| 1 | //! Drive `parse_sse_chunk` (the in-place SSE event extractor) over canned |
| 2 | //! chunk sequences. The full `handle_chat_completion_stream` path needs a |
| 3 | //! live `reqwest::Response` so it isn't unit-testable without a mock HTTP |
| 4 | //! harness (issue #69 tracks that). For #103 we exercise the chunk decoder |
| 5 | //! directly to verify each "class of stream failure" the engine relies on. |
| 6 | use super::*; |
| 7 | use crate::client::wire::{InvalidSseUtf8, SseLineDecoder}; |
| 8 | use codewhale_models::{ContentBlockStart, Delta, StreamEvent}; |
| 9 | |
| 10 | /// Decode a raw SSE-data JSON chunk into our internal events, mirroring |
| 11 | /// the per-event call shape used by `handle_chat_completion_stream`. |
| 12 | fn decode_chunk(json_text: &str) -> Vec<StreamEvent> { |
| 13 | decode_chunk_with_reasoning(json_text, true) |
| 14 | } |
| 15 | |
| 16 | fn decode_chunk_with_reasoning(json_text: &str, is_reasoning_model: bool) -> Vec<StreamEvent> { |
| 17 | let chunk: Value = serde_json::from_str(json_text).expect("valid SSE JSON"); |
| 18 | let mut content_index = 0u32; |
| 19 | let mut text_started = false; |
| 20 | let mut thinking_started = false; |
| 21 | let mut tool_indices = std::collections::HashMap::new(); |
| 22 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 23 | parse_sse_chunk( |
| 24 | &chunk, |
| 25 | &mut content_index, |
| 26 | &mut text_started, |
| 27 | &mut thinking_started, |
| 28 | &mut tool_indices, |
| 29 | &mut reasoning_detail_buffers, |
| 30 | is_reasoning_model, |
| 31 | ) |
| 32 | } |
| 33 | |
| 34 | fn decode_chunks_with_style( |
| 35 | chunks: &[&str], |
| 36 | reasoning_stream_style: ReasoningStreamStyle, |
| 37 | ) -> Vec<StreamEvent> { |
| 38 | let mut content_index = 0u32; |
| 39 | let mut text_started = false; |
| 40 | let mut thinking_started = false; |
| 41 | let mut tool_indices = std::collections::HashMap::new(); |
| 42 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 43 | let mut inline_reasoning_tags = InlineReasoningTagState::default(); |
| 44 | let mut events = Vec::new(); |
| 45 | |
| 46 | for chunk in chunks { |
| 47 | let value: Value = serde_json::from_str(chunk).expect("valid SSE JSON"); |
| 48 | events.extend(parse_sse_chunk_with_reasoning_style( |
| 49 | &value, |
| 50 | &mut content_index, |
| 51 | &mut text_started, |
| 52 | &mut thinking_started, |
| 53 | &mut tool_indices, |
| 54 | &mut reasoning_detail_buffers, |
| 55 | &mut inline_reasoning_tags, |
| 56 | reasoning_stream_style, |
| 57 | )); |
| 58 | } |
| 59 | events |
| 60 | } |
| 61 | |
| 62 | /// Drive the Chat Completions SSE path with raw byte chunks so tests can |
| 63 | /// split a multi-byte UTF-8 character across HTTP/2-style DATA boundaries. |
| 64 | fn decode_sse_byte_chunks(chunks: &[&[u8]]) -> Result<Vec<StreamEvent>, InvalidSseUtf8> { |
| 65 | struct FrameState { |
| 66 | line_buf: String, |
| 67 | content_index: u32, |
| 68 | text_started: bool, |
| 69 | thinking_started: bool, |
| 70 | tool_indices: std::collections::HashMap<u32, u32>, |
| 71 | reasoning_detail_buffers: std::collections::HashMap<u32, String>, |
| 72 | inline_reasoning_tags: InlineReasoningTagState, |
| 73 | events: Vec<StreamEvent>, |
| 74 | } |
| 75 | |
| 76 | impl FrameState { |
| 77 | fn new() -> Self { |
| 78 | Self { |
| 79 | line_buf: String::new(), |
| 80 | content_index: 0, |
| 81 | text_started: false, |
| 82 | thinking_started: false, |
| 83 | tool_indices: std::collections::HashMap::new(), |
| 84 | reasoning_detail_buffers: std::collections::HashMap::new(), |
| 85 | inline_reasoning_tags: InlineReasoningTagState::default(), |
| 86 | events: Vec::new(), |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | fn handle_line(&mut self, line: &str) -> bool { |
| 91 | if line.is_empty() { |
| 92 | return matches!(self.flush_frame(), SseDataFrame::Done); |
| 93 | } |
| 94 | if let Some(data) = extract_sse_data_value(line) { |
| 95 | if !self.line_buf.is_empty() { |
| 96 | self.line_buf.push('\n'); |
| 97 | } |
| 98 | self.line_buf.push_str(data); |
| 99 | } |
| 100 | false |
| 101 | } |
| 102 | |
| 103 | fn flush_frame(&mut self) -> SseDataFrame { |
| 104 | if self.line_buf.is_empty() { |
| 105 | return SseDataFrame::Events(Vec::new()); |
| 106 | } |
| 107 | let data = std::mem::take(&mut self.line_buf); |
| 108 | match parse_sse_data_frame( |
| 109 | &data, |
| 110 | &mut self.content_index, |
| 111 | &mut self.text_started, |
| 112 | &mut self.thinking_started, |
| 113 | &mut self.tool_indices, |
| 114 | &mut self.reasoning_detail_buffers, |
| 115 | &mut self.inline_reasoning_tags, |
| 116 | ReasoningStreamStyle::SeparateField, |
| 117 | ) { |
| 118 | SseDataFrame::Done => SseDataFrame::Done, |
| 119 | SseDataFrame::Events(frame_events) => { |
| 120 | self.events.extend(frame_events); |
| 121 | SseDataFrame::Events(Vec::new()) |
| 122 | } |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | let mut decoder = SseLineDecoder::new(); |
| 128 | let mut state = FrameState::new(); |
| 129 | for chunk in chunks { |
| 130 | for line in decoder.push(chunk)? { |
| 131 | if state.handle_line(&line) { |
| 132 | return Ok(state.events); |
| 133 | } |
| 134 | } |
| 135 | } |
| 136 | if let Some(line) = decoder.finish()? |
| 137 | && state.handle_line(&line) |
| 138 | { |
| 139 | return Ok(state.events); |
| 140 | } |
| 141 | state.flush_frame(); |
| 142 | Ok(state.events) |
| 143 | } |
| 144 | |
| 145 | fn cjk_content_sse(text: &str) -> Vec<u8> { |
| 146 | let payload = serde_json::json!({ |
| 147 | "choices": [{ "delta": { "content": text } }] |
| 148 | }); |
| 149 | format!("data: {payload}\n\n").into_bytes() |
| 150 | } |
| 151 | |
| 152 | fn mid_char_split(bytes: &[u8], ch: char) -> usize { |
| 153 | let needle = ch.to_string(); |
| 154 | let start = bytes |
| 155 | .windows(needle.len()) |
| 156 | .position(|window| window == needle.as_bytes()) |
| 157 | .unwrap_or_else(|| panic!("{ch:?} present in SSE frame")); |
| 158 | start + 1 |
| 159 | } |
| 160 | |
| 161 | fn text_delta_text(events: &[StreamEvent]) -> String { |
| 162 | events |
| 163 | .iter() |
| 164 | .filter_map(|event| match event { |
| 165 | StreamEvent::ContentBlockDelta { |
| 166 | delta: Delta::TextDelta { text }, |
| 167 | .. |
| 168 | } => Some(text.as_str()), |
| 169 | _ => None, |
| 170 | }) |
| 171 | .collect() |
| 172 | } |
| 173 | |
| 174 | fn thinking_delta_text(events: &[StreamEvent]) -> String { |
| 175 | events |
| 176 | .iter() |
| 177 | .filter_map(|event| match event { |
| 178 | StreamEvent::ContentBlockDelta { |
| 179 | delta: Delta::ThinkingDelta { thinking }, |
| 180 | .. |
| 181 | } => Some(thinking.as_str()), |
| 182 | _ => None, |
| 183 | }) |
| 184 | .collect() |
| 185 | } |
| 186 | |
| 187 | #[test] |
| 188 | fn decoder_reassembles_cjk_split_across_byte_chunks() { |
| 189 | let frame = cjk_content_sse("你好世界"); |
| 190 | let split = mid_char_split(&frame, '好'); |
| 191 | let events = decode_sse_byte_chunks(&[&frame[..split], &frame[split..]]).expect("valid utf-8"); |
| 192 | let text = text_delta_text(&events); |
| 193 | assert_eq!(text, "你好世界"); |
| 194 | assert!( |
| 195 | !text.contains('\u{FFFD}'), |
| 196 | "HTTP/2 mid-character split must not substitute U+FFFD; got {text:?}" |
| 197 | ); |
| 198 | } |
| 199 | |
| 200 | #[test] |
| 201 | fn decoder_reassembles_emoji_and_cjk_fed_one_byte_at_a_time() { |
| 202 | let frame = cjk_content_sse("你好🌊世界"); |
| 203 | let chunks: Vec<&[u8]> = frame.chunks(1).collect(); |
| 204 | let events = decode_sse_byte_chunks(&chunks).expect("valid utf-8"); |
| 205 | let text = text_delta_text(&events); |
| 206 | assert_eq!(text, "你好🌊世界"); |
