| 1 | //! Narrow provider-native web-search client. |
| 2 | //! |
| 3 | //! This adapter reuses the active route's authenticated HTTP client without |
| 4 | //! exposing credentials to tool code. Route capability facts decide whether |
| 5 | //! the adapter is attached; this module only speaks the three documented |
| 6 | //! first-party wire contracts. |
| 7 | |
| 8 | use anyhow::{Context, Result, bail}; |
| 9 | use reqwest::header::{HeaderName, HeaderValue}; |
| 10 | use serde_json::{Value, json}; |
| 11 | |
| 12 | use crate::config::ProviderKind; |
| 13 | |
| 14 | use super::{CodewhaleClient, api_url, responses_api_url}; |
| 15 | |
| 16 | mod zai; |
| 17 | |
| 18 | mod kimi; |
| 19 | |
| 20 | #[derive(Clone)] |
| 21 | pub(crate) struct ProviderNativeSearchClient { |
| 22 | pub(super) inner: CodewhaleClient, |
| 23 | } |
| 24 | |
| 25 | #[derive(Clone)] |
| 26 | pub(crate) struct ProviderNativeSearchRequest { |
| 27 | pub(crate) query: String, |
| 28 | pub(crate) max_results: u8, |
| 29 | pub(crate) domains: Vec<String>, |
| 30 | } |
| 31 | |
| 32 | #[derive(Clone, PartialEq, Eq)] |
| 33 | pub(crate) struct ProviderNativeCitation { |
| 34 | pub(crate) url: String, |
| 35 | pub(crate) title: String, |
| 36 | pub(crate) snippet: Option<String>, |
| 37 | pub(crate) published: Option<String>, |
| 38 | } |
| 39 | |
| 40 | #[derive(Clone, PartialEq, Eq)] |
| 41 | pub(crate) struct ProviderNativeSearchResponse { |
| 42 | pub(crate) answer: Option<String>, |
| 43 | pub(crate) citations: Vec<ProviderNativeCitation>, |
| 44 | /// The provider stopped the answer at its output limit (#6508). |
| 45 | pub(crate) truncated: bool, |
| 46 | } |
| 47 | |
| 48 | /// What the Anthropic and MiMo adapters requested before #6508, kept for a |
| 49 | /// model whose output ceiling the catalogue does not document. |
| 50 | const PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS: u32 = 2_048; |
| 51 | |
| 52 | /// Ceiling on the output a native-search request asks for (#6508). These are |
| 53 | /// single-shot, non-streaming requests: nothing arrives until the whole answer |
| 54 | /// is generated, so web_search gives the attempt enough time to generate what |
| 55 | /// was asked for (`tools::web_search::native_answer_time_budget`). The clamp |
| 56 | /// bounds that wait: at the budget's assumed generation rate, 8,192 tokens is |
| 57 | /// about four minutes before a stalled provider falls back to another backend. |
| 58 | /// It is not a context or API limit; the model's own output ceiling still |
| 59 | /// applies beneath it. |
| 60 | const NATIVE_SEARCH_MAX_OUTPUT_TOKENS: u32 = 8_192; |
| 61 | |
| 62 | impl ProviderNativeSearchClient { |
| 63 | #[must_use] |
| 64 | pub(crate) fn new(inner: CodewhaleClient) -> Option<Self> { |
| 65 | matches!( |
| 66 | inner.api_provider, |
| 67 | ProviderKind::Openai |
| 68 | | ProviderKind::Anthropic |
| 69 | | ProviderKind::Xai |
| 70 | | ProviderKind::XiaomiMimo |
| 71 | | ProviderKind::Zai |
| 72 | | ProviderKind::ModelstudioTokenPlan |
| 73 | | ProviderKind::Deepseek |
| 74 | | ProviderKind::Moonshot |
| 75 | ) |
| 76 | .then_some(Self { inner }) |
| 77 | } |
| 78 | |
| 79 | #[must_use] |
| 80 | pub(crate) fn provider(&self) -> ProviderKind { |
| 81 | self.inner.api_provider |
| 82 | } |
| 83 | |
| 84 | #[must_use] |
| 85 | pub(crate) fn model(&self) -> &str { |
| 86 | &self.inner.default_model |
| 87 | } |
| 88 | |
| 89 | /// Output tokens to request for a native search answer: the search |
| 90 | /// model's route ceiling up to [`NATIVE_SEARCH_MAX_OUTPUT_TOKENS`]. When |
| 91 | /// the catalogue does not document that ceiling, keep `fallback` (the |
| 92 | /// value this adapter always sent) rather than risk a provider rejecting |
| 93 | /// a larger request. |
| 94 | #[must_use] |
| 95 | pub(super) fn answer_output_tokens(&self, fallback: u32) -> u32 { |
| 96 | let model = &self.inner.default_model; |
| 97 | let route_cap = self.inner.effective_max_output_tokens(model); |
| 98 | let documented = matches!( |
| 99 | crate::route_budget::output_ceiling_source(self.inner.api_provider, model), |
| 100 | crate::route_budget::OutputCeilingSource::Documented(_) |
| 101 | ); |
| 102 | let wanted = if documented { |
| 103 | NATIVE_SEARCH_MAX_OUTPUT_TOKENS |
| 104 | } else { |
| 105 | fallback |
| 106 | }; |
| 107 | wanted.min(route_cap) |
| 108 | } |
| 109 | |
| 110 | /// Output tokens this client asks for per native-search answer, or `None` |
| 111 | /// for adapters that leave the answer length to the provider's default |
| 112 | /// (the Responses and Z.ai adapters). web_search sizes the attempt's time |
| 113 | /// budget from this, so the request and the wait cannot disagree. |
| 114 | #[must_use] |
| 115 | pub(crate) fn requested_answer_output_tokens(&self) -> Option<u32> { |
| 116 | match self.inner.api_provider { |
| 117 | ProviderKind::Anthropic | ProviderKind::XiaomiMimo => { |
| 118 | Some(self.answer_output_tokens(PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS)) |
| 119 | } |
| 120 | ProviderKind::Moonshot => { |
| 121 | Some(self.answer_output_tokens(kimi::PRIOR_NATIVE_SEARCH_MAX_COMPLETION_TOKENS)) |
| 122 | } |
| 123 | _ => None, |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | #[must_use] |
| 128 | pub(crate) fn base_url(&self) -> &str { |
| 129 | &self.inner.base_url |
| 130 | } |
| 131 | |
| 132 | #[must_use] |
| 133 | pub(crate) fn host(&self) -> Option<String> { |
| 134 | reqwest::Url::parse(&self.inner.base_url) |
| 135 | .ok() |
| 136 | .and_then(|url| url.host_str().map(str::to_ascii_lowercase)) |
| 137 | } |
| 138 | |
| 139 | #[must_use] |
| 140 | pub(crate) fn cache_identity(&self) -> String { |
| 141 | format!( |
| 142 | "provider-native://{}/{}/{}", |
| 143 | self.inner.api_provider.as_str(), |
| 144 | self.host().as_deref().unwrap_or("unknown-host"), |
| 145 | self.inner.default_model |
| 146 | ) |
| 147 | } |
| 148 | |
| 149 | #[must_use] |
| 150 | pub(crate) const fn maximum_domain_count(&self) -> Option<usize> { |
| 151 | match self.inner.api_provider { |
| 152 | ProviderKind::Xai => Some(5), |
| 153 | ProviderKind::Openai => Some(100), |
