补齐 Anthropic 原生工具与三协议流式工具调用

Anthropic 请求发送 tools 与对象形态 tool_choice 并解析 tool_use block，解除本地对 function tools 的拦截
Chat / Responses / Anthropic 三种协议的流式工具增量按槽位聚合，收尾校验参数为完整 JSON，截断流不返回半截参数
解除 App 侧 Anthropic 降级为文本 JSON 协议的两处守卫与对应提示词分支
新增依赖真实凭据的流式工具验收用例，默认 ignore

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

apps/ai-game-creator-shell/src-tauri/src/agent/interaction.rs

@@ -93,17 +93,13 @@ fn agent_interaction_function_tools() -> Vec<platform_llm::LlmFunctionTool> {
         .collect()
 }
 
-fn agent_interaction_system_prompt(agent_id: &str, native_tools: bool) -> String {
+fn agent_interaction_system_prompt(agent_id: &str) -> String {
     let role_prompt = if agent_id == GAME_CREATOR_PROJECT_SUPERVISOR_AGENT_ID {
         game_creator_project_supervisor_chat_system_prompt()
     } else {
         game_creator_role_agent_chat_system_prompt()
     };
-    let protocol = if native_tools {
+    let protocol = "普通问答、身份说明、架构解释、方案讨论和必要澄清直接用自然语言回复。只有确实需要宿主持久能力时才调用一个 function tool，调用工具时不要同时输出回复文本。不要根据单个关键词决定是否执行，要理解整句的否定、假设、范围和上下文。";
-        "普通问答、身份说明、架构解释、方案讨论和必要澄清直接用自然语言回复。只有确实需要宿主持久能力时才调用一个 function tool，调用工具时不要同时输出回复文本。不要根据单个关键词决定是否执行，要理解整句的否定、假设、范围和上下文。"
-    } else {
-        "你必须只输出一个 JSON 对象，不要代码块或额外文字。允许的结构为：{\"action\":\"reply\",\"reply\":\"自然语言回复\"}、{\"action\":\"execute\"}、{\"action\":\"resume\"}、{\"action\":\"project_location\"}。不要根据单个关键词决定 action，要理解整句的否定、假设、范围和上下文。"
-    };
     format!(
         "{role_prompt}\n\n你现在位于统一的 Agent interaction loop。{protocol} 高影响请求仍不明确时直接追问，不要擅自启动 Runtime。"
     )
@@ -114,7 +110,7 @@ fn build_agent_interaction_request_for_session(
     agent_id: &str,
     session_id: &str,
     prompt: &str,
-) -> Result<(GameCreatorLlmConfig, String, LlmRunRequest, bool), String> {
+) -> Result<(GameCreatorLlmConfig, String, LlmRunRequest), String> {
     let prompt = prompt.trim();
     if prompt.is_empty() {
         return Err("交互内容不能为空".to_string());
@@ -125,7 +121,6 @@ fn build_agent_interaction_request_for_session(
     let (llm, config_path, context) =
         build_game_creator_role_agent_context_for_session(root, agent_id, Some(session_id))?;
     let api_kind = parse_game_creator_llm_api_kind(&llm.api_kind)?;
-    let native_tools = api_kind != LlmApiKind::Anthropic;
     let user_prompt = if context.trim().is_empty() {
         format!("用户这轮输入：\n{prompt}")
     } else {
@@ -133,18 +128,15 @@ fn build_agent_interaction_request_for_session(
             "项目上下文如下。只把它当作背景，不要逐字复述。\n\n{context}\n\n用户这轮输入：\n{prompt}"
         )
     };
-    let mut request = LlmRunRequest::new(vec![
+    let request = LlmRunRequest::new(vec![
-        LlmMessage::system(agent_interaction_system_prompt(agent_id, native_tools)),
+        LlmMessage::system(agent_interaction_system_prompt(agent_id)),
         LlmMessage::user(user_prompt),
     ])
     .with_api_kind(api_kind)
-    .with_max_output_tokens(AGENT_INTERACTION_MAX_OUTPUT_TOKENS);
+    .with_max_output_tokens(AGENT_INTERACTION_MAX_OUTPUT_TOKENS)
-    if native_tools {
+    .with_function_tools(agent_interaction_function_tools())
-        request = request
+    .with_tool_choice(platform_llm::LlmToolChoice::Auto);
-            .with_function_tools(agent_interaction_function_tools())
+    Ok((llm, config_path, request))
-            .with_tool_choice(platform_llm::LlmToolChoice::Auto);
-    }
-    Ok((llm, config_path, request, native_tools))
 }
 