| 207 | assert!( |
| 208 | !text.contains('\u{FFFD}'), |
| 209 | "byte-at-a-time feed garbled: {text:?}" |
| 210 | ); |
| 211 | } |
| 212 | |
| 213 | #[test] |
| 214 | fn decoder_rejects_invalid_sse_bytes_without_replacement() { |
| 215 | let mut frame = cjk_content_sse("ok"); |
| 216 | // Bare 0xFF is never valid UTF-8. Insert it inside the first SSE line. |
| 217 | let newline = frame.iter().position(|&b| b == b'\n').expect("SSE line"); |
| 218 | frame.insert(newline, 0xFF); |
| 219 | let result = decode_sse_byte_chunks(&[&frame]); |
| 220 | let err = result.expect_err("invalid SSE bytes must fail closed"); |
| 221 | assert!( |
| 222 | !err.to_string().contains('\u{FFFD}'), |
| 223 | "error path must not substitute U+FFFD: {err}" |
| 224 | ); |
| 225 | |
| 226 | // Unterminated tail of continuation bytes: fail on flush, no replacement. |
| 227 | let result = decode_sse_byte_chunks(&[&[0x80, 0xBF]]); |
| 228 | let err = result.expect_err("invalid unterminated flush must fail closed"); |
| 229 | assert!(!err.to_string().contains('\u{FFFD}')); |
| 230 | } |
| 231 | |
| 232 | #[test] |
| 233 | fn decoder_emits_text_delta_for_content_chunk() { |
| 234 | // The "happy" first chunk: a normal content delta. The engine treats |
| 235 | // this as `any_content_received = true` and would NOT transparently |
| 236 | // retry on a subsequent error. |
| 237 | let events = decode_chunk(r#"{"choices":[{"delta":{"content":"hello"}}]}"#); |
| 238 | assert!( |
| 239 | matches!( |
| 240 | events.first(), |
| 241 | Some(StreamEvent::ContentBlockStart { |
| 242 | content_block: ContentBlockStart::Text { .. }, |
| 243 | .. |
| 244 | }) |
| 245 | ), |
| 246 | "first event should open a text block; got {events:?}" |
| 247 | ); |
| 248 | assert!( |
| 249 | events |
| 250 | .iter() |
| 251 | .any(|e| matches!(e, StreamEvent::ContentBlockDelta { |
| 252 | delta: Delta::TextDelta { text }, |
| 253 | .. |
| 254 | } if text == "hello")), |
| 255 | "should yield a TextDelta carrying 'hello'; got {events:?}" |
| 256 | ); |
| 257 | } |
| 258 | |
| 259 | #[test] |
| 260 | fn decoder_emits_thinking_delta_for_reasoning_chunk() { |
| 261 | // V4 thinking models surface reasoning_content first — the engine |
| 262 | // also counts these as content received (so a subsequent stream error |
| 263 | // surfaces rather than retrying transparently). |
| 264 | let events = decode_chunk(r#"{"choices":[{"delta":{"reasoning_content":"plan..."}}]}"#); |
| 265 | assert!( |
| 266 | matches!( |
| 267 | events.first(), |
| 268 | Some(StreamEvent::ContentBlockStart { |
| 269 | content_block: ContentBlockStart::Thinking { .. }, |
| 270 | .. |
| 271 | }) |
| 272 | ), |
| 273 | "first event should open a thinking block; got {events:?}" |
| 274 | ); |
| 275 | assert!( |
| 276 | events |
| 277 | .iter() |
| 278 | .any(|e| matches!(e, StreamEvent::ContentBlockDelta { |
| 279 | delta: Delta::ThinkingDelta { thinking }, |
| 280 | .. |
| 281 | } if thinking == "plan...")), |
| 282 | "should yield a ThinkingDelta carrying 'plan...'; got {events:?}" |
| 283 | ); |
| 284 | } |
| 285 | |
| 286 | #[test] |
| 287 | fn decoder_streams_moonshot_multi_chunk_reasoning_as_thinking() { |
| 288 | // #3016: recorded shape from Moonshot's native endpoint — kimi-k2.6 |
| 289 | // streams `reasoning_content` deltas before the answer text. The |
| 290 | // thinking deltas must accumulate into ONE thinking block and the |
| 291 | // answer must arrive as text, not be glued into the trace. |
| 292 | let chunks = [ |
| 293 | r#"{"id":"cmpl-kimi","model":"kimi-k2.6","choices":[{"index":0,"delta":{"role":"assistant","reasoning_content":"Let me check"}}]}"#, |
| 294 | r#"{"id":"cmpl-kimi","model":"kimi-k2.6","choices":[{"index":0,"delta":{"reasoning_content":" the config."}}]}"#, |
| 295 | r#"{"id":"cmpl-kimi","model":"kimi-k2.6","choices":[{"index":0,"delta":{"content":"The answer is 42."}}]}"#, |
| 296 | ]; |
| 297 | |
| 298 | let is_reasoning = |
| 299 | is_reasoning_model_for_stream(crate::config::ApiProvider::Moonshot, "kimi-k2.6"); |
| 300 | let mut content_index = 0u32; |
| 301 | let mut text_started = false; |
| 302 | let mut thinking_started = false; |
| 303 | let mut tool_indices = std::collections::HashMap::new(); |
| 304 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 305 | let mut events = Vec::new(); |
| 306 | for chunk in chunks { |
| 307 | let value: Value = serde_json::from_str(chunk).expect("valid SSE JSON"); |
| 308 | events.extend(parse_sse_chunk( |
| 309 | &value, |
| 310 | &mut content_index, |
| 311 | &mut text_started, |
| 312 | &mut thinking_started, |
| 313 | &mut tool_indices, |
| 314 | &mut reasoning_detail_buffers, |
| 315 | is_reasoning, |
| 316 | )); |
| 317 | } |
| 318 | |
| 319 | let thinking: String = events |
| 320 | .iter() |
| 321 | .filter_map(|event| match event { |
| 322 | StreamEvent::ContentBlockDelta { |
| 323 | delta: Delta::ThinkingDelta { thinking }, |
| 324 | .. |
| 325 | } => Some(thinking.as_str()), |
| 326 | _ => None, |
| 327 | }) |
| 328 | .collect(); |
| 329 | assert_eq!(thinking, "Let me check the config."); |
| 330 | |
| 331 | let thinking_starts = events |
| 332 | .iter() |
| 333 | .filter(|event| { |
| 334 | matches!( |
| 335 | event, |
| 336 | StreamEvent::ContentBlockStart { |
| 337 | content_block: ContentBlockStart::Thinking { .. }, |
| 338 | .. |
| 339 | } |
| 340 | ) |
| 341 | }) |
| 342 | .count(); |
| 343 | assert_eq!(thinking_starts, 1, "one thinking block: {events:?}"); |
| 344 | |
| 345 | let text: String = events |
| 346 | .iter() |
| 347 | .filter_map(|event| match event { |
| 348 | StreamEvent::ContentBlockDelta { |
| 349 | delta: Delta::TextDelta { text }, |
| 350 | .. |
| 351 | } => Some(text.as_str()), |
| 352 | _ => None, |
| 353 | }) |
| 354 | .collect(); |
| 355 | assert_eq!(text, "The answer is 42."); |
| 356 | } |
| 357 | |
| 358 | #[test] |
| 359 | fn decoder_accepts_openrouter_reasoning_delta_with_extra_fields() { |
| 360 | let events = decode_chunk( |
| 361 | r#"{"id":"or-1","choices":[{"delta":{"reasoning":"openrouter thought","reasoning_details":[{"type":"summary","text":"extra"}],"native_finish_reason":null}}],"usage":{"completion_tokens_details":{"reasoning_tokens":3}}}"#, |
| 362 | ); |
| 363 | |
| 364 | assert!( |
| 365 | events.iter().any(|e| matches!( |
| 366 | e, |
| 367 | StreamEvent::ContentBlockDelta { |
| 368 | delta: Delta::ThinkingDelta { thinking }, |
| 369 | .. |
| 370 | } if thinking == "openrouter thought" |
| 371 | )), |
| 372 | "OpenRouter-style reasoning deltas with extra fields should not crash decoding; got {events:?}" |
| 373 | ); |
| 374 | } |
| 375 | |
| 376 | #[test] |
| 377 | fn decoder_streams_minimax_reasoning_details_as_incremental_thinking() { |
| 378 | // MiniMax's reasoning_split stream reports reasoning_details text as |
| 379 | // a cumulative buffer. Emit only the suffix so the Thinking cell does |
| 380 | // not duplicate earlier reasoning chunks. |
| 381 | let chunks = [ |
| 382 | r#"{"id":"minimax-1","choices":[{"index":0,"delta":{"reasoning_details":[{"type":"text","text":"Inspect"}]}}]}"#, |
| 383 | r#"{"id":"minimax-1","choices":[{"index":0,"delta":{"reasoning_details":[{"type":"text","text":"Inspect config"}]}}]}"#, |
| 384 | r#"{"id":"minimax-1","choices":[{"index":0,"delta":{"content":"Done."}}]}"#, |
| 385 | ]; |
| 386 | |
| 387 | let is_reasoning = is_reasoning_model_for_stream(ApiProvider::Minimax, "MiniMax-M3"); |
| 388 | let mut content_index = 0u32; |
| 389 | let mut text_started = false; |
| 390 | let mut thinking_started = false; |
| 391 | let mut tool_indices = std::collections::HashMap::new(); |
