| 1 | use futures_util::StreamExt; |
| 2 | |
| 3 | use crate::llm_client::LlmClient; |
| 4 | use crate::llm_client::mock::{MockLlmClient, canned}; |
| 5 | use codewhale_models::Role; |
| 6 | use codewhale_models::{ContentBlock, Message, MessageRequest}; |
| 7 | |
| 8 | fn user_message(text: &str) -> Message { |
| 9 | Message { |
| 10 | role: Role::User, |
| 11 | content: vec![ContentBlock::Text { |
| 12 | text: text.to_string(), |
| 13 | cache_control: None, |
| 14 | }], |
| 15 | } |
| 16 | } |
| 17 | |
| 18 | fn assistant_thinking_tool_call( |
| 19 | thinking: &str, |
| 20 | id: &str, |
| 21 | name: &str, |
| 22 | input: serde_json::Value, |
| 23 | ) -> Message { |
| 24 | Message { |
| 25 | role: Role::Assistant, |
| 26 | content: vec![ |
| 27 | ContentBlock::Thinking { |
| 28 | thinking: thinking.to_string(), |
| 29 | signature: None, |
| 30 | state: None, |
| 31 | }, |
| 32 | ContentBlock::ToolUse { |
| 33 | execution_id: None, |
| 34 | id: id.to_string(), |
| 35 | name: name.to_string(), |
| 36 | input, |
| 37 | caller: None, |
| 38 | thought_signature: None, |
| 39 | }, |
| 40 | ], |
| 41 | } |
| 42 | } |
| 43 | |
| 44 | fn tool_result_message(tool_use_id: &str, content: &str) -> Message { |
| 45 | Message { |
| 46 | role: Role::User, |
| 47 | content: vec![ContentBlock::ToolResult { |
| 48 | execution_id: None, |
| 49 | tool_use_id: tool_use_id.to_string(), |
| 50 | content: content.to_string(), |
| 51 | is_error: None, |
| 52 | content_blocks: None, |
| 53 | }], |
| 54 | } |
| 55 | } |
| 56 | |
| 57 | fn make_request(messages: Vec<Message>) -> MessageRequest { |
| 58 | MessageRequest { |
| 59 | model: "deepseek-v4-pro".to_string(), |
| 60 | messages, |
| 61 | max_tokens: 4096, |
| 62 | system: None, |
| 63 | tools: None, |
| 64 | tool_choice: None, |
| 65 | metadata: None, |
| 66 | thinking: None, |
| 67 | reasoning_effort: Some("high".to_string()), |
| 68 | stream: Some(true), |
| 69 | temperature: None, |
| 70 | top_p: None, |
| 71 | } |
| 72 | } |
| 73 | |
| 74 | #[tokio::test] |
| 75 | async fn reasoning_content_is_replayed_after_thinking_tool_call() { |
| 76 | let mock = MockLlmClient::new(vec![]); |
| 77 | |
| 78 | mock.push_turn(vec![ |
| 79 | canned::message_start("r1"), |
| 80 | canned::thinking_delta(0, "I should inspect /tmp before answering."), |
| 81 | canned::tool_use_block_start(1, "call_a", "list_dir"), |
| 82 | canned::tool_input_delta(1, r#"{"path":"/tmp"}"#), |
| 83 | canned::block_stop(1), |
| 84 | canned::message_delta("tool_use", None), |
| 85 | canned::message_stop(), |
| 86 | ]); |
| 87 | |
| 88 | mock.push_factory(|request| { |
| 89 | let assistant = request |
| 90 | .messages |
| 91 | .iter() |
| 92 | .rev() |
| 93 | .find(|message| message.role == "assistant") |
| 94 | .expect("follow-up request must include the prior assistant tool-call turn"); |
| 95 | |
| 96 | assert!( |
| 97 | assistant |
| 98 | .content |
| 99 | .iter() |
| 100 | .any(|block| matches!(block, ContentBlock::Thinking { .. })), |
| 101 | "DeepSeek V4 follow-up requests must replay reasoning_content on the assistant tool-call turn" |
| 102 | ); |
| 103 | |
| 104 | canned::simple_text_turn("I see the /tmp entries.") |
| 105 | }); |
| 106 | |
| 107 | let mut first = mock |
| 108 | .create_message_stream(make_request(vec![user_message("list /tmp")])) |
| 109 | .await |
| 110 | .expect("first stream opens"); |
| 111 | while first.next().await.is_some() {} |
| 112 | |
| 113 | let mut second = mock |
| 114 | .create_message_stream(make_request(vec![ |
| 115 | user_message("list /tmp"), |
| 116 | assistant_thinking_tool_call( |
| 117 | "I should inspect /tmp before answering.", |
| 118 | "call_a", |
| 119 | "list_dir", |
| 120 | serde_json::json!({ "path": "/tmp" }), |
| 121 | ), |
| 122 | tool_result_message("call_a", "/tmp/file1\n/tmp/file2"), |
| 123 | ])) |
| 124 | .await |
| 125 | .expect("second stream opens"); |
| 126 | while second.next().await.is_some() {} |
| 127 | } |
| 128 |