GLM-5.1: bump inference-proxy + wire web_context_search env

Image bump
- 05ad3e830dfca99e499705deab54616db483cf0b222be83b591b6037b7fd2abc
+ 59e42dd68faa15eb0c23521029a2fc3d80d86a4143f9f766542357918be33a8c

(nearai/inference-proxy#144 — server-side `web_context_search` agent
loop for /v1/chat/completions; PR merged on main as 177c58c.)

Env wiring
- WEB_CONTEXT_SEARCH_URL=${WEB_CONTEXT_SEARCH_URL}
- WEB_CONTEXT_SEARCH_API_KEY=${WEB_CONTEXT_SEARCH_API_KEY}

Both gated by compose interpolation so the feature stays dormant
until the host's .env defines them (the inference-proxy startup
validation requires both set or both unset, so leaving one of them
hardcoded while the other comes from .env would block startup on
hosts that haven't been provisioned with a Brave key).

To enable on a deployed host, populate .env with:
  WEB_CONTEXT_SEARCH_URL=https://api.search.brave.com/res/v1/llm/context
  WEB_CONTEXT_SEARCH_API_KEY=<brave-subscription-token>

Until then, this is a pure image bump — no behavior change.

Author: Evrard-Nil

GLM-5.1.yaml

@@ -15,7 +15,7 @@ x-nvidia: &nvidia
       hard: 65535
 
 x-vllm-proxy-common: &vllm-proxy-common
-  image: nearaidev/vllm-proxy-rs@sha256:05ad3e830dfca99e499705deab54616db483cf0b222be83b591b6037b7fd2abc
+  image: nearaidev/vllm-proxy-rs@sha256:59e42dd68faa15eb0c23521029a2fc3d80d86a4143f9f766542357918be33a8c
   user: root
   privileged: true
   <<: *nvidia
@@ -94,6 +94,21 @@ services:
       - VLLM_BASE_URL=http://glm51:8000
       - TLS_CERT_PATH=/etc/letsencrypt/live/completions.near.ai/fullchain.pem
       - USE_NV_ATTESTATION_SDK=true
+      # Server-side `web_context_search` tool loop (inference-proxy#144).
+      # Activated only when a client sends
+      # `tools: [{"type":"web_context_search"}]` on /v1/chat/completions
+      # with stream:true. The tool runs inside this CVM, hitting Brave's
+      # LLM Context endpoint directly; tool args + results stay within the
+      # E2EE perimeter. Both URL and key are gated by env interpolation —
+      # if the host's `.env` doesn't define them, compose substitutes
+      # empty strings, the inference-proxy treats the feature as
+      # unconfigured, and existing chat-completion behavior is unchanged.
+      # To enable on a host, set both
+      # `WEB_CONTEXT_SEARCH_URL=https://api.search.brave.com/res/v1/llm/context`
+      # and `WEB_CONTEXT_SEARCH_API_KEY=<brave-subscription-token>` in
+      # the host .env before compose up.
+      - WEB_CONTEXT_SEARCH_URL=${WEB_CONTEXT_SEARCH_URL}
+      - WEB_CONTEXT_SEARCH_API_KEY=${WEB_CONTEXT_SEARCH_API_KEY}
 
   glm51:
     <<: *nvidia