| 1 | //! Chat Completions API helpers for DeepSeek's OpenAI-compatible endpoint. |
| 2 | //! |
| 3 | //! This is the production code path. Streaming (`create_message_stream`), |
| 4 | //! request building (`build_chat_messages*`), and SSE parsing |
| 5 | //! (`parse_sse_chunk_with_reasoning_style`) all live here. |
| 6 | |
| 7 | use std::collections::HashMap; |
| 8 | use std::io::Write; |
| 9 | use std::pin::Pin; |
| 10 | use std::time::Duration; |
| 11 | |
| 12 | use anyhow::{Context, Result, bail}; |
| 13 | use serde::{Deserialize, Serialize}; |
| 14 | use serde_json::{Value, json}; |
| 15 | use tokio::time::timeout as tokio_timeout; |
| 16 | |
| 17 | use crate::config::{ |
| 18 | TOGETHER_INKLING_MODEL, is_exact_direct_moonshot_k3_route, is_exact_kimi_code_k3_route, |
| 19 | is_exact_zai_chat_route, is_exact_zai_forced_thinking_route, is_exact_zai_tiered_effort_route, |
| 20 | is_kimi_code_membership_model, minimax_m3_route_uses_max_completion_tokens, |
| 21 | moonshot_base_url_is_exact_kimi_code, wire_model_for_provider_route, |
| 22 | }; |
| 23 | |
| 24 | use crate::config::ProviderKind; |
| 25 | use crate::llm_client::StreamEventBox; |
| 26 | use crate::llm_client::sanitize_http_error_body; |
| 27 | use crate::logging; |
| 28 | use codewhale_models::{ |
| 29 | ContentBlock, ContentBlockStart, Delta, Message, MessageDelta, MessageRequest, MessageResponse, |
| 30 | StreamEvent, SystemPrompt, Tool, ToolCaller, Usage, is_openai_gpt_56_api_model, |
| 31 | model_is_openai_reasoning_family, model_supports_reasoning, |
| 32 | }; |
| 33 | |
| 34 | use super::prepared::WireDialect; |
| 35 | use super::role_placement::{RolePlacement, role_placement}; |
| 36 | use super::wire::{extract_sse_data_value, flush_sse_line, push_sse_event_data, take_sse_line}; |
| 37 | use super::{ |
| 38 | CodewhaleClient, ERROR_BODY_MAX_BYTES, SSE_BACKPRESSURE_HIGH_WATERMARK, |
| 39 | SSE_BACKPRESSURE_SLEEP_MS, SSE_MAX_LINES_PER_CHUNK, acquire_stream_buffer, |
| 40 | apply_reasoning_effort, bounded_error_text, from_api_tool_name, parse_usage, |
| 41 | release_stream_buffer, system_to_instructions, to_api_tool_name, |
| 42 | }; |
| 43 | use codewhale_config::route::RouteLimits; |
| 44 | use codewhale_models::Role; |
| 45 | |
| 46 | fn apply_provider_token_limit( |
| 47 | body: &mut Value, |
| 48 | provider: ProviderKind, |
| 49 | base_url: &str, |
| 50 | model: &str, |
| 51 | max_tokens: u32, |
| 52 | ) { |
| 53 | let use_max_completion_tokens = provider == ProviderKind::XiaomiMimo |
| 54 | || (provider == ProviderKind::Openai && model_is_openai_reasoning_family(model)) |
| 55 | || minimax_m3_route_uses_max_completion_tokens(provider, base_url, model) |
| 56 | || is_exact_direct_moonshot_k3_route(provider, base_url, model); |
| 57 | if !use_max_completion_tokens { |
| 58 | return; |
| 59 | } |
| 60 | |
| 61 | if let Some(object) = body.as_object_mut() { |
| 62 | object.remove("max_tokens"); |
| 63 | } |
| 64 | body["max_completion_tokens"] = json!(max_tokens); |
| 65 | } |
| 66 | |
| 67 | fn apply_openai_reasoning_effort( |
| 68 | body: &mut Value, |
| 69 | provider: ProviderKind, |
| 70 | model: &str, |
| 71 | effort: Option<&str>, |
| 72 | ) { |
| 73 | let model_lower = model.trim().to_ascii_lowercase(); |
| 74 | let is_gpt_56 = |
| 75 | provider == ProviderKind::Openai && is_openai_gpt_56_api_model(model_lower.as_str()); |
| 76 | let is_openai_reasoning = |
| 77 | provider == ProviderKind::Openai && model_is_openai_reasoning_family(model); |
| 78 | let is_muse_spark = provider == ProviderKind::Meta |
| 79 | && (model_lower == "muse-spark" || model_lower.starts_with("muse-spark-")); |
| 80 | if !is_openai_reasoning && !is_muse_spark { |
| 81 | return; |
| 82 | } |
| 83 | let Some(effort) = |
| 84 | effort.and_then(|value| openai_compatible_reasoning_effort(value, is_gpt_56, !is_gpt_56)) |
| 85 | else { |
| 86 | return; |
| 87 | }; |
| 88 | body["reasoning_effort"] = json!(effort); |
| 89 | } |
| 90 | |
| 91 | /// xAI's first-party `reasoning_effort` ladder, driven by the bundled |
| 92 | /// catalog row for the exact model id: a row that documents an `effort` |
| 93 | /// option gets the field (`xhigh` only where the row lists it — grok-4.7 and |
| 94 | /// grok-4.6 do, grok-4.5 maps it to `high`); a row without one (grok-4.3, |
| 95 | /// grok-build) or no row at all sends nothing. Grok reasoning cannot be |
| 96 | /// disabled, so `off` is sent as the documented default `high` |
| 97 | /// (<https://docs.x.ai/docs/guides/reasoning>). |
| 98 | fn apply_xai_grok_reasoning_effort( |
| 99 | body: &mut Value, |
| 100 | provider: ProviderKind, |
| 101 | base_url: &str, |
| 102 | model: &str, |
| 103 | effort: Option<&str>, |
| 104 | ) { |
| 105 | if provider != ProviderKind::Xai |
| 106 | || !codewhale_config::provider::is_exact_xai_platform_route( |
| 107 | codewhale_config::ProviderKind::Xai, |
| 108 | base_url, |
| 109 | ) |
| 110 | { |
| 111 | return; |
| 112 | } |
| 113 | let Some(effort) = effort else { |
| 114 | return; |
| 115 | }; |
| 116 | let model = model.trim().to_ascii_lowercase(); |
| 117 | // The bundled row with Codewhale's corrections applied (#6396): the raw |
| 118 | // seed row carries Models.dev's ladder, which lists an effort control for |
| 119 | // grok-4.3 that xAI does not document. |
| 120 | let Some(row) = |
| 121 | crate::provider_lake::bundled_catalog_offering_for_model(ProviderKind::Xai, &model) |
| 122 | else { |
| 123 | return; |
| 124 | }; |
| 125 | let Some(documented) = row |
| 126 | .reasoning_options |
| 127 | .iter() |
| 128 | .find(|option| option["type"] == "effort") |
| 129 | .and_then(|option| option["values"].as_array()) |
| 130 | else { |
| 131 | return; |
| 132 | }; |
| 133 | let supports_xhigh = documented.iter().any(|value| value == "xhigh"); |
| 134 | let wire_effort = match effort.trim().to_ascii_lowercase().as_str() { |
| 135 | "auto" | "automatic" | "" => return, |
| 136 | "off" | "disabled" | "none" | "false" | "high" => "high", |
| 137 | "minimal" | "minimum" | "low" | "light" => "low", |
| 138 | "medium" | "mid" => "medium", |
| 139 | "xhigh" | "max" | "maximum" | "highest" | "ultra" | "ultracode" => { |
| 140 | if supports_xhigh { |
| 141 | "xhigh" |
| 142 | } else { |
| 143 | "high" |
| 144 | } |
| 145 | } |
| 146 | _ => return, |
| 147 | }; |
| 148 | body["reasoning_effort"] = json!(wire_effort); |
| 149 | } |
| 150 | |
| 151 | fn apply_inkling_reasoning_effort( |
| 152 | body: &mut Value, |
| 153 | provider: ProviderKind, |
| 154 | model: &str, |
| 155 | effort: Option<&str>, |
| 156 | ) { |
| 157 | if provider != ProviderKind::Together |
| 158 | || !model.trim().eq_ignore_ascii_case(TOGETHER_INKLING_MODEL) |
| 159 | { |
| 160 | return; |
| 161 | } |
| 162 | |
| 163 | // Inkling's official chat template accepts OpenAI's top-level |
| 164 | // `reasoning_effort` field with this exact vocabulary. It does not use |
| 165 | // Together's generic `thinking` extension or the `xhigh` wire value. |
| 166 | if let Some(object) = body.as_object_mut() { |
| 167 | object.remove("thinking"); |
| 168 | } |
| 169 | let Some(effort) = effort else { |
| 170 | return; |
| 171 | }; |
| 172 | let wire_effort = match effort.trim().to_ascii_lowercase().as_str() { |
| 173 | "off" | "disabled" | "none" | "false" => "none", |
| 174 | "minimal" => "minimal", |
| 175 | "low" => "low", |
| 176 | "medium" | "mid" | "" => "medium", |
| 177 | "high" => "high", |
| 178 | "max" | "xhigh" | "highest" | "ultra" | "ultracode" => "max", |
| 179 | _ => return, |
| 180 | }; |
| 181 | body["reasoning_effort"] = json!(wire_effort); |
| 182 | } |
| 183 | |
| 184 | /// Apply Kimi Code K3's route-specific nested thinking effort after the |
| 185 | /// generic Moonshot shaping. Other Moonshot and Kimi-compatible routes accept |
| 186 | /// only the generic enabled/disabled form, so the exact endpoint and bare |
| 187 | /// model identifier are both part of this guard. |
| 188 | fn apply_kimi_code_k3_reasoning_effort( |
| 189 | body: &mut Value, |
| 190 | provider: ProviderKind, |
| 191 | base_url: &str, |
| 192 | model: &str, |
| 193 | effort: Option<&str>, |
| 194 | ) { |
| 195 | if !is_exact_kimi_code_k3_route(provider, base_url, model) { |
| 196 | return; |
| 197 | } |
| 198 | let Some(effort) = effort else { |
| 199 | return; |
| 200 | }; |
| 201 | |
| 202 | let thinking = match effort.trim().to_ascii_lowercase().as_str() { |
| 203 | "off" | "none" | "disabled" | "false" | "low" | "minimum" | "minimal" | "light" => { |
| 204 | json!({ "type": "enabled", "effort": "low" }) |
| 205 | } |
| 206 | "medium" | "high" => json!({ "type": "enabled", "effort": "high" }), |
| 207 | "xhigh" | "ultra" | "max" => json!({ "type": "enabled", "effort": "max" }), |
| 208 | _ => return, |
| 209 | }; |
| 210 | |
| 211 | // K3 uses the nested `thinking.effort` dialect. Do not leave an |
| 212 | // OpenAI-style effort value behind if another shaping layer was added |
| 213 | // before this route-specific override. |
| 214 | if let Some(object) = body.as_object_mut() { |
| 215 | object.remove("reasoning_effort"); |
| 216 | } |
| 217 | body["thinking"] = thinking; |
| 218 | } |
| 219 | |
| 220 | /// Apply Moonshot's direct K3 reasoning dialect. |
| 221 | /// |
| 222 | /// The pay-as-you-go K3 endpoint is always-thinking and accepts only the |
| 223 | /// top-level `reasoning_effort` values low/high/max. In particular, a generic |
| 224 | /// Moonshot `thinking: {type: disabled}` payload is not truthful for this |
| 225 | /// route. Treat a legacy raw `off` as the lowest supported tier defensively; |
| 226 | /// route-aware callers normalize it before it reaches this layer. |
| 227 | fn apply_direct_moonshot_k3_reasoning_effort( |
| 228 | body: &mut Value, |
| 229 | provider: ProviderKind, |
| 230 | base_url: &str, |
| 231 | model: &str, |
| 232 | effort: Option<&str>, |
| 233 | ) { |
| 234 | if !is_exact_direct_moonshot_k3_route(provider, base_url, model) { |
| 235 | return; |
| 236 | } |
| 237 | |
| 238 | if let Some(object) = body.as_object_mut() { |
| 239 | object.remove("thinking"); |
| 240 | object.remove("reasoning_effort"); |
| 241 | } |
| 242 | let Some(effort) = effort else { |
| 243 | return; |
| 244 | }; |
| 245 | let wire_effort = match effort.trim().to_ascii_lowercase().as_str() { |
| 246 | "off" | "none" | "disabled" | "false" | "low" | "minimum" | "minimal" | "light" => "low", |
| 247 | "medium" | "mid" | "high" | "" => "high", |
| 248 | "xhigh" | "ultra" | "max" | "highest" | "ultracode" => "max", |
| 249 | // `auto` and unknown legacy values leave the field omitted so the |
| 250 | // direct API owns its documented default (`max`). |
| 251 | _ => return, |
| 252 | }; |
| 253 | body["reasoning_effort"] = json!(wire_effort); |
| 254 | } |
| 255 | |
| 256 | /// Keep Z.ai controls on exact first-party routes only. The tiered-effort GLM |
| 257 | /// models (5.2, and the forced-thinking 5.3 family) receive the documented |
| 258 | /// top-level effort, GLM-5.1 and GLM-5-Turbo keep only the generic thinking |
| 259 | /// toggle, and compatible gateways receive neither field because their |
| 260 | /// request dialect is not known from provider/model selection alone. |
| 261 | fn apply_zai_route_reasoning_controls( |
| 262 | body: &mut Value, |
| 263 | provider: ProviderKind, |
| 264 | base_url: &str, |
| 265 | model: &str, |
| 266 | effort: Option<&str>, |
| 267 | ) { |
| 268 | if provider != ProviderKind::Zai { |
| 269 | return; |
| 270 | } |
| 271 | |
| 272 | if let Some(object) = body.as_object_mut() { |
| 273 | object.remove("reasoning_effort"); |
| 274 | if !is_exact_zai_chat_route(provider, base_url) { |
| 275 | // A compatible gateway owns its own request dialect. Provider/model |
| 276 | // selection alone is not evidence that Z.ai's `thinking` object is |
| 277 | // supported there, so fail closed instead of leaking it. |
| 278 | object.remove("thinking"); |
| 279 | return; |
| 280 | } |
| 281 | } |
| 282 | if !crate::config::is_exact_known_zai_reasoning_route(provider, base_url, model) { |
| 283 | if let Some(object) = body.as_object_mut() { |
| 284 | object.remove("thinking"); |
| 285 | } |
| 286 | return; |
| 287 | } |
| 288 | if !is_exact_zai_tiered_effort_route(provider, base_url, model) { |
| 289 | // Exact first-party GLM-5-Turbo and GLM-5.1 keep only the generic |
| 290 | // enabled/disabled thinking control. |
| 291 | return; |
| 292 | } |
| 293 | if is_exact_zai_forced_thinking_route(provider, base_url, model) { |
| 294 | apply_zai_forced_thinking_effort(body, effort); |
| 295 | return; |
| 296 | } |
| 297 | match effort |
| 298 | .map(|value| value.trim().to_ascii_lowercase()) |
| 299 | .as_deref() |
| 300 | { |
| 301 | Some("high") => body["reasoning_effort"] = json!("high"), |
| 302 | Some("xhigh") | Some("max") | Some("highest") | Some("ultra") | Some("ultracode") => { |
| 303 | body["reasoning_effort"] = json!("max"); |
| 304 | } |
| 305 | // Off, lower tiers, omitted effort, and unknown legacy values retain |
| 306 | // only the generic Z.ai thinking control. |
| 307 | _ => {} |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | /// GLM-5.3 and GLM-5.3-Flash are forced-thinking on the exact first-party |
| 312 | /// Z.ai route: `thinking.type: "disabled"` is rejected with an error and |
| 313 | /// `reasoning_effort` accepts only low/high/max. The generic Z.ai layer emits |
