feat(caddy): expose LiteLLM at /llm/* for tailnet API clients

auth/caddy/Caddyfile

@@ -45,6 +45,16 @@
         reverse_proxy n8n:5678
     }
 
+    # ---- LiteLLM model gateway (OpenAI-compatible API) for tailnet API clients ----
+    # Bypasses SSO: programmatic clients (Cline, Cursor, scripts) can't do an
+    # interactive Google login. Auth is LiteLLM's own Bearer key
+    # (LITELLM_MASTER_KEY), which model-gateway enforces — so this is safe to
+    # expose over the tailnet (Caddy binds the tailnet interface only).
+    # handle_path strips /llm so LiteLLM sees /v1/... (e.g. /llm/v1/chat/completions).
+    handle_path /llm/* {
+        reverse_proxy model-gateway:11435
+    }
+
     # ---- Open WebUI at ROOT (no auth needed for the redirect) ----
     # Open WebUI's upstream v0.9.2 image is a prebuilt SvelteKit SPA with
     # base="" (assets are root-absolute: /_app, /static, /api, /ws). It MUST be

docker-compose.yml

@@ -160,10 +160,11 @@ services:
       - DASHBOARD_URL=http://dashboard:8080
       - THROUGHPUT_RECORD_TOKEN=${THROUGHPUT_RECORD_TOKEN:-}
     ports:
-      # 127.0.0.1 bind: localhost-only host publish. Tailnet peers reach this
-      # service through Caddy (see auth/caddy/Caddyfile); host apps (Cline,
-      # VS Code, MCP clients, Hermes auth.json) keep their `localhost:11435`
-      # connectivity. Removes the prior 0.0.0.0 LAN exposure.
+      # 127.0.0.1 bind: localhost-only host publish. Tailnet peers reach the
+      # OpenAI-compatible API via Caddy at https://<tailnet-host>/llm/* — gated
+      # by the LiteLLM master key, no SSO (see auth/caddy/Caddyfile). Host apps
+      # (Cline, VS Code, MCP clients, Hermes auth.json) keep their
+      # `localhost:11435` connectivity. Removes the prior 0.0.0.0 LAN exposure.
       - "127.0.0.1:${MODEL_GATEWAY_PORT:-11435}:11435"
     healthcheck:
       test: ["CMD-SHELL", "python3 -c \"import os, urllib.request; req = urllib.request.Request('http://localhost:11435/v1/models', headers={'Authorization': 'Bearer ' + os.environ.get('LITELLM_MASTER_KEY', 'local')}); urllib.request.urlopen(req)\""]
@@ -178,6 +179,8 @@ services:
     networks:
       - frontend
       - backend
+      # proxy-net: lets Caddy (front door) reach this for the /llm/* API route.
+      - proxy-net
 
   ops-controller:
     build: ./ops-controller