| 154 | ProviderKind::Anthropic => None, |
| 155 | _ => Some(0), |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | pub(crate) async fn search( |
| 160 | &self, |
| 161 | request: &ProviderNativeSearchRequest, |
| 162 | ) -> Result<ProviderNativeSearchResponse> { |
| 163 | // This adapter performs model-backed inference directly instead of |
| 164 | // calling `CodewhaleClient::create_message*`. It must therefore join |
| 165 | // the same attached-run ownership boundary explicitly. The guard is |
| 166 | // retained through response decode so a relay writer cannot start |
| 167 | // while this result is still able to feed the interactive turn. |
| 168 | let _inference = self.inner.acquire_remote_control_inference_permit().await; |
| 169 | if self.inner.api_provider == ProviderKind::Moonshot { |
| 170 | // Kimi/Moonshot runs a bounded multi-round agentic search with its |
| 171 | // own request/reply loop, so it cannot share the single-shot body |
| 172 | // dispatch below. It still runs under the inference permit above. |
| 173 | let mut parsed = kimi::search(self, request).await?; |
| 174 | parsed.citations.truncate(usize::from(request.max_results)); |
| 175 | return Ok(parsed); |
| 176 | } |
| 177 | let body = match self.inner.api_provider { |
| 178 | ProviderKind::Openai => build_responses_search_body( |
| 179 | &self.inner.default_model, |
| 180 | request, |
| 181 | ResponsesSearchDialect::Openai, |
| 182 | ), |
| 183 | ProviderKind::Xai => build_responses_search_body( |
| 184 | &self.inner.default_model, |
| 185 | request, |
| 186 | ResponsesSearchDialect::Xai, |
| 187 | ), |
| 188 | ProviderKind::ModelstudioTokenPlan => build_responses_search_body( |
| 189 | &self.inner.default_model, |
| 190 | request, |
| 191 | ResponsesSearchDialect::ModelStudio, |
| 192 | ), |
| 193 | ProviderKind::Deepseek => build_responses_search_body( |
| 194 | &self.inner.default_model, |
| 195 | request, |
| 196 | ResponsesSearchDialect::Deepseek, |
| 197 | ), |
| 198 | ProviderKind::Anthropic => build_anthropic_search_body( |
| 199 | &self.inner.default_model, |
| 200 | request, |
| 201 | self.requested_answer_output_tokens() |
| 202 | .unwrap_or(PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS), |
| 203 | ), |
| 204 | ProviderKind::XiaomiMimo => build_mimo_search_body( |
| 205 | &self.inner.default_model, |
| 206 | request, |
| 207 | self.requested_answer_output_tokens() |
| 208 | .unwrap_or(PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS), |
| 209 | ), |
| 210 | ProviderKind::Zai => zai::build_body(request, &self.inner.base_url)?, |
| 211 | _ => bail!("active provider has no native web-search adapter"), |
| 212 | }; |
| 213 | let url = match self.inner.api_provider { |
| 214 | ProviderKind::Openai | ProviderKind::Xai | ProviderKind::ModelstudioTokenPlan => { |
| 215 | api_url(&self.inner.base_url, "responses") |
| 216 | } |
| 217 | ProviderKind::Deepseek => { |
| 218 | responses_api_url(&self.inner.base_url, self.inner.api_provider) |
| 219 | } |
| 220 | ProviderKind::Anthropic => anthropic_messages_url(&self.inner.base_url), |
| 221 | ProviderKind::XiaomiMimo => api_url(&self.inner.base_url, "chat/completions"), |
| 222 | ProviderKind::Zai => api_url(&self.inner.base_url, "web_search"), |
| 223 | _ => unreachable!("provider checked above"), |
| 224 | }; |
| 225 | let body_bytes = serde_json::to_vec(&body) |
| 226 | .context("failed to serialize provider-native web-search request")?; |
| 227 | let response = self |
| 228 | .inner |
| 229 | .send_with_retry(|| { |
| 230 | self.inner |
| 231 | .http_client |
| 232 | .post(&url) |
| 233 | .header("Accept", "application/json") |
| 234 | .body(body_bytes.clone()) |
| 235 | }) |
| 236 | .await |
| 237 | .context("provider-native web search request failed")?; |
| 238 | let payload = response |
| 239 | .json::<Value>() |
| 240 | .await |
| 241 | .context("provider-native web search returned invalid JSON")?; |
| 242 | let mut parsed = match self.inner.api_provider { |
| 243 | ProviderKind::Openai |
| 244 | | ProviderKind::Xai |
| 245 | | ProviderKind::ModelstudioTokenPlan |
| 246 | | ProviderKind::Deepseek => parse_responses_search(&payload), |
| 247 | ProviderKind::Anthropic => parse_anthropic_search(&payload), |
| 248 | ProviderKind::XiaomiMimo => parse_mimo_search(&payload), |
| 249 | ProviderKind::Zai => zai::parse(&payload), |
| 250 | _ => unreachable!("provider checked above"), |
| 251 | }; |
| 252 | parsed.citations.truncate(usize::from(request.max_results)); |
| 253 | Ok(parsed) |
| 254 | } |
| 255 | |
| 256 | pub(super) async fn post_json( |
| 257 | &self, |
| 258 | url: &str, |
| 259 | body: &Value, |
| 260 | headers: &[(HeaderName, HeaderValue)], |
| 261 | ) -> Result<Value> { |
| 262 | let body_bytes = serde_json::to_vec(&body) |
| 263 | .context("failed to serialize provider-native web-search request")?; |
| 264 | let headers = headers.to_vec(); |
| 265 | let response = self |
| 266 | .inner |
| 267 | .send_with_retry(|| { |
| 268 | let mut request = self |
| 269 | .inner |
| 270 | .http_client |
| 271 | .post(url) |
| 272 | .header("Accept", "application/json") |
| 273 | .body(body_bytes.clone()); |
| 274 | for (name, value) in &headers { |
| 275 | request = request.header(name, value); |
| 276 | } |
| 277 | request |
| 278 | }) |
| 279 | .await |
| 280 | .context("provider-native web search request failed")?; |
| 281 | response |
| 282 | .json::<Value>() |
| 283 | .await |
| 284 | .context("provider-native web search returned invalid JSON") |
| 285 | } |
| 286 | |
| 287 | pub(super) async fn get_json(&self, url: &str) -> Result<Value> { |
| 288 | let response = self |
| 289 | .inner |
| 290 | .send_with_retry(|| { |
| 291 | self.inner |
| 292 | .http_client |
| 293 | .get(url) |
| 294 | .header("Accept", "application/json") |
| 295 | }) |
| 296 | .await |
| 297 | .context("provider-native web search request failed")?; |
| 298 | response |
| 299 | .json::<Value>() |
| 300 | .await |
| 301 | .context("provider-native web search returned invalid JSON") |
| 302 | } |
| 303 | } |
| 304 | |
| 305 | #[derive(Clone, Copy)] |
| 306 | enum ResponsesSearchDialect { |
| 307 | Openai, |
