fix(llm): omit empty tools and keep tool_choice off the wire (#21)

Co-authored-by: Alice <alice@prismshadow.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/core/src/llm/generative-model.ts

@@ -1095,11 +1095,21 @@ export function toolDefinitionsToSchemas(tools: ToolDefinition[]): ToolSchema[]
   }));
 }
 
-/** Pre-builds UniConfig from GenerativeModelConfig (called once at construction time). */
+/**
+ * Pre-builds UniConfig from GenerativeModelConfig (called once at construction time).
+ *
+ * When the tool list is empty (connectivity probe, bare/meta LLM, vision describer), `tools`
+ * is omitted entirely instead of set to `[]`: strict OpenAI-compatible servers (e.g. vLLM)
+ * reject an empty array with a 400 ("tools must not be an empty array"), and omission is the
+ * protocol equivalent. `tool_choice` is likewise never set — AgentHub only puts it on the wire
+ * when UniConfig defines it, and leaving it off preserves the protocol default ("auto" when
+ * tools are present).
+ */
 export function buildUniConfig(config: GenerativeModelConfig): UniConfig {
-  const uniConfig: UniConfig = {
-    tools: toolDefinitionsToSchemas(config.tools),
-  };
+  const uniConfig: UniConfig = {};
+  if (config.tools.length > 0) {
+    uniConfig.tools = toolDefinitionsToSchemas(config.tools);
+  }
   if (config.systemPrompt !== undefined) {
     uniConfig.system_prompt = config.systemPrompt;
   }

packages/core/test/llm.test.ts

@@ -1048,11 +1048,24 @@ describe("config helpers", () => {
     expect(cfg.tools).toEqual([{ name: "t", description: "d" }]);
 
     const minimal = buildUniConfig({ modelId: "m", tools: [] });
-    expect(minimal.tools).toEqual([]);
+    expect("tools" in minimal).toBe(false);
     expect("system_prompt" in minimal).toBe(false);
     expect("max_tokens" in minimal).toBe(false);
     expect("thinking_level" in minimal).toBe(false);
   });
+
+  it("omits tools when empty and never sets tool_choice (strict endpoints reject both)", () => {
+    // Empty tool list (connectivity probe, bare/meta LLM, vision describer): the `tools` key
+    // must be absent, not `[]` — AgentHub forwards any defined array verbatim, and strict
+    // OpenAI-compatible servers (e.g. vLLM) reject `tools: []` with a 400.
+    const empty = buildUniConfig({ modelId: "m", tools: [] });
+    expect("tools" in empty).toBe(false);
+    // `tool_choice` must never be set: AgentHub only emits it on the wire when UniConfig
+    // defines it, and leaving it off preserves the protocol default.
+    expect("tool_choice" in empty).toBe(false);
+    const withTools = buildUniConfig({ modelId: "m", tools: [{ name: "t", description: "d" }] });
+    expect("tool_choice" in withTools).toBe(false);
+  });
 });
 
 describe("isRetryableError", () => {