| 314 | /// `disabled` for `off`, so a request that was valid for GLM-5.2 fails on |
| 315 | /// 5.3. Rewrite that payload the way the vendor migration note prescribes — |
| 316 | /// keep thinking enabled and send the lowest tier — and map the remaining |
| 317 | /// aliases onto the three documented values, leaving unknown legacy values |
| 318 | /// omitted so the API owns its documented default (`max`). |
| 319 | fn apply_zai_forced_thinking_effort(body: &mut Value, effort: Option<&str>) { |
| 320 | let thinking_disabled = body |
| 321 | .get("thinking") |
| 322 | .and_then(|thinking| thinking.get("type")) |
| 323 | .and_then(Value::as_str) |
| 324 | == Some("disabled"); |
| 325 | if thinking_disabled { |
| 326 | body["thinking"] = json!({ |
| 327 | "type": "enabled", |
| 328 | "clear_thinking": false, |
| 329 | }); |
| 330 | } |
| 331 | let Some(effort) = effort else { |
| 332 | return; |
| 333 | }; |
| 334 | let wire_effort = match effort.trim().to_ascii_lowercase().as_str() { |
| 335 | "off" | "none" | "disabled" | "false" | "low" | "minimum" | "minimal" | "light" => "low", |
| 336 | "medium" | "mid" | "high" => "high", |
| 337 | "xhigh" | "max" | "highest" | "ultra" | "ultracode" => "max", |
| 338 | _ => return, |
| 339 | }; |
| 340 | body["reasoning_effort"] = json!(wire_effort); |
| 341 | } |
| 342 | |
| 343 | /// Add MiniMax's Chat-only reasoning controls only when endpoint and model |
| 344 | /// prove the exact first-party M3 route. A provider label alone is not enough |
| 345 | /// to send MiniMax-specific fields to a compatible gateway or unknown model. |
| 346 | fn apply_minimax_route_reasoning_controls( |
| 347 | body: &mut Value, |
| 348 | provider: ProviderKind, |
| 349 | base_url: &str, |
| 350 | model: &str, |
| 351 | effort: Option<&str>, |
| 352 | ) { |
| 353 | if provider != ProviderKind::Minimax { |
| 354 | return; |
| 355 | } |
| 356 | if let Some(object) = body.as_object_mut() { |
| 357 | object.remove("reasoning_split"); |
| 358 | object.remove("thinking"); |
| 359 | } |
| 360 | if !crate::config::is_exact_minimax_m3_route(provider, base_url, model) { |
| 361 | return; |
| 362 | } |
| 363 | |
| 364 | body["reasoning_split"] = json!(true); |
| 365 | match effort |
| 366 | .map(|value| value.trim().to_ascii_lowercase()) |
| 367 | .as_deref() |
| 368 | { |
| 369 | Some("off" | "disabled" | "none" | "false") => { |
| 370 | body["thinking"] = json!({ "type": "disabled" }); |
| 371 | } |
| 372 | Some( |
| 373 | "low" | "minimal" | "medium" | "mid" | "high" | "xhigh" | "max" | "highest" | "ultra" |
| 374 | | "ultracode" | "", |
| 375 | ) => { |
| 376 | body["thinking"] = json!({ "type": "adaptive" }); |
| 377 | } |
| 378 | _ => {} |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | /// Model Studio's OpenAI-compatible API uses its own top-level reasoning |
| 383 | /// controls. Keep them on verified Alibaba Chat Completions routes: a custom |
| 384 | /// `base_url` points the same provider identity at an arbitrary gateway, and |
| 385 | /// that gateway must not be handed Alibaba's dialect. |
| 386 | /// |
| 387 | /// This is the *sole* writer of Model Studio reasoning fields — |
| 388 | /// `apply_reasoning_effort` deliberately writes nothing for the `Modelstudio*` |
| 389 | /// identities — so the strip below runs for all four variants, including the |
| 390 | /// two Anthropic-dialect ones. Those normally reach the Messages adapter |
| 391 | /// instead, but `wire = "openai"` can route them here, and an unmatched |
| 392 | /// `enable_thinking` left in the body would then go out unguarded. |
| 393 | fn apply_modelstudio_route_reasoning_controls( |
| 394 | body: &mut Value, |
| 395 | provider: ProviderKind, |
| 396 | base_url: &str, |
| 397 | model: &str, |
| 398 | effort: Option<&str>, |
| 399 | ) { |
| 400 | if !matches!( |
| 401 | provider, |
| 402 | ProviderKind::ModelstudioTokenPlan |
| 403 | | ProviderKind::ModelstudioTokenPlanAnthropic |
| 404 | | ProviderKind::ModelstudioCodingPlan |
| 405 | | ProviderKind::ModelstudioCodingPlanAnthropic |
| 406 | ) { |
| 407 | return; |
| 408 | } |
| 409 | |
| 410 | if let Some(object) = body.as_object_mut() { |
| 411 | object.remove("thinking"); |
| 412 | object.remove("enable_thinking"); |
| 413 | object.remove("preserve_thinking"); |
| 414 | object.remove("reasoning_effort"); |
| 415 | } |
| 416 | if !is_exact_modelstudio_chat_route(provider, base_url) { |
| 417 | return; |
| 418 | } |
| 419 | |
| 420 | let thinking_only = modelstudio_model_is_thinking_only(model); |
| 421 | if !thinking_only && !modelstudio_model_is_hybrid(model) { |
| 422 | return; |
| 423 | } |
| 424 | |
| 425 | let thinking_enabled = !modelstudio_effort_disables_thinking(effort); |
| 426 | // Thinking-only models emit `reasoning_content` but reject an |
| 427 | // enable/disable control. Hybrid models use `enable_thinking`. |
| 428 | if !thinking_only { |
| 429 | body["enable_thinking"] = json!(thinking_enabled); |
| 430 | } |
| 431 | if modelstudio_model_supports_preserve_thinking(model) { |
| 432 | // Model Studio otherwise drops assistant `reasoning_content` from the |
| 433 | // next turn's context. This applies even when the provider default |
| 434 | // leaves thinking enabled and no explicit UI effort was selected. |
| 435 | body["preserve_thinking"] = json!(thinking_only || thinking_enabled); |
| 436 | } |
| 437 | if !thinking_only |
| 438 | && thinking_enabled |
| 439 | && let Some(effort) = effort.and_then(modelstudio_reasoning_effort_for_model) |
| 440 | && modelstudio_model_supports_reasoning_effort(model) |
| 441 | { |
| 442 | body["reasoning_effort"] = json!(effort); |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | /// Fail-closed host guard: only Alibaba's own OpenAI-compatible Chat |
| 447 | /// Completions URL shapes count. Anything else (a proxy, a self-hosted |
| 448 | /// gateway, a typo) gets the Model Studio fields stripped and nothing added. |
| 449 | fn is_exact_modelstudio_chat_route(provider: ProviderKind, base_url: &str) -> bool { |
| 450 | let trimmed = base_url.trim().trim_end_matches('/').to_ascii_lowercase(); |
| 451 | let Some((host, path)) = trimmed |
| 452 | .strip_prefix("https://") |
| 453 | .and_then(|rest| rest.split_once('/')) |
| 454 | else { |
| 455 | return false; |
| 456 | }; |
| 457 | |
| 458 | // Includes Token Plan's default and workspace-scoped |
| 459 | // `{workspace}.<region>.maas.aliyuncs.com/compatible-mode/v1` hosts. |
| 460 | let token_plan_chat = host.ends_with(".maas.aliyuncs.com") && path == "compatible-mode/v1"; |
| 461 | let coding_plan_chat = host == "coding-intl.dashscope.aliyuncs.com" && path == "v1"; |
| 462 | // Alibaba's classic pay-as-you-go DashScope endpoints serve the same |
| 463 | // models and the same dialect; leaving them off the allowlist silently |
| 464 | // stripped every reasoning control on a genuine Alibaba host |
| 465 | // (2026-08-04 review). The intl spelling matches the repo's own |
| 466 | // provider defaults. |
| 467 | let classic_dashscope_chat = matches!( |
| 468 | host, |
| 469 | "dashscope.aliyuncs.com" | "dashscope-intl.aliyuncs.com" |
| 470 | ) && path == "compatible-mode/v1"; |
| 471 | |
| 472 | match provider { |
| 473 | // The primary Model Studio provider selects Coding Plan through |
| 474 | // `mode = "coding-plan"`, which resolves this base URL without |
| 475 | // changing the provider enum. Legacy Coding Plan identities remain |
| 476 | // supported as well, so recognize either official Chat route for the |
| 477 | // complete Model Studio OpenAI family. The `*Anthropic` identities |
| 478 | // speak the Messages dialect and are never verified here. |
| 479 | ProviderKind::ModelstudioTokenPlan | ProviderKind::ModelstudioCodingPlan => { |
| 480 | token_plan_chat || coding_plan_chat || classic_dashscope_chat |
| 481 | } |
| 482 | _ => false, |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | fn is_exact_modelstudio_thinking_only_route( |
| 487 | provider: ProviderKind, |
| 488 | base_url: &str, |
| 489 | model: &str, |
| 490 | ) -> bool { |
| 491 | is_exact_modelstudio_chat_route(provider, base_url) && modelstudio_model_is_thinking_only(model) |
| 492 | } |
| 493 | |
| 494 | fn modelstudio_effort_disables_thinking(effort: Option<&str>) -> bool { |
| 495 | effort.is_some_and(|value| { |
| 496 | matches!( |
| 497 | value.trim().to_ascii_lowercase().as_str(), |
| 498 | "off" | "disabled" | "none" | "false" |
| 499 | ) |
| 500 | }) |
| 501 | } |
| 502 | |
| 503 | /// Models with no enable/disable control at all. `models_dev.bundled.json` |
| 504 | /// lists `qwen3.8-max` as `thinking: always_on` and gives `qwen3.8-max-preview` |
| 505 | /// effort/budget options with no `toggle`, so sending `enable_thinking` to |
| 506 | /// either is at best ignored and at worst a 400. |
| 507 | fn modelstudio_model_is_thinking_only(model: &str) -> bool { |
| 508 | let model = model.trim().to_ascii_lowercase(); |
| 509 | matches!( |
| 510 | model.as_str(), |
| 511 | "qwen3.8-max" |
| 512 | | "qwen3.8-max-preview" |
| 513 | // Kimi K2.7 Code is always-thinking. Keep both Alibaba-hosted and |
| 514 | // Moonshot-supplied exact IDs separate from hybrid Kimi variants |
| 515 | // so we never send the unsupported enable_thinking switch. |
| 516 | | "kimi-k2.7-code" |
| 517 | | "kimi/kimi-k2.7-code" |
| 518 | | "kimi/kimi-k2.7-code-highspeed" |
| 519 | ) |
| 520 | } |
| 521 | |
| 522 | fn modelstudio_model_is_hybrid(model: &str) -> bool { |
| 523 | let model = model.trim().to_ascii_lowercase(); |
| 524 | model.starts_with("qwen3.7-") |
| 525 | || model.starts_with("qwen3.6-") |
| 526 | || model.starts_with("qwen3.5-") |
| 527 | || model.starts_with("qwen3-") |
| 528 | || model.starts_with("deepseek-v4") |
| 529 | || model.starts_with("deepseek-v3.2") |
| 530 | || model.starts_with("deepseek-v3.1") |
| 531 | || model.starts_with("kimi-k2.6") |
| 532 | || matches!(model.as_str(), "kimi/kimi-k2.6") |
| 533 | || model.starts_with("kimi-k2.5") |
| 534 | || model.starts_with("glm-") |
| 535 | } |
| 536 | |
| 537 | fn modelstudio_model_supports_preserve_thinking(model: &str) -> bool { |
| 538 | let model = model.trim().to_ascii_lowercase(); |
| 539 | matches!( |
| 540 | model.as_str(), |
| 541 | "qwen3.7-max" |
| 542 | | "qwen3.7-max-us" |
| 543 | | "qwen3.7-max-2026-05-17" |
| 544 | | "qwen3.7-max-2026-05-20" |
| 545 | | "qwen3.7-max-2026-06-08" |
| 546 | | "qwen3.7-max-preview" |
| 547 | | "qwen3.7-plus" |
| 548 | | "qwen3.7-plus-us" |
| 549 | | "qwen3.7-plus-2026-05-26" |
| 550 | | "qwen3.6-max-preview" |
| 551 | | "qwen3.6-plus" |
| 552 | | "qwen3.6-plus-2026-04-02" |
| 553 | | "qwen3.6-flash" |
| 554 | | "qwen3.6-flash-2026-04-16" |
| 555 | | "kimi-k2.6" |
| 556 | | "kimi-k2.7-code" |
| 557 | | "kimi/kimi-k2.6" |
| 558 | | "kimi/kimi-k2.7-code" |
| 559 | | "kimi/kimi-k2.7-code-highspeed" |
| 560 | ) |
| 561 | } |
| 562 | |
| 563 | fn modelstudio_model_supports_reasoning_effort(model: &str) -> bool { |
| 564 | let model = model.trim().to_ascii_lowercase(); |
| 565 | model.starts_with("deepseek-v4") || matches!(model.as_str(), "glm-5.2" | "glm-5.1" | "glm-5") |
| 566 | } |
| 567 | |
| 568 | fn modelstudio_reasoning_effort_for_model(effort: &str) -> Option<&'static str> { |
| 569 | match effort.trim().to_ascii_lowercase().as_str() { |
| 570 | // Model Studio documents low and medium as aliases for high. |
| 571 | "minimal" | "low" | "medium" | "mid" | "high" | "" => Some("high"), |
| 572 | "xhigh" | "max" | "highest" | "ultra" | "ultracode" => Some("max"), |
| 573 | _ => None, |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | /// Final reasoning-control pass shared by streaming and non-streaming Chat |
| 578 | /// Completions requests. Route-specific shapers run after the generic provider |
| 579 | /// layer so they can remove fields that are invalid for their exact endpoint. |
| 580 | pub(super) fn apply_route_reasoning_controls( |
| 581 | body: &mut Value, |
| 582 | provider: ProviderKind, |
| 583 | base_url: &str, |
| 584 | model: &str, |
| 585 | effort: Option<&str>, |
| 586 | ) { |
| 587 | apply_reasoning_effort(body, effort, provider); |
| 588 | apply_modelstudio_route_reasoning_controls(body, provider, base_url, model, effort); |
| 589 | apply_minimax_route_reasoning_controls(body, provider, base_url, model, effort); |
| 590 | apply_inkling_reasoning_effort(body, provider, model, effort); |
| 591 | apply_openai_reasoning_effort(body, provider, model, effort); |
| 592 | apply_xai_grok_reasoning_effort(body, provider, base_url, model, effort); |
| 593 | apply_direct_moonshot_k3_reasoning_effort(body, provider, base_url, model, effort); |
| 594 | apply_kimi_code_k3_reasoning_effort(body, provider, base_url, model, effort); |
| 595 | apply_zai_route_reasoning_controls(body, provider, base_url, model, effort); |
| 596 | apply_mistral_route_reasoning_controls(body, provider, base_url, model, effort); |
| 597 | apply_google_reasoning_effort(body, base_url, model, effort); |
| 598 | } |
| 599 | |
| 600 | /// Mistral's polymorphic reasoning-content contract is only proven on its |
| 601 | /// first-party Chat Completions endpoints. A configured `mistral` provider may |
| 602 | /// point at an arbitrary OpenAI-compatible gateway, so provider identity alone |
| 603 | /// is not enough to opt that route into Mistral's request or response dialect. |