| 308 | Xai, |
| 309 | ModelStudio, |
| 310 | Deepseek, |
| 311 | } |
| 312 | |
| 313 | fn search_prompt(request: &ProviderNativeSearchRequest) -> String { |
| 314 | format!( |
| 315 | "Search the web for the following query and answer only from web sources. \ |
| 316 | Use concise prose with citations and prefer at most {} distinct sources.\n\n{}", |
| 317 | request.max_results, request.query |
| 318 | ) |
| 319 | } |
| 320 | |
| 321 | fn build_responses_search_body( |
| 322 | model: &str, |
| 323 | request: &ProviderNativeSearchRequest, |
| 324 | dialect: ResponsesSearchDialect, |
| 325 | ) -> Value { |
| 326 | let mut tool = json!({ "type": "web_search" }); |
| 327 | if !request.domains.is_empty() |
| 328 | && matches!( |
| 329 | dialect, |
| 330 | ResponsesSearchDialect::Openai | ResponsesSearchDialect::Xai |
| 331 | ) |
| 332 | { |
| 333 | tool["filters"] = json!({ "allowed_domains": request.domains }); |
| 334 | } |
| 335 | let mut body = json!({ |
| 336 | "model": model, |
| 337 | "input": search_prompt(request), |
| 338 | "tools": [tool], |
| 339 | }); |
| 340 | match dialect { |
| 341 | ResponsesSearchDialect::Openai => { |
| 342 | body["tool_choice"] = json!("required"); |
| 343 | body["store"] = json!(false); |
| 344 | body["include"] = json!(["web_search_call.action.sources"]); |
| 345 | } |
| 346 | ResponsesSearchDialect::Xai => { |
| 347 | body["tool_choice"] = json!("required"); |
| 348 | } |
| 349 | ResponsesSearchDialect::ModelStudio => { |
| 350 | body["tool_choice"] = json!("required"); |
| 351 | } |
| 352 | ResponsesSearchDialect::Deepseek => { |
| 353 | body["tool_choice"] = json!({ "type": "web_search" }); |
| 354 | } |
| 355 | } |
| 356 | body |
| 357 | } |
| 358 | |
| 359 | fn build_anthropic_search_body( |
| 360 | model: &str, |
| 361 | request: &ProviderNativeSearchRequest, |
| 362 | max_tokens: u32, |
| 363 | ) -> Value { |
| 364 | let mut tool = json!({ |
| 365 | "type": "web_search_20250305", |
| 366 | "name": "web_search", |
| 367 | "max_uses": 1, |
| 368 | }); |
| 369 | if !request.domains.is_empty() { |
| 370 | tool["allowed_domains"] = json!(request.domains); |
| 371 | } |
| 372 | json!({ |
| 373 | "model": model, |
| 374 | "max_tokens": max_tokens, |
| 375 | "messages": [{ "role": "user", "content": search_prompt(request) }], |
| 376 | "tools": [tool], |
| 377 | }) |
| 378 | } |
| 379 | |
| 380 | fn build_mimo_search_body( |
| 381 | model: &str, |
| 382 | request: &ProviderNativeSearchRequest, |
| 383 | max_completion_tokens: u32, |
| 384 | ) -> Value { |
| 385 | json!({ |
| 386 | "model": model, |
| 387 | "messages": [{ "role": "user", "content": search_prompt(request) }], |
| 388 | "tools": [{ |
| 389 | "type": "web_search", |
| 390 | "max_keyword": 1, |
| 391 | "force_search": true, |
| 392 | "limit": request.max_results, |
| 393 | }], |
| 394 | "tool_choice": "auto", |
| 395 | "max_completion_tokens": max_completion_tokens, |
| 396 | "stream": false, |
| 397 | "thinking": { "type": "disabled" }, |
| 398 | }) |
| 399 | } |
| 400 | |
| 401 | fn anthropic_messages_url(base_url: &str) -> String { |
| 402 | let base = base_url.trim_end_matches('/'); |
| 403 | if base.ends_with("/v1") { |
| 404 | format!("{base}/messages") |
| 405 | } else { |
| 406 | format!("{base}/v1/messages") |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | fn parse_responses_search(payload: &Value) -> ProviderNativeSearchResponse { |
| 411 | let mut answer_parts = Vec::new(); |
| 412 | let mut citations = Vec::new(); |
| 413 | if let Some(output) = payload.get("output").and_then(Value::as_array) { |
| 414 | for item in output { |
| 415 | let item_type = item.get("type").and_then(Value::as_str); |
| 416 | if item_type == Some("web_search_call") |
| 417 | && let Some(action) = item.get("action") |
| 418 | { |
| 419 | if let Some(sources) = action.get("sources").and_then(Value::as_array) { |
| 420 | for source in sources { |
| 421 | push_citation(&mut citations, citation_from_value(source, None, None)); |
| 422 | } |
| 423 | } |
| 424 | push_citation(&mut citations, citation_from_value(action, None, None)); |
| 425 | } |
| 426 | |
| 427 | if item_type == Some("message") |
| 428 | && let Some(content) = item.get("content").and_then(Value::as_array) |
| 429 | { |
| 430 | for block in content { |
| 431 | if matches!( |
| 432 | block.get("type").and_then(Value::as_str), |
| 433 | Some("output_text" | "text") |
| 434 | ) && let Some(text) = block.get("text").and_then(Value::as_str) |
| 435 | && !text.trim().is_empty() |
| 436 | { |
| 437 | answer_parts.push(text.trim().to_string()); |
| 438 | } |
| 439 | if let Some(annotations) = block.get("annotations").and_then(Value::as_array) { |
| 440 | for annotation in annotations { |
| 441 | push_citation( |
| 442 | &mut citations, |
| 443 | citation_from_value(annotation, None, None), |
| 444 | ); |
| 445 | } |
| 446 | } |
| 447 | } |
| 448 | } |
| 449 | } |
| 450 | } |
| 451 | if answer_parts.is_empty() |
| 452 | && let Some(output_text) = payload.get("output_text").and_then(Value::as_str) |
| 453 | && !output_text.trim().is_empty() |
| 454 | { |
| 455 | answer_parts.push(output_text.trim().to_string()); |
| 456 | } |
| 457 | for answer in &answer_parts { |
| 458 | for citation in citations_from_text(answer) { |
| 459 | push_citation(&mut citations, Some(citation)); |
| 460 | } |
| 461 | } |
| 462 | if let Some(top_level) = payload.get("citations").and_then(Value::as_array) { |
| 463 | for citation in top_level { |
| 464 | let parsed = citation |
| 465 | .as_str() |
| 466 | .and_then(|url| citation_from_url(url, None, None, None)) |
| 467 | .or_else(|| citation_from_value(citation, None, None)); |
| 468 | push_citation(&mut citations, parsed); |
| 469 | } |
| 470 | } |
| 471 | // A Responses reply cut at its output limit is `incomplete` with reason |
| 472 | // `max_output_tokens`; the text it did return is a partial answer. |
| 473 | let truncated = payload.get("status").and_then(Value::as_str) == Some("incomplete") |
| 474 | && payload |
| 475 | .pointer("/incomplete_details/reason") |
| 476 | .and_then(Value::as_str) |
| 477 | == Some("max_output_tokens"); |
