fix: Add test coverage for llm-php-ext to verify that CHAT MODEL now …

manual/english/Searching/Conversational_search.md

@@ -108,8 +108,8 @@ You can also set provider options and retrieval limits:
 ```sql
 CREATE CHAT MODEL support_assistant (
     model='openai:gpt-4o-mini',
-    api_key='your-provider-api-key',
-    base_url='http://host.docker.internal:8787/v1',
+    api_key='your-azure-entra-token',
+    base_url='https://YOUR-RESOURCE-NAME.openai.azure.com/openai/v1/responses',
     timeout=60,
     retrieval_limit=5,
     max_document_length=3000
@@ -125,7 +125,7 @@ CREATE CHAT MODEL support_assistant (
 ```bash
 curl -s -X POST 'http://localhost:9308/sql?mode=raw' \
   -H 'Content-Type: text/plain' \
-  -d "CREATE CHAT MODEL support_assistant (model='openai:gpt-4o-mini', api_key='your-provider-api-key', base_url='http://host.docker.internal:8787/v1', timeout=60, retrieval_limit=5, max_document_length=3000)"
+  -d "CREATE CHAT MODEL support_assistant (model='openai:gpt-4o-mini', api_key='your-azure-entra-token', base_url='https://YOUR-RESOURCE-NAME.openai.azure.com/openai/v1/responses', timeout=60, retrieval_limit=5, max_document_length=3000)"
 ```
 
 <!-- end -->
@@ -161,6 +161,31 @@ environment:
 
 If `api_key` is not set in `CREATE CHAT MODEL`, the `llm` extension can use the matching provider environment variable. Set `api_key` in the chat model only when you need this model to use a different key.
 
+## Local OpenAI-compatible models
+
+LM Studio can run local models with Conversational Search. For a local model identifier that is not in the `openai:*` catalog, use the `openrouter:*` transport with LM Studio's Chat Completions endpoint. When Buddy runs in Docker, use `host.docker.internal` to reach the server on the host.
+
+<!-- example conversational_search_create_local_model -->
+
+<!-- intro -->
+##### SQL:
+
+<!-- request SQL -->
+
+```sql
+CREATE CHAT MODEL local_assistant (
+    model='openrouter:google_gemma-4-e4b-it',
+    api_key='lm-studio',
+    base_url='http://host.docker.internal:1234/v1/chat/completions',
+    timeout=60,
+    retrieval_limit=5
+);
+```
+
+<!-- end -->
+
+`api_key` is a non-secret placeholder accepted by LM Studio; do not use a real provider key for a local server. The `llm` extension validates the model portion of `openai:*` against its supported OpenAI model names, so `openai:google_gemma-4-e4b-it` is rejected even when LM Studio has that model loaded. The `openrouter:*` transport accepts the local identifier and can target an OpenAI-compatible Chat Completions server through `base_url`. Conversational Search requires the local model to reliably return an OpenAI function call for Buddy's routing schema. Test it with `CALL CHAT`; basic text completion or a simple tool-call test alone is not sufficient.
+
 ## CALL CHAT syntax
 
 ```sql

test/clt-tests/buddy-plugins/test-conversational-basic.rec

@@ -191,6 +191,55 @@ mysql -h0 -P9306 -e "CALL CHAT('Can you explain more?', 'docs', 'test_assistant'
 #!/.*response: .+/!#
 #!/.*sources: \[\]/!#
 ––– comment –––
+Create a dedicated chat model for the endpoint probe. It is intentionally
+allowed to receive HTTP 502 responses and is never used by normal chat cases.
+––– input –––
+rm -f /tmp/chat-model-requests.log /tmp/mock-chat-model-server.php
+cat > /tmp/mock-chat-model-server.php <<'PHP'
+<?php
+$logFile = getenv('CHAT_MODEL_REQUEST_LOG') ?: '';
+if ($logFile !== '') {
+    file_put_contents($logFile, ($_SERVER['REQUEST_URI'] ?? '') . PHP_EOL, FILE_APPEND | LOCK_EX);
+}
+http_response_code(502);
+header('Content-Type: application/json');
+echo '{"error":{"message":"CLT endpoint probe"}}';
+PHP
+CHAT_MODEL_REQUEST_LOG=/tmp/chat-model-requests.log php -S 127.0.0.1:18087 /tmp/mock-chat-model-server.php > /tmp/mock-chat-model-server.log 2>&1 &
+MOCK_CHAT_PID=$!
+sleep 1
+mysql -h0 -P9306 << 'EOF'
+CREATE CHAT MODEL endpoint_probe_assistant (
+    model='openai:gpt-4o-mini',
+    api_key='chat-model-key',
+    base_url='http://127.0.0.1:18087/configured/v1/responses',
+    timeout=60
+);
+EOF
+––– output –––
++--------------------------+
+| name                     |
++--------------------------+
+| endpoint_probe_assistant |
++--------------------------+
+––– comment –––
+Test CALL CHAT uses the dedicated model-specific base_url rather than the environment default
+––– input –––
+mysql -h0 -P9306 -e "CALL CHAT('What is vector search?', 'docs', 'endpoint_probe_assistant')" > /dev/null 2>&1 || true
+sort -u /tmp/chat-model-requests.log
+––– output –––
+/configured/v1/responses
+––– comment –––
+Stop the local endpoint probe
+––– input –––
+kill "$MOCK_CHAT_PID" 2>/dev/null || true
+––– output –––
+––– comment –––
+Remove the dedicated endpoint-probe model
+––– input –––
+mysql -h0 -P9306 -e "DROP CHAT MODEL endpoint_probe_assistant;"
+––– output –––
+––– comment –––
 Test DROP CHAT MODEL - cleanup
 ––– input –––
 mysql -h0 -P9306 -e "DROP CHAT MODEL test_assistant;"