| 604 | fn is_exact_mistral_chat_route(provider: ProviderKind, base_url: &str) -> bool { |
| 605 | if provider != ProviderKind::Mistral { |
| 606 | return false; |
| 607 | } |
| 608 | let trimmed = base_url.trim().trim_end_matches('/').to_ascii_lowercase(); |
| 609 | let Some((host, path)) = trimmed |
| 610 | .strip_prefix("https://") |
| 611 | .and_then(|rest| rest.split_once('/')) |
| 612 | else { |
| 613 | return false; |
| 614 | }; |
| 615 | matches!( |
| 616 | host, |
| 617 | "api.mistral.ai" | "api.eu.mistral.ai" | "api.us.mistral.ai" |
| 618 | ) && path == "v1" |
| 619 | } |
| 620 | |
| 621 | /// Google's OpenAI-compatibility route, identified by the **resolved base |
| 622 | /// URL** rather than by provider identity. Thought signatures are captured |
| 623 | /// from tool-call `extra_content.google.thought_signature` and replayed on |
| 624 | /// the assistant tool-call messages of later turns; thinking models fail |
| 625 | /// closed when a replayed call has no signature. |
| 626 | /// |
| 627 | /// The endpoint carries the signature contract, not the config row that |
| 628 | /// happens to name it: a manually configured `kind="openai-compatible"` |
| 629 | /// provider ([`ProviderKind::Custom`]) pointed at this exact host and path is |
| 630 | /// byte-for-byte the same endpoint as the built-in `google` row, so it must |
| 631 | /// preserve and replay signatures the same way. The converse still holds — |
| 632 | /// a `google` row pointed at some other gateway is not this route and never |
| 633 | /// carries Google-only fields off-endpoint. |
| 634 | fn is_google_openai_compat_chat_route(base_url: &str) -> bool { |
| 635 | let trimmed = base_url.trim().trim_end_matches('/').to_ascii_lowercase(); |
| 636 | let Some((host, path)) = trimmed |
| 637 | .strip_prefix("https://") |
| 638 | .and_then(|rest| rest.split_once('/')) |
| 639 | else { |
| 640 | return false; |
| 641 | }; |
| 642 | host == "generativelanguage.googleapis.com" && path == "v1beta/openai" |
| 643 | } |
| 644 | |
| 645 | /// Gemini models whose thinking makes thought signatures load-bearing on |
| 646 | /// the OpenAI-compat route. Gemini 2.5 Flash-Lite ships with thinking off |
| 647 | /// by default, so a missing signature there degrades with a warning |
| 648 | /// instead of failing the turn. |
| 649 | fn google_model_requires_thought_signatures(model: &str) -> bool { |
| 650 | let model = model.trim().to_ascii_lowercase(); |
| 651 | // Google names the same model both ways on this endpoint, and a route |
| 652 | // configured as `models/gemini-3-pro` matched none of the prefixes below: |
| 653 | // the model that most needs a signature looked like one that needs none, |
| 654 | // so the fail-closed check waved it through and Google rejected the replay |
| 655 | // instead (#6018). |
| 656 | let model = model.strip_prefix("models/").unwrap_or(&model); |
| 657 | if model.starts_with("gemini-3") { |
| 658 | return true; |
| 659 | } |
| 660 | if model.starts_with("gemini-2.5-pro") { |
| 661 | return true; |
| 662 | } |
| 663 | model.starts_with("gemini-2.5-flash") && !model.starts_with("gemini-2.5-flash-lite") |
| 664 | } |
| 665 | |
| 666 | /// Google's compatibility endpoint accepts the ordinary `reasoning_effort` |
| 667 | /// field across Gemini 2.5 and 3. A top-level `google` object is rejected; |
| 668 | /// native thinking controls would require `extra_body.google` instead. |
| 669 | /// Use one control, since the endpoint rejects overlapping effort and native |
| 670 | /// thinking settings. https://ai.google.dev/gemini-api/docs/openai#thinking |
| 671 | fn apply_google_reasoning_effort( |
| 672 | body: &mut serde_json::Value, |
| 673 | base_url: &str, |
| 674 | model: &str, |
| 675 | effort: Option<&str>, |
| 676 | ) { |
| 677 | if !is_google_openai_compat_chat_route(base_url) { |
| 678 | return; |
| 679 | } |
| 680 | let Some(effort) = effort else { |
| 681 | return; |
| 682 | }; |
| 683 | let model = model.trim().to_ascii_lowercase(); |
| 684 | let model = model.strip_prefix("models/").unwrap_or(&model); |
| 685 | let can_disable = model.starts_with("gemini-2.5-") && !model.starts_with("gemini-2.5-pro"); |
| 686 | let effort = match effort.trim().to_ascii_lowercase().as_str() { |
| 687 | "off" | "disabled" | "none" | "false" if can_disable => "none", |
| 688 | // Gemini 3 and 2.5 Pro cannot disable thinking. The compatibility |
| 689 | // layer maps minimal to the selected model's lowest supported level. |
| 690 | "off" | "disabled" | "none" | "false" | "minimal" => "minimal", |
| 691 | "low" => "low", |
| 692 | "medium" | "mid" | "" => "medium", |
| 693 | "high" | "xhigh" | "max" | "highest" | "ultra" | "ultracode" => "high", |
| 694 | _ => return, |
| 695 | }; |
| 696 | body["reasoning_effort"] = json!(effort); |
| 697 | } |
| 698 | |
| 699 | /// Fail closed before transport when Google's OpenAI-compat route would |
| 700 | /// replay tool calls without the thought signatures Google's thinking models |
| 701 | /// require. The error names the model and the tool call and tells the |
| 702 | /// operator how to recover instead of letting Google reject or corrupt the |
| 703 | /// tool loop. |
| 704 | /// |
| 705 | /// Models whose thinking is off by default (Gemini 2.5 Flash-Lite) degrade |
| 706 | /// instead of failing — but never silently: the unsigned replay is reported |
| 707 | /// through the same warning path the reasoning-replay sanitizer uses, so a |
| 708 | /// later tool-turn failure has a receipt. Only tool-call identifiers and the |
| 709 | /// model id are logged; signature bytes never are. |
| 710 | fn validate_google_thought_signature_replay( |
| 711 | base_url: &str, |
| 712 | model: &str, |
| 713 | messages: &[Value], |
| 714 | ) -> Result<()> { |
| 715 | if !is_google_openai_compat_chat_route(base_url) { |
| 716 | return Ok(()); |
| 717 | } |
| 718 | let requires_signatures = google_model_requires_thought_signatures(model); |
| 719 | let mut unsigned_call_ids: Vec<&str> = Vec::new(); |
| 720 | for message in messages { |
| 721 | let Some(tool_calls) = message.get("tool_calls").and_then(Value::as_array) else { |
| 722 | continue; |
| 723 | }; |
| 724 | for call in tool_calls { |
| 725 | let missing = call |
| 726 | .pointer("/extra_content/google/thought_signature") |
| 727 | .and_then(Value::as_str) |
| 728 | .is_none(); |
| 729 | if missing { |
| 730 | let id = call.get("id").and_then(Value::as_str).unwrap_or("?"); |
| 731 | if requires_signatures { |
| 732 | anyhow::bail!( |
| 733 | "Gemini model `{model}` requires a thought signature to replay tool call \ |
| 734 | `{id}`, but none was captured (the turn predates signature capture, or \ |
| 735 | the provider omitted it). Start a new session before using tools on \ |
| 736 | this route." |
| 737 | ); |
| 738 | } |
| 739 | unsigned_call_ids.push(id); |
| 740 | } |
| 741 | } |
| 742 | } |
| 743 | if !unsigned_call_ids.is_empty() { |
| 744 | // Bounded: identifiers only, and only the first few of them. |
| 745 | let sample = unsigned_call_ids |
| 746 | .iter() |
| 747 | .take(3) |
| 748 | .copied() |
| 749 | .collect::<Vec<_>>() |
| 750 | .join(", "); |
| 751 | tracing::warn!( |
| 752 | model = %model, |
| 753 | unsigned_tool_calls = unsigned_call_ids.len(), |
| 754 | sample_tool_call_ids = %sample, |
| 755 | "replaying tool calls without Google thought signatures on the Gemini \ |
| 756 | OpenAI-compatible route; later signed tool turns may be rejected" |
| 757 | ); |
| 758 | } |
| 759 | Ok(()) |
| 760 | } |
| 761 | |
| 762 | /// Captured Google signatures ride on tool calls as |
| 763 | /// `extra_content.google.thought_signature`. Only Google's OpenAI-compat |
| 764 | /// endpoint may carry them on the wire; every other route gets them stripped |
| 765 | /// so a route switch never leaks Google-only fields to a foreign gateway. |
| 766 | /// |
| 767 | /// Returns how many tool calls lost a signature, so the caller can report a |
| 768 | /// route switch that silently drops signed history instead of dropping it |
| 769 | /// without a receipt. Never returns or logs the signature bytes. |
| 770 | fn strip_google_tool_call_extra_content(messages: &mut [Value]) -> usize { |
| 771 | let mut stripped = 0usize; |
| 772 | for message in messages { |
| 773 | let Some(tool_calls) = message.get_mut("tool_calls").and_then(Value::as_array_mut) else { |
| 774 | continue; |
| 775 | }; |
| 776 | for call in tool_calls { |
| 777 | if let Some(extra) = call.get_mut("extra_content") |
| 778 | && let Some(obj) = extra.as_object_mut() |
| 779 | { |
| 780 | if obj.remove("google").is_some() { |
| 781 | stripped += 1; |
| 782 | } |
| 783 | if obj.is_empty() { |
| 784 | call.as_object_mut().map(|c| c.remove("extra_content")); |
| 785 | } |
| 786 | } |
| 787 | } |
| 788 | } |
| 789 | stripped |
| 790 | } |
| 791 | |
| 792 | fn mistral_model_has_adjustable_reasoning(model: &str) -> bool { |
| 793 | let model = model.trim().to_ascii_lowercase(); |
| 794 | model.starts_with("mistral-medium") || model.starts_with("mistral-small") |
| 795 | } |
| 796 | |
| 797 | fn mistral_model_has_native_reasoning(model: &str) -> bool { |
| 798 | model.trim().to_ascii_lowercase().starts_with("magistral") |
| 799 | } |
| 800 | |
| 801 | fn mistral_model_supports_reasoning(model: &str) -> bool { |
| 802 | mistral_model_has_adjustable_reasoning(model) || mistral_model_has_native_reasoning(model) |
| 803 | } |
| 804 | |
| 805 | fn mistral_reasoning_effort_wire_value(effort: &str) -> Option<&'static str> { |
| 806 | match effort.trim().to_ascii_lowercase().as_str() { |
| 807 | "off" | "disabled" | "none" | "false" => Some("none"), |
| 808 | "high" | "xhigh" | "max" | "highest" | "ultra" | "ultracode" => Some("high"), |
| 809 | _ => None, |
| 810 | } |
| 811 | } |
| 812 | |
| 813 | /// Rewrite assistant messages that carry `reasoning_content` back into the |
| 814 | /// polymorphic `content: [{type: thinking, thinking: [{type: text, text: ...}], |
| 815 | /// closed: bool}, {type: text, text: ...}]` shape that Mistral la Plateforme |
| 816 | /// emits and accepts on replay. Mistral tolerates plain-string history in a |
| 817 | /// thinking-capable conversation, but replaying the original thinking trace |
| 818 | /// keeps multi-turn reasoning quality high per the official docs |
| 819 | /// (docs.mistral.ai/capabilities/reasoning). Non-assistant messages and |
| 820 | /// assistant messages without stored thinking are left untouched. |
| 821 | fn reshape_mistral_messages_for_reasoning_replay(messages: &mut [Value]) { |
| 822 | for message in messages.iter_mut() { |
| 823 | let Some(object) = message.as_object_mut() else { |
| 824 | continue; |
| 825 | }; |
| 826 | if object.get("role").and_then(Value::as_str) != Some("assistant") { |
| 827 | continue; |
| 828 | } |
| 829 | let Some(reasoning) = object.remove("reasoning_content") else { |
| 830 | continue; |
| 831 | }; |
| 832 | let reasoning_text = reasoning |
| 833 | .as_str() |
| 834 | .map(str::to_string) |
| 835 | .filter(|s| !s.trim().is_empty()); |
| 836 | let Some(reasoning_text) = reasoning_text else { |
| 837 | continue; |
| 838 | }; |
| 839 | let text_content = object |
| 840 | .get("content") |
| 841 | .and_then(Value::as_str) |
| 842 | .map(str::to_string); |
| 843 | let mut blocks = vec![json!({ |
| 844 | "type": "thinking", |
| 845 | "thinking": [{"type": "text", "text": reasoning_text}], |
| 846 | "closed": true, |
| 847 | })]; |
| 848 | if let Some(text) = text_content.filter(|s| !s.trim().is_empty()) { |
| 849 | blocks.push(json!({"type": "text", "text": text})); |
| 850 | } |
| 851 | object.insert("content".to_string(), Value::Array(blocks)); |
| 852 | } |
| 853 | } |
| 854 | |
| 855 | /// Extract thinking and text content from a Mistral polymorphic `content` |
| 856 | /// value. Mistral la Plateforme returns `content` as either a plain string |
| 857 | /// (default) or an array of typed blocks (`{type: "thinking", thinking: |
| 858 | /// [{type: "text", text: "..."}], closed: bool}` and `{type: "text", text: |
| 859 | /// "..."}`). This helper flattens the thinking sub-array into a single |
| 860 | /// string and returns any inline text separately. It ignores plain-string |
| 861 | /// `content` (returns `(None, None)`) so the shared string fallback still |
| 862 | /// runs for non-reasoning responses. |
| 863 | fn extract_mistral_polymorphic_content(value: &Value) -> (Option<String>, Option<String>) { |
| 864 | let Some(array) = value.get("content").and_then(Value::as_array) else { |
| 865 | return (None, None); |
| 866 | }; |
| 867 | let mut thinking = String::new(); |
| 868 | let mut text = String::new(); |
| 869 | for block in array { |
| 870 | let Some(kind) = block.get("type").and_then(Value::as_str) else { |
| 871 | continue; |
| 872 | }; |
| 873 | match kind { |
| 874 | "thinking" => { |
| 875 | if let Some(inner) = block.get("thinking").and_then(Value::as_array) { |
| 876 | for sub in inner { |
| 877 | if let Some(sub_text) = sub |
| 878 | .get("text") |
| 879 | .and_then(Value::as_str) |
| 880 | .filter(|s| !s.is_empty()) |
| 881 | { |
| 882 | thinking.push_str(sub_text); |
| 883 | } |
| 884 | } |
| 885 | } else if let Some(inline) = block |
| 886 | .get("thinking") |
| 887 | .and_then(Value::as_str) |
| 888 | .filter(|s| !s.is_empty()) |
| 889 | { |
| 890 | thinking.push_str(inline); |
| 891 | } |
| 892 | } |
| 893 | "text" => { |