| 392 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 393 | let mut events = Vec::new(); |
| 394 | for chunk in chunks { |
| 395 | let value: Value = serde_json::from_str(chunk).expect("valid SSE JSON"); |
| 396 | events.extend(parse_sse_chunk( |
| 397 | &value, |
| 398 | &mut content_index, |
| 399 | &mut text_started, |
| 400 | &mut thinking_started, |
| 401 | &mut tool_indices, |
| 402 | &mut reasoning_detail_buffers, |
| 403 | is_reasoning, |
| 404 | )); |
| 405 | } |
| 406 | |
| 407 | let thinking: String = events |
| 408 | .iter() |
| 409 | .filter_map(|event| match event { |
| 410 | StreamEvent::ContentBlockDelta { |
| 411 | delta: Delta::ThinkingDelta { thinking }, |
| 412 | .. |
| 413 | } => Some(thinking.as_str()), |
| 414 | _ => None, |
| 415 | }) |
| 416 | .collect(); |
| 417 | assert_eq!(thinking, "Inspect config"); |
| 418 | |
| 419 | assert!(!events.iter().any(|event| matches!( |
| 420 | event, |
| 421 | StreamEvent::ContentBlockDelta { |
| 422 | delta: Delta::TextDelta { text }, |
| 423 | .. |
| 424 | } if text == "Inspect" || text == "Inspect config" |
| 425 | ))); |
| 426 | } |
| 427 | |
| 428 | #[test] |
| 429 | fn modelstudio_streams_reasoning_content_as_thinking() { |
| 430 | // Recorded-style DashScope OpenAI-compatible frames (shape lifted from |
| 431 | // Model Studio's deep-thinking docs): reasoning streams in |
| 432 | // `delta.reasoning_content`, the answer in `delta.content`, and a |
| 433 | // trailing usage-only chunk closes the stream. |
| 434 | let chunks = [ |
| 435 | r#"{"choices":[{"delta":{"content":null,"role":"assistant","reasoning_content":""},"index":0,"logprobs":null,"finish_reason":null}],"object":"chat.completion.chunk","usage":null,"model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 436 | r#"{"choices":[{"delta":{"reasoning_content":"Let me think"},"index":0}],"object":"chat.completion.chunk","model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 437 | r#"{"choices":[{"delta":{"reasoning_content":" about this."},"index":0}],"object":"chat.completion.chunk","model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 438 | r#"{"choices":[{"delta":{"content":"The answer."},"index":0}],"object":"chat.completion.chunk","model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 439 | r#"{"choices":[{"finish_reason":"stop","delta":{"content":"","reasoning_content":null},"index":0}],"object":"chat.completion.chunk","model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 440 | r#"{"choices":[],"object":"chat.completion.chunk","usage":{"prompt_tokens":10,"completion_tokens":30,"total_tokens":40,"completion_tokens_details":{"reasoning_tokens":20}},"model":"qwen3.8-max","id":"chatcmpl-ms-1"}"#, |
| 441 | ]; |
| 442 | |
| 443 | // Both OpenAI-dialect plans classify their reasoning catalog. |
| 444 | for (provider, base_url, model) in [ |
| 445 | ( |
| 446 | ApiProvider::ModelstudioTokenPlan, |
| 447 | crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 448 | "qwen3.8-max", |
| 449 | ), |
| 450 | ( |
| 451 | ApiProvider::ModelstudioCodingPlan, |
| 452 | crate::config::DEFAULT_MODELSTUDIO_CODING_PLAN_BASE_URL, |
| 453 | "qwen3.7-plus", |
| 454 | ), |
| 455 | ] { |
| 456 | let style = reasoning_stream_style_for_route(provider, base_url, model, None); |
| 457 | assert_eq!(style, ReasoningStreamStyle::SeparateField, "{provider:?}"); |
| 458 | |
| 459 | let mut content_index = 0u32; |
| 460 | let mut text_started = false; |
| 461 | let mut thinking_started = false; |
| 462 | let mut tool_indices = std::collections::HashMap::new(); |
| 463 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 464 | let mut inline_reasoning_tags = InlineReasoningTagState::default(); |
| 465 | let mut events = Vec::new(); |
| 466 | for chunk in chunks { |
| 467 | let value: Value = serde_json::from_str(chunk).expect("valid SSE JSON"); |
| 468 | events.extend(parse_sse_chunk_with_reasoning_style( |
| 469 | &value, |
| 470 | &mut content_index, |
| 471 | &mut text_started, |
| 472 | &mut thinking_started, |
| 473 | &mut tool_indices, |
| 474 | &mut reasoning_detail_buffers, |
| 475 | &mut inline_reasoning_tags, |
| 476 | style, |
| 477 | )); |
| 478 | } |
| 479 | |
| 480 | let thinking: String = events |
| 481 | .iter() |
| 482 | .filter_map(|event| match event { |
| 483 | StreamEvent::ContentBlockDelta { |
| 484 | delta: Delta::ThinkingDelta { thinking }, |
| 485 | .. |
| 486 | } => Some(thinking.as_str()), |
| 487 | _ => None, |
| 488 | }) |
| 489 | .collect(); |
| 490 | assert_eq!(thinking, "Let me think about this.", "{provider:?}"); |
| 491 | |
| 492 | let text: String = events |
| 493 | .iter() |
| 494 | .filter_map(|event| match event { |
| 495 | StreamEvent::ContentBlockDelta { |
| 496 | delta: Delta::TextDelta { text }, |
| 497 | .. |
| 498 | } => Some(text.as_str()), |
| 499 | _ => None, |
| 500 | }) |
| 501 | .collect(); |
| 502 | assert_eq!(text, "The answer.", "{provider:?}"); |
| 503 | |
| 504 | // The trailing usage chunk still surfaces token accounting. |
| 505 | assert!( |
| 506 | events.iter().any(|event| matches!( |
| 507 | event, |
| 508 | StreamEvent::MessageDelta { usage: Some(usage), .. } |
| 509 | if usage.output_tokens == 30 |
| 510 | )), |
| 511 | "{provider:?}: {events:?}" |
| 512 | ); |
| 513 | } |
| 514 | |
| 515 | // A non-reasoning model id on the same route keeps the old |
| 516 | // pass-through semantics (no fabricated Thinking surface). |
| 517 | let style = reasoning_stream_style_for_route( |
| 518 | ApiProvider::ModelstudioTokenPlan, |
| 519 | crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 520 | "qwen3.8-max-lite-unknown", |
| 521 | None, |
| 522 | ); |
| 523 | assert_eq!(style, ReasoningStreamStyle::None); |
| 524 | } |
| 525 | |
| 526 | #[test] |
| 527 | fn decoder_does_not_render_reasoning_as_text_for_known_provider_models() { |
| 528 | let mut content_index = 0u32; |
| 529 | let mut text_started = false; |
| 530 | let mut thinking_started = false; |
| 531 | let mut tool_indices = std::collections::HashMap::new(); |
| 532 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 533 | let is_reasoning_model = |
| 534 | is_reasoning_model_for_stream(ApiProvider::XiaomiMimo, "mimo-v2.5-pro"); |
| 535 | let events = parse_sse_chunk( |
| 536 | &serde_json::json!({ |
| 537 | "choices": [{ |
| 538 | "delta": { |
| 539 | "reasoning_content": "private plan" |
| 540 | } |
| 541 | }] |
| 542 | }), |
| 543 | &mut content_index, |
| 544 | &mut text_started, |
| 545 | &mut thinking_started, |
| 546 | &mut tool_indices, |
| 547 | &mut reasoning_detail_buffers, |
| 548 | is_reasoning_model, |
| 549 | ); |
| 550 | |
| 551 | assert!(events.iter().any(|event| matches!( |
| 552 | event, |
| 553 | StreamEvent::ContentBlockDelta { |
| 554 | delta: Delta::ThinkingDelta { thinking }, |
| 555 | .. |
| 556 | } if thinking == "private plan" |
| 557 | ))); |
| 558 | assert!(!events.iter().any(|event| matches!( |
| 559 | event, |
| 560 | StreamEvent::ContentBlockDelta { |
| 561 | delta: Delta::TextDelta { text }, |
| 562 | .. |
| 563 | } if text == "private plan" |
| 564 | ))); |
| 565 | } |
| 566 | |
| 567 | #[test] |
| 568 | fn decoder_treats_reasoning_content_as_text_when_provider_does_not_support_reasoning() { |
| 569 | let events = decode_chunk_with_reasoning( |
| 570 | r#"{"choices":[{"delta":{"reasoning_content":"hello"}}]}"#, |
| 571 | false, |
| 572 | ); |
| 573 | |
| 574 | assert!( |
| 575 | matches!( |
| 576 | events.first(), |
| 577 | Some(StreamEvent::ContentBlockStart { |
| 578 | content_block: ContentBlockStart::Text { .. }, |
| 579 | .. |
| 580 | }) |
| 581 | ), |
| 582 | "first event should open a text block; got {events:?}" |
| 583 | ); |
| 584 | assert!( |