| 478 | ProviderNativeSearchResponse { |
| 479 | answer: joined_answer(answer_parts), |
| 480 | citations, |
| 481 | truncated, |
| 482 | } |
| 483 | } |
| 484 | |
| 485 | fn parse_anthropic_search(payload: &Value) -> ProviderNativeSearchResponse { |
| 486 | let mut answer_parts = Vec::new(); |
| 487 | let mut citations = Vec::new(); |
| 488 | if let Some(content) = payload.get("content").and_then(Value::as_array) { |
| 489 | for block in content { |
| 490 | match block.get("type").and_then(Value::as_str) { |
| 491 | Some("web_search_tool_result") => { |
| 492 | if let Some(results) = block.get("content").and_then(Value::as_array) { |
| 493 | for result in results { |
| 494 | let published = result |
| 495 | .get("page_age") |
| 496 | .and_then(Value::as_str) |
| 497 | .map(str::to_string); |
| 498 | push_citation( |
| 499 | &mut citations, |
| 500 | citation_from_value(result, None, published), |
| 501 | ); |
| 502 | } |
| 503 | } |
| 504 | } |
| 505 | Some("text") => { |
| 506 | if let Some(text) = block.get("text").and_then(Value::as_str) |
| 507 | && !text.trim().is_empty() |
| 508 | { |
| 509 | answer_parts.push(text.trim().to_string()); |
| 510 | } |
| 511 | if let Some(block_citations) = block.get("citations").and_then(Value::as_array) |
| 512 | { |
| 513 | for citation in block_citations { |
| 514 | let snippet = citation |
| 515 | .get("cited_text") |
| 516 | .and_then(Value::as_str) |
| 517 | .map(str::to_string); |
| 518 | push_citation( |
| 519 | &mut citations, |
| 520 | citation_from_value(citation, snippet, None), |
| 521 | ); |
| 522 | } |
| 523 | } |
| 524 | } |
| 525 | _ => {} |
| 526 | } |
| 527 | } |
| 528 | } |
| 529 | ProviderNativeSearchResponse { |
| 530 | answer: joined_answer(answer_parts), |
| 531 | citations, |
| 532 | truncated: payload.get("stop_reason").and_then(Value::as_str) == Some("max_tokens"), |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | fn parse_mimo_search(payload: &Value) -> ProviderNativeSearchResponse { |
| 537 | let message = payload.pointer("/choices/0/message"); |
| 538 | let answer = message |
| 539 | .and_then(|value| value.get("content")) |
| 540 | .and_then(Value::as_str) |
| 541 | .map(str::trim) |
| 542 | .filter(|text| !text.is_empty()) |
| 543 | .map(str::to_string); |
| 544 | let mut citations = Vec::new(); |
| 545 | if let Some(annotations) = message |
| 546 | .and_then(|value| value.get("annotations")) |
| 547 | .and_then(Value::as_array) |
| 548 | .or_else(|| payload.get("annotations").and_then(Value::as_array)) |
| 549 | { |
| 550 | for annotation in annotations { |
| 551 | let Some(url) = annotation.get("url").and_then(Value::as_str) else { |
| 552 | continue; |
| 553 | }; |
| 554 | let title = annotation |
| 555 | .get("title") |
| 556 | .and_then(Value::as_str) |
| 557 | .map(str::to_string); |
| 558 | let snippet = annotation |
| 559 | .get("summary") |
| 560 | .and_then(Value::as_str) |
| 561 | .map(str::to_string); |
| 562 | let published = annotation |
| 563 | .get("publish_time") |
| 564 | .and_then(Value::as_str) |
| 565 | .map(str::to_string); |
| 566 | push_citation( |
| 567 | &mut citations, |
| 568 | citation_from_url(url, title, snippet, published), |
| 569 | ); |
| 570 | } |
| 571 | } |
| 572 | ProviderNativeSearchResponse { |
| 573 | answer: joined_answer(answer.into_iter().collect()), |
| 574 | citations, |
| 575 | truncated: finish_reason_is_length(payload.pointer("/choices/0")), |
| 576 | } |
| 577 | } |
| 578 | |
| 579 | /// A Chat Completions choice that stopped at its output limit. |
| 580 | fn finish_reason_is_length(choice: Option<&Value>) -> bool { |
| 581 | choice |
| 582 | .and_then(|choice| choice.get("finish_reason")) |
| 583 | .and_then(Value::as_str) |
| 584 | == Some("length") |
| 585 | } |
| 586 | |
| 587 | fn citation_from_value( |
| 588 | value: &Value, |
| 589 | snippet: Option<String>, |
| 590 | published: Option<String>, |
| 591 | ) -> Option<ProviderNativeCitation> { |
| 592 | let url = value.get("url").and_then(Value::as_str)?.trim(); |
| 593 | let title = value |
| 594 | .get("title") |
| 595 | .and_then(Value::as_str) |
| 596 | .map(str::trim) |
| 597 | .filter(|title| !title.is_empty()) |
| 598 | .map(str::to_string); |
| 599 | citation_from_url(url, title, snippet, published) |
| 600 | } |
| 601 | |
| 602 | fn citation_from_url( |
| 603 | url: &str, |
| 604 | title: Option<String>, |
| 605 | snippet: Option<String>, |
| 606 | published: Option<String>, |
| 607 | ) -> Option<ProviderNativeCitation> { |
| 608 | let parsed = reqwest::Url::parse(url).ok()?; |
| 609 | if !matches!(parsed.scheme(), "http" | "https") || parsed.host_str().is_none() { |
| 610 | return None; |
| 611 | } |
| 612 | Some(ProviderNativeCitation { |
| 613 | url: url.to_string(), |
| 614 | title: title.unwrap_or_else(|| fallback_title(url)), |
| 615 | snippet, |
| 616 | published, |
| 617 | }) |
| 618 | } |
| 619 | |
| 620 | fn push_citation( |
| 621 | citations: &mut Vec<ProviderNativeCitation>, |
| 622 | candidate: Option<ProviderNativeCitation>, |
| 623 | ) { |
| 624 | let Some(candidate) = candidate else { |
| 625 | return; |
| 626 | }; |
| 627 | if let Some(existing) = citations |
| 628 | .iter_mut() |
| 629 | .find(|existing| existing.url == candidate.url) |
| 630 | { |
| 631 | if existing.title == fallback_title(&existing.url) |
| 632 | && candidate.title != fallback_title(&candidate.url) |
| 633 | { |
| 634 | existing.title = candidate.title; |
| 635 | } |
| 636 | if existing.snippet.is_none() { |
| 637 | existing.snippet = candidate.snippet; |
| 638 | } |
| 639 | if existing.published.is_none() { |
| 640 | existing.published = candidate.published; |
| 641 | } |
| 642 | return; |
| 643 | } |
| 644 | citations.push(candidate); |
| 645 | } |
| 646 | |
| 647 | fn citations_from_text(text: &str) -> Vec<ProviderNativeCitation> { |
| 648 | let mut citations = Vec::new(); |
| 649 | let mut offset = 0; |
| 650 | while offset < text.len() { |
| 651 | let remaining = &text[offset..]; |
| 652 | let relative_start = match (remaining.find("https://"), remaining.find("http://")) { |
| 653 | (Some(https), Some(http)) => Some(https.min(http)), |
| 654 | (Some(https), None) => Some(https), |