| 894 | if let Some(sub_text) = block |
| 895 | .get("text") |
| 896 | .and_then(Value::as_str) |
| 897 | .filter(|s| !s.is_empty()) |
| 898 | { |
| 899 | text.push_str(sub_text); |
| 900 | } |
| 901 | } |
| 902 | _ => {} |
| 903 | } |
| 904 | } |
| 905 | let thinking = (!thinking.is_empty()).then_some(thinking); |
| 906 | let text = (!text.is_empty()).then_some(text); |
| 907 | (thinking, text) |
| 908 | } |
| 909 | |
| 910 | fn apply_mistral_route_reasoning_controls( |
| 911 | body: &mut Value, |
| 912 | provider: ProviderKind, |
| 913 | base_url: &str, |
| 914 | model: &str, |
| 915 | effort: Option<&str>, |
| 916 | ) { |
| 917 | if provider != ProviderKind::Mistral { |
| 918 | return; |
| 919 | } |
| 920 | if let Some(object) = body.as_object_mut() { |
| 921 | object.remove("thinking"); |
| 922 | object.remove("reasoning_effort"); |
| 923 | } |
| 924 | if !is_exact_mistral_chat_route(provider, base_url) |
| 925 | || !mistral_model_has_adjustable_reasoning(model) |
| 926 | { |
| 927 | return; |
| 928 | } |
| 929 | let Some(effort) = effort else { |
| 930 | return; |
| 931 | }; |
| 932 | if let Some(wire) = mistral_reasoning_effort_wire_value(effort) { |
| 933 | body["reasoning_effort"] = json!(wire); |
| 934 | } |
| 935 | } |
| 936 | |
| 937 | /// The direct K3 Chat Completions schema exposes fixed sampling behavior and |
| 938 | /// omits `temperature` and `top_p`. Strip legacy/generic values only from the |
| 939 | /// exact first-party route so compatible gateways keep their own contract. |
| 940 | /// Source: <https://platform.kimi.ai/docs/guide/kimi-k3-quickstart> (verified 2026-07-20). |
| 941 | fn apply_direct_moonshot_k3_fixed_sampling( |
| 942 | body: &mut Value, |
| 943 | provider: ProviderKind, |
| 944 | base_url: &str, |
| 945 | model: &str, |
| 946 | ) { |
| 947 | if !is_exact_direct_moonshot_k3_route(provider, base_url, model) { |
| 948 | return; |
| 949 | } |
| 950 | if let Some(object) = body.as_object_mut() { |
| 951 | object.remove("temperature"); |
| 952 | object.remove("top_p"); |
| 953 | } |
| 954 | } |
| 955 | |
| 956 | /// Kimi Code's documented membership models own their sampling behavior. |
| 957 | /// Strip generic controls only on the exact first-party membership route; |
| 958 | /// custom gateways and unknown model ids retain their own wire contract. |
| 959 | /// Source: <https://www.kimi.com/code/docs/en/third-party-tools/codex.html> |
| 960 | /// (verified 2026-08-26). |
| 961 | fn apply_kimi_code_fixed_sampling( |
| 962 | body: &mut Value, |
| 963 | provider: ProviderKind, |
| 964 | base_url: &str, |
| 965 | model: &str, |
| 966 | ) { |
| 967 | if provider != ProviderKind::Moonshot |
| 968 | || !moonshot_base_url_is_exact_kimi_code(base_url) |
| 969 | || !is_kimi_code_membership_model(model) |
| 970 | { |
| 971 | return; |
| 972 | } |
| 973 | if let Some(object) = body.as_object_mut() { |
| 974 | object.remove("temperature"); |
| 975 | object.remove("top_p"); |
| 976 | } |
| 977 | } |
| 978 | |
| 979 | fn openai_compatible_reasoning_effort( |
| 980 | effort: &str, |
| 981 | supports_max: bool, |
| 982 | supports_minimal: bool, |
| 983 | ) -> Option<&'static str> { |
| 984 | match effort.trim().to_ascii_lowercase().as_str() { |
| 985 | "off" | "disabled" | "none" | "false" => Some("none"), |
| 986 | "minimal" if supports_minimal => Some("minimal"), |
| 987 | "minimal" => Some("low"), |
| 988 | "low" => Some("low"), |
| 989 | "medium" | "mid" | "" => Some("medium"), |
| 990 | "high" => Some("high"), |
| 991 | "xhigh" => Some("xhigh"), |
| 992 | "max" | "highest" | "ultra" | "ultracode" if supports_max => Some("max"), |
| 993 | "max" | "highest" | "ultra" | "ultracode" => Some("xhigh"), |
| 994 | _ => None, |
| 995 | } |
| 996 | } |
| 997 | |
| 998 | fn mirror_minimax_reasoning_details_for_messages(messages: &mut [Value]) { |
| 999 | for message in messages { |
| 1000 | if message.get("role").and_then(Value::as_str) != Some("assistant") { |
| 1001 | continue; |
| 1002 | } |
| 1003 | if message.get("reasoning_details").is_some() { |
| 1004 | continue; |
| 1005 | } |
| 1006 | let Some(reasoning) = message |
| 1007 | .get("reasoning_content") |
| 1008 | .and_then(Value::as_str) |
| 1009 | .filter(|reasoning| !reasoning.trim().is_empty()) |
| 1010 | .map(str::to_string) |
| 1011 | else { |
| 1012 | continue; |
| 1013 | }; |
| 1014 | message["reasoning_details"] = json!([ |
| 1015 | { |
| 1016 | "type": "text", |
| 1017 | "text": reasoning, |
| 1018 | } |
| 1019 | ]); |
| 1020 | } |
| 1021 | } |
| 1022 | |
| 1023 | fn mirror_minimax_reasoning_details_for_body(body: &mut Value, provider: ProviderKind) { |
| 1024 | if provider != ProviderKind::Minimax { |
| 1025 | return; |
| 1026 | } |
| 1027 | let Some(messages) = body.get_mut("messages").and_then(Value::as_array_mut) else { |
| 1028 | return; |
| 1029 | }; |
| 1030 | mirror_minimax_reasoning_details_for_messages(messages); |
| 1031 | } |
| 1032 | |
| 1033 | /// Sanitize every Moonshot chat tool in place, dropping only the tools whose |
| 1034 | /// parameters cannot pass MFJS compatibility validation. |
| 1035 | /// |
| 1036 | /// Per-tool degradation: a single incompatible tool (e.g. a third-party MCP |
| 1037 | /// server whose schema uses keywords outside the MFJS whitelist) is excluded |
| 1038 | /// from this request with a warning instead of failing the whole request |
| 1039 | /// before transport. The tool name is safe to log — it is already visible in |
| 1040 | /// the UI — while the error's `Display` deliberately carries no schema values. |
| 1041 | /// |
| 1042 | /// Returns the names of the dropped tools, in catalog order. |
| 1043 | fn sanitize_moonshot_chat_tools(chat_tools: &mut Vec<Value>) -> Vec<String> { |
| 1044 | let mut dropped = Vec::new(); |
| 1045 | chat_tools.retain_mut(|tool| { |
| 1046 | let Some(function) = tool |
| 1047 | .as_object_mut() |
| 1048 | .and_then(|tool| tool.get_mut("function")) |
| 1049 | .and_then(Value::as_object_mut) |
| 1050 | else { |
| 1051 | return true; |
| 1052 | }; |
| 1053 | let Some(parameters) = function.get_mut("parameters") else { |
| 1054 | return true; |
| 1055 | }; |
| 1056 | match crate::tools::schema_sanitize::sanitize_for_kimi_parameters(parameters) { |
| 1057 | Ok(note) => { |
| 1058 | if let Some(note) = note { |
| 1059 | let description = function |
| 1060 | .get("description") |
| 1061 | .and_then(Value::as_str) |
| 1062 | .unwrap_or_default(); |
| 1063 | let description = if description.is_empty() { |
| 1064 | note |
| 1065 | } else { |
| 1066 | format!("{description} {note}") |
| 1067 | }; |
| 1068 | function.insert("description".to_string(), json!(description)); |
| 1069 | } |
| 1070 | true |
| 1071 | } |
| 1072 | Err(error) => { |
| 1073 | let name = function |
| 1074 | .get("name") |
| 1075 | .and_then(Value::as_str) |
| 1076 | .unwrap_or("<unnamed>") |
| 1077 | .to_string(); |
| 1078 | tracing::warn!( |
| 1079 | tool = %name, |
| 1080 | error = %error, |
| 1081 | "dropping Moonshot tool from this request: parameters failed safe compatibility validation" |
| 1082 | ); |
| 1083 | dropped.push(name); |
| 1084 | false |
| 1085 | } |
| 1086 | } |
| 1087 | }); |
| 1088 | dropped |
| 1089 | } |
| 1090 | |
| 1091 | /// The final Chat Completions wire payload for one request. |
| 1092 | /// |
| 1093 | /// Produced by [`build_chat_wire_body`], the single place where a |
| 1094 | /// `MessageRequest` becomes Chat-shaped JSON. It is reached only through |
| 1095 | /// [`super::CodewhaleClient::prepare_outbound_request`], the shared outbound |
| 1096 | /// seam that the blocking transport, the streaming transport, and |
| 1097 | /// `/preview-request` all consume — so a preview cannot drift from what would |
| 1098 | /// be sent, and no other dialect is projected through this builder. |
| 1099 | /// |
| 1100 | /// Seam concept harvested from PR #1099 (`build_sanitized_chat_completion_body`) |
| 1101 | /// by TaoMu (GTC2080); re-implemented against the current client shape. |
| 1102 | pub(crate) struct ChatWireBody { |
| 1103 | /// Provider-shaped JSON body, post-sanitizers. |
| 1104 | pub(crate) body: Value, |
| 1105 | /// The model id actually placed on the wire (may differ from the |
| 1106 | /// configured/display model for routed providers). |
| 1107 | pub(crate) model: String, |
| 1108 | /// Tokens re-sent because thinking-mode replay substituted |
| 1109 | /// `reasoning_content`. Only computed on the streaming path, which is the |
| 1110 | /// only path that runs the replay sanitizer today. |
| 1111 | pub(crate) replay_input_tokens: Option<u32>, |
| 1112 | /// Wire-normalized tool names omitted because Moonshot's MFJS validator |
| 1113 | /// rejected their parameter schemas. Kept outside the wire body so the |
| 1114 | /// caller can surface one bounded diagnostic without leaking schema data. |
| 1115 | pub(crate) omitted_tool_names: Vec<String>, |
| 1116 | } |
| 1117 | |
| 1118 | /// Build the Chat Completions wire body for `request`. |
| 1119 | /// |
| 1120 | /// `stream` selects the streaming shape (`stream` + `stream_options`) and, to |
| 1121 | /// preserve historical behavior exactly, also gates the thinking-mode replay |
| 1122 | /// sanitizer — the blocking path has never run it. |
| 1123 | pub(crate) fn build_chat_wire_body( |
| 1124 | request: &MessageRequest, |
| 1125 | provider: ProviderKind, |
| 1126 | base_url: &str, |
| 1127 | stream: bool, |
| 1128 | route_limits: Option<RouteLimits>, |
| 1129 | ) -> Result<ChatWireBody> { |
| 1130 | let messages = PromptBuilder::for_request(request).build_for_provider_and_route( |
| 1131 | provider, |
| 1132 | base_url, |
| 1133 | route_limits, |
| 1134 | ); |
| 1135 | let model = { |
| 1136 | let wire = wire_model_for_provider_route(provider, base_url, &request.model); |
| 1137 | codewhale_models::effective_muse_wire_id(&wire).to_string() |
| 1138 | }; |
| 1139 | validate_google_thought_signature_replay(base_url, &model, &messages)?; |
| 1140 | let mut body = if stream { |
| 1141 | json!({ |
| 1142 | "model": model.clone(), |
| 1143 | "messages": messages, |
| 1144 | "max_tokens": request.max_tokens, |
| 1145 | "stream": true, |
| 1146 | "stream_options": { |
| 1147 | "include_usage": true |
| 1148 | }, |
| 1149 | }) |
| 1150 | } else { |
| 1151 | json!({ |
| 1152 | "model": model.clone(), |
| 1153 | "messages": messages, |
| 1154 | "max_tokens": request.max_tokens, |
| 1155 | }) |
| 1156 | }; |
| 1157 | apply_provider_token_limit(&mut body, provider, base_url, &model, request.max_tokens); |
| 1158 | |
| 1159 | if let Some(temperature) = request.temperature { |
| 1160 | body["temperature"] = json!(temperature); |
| 1161 | } |
| 1162 | if let Some(top_p) = request.top_p { |
| 1163 | body["top_p"] = json!(top_p); |
| 1164 | } |
| 1165 | let mut omitted_tool_names = Vec::new(); |
| 1166 | if let Some(tools) = request.tools.as_ref() { |
| 1167 | let mut chat_tools: Vec<_> = tools |
| 1168 | .iter() |
| 1169 | .map(|tool| tool_to_chat_for_base_url(tool, base_url)) |
| 1170 | .collect(); |
| 1171 | // Moonshot function parameters must end at a plain object root. |
| 1172 | // Flatten root composition, preserve valid nested anyOf, and drop |
| 1173 | // only the tools whose parameters cannot pass MFJS validation so one |
| 1174 | // incompatible tool never sinks the whole request. |
| 1175 | if matches!(provider, crate::config::ProviderKind::Moonshot) { |
| 1176 | omitted_tool_names = sanitize_moonshot_chat_tools(&mut chat_tools); |
| 1177 | } |
| 1178 | // xAI rejects a parameters root that is not a plain object schema |
| 1179 | // (e.g. apply_patch's root `oneOf` required-groups) with a 400. |
| 1180 | if matches!(provider, crate::config::ProviderKind::Xai) { |
| 1181 | for t in &mut chat_tools { |
| 1182 | let Some(function) = t |
| 1183 | .as_object_mut() |
| 1184 | .and_then(|t| t.get_mut("function")) |
| 1185 | .and_then(|f| f.as_object_mut()) |
| 1186 | else { |
| 1187 | continue; |
| 1188 | }; |
| 1189 | let note = function.get_mut("parameters").and_then(|parameters| { |
| 1190 | crate::tools::schema_sanitize::sanitize_for_xai_parameters(parameters) |
| 1191 | }); |
| 1192 | if let Some(note) = note |
| 1193 | && let Some(description) = function |
| 1194 | .get_mut("description") |
| 1195 | .and_then(|d| d.as_str().map(str::to_string)) |
| 1196 | { |
| 1197 | function.insert( |
| 1198 | "description".to_string(), |
| 1199 | json!(format!("{description} {note}")), |
| 1200 | ); |
| 1201 | } |
| 1202 | } |
| 1203 | } |
| 1204 | // When per-tool degradation (or the caller) left no tools, omit the |
| 1205 | // key entirely: an empty `tools` array — or a `tool_choice` pointing |
| 1206 | // at a dropped tool — is itself a fresh 400 on strict providers. |
| 1207 | if !chat_tools.is_empty() { |
| 1208 | body["tools"] = json!(chat_tools); |
| 1209 | } |
| 1210 | } |
| 1211 | if should_send_tool_choice_for_chat(provider, request.reasoning_effort.as_deref()) |
| 1212 | && let Some(choice) = request.tool_choice.as_ref() |
| 1213 | && let Some(mapped) = map_tool_choice_for_chat(choice) |
| 1214 | { |
| 1215 | if matches!(provider, crate::config::ProviderKind::Moonshot) |
| 1216 | && let Some(name) = mapped.pointer("/function/name").and_then(Value::as_str) |
| 1217 | && omitted_tool_names.iter().any(|omitted| omitted == name) |
| 1218 | { |
| 1219 | bail!( |