| 585 | events.iter().any(|e| matches!( |
| 586 | e, |
| 587 | StreamEvent::ContentBlockDelta { |
| 588 | delta: Delta::TextDelta { text }, |
| 589 | .. |
| 590 | } if text == "hello" |
| 591 | )), |
| 592 | "should yield a TextDelta carrying 'hello'; got {events:?}" |
| 593 | ); |
| 594 | assert!( |
| 595 | !events.iter().any(|e| matches!( |
| 596 | e, |
| 597 | StreamEvent::ContentBlockDelta { |
| 598 | delta: Delta::ThinkingDelta { .. }, |
| 599 | .. |
| 600 | } |
| 601 | )), |
| 602 | "should not emit thinking deltas for generic providers; got {events:?}" |
| 603 | ); |
| 604 | } |
| 605 | |
| 606 | #[test] |
| 607 | fn reasoning_style_separate_field_routes_reasoning_to_thinking() { |
| 608 | let events = decode_chunks_with_style( |
| 609 | &[ |
| 610 | r#"{"choices":[{"delta":{"reasoning_content":"private plan"}}]}"#, |
| 611 | r#"{"choices":[{"delta":{"content":"Public answer."}}]}"#, |
| 612 | ], |
| 613 | ReasoningStreamStyle::SeparateField, |
| 614 | ); |
| 615 | |
| 616 | assert_eq!(thinking_delta_text(&events), "private plan"); |
| 617 | assert_eq!(text_delta_text(&events), "Public answer."); |
| 618 | } |
| 619 | |
| 620 | #[test] |
| 621 | fn exact_kimi_code_k3_streams_reasoning_content_as_thinking() { |
| 622 | let style = reasoning_stream_style_for_route( |
| 623 | ApiProvider::Moonshot, |
| 624 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 625 | crate::config::KIMI_CODE_K3_MODEL, |
| 626 | None, |
| 627 | ); |
| 628 | assert_eq!(style, ReasoningStreamStyle::SeparateField); |
| 629 | |
| 630 | let events = decode_chunks_with_style( |
| 631 | &[r#"{"choices":[{"delta":{"reasoning_content":"private K3 plan"}}]}"#], |
| 632 | style, |
| 633 | ); |
| 634 | assert_eq!(thinking_delta_text(&events), "private K3 plan"); |
| 635 | assert_eq!(text_delta_text(&events), ""); |
| 636 | |
| 637 | let generic_style = reasoning_stream_style_for_route( |
| 638 | ApiProvider::Moonshot, |
| 639 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 640 | crate::config::KIMI_CODE_K3_MODEL, |
| 641 | None, |
| 642 | ); |
| 643 | assert_eq!(generic_style, ReasoningStreamStyle::None); |
| 644 | } |
| 645 | |
| 646 | #[test] |
| 647 | fn reasoning_style_inline_tags_routes_think_blocks_to_thinking() { |
| 648 | let events = decode_chunks_with_style( |
| 649 | &[ |
| 650 | r#"{"choices":[{"delta":{"content":"Before <thi"}}]}"#, |
| 651 | r#"{"choices":[{"delta":{"content":"nk>private plan</thi"}}]}"#, |
| 652 | r#"{"choices":[{"delta":{"content":"nk> after."}}]}"#, |
| 653 | ], |
| 654 | ReasoningStreamStyle::InlineTags, |
| 655 | ); |
| 656 | |
| 657 | assert_eq!(thinking_delta_text(&events), "private plan"); |
| 658 | assert_eq!(text_delta_text(&events), "Before after."); |
| 659 | assert!( |
| 660 | !text_delta_text(&events).contains("<think>"), |
| 661 | "inline reasoning tags must not leak into visible text: {events:?}" |
| 662 | ); |
| 663 | } |
| 664 | |
| 665 | #[test] |
| 666 | fn reasoning_style_inline_tags_flushes_unclosed_think_at_stream_end() { |
| 667 | let events = decode_chunks_with_style( |
| 668 | &[ |
| 669 | r#"{"choices":[{"delta":{"content":"Before <think>partial reasoning"}}]}"#, |
| 670 | r#"{"choices":[{"finish_reason":"stop"}]}"#, |
| 671 | ], |
| 672 | ReasoningStreamStyle::InlineTags, |
| 673 | ); |
| 674 | |
| 675 | assert_eq!(thinking_delta_text(&events), "partial reasoning"); |
| 676 | assert_eq!(text_delta_text(&events), "Before "); |
| 677 | } |
| 678 | |
| 679 | #[test] |
| 680 | fn reasoning_style_inline_tags_ignores_separate_reasoning_field() { |
| 681 | let events = decode_chunks_with_style( |
| 682 | &[ |
| 683 | r#"{"choices":[{"delta":{"reasoning_content":"metadata","content":"<think>tagged</think> answer"}}]}"#, |
| 684 | ], |
| 685 | ReasoningStreamStyle::InlineTags, |
| 686 | ); |
| 687 | |
| 688 | assert_eq!(thinking_delta_text(&events), "tagged"); |
| 689 | assert_eq!(text_delta_text(&events), " answer"); |
| 690 | } |
| 691 | |
| 692 | #[test] |
| 693 | fn reasoning_style_none_keeps_inline_tags_visible_text() { |
| 694 | let events = decode_chunks_with_style( |
| 695 | &[r#"{"choices":[{"delta":{"content":"<think>visible</think> answer"}}]}"#], |
| 696 | ReasoningStreamStyle::None, |
| 697 | ); |
| 698 | |
| 699 | assert_eq!(thinking_delta_text(&events), ""); |
| 700 | assert_eq!(text_delta_text(&events), "<think>visible</think> answer"); |
| 701 | } |
| 702 | |
| 703 | #[test] |
| 704 | fn configured_reasoning_style_overrides_route_default() { |
| 705 | assert_eq!( |
| 706 | reasoning_stream_style_for_stream(ApiProvider::Openai, "custom-minimax", None), |
| 707 | ReasoningStreamStyle::None |
| 708 | ); |
| 709 | assert_eq!( |
| 710 | reasoning_stream_style_for_stream( |
| 711 | ApiProvider::Openai, |
| 712 | "custom-minimax", |
| 713 | Some("inline-tags") |
| 714 | ), |
| 715 | ReasoningStreamStyle::InlineTags |
| 716 | ); |
| 717 | assert_eq!( |
| 718 | reasoning_stream_style_for_stream(ApiProvider::XiaomiMimo, "mimo-v2.5-pro", None), |
| 719 | ReasoningStreamStyle::SeparateField |
| 720 | ); |
| 721 | assert_eq!( |
| 722 | reasoning_stream_style_for_stream(ApiProvider::XiaomiMimo, "mimo-v2.5-pro", Some("none")), |
| 723 | ReasoningStreamStyle::None |
| 724 | ); |
| 725 | } |
| 726 | |
| 727 | #[test] |
| 728 | fn decoder_yields_no_events_for_keepalive_chunk() { |
| 729 | // DeepSeek often sends `{"choices":[]}` keepalive chunks before |
| 730 | // emitting real content. The engine MUST treat a stream error after |
| 731 | // these as "no content received" and be eligible for transparent |
| 732 | // retry — assert here that the decoder yields no payload events. |
| 733 | let events = decode_chunk(r#"{"choices":[]}"#); |
| 734 | assert!( |
| 735 | events.is_empty(), |
| 736 | "empty-choices chunk must produce no events; got {events:?}" |
| 737 | ); |
| 738 | } |
| 739 | |
| 740 | #[test] |
| 741 | fn decoder_treats_done_frame_as_terminal() { |
| 742 | let mut content_index = 0u32; |
| 743 | let mut text_started = false; |
| 744 | let mut thinking_started = false; |
| 745 | let mut tool_indices = std::collections::HashMap::new(); |
| 746 | let mut reasoning_detail_buffers = std::collections::HashMap::new(); |
| 747 | let mut inline_reasoning_tags = InlineReasoningTagState::default(); |
| 748 | |
| 749 | let outcome = parse_sse_data_frame( |
| 750 | " [DONE] ", |
| 751 | &mut content_index, |
| 752 | &mut text_started, |
| 753 | &mut thinking_started, |
| 754 | &mut tool_indices, |
| 755 | &mut reasoning_detail_buffers, |
| 756 | &mut inline_reasoning_tags, |
| 757 | ReasoningStreamStyle::SeparateField, |
| 758 | ); |
| 759 | |
| 760 | assert!( |
| 761 | matches!(outcome, SseDataFrame::Done), |
| 762 | "`data: [DONE]` must terminate the stream instead of waiting for the HTTP connection to close" |
| 763 | ); |
| 764 | assert_eq!(content_index, 0); |
| 765 | assert!(!text_started); |
| 766 | assert!(!thinking_started); |
| 767 | assert!(tool_indices.is_empty()); |
| 768 | } |
| 769 | |
| 770 | #[test] |
| 771 | fn decoder_emits_tool_use_block_for_tool_call_delta() { |
| 772 | // Tool-call deltas are content too — once one arrives, transparent |
| 773 | // retry must be off (the model has committed to a tool invocation |
| 774 | // path that DeepSeek has billed for). |
| 775 | let events = decode_chunk( |
| 776 | r#"{"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_1","function":{"name":"grep_files","arguments":"{\"pattern\":\"foo\"}"}}]}}]}"#, |
| 777 | ); |
| 778 | assert!( |
| 779 | events.iter().any(|e| matches!( |
| 780 | e, |
| 781 | StreamEvent::ContentBlockStart { |
| 782 | content_block: ContentBlockStart::ToolUse { name, ..}, |
| 783 | .. |