| 655 | (None, Some(http)) => Some(http), |
| 656 | (None, None) => None, |
| 657 | }; |
| 658 | let Some(relative_start) = relative_start else { |
| 659 | break; |
| 660 | }; |
| 661 | let start = offset + relative_start; |
| 662 | let tail = &text[start..]; |
| 663 | // Balanced parentheses belong to the URL (Wikipedia titles such as |
| 664 | // `Foo_(bar)` keep their closing paren); an unmatched closer ends it. |
| 665 | let mut open_parens = 0_usize; |
| 666 | let end = tail |
| 667 | .char_indices() |
| 668 | .find_map(|(index, ch)| match ch { |
| 669 | '(' => { |
| 670 | open_parens += 1; |
| 671 | None |
| 672 | } |
| 673 | ')' if open_parens > 0 => { |
| 674 | open_parens -= 1; |
| 675 | None |
| 676 | } |
| 677 | _ => (index > 0 |
| 678 | && (ch.is_whitespace() |
| 679 | || matches!(ch, ')' | ']' | '}' | '>' | '"' | '\'' | '`'))) |
| 680 | .then_some(index), |
| 681 | }) |
| 682 | .unwrap_or(tail.len()); |
| 683 | let url = tail[..end].trim_end_matches(['.', ',', ';', ':', '!', '?']); |
| 684 | push_citation(&mut citations, citation_from_url(url, None, None, None)); |
| 685 | offset = start + end.max(1); |
| 686 | } |
| 687 | citations |
| 688 | } |
| 689 | |
| 690 | fn fallback_title(url: &str) -> String { |
| 691 | reqwest::Url::parse(url) |
| 692 | .ok() |
| 693 | .and_then(|parsed| parsed.host_str().map(str::to_string)) |
| 694 | .unwrap_or_else(|| "Web source".to_string()) |
| 695 | } |
| 696 | |
| 697 | /// Join the answer parts whole. No per-adapter cap: the provider already |
| 698 | /// bounded its output tokens, and the engine's route budget bounds what |
| 699 | /// reaches the model inline while keeping the rest recoverable (#6508). |
| 700 | fn joined_answer(parts: Vec<String>) -> Option<String> { |
| 701 | let joined = parts.join("\n\n"); |
| 702 | let trimmed = joined.trim(); |
| 703 | (!trimmed.is_empty()).then(|| trimmed.to_string()) |
| 704 | } |
| 705 | |
| 706 | #[cfg(test)] |
| 707 | mod tests { |
| 708 | use super::*; |
| 709 | use crate::config::{Config, ProviderConfig, ProvidersConfig}; |
| 710 | use wiremock::matchers::{body_partial_json, header, method, path}; |
| 711 | use wiremock::{Mock, MockServer, ResponseTemplate}; |
| 712 | |
| 713 | fn request() -> ProviderNativeSearchRequest { |
| 714 | ProviderNativeSearchRequest { |
| 715 | query: "current release".to_string(), |
| 716 | max_results: 3, |
| 717 | domains: vec!["example.com".to_string()], |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | fn xai_client_with_boundary( |
| 722 | server: &MockServer, |
| 723 | isolated: bool, |
| 724 | unrelated: bool, |
| 725 | ) -> ProviderNativeSearchClient { |
| 726 | let config = Config { |
| 727 | provider: Some("xai".to_string()), |
| 728 | providers: Some(ProvidersConfig { |
| 729 | xai: ProviderConfig { |
| 730 | api_key: Some("xai-test-key".to_string()), |
| 731 | base_url: Some(format!("{}/v1", server.uri())), |
| 732 | model: Some("grok-4.5".to_string()), |
| 733 | ..ProviderConfig::default() |
| 734 | }, |
| 735 | ..ProvidersConfig::default() |
| 736 | }), |
| 737 | runtime_chat_isolated: isolated, |
| 738 | runtime_thread_inference_unrelated: unrelated, |
| 739 | ..Config::default() |
| 740 | }; |
| 741 | ProviderNativeSearchClient::new(CodewhaleClient::new(&config).expect("test xAI client")) |
| 742 | .expect("xAI native adapter") |
| 743 | } |
| 744 | |
| 745 | #[test] |
| 746 | fn responses_payload_requires_search_and_keeps_domains_provider_side() { |
| 747 | let body = |
| 748 | build_responses_search_body("gpt-5.6", &request(), ResponsesSearchDialect::Openai); |
| 749 | assert_eq!(body["tools"][0]["type"], "web_search"); |
| 750 | assert_eq!( |
| 751 | body["tools"][0]["filters"]["allowed_domains"][0], |
| 752 | "example.com" |
| 753 | ); |
| 754 | assert_eq!(body["tool_choice"], "required"); |
| 755 | assert_eq!(body["include"][0], "web_search_call.action.sources"); |
| 756 | } |
| 757 | |
| 758 | #[test] |
| 759 | fn modelstudio_payload_uses_required_harness_search_without_filters() { |
| 760 | let body = build_responses_search_body( |
| 761 | "qwen3.8-max", |
| 762 | &request(), |
| 763 | ResponsesSearchDialect::ModelStudio, |
| 764 | ); |
| 765 | assert_eq!(body["tools"][0]["type"], "web_search"); |
| 766 | assert!(body["tools"][0].get("filters").is_none()); |
| 767 | assert_eq!(body["tool_choice"], "required"); |
| 768 | assert!(body.get("include").is_none()); |
| 769 | assert!(body.get("store").is_none()); |
| 770 | } |
| 771 | |
| 772 | #[test] |
| 773 | fn deepseek_payload_uses_its_responses_search_contract() { |
| 774 | let body = build_responses_search_body( |
| 775 | "deepseek-v4-flash", |
| 776 | &request(), |
| 777 | ResponsesSearchDialect::Deepseek, |
| 778 | ); |
| 779 | assert_eq!(body["tools"][0]["type"], "web_search"); |
| 780 | assert!(body["tools"][0].get("filters").is_none()); |
| 781 | assert_eq!(body["tool_choice"]["type"], "web_search"); |
| 782 | assert!(body.get("include").is_none()); |
| 783 | assert!(body.get("store").is_none()); |
| 784 | } |
| 785 | |
| 786 | #[test] |
| 787 | fn anthropic_payload_uses_basic_direct_search_contract() { |
| 788 | let body = build_anthropic_search_body("claude-opus-4-8", &request(), 2_048); |
| 789 | assert_eq!(body["tools"][0]["type"], "web_search_20250305"); |
| 790 | assert_eq!(body["tools"][0]["max_uses"], 1); |
| 791 | assert_eq!(body["tools"][0]["allowed_domains"][0], "example.com"); |
| 792 | assert_eq!(body["max_tokens"], 2_048); |
| 793 | |
| 794 | let tiny_route = build_anthropic_search_body("claude-opus-4-8", &request(), 128); |
| 795 | assert_eq!(tiny_route["max_tokens"], 128); |
| 796 | } |
| 797 | |
| 798 | #[test] |
| 799 | fn mimo_payload_forces_bounded_web_search_plugin() { |
| 800 | let body = build_mimo_search_body("mimo-v2.5-pro", &request(), 8_192); |
| 801 | assert_eq!(body["tools"][0]["type"], "web_search"); |
| 802 | assert_eq!(body["tools"][0]["force_search"], true); |
| 803 | assert_eq!(body["tools"][0]["limit"], 3); |
| 804 | assert_eq!(body["max_completion_tokens"], 8_192); |
| 805 | assert_eq!(body["thinking"]["type"], "disabled"); |
| 806 | } |
| 807 | |