 pub(crate) async fn decide_game_creator_agent_interaction_turn_for_session_at<F>(
@@ -157,10 +149,10 @@ pub(crate) async fn decide_game_creator_agent_interaction_turn_for_session_at<F>
 where
     F: FnMut(&platform_llm::LlmStreamDelta),
 {
-    let (llm, config_path, request, native_tools) =
+    let (llm, config_path, request) =
         build_agent_interaction_request_for_session(root, agent_id, session_id, prompt)?;
     let client = build_game_creator_agent_runtime_llm_client(&llm, &config_path)?;
-    let response = if native_tools && llm.stream {
+    let response = if llm.stream {
         let fallback_request = request.clone();
         match client.stream_run(request, |delta| on_delta(delta)).await {
             Ok(response) => response,

apps/ai-game-creator-shell/src-tauri/src/agent/runtime_actions/provider_request_builders.rs

@@ -139,11 +139,7 @@ pub(in crate::agent) fn build_game_creator_agent_background_tool_plan_request(
         "command.start 使用 {\"program\":\"受信任 PATH 中的裸可执行名\"",
     );
     let api_kind = parse_game_creator_llm_api_kind(&llm.api_kind)?;
-    let protocol_prompt = if api_kind == platform_llm::LlmApiKind::Anthropic {
+    let protocol_prompt = "必须直接调用当前请求提供的原生函数：需要更新持久计划时调用 update_agent_plan，需要行动时调用对应动作工具，已有观察足够时调用 respond_to_user。只有步骤或状态真实变化时才单独调用 update_agent_plan；当前 in_progress 步骤已具备执行条件时必须在同一响应调用对应动作工具，不能只改计划解释。不要调用未广告的旧 submit_agent_tool_plan，也不要把计划或动作放在普通文本中。";
-        "当前 Provider 不提供 function tools，请返回上述 schema 的单个完整 JSON object；不要解释、markdown 或代码围栏。"
-    } else {
-        "必须直接调用当前请求提供的原生函数：需要更新持久计划时调用 update_agent_plan，需要行动时调用对应动作工具，已有观察足够时调用 respond_to_user。只有步骤或状态真实变化时才单独调用 update_agent_plan；当前 in_progress 步骤已具备执行条件时必须在同一响应调用对应动作工具，不能只改计划解释。不要调用未广告的旧 submit_agent_tool_plan，也不要把计划或动作放在普通文本中。"
-    };
     let mut system_prompt = game_creator_agent_runtime_tool_plan_system_prompt_for_agent(agent_id);
     if autonomous_game_build {
         system_prompt.push_str(
@@ -167,12 +163,9 @@ pub(in crate::agent) fn build_game_creator_agent_background_tool_plan_request(
     ])
     .with_api_kind(api_kind)
     .with_max_output_tokens(AGENT_RUNTIME_TOOL_PLAN_MAX_OUTPUT_TOKENS)
-    .with_response_text_verbosity(platform_llm::LlmResponseTextVerbosity::Low);
+    .with_response_text_verbosity(platform_llm::LlmResponseTextVerbosity::Low)
-    if api_kind != platform_llm::LlmApiKind::Anthropic {
+    .with_function_tools(build_agent_runtime_native_function_tools(mcp_catalog)?)
-        request = request
+    .with_tool_choice(platform_llm::LlmToolChoice::Required);
-            .with_function_tools(build_agent_runtime_native_function_tools(mcp_catalog)?)
-            .with_tool_choice(platform_llm::LlmToolChoice::Required);
-    }
     request = apply_game_creator_llm_web_search(
         apply_game_creator_llm_reasoning_effort(request, &llm)?,
         &llm,