| 784 | } if name == "grep_files" |
| 785 | )), |
| 786 | "should open a ToolUse block for grep_files; got {events:?}" |
| 787 | ); |
| 788 | assert!( |
| 789 | events.iter().any(|e| matches!( |
| 790 | e, |
| 791 | StreamEvent::ContentBlockDelta { |
| 792 | delta: Delta::InputJsonDelta { partial_json }, |
| 793 | .. |
| 794 | } if partial_json.contains("\"pattern\"") |
| 795 | )), |
| 796 | "should yield InputJsonDelta carrying the tool args; got {events:?}" |
| 797 | ); |
| 798 | } |
| 799 | |
| 800 | #[test] |
| 801 | fn decoder_uses_fallback_name_for_empty_streaming_tool_name() { |
| 802 | let events = decode_chunk( |
| 803 | r#"{"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_empty","function":{"name":"","arguments":"{}"}}]}}]}"#, |
| 804 | ); |
| 805 | |
| 806 | assert!( |
| 807 | events.iter().any(|event| matches!( |
| 808 | event, |
| 809 | StreamEvent::ContentBlockStart { |
| 810 | content_block: ContentBlockStart::ToolUse { name, ..}, |
| 811 | .. |
| 812 | } if name == "unknown_tool" |
| 813 | )), |
| 814 | "empty upstream tool names should render as unknown_tool; got {events:?}" |
| 815 | ); |
| 816 | } |
| 817 | |
| 818 | #[test] |
| 819 | fn non_streaming_response_uses_fallback_name_for_missing_tool_name() { |
| 820 | let payload: Value = serde_json::from_str( |
| 821 | r#"{ |
| 822 | "id": "chatcmpl_1", |
| 823 | "model": "deepseek-v4-pro", |
| 824 | "choices": [{ |
| 825 | "message": { |
| 826 | "role": "assistant", |
| 827 | "tool_calls": [{ |
| 828 | "id": "call_missing", |
| 829 | "function": { "arguments": "{}" } |
| 830 | }] |
| 831 | }, |
| 832 | "finish_reason": "tool_calls" |
| 833 | }] |
| 834 | }"#, |
| 835 | ) |
| 836 | .expect("valid response"); |
| 837 | |
| 838 | let parsed = parse_chat_message(&payload).expect("message parses"); |
| 839 | let tool_name = parsed.content.iter().find_map(|block| match block { |
| 840 | ContentBlock::ToolUse { name, .. } => Some(name.as_str()), |
| 841 | _ => None, |
| 842 | }); |
| 843 | |
| 844 | assert_eq!(tool_name, Some("unknown_tool")); |
| 845 | } |
| 846 | |
| 847 | /// Regression for the parallel-tool-calls-without-id collision (audit |
| 848 | /// Finding 8): when the upstream chunk omits the `id` field, the |
| 849 | /// fallback used to be the literal string `"tool_call"` for every |
| 850 | /// parallel call, so two tool calls in one delta ended up sharing an |
| 851 | /// id. Downstream routing then matched the first call's tool_result |
| 852 | /// twice and the second call hung. The fallback is now indexed by the |
| 853 | /// content-block position, keeping each call unique within the |
| 854 | /// response. |
| 855 | #[test] |
| 856 | fn decoder_assigns_unique_fallback_ids_to_parallel_tool_calls_missing_id() { |
| 857 | let events = decode_chunk( |
| 858 | r#"{"choices":[{"delta":{"tool_calls":[ |
| 859 | {"index":0,"function":{"name":"grep_files","arguments":"{\"pattern\":\"a\"}"}}, |
| 860 | {"index":1,"function":{"name":"read_file","arguments":"{\"path\":\"x\"}"}} |
| 861 | ]}}]}"#, |
| 862 | ); |
| 863 | |
| 864 | let ids: Vec<&str> = events |
| 865 | .iter() |
| 866 | .filter_map(|e| match e { |
| 867 | StreamEvent::ContentBlockStart { |
| 868 | content_block: ContentBlockStart::ToolUse { id, .. }, |
| 869 | .. |
| 870 | } => Some(id.as_str()), |
| 871 | _ => None, |
| 872 | }) |
| 873 | .collect(); |
| 874 | |
| 875 | assert_eq!( |
| 876 | ids.len(), |
| 877 | 2, |
| 878 | "expected two tool-use blocks for parallel tool calls; got {events:?}" |
| 879 | ); |
| 880 | assert_ne!( |
| 881 | ids[0], ids[1], |
| 882 | "parallel tool calls without upstream `id` must get distinct fallback ids; got {ids:?}" |
| 883 | ); |
| 884 | } |
| 885 | |
| 886 | #[test] |
| 887 | fn decoder_preserves_upstream_tool_call_id_when_present() { |
| 888 | // Counter-test to the fallback regression: when the upstream chunk |
| 889 | // does include `id`, we forward it verbatim — we shouldn't quietly |
| 890 | // rewrite ids the API gave us just because we have a fallback path. |
| 891 | let events = decode_chunk( |
| 892 | r#"{"choices":[{"delta":{"tool_calls":[{"index":0,"id":"call_xyz","function":{"name":"grep_files","arguments":"{}"}}]}}]}"#, |
| 893 | ); |
| 894 | let id = events |
| 895 | .iter() |
| 896 | .find_map(|e| match e { |
| 897 | StreamEvent::ContentBlockStart { |
| 898 | content_block: ContentBlockStart::ToolUse { id, .. }, |
| 899 | .. |
| 900 | } => Some(id.as_str()), |
| 901 | _ => None, |
| 902 | }) |
| 903 | .expect("tool-use block present"); |
| 904 | assert_eq!(id, "call_xyz"); |
| 905 | } |
| 906 | |
| 907 | #[test] |
| 908 | fn request_builder_preserves_internal_system_messages() { |
| 909 | let messages = vec![Message { |
| 910 | role: Role::System, |
| 911 | content: vec![ContentBlock::Text { |
| 912 | text: "internal runtime event".to_string(), |
| 913 | cache_control: None, |
| 914 | }], |
| 915 | }]; |
| 916 | |
| 917 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 918 | |
| 919 | assert_eq!(built.len(), 1); |
| 920 | assert_eq!(built[0]["role"], "system"); |
| 921 | assert_eq!(built[0]["content"], "internal runtime event"); |
| 922 | } |
| 923 | |
| 924 | fn tool_use_message(id: &str, name: &str, input: Value) -> Message { |
| 925 | Message { |
| 926 | role: Role::Assistant, |
| 927 | content: vec![ContentBlock::ToolUse { |
| 928 | id: id.to_string(), |
| 929 | name: name.to_string(), |
| 930 | input, |
| 931 | caller: None, |
| 932 | thought_signature: None, |
| 933 | }], |
| 934 | } |
| 935 | } |
| 936 | |
| 937 | fn tool_result_message(id: &str, content: &str) -> Message { |
| 938 | Message { |
| 939 | role: Role::User, |
| 940 | content: vec![ContentBlock::ToolResult { |
| 941 | tool_use_id: id.to_string(), |
| 942 | content: content.to_string(), |
| 943 | is_error: None, |
| 944 | content_blocks: None, |
| 945 | }], |
| 946 | } |
| 947 | } |
| 948 | |
| 949 | fn user_message_with_turn_meta(turn_meta: &str, task: &str) -> Message { |
| 950 | Message { |
| 951 | role: Role::User, |
| 952 | content: vec![ |
| 953 | ContentBlock::Text { |
| 954 | text: turn_meta.to_string(), |
| 955 | cache_control: None, |
| 956 | }, |
| 957 | ContentBlock::Text { |
| 958 | text: task.to_string(), |
| 959 | cache_control: None, |
| 960 | }, |
| 961 | ], |
| 962 | } |
| 963 | } |
| 964 | |
| 965 | fn user_message_with_tail_turn_meta(task: &str, turn_meta: &str) -> Message { |
| 966 | Message { |
| 967 | role: Role::User, |
| 968 | content: vec![ |
| 969 | ContentBlock::Text { |
| 970 | text: task.to_string(), |
| 971 | cache_control: None, |
| 972 | }, |
| 973 | ContentBlock::Text { |
| 974 | text: turn_meta.to_string(), |
| 975 | cache_control: None, |
| 976 | }, |
| 977 | ], |
| 978 | } |
| 979 | } |
| 980 | |
| 981 | fn tool_message_content(messages: &[Value], index: usize) -> &str { |
| 982 | messages |
| 983 | .iter() |
| 984 | .filter(|message| message.get("role").and_then(Value::as_str) == Some("tool")) |
| 985 | .nth(index) |
| 986 | .and_then(|message| message.get("content").and_then(Value::as_str)) |
| 987 | .expect("tool message content") |
| 988 | } |
| 989 | |
| 990 | fn user_message_content(messages: &[Value], index: usize) -> &str { |
| 991 | messages |
| 992 | .iter() |
| 993 | .filter(|message| message.get("role").and_then(Value::as_str) == Some("user")) |
| 994 | .nth(index) |
| 995 | .and_then(|message| message.get("content").and_then(Value::as_str)) |
| 996 | .expect("user message content") |
| 997 | } |
| 998 | |
| 999 | fn with_tool_result_sha_spillover_root<T>(f: impl FnOnce() -> T) -> T { |
| 1000 | let _guard = crate::tools::truncate::TEST_SPILLOVER_GUARD |
| 1001 | .lock() |