| 1220 | "Moonshot cannot force tool '{name}' because its input schema is incompatible with this route" |
| 1221 | ); |
| 1222 | } |
| 1223 | if body.get("tools").is_some() { |
| 1224 | body["tool_choice"] = mapped; |
| 1225 | } |
| 1226 | } |
| 1227 | apply_route_reasoning_controls( |
| 1228 | &mut body, |
| 1229 | provider, |
| 1230 | base_url, |
| 1231 | &model, |
| 1232 | request.reasoning_effort.as_deref(), |
| 1233 | ); |
| 1234 | apply_direct_moonshot_k3_fixed_sampling(&mut body, provider, base_url, &model); |
| 1235 | apply_kimi_code_fixed_sampling(&mut body, provider, base_url, &model); |
| 1236 | |
| 1237 | // Bulletproof final sanitizer: walk the wire payload and force |
| 1238 | // `reasoning_content` onto any assistant message that has tool_calls |
| 1239 | // but no reasoning_content. DeepSeek's thinking-mode API rejects |
| 1240 | // such messages with a 400. This is the last line of defense after |
| 1241 | // engine-side and build-side substitution; if either upstream path |
| 1242 | // misses a case (e.g. a session restored from disk, a sub-agent |
| 1243 | // adding messages directly, or a cached prefix mismatch), this pass |
| 1244 | // still produces a valid request. |
| 1245 | let replay_input_tokens = if stream { |
| 1246 | sanitize_thinking_mode_messages_for_route( |
| 1247 | &mut body, |
| 1248 | &model, |
| 1249 | request.reasoning_effort.as_deref(), |
| 1250 | provider, |
| 1251 | base_url, |
| 1252 | ) |
| 1253 | } else { |
| 1254 | None |
| 1255 | }; |
| 1256 | mirror_minimax_reasoning_details_for_body(&mut body, provider); |
| 1257 | |
| 1258 | Ok(ChatWireBody { |
| 1259 | body, |
| 1260 | model, |
| 1261 | replay_input_tokens, |
| 1262 | omitted_tool_names, |
| 1263 | }) |
| 1264 | } |
| 1265 | |
| 1266 | impl CodewhaleClient { |
| 1267 | pub(super) async fn create_message_chat( |
| 1268 | &self, |
| 1269 | prepared: &super::PreparedOutboundRequest, |
| 1270 | cacheable: bool, |
| 1271 | ) -> Result<MessageResponse> { |
| 1272 | let body = &prepared.body; |
| 1273 | |
| 1274 | let response_cache_key = if cacheable { |
| 1275 | let wire_body = |
| 1276 | serde_json::to_vec(&body).context("Failed to serialize Chat API cache key")?; |
| 1277 | let key = crate::llm_response_cache::ResponseCache::make_key( |
| 1278 | self.api_provider.as_str(), |
| 1279 | &self.base_url, |
| 1280 | self.path_suffix.as_deref(), |
| 1281 | &self.api_key, |
| 1282 | &wire_body, |
| 1283 | ); |
| 1284 | if let Some(cached) = crate::llm_response_cache::response_cache().get(&key) { |
| 1285 | return Ok(cached); |
| 1286 | } |
| 1287 | Some(key) |
| 1288 | } else { |
| 1289 | None |
| 1290 | }; |
| 1291 | |
| 1292 | // The endpoint was resolved by the shared seam alongside the body, so |
| 1293 | // a route-shape decision (e.g. DeepSeek's strict-tools `/beta` path) |
| 1294 | // cannot be made twice with two different answers. |
| 1295 | let url = prepared.endpoint.url.as_str(); |
| 1296 | let response = self.send_json_with_retry(url, body).await?; |
| 1297 | |
| 1298 | let status = response.status(); |
| 1299 | crate::client::record_provider_response(self.api_provider, status.as_u16()); |
| 1300 | if !status.is_success() { |
| 1301 | let raw_error_text = bounded_error_text(response, ERROR_BODY_MAX_BYTES).await; |
| 1302 | let error_text = sanitize_http_error_body( |
| 1303 | Some(self.api_provider.provider().display_name()), |
| 1304 | status.as_u16(), |
| 1305 | &raw_error_text, |
| 1306 | ); |
| 1307 | anyhow::bail!( |
| 1308 | "Failed to call {} Chat Completions API: HTTP {status}: {error_text}", |
| 1309 | self.api_provider.provider().display_name() |
| 1310 | ); |
| 1311 | } |
| 1312 | |
| 1313 | let response_text = response |
| 1314 | .text() |
| 1315 | .await |
| 1316 | .context("Failed to read Chat API response body")?; |
| 1317 | let value: Value = |
| 1318 | serde_json::from_str(&response_text).context("Failed to parse Chat API JSON")?; |
| 1319 | let parsed = parse_chat_message_for_route(&value, self.api_provider, &self.base_url)?; |
| 1320 | if let Some(key) = response_cache_key { |
| 1321 | crate::llm_response_cache::response_cache().put(key, parsed.clone()); |
| 1322 | } |
| 1323 | Ok(parsed) |
| 1324 | } |
| 1325 | } |
| 1326 | |
| 1327 | impl CodewhaleClient { |
| 1328 | async fn open_chat_stream_response( |
| 1329 | &self, |
| 1330 | url: &str, |
| 1331 | body: &Value, |
| 1332 | ) -> Result<(reqwest::Response, Duration)> { |
| 1333 | let open_req = self.stream_open_request(); |
| 1334 | let idle_timeout = open_req.idle_timeout; |
| 1335 | let response = super::stream_entry::open_sse_response(&open_req, |policy| async move { |
| 1336 | match policy { |
| 1337 | // The prebuilt HTTP/1.1 twin carries the same default |
| 1338 | // headers/auth; send once, without the JSON retry loop |
| 1339 | // (matching the pre-seam H1-pin behavior). |
| 1340 | super::stream_entry::StreamHttpPolicy::Http1Only => { |
| 1341 | let client = super::stream_entry::client_for_policy( |
| 1342 | &self.http_client, |
| 1343 | self.http1_fallback_client(), |
| 1344 | policy, |
| 1345 | ); |
| 1346 | Ok(client |
| 1347 | .post(url) |
| 1348 | .header(reqwest::header::CONTENT_TYPE, "application/json") |
| 1349 | .json(body) |
| 1350 | .send() |
| 1351 | .await?) |
| 1352 | } |
| 1353 | super::stream_entry::StreamHttpPolicy::DualWithH1Fallback => { |
| 1354 | // Stream open, not a JSON retry: the response body outlives |
| 1355 | // the open, so this path must not carry any total deadline |
| 1356 | // (`open_stream_json_with_retry`, not `send_json_with_retry`). |
| 1357 | self.open_stream_json_with_retry(url, body).await |
| 1358 | } |
| 1359 | } |
| 1360 | }) |
| 1361 | .await?; |
| 1362 | Ok((response, idle_timeout)) |
| 1363 | } |
| 1364 | |
| 1365 | pub(super) async fn handle_chat_completion_stream( |
| 1366 | &self, |
| 1367 | prepared: super::PreparedOutboundRequest, |
| 1368 | ) -> Result<StreamEventBox> { |
| 1369 | // Try true SSE streaming via chat completions (widely supported). |
| 1370 | // Body and endpoint both come from the shared prepared-request seam, |
| 1371 | // so a preview or a non-stream call can never diverge from the |
| 1372 | // streamed request. |
| 1373 | let super::PreparedOutboundRequest { |
| 1374 | body, |
| 1375 | wire_model: model, |
| 1376 | replay_input_tokens, |
| 1377 | endpoint, |
| 1378 | .. |
| 1379 | } = prepared; |
| 1380 | let url = endpoint.url; |
| 1381 | |
| 1382 | let (response, stream_idle_timeout) = self.open_chat_stream_response(&url, &body).await?; |
| 1383 | |
| 1384 | let status = response.status(); |
| 1385 | crate::client::record_provider_response(self.api_provider, status.as_u16()); |
| 1386 | if !status.is_success() { |
| 1387 | let raw_error_text = bounded_error_text(response, ERROR_BODY_MAX_BYTES).await; |
| 1388 | let error_text = sanitize_http_error_body( |
| 1389 | Some(self.api_provider.provider().display_name()), |
| 1390 | status.as_u16(), |
| 1391 | &raw_error_text, |
| 1392 | ); |
| 1393 | // If DeepSeek rejected for missing reasoning_content despite the |
| 1394 | // sanitizer, dump the offending indices so we can diagnose where |
| 1395 | // they came from on the next failure. |
| 1396 | if error_text.contains("reasoning_content") { |
| 1397 | log_thinking_mode_violations(&body); |
| 1398 | } |
| 1399 | anyhow::bail!("SSE stream request failed: HTTP {status}: {error_text}"); |
| 1400 | } |
| 1401 | |
| 1402 | let api_provider = self.api_provider; |
| 1403 | let base_url = self.base_url.clone(); |
| 1404 | |
| 1405 | // Capture transport-shape headers before we consume `response` into |
| 1406 | // `bytes_stream()`. They are surfaced in the decode-error log path so |
| 1407 | // we can tell HTTP/2 RST_STREAM from chunked-encoding corruption from |
| 1408 | // gzip-compressor failure when investigating #103. |
| 1409 | let response_headers = format_stream_headers(response.headers()); |
| 1410 | let byte_stream = response.bytes_stream(); |
| 1411 | let configured_reasoning_stream_style = self.reasoning_stream_style.clone(); |
| 1412 | |
| 1413 | let stream = async_stream::stream! { |
| 1414 | use futures_util::StreamExt; |
| 1415 | |
| 1416 | // Emit a synthetic MessageStart |
| 1417 | yield Ok(StreamEvent::MessageStart { |
| 1418 | message: MessageResponse { |
| 1419 | id: String::new(), |
| 1420 | r#type: "message".to_string(), |
| 1421 | role: "assistant".to_string(), |
| 1422 | content: Vec::new(), |
| 1423 | model: model.clone(), |
| 1424 | stop_reason: None, |
| 1425 | stop_sequence: None, |
| 1426 | container: None, |
| 1427 | usage: Usage { |
| 1428 | input_tokens: 0, |
| 1429 | output_tokens: 0, |
| 1430 | ..Usage::default() |
| 1431 | }, |
| 1432 | }, |
| 1433 | }); |
| 1434 | |
| 1435 | let mut line_buf = String::new(); |
| 1436 | let mut byte_buf = acquire_stream_buffer(); |
| 1437 | let mut content_index: u32 = 0; |
| 1438 | let mut text_started = false; |
| 1439 | let mut thinking_started = false; |
| 1440 | let mut tool_indices: std::collections::HashMap<u32, u32> = std::collections::HashMap::new(); |
| 1441 | let mut reasoning_detail_buffers: std::collections::HashMap<u32, String> = std::collections::HashMap::new(); |
| 1442 | let mut inline_reasoning_tags = InlineReasoningTagState::default(); |
| 1443 | let reasoning_stream_style = reasoning_stream_style_for_route( |
| 1444 | api_provider, |
| 1445 | &base_url, |
| 1446 | &model, |
| 1447 | configured_reasoning_stream_style.as_deref(), |
| 1448 | ); |
| 1449 | |
| 1450 | let mut byte_stream = std::pin::pin!(byte_stream); |
| 1451 | let idle = stream_idle_timeout; |
| 1452 | |
| 1453 | // Telemetry for #103 stream-decode diagnostics: bytes received |
| 1454 | // since the start of this stream and last successful event time. |
| 1455 | // Surfaces in the error log when reqwest yields a chunk error so |
| 1456 | // we can tell HTTP/2 RST_STREAM from chunk-decode-failure from |
| 1457 | // gzip-corruption when investigating a flaky session. |
| 1458 | let stream_start = std::time::Instant::now(); |
| 1459 | let mut last_event_at = std::time::Instant::now(); |
| 1460 | let mut bytes_received: usize = 0; |
| 1461 | // Set when a `[DONE]` sentinel was seen, so the post-loop flush does |
| 1462 | // not re-process trailing post-DONE bytes. |
| 1463 | let mut saw_done = false; |
| 1464 | // A number of OpenAI-compatible providers omit `[DONE]` but send a |
| 1465 | // terminal `finish_reason`. Either is valid terminal proof. A raw |
| 1466 | // HTTP EOF with neither is not: treating that as MessageStop turns |
| 1467 | // a truncated provider response into a successful empty turn. |
| 1468 | let mut saw_finish_reason = false; |
| 1469 | // Once an error has been emitted, do not follow it with a synthetic |
| 1470 | // MessageStop (or a second, less-specific premature-EOF error). |
| 1471 | let mut stream_failed = false; |
| 1472 | // Set when a complete line or unterminated flush failed UTF-8. |
| 1473 | // Skip further data-frame parsing so U+FFFD cannot enter the transcript. |
| 1474 | let mut decode_failed = false; |
| 1475 | |
| 1476 | let first_byte = super::stream_entry::first_byte_timeout(idle); |
| 1477 | 'stream: loop { |
| 1478 | let wait = super::stream_entry::next_chunk_timeout(idle, first_byte, bytes_received); |
| 1479 | let chunk_result = match tokio_timeout(wait, byte_stream.next()).await { |
| 1480 | Ok(Some(result)) => result, |
| 1481 | Ok(None) => break, // Stream ended normally |
| 1482 | Err(_elapsed) => { |
| 1483 | stream_failed = true; |
| 1484 | yield Err(anyhow::anyhow!(super::stream_entry::body_timeout_message( |
| 1485 | wait, |
| 1486 | bytes_received, |
| 1487 | stream_start.elapsed(), |
| 1488 | last_event_at.elapsed(), |
| 1489 | api_provider.provider().display_name(), |
| 1490 | ))); |
| 1491 | break; |
| 1492 | } |
| 1493 | }; |
| 1494 | let chunk = match chunk_result { |
| 1495 | Ok(bytes) => bytes, |
| 1496 | Err(e) => { |
| 1497 | stream_failed = true; |
| 1498 | // Walk the error source chain so reqwest's underlying |
| 1499 | // hyper / h2 / io error is visible — without this the |
| 1500 | // outer "error decoding response body" message tells |
| 1501 | // us nothing about WHY the stream died. |
| 1502 | let mut error_chain = format!("{e}"); |
| 1503 | let mut current: Option<&(dyn std::error::Error + 'static)> = |
| 1504 | std::error::Error::source(&e); |
| 1505 | while let Some(source) = current { |
| 1506 | error_chain.push_str(&format!(" -> {source}")); |
| 1507 | current = std::error::Error::source(source); |
| 1508 | } |
| 1509 | crate::logging::warn(format!( |
| 1510 | "Stream read error: {error_chain} \ |
| 1511 | (elapsed: {}ms, bytes_received: {}, ms_since_last_event: {}, headers: {})", |
| 1512 | stream_start.elapsed().as_millis(), |
| 1513 | bytes_received, |
| 1514 | last_event_at.elapsed().as_millis(), |
| 1515 | response_headers, |
| 1516 | )); |
| 1517 | yield Err(anyhow::anyhow!("Stream read error: {e}")); |
| 1518 | break; |
| 1519 | } |
| 1520 | }; |
| 1521 | |
| 1522 | bytes_received = bytes_received.saturating_add(chunk.len()); |
| 1523 | last_event_at = std::time::Instant::now(); |
| 1524 | byte_buf.extend_from_slice(&chunk); |
| 1525 | |
| 1526 | // Guard against unbounded buffer growth (e.g., malformed stream without newlines) |
| 1527 | const MAX_SSE_BUF: usize = 10 * 1024 * 1024; // 10 MB |
| 1528 | if byte_buf.len() > MAX_SSE_BUF { |