| 808 | fn anthropic_client(model: &str) -> ProviderNativeSearchClient { |
| 809 | let config = Config { |
| 810 | provider: Some("anthropic".to_string()), |
| 811 | providers: Some(ProvidersConfig { |
| 812 | anthropic: ProviderConfig { |
| 813 | api_key: Some("anthropic-test-key".to_string()), |
| 814 | base_url: Some("https://api.anthropic.com".to_string()), |
| 815 | model: Some(model.to_string()), |
| 816 | ..ProviderConfig::default() |
| 817 | }, |
| 818 | ..ProvidersConfig::default() |
| 819 | }), |
| 820 | ..Config::default() |
| 821 | }; |
| 822 | ProviderNativeSearchClient::new(CodewhaleClient::new(&config).expect("Anthropic client")) |
| 823 | .expect("Anthropic native adapter") |
| 824 | } |
| 825 | |
| 826 | #[test] |
| 827 | fn answer_output_tokens_is_the_model_ceiling_up_to_the_clamp() { |
| 828 | // #6508: native answers were requested at 2,048 tokens and silently |
| 829 | // cut. A catalogued model now gets its route ceiling, up to the clamp. |
| 830 | let client = anthropic_client("claude-opus-4-8"); |
| 831 | let route_cap = client.inner.effective_max_output_tokens("claude-opus-4-8"); |
| 832 | assert!(route_cap > PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS); |
| 833 | assert_eq!( |
| 834 | client.answer_output_tokens(PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS), |
| 835 | NATIVE_SEARCH_MAX_OUTPUT_TOKENS.min(route_cap) |
| 836 | ); |
| 837 | // web_search budgets time from the same number the body carries. |
| 838 | assert_eq!( |
| 839 | client.requested_answer_output_tokens(), |
| 840 | Some(NATIVE_SEARCH_MAX_OUTPUT_TOKENS.min(route_cap)) |
| 841 | ); |
| 842 | |
| 843 | // A model the catalogue does not describe keeps the old request size |
| 844 | // rather than risk a rejection. |
| 845 | let unknown = anthropic_client("claude-unlisted-test-model"); |
| 846 | assert_eq!( |
| 847 | unknown.answer_output_tokens(PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS), |
| 848 | PRIOR_NATIVE_SEARCH_OUTPUT_TOKENS |
| 849 | ); |
| 850 | } |
| 851 | |
| 852 | #[test] |
| 853 | fn anthropic_stop_reason_max_tokens_sets_truncated() { |
| 854 | let cut = parse_anthropic_search(&json!({ |
| 855 | "stop_reason": "max_tokens", |
| 856 | "content": [{ "type": "text", "text": "Partial answer" }] |
| 857 | })); |
| 858 | assert!(cut.truncated); |
| 859 | assert_eq!(cut.answer.as_deref(), Some("Partial answer")); |
| 860 | |
| 861 | let whole = parse_anthropic_search(&json!({ |
| 862 | "stop_reason": "end_turn", |
| 863 | "content": [{ "type": "text", "text": "Whole answer" }] |
| 864 | })); |
| 865 | assert!(!whole.truncated); |
| 866 | } |
| 867 | |
| 868 | #[test] |
| 869 | fn responses_incomplete_max_output_tokens_sets_truncated() { |
| 870 | let payload = |status: &str, reason: &str| { |
| 871 | json!({ |
| 872 | "status": status, |
| 873 | "incomplete_details": { "reason": reason }, |
| 874 | "output": [{ |
| 875 | "type": "message", |
| 876 | "content": [{ "type": "output_text", "text": "Partial answer" }] |
| 877 | }] |
| 878 | }) |
| 879 | }; |
| 880 | assert!(parse_responses_search(&payload("incomplete", "max_output_tokens")).truncated); |
| 881 | assert!(!parse_responses_search(&payload("incomplete", "content_filter")).truncated); |
| 882 | assert!(!parse_responses_search(&payload("completed", "")).truncated); |
| 883 | } |
| 884 | |
| 885 | #[test] |
| 886 | fn mimo_finish_reason_length_sets_truncated() { |
| 887 | let payload = |finish_reason: &str| { |
| 888 | json!({ |
| 889 | "choices": [{ |
| 890 | "finish_reason": finish_reason, |
| 891 | "message": { "content": "Partial answer" } |
| 892 | }] |
| 893 | }) |
| 894 | }; |
| 895 | assert!(parse_mimo_search(&payload("length")).truncated); |
| 896 | assert!(!parse_mimo_search(&payload("stop")).truncated); |
| 897 | } |
| 898 | |
| 899 | #[test] |
| 900 | fn responses_parser_separates_answer_and_deduplicated_citations() { |
| 901 | let payload = json!({ |
| 902 | "output": [ |
| 903 | { |
| 904 | "type": "web_search_call", |
| 905 | "action": { "sources": [ |
| 906 | { "url": "https://example.com/a", "title": "Source A" } |
| 907 | ] } |
| 908 | }, |
| 909 | { |
| 910 | "type": "message", |
| 911 | "content": [{ |
| 912 | "type": "output_text", |
| 913 | "text": "Grounded answer.", |
| 914 | "annotations": [ |
| 915 | { "type": "url_citation", "url": "https://example.com/a", "title": "Source A" }, |
| 916 | { "type": "url_citation", "url": "https://example.org/b", "title": "Source B" } |
| 917 | ] |
| 918 | }] |
| 919 | } |
| 920 | ] |
| 921 | }); |
| 922 | let parsed = parse_responses_search(&payload); |
| 923 | assert_eq!(parsed.answer.as_deref(), Some("Grounded answer.")); |
| 924 | assert_eq!(parsed.citations.len(), 2); |
| 925 | assert_eq!(parsed.citations[0].title, "Source A"); |
| 926 | assert_eq!(parsed.citations[1].url, "https://example.org/b"); |
| 927 | } |
| 928 | |
| 929 | #[test] |
| 930 | fn long_native_answers_are_returned_whole() { |
| 931 | let long = format!("{} end-of-answer", "grounded sentence. ".repeat(600)); |
| 932 | assert!(long.chars().count() > 10_000); |
| 933 | let payload = json!({ |
| 934 | "output": [{ |
| 935 | "type": "message", |
| 936 | "content": [{ "type": "output_text", "text": long.clone() }] |
| 937 | }] |
| 938 | }); |
| 939 | let parsed = parse_responses_search(&payload); |
| 940 | assert_eq!(parsed.answer.as_deref(), Some(long.trim())); |
| 941 | assert!(parsed.answer.unwrap().ends_with("end-of-answer")); |
| 942 | } |
| 943 | |
| 944 | #[test] |
| 945 | fn responses_parser_keeps_final_message_and_opened_pages_only() { |
| 946 | let payload = json!({ |
| 947 | "output": [ |
| 948 | { |
| 949 | "type": "reasoning", |
| 950 | "content": [{ |
| 951 | "type": "reasoning_text", |
| 952 | "text": "private analysis https://reasoning.example/ must stay hidden" |
| 953 | }] |
| 954 | }, |
| 955 | { |
| 956 | "type": "web_search_call", |
| 957 | "action": { |
| 958 | "type": "open_page", |
| 959 | "url": "https://github.com/codewhale-hq/CodeWhale" |
| 960 | } |
| 961 | }, |
| 962 | { |
| 963 | "type": "message", |
| 964 | "content": [{ |
| 965 | "type": "output_text", |