| 1002 | .unwrap_or_else(|err| err.into_inner()); |
| 1003 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 1004 | let prior = crate::tools::truncate::set_test_spillover_root(Some( |
| 1005 | tmp.path().join(".deepseek").join("tool_outputs"), |
| 1006 | )); |
| 1007 | struct Restore(Option<std::path::PathBuf>); |
| 1008 | impl Drop for Restore { |
| 1009 | fn drop(&mut self) { |
| 1010 | crate::tools::truncate::set_test_spillover_root(self.0.take()); |
| 1011 | } |
| 1012 | } |
| 1013 | let _restore = Restore(prior); |
| 1014 | f() |
| 1015 | } |
| 1016 | |
| 1017 | #[test] |
| 1018 | fn request_builder_deduplicates_consecutive_identical_turn_meta_for_wire() { |
| 1019 | let turn_meta = "<turn_meta>\nCurrent local date: 2026-05-09\n</turn_meta>"; |
| 1020 | let messages = vec![ |
| 1021 | user_message_with_turn_meta(turn_meta, "first task"), |
| 1022 | Message { |
| 1023 | role: Role::Assistant, |
| 1024 | content: vec![ContentBlock::Text { |
| 1025 | text: "first answer".to_string(), |
| 1026 | cache_control: None, |
| 1027 | }], |
| 1028 | }, |
| 1029 | user_message_with_turn_meta(turn_meta, "second task"), |
| 1030 | ]; |
| 1031 | |
| 1032 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1033 | let first = user_message_content(&built, 0); |
| 1034 | let second = user_message_content(&built, 1); |
| 1035 | let expected_ref = "<turn_meta_unchanged />"; |
| 1036 | |
| 1037 | assert!(first.starts_with(turn_meta), "got: {first}"); |
| 1038 | assert!(second.starts_with(expected_ref), "got: {second}"); |
| 1039 | assert!(second.ends_with("second task"), "got: {second}"); |
| 1040 | assert_eq!( |
| 1041 | second, |
| 1042 | format!("{expected_ref}\nsecond task"), |
| 1043 | "ref text must stay stable" |
| 1044 | ); |
| 1045 | } |
| 1046 | |
| 1047 | #[test] |
| 1048 | fn request_builder_keeps_tail_turn_meta_after_user_text_for_wire() { |
| 1049 | let turn_meta = "<turn_meta>\nCurrent local date: 2026-05-09\n</turn_meta>"; |
| 1050 | let messages = vec![ |
| 1051 | user_message_with_tail_turn_meta("first task", turn_meta), |
| 1052 | Message { |
| 1053 | role: Role::Assistant, |
| 1054 | content: vec![ContentBlock::Text { |
| 1055 | text: "first answer".to_string(), |
| 1056 | cache_control: None, |
| 1057 | }], |
| 1058 | }, |
| 1059 | user_message_with_tail_turn_meta("second task", turn_meta), |
| 1060 | ]; |
| 1061 | |
| 1062 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1063 | let first = user_message_content(&built, 0); |
| 1064 | let second = user_message_content(&built, 1); |
| 1065 | let expected_ref = "<turn_meta_unchanged />"; |
| 1066 | |
| 1067 | assert_eq!(first, format!("first task\n{turn_meta}")); |
| 1068 | assert_eq!(second, format!("second task\n{expected_ref}")); |
| 1069 | } |
| 1070 | |
| 1071 | #[test] |
| 1072 | fn request_builder_keeps_changed_turn_meta_full_and_updates_recent_hash() { |
| 1073 | let first_meta = "<turn_meta>\nCurrent local date: 2026-05-09\n</turn_meta>"; |
| 1074 | let second_meta = |
| 1075 | "<turn_meta>\nCurrent local date: 2026-05-09\nWorking set: src/lib.rs\n</turn_meta>"; |
| 1076 | let messages = vec![ |
| 1077 | user_message_with_turn_meta(first_meta, "first task"), |
| 1078 | user_message_with_turn_meta(second_meta, "second task"), |
| 1079 | ]; |
| 1080 | |
| 1081 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1082 | let first = user_message_content(&built, 0); |
| 1083 | let second = user_message_content(&built, 1); |
| 1084 | |
| 1085 | assert!(first.starts_with(first_meta), "got: {first}"); |
| 1086 | assert!(second.starts_with(second_meta), "got: {second}"); |
| 1087 | assert!(!second.contains("<TURN_META_REF"), "got: {second}"); |
| 1088 | } |
| 1089 | |
| 1090 | #[test] |
| 1091 | fn turn_meta_dedup_is_wire_only_and_does_not_mutate_session_message() { |
| 1092 | let turn_meta = "<turn_meta>\nCurrent local date: 2026-05-09\n</turn_meta>"; |
| 1093 | let messages = vec![ |
| 1094 | user_message_with_turn_meta(turn_meta, "first task"), |
| 1095 | user_message_with_turn_meta(turn_meta, "second task"), |
| 1096 | ]; |
| 1097 | |
| 1098 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1099 | assert!( |
| 1100 | user_message_content(&built, 1).starts_with("<turn_meta_unchanged />"), |
| 1101 | "got: {}", |
| 1102 | user_message_content(&built, 1) |
| 1103 | ); |
| 1104 | |
| 1105 | match &messages[1].content[0] { |
| 1106 | ContentBlock::Text { text, .. } => assert_eq!(text, turn_meta), |
| 1107 | other => panic!("expected text block, got {other:?}"), |
| 1108 | } |
| 1109 | } |
| 1110 | |
| 1111 | #[test] |
| 1112 | fn cache_inspect_reports_turn_meta_dedup_metadata() { |
| 1113 | let turn_meta = format!( |
| 1114 | "<turn_meta>\nCurrent local date: 2026-05-09\n{}\n</turn_meta>", |
| 1115 | "Working set: src/lib.rs\n".repeat(20) |
| 1116 | ); |
| 1117 | let request = MessageRequest { |
| 1118 | model: "deepseek-v4-flash".to_string(), |
| 1119 | messages: vec![ |
| 1120 | user_message_with_turn_meta(&turn_meta, "first task"), |
| 1121 | user_message_with_turn_meta(&turn_meta, "second task"), |
| 1122 | ], |
| 1123 | max_tokens: 0, |
| 1124 | system: None, |
| 1125 | tools: None, |
| 1126 | tool_choice: None, |
| 1127 | metadata: None, |
| 1128 | thinking: None, |
| 1129 | reasoning_effort: None, |
| 1130 | stream: None, |
| 1131 | temperature: None, |
| 1132 | top_p: None, |
| 1133 | }; |
| 1134 | |
| 1135 | let inspection = inspect_prompt_for_request(&request); |
| 1136 | let turn_meta_layers: Vec<_> = inspection |
| 1137 | .layers |
| 1138 | .iter() |
| 1139 | .filter_map(|layer| layer.turn_meta.as_ref()) |
| 1140 | .collect(); |
| 1141 | |
| 1142 | assert_eq!(turn_meta_layers.len(), 2); |
| 1143 | assert_eq!( |
| 1144 | turn_meta_layers[0].original_chars, |
| 1145 | turn_meta.chars().count() |
| 1146 | ); |
| 1147 | assert_eq!(turn_meta_layers[0].sent_chars, turn_meta.chars().count()); |
| 1148 | assert!(!turn_meta_layers[0].deduplicated); |
| 1149 | assert_eq!(turn_meta_layers[0].sha256, sha256_hex(turn_meta.as_bytes())); |
| 1150 | assert_eq!( |
| 1151 | turn_meta_layers[1].original_chars, |
| 1152 | turn_meta.chars().count() |
| 1153 | ); |
| 1154 | assert!(turn_meta_layers[1].sent_chars < turn_meta_layers[1].original_chars); |
| 1155 | assert!(turn_meta_layers[1].deduplicated); |
| 1156 | assert_eq!(turn_meta_layers[1].sha256, turn_meta_layers[0].sha256); |
| 1157 | } |
| 1158 | |
| 1159 | #[test] |
| 1160 | fn request_builder_truncates_large_tool_result_for_wire() { |
| 1161 | let long_output = format!("{}{}", "A".repeat(7_000), "Z".repeat(7_000)); |
| 1162 | let messages = vec![ |
| 1163 | tool_use_message( |
| 1164 | "tool-long", |
| 1165 | "shell_command", |
| 1166 | json!({"command": "cargo test"}), |
| 1167 | ), |
| 1168 | tool_result_message("tool-long", &long_output), |
| 1169 | ]; |
| 1170 | |
| 1171 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1172 | let sent = tool_message_content(&built, 0); |
| 1173 | |
| 1174 | assert!(sent.contains("[TOOL_RESULT_TRUNCATED]"), "got: {sent}"); |
| 1175 | assert!(sent.contains("tool_name: shell_command"), "got: {sent}"); |
| 1176 | assert!(sent.contains("command_or_query: cargo test"), "got: {sent}"); |
| 1177 | assert!(sent.contains("original_chars: 14000"), "got: {sent}"); |
| 1178 | assert!(sent.contains("sha256:"), "got: {sent}"); |
| 1179 | assert!( |
| 1180 | sent.contains("exact_detail: unavailable; no session-owned artifact was recorded"), |
| 1181 | "got: {sent}" |
| 1182 | ); |
| 1183 | assert!(!sent.contains("retrieve_tool_result"), "got: {sent}"); |
| 1184 | assert!(sent.contains(&"A".repeat(4_000)), "got: {sent}"); |
| 1185 | assert!(sent.contains(&"Z".repeat(4_000)), "got: {sent}"); |
