feat: add codex agent support to sandbox container

- Install @openai/codex in Dockerfile.sandbox
- Pass developer_instructions via -c flag with proper shell escaping
- Use --yolo flag for sandbox profile (container is the sandbox)
- Mount ~/.codex into container via workmux extra_mounts config

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile.sandbox

@@ -144,6 +144,9 @@ RUN curl -fsSL https://claude.ai/install.sh | bash && \
     ln -s /usr/local/bin/claude /tmp/.local/bin/claude && \
     chmod 777 /tmp/.local
 
+# --- codex agent ---
+RUN npm i -g @openai/codex
+
 # Entrypoint that registers the dynamic UID in /etc/passwd so sudo works.
 # workmux passes --user UID:GID directly, bypassing network-init.sh.
 RUN cat <<'ENTRY' > /usr/local/bin/entrypoint.sh

dev-dashboard/backend/src/workmux.ts

@@ -105,9 +105,11 @@ function buildAgentCmd(profile: Profile, env: Record<string, string>, agent: Age
 
   if (agent === "codex") {
     if (profile === "agent-yolo") {
-      return `workmux sandbox agent -- codex --yolo`;
+      const innerEscaped = prompt.replace(/["\\$`]/g, "\\$&");
+      return `workmux sandbox agent -- codex --yolo -c '"developer_instructions=${innerEscaped}"'`;
     }
-    return "codex";
+    const escapedPrompt = prompt.replace(/'/g, "'\\''");
+    return `codex -c 'developer_instructions=${escapedPrompt}'`;
   }
 
   // Claude