server-rs/crates/platform-llm/tests/live_stream_tool_calls.rs

@@ -0,0 +1,101 @@
+//! 真实端点的流式工具调用验收。默认 `#[ignore]`，只在显式指定环境变量时运行：
+//!
+//! ```powershell
+//! $env:PLATFORM_LLM_LIVE_BASE_URL = 'https://api.minimaxi.com/anthropic'
+//! $env:PLATFORM_LLM_LIVE_API_KEY  = '...'
+//! $env:PLATFORM_LLM_LIVE_MODEL    = 'MiniMax-M3'
+//! $env:PLATFORM_LLM_LIVE_API_KIND = 'anthropic'   # 或 openai_chat / openai_responses
+//! cargo test -p platform-llm --test live_stream_tool_calls -- --ignored --nocapture
+//! ```
+//!
+//! 单测里的 SSE 是转录的真实报文，这个用例负责证明转录没有偏差。
+
+use platform_llm::{
+    LlmApiKind, LlmClient, LlmConfig, LlmFunctionTool, LlmMessage, LlmProvider, LlmRunRequest,
+    LlmToolChoice,
+};
+
+fn env_var(name: &str) -> Option<String> {
+    std::env::var(name).ok().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,
+    }
+}
+
+#[tokio::test]
+#[ignore = "需要真实 Provider 凭据，用 --ignored 显式运行"]
+async fn live_stream_run_returns_native_tool_calls() {
+    let (Some(base_url), Some(api_key), Some(model)) = (
+        env_var("PLATFORM_LLM_LIVE_BASE_URL"),
+        env_var("PLATFORM_LLM_LIVE_API_KEY"),
+        env_var("PLATFORM_LLM_LIVE_MODEL"),
+    ) 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 config = LlmConfig::new(
+        LlmProvider::OpenAiCompatible,
+        base_url,
+        api_key,
+        model,
+        120_000,
+        0,
+        1_000,
+    )
+    .expect("live config should be valid");
+    let client = LlmClient::new(config).expect("live client should be created");
+
+    let request = LlmRunRequest::new(vec![
+        LlmMessage::system("你可以使用工具。需要外部数据时必须调用工具，不要凭空回答。"),
+        LlmMessage::user("杭州现在天气怎么样？"),
+    ])
+    .with_api_kind(api_kind)
+    .with_max_output_tokens(512)
+    .with_function_tools(vec![LlmFunctionTool::new(
+        "get_weather",
+        "查询指定城市的当前天气。",
+        serde_json::json!({
+            "type": "object",
+            "properties": { "city": { "type": "string" } },
+            "required": ["city"]
+        }),
+    )])
+    .with_tool_choice(LlmToolChoice::Required);
+
+    let mut streamed_chars = 0usize;
+    let response = client
+        .stream_run(request, |delta| {
+            streamed_chars += delta.delta_text.chars().count();
+        })
+        .await
+        .expect("live stream_run should succeed");
+
+    println!(
+        "api_kind={api_kind:?} finish_reason={:?} streamed_chars={streamed_chars} text={:?}",
+        response.finish_reason, response.text
+    );
+    for call in &response.tool_calls {
+        println!("tool_call id={} name={} args={}", call.id, call.name, call.arguments);
+    }
+
+    assert!(
+        !response.tool_calls.is_empty(),
+        "流式必须解析出工具调用，实际 finish_reason={:?}",
+        response.finish_reason
+    );
+    let call = &response.tool_calls[0];
+    assert_eq!(call.name, "get_weather");
+    assert!(!call.id.trim().is_empty(), "工具调用必须带 id");
+    let arguments: serde_json::Value =
+        serde_json::from_str(&call.arguments).expect("参数必须是完整 JSON");
+    assert!(
+        arguments.get("city").is_some(),
+        "参数应包含 city，实际为 {arguments}"
+    );
+}