| 1186 | assert!( |
| 1187 | sent.contains("truncated 6000 chars from middle"), |
| 1188 | "got: {sent}" |
| 1189 | ); |
| 1190 | assert_ne!(sent, long_output); |
| 1191 | } |
| 1192 | |
| 1193 | #[test] |
| 1194 | fn request_builder_keeps_unowned_extreme_tool_output_bounded_without_false_hint() { |
| 1195 | with_tool_result_sha_spillover_root(|| { |
| 1196 | let huge_output = format!( |
| 1197 | "{}{}{}", |
| 1198 | "DIFF_HEAD\n".repeat(10_000), |
| 1199 | "MIDDLE_POISON\n".repeat(10_000), |
| 1200 | "DIFF_TAIL\n".repeat(10_000) |
| 1201 | ); |
| 1202 | let sha = sha256_hex(huge_output.as_bytes()); |
| 1203 | let messages = vec![ |
| 1204 | tool_use_message("tool-huge", "exec_shell", json!({"command": "git diff"})), |
| 1205 | tool_result_message("tool-huge", &huge_output), |
| 1206 | ]; |
| 1207 | |
| 1208 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1209 | let sent = tool_message_content(&built, 0); |
| 1210 | |
| 1211 | assert!(sent.contains("[TOOL_RESULT_TRUNCATED]"), "got: {sent}"); |
| 1212 | assert!(sent.contains("tool_name: exec_shell"), "got: {sent}"); |
| 1213 | assert!(sent.contains("command_or_query: git diff"), "got: {sent}"); |
| 1214 | assert!(sent.contains(&format!("sha256: {sha}")), "got: {sent}"); |
| 1215 | assert!(sent.contains("exact_detail: unavailable"), "got: {sent}"); |
| 1216 | assert!(!sent.contains("retrieve_tool_result"), "got: {sent}"); |
| 1217 | assert!( |
| 1218 | sent.chars().count() <= TOOL_RESULT_SENT_CHAR_BUDGET, |
| 1219 | "truncated result should stay bounded, sent {} chars", |
| 1220 | sent.chars().count() |
| 1221 | ); |
| 1222 | assert!( |
| 1223 | !sent.contains("MIDDLE_POISON"), |
| 1224 | "omitted middle should not be sent to the next model turn" |
| 1225 | ); |
| 1226 | assert_ne!(sent, huge_output); |
| 1227 | }); |
| 1228 | } |
| 1229 | |
| 1230 | #[test] |
| 1231 | fn request_builder_does_not_dedup_short_tool_results_for_wire() { |
| 1232 | let output = "same tool output"; |
| 1233 | let messages = vec![ |
| 1234 | tool_use_message("tool-1", "read_file", json!({"path": "README.md"})), |
| 1235 | tool_result_message("tool-1", output), |
| 1236 | tool_use_message("tool-2", "read_file", json!({"path": "README.md"})), |
| 1237 | tool_result_message("tool-2", output), |
| 1238 | ]; |
| 1239 | |
| 1240 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1241 | let first = tool_message_content(&built, 0); |
| 1242 | let second = tool_message_content(&built, 1); |
| 1243 | |
| 1244 | assert_eq!(first, output); |
| 1245 | assert_eq!(second, output); |
| 1246 | assert!(!second.contains("<TOOL_RESULT_REF"), "got: {second}"); |
| 1247 | } |
| 1248 | |
| 1249 | #[test] |
| 1250 | fn request_builder_deduplicates_medium_identical_tool_results_to_earlier_message() { |
| 1251 | with_tool_result_sha_spillover_root(|| { |
| 1252 | // 2,000 chars is intentionally above TOOL_RESULT_DEDUP_MIN_CHARS |
| 1253 | // (1,024) but below TOOL_RESULT_SENT_CHAR_BUDGET (12,000). This |
| 1254 | // verifies the cache-saving path for repeated medium outputs that |
| 1255 | // do not otherwise need truncation. |
| 1256 | let output = "A".repeat(2_000); |
| 1257 | let messages = vec![ |
| 1258 | tool_use_message("tool-1", "read_file", json!({"path": "README.md"})), |
| 1259 | tool_result_message("tool-1", &output), |
| 1260 | tool_use_message("tool-2", "read_file", json!({"path": "README.md"})), |
| 1261 | tool_result_message("tool-2", &output), |
| 1262 | ]; |
| 1263 | |
| 1264 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1265 | let first = tool_message_content(&built, 0); |
| 1266 | let second = tool_message_content(&built, 1); |
| 1267 | |
| 1268 | assert_eq!(first, output); |
| 1269 | assert!(!first.contains("[TOOL_RESULT_TRUNCATED]"), "got: {first}"); |
| 1270 | assert!( |
| 1271 | second.starts_with("<TOOL_RESULT_REF sha=\""), |
| 1272 | "got: {second}" |
| 1273 | ); |
| 1274 | assert!( |
| 1275 | second.contains("original_message=\"Message #1\""), |
| 1276 | "got: {second}" |
| 1277 | ); |
| 1278 | assert!(second.contains("chars=\"2000\""), "got: {second}"); |
| 1279 | assert!( |
| 1280 | second.contains("source: full content appears in Message #1 earlier in this request"), |
| 1281 | "got: {second}" |
| 1282 | ); |
| 1283 | assert!(!second.contains("retrieve_tool_result"), "got: {second}"); |
| 1284 | }); |
| 1285 | } |
| 1286 | |
| 1287 | #[test] |
| 1288 | fn request_builder_never_dedups_large_identical_write_file_confirmations() { |
| 1289 | with_tool_result_sha_spillover_root(|| { |
| 1290 | // A `write_file` result embeds the unified diff + summary; it is a |
| 1291 | // confirmation, not retrievable data. Two identical >1024-char |
| 1292 | // write_file results must BOTH stay inline — collapsing the second |
| 1293 | // to a SHA ref makes the model lose write-success context and |
| 1294 | // report the file as missing (#1695). |
| 1295 | let output = "A".repeat(2_000); |
| 1296 | let messages = vec![ |
| 1297 | tool_use_message("tool-1", "write_file", json!({"path": "big.txt"})), |
| 1298 | tool_result_message("tool-1", &output), |
| 1299 | tool_use_message("tool-2", "write_file", json!({"path": "big.txt"})), |
| 1300 | tool_result_message("tool-2", &output), |
| 1301 | ]; |
| 1302 | |
| 1303 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1304 | let first = tool_message_content(&built, 0); |
| 1305 | let second = tool_message_content(&built, 1); |
| 1306 | |
| 1307 | assert_eq!(first, output); |
| 1308 | assert_eq!(second, output); |
| 1309 | assert!(!second.contains("<TOOL_RESULT_REF"), "got: {second}"); |
| 1310 | |
| 1311 | // Non-mutation tools still dedup: an identical medium read_file |
| 1312 | // result points back to the first full message in this request. |
| 1313 | let read_messages = vec![ |
| 1314 | tool_use_message("read-1", "read_file", json!({"path": "README.md"})), |
| 1315 | tool_result_message("read-1", &output), |
| 1316 | tool_use_message("read-2", "read_file", json!({"path": "README.md"})), |
| 1317 | tool_result_message("read-2", &output), |
| 1318 | ]; |
| 1319 | let read_built = build_chat_messages(None, &read_messages, "deepseek-v4-flash"); |
| 1320 | let read_first = tool_message_content(&read_built, 0); |
| 1321 | let read_second = tool_message_content(&read_built, 1); |
| 1322 | assert_eq!(read_first, output); |
| 1323 | assert!( |
| 1324 | read_second.starts_with("<TOOL_RESULT_REF sha=\""), |
| 1325 | "got: {read_second}" |
| 1326 | ); |
| 1327 | assert!(read_second.contains("source: full content appears in Message #1")); |
| 1328 | assert!(!read_second.contains("retrieve_tool_result")); |
| 1329 | }); |
| 1330 | } |
| 1331 | |
| 1332 | #[test] |
| 1333 | fn large_unowned_results_stay_bounded_without_false_retrieval_handles() { |
| 1334 | // The adaptive router normally replaces a large result with a |
| 1335 | // session-owned artifact receipt before this provider-wire fallback. |
| 1336 | // If legacy/raw history reaches here, it may be excerpted but must not |
| 1337 | // advertise the process-wide SHA store as retrievable. |
| 1338 | let big_diff = "D".repeat(20_000); |
| 1339 | let sha = sha256_hex(big_diff.as_bytes()); |
| 1340 | |
| 1341 | let messages = vec![ |
| 1342 | tool_use_message("w-1", "write_file", json!({"path": "huge.rs"})), |
| 1343 | tool_result_message("w-1", &big_diff), |
| 1344 | tool_use_message("w-2", "write_file", json!({"path": "huge.rs"})), |
| 1345 | tool_result_message("w-2", &big_diff), |
| 1346 | ]; |
| 1347 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1348 | let first = tool_message_content(&built, 0); |
| 1349 | let second = tool_message_content(&built, 1); |