| 1529 | stream_failed = true; |
| 1530 | yield Err(anyhow::anyhow!("SSE buffer exceeded {MAX_SSE_BUF} bytes — aborting stream")); |
| 1531 | break; |
| 1532 | } |
| 1533 | |
| 1534 | if byte_buf.len() > SSE_BACKPRESSURE_HIGH_WATERMARK { |
| 1535 | tokio::time::sleep(Duration::from_millis(SSE_BACKPRESSURE_SLEEP_MS)).await; |
| 1536 | } |
| 1537 | |
| 1538 | // Process complete SSE lines from the buffer. Decode only after |
| 1539 | // a `\n` so an HTTP/2 DATA split mid-character cannot become |
| 1540 | // U+FFFD; genuine invalid bytes fail closed. |
| 1541 | let mut lines_processed = 0usize; |
| 1542 | loop { |
| 1543 | let line = match take_sse_line(&mut byte_buf) { |
| 1544 | Ok(Some(line)) => line, |
| 1545 | Ok(None) => break, |
| 1546 | Err(err) => { |
| 1547 | decode_failed = true; |
| 1548 | stream_failed = true; |
| 1549 | yield Err(anyhow::anyhow!("{err}")); |
| 1550 | break 'stream; |
| 1551 | } |
| 1552 | }; |
| 1553 | |
| 1554 | if line.is_empty() { |
| 1555 | // Empty line = event boundary, process accumulated data |
| 1556 | if !line_buf.is_empty() { |
| 1557 | let data = std::mem::take(&mut line_buf); |
| 1558 | match parse_sse_data_frame( |
| 1559 | &data, |
| 1560 | &mut content_index, |
| 1561 | &mut text_started, |
| 1562 | &mut thinking_started, |
| 1563 | &mut tool_indices, |
| 1564 | &mut reasoning_detail_buffers, |
| 1565 | &mut inline_reasoning_tags, |
| 1566 | reasoning_stream_style, |
| 1567 | ) { |
| 1568 | SseDataFrame::Done => { |
| 1569 | saw_done = true; |
| 1570 | break 'stream; |
| 1571 | } |
| 1572 | SseDataFrame::Events(events) => { |
| 1573 | for mut event in events { |
| 1574 | saw_finish_reason |= matches!( |
| 1575 | &event, |
| 1576 | StreamEvent::MessageDelta { delta, .. } |
| 1577 | if delta.stop_reason.as_deref().is_some_and(|reason| !reason.trim().is_empty()) |
| 1578 | ); |
| 1579 | // Stamp the client-side replay-token estimate |
| 1580 | // onto the final usage so the UI can surface |
| 1581 | // it (#30). We compute it pre-request and |
| 1582 | // overlay it on the server-reported usage at |
| 1583 | // stream completion. |
| 1584 | if let Some(tokens) = replay_input_tokens |
| 1585 | && let StreamEvent::MessageDelta { |
| 1586 | usage: Some(usage), |
| 1587 | .. |
| 1588 | } = &mut event |
| 1589 | { |
| 1590 | usage.reasoning_replay_tokens = Some(tokens); |
| 1591 | } |
| 1592 | yield Ok(event); |
| 1593 | } |
| 1594 | } |
| 1595 | } |
| 1596 | } |
| 1597 | continue; |
| 1598 | } |
| 1599 | |
| 1600 | if line.starts_with(':') { |
| 1601 | // SSE comment (`: keep-alive`, `: OPENROUTER PROCESSING`). |
| 1602 | // Surface it as a ping so the engine counts a provider |
| 1603 | // that is alive but queued/thinking as progress |
| 1604 | // (#6184) instead of timing out on a live stream. |
| 1605 | yield Ok(StreamEvent::Ping); |
| 1606 | continue; |
| 1607 | } |
| 1608 | |
| 1609 | if let Some(data) = extract_sse_data_value(&line) |
| 1610 | && let Err(err) = push_sse_event_data(&mut line_buf, data) |
| 1611 | { |
| 1612 | decode_failed = true; |
| 1613 | stream_failed = true; |
| 1614 | yield Err(anyhow::anyhow!("{err}")); |
| 1615 | break 'stream; |
| 1616 | } |
| 1617 | // Ignore other SSE fields (event:, id:, retry:) |
| 1618 | |
| 1619 | lines_processed = lines_processed.saturating_add(1); |
| 1620 | if lines_processed >= SSE_MAX_LINES_PER_CHUNK { |
| 1621 | // Backpressure relief: hand the executor a turn so a |
| 1622 | // slow consumer is not starved. Keep draining after |
| 1623 | // that — leaving complete lines buffered would strand |
| 1624 | // them, because the outer loop only resumes draining |
| 1625 | // once ANOTHER chunk arrives and the end-of-stream |
| 1626 | // flush treats the whole remainder as a single |
| 1627 | // unterminated line. |
| 1628 | lines_processed = 0; |
| 1629 | tokio::task::yield_now().await; |
| 1630 | } |
| 1631 | } |
| 1632 | } |
| 1633 | |
| 1634 | // Flush a final SSE frame that arrived without a terminating blank |
| 1635 | // line (the stream closed straight after the last `data:` line, or |
| 1636 | // that line lacked a trailing newline). Without this the final delta |
| 1637 | // — last tokens, finish_reason, and usage — is silently dropped. |
| 1638 | // Skipped after `[DONE]`, whose frame was already processed, and |
| 1639 | // after a fail-closed UTF-8 error. |
| 1640 | if !saw_done && !decode_failed { |
| 1641 | match flush_sse_line(&mut byte_buf) { |
| 1642 | Ok(Some(line)) => { |
| 1643 | if let Some(data) = extract_sse_data_value(&line) |
| 1644 | && let Err(err) = push_sse_event_data(&mut line_buf, data) |
| 1645 | { |
| 1646 | decode_failed = true; |
| 1647 | stream_failed = true; |
| 1648 | yield Err(anyhow::anyhow!("{err}")); |
| 1649 | } |
| 1650 | } |
| 1651 | Ok(None) => {} |
| 1652 | Err(err) => { |
| 1653 | decode_failed = true; |
| 1654 | stream_failed = true; |
| 1655 | yield Err(anyhow::anyhow!("{err}")); |
| 1656 | } |
| 1657 | } |
| 1658 | if !decode_failed && !line_buf.is_empty() { |
| 1659 | let data = std::mem::take(&mut line_buf); |
| 1660 | match parse_sse_data_frame( |
| 1661 | &data, |
| 1662 | &mut content_index, |
| 1663 | &mut text_started, |
| 1664 | &mut thinking_started, |
| 1665 | &mut tool_indices, |
| 1666 | &mut reasoning_detail_buffers, |
| 1667 | &mut inline_reasoning_tags, |
| 1668 | reasoning_stream_style, |
| 1669 | ) { |
| 1670 | SseDataFrame::Done => saw_done = true, |
| 1671 | SseDataFrame::Events(events) => { |
| 1672 | for mut event in events { |
| 1673 | saw_finish_reason |= matches!( |
| 1674 | &event, |
| 1675 | StreamEvent::MessageDelta { delta, .. } |
| 1676 | if delta.stop_reason.as_deref().is_some_and(|reason| !reason.trim().is_empty()) |
| 1677 | ); |
| 1678 | if let Some(tokens) = replay_input_tokens |
| 1679 | && let StreamEvent::MessageDelta { |
| 1680 | usage: Some(usage), .. |
| 1681 | } = &mut event |
| 1682 | { |
| 1683 | usage.reasoning_replay_tokens = Some(tokens); |
| 1684 | } |
| 1685 | yield Ok(event); |
| 1686 | } |
| 1687 | } |
| 1688 | } |
| 1689 | } |
| 1690 | } |
| 1691 | |
| 1692 | // Close any open blocks — content_index points to the |
| 1693 | // currently active open block (it is only incremented |
| 1694 | // *after* a block is closed, not when opened). |
| 1695 | if thinking_started || text_started { |
| 1696 | yield Ok(StreamEvent::ContentBlockStop { index: content_index }); |
| 1697 | } |
| 1698 | |
| 1699 | release_stream_buffer(byte_buf); |
| 1700 | if !stream_failed && (saw_done || saw_finish_reason) { |
| 1701 | yield Ok(StreamEvent::MessageStop); |
| 1702 | } else if !stream_failed { |
| 1703 | yield Err(anyhow::anyhow!( |
| 1704 | "Chat Completions stream closed before [DONE] or finish_reason" |
| 1705 | )); |
| 1706 | } |
| 1707 | }; |
| 1708 | |
| 1709 | Ok(Pin::from(Box::new(stream) |
| 1710 | as Box< |
| 1711 | dyn futures_util::Stream<Item = Result<StreamEvent>> + Send, |
| 1712 | >)) |
| 1713 | } |
| 1714 | } |
| 1715 | |
| 1716 | // === Chat Completions Helpers === |
| 1717 | |
| 1718 | #[cfg(test)] |
| 1719 | pub(super) fn build_chat_messages( |
| 1720 | system: Option<&SystemPrompt>, |
| 1721 | messages: &[Message], |
| 1722 | model: &str, |
| 1723 | ) -> Vec<Value> { |
| 1724 | build_chat_messages_with_reasoning( |
| 1725 | system, |
| 1726 | messages, |
| 1727 | tool_result_sent_char_budget(model), |
| 1728 | should_replay_reasoning_content(model, None), |
| 1729 | false, |
| 1730 | ) |
| 1731 | } |
| 1732 | |
| 1733 | #[cfg(test)] |
| 1734 | pub(super) fn build_chat_messages_for_request(request: &MessageRequest) -> Vec<Value> { |
| 1735 | PromptBuilder::for_request(request).build() |
| 1736 | } |
| 1737 | |
| 1738 | #[cfg(test)] |
| 1739 | pub(super) fn build_chat_messages_for_request_and_provider( |
| 1740 | request: &MessageRequest, |
| 1741 | provider: ProviderKind, |
| 1742 | ) -> Vec<Value> { |
| 1743 | build_chat_messages_for_request_and_provider_and_route(request, provider, "") |
| 1744 | } |
| 1745 | |
| 1746 | /// Build a wire prompt for one fully resolved provider route. |
| 1747 | /// |
| 1748 | /// Most provider behavior is keyed only by the provider kind and model. Kimi |
| 1749 | /// Code K3 is deliberately narrower: the bare `k3` model owns reasoning |
| 1750 | /// replay only on its official membership-plan endpoint, so callers that have |
| 1751 | /// a concrete base URL must retain it through prompt construction. |
| 1752 | #[cfg(test)] |
| 1753 | pub(super) fn build_chat_messages_for_request_and_provider_and_route( |
| 1754 | request: &MessageRequest, |
| 1755 | provider: ProviderKind, |
| 1756 | base_url: &str, |
| 1757 | ) -> Vec<Value> { |
| 1758 | PromptBuilder::for_request(request).build_for_provider_and_route(provider, base_url, None) |
| 1759 | } |
| 1760 | |
| 1761 | pub(crate) fn inspect_prompt_for_request(request: &MessageRequest) -> PromptInspection { |
| 1762 | PromptBuilder::for_request(request).inspect() |
| 1763 | } |
| 1764 | |
| 1765 | pub(crate) fn build_cache_warmup_request(request: &MessageRequest) -> MessageRequest { |
| 1766 | PromptBuilder::for_request(request).build_cache_warmup_request() |
| 1767 | } |
| 1768 | |
| 1769 | struct PromptBuilder<'a> { |
| 1770 | system: Option<&'a SystemPrompt>, |
| 1771 | messages: &'a [Message], |
| 1772 | tools: Option<&'a [Tool]>, |
| 1773 | model: &'a str, |
| 1774 | reasoning_effort: Option<&'a str>, |
| 1775 | } |
| 1776 | |
| 1777 | impl<'a> PromptBuilder<'a> { |
| 1778 | fn for_request(request: &'a MessageRequest) -> Self { |
| 1779 | Self { |
| 1780 | system: request.system.as_ref(), |
| 1781 | messages: &request.messages, |
| 1782 | tools: request.tools.as_deref(), |
| 1783 | model: &request.model, |
| 1784 | reasoning_effort: request.reasoning_effort.as_deref(), |
| 1785 | } |
| 1786 | } |
| 1787 | |
| 1788 | #[cfg(test)] |
| 1789 | fn build(self) -> Vec<Value> { |
| 1790 | build_chat_messages_with_reasoning( |
| 1791 | self.system, |
| 1792 | self.messages, |
| 1793 | tool_result_sent_char_budget(self.model), |
| 1794 | should_replay_reasoning_content(self.model, self.reasoning_effort), |
| 1795 | false, |
| 1796 | ) |
| 1797 | } |
| 1798 | |
| 1799 | fn build_for_provider_and_route( |
| 1800 | self, |
| 1801 | provider: ProviderKind, |
| 1802 | base_url: &str, |
| 1803 | route_limits: Option<RouteLimits>, |
| 1804 | ) -> Vec<Value> { |
| 1805 | let mut messages = build_chat_messages_with_reasoning( |
| 1806 | self.system, |
| 1807 | self.messages, |
| 1808 | crate::route_budget::route_inline_char_budget_for_route( |
| 1809 | provider, |
| 1810 | self.model, |
| 1811 | route_limits, |
| 1812 | ), |
| 1813 | should_replay_reasoning_content_for_provider_on_route( |
| 1814 | provider, |
| 1815 | base_url, |
| 1816 | self.model, |
| 1817 | self.reasoning_effort, |
| 1818 | ), |
| 1819 | false, |
| 1820 | ); |
| 1821 | dump_system_prompt_if_requested(&messages); |
| 1822 | if provider == ProviderKind::Arcee { |
| 1823 | apply_arcee_waf_safe_message_encoding(&mut messages); |
| 1824 | } |
| 1825 | if provider == ProviderKind::Minimax { |
| 1826 | mirror_minimax_reasoning_details_for_messages(&mut messages); |
| 1827 | } |
| 1828 | if is_exact_mistral_chat_route(provider, base_url) { |
| 1829 | reshape_mistral_messages_for_reasoning_replay(&mut messages); |
| 1830 | } |
| 1831 | if !is_google_openai_compat_chat_route(base_url) { |
| 1832 | // A signature captured on Google's endpoint is meaningless — and |
| 1833 | // potentially a leak — anywhere else, so it is stripped. Say so: |
| 1834 | // the model will behave differently on the replayed tool history, |
| 1835 | // and a silent strip is exactly what made this defect invisible. |
| 1836 | let stripped = strip_google_tool_call_extra_content(&mut messages); |
| 1837 | if stripped > 0 { |
| 1838 | tracing::warn!( |
| 1839 | provider = ?provider, |
| 1840 | stripped_tool_calls = stripped, |
| 1841 | "dropping captured Google thought signatures: this route is not Google's \ |
| 1842 | OpenAI-compatible endpoint, so the replayed tool history is unsigned" |
| 1843 | ); |
| 1844 | } |
| 1845 | } |
| 1846 | messages |
| 1847 | } |
| 1848 | |
| 1849 | fn inspect(self) -> PromptInspection { |
| 1850 | let messages = build_chat_messages_with_reasoning( |
| 1851 | self.system, |
| 1852 | self.messages, |
| 1853 | tool_result_sent_char_budget(self.model), |
| 1854 | should_replay_reasoning_content(self.model, self.reasoning_effort), |
| 1855 | true, |
| 1856 | ); |
| 1857 | inspect_wire_request(self.tools, &messages) |
| 1858 | } |
| 1859 | |
| 1860 | fn build_cache_warmup_request(self) -> MessageRequest { |
| 1861 | let system = stable_system_prompt(self.system); |
| 1862 | let mut messages = stable_history_messages(self.messages); |
| 1863 | let tools = self |
| 1864 | .tools |
| 1865 | .filter(|tools| !tools.is_empty()) |
| 1866 | .map(<[Tool]>::to_vec); |
| 1867 | let tool_choice = tools.as_ref().map(|_| json!("none")); |
| 1868 | messages.push(Message { |
| 1869 | role: Role::User, |
| 1870 | content: vec![ContentBlock::Text { |
| 1871 | text: CACHE_WARMUP_USER_TAIL.to_string(), |
| 1872 | cache_control: None, |
| 1873 | }], |
| 1874 | }); |
| 1875 | |
| 1876 | MessageRequest { |
| 1877 | model: self.model.to_string(), |
| 1878 | messages, |
| 1879 | max_tokens: CACHE_WARMUP_MAX_TOKENS, |
| 1880 | system, |
| 1881 | tools, |
| 1882 | tool_choice, |
| 1883 | metadata: None, |
