fix(render): let the native tool protocol allow a lead-in line before tool calls

crates/tinytools-agent/src/dialect/test.rs

@@ -835,3 +835,17 @@ fn json_call_rendering_round_trips_through_the_parser() {
     assert_eq!(calls[0].arguments["path"], "a.txt");
     assert_eq!(calls[1].arguments, serde_json::json!({}));
 }
+
+#[test]
+fn native_instructions_allow_a_lead_in_but_require_the_call_in_the_same_message() {
+    let text = crate::render::native_instructions();
+    assert!(text.contains("same message"), "{text}");
+    assert!(
+        text.contains("never end a turn on an announcement"),
+        "{text}"
+    );
+    // The old wording read as a ban on any lead-in text, which stopped models
+    // from streaming a one-line "looking that up" before their tool calls.
+    assert!(!text.contains("Let me check"), "{text}");
+    assert!(!text.contains("narrate intent"), "{text}");
+}

crates/tinytools-agent/src/render/instructions.rs

@@ -71,9 +71,9 @@ pub fn native_instructions() -> String {
     [
         "## Tool Use Protocol",
         "",
-        "When a tool is needed, emit tool calls directly via the model's native tool-calling output.",
-        "Do not only narrate intent (for example, avoid \"Let me check...\") without emitting the tool call.",
-        "After tool results are provided, continue reasoning and then produce the final answer.",
+        "When you need a tool, call it in this response through native tool-calling.",
+        "A short line saying what you are about to do is fine, but the tool call must be in the same message; never end a turn on an announcement.",
+        "After the results arrive, keep going until you can give the final answer.",
         "",
     ]
     .join("\n")