| 1350 | |
| 1351 | // Mutation confirmations are independently excerpted, never deduped. |
| 1352 | assert!( |
| 1353 | first.contains("[TOOL_RESULT_TRUNCATED]"), |
| 1354 | "first should be truncated, got: {first}" |
| 1355 | ); |
| 1356 | assert!( |
| 1357 | !first.contains("<TOOL_RESULT_REF"), |
| 1358 | "first must not be a dedup ref, got: {first}" |
| 1359 | ); |
| 1360 | assert!( |
| 1361 | !second.contains("<TOOL_RESULT_REF"), |
| 1362 | "second identical write_file must stay inline (#1695), got: {second}" |
| 1363 | ); |
| 1364 | assert!( |
| 1365 | second.contains("[TOOL_RESULT_TRUNCATED]"), |
| 1366 | "second should also be inline-truncated, got: {second}" |
| 1367 | ); |
| 1368 | assert!( |
| 1369 | first.contains(&format!("sha256: {sha}")), |
| 1370 | "truncation block should retain an integrity digest, got: {first}" |
| 1371 | ); |
| 1372 | assert!(first.contains("exact_detail: unavailable")); |
| 1373 | assert!(!first.contains("retrieve_tool_result")); |
| 1374 | |
| 1375 | // A huge non-mutation result cannot refer to an earlier *full* message, |
| 1376 | // because both wire messages are excerpts. It therefore stays a |
| 1377 | // truthful bounded excerpt too. |
| 1378 | let read_messages = vec![ |
| 1379 | tool_use_message("r-1", "read_file", json!({"path": "huge.rs"})), |
| 1380 | tool_result_message("r-1", &big_diff), |
| 1381 | tool_use_message("r-2", "read_file", json!({"path": "huge.rs"})), |
| 1382 | tool_result_message("r-2", &big_diff), |
| 1383 | ]; |
| 1384 | let read_built = build_chat_messages(None, &read_messages, "deepseek-v4-flash"); |
| 1385 | let read_second = tool_message_content(&read_built, 1); |
| 1386 | assert!(read_second.contains("[TOOL_RESULT_TRUNCATED]")); |
| 1387 | assert!(!read_second.contains("<TOOL_RESULT_REF")); |
| 1388 | assert!(!read_second.contains("retrieve_tool_result")); |
| 1389 | } |
| 1390 | |
| 1391 | #[test] |
| 1392 | fn tool_result_budget_is_wire_only_and_does_not_mutate_session_message() { |
| 1393 | let long_output = format!("{}{}", "A".repeat(7_000), "Z".repeat(7_000)); |
| 1394 | let messages = vec![ |
| 1395 | tool_use_message( |
| 1396 | "tool-long", |
| 1397 | "shell_command", |
| 1398 | json!({"command": "cargo test"}), |
| 1399 | ), |
| 1400 | tool_result_message("tool-long", &long_output), |
| 1401 | ]; |
| 1402 | |
| 1403 | let built = build_chat_messages(None, &messages, "deepseek-v4-flash"); |
| 1404 | let sent = tool_message_content(&built, 0); |
| 1405 | assert_ne!(sent, long_output); |
| 1406 | |
| 1407 | match &messages[1].content[0] { |
| 1408 | ContentBlock::ToolResult { content, .. } => assert_eq!(content, &long_output), |
| 1409 | other => panic!("expected tool result, got {other:?}"), |
| 1410 | } |
| 1411 | } |
| 1412 | |
| 1413 | #[test] |
| 1414 | fn cache_inspect_reports_bounded_unowned_tool_result_metadata() { |
| 1415 | let long_output = format!("{}{}", "A".repeat(7_000), "Z".repeat(7_000)); |
| 1416 | let request = MessageRequest { |
| 1417 | model: "deepseek-v4-flash".to_string(), |
| 1418 | messages: vec![ |
| 1419 | tool_use_message("tool-1", "shell_command", json!({"command": "cargo test"})), |
| 1420 | tool_result_message("tool-1", &long_output), |
| 1421 | tool_use_message("tool-2", "shell_command", json!({"command": "cargo test"})), |
| 1422 | tool_result_message("tool-2", &long_output), |
| 1423 | ], |
| 1424 | max_tokens: 0, |
| 1425 | system: None, |
| 1426 | tools: None, |
| 1427 | tool_choice: None, |
| 1428 | metadata: None, |
| 1429 | thinking: None, |
| 1430 | reasoning_effort: None, |
| 1431 | stream: None, |
| 1432 | temperature: None, |
| 1433 | top_p: None, |
| 1434 | }; |
| 1435 | |
| 1436 | let inspection = inspect_prompt_for_request(&request); |
| 1437 | let tool_layers: Vec<_> = inspection |
| 1438 | .layers |
| 1439 | .iter() |
| 1440 | .filter_map(|layer| layer.tool_result.as_ref()) |
| 1441 | .collect(); |
| 1442 | |
| 1443 | assert_eq!(tool_layers.len(), 2); |
| 1444 | for layer in tool_layers { |
| 1445 | assert_eq!(layer.original_chars, 14_000); |
| 1446 | assert!(layer.sent_chars < layer.original_chars); |
| 1447 | assert!(layer.truncated); |
| 1448 | assert!(!layer.deduplicated); |
| 1449 | } |
| 1450 | } |
| 1451 | |
| 1452 | #[test] |
| 1453 | fn mistral_stream_blocks_are_decoded_only_by_the_mistral_style() { |
| 1454 | let chunk = r#"{ |
| 1455 | "choices": [{ |
| 1456 | "index": 0, |
| 1457 | "delta": {"content": [ |
| 1458 | {"type": "thinking", "thinking": [ |
| 1459 | {"type": "text", "text": "private trace"} |
| 1460 | ], "closed": true}, |
| 1461 | {"type": "text", "text": "public answer"} |
| 1462 | ]}, |
| 1463 | "finish_reason": null |
| 1464 | }] |
| 1465 | }"#; |
| 1466 | |
| 1467 | let mistral = decode_chunks_with_style(&[chunk], ReasoningStreamStyle::MistralBlocks); |
| 1468 | assert!(mistral.iter().any(|event| matches!( |
| 1469 | event, |
| 1470 | StreamEvent::ContentBlockDelta { |
| 1471 | delta: Delta::ThinkingDelta { thinking }, |
| 1472 | .. |
| 1473 | } if thinking == "private trace" |
| 1474 | ))); |
| 1475 | assert!(mistral.iter().any(|event| matches!( |
| 1476 | event, |
| 1477 | StreamEvent::ContentBlockDelta { |
| 1478 | delta: Delta::TextDelta { text }, |
| 1479 | .. |
| 1480 | } if text == "public answer" |
| 1481 | ))); |
| 1482 | |
| 1483 | let generic = decode_chunks_with_style(&[chunk], ReasoningStreamStyle::None); |
| 1484 | assert!(!generic.iter().any(|event| matches!( |
| 1485 | event, |
| 1486 | StreamEvent::ContentBlockDelta { |
| 1487 | delta: Delta::ThinkingDelta { .. }, |
| 1488 | .. |
| 1489 | } |
| 1490 | ))); |
| 1491 | assert!(!generic.iter().any(|event| matches!( |
| 1492 | event, |
| 1493 | StreamEvent::ContentBlockDelta { |
| 1494 | delta: Delta::TextDelta { .. }, |
| 1495 | .. |
| 1496 | } |
| 1497 | ))); |
| 1498 | } |
| 1499 | |
| 1500 | #[test] |
| 1501 | fn deepseek_flash_v41_classifies_reasoning_through_the_catalog() { |
| 1502 | // #6044: V4.1's official id dropped the version number, so the literal |
| 1503 | // `deepseek-v4` arms cannot see it. The catalog owns the capability and |
| 1504 | // every classifier — stream style, wire replay, prompt inspection — must |
| 1505 | // read it there instead of relying on another classifier's fallback. |
| 1506 | let base_url = "https://api.deepseek.com"; |
| 1507 | assert!( |
| 1508 | requires_reasoning_content("deepseek-flash"), |
| 1509 | "the name gate must recognize the official V4.1 id through the catalog" |
| 1510 | ); |
| 1511 | assert!( |
| 1512 | should_replay_reasoning_content("deepseek-flash", None), |
| 1513 | "prompt inspection must agree with the wire request" |
| 1514 | ); |
| 1515 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 1516 | ApiProvider::Deepseek, |
| 1517 | base_url, |
| 1518 | "deepseek-flash", |
| 1519 | None, |
| 1520 | )); |
| 1521 | |
| 1522 | let style = |
| 1523 | reasoning_stream_style_for_route(ApiProvider::Deepseek, base_url, "deepseek-flash", None); |
| 1524 | assert_eq!(style, ReasoningStreamStyle::SeparateField); |
| 1525 | let events = decode_chunks_with_style( |
| 1526 | &[r#"{"choices":[{"delta":{"reasoning_content":"private flash plan"}}]}"#], |
| 1527 | style, |
| 1528 | ); |
| 1529 | assert_eq!(thinking_delta_text(&events), "private flash plan"); |
| 1530 | assert_eq!( |
| 1531 | text_delta_text(&events), |
| 1532 | "", |
| 1533 | "reasoning must never leak into visible prose" |
| 1534 | ); |
| 1535 | |
| 1536 | // A non-reasoning DeepSeek name stays literal: the prefix alone is not |
| 1537 | // evidence, the catalog entry is. |
| 1538 | assert!(!requires_reasoning_content("deepseek-coder")); |
| 1539 | } |
| 1540 |