| 966 | "text": "Official repository: https://github.com/codewhale-hq/CodeWhale", |
| 967 | "annotations": [] |
| 968 | }] |
| 969 | } |
| 970 | ] |
| 971 | }); |
| 972 | |
| 973 | let parsed = parse_responses_search(&payload); |
| 974 | |
| 975 | assert_eq!( |
| 976 | parsed.answer.as_deref(), |
| 977 | Some("Official repository: https://github.com/codewhale-hq/CodeWhale") |
| 978 | ); |
| 979 | assert_eq!(parsed.citations.len(), 1); |
| 980 | assert_eq!( |
| 981 | parsed.citations[0].url, |
| 982 | "https://github.com/codewhale-hq/CodeWhale" |
| 983 | ); |
| 984 | } |
| 985 | |
| 986 | #[test] |
| 987 | fn anthropic_parser_keeps_result_metadata_and_cited_text_separate() { |
| 988 | let payload = json!({ |
| 989 | "content": [ |
| 990 | { |
| 991 | "type": "web_search_tool_result", |
| 992 | "content": [{ |
| 993 | "type": "web_search_result", |
| 994 | "url": "https://example.com/a", |
| 995 | "title": "Source A", |
| 996 | "page_age": "July 18, 2026" |
| 997 | }] |
| 998 | }, |
| 999 | { |
| 1000 | "type": "text", |
| 1001 | "text": "Grounded answer.", |
| 1002 | "citations": [{ |
| 1003 | "type": "web_search_result_location", |
| 1004 | "url": "https://example.com/a", |
| 1005 | "title": "Source A", |
| 1006 | "cited_text": "Supporting passage" |
| 1007 | }] |
| 1008 | } |
| 1009 | ] |
| 1010 | }); |
| 1011 | let parsed = parse_anthropic_search(&payload); |
| 1012 | assert_eq!(parsed.answer.as_deref(), Some("Grounded answer.")); |
| 1013 | assert_eq!(parsed.citations.len(), 1); |
| 1014 | assert_eq!( |
| 1015 | parsed.citations[0].published.as_deref(), |
| 1016 | Some("July 18, 2026") |
| 1017 | ); |
| 1018 | assert_eq!( |
| 1019 | parsed.citations[0].snippet.as_deref(), |
| 1020 | Some("Supporting passage") |
| 1021 | ); |
| 1022 | } |
| 1023 | |
| 1024 | #[test] |
| 1025 | fn mimo_parser_keeps_non_streaming_annotations() { |
| 1026 | let parsed = parse_mimo_search(&json!({ |
| 1027 | "choices": [{ |
| 1028 | "message": { |
| 1029 | "content": "Grounded answer.", |
| 1030 | "annotations": [{ |
| 1031 | "type": "url_citation", |
| 1032 | "url": "https://example.com/weather", |
| 1033 | "title": "Weather", |
| 1034 | "summary": "Forecast", |
| 1035 | "publish_time": "2026-08-28" |
| 1036 | }] |
| 1037 | } |
| 1038 | }] |
| 1039 | })); |
| 1040 | assert_eq!(parsed.answer.as_deref(), Some("Grounded answer.")); |
| 1041 | assert_eq!(parsed.citations.len(), 1); |
| 1042 | assert_eq!(parsed.citations[0].snippet.as_deref(), Some("Forecast")); |
| 1043 | assert_eq!(parsed.citations[0].published.as_deref(), Some("2026-08-28")); |
| 1044 | } |
| 1045 | |
| 1046 | #[test] |
| 1047 | fn non_http_citations_are_rejected() { |
| 1048 | let payload = json!({ "citations": ["javascript:alert(1)"] }); |
| 1049 | assert!(parse_responses_search(&payload).citations.is_empty()); |
| 1050 | } |
| 1051 | |
| 1052 | #[test] |
| 1053 | fn answer_links_preserve_mixed_scheme_source_order() { |
| 1054 | let citations = |
| 1055 | citations_from_text("First http://legacy.example/a, then https://secure.example/b."); |
| 1056 | assert_eq!(citations.len(), 2); |
| 1057 | assert_eq!(citations[0].url, "http://legacy.example/a"); |
| 1058 | assert_eq!(citations[1].url, "https://secure.example/b"); |
| 1059 | } |
| 1060 | |
| 1061 | #[test] |
| 1062 | fn answer_links_keep_balanced_parenthesis_segments() { |
| 1063 | let citations = citations_from_text( |
| 1064 | "See https://en.wikipedia.org/wiki/Foo_(bar) and https://en.wikipedia.org/wiki/Baz_(qux_(nested)) for details.", |
| 1065 | ); |
| 1066 | assert_eq!(citations.len(), 2); |
| 1067 | assert_eq!( |
| 1068 | citations[0].url, "https://en.wikipedia.org/wiki/Foo_(bar)", |
| 1069 | "a balanced closing paren is part of the URL" |
| 1070 | ); |
| 1071 | assert_eq!( |
| 1072 | citations[1].url, "https://en.wikipedia.org/wiki/Baz_(qux_(nested))", |
| 1073 | "nested balanced parens stay intact" |
| 1074 | ); |
| 1075 | |
| 1076 | let unbalanced = citations_from_text("Broken https://en.wikipedia.org/wiki/Foo_(bar here."); |
| 1077 | assert_eq!( |
| 1078 | unbalanced[0].url, "https://en.wikipedia.org/wiki/Foo_(bar", |
| 1079 | "an unclosed paren cannot extend past the next whitespace" |
| 1080 | ); |
| 1081 | } |
| 1082 | |
| 1083 | #[tokio::test] |
| 1084 | async fn xai_adapter_reuses_active_authenticated_transport() { |
| 1085 | let server = MockServer::start().await; |
| 1086 | Mock::given(method("POST")) |
| 1087 | .and(path("/v1/responses")) |
| 1088 | .and(header("authorization", "Bearer xai-test-key")) |
| 1089 | .and(body_partial_json(json!({ |
| 1090 | "model": "grok-4.5", |
| 1091 | "tools": [{ |
| 1092 | "type": "web_search", |
| 1093 | "filters": { "allowed_domains": ["example.com"] } |
| 1094 | }], |
| 1095 | "tool_choice": "required" |
| 1096 | }))) |
| 1097 | .respond_with(ResponseTemplate::new(200).set_body_json(json!({ |
| 1098 | "output_text": "Grounded answer.", |
| 1099 | "citations": ["https://example.com/source"] |
| 1100 | }))) |
| 1101 | .expect(1) |
| 1102 | .mount(&server) |
| 1103 | .await; |
| 1104 | let config = Config { |
| 1105 | provider: Some("xai".to_string()), |
| 1106 | providers: Some(ProvidersConfig { |
| 1107 | xai: ProviderConfig { |
| 1108 | api_key: Some("xai-test-key".to_string()), |
| 1109 | base_url: Some(format!("{}/v1", server.uri())), |
| 1110 | model: Some("grok-4.5".to_string()), |
| 1111 | ..ProviderConfig::default() |
| 1112 | }, |
| 1113 | ..ProvidersConfig::default() |
| 1114 | }), |
| 1115 | ..Config::default() |
| 1116 | }; |
| 1117 | let inner = CodewhaleClient::new(&config).expect("test xAI client"); |
| 1118 | let client = ProviderNativeSearchClient::new(inner).expect("xAI native adapter"); |
| 1119 | let cache_identity = client.cache_identity(); |
| 1120 | assert!(cache_identity.contains("provider-native://xai/")); |
| 1121 | assert!(cache_identity.ends_with("/grok-4.5")); |
| 1122 | assert!(!cache_identity.contains("xai-test-key")); |
| 1123 | |
| 1124 | let response = client.search(&request()).await.expect("native search"); |
| 1125 | |
| 1126 | assert_eq!(response.answer.as_deref(), Some("Grounded answer.")); |
| 1127 | assert_eq!(response.citations.len(), 1); |
