补齐Anthropic推理捕获

传递capture_reasoning并解析thinking内容增加非流式与流式推理回归测试
2026-09-14 20:50:20 +08:00
parent 59289199ed
commit 6ba4b46b4f
2 changed files with 107 additions and 9 deletions
@@ -19,7 +19,7 @@
 2. `OpenAiChat`、`OpenAiResponses` 与 `Anthropic` 三类 API kind 都支持 JSON 请求、非流式响应和 SSE 流式响应；默认 API kind 仍为 `OpenAiResponses`。
 3. 三类协议都使用统一的 `function_tools` / `tool_choice` 输入和 `LlmRunResponse.tool_calls` 输出。Anthropic 请求使用顶层 `tools[].input_schema` 与对象形态 `tool_choice`；Anthropic URL 默认在 base URL 后拼 `/v1/messages`，如果 base URL 已以 `/v1` 结尾则只拼 `/messages`。
 4. Anthropic 当前仍不支持 `web_search`、图片内容和纯 system 消息；至少需要一条非 system 文本消息。角色动画、图片、视频、资产轮询仍留在其他平台适配和业务模块任务里。
-5. 流式 `on_delta` 发送正文增量、可选的独立 reasoning 增量与完成原因；reasoning 只有在 `LlmRunRequest.capture_reasoning=true` 时才累计，默认关闭。Responses 的 reasoning summary 按 `summary_index` 分段累计，单段 `.done` 只校正对应段，不覆盖其它段。工具调用增量在 crate 内按 slot 聚合，完整调用只从最终 `LlmRunResponse.tool_calls` 读取。reasoning 不进入 `delta_text`、`accumulated_text`、正式 assistant message 或工具参数；上下文管理、后台执行和业务状态不写回本 crate。
+5. 流式 `on_delta` 发送正文增量、可选的独立 reasoning 增量与完成原因；reasoning 只有在 `LlmRunRequest.capture_reasoning=true` 时才累计，默认关闭。Responses 的 reasoning summary 按 `summary_index` 分段累计，单段 `.done` 只校正对应段，不覆盖其它段；Anthropic 的 `thinking_delta` 同样进入独立 reasoning 通道。工具调用增量在 crate 内按 slot 聚合，完整调用只从最终 `LlmRunResponse.tool_calls` 读取。reasoning 不进入 `delta_text`、`accumulated_text`、正式 assistant message 或工具参数；上下文管理、后台执行和业务状态不写回本 crate。
 6. 支持按 provider 打标签，但不把业务 prompt、SSE 转发和模块状态写回本 crate。
 7. `DashScope` 当前只通过“调用方显式提供兼容文本网关 base url”的方式接入，不复用图像 API。
 8. 角色动画、图片、视频、资产轮询仍留在后续 `platform-llm` / `platform-oss` / 业务模块任务里另行实现。
@@ -667,6 +667,8 @@ struct AnthropicContentBlock {
    block_type: Option<String>,
    #[serde(default)]
    text: Option<String>,
+    #[serde(default)]
+    thinking: Option<String>,
    // tool_use block 字段：id 与 name 标识调用，input 是已解析的 JSON object。
    #[serde(default)]
    id: Option<String>,
@@ -3335,7 +3337,9 @@ fn parse_text_response(
            capture_reasoning,
            raw_text,
        ),
-        LlmApiKind::Anthropic => parse_anthropic_response(provider, fallback_model, raw_text),
+        LlmApiKind::Anthropic => {
+            parse_anthropic_response(provider, fallback_model, capture_reasoning, raw_text)
+        }
    }
 }

@@ -3470,6 +3474,7 @@ fn parse_responses_response_with_capture(
 fn parse_anthropic_response(
    provider: LlmProvider,
    fallback_model: &str,
+    capture_reasoning: bool,
    raw_text: &str,
 ) -> Result<LlmRunResponse, LlmError> {
    let parsed: AnthropicResponseEnvelope = serde_json::from_str(raw_text).map_err(|error| {
@@ -3494,9 +3499,16 @@ fn parse_anthropic_response(

    Ok(LlmRunResponse {
        provider,
-        model: parsed.model.unwrap_or_else(|| fallback_model.to_string()),
+        model: parsed
+            .model
+            .clone()
+            .unwrap_or_else(|| fallback_model.to_string()),
        text: content,
-        reasoning: String::new(),
+        reasoning: if capture_reasoning {
+            extract_anthropic_reasoning(&parsed).unwrap_or_default()
+        } else {
+            String::new()
+        },
        finish_reason: parsed.stop_reason,
        response_id: parsed.id,
        usage: parsed.usage.map(map_anthropic_usage),
@@ -3606,6 +3618,16 @@ fn extract_anthropic_text(parsed: &AnthropicResponseEnvelope) -> Option<String>
    if text.is_empty() { None } else { Some(text) }
 }