| 1884 | thinking: None, |
| 1885 | // Warmup has an intentionally tiny answer contract ("OK"). Do not |
| 1886 | // let hidden reasoning consume that allowance before the cacheable |
| 1887 | // prefix is accepted by the provider. |
| 1888 | reasoning_effort: Some("off".to_string()), |
| 1889 | stream: None, |
| 1890 | temperature: None, |
| 1891 | top_p: None, |
| 1892 | } |
| 1893 | } |
| 1894 | } |
| 1895 | |
| 1896 | const SYSTEM_PROMPT_DUMP_ENV: &str = "CODEWHALE_DUMP_SYSTEM_PROMPT"; |
| 1897 | const SYSTEM_PROMPT_DUMP_BEGIN: &str = "<<<CODEWHALE_SYSTEM_PROMPT_BEGIN>>>"; |
| 1898 | const SYSTEM_PROMPT_DUMP_END: &str = "<<<CODEWHALE_SYSTEM_PROMPT_END>>>"; |
| 1899 | const ARCEE_WAF_TEXT_SPLIT_TRIGGERS: &[(&str, &str, &str)] = &[("python -c", "python ", "-c")]; |
| 1900 | |
| 1901 | fn dump_system_prompt_if_requested(messages: &[Value]) { |
| 1902 | let Ok(flag) = std::env::var(SYSTEM_PROMPT_DUMP_ENV) else { |
| 1903 | return; |
| 1904 | }; |
| 1905 | if !matches!(flag.trim(), "1" | "true" | "TRUE" | "yes" | "YES") { |
| 1906 | return; |
| 1907 | } |
| 1908 | let Some(prompt) = messages.iter().find_map(system_message_text) else { |
| 1909 | return; |
| 1910 | }; |
| 1911 | let mut stderr = std::io::stderr().lock(); |
| 1912 | let _ = writeln!(stderr, "{SYSTEM_PROMPT_DUMP_BEGIN}"); |
| 1913 | let _ = writeln!(stderr, "{prompt}"); |
| 1914 | let _ = writeln!(stderr, "{SYSTEM_PROMPT_DUMP_END}"); |
| 1915 | } |
| 1916 | |
| 1917 | fn system_message_text(message: &Value) -> Option<String> { |
| 1918 | if message.get("role").and_then(Value::as_str) != Some("system") { |
| 1919 | return None; |
| 1920 | } |
| 1921 | match message.get("content")? { |
| 1922 | Value::String(text) => Some(text.clone()), |
| 1923 | Value::Array(parts) => { |
| 1924 | let text = parts |
| 1925 | .iter() |
| 1926 | .filter_map(|part| part.get("text").and_then(Value::as_str)) |
| 1927 | .collect::<Vec<_>>() |
| 1928 | .join(""); |
| 1929 | (!text.is_empty()).then_some(text) |
| 1930 | } |
| 1931 | _ => None, |
| 1932 | } |
| 1933 | } |
| 1934 | |
| 1935 | fn apply_arcee_waf_safe_message_encoding(messages: &mut [Value]) { |
| 1936 | for message in messages { |
| 1937 | if message.get("role").and_then(Value::as_str) != Some("system") { |
| 1938 | continue; |
| 1939 | } |
| 1940 | let Some(content) = message.get("content").and_then(Value::as_str) else { |
| 1941 | continue; |
| 1942 | }; |
| 1943 | let Some(parts) = arcee_waf_safe_text_parts(content) else { |
| 1944 | continue; |
| 1945 | }; |
| 1946 | message["content"] = json!(parts); |
| 1947 | } |
| 1948 | } |
| 1949 | |
| 1950 | fn arcee_waf_safe_text_parts(content: &str) -> Option<Vec<Value>> { |
| 1951 | let mut parts = Vec::new(); |
| 1952 | let mut cursor = 0usize; |
| 1953 | let mut split_any = false; |
| 1954 | |
| 1955 | while cursor < content.len() { |
| 1956 | let Some((trigger_start, trigger, left, right)) = next_arcee_waf_trigger(content, cursor) |
| 1957 | else { |
| 1958 | push_text_part(&mut parts, &content[cursor..]); |
| 1959 | break; |
| 1960 | }; |
| 1961 | |
| 1962 | push_text_part(&mut parts, &content[cursor..trigger_start]); |
| 1963 | push_text_part(&mut parts, left); |
| 1964 | push_text_part(&mut parts, right); |
| 1965 | cursor = trigger_start + trigger.len(); |
| 1966 | split_any = true; |
| 1967 | } |
| 1968 | |
| 1969 | split_any.then_some(parts) |
| 1970 | } |
| 1971 | |
| 1972 | fn next_arcee_waf_trigger(content: &str, cursor: usize) -> Option<(usize, &str, &str, &str)> { |
| 1973 | ARCEE_WAF_TEXT_SPLIT_TRIGGERS |
| 1974 | .iter() |
| 1975 | .filter_map(|(trigger, left, right)| { |
| 1976 | content[cursor..] |
| 1977 | .find(trigger) |
| 1978 | .map(|offset| (cursor + offset, *trigger, *left, *right)) |
| 1979 | }) |
| 1980 | .min_by_key(|(start, _, _, _)| *start) |
| 1981 | } |
| 1982 | |
| 1983 | fn push_text_part(parts: &mut Vec<Value>, text: &str) { |
| 1984 | if !text.is_empty() { |
| 1985 | parts.push(json!({ |
| 1986 | "type": "text", |
| 1987 | "text": text, |
| 1988 | })); |
| 1989 | } |
| 1990 | } |
| 1991 | |
| 1992 | pub(crate) const CACHE_WARMUP_USER_TAIL: &str = "请只回复 OK"; |
| 1993 | pub(crate) const CACHE_WARMUP_MAX_TOKENS: u32 = 8; |
| 1994 | /// Wire backstop for tool results (#6508). The engine already fits every |
| 1995 | /// result it gives the model to the route's inline budget, and marks any cut |
| 1996 | /// with a recovery footer. This pass only catches history that never went |
| 1997 | /// through the engine (legacy or restored raw results). Model-only inspection |
| 1998 | /// uses the catalog window; outbound requests pass their resolved route budget. |
| 1999 | /// Results with an engine recovery footer remain intact. |
| 2000 | fn tool_result_sent_char_budget(model: &str) -> usize { |
| 2001 | crate::route_budget::route_inline_char_budget(codewhale_models::context_window_for_model(model)) |
| 2002 | } |
| 2003 | /// Characters of an excerpted wire result spent on its labelled header. |
| 2004 | const TOOL_RESULT_EXCERPT_FRAME_CHARS: usize = 1_024; |
| 2005 | /// Tool results shorter than this stay inline even when repeated. The |
| 2006 | /// extra prompt bytes are cheaper than adding an earlier-message reference |
| 2007 | /// for tiny command outputs. |
| 2008 | const TOOL_RESULT_DEDUP_MIN_CHARS: usize = 1_024; |
| 2009 | |
| 2010 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 2011 | pub(crate) struct PromptInspection { |
| 2012 | pub base_static_prefix_hash: String, |
| 2013 | pub full_request_prefix_hash: String, |
| 2014 | /// Hash of the rendered tool catalog JSON, or empty when no tools were supplied. |
| 2015 | pub tool_catalog_hash: String, |
| 2016 | pub layers: Vec<PromptLayerInspection>, |
| 2017 | } |
| 2018 | |
| 2019 | /// Identifies the stable prefix that a cache warmup primes. |
| 2020 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 2021 | pub(crate) struct CacheWarmupKey { |
| 2022 | pub provider: String, |
| 2023 | pub model: String, |
| 2024 | pub base_url: String, |
| 2025 | pub static_prefix_hash: String, |
| 2026 | pub tool_catalog_hash: String, |
| 2027 | pub project_pack_hash: String, |
| 2028 | pub skills_hash: String, |
| 2029 | } |
| 2030 | |
| 2031 | impl CacheWarmupKey { |
| 2032 | pub(crate) fn from_inspection( |
| 2033 | provider: &str, |
| 2034 | model: &str, |
| 2035 | base_url: &str, |
| 2036 | inspection: &PromptInspection, |
| 2037 | ) -> Self { |
| 2038 | Self { |
| 2039 | provider: provider.to_string(), |
| 2040 | model: model.to_string(), |
| 2041 | base_url: base_url.to_string(), |
| 2042 | static_prefix_hash: inspection.base_static_prefix_hash.clone(), |
| 2043 | tool_catalog_hash: inspection.tool_catalog_hash.clone(), |
| 2044 | project_pack_hash: layer_hash(inspection, "Project context pack"), |
| 2045 | skills_hash: layer_hash(inspection, "Skills"), |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | pub(crate) fn hash_short(&self) -> String { |
| 2050 | let json = serde_json::to_string(self).unwrap_or_default(); |
| 2051 | let hash = sha256_hex(json.as_bytes()); |
| 2052 | hash[..hash.len().min(12)].to_string() |
| 2053 | } |
| 2054 | } |
| 2055 | |
| 2056 | fn layer_hash(inspection: &PromptInspection, name: &str) -> String { |
| 2057 | inspection |
| 2058 | .layers |
| 2059 | .iter() |
| 2060 | .find(|layer| layer.name == name) |
| 2061 | .map(|layer| layer.sha256.clone()) |
| 2062 | .unwrap_or_default() |
| 2063 | } |
| 2064 | |
| 2065 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 2066 | pub(crate) struct PromptLayerInspection { |
| 2067 | pub name: String, |
| 2068 | pub stability: PromptLayerStability, |
| 2069 | pub char_len: usize, |
| 2070 | pub byte_len: usize, |
| 2071 | /// Rough token estimate for quick before/after cache-hit reports. |
| 2072 | pub token_estimate: usize, |
| 2073 | pub sha256: String, |
| 2074 | pub tool_result: Option<ToolResultInspection>, |
| 2075 | pub turn_meta: Option<TurnMetaInspection>, |
| 2076 | } |
| 2077 | |
| 2078 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 2079 | pub(crate) struct ToolResultInspection { |
| 2080 | pub original_chars: usize, |
| 2081 | pub sent_chars: usize, |
| 2082 | pub truncated: bool, |
| 2083 | pub deduplicated: bool, |
| 2084 | } |
| 2085 | |
| 2086 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 2087 | pub(crate) struct TurnMetaInspection { |
| 2088 | pub original_chars: usize, |
| 2089 | pub sent_chars: usize, |
| 2090 | pub deduplicated: bool, |
| 2091 | pub sha256: String, |
| 2092 | } |
| 2093 | |
| 2094 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 2095 | pub(crate) enum PromptLayerStability { |
| 2096 | Static, |
| 2097 | History, |
| 2098 | Dynamic, |
| 2099 | } |
| 2100 | |
| 2101 | #[cfg(test)] |
| 2102 | impl PromptLayerStability { |
| 2103 | pub(crate) fn label(self) -> &'static str { |
| 2104 | match self { |
| 2105 | Self::Static => "static", |
| 2106 | Self::History => "history", |
| 2107 | Self::Dynamic => "dynamic", |
| 2108 | } |
| 2109 | } |
| 2110 | } |
| 2111 | |
| 2112 | fn inspect_wire_request(tools: Option<&[Tool]>, messages: &[Value]) -> PromptInspection { |
| 2113 | let mut layers = Vec::new(); |
| 2114 | let mut base_static_prefix_parts = Vec::new(); |
| 2115 | let mut full_request_prefix_parts = Vec::new(); |
| 2116 | let mut tool_catalog_hash = String::new(); |
| 2117 | let mut start_index = 0; |
| 2118 | |
| 2119 | if let Some(message) = messages.first() { |
| 2120 | let role = message |
| 2121 | .get("role") |
| 2122 | .and_then(Value::as_str) |
| 2123 | .unwrap_or("unknown"); |
| 2124 | let content = message_content_for_inspect(message); |
| 2125 | if role == "system" { |
| 2126 | for (name, stability, body) in split_system_layers(&content) { |
| 2127 | if stability == PromptLayerStability::Static { |
| 2128 | base_static_prefix_parts.push(body.to_string()); |
| 2129 | } |
| 2130 | if stability != PromptLayerStability::Dynamic { |
| 2131 | full_request_prefix_parts.push(body.to_string()); |
| 2132 | } |
| 2133 | layers.push(prompt_layer(name, stability, body)); |
| 2134 | } |
| 2135 | start_index = 1; |
| 2136 | } |
| 2137 | } |
| 2138 | |
| 2139 | if let Some(tool_catalog) = tool_catalog_for_inspect(tools) { |
| 2140 | tool_catalog_hash = sha256_hex(tool_catalog.as_bytes()); |
| 2141 | base_static_prefix_parts.push(tool_catalog.clone()); |
| 2142 | full_request_prefix_parts.push(tool_catalog.clone()); |
| 2143 | layers.push(prompt_layer( |
| 2144 | "Tool catalog".to_string(), |
| 2145 | PromptLayerStability::Static, |
| 2146 | &tool_catalog, |
| 2147 | )); |
| 2148 | } |
| 2149 | |
| 2150 | for (index, message) in messages.iter().enumerate().skip(start_index) { |
| 2151 | let role = message |
| 2152 | .get("role") |
| 2153 | .and_then(Value::as_str) |
| 2154 | .unwrap_or("unknown"); |
| 2155 | let content = message_content_for_inspect(message); |
| 2156 | let is_last = index + 1 == messages.len(); |
| 2157 | let stability = if (is_last && role == "user") || role == "tool" { |
| 2158 | PromptLayerStability::Dynamic |
| 2159 | } else { |
| 2160 | PromptLayerStability::History |
| 2161 | }; |
| 2162 | let name = if is_last && role == "user" { |
| 2163 | "User task".to_string() |
| 2164 | } else { |
| 2165 | format!("Message #{index} {role}") |
| 2166 | }; |
| 2167 | if stability != PromptLayerStability::Dynamic { |
| 2168 | full_request_prefix_parts.push(content.clone()); |
| 2169 | } |
| 2170 | let mut layer = prompt_layer(name, stability, &content); |
| 2171 | layer.tool_result = tool_result_inspection_for_message(message); |
| 2172 | layer.turn_meta = turn_meta_inspection_for_message(message); |
| 2173 | layers.push(layer); |
| 2174 | } |
| 2175 | |
| 2176 | let base_static_prefix = base_static_prefix_parts.join("\n"); |
| 2177 | let full_request_prefix = full_request_prefix_parts.join("\n"); |
| 2178 | |
| 2179 | PromptInspection { |
| 2180 | base_static_prefix_hash: sha256_hex(base_static_prefix.as_bytes()), |
| 2181 | full_request_prefix_hash: sha256_hex(full_request_prefix.as_bytes()), |
| 2182 | tool_catalog_hash, |
| 2183 | layers, |
| 2184 | } |
| 2185 | } |
| 2186 | |
| 2187 | fn tool_catalog_for_inspect(tools: Option<&[Tool]>) -> Option<String> { |
| 2188 | let tools = tools.filter(|tools| !tools.is_empty())?; |
| 2189 | serde_json::to_string(&tools.iter().map(tool_to_chat).collect::<Vec<_>>()).ok() |
| 2190 | } |
| 2191 | |
| 2192 | fn message_content_for_inspect(message: &Value) -> String { |
| 2193 | let mut parts = Vec::new(); |
| 2194 | if let Some(content) = message.get("content").and_then(Value::as_str) |
| 2195 | && !content.is_empty() |
| 2196 | { |
| 2197 | parts.push(content.to_string()); |
| 2198 | } |
| 2199 | if let Some(content) = message.get("content").and_then(Value::as_array) { |
| 2200 | for part in content { |
| 2201 | match part.get("type").and_then(Value::as_str) { |
| 2202 | Some("text") => { |
| 2203 | if let Some(text) = part.get("text").and_then(Value::as_str) |
| 2204 | && !text.is_empty() |
| 2205 | { |
| 2206 | parts.push(text.to_string()); |
| 2207 | } |
| 2208 | } |
| 2209 | Some("image_url") => { |
| 2210 | let url = part |
| 2211 | .get("image_url") |
| 2212 | .and_then(|image_url| image_url.get("url")) |
| 2213 | .and_then(Value::as_str) |
| 2214 | .unwrap_or(""); |
| 2215 | parts.push(format!( |
| 2216 | "[image_url:{}]", |
| 2217 | summarize_image_url_for_inspect(url) |