| 1128 | assert_eq!(response.citations[0].url, "https://example.com/source"); |
| 1129 | } |
| 1130 | |
| 1131 | #[tokio::test] |
| 1132 | async fn modelstudio_adapter_uses_token_plan_responses_contract() { |
| 1133 | let server = MockServer::start().await; |
| 1134 | Mock::given(method("POST")) |
| 1135 | .and(path("/v1/responses")) |
| 1136 | .and(header("authorization", "Bearer modelstudio-test-key")) |
| 1137 | .and(body_partial_json(json!({ |
| 1138 | "model": "qwen3.8-max", |
| 1139 | "tools": [{ "type": "web_search" }], |
| 1140 | "tool_choice": "required" |
| 1141 | }))) |
| 1142 | .respond_with(ResponseTemplate::new(200).set_body_json(json!({ |
| 1143 | "output": [{ |
| 1144 | "type": "web_search_call", |
| 1145 | "action": { |
| 1146 | "sources": [{ |
| 1147 | "url": "https://example.com/qwen", |
| 1148 | "title": "Qwen source" |
| 1149 | }] |
| 1150 | } |
| 1151 | }] |
| 1152 | }))) |
| 1153 | .expect(1) |
| 1154 | .mount(&server) |
| 1155 | .await; |
| 1156 | let config = Config { |
| 1157 | provider: Some("modelstudio-token-plan".to_string()), |
| 1158 | providers: Some(ProvidersConfig { |
| 1159 | modelstudio_token_plan: ProviderConfig { |
| 1160 | api_key: Some("modelstudio-test-key".to_string()), |
| 1161 | base_url: Some(format!("{}/v1", server.uri())), |
| 1162 | model: Some("qwen3.8-max".to_string()), |
| 1163 | ..ProviderConfig::default() |
| 1164 | }, |
| 1165 | ..ProvidersConfig::default() |
| 1166 | }), |
| 1167 | ..Config::default() |
| 1168 | }; |
| 1169 | let inner = CodewhaleClient::new(&config).expect("test ModelStudio client"); |
| 1170 | let client = ProviderNativeSearchClient::new(inner).expect("Qwen native adapter"); |
| 1171 | |
| 1172 | let response = client.search(&request()).await.expect("native search"); |
| 1173 | |
| 1174 | assert_eq!(response.citations.len(), 1); |
| 1175 | assert_eq!(response.citations[0].url, "https://example.com/qwen"); |
| 1176 | } |
| 1177 | #[tokio::test] |
| 1178 | async fn deepseek_adapter_uses_authenticated_responses_endpoint() { |
| 1179 | let server = MockServer::start().await; |
| 1180 | Mock::given(method("POST")) |
| 1181 | .and(path("/v1/responses")) |
| 1182 | .and(header("authorization", "Bearer deepseek-test-key")) |
| 1183 | .and(body_partial_json(json!({ |
| 1184 | "model": "deepseek-v4-flash", |
| 1185 | "tools": [{ "type": "web_search" }], |
| 1186 | "tool_choice": { "type": "web_search" } |
| 1187 | }))) |
| 1188 | .respond_with(ResponseTemplate::new(200).set_body_json(json!({ |
| 1189 | "output": [ |
| 1190 | { |
| 1191 | "type": "web_search_call", |
| 1192 | "action": { |
| 1193 | "type": "open_page", |
| 1194 | "url": "https://example.com/deepseek" |
| 1195 | } |
| 1196 | }, |
| 1197 | { |
| 1198 | "type": "message", |
| 1199 | "content": [{ |
| 1200 | "type": "output_text", |
| 1201 | "text": "Grounded answer.", |
| 1202 | "annotations": [] |
| 1203 | }] |
| 1204 | } |
| 1205 | ] |
| 1206 | }))) |
| 1207 | .expect(1) |
| 1208 | .mount(&server) |
| 1209 | .await; |
| 1210 | let config = Config { |
| 1211 | provider: Some("deepseek".to_string()), |
| 1212 | providers: Some(ProvidersConfig { |
| 1213 | deepseek: ProviderConfig { |
| 1214 | api_key: Some("deepseek-test-key".to_string()), |
| 1215 | base_url: Some(format!("{}/v1", server.uri())), |
| 1216 | model: Some("deepseek-v4-flash".to_string()), |
| 1217 | ..ProviderConfig::default() |
| 1218 | }, |
| 1219 | ..ProvidersConfig::default() |
| 1220 | }), |
| 1221 | ..Config::default() |
| 1222 | }; |
| 1223 | let inner = CodewhaleClient::new(&config).expect("test DeepSeek client"); |
| 1224 | let client = ProviderNativeSearchClient::new(inner).expect("DeepSeek native adapter"); |
| 1225 | |
| 1226 | let response = client.search(&request()).await.expect("native search"); |
| 1227 | |
| 1228 | assert_eq!(response.answer.as_deref(), Some("Grounded answer.")); |
| 1229 | assert_eq!(response.citations.len(), 1); |
| 1230 | assert_eq!(response.citations[0].url, "https://example.com/deepseek"); |
| 1231 | } |
| 1232 | |
| 1233 | #[tokio::test] |
| 1234 | async fn native_search_obeys_attached_run_ownership_without_blocking_unrelated_runtime() { |
| 1235 | let server = MockServer::start().await; |
| 1236 | Mock::given(method("POST")) |
| 1237 | .and(path("/v1/responses")) |
| 1238 | .respond_with(ResponseTemplate::new(200).set_body_json(json!({ |
| 1239 | "output_text": "Grounded answer.", |
| 1240 | "citations": ["https://example.com/source"] |
| 1241 | }))) |
| 1242 | .expect(3) |
| 1243 | .mount(&server) |
| 1244 | .await; |
| 1245 | let participant = xai_client_with_boundary(&server, false, false); |
| 1246 | let isolated = xai_client_with_boundary(&server, true, false); |
| 1247 | let unrelated = xai_client_with_boundary(&server, false, true); |
| 1248 | |
| 1249 | let ownership = crate::client::acquire_runtime_chat_inference_ownership().await; |
| 1250 | let participant_request = request(); |
| 1251 | let mut waiting = |
| 1252 | tokio::spawn(async move { participant.search(&participant_request).await }); |
| 1253 | assert!( |
| 1254 | tokio::time::timeout(std::time::Duration::from_millis(40), &mut waiting) |
| 1255 | .await |
| 1256 | .is_err(), |
| 1257 | "provider-native inference from the attached run must wait behind Runtime Chat" |
| 1258 | ); |
| 1259 | |
| 1260 | tokio::time::timeout( |
| 1261 | std::time::Duration::from_secs(1), |
| 1262 | isolated.search(&request()), |
| 1263 | ) |
| 1264 | .await |
| 1265 | .expect("isolated relay request must not self-deadlock") |
| 1266 | .expect("isolated native search fixture"); |
| 1267 | tokio::time::timeout( |
| 1268 | std::time::Duration::from_secs(1), |
| 1269 | unrelated.search(&request()), |
| 1270 | ) |
| 1271 | .await |
| 1272 | .expect("unrelated Runtime manager stays concurrent") |
| 1273 | .expect("unrelated native search fixture"); |
| 1274 | |
| 1275 | drop(ownership); |
| 1276 | tokio::time::timeout(std::time::Duration::from_secs(1), waiting) |
| 1277 | .await |
| 1278 | .expect("attached participant resumes after relay settlement") |
| 1279 | .expect("participant task") |
| 1280 | .expect("participant native search fixture"); |
| 1281 | } |
| 1282 | } |
| 1283 |