+fn extract_anthropic_reasoning(parsed: &AnthropicResponseEnvelope) -> Option<String> {
+    let mut reasoning = String::new();
+    for block in &parsed.content {
+        if block.block_type.as_deref() == Some("thinking") {
+            append_reasoning(&mut reasoning, block.thinking.as_deref());
+        }
+    }
+    (!reasoning.is_empty()).then_some(reasoning)
+}
+
 fn extract_message_text(choice: &ChatCompletionsChoice) -> Option<String> {
    choice
        .message
@@ -4245,6 +4267,16 @@ fn parse_anthropic_sse_event(data: &str) -> Result<Option<ParsedStreamEvent>, Ll
                }));
            }

+            if delta_type == "thinking_delta" {
+                return Ok(Some(ParsedStreamEvent {
+                    reasoning_delta: delta
+                        .and_then(|value| value.get("thinking"))
+                        .and_then(serde_json::Value::as_str)
+                        .map(str::to_string),
+                    ..Default::default()
+                }));
+            }
+
            if delta_type != "text_delta" {
                return Ok(None);
            }
@@ -5095,8 +5127,9 @@ mod tests {
            ]
        }"#;

-        let response = parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", raw)
-            .expect("tool-only response should parse");
+        let response =
+            parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", false, raw)
+                .expect("tool-only response should parse");

        assert_eq!(response.text, "");
        assert!(response.reasoning.is_empty());
@@ -5123,19 +5156,44 @@ mod tests {
            ]
        }"#;

-        let response = parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", raw)
-            .expect("mixed response should parse");
+        let response =
+            parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", false, raw)
+                .expect("mixed response should parse");

        assert_eq!(response.text, "我来帮你查询。");
        assert_eq!(response.tool_calls.len(), 1);
        assert_eq!(response.tool_calls[0].arguments, "{}");
    }

+    #[test]
+    fn anthropic_response_captures_thinking_only_when_enabled() {
+        let raw = r#"{
+            "id": "msg_thinking",
+            "model": "model-a",
+            "content": [
+                { "type": "thinking", "thinking": "先分析需求。" },
+                { "type": "text", "text": "最终答案" }
+            ],
+            "stop_reason": "end_turn"
+        }"#;
+
+        let captured =
+            parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", true, raw)
+                .expect("Anthropic thinking should parse");
+        assert_eq!(captured.text, "最终答案");
+        assert_eq!(captured.reasoning, "先分析需求。");
+
+        let hidden =
+            parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", false, raw)
+                .expect("Anthropic response should parse without capture");
+        assert!(hidden.reasoning.is_empty());
+    }
+
    #[test]
    fn anthropic_response_without_text_or_tool_calls_is_empty() {
        let raw = r#"{ "id": "msg_3", "model": "model-a", "content": [] }"#;

-        let error = parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", raw)
+        let error = parse_anthropic_response(LlmProvider::OpenAiCompatible, "fallback", false, raw)
            .expect_err("empty content should fail");

        assert_eq!(error, LlmError::EmptyResponse);
@@ -7297,6 +7355,46 @@ mod tests {
        );
    }

+    #[tokio::test]
+    async fn stream_run_captures_anthropic_thinking_separately_when_enabled() {
+        let server_url = spawn_mock_server(vec![MockResponse {
+            status_line: "200 OK",
+            content_type: "text/event-stream; charset=utf-8",
+            body: concat!(
+                r#"data: {"type":"content_block_delta","index":0,"delta":{"type":"thinking_delta","thinking":"先分析。"}}"#, "\n\n",
+                r#"data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"答案"}}"#, "\n\n",
+                r#"data: {"type":"message_delta","delta":{"stop_reason":"end_turn"}}"#, "\n\n",
+                r#"data: {"type":"message_stop"}"#, "\n\n"
+            )
+            .to_string(),
+            extra_headers: Vec::new(),
+        }]);
+
+        let mut updates = Vec::new();
+        let response = build_test_client(server_url, 0)
+            .stream_run(
+                LlmRunRequest::single_turn("系统", "用户")
+                    .with_anthropic()
+                    .with_reasoning_capture(true),
+                |delta| updates.push((delta.delta_text.clone(), delta.reasoning_delta.clone())),
+            )
+            .await
+            .expect("Anthropic thinking stream should parse");
+
+        assert_eq!(response.text, "答案");
+        assert_eq!(response.reasoning, "先分析。");
+        assert!(
+            updates
+                .iter()
+                .any(|(text, reasoning)| text.is_empty() && reasoning == "先分析。")
+        );
+        assert!(
+            updates
+                .iter()
+                .any(|(text, reasoning)| text == "答案" && reasoning.is_empty())
+        );
+    }
+
    #[tokio::test]
    async fn stream_run_keeps_multiple_responses_reasoning_summary_parts() {
        let server_url = spawn_mock_server(vec![MockResponse {