@jambonz/schema 0.4.27 → 0.4.29

package/docs/verbs/agent.md

@@ -327,6 +327,30 @@ The `llm` property is the only required field. It configures the text-to-text LL
 
 For Anthropic models, use `"vendor": "anthropic"` and structure messages accordingly (Anthropic uses `"role": "user"` for the system-level instruction).
 
+### Vendor-specific parameters (`vendorParams`)
+
+`llmOptions.vendorParams` passes vendor-specific request parameters that jambonz has no first-class option for. For example, OpenRouter's provider routing and fallback models:
+
+```json
+{
+  "llm": {
+    "vendor": "openrouter",
+    "model": "meta-llama/llama-3.3-70b-instruct",
+    "llmOptions": {
+      "systemPrompt": "You are a helpful voice assistant.",
+      "vendorParams": {
+        "provider": { "sort": "latency", "data_collection": "deny" },
+        "models": ["openai/gpt-4.1-mini", "google/gemini-3.5-flash"]
+      }
+    }
+  }
+}
+```
+
+- Fields are added to the vendor request: the request body for OpenAI-compatible vendors (openai, openrouter, groq, deepseek, xai, azure-openai, ...) and the generation config for `google` and `vertex-gemini`. `anthropic` and `bedrock` ignore it.
+- Add-only: a field jambonz already sets (`model`, `messages`, `tools`, max tokens, `temperature`, ...) is never overridden, so use the first-class option for those.
+- Not validated by jambonz. A misspelled field, or one the vendor rejects, surfaces as a vendor error on the first turn.
+
 ## Greeting
 
 By default (`greeting: true`), the agent prompts the LLM to generate an initial greeting before the user speaks. Set `greeting: false` if you want the agent to wait silently for the user to speak first.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jambonz/schema",
-  "version": "0.4.27",
+  "version": "0.4.29",
   "description": "JSON Schema definitions and validation for jambonz verb applications",
   "main": "index.js",
   "scripts": {

package/verbs/agent.schema.json

@@ -194,7 +194,13 @@
             "reasoningEffort": {
               "type": "string",
               "enum": ["minimal", "low", "medium", "high"],
-              "description": "Vendor-neutral thinking/reasoning effort. Mapped per-vendor by the LLM adapter (Gemini thinkingLevel, OpenAI reasoning_effort, Anthropic extended thinking); ignored by vendors without a native equivalent. 'minimal' minimizes thinking for lowest TTFT on latency-sensitive voice turns."
+              "deprecated": true,
+              "description": "Deprecated: has no effect. Accepted so existing applications keep validating. To control thinking, set the vendor's own parameter in vendorParams, e.g. {\"reasoning_effort\": \"none\"} for openai, {\"thinkingConfig\": {\"thinkingLevel\": \"minimal\"}} for google, {\"thinking\": {\"type\": \"disabled\"}} for deepseek, {\"reasoning\": {\"effort\": \"minimal\"}} for openrouter."
+            },
+            "vendorParams": {
+              "type": "object",
+              "description": "Vendor-specific request parameters, passed through as-is. Fields are added to the vendor request (the request body for OpenAI-compatible vendors such as openai and openrouter, the generation config for google and vertex-gemini) but never override a field jambonz sets itself (model, messages, tools, max tokens, temperature, ...). Not validated by jambonz; ignored by anthropic and bedrock. Example for openrouter: {\"provider\": {\"sort\": \"latency\"}, \"models\": [\"openai/gpt-4.1-mini\"]}.",
+              "additionalProperties": true
             },
             "tools": {
               "type": "array",