diff --git a/server-rs/crates/platform-llm/README.md b/server-rs/crates/platform-llm/README.md
index 8a9313433..9fcacd5c2 100644
--- a/server-rs/crates/platform-llm/README.md
+++ b/server-rs/crates/platform-llm/README.md
@@ -88,5 +88,5 @@ Responses 如果只发送 `response.completed` 或 `response.incomplete`，解
 ## 9. 验收证据边界
 
 1. `cargo test -p platform-llm` 的确定性用例把 checked-in SSE fixture 交给 parser，验证归一后的文本、工具调用、slot 聚合、Responses 仅有 completed / incomplete 终态事件时的恢复、参数 JSON 完整性和错误边界。fixture 可以来源于真实端点抓包，但测试不保存原始 SSE，也不逐事件与端点报文比较，因此不能证明抓包转录无偏差。
-2. `tests/live_stream_tool_calls.rs` 是默认 `#[ignore]` 的真实端点工具调用 smoke。它只验证最终归一结果中的工具名、id 和完整参数 JSON；文本增量字符数仅用于打印观测，工具调用不进入 `on_delta`，也没有原始 SSE 录制或逐事件对比能力。
+2. `tests/live_stream_tool_calls.rs` 是默认 `#[ignore]` 的真实端点工具调用 smoke。`PLATFORM_LLM_LIVE_API_KIND` 支持 `openai_responses`、`openai_chat` 和 `anthropic`，未设置或空白时默认 `openai_responses`，未知非空值会直接使验收失败。它只验证最终归一结果中的工具名、id 和完整参数 JSON；文本增量字符数仅用于打印观测，工具调用不进入 `on_delta`，也没有原始 SSE 录制或逐事件对比能力。
 3. 因此验收应分别称为“固定 SSE fixture parser 覆盖”和“真实端点归一工具调用 smoke”，不能把后者描述为原始 SSE fidelity 或转录一致性证明。
diff --git a/server-rs/crates/platform-llm/tests/live_stream_tool_calls.rs b/server-rs/crates/platform-llm/tests/live_stream_tool_calls.rs
index 8307815e8..d40ee8def 100644
--- a/server-rs/crates/platform-llm/tests/live_stream_tool_calls.rs
+++ b/server-rs/crates/platform-llm/tests/live_stream_tool_calls.rs
@@ -26,11 +26,58 @@ fn env_var(name: &str) -> Option<String> {
         .filter(|value| !value.trim().is_empty())
 }
 
-fn parse_api_kind(value: &str) -> LlmApiKind {
-    match value.trim().to_ascii_lowercase().replace('-', "_").as_str() {
-        "anthropic" => LlmApiKind::Anthropic,
-        "openai_chat" => LlmApiKind::OpenAiChat,
-        _ => LlmApiKind::OpenAiResponses,
+fn parse_api_kind(value: &str) -> Result<LlmApiKind, String> {
+    let normalized = value.trim().to_ascii_lowercase().replace('-', "_");
+    if normalized.is_empty() {
+        return Ok(LlmApiKind::OpenAiResponses);
+    }
+
+    match normalized.as_str() {
+        "anthropic" => Ok(LlmApiKind::Anthropic),
+        "openai_chat" => Ok(LlmApiKind::OpenAiChat),
+        "openai_responses" => Ok(LlmApiKind::OpenAiResponses),
+        value => Err(format!(
+            "PLATFORM_LLM_LIVE_API_KIND 无效：{value}，请使用 openai_responses、openai_chat 或 anthropic"
+        )),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parse_api_kind_defaults_only_for_empty_value() {
+        assert_eq!(
+            parse_api_kind("").expect("empty api kind should default"),
+            LlmApiKind::OpenAiResponses
+        );
+        assert_eq!(
+            parse_api_kind("  ").expect("whitespace api kind should default"),
+            LlmApiKind::OpenAiResponses
+        );
+    }
+
+    #[test]
+    fn parse_api_kind_accepts_supported_values() {
+        assert_eq!(
+            parse_api_kind("anthropic").expect("anthropic should parse"),
+            LlmApiKind::Anthropic
+        );
+        assert_eq!(
+            parse_api_kind("OPENAI-CHAT").expect("openai chat should parse"),
+            LlmApiKind::OpenAiChat
+        );
+        assert_eq!(
+            parse_api_kind("openai_responses").expect("openai responses should parse"),
+            LlmApiKind::OpenAiResponses
+        );
+    }
+
+    #[test]
+    fn parse_api_kind_rejects_unknown_non_empty_value() {
+        let error = parse_api_kind("anthopic").expect_err("misspelled api kind must fail");
+        assert!(error.contains("anthopic"));
     }
 }
 
@@ -44,7 +91,12 @@ async fn live_stream_run_returns_native_tool_calls() {
     ) else {
         panic!("缺少 PLATFORM_LLM_LIVE_BASE_URL / _API_KEY / _MODEL");
     };
-    let api_kind = parse_api_kind(&env_var("PLATFORM_LLM_LIVE_API_KIND").unwrap_or_default());
+    let api_kind = parse_api_kind(
+        env_var("PLATFORM_LLM_LIVE_API_KIND")
+            .as_deref()
+            .unwrap_or_default(),
+    )
+    .unwrap_or_else(|error| panic!("{error}"));
 
     let config = LlmConfig::new(
         LlmProvider::OpenAiCompatible,