| 2218 | )); |
| 2219 | } |
| 2220 | _ => {} |
| 2221 | } |
| 2222 | } |
| 2223 | } |
| 2224 | if let Some(reasoning) = message.get("reasoning_content").and_then(Value::as_str) |
| 2225 | && !reasoning.is_empty() |
| 2226 | { |
| 2227 | parts.push(reasoning.to_string()); |
| 2228 | } |
| 2229 | if let Some(tool_calls) = message.get("tool_calls") { |
| 2230 | parts.push(tool_calls.to_string()); |
| 2231 | } |
| 2232 | parts.join("\n") |
| 2233 | } |
| 2234 | |
| 2235 | fn summarize_image_url_for_inspect(url: &str) -> String { |
| 2236 | let Some((prefix, encoded)) = url.split_once(";base64,") else { |
| 2237 | return first_chars(url, 96); |
| 2238 | }; |
| 2239 | format!("{prefix};base64,<{} chars>", encoded.len()) |
| 2240 | } |
| 2241 | |
| 2242 | fn tool_result_inspection_for_message(message: &Value) -> Option<ToolResultInspection> { |
| 2243 | if message.get("role").and_then(Value::as_str) != Some("tool") { |
| 2244 | return None; |
| 2245 | } |
| 2246 | let budget = message.get("_tool_result_budget")?; |
| 2247 | Some(ToolResultInspection { |
| 2248 | original_chars: budget |
| 2249 | .get("original_chars") |
| 2250 | .and_then(Value::as_u64) |
| 2251 | .and_then(|n| usize::try_from(n).ok())?, |
| 2252 | sent_chars: budget |
| 2253 | .get("sent_chars") |
| 2254 | .and_then(Value::as_u64) |
| 2255 | .and_then(|n| usize::try_from(n).ok())?, |
| 2256 | truncated: budget |
| 2257 | .get("truncated") |
| 2258 | .and_then(Value::as_bool) |
| 2259 | .unwrap_or(false), |
| 2260 | deduplicated: budget |
| 2261 | .get("deduplicated") |
| 2262 | .and_then(Value::as_bool) |
| 2263 | .unwrap_or(false), |
| 2264 | }) |
| 2265 | } |
| 2266 | |
| 2267 | fn turn_meta_inspection_for_message(message: &Value) -> Option<TurnMetaInspection> { |
| 2268 | let budget = message.get("_turn_meta_budget")?; |
| 2269 | Some(TurnMetaInspection { |
| 2270 | original_chars: budget |
| 2271 | .get("original_chars") |
| 2272 | .and_then(Value::as_u64) |
| 2273 | .and_then(|n| usize::try_from(n).ok())?, |
| 2274 | sent_chars: budget |
| 2275 | .get("sent_chars") |
| 2276 | .and_then(Value::as_u64) |
| 2277 | .and_then(|n| usize::try_from(n).ok())?, |
| 2278 | deduplicated: budget |
| 2279 | .get("deduplicated") |
| 2280 | .and_then(Value::as_bool) |
| 2281 | .unwrap_or(false), |
| 2282 | sha256: budget |
| 2283 | .get("sha256") |
| 2284 | .and_then(Value::as_str) |
| 2285 | .map(str::to_string)?, |
| 2286 | }) |
| 2287 | } |
| 2288 | |
| 2289 | fn split_system_layers(content: &str) -> Vec<(String, PromptLayerStability, &str)> { |
| 2290 | let markers = [ |
| 2291 | ("Project context", "<project_instructions"), |
| 2292 | ("Project context pack", "## Project Context Pack"), |
| 2293 | ("Environment", "## Environment"), |
| 2294 | ("Configured instructions", "<instructions "), |
| 2295 | ("User memory", "## User Memory"), |
| 2296 | ("Current session goal", "## Current Session Goal"), |
| 2297 | ("Skills", "## Skills"), |
| 2298 | ("Core execution", "## Core Execution"), |
| 2299 | ("Compact template", "## Compact"), |
| 2300 | ("Previous session relay", "## Previous Session Relay"), |
| 2301 | ]; |
| 2302 | |
| 2303 | let mut starts: Vec<(usize, &str)> = markers |
| 2304 | .iter() |
| 2305 | .filter_map(|(name, marker)| content.find(marker).map(|idx| (idx, *name))) |
| 2306 | .collect(); |
| 2307 | starts.sort_by_key(|(idx, _)| *idx); |
| 2308 | |
| 2309 | let mut layers = Vec::new(); |
| 2310 | let first_marker = starts.first().map_or(content.len(), |(idx, _)| *idx); |
| 2311 | if first_marker > 0 { |
| 2312 | layers.push(( |
| 2313 | "Global system prefix".to_string(), |
| 2314 | PromptLayerStability::Static, |
| 2315 | content[..first_marker].trim(), |
| 2316 | )); |
| 2317 | } |
| 2318 | |
| 2319 | for (i, (start, name)) in starts.iter().enumerate() { |
| 2320 | let end = starts.get(i + 1).map_or(content.len(), |(idx, _)| *idx); |
| 2321 | let stability = if *name == "Previous session relay" { |
| 2322 | PromptLayerStability::Dynamic |
| 2323 | } else if is_static_base_layer(name) { |
| 2324 | PromptLayerStability::Static |
| 2325 | } else { |
| 2326 | PromptLayerStability::History |
| 2327 | }; |
| 2328 | layers.push(((*name).to_string(), stability, content[*start..end].trim())); |
| 2329 | } |
| 2330 | |
| 2331 | if layers.is_empty() { |
| 2332 | layers.push(( |
| 2333 | "Global system prefix".to_string(), |
| 2334 | PromptLayerStability::Static, |
| 2335 | content.trim(), |
| 2336 | )); |
| 2337 | } |
| 2338 | layers |
| 2339 | } |
| 2340 | |
| 2341 | fn is_static_base_layer(name: &str) -> bool { |
| 2342 | matches!( |
| 2343 | name, |
| 2344 | "Global system prefix" |
| 2345 | | "Environment" |
| 2346 | | "Skills" |
| 2347 | | "Project context" |
| 2348 | | "Project context pack" |
| 2349 | | "Core execution" |
| 2350 | | "Compact template" |
| 2351 | ) |
| 2352 | } |
| 2353 | |
| 2354 | fn stable_system_prompt(system: Option<&SystemPrompt>) -> Option<SystemPrompt> { |
| 2355 | let instructions = system_to_instructions(system.cloned())?; |
| 2356 | let stable = split_system_layers(&instructions) |
| 2357 | .into_iter() |
| 2358 | .filter_map(|(_, stability, body)| { |
| 2359 | (stability == PromptLayerStability::Static).then_some(body) |
| 2360 | }) |
| 2361 | .collect::<Vec<_>>() |
| 2362 | .join("\n\n"); |
| 2363 | if stable.trim().is_empty() { |
| 2364 | None |
| 2365 | } else { |
| 2366 | Some(SystemPrompt::Text(stable)) |
| 2367 | } |
| 2368 | } |
| 2369 | |
| 2370 | fn stable_history_messages(messages: &[Message]) -> Vec<Message> { |
| 2371 | let mut end = messages.len(); |
| 2372 | if messages |
| 2373 | .last() |
| 2374 | .is_some_and(|message| message.role.as_str() == "user") |
| 2375 | { |
| 2376 | end = end.saturating_sub(1); |
| 2377 | } |
| 2378 | messages[..end].to_vec() |
| 2379 | } |
| 2380 | |
| 2381 | fn prompt_layer( |
| 2382 | name: String, |
| 2383 | stability: PromptLayerStability, |
| 2384 | content: &str, |
| 2385 | ) -> PromptLayerInspection { |
| 2386 | let char_len = content.chars().count(); |
| 2387 | let token_estimate = if char_len == 0 { |
| 2388 | 0 |
| 2389 | } else if content.is_ascii() { |
| 2390 | (char_len / 4).max(1) |
| 2391 | } else { |
| 2392 | char_len.max(1) |
| 2393 | }; |
| 2394 | PromptLayerInspection { |
| 2395 | name, |
| 2396 | stability, |
| 2397 | char_len, |
| 2398 | byte_len: content.len(), |
| 2399 | token_estimate, |
| 2400 | sha256: sha256_hex(content.as_bytes()), |
| 2401 | tool_result: None, |
| 2402 | turn_meta: None, |
| 2403 | } |
| 2404 | } |
| 2405 | |
| 2406 | fn sha256_hex(bytes: &[u8]) -> String { |
| 2407 | crate::hashing::sha256_hex(bytes) |
| 2408 | } |
| 2409 | |
| 2410 | #[derive(Clone)] |
| 2411 | struct PendingToolCallInfo { |
| 2412 | tool_name: String, |
| 2413 | input: Value, |
| 2414 | } |
| 2415 | |
| 2416 | struct SeenToolResult { |
| 2417 | message_label: String, |
| 2418 | original_chars: usize, |
| 2419 | } |
| 2420 | |
| 2421 | struct WireToolResult { |
| 2422 | content: String, |
| 2423 | original_chars: usize, |
| 2424 | sent_chars: usize, |
| 2425 | truncated: bool, |
| 2426 | deduplicated: bool, |
| 2427 | } |
| 2428 | |
| 2429 | #[derive(Clone)] |
| 2430 | struct TurnMetaBudget { |
| 2431 | original_chars: usize, |
| 2432 | sent_chars: usize, |
| 2433 | deduplicated: bool, |
| 2434 | sha256: String, |
| 2435 | } |
| 2436 | |
| 2437 | struct LastFullTurnMeta { |
| 2438 | sha256: String, |
| 2439 | } |
| 2440 | |
| 2441 | fn render_turn_meta_for_wire( |
| 2442 | text: &str, |
| 2443 | last_full_turn_meta: &mut Option<LastFullTurnMeta>, |
| 2444 | ) -> (String, TurnMetaBudget) { |
| 2445 | let original_chars = text.chars().count(); |
| 2446 | let sha = sha256_hex(text.as_bytes()); |
| 2447 | |
| 2448 | if last_full_turn_meta |
| 2449 | .as_ref() |
| 2450 | .is_some_and(|previous| previous.sha256 == sha) |
| 2451 | { |
| 2452 | // Keep the repeated metadata slot short without surfacing an |
| 2453 | // opaque hash the model cannot resolve. |
| 2454 | let rendered = "<turn_meta_unchanged />".to_string(); |
| 2455 | let budget = TurnMetaBudget { |
| 2456 | original_chars, |
| 2457 | sent_chars: rendered.chars().count(), |
| 2458 | deduplicated: true, |
| 2459 | sha256: sha, |
| 2460 | }; |
| 2461 | return (rendered, budget); |
| 2462 | } |
| 2463 | |
| 2464 | *last_full_turn_meta = Some(LastFullTurnMeta { |
| 2465 | sha256: sha.clone(), |
| 2466 | }); |
| 2467 | ( |
| 2468 | text.to_string(), |
| 2469 | TurnMetaBudget { |
| 2470 | original_chars, |
| 2471 | sent_chars: original_chars, |
| 2472 | deduplicated: false, |
| 2473 | sha256: sha, |
| 2474 | }, |
| 2475 | ) |
| 2476 | } |
| 2477 | |
| 2478 | fn is_turn_meta_text(text: &str) -> bool { |
| 2479 | text.trim_start().starts_with("<turn_meta>") |
| 2480 | } |
| 2481 | |
| 2482 | fn turn_meta_budget_json(turn_meta: &TurnMetaBudget) -> Value { |
| 2483 | json!({ |
| 2484 | "original_chars": turn_meta.original_chars, |
| 2485 | "sent_chars": turn_meta.sent_chars, |
| 2486 | "deduplicated": turn_meta.deduplicated, |
| 2487 | "sha256": turn_meta.sha256, |
| 2488 | }) |
| 2489 | } |
| 2490 | |
| 2491 | /// Mutating/write tools whose result body is a *confirmation* (it embeds |
| 2492 | /// the unified diff + summary of what was just written), not retrievable |
| 2493 | /// reference data. Two identical large `write_file` calls must each keep |
| 2494 | /// their full confirmation inline: collapsing the later one to a |
| 2495 | /// `<TOOL_RESULT_REF sha="..." />` makes the model lose the write-success |
| 2496 | /// context and behave as if the file is missing (issue #1695). Read-style |
| 2497 | /// tools (`read_file`, `grep_files`, `exec_shell`, …) may deduplicate medium |
| 2498 | /// outputs by pointing at an earlier full message in the same request. They |
| 2499 | /// never advertise a process-wide SHA as retrievable: that store cannot prove |
| 2500 | /// session ownership. |
| 2501 | fn is_mutation_tool(tool_name: &str) -> bool { |
| 2502 | matches!( |
| 2503 | tool_name, |
| 2504 | "write" | "edit" | "write_file" | "edit_file" | "apply_patch" |
| 2505 | ) |
| 2506 | } |
| 2507 | |
| 2508 | fn compact_tool_result_for_wire( |
| 2509 | tool_name: &str, |
| 2510 | input: &Value, |
| 2511 | content: &str, |
| 2512 | message_label: &str, |
| 2513 | sent_budget: usize, |
| 2514 | seen_tool_results: &mut HashMap<String, SeenToolResult>, |
| 2515 | ) -> WireToolResult { |
| 2516 | let original_chars = content.chars().count(); |
| 2517 | let sha = sha256_hex(content.as_bytes()); |
| 2518 | |
| 2519 | // Only medium, non-mutation results can point back to a full earlier |
| 2520 | // message in this one request. Oversized results are already excerpts, so |
| 2521 | // a back-reference would falsely imply the exact bytes remain available. |
| 2522 | let dedup_eligible = (TOOL_RESULT_DEDUP_MIN_CHARS..=sent_budget).contains(&original_chars) |
| 2523 | && !is_mutation_tool(tool_name); |
| 2524 | |
| 2525 | if dedup_eligible && let Some(previous) = seen_tool_results.get(&sha) { |
| 2526 | let content = format!( |
| 2527 | "<TOOL_RESULT_REF sha=\"{sha}\" original_message=\"{label}\" chars=\"{chars}\">\n\ |
| 2528 | source: full content appears in {label} earlier in this request\n\ |
| 2529 | </TOOL_RESULT_REF>", |
| 2530 | label = previous.message_label, |
| 2531 | chars = previous.original_chars, |
| 2532 | ); |
| 2533 | return WireToolResult { |
| 2534 | sent_chars: content.chars().count(), |
| 2535 | content, |
| 2536 | original_chars, |
| 2537 | truncated: false, |
| 2538 | deduplicated: true, |
| 2539 | }; |
| 2540 | } |
| 2541 | |
| 2542 | if dedup_eligible { |
| 2543 | seen_tool_results.insert( |
| 2544 | sha.clone(), |
| 2545 | SeenToolResult { |
| 2546 | message_label: message_label.to_string(), |
| 2547 | original_chars, |
| 2548 | }, |
| 2549 | ); |
| 2550 | } |
| 2551 | |
| 2552 | if original_chars <= sent_budget { |
| 2553 | return WireToolResult { |
| 2554 | content: content.to_string(), |
| 2555 | original_chars, |
| 2556 | sent_chars: original_chars, |
| 2557 | truncated: false, |
| 2558 | deduplicated: false, |
| 2559 | }; |
| 2560 | } |
| 2561 | |
| 2562 | // Content already bounded by the adaptive evidence envelope carries its |
| 2563 | // own honest footer: the omitted count, the on-disk artifact path, and a |
| 2564 | // recovery instruction. Truncating it again here would destroy that |
| 2565 | // recovery contract and falsely report that no session-owned artifact |
| 2566 | // was recorded, so pass it through untouched. |
| 2567 | if content.contains(crate::tools::truncate::SPILLOVER_RECOVERY_HINT) { |
| 2568 | return WireToolResult { |
| 2569 | content: content.to_string(), |
| 2570 | original_chars, |
| 2571 | sent_chars: original_chars, |
| 2572 | truncated: false, |
| 2573 | deduplicated: false, |
| 2574 | }; |
| 2575 | } |
| 2576 | |
| 2577 | let excerpt_chars = sent_budget.saturating_sub(TOOL_RESULT_EXCERPT_FRAME_CHARS); |
| 2578 | let head_chars = excerpt_chars * 2 / 3; |
| 2579 | let head = first_chars(content, head_chars); |
| 2580 | let tail = last_chars(content, excerpt_chars - head_chars); |
| 2581 | let kept = head.chars().count() + tail.chars().count(); |
| 2582 | let omitted = original_chars.saturating_sub(kept); |
| 2583 | let compacted = format!( |
| 2584 | "[TOOL_RESULT_TRUNCATED]\n\ |
| 2585 | tool_name: {tool_name}\n\ |
| 2586 | command_or_query: {}\