr/local/bin/npm and is denied (ECONNRESET).
Add /usr/local/bin/npm and /usr/bin/npm to the binaries allowlist,
matching the dedicated npm policies in the droid and ollama sandboxes.

The policy also had no Google egress, so running opencode against a
google-vertex-* provider could not reach Vertex AI or the Google token
endpoints. Add the Vertex AI + token hosts (mirroring the gemini policy)
plus sts.googleapis.com, which GitHub-OIDC Workload Identity Federation
requires for token exchange. Also add models.opencode.ai so opencode
does not silently fall back to its built-in default model.

Applied identically to the base and gemini sandbox policies.

Signed-off-by: Yvonne Devlin <ydevlin@redhat.com>

Author: yvonnedevlinrh

sandboxes/base/policy.yaml

@@ -177,12 +177,32 @@ network_policies:
       port: 443
     - host: opencode.ai
       port: 443
+    - host: models.opencode.ai
+      port: 443
     - host: integrate.api.nvidia.com
       port: 443
+    # Google Vertex AI + token endpoints (mirrors gemini policy). sts is
+    # additionally required for GitHub-OIDC Workload Identity Federation.
+    - host: "*-aiplatform.googleapis.com"
+      port: 443
+    - host: sts.googleapis.com
+      port: 443
+    - host: oauth2.googleapis.com
+      port: 443
+    - host: www.googleapis.com
+      port: 443
+    - host: iamcredentials.googleapis.com
+      port: 443
+    - host: accounts.google.com
+      port: 443
     binaries:
     - path: /usr/lib/node_modules/opencode-ai/bin/.opencode
     - path: /usr/bin/node
     - path: /usr/local/bin/opencode
+    # npm child process (background dependency install / arborist) connects
+    # as npm, not opencode/node, so it must be allowlisted explicitly.
+    - path: /usr/local/bin/npm
+    - path: /usr/bin/npm
 
   codex:
     name: codex

sandboxes/gemini/policy.yaml

@@ -160,12 +160,32 @@ network_policies:
         port: 443
       - host: opencode.ai
         port: 443
+      - host: models.opencode.ai
+        port: 443
       - host: integrate.api.nvidia.com
         port: 443
+      # Google Vertex AI + token endpoints. sts is additionally required
+      # for GitHub-OIDC Workload Identity Federation.
+      - host: "*-aiplatform.googleapis.com"
+        port: 443
+      - host: sts.googleapis.com
+        port: 443
+      - host: oauth2.googleapis.com
+        port: 443
+      - host: www.googleapis.com
+        port: 443
+      - host: iamcredentials.googleapis.com
+        port: 443
+      - host: accounts.google.com
+        port: 443
     binaries:
       - path: /usr/lib/node_modules/opencode-ai/bin/.opencode
       - path: /usr/bin/node
       - path: /usr/local/bin/opencode
+      # npm child process (background dependency install / arborist)
+      # connects as npm, not opencode/node, so allowlist it explicitly.
+      - path: /usr/local/bin/npm
+      - path: /usr/bin/npm
 
   copilot:
     name: copilot