security: drop guessable LITELLM_MASTER_KEY fallback on gateway clients

hermes/entrypoint.sh

@@ -56,11 +56,16 @@ fi
 
 HERMES_BIN=/opt/hermes-agent/.venv/bin/hermes
 
+# Fail loud: LITELLM_MASTER_KEY is the gateway bearer this agent authenticates with. It arrives
+# from secrets.env (SOPS) via the service's env_file. A guessable `local` default silently locks
+# the agent out once the gateway rejects it — refuse to start with no key rather than seed a bad one.
+: "${LITELLM_MASTER_KEY:?LITELLM_MASTER_KEY must be set (SOPS/secrets.env) — refusing to seed a guessable default}"
+
 # Seed model + MCP endpoints to Docker-network DNS. hermes config set is idempotent
 # and overwrites stale values (e.g. localhost: from a prior host-mode install).
 gosu hermes "$HERMES_BIN" config set model.provider        "custom"                        >/dev/null
 gosu hermes "$HERMES_BIN" config set model.base_url        "http://model-gateway:11435/v1" >/dev/null
-gosu hermes "$HERMES_BIN" config set model.api_key         "${LITELLM_MASTER_KEY:-local}"  >/dev/null
+gosu hermes "$HERMES_BIN" config set model.api_key         "${LITELLM_MASTER_KEY}"         >/dev/null
 gosu hermes "$HERMES_BIN" config set model.default         "local-chat"                    >/dev/null
 # Context window: single source of truth is LLAMACPP_CTX_SIZE in .env. The
 # compose file plumbs it into this container's env; the seed below overwrites

services/automation/plugin.yaml

@@ -25,7 +25,9 @@ services:
       N8N_PROXY_HOPS: "1"
       N8N_USER_MANAGEMENT_DISABLED: "true"
       OPENAI_API_BASE_URL: ${OPENAI_API_BASE:-http://model-gateway:11435/v1}
-      OPENAI_API_KEY: local
+      # Interpolated from secrets.env at compose-up (both --env-file passed); no guessable
+      # default — an unset key fails loud (gateway rejects) rather than silently sending `local`.
+      OPENAI_API_KEY: ${LITELLM_MASTER_KEY}
       WEBHOOK_URL: ${N8N_WEBHOOK_URL:-}
       N8N_EDITOR_BASE_URL: ${N8N_WEBHOOK_URL:-}
       N8N_AUTH_EXCLUDE_ENDPOINTS: rest/oauth2-credential/callback,webhook

services/comfyui/plugin.yaml

@@ -38,7 +38,7 @@ services:
       PYTORCH_CUDA_ALLOC_CONF: ${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True,pinned_use_cuda_host_register:True}
       OPS_CONTROLLER_URL: http://ops-controller:9000
       OPS_CONTROLLER_TOKEN: ${OPS_CONTROLLER_TOKEN:-}
-      LITELLM_MASTER_KEY: ${LITELLM_MASTER_KEY:-local}
+      LITELLM_MASTER_KEY: ${LITELLM_MASTER_KEY}
       HF_TOKEN: ${HF_TOKEN:-}
       GITHUB_TOKEN: ${GITHUB_PERSONAL_ACCESS_TOKEN:-}
     volumes:

services/open-webui/plugin.yaml

@@ -22,15 +22,15 @@ services:
     env:
       ENABLE_OLLAMA_API: "false"
       OPENAI_API_BASE_URL: ${OPENAI_API_BASE:-http://model-gateway:11435/v1}
-      OPENAI_API_KEY: ${LITELLM_MASTER_KEY:-local}
+      OPENAI_API_KEY: ${LITELLM_MASTER_KEY}
       WEBUI_AUTH: ${WEBUI_AUTH:-False}
       DEFAULT_MODELS: ${OPEN_WEBUI_DEFAULT_MODEL:-local-chat}
       VECTOR_DB: qdrant
       QDRANT_URI: http://qdrant:6333
       QDRANT_URL: http://qdrant:6333
       RAG_EMBEDDING_ENGINE: openai
       RAG_OPENAI_API_BASE_URL: http://model-gateway:11435/v1
-      RAG_OPENAI_API_KEY: ${LITELLM_MASTER_KEY:-local}
+      RAG_OPENAI_API_KEY: ${LITELLM_MASTER_KEY}
       RAG_EMBEDDING_MODEL: ${EMBED_MODEL:-${LLAMACPP_EMBED_MODEL:-nomic-embed-text-v1.5.Q4_K_M.gguf}}
     volumes:
       - ${DATA_PATH:-./data}/open-webui:/app/backend/data

services/v1-parity/dashboard.yaml

@@ -41,7 +41,7 @@ environment:
   # Backend service URLs the dashboard reads at runtime (all V2 service names).
   LLAMACPP_URL: http://llamacpp:8080
   MODEL_GATEWAY_URL: http://model-gateway:11435
-  MODEL_GATEWAY_API_KEY: ${LITELLM_MASTER_KEY:-local}
+  MODEL_GATEWAY_API_KEY: ${LITELLM_MASTER_KEY}
   MCP_GATEWAY_URL: http://mcp-gateway:8811
   # Single source of truth for the MCP server list: the SAME out/mcp/servers.txt the
   # gateway reads (mounted below at /mcp-config). Without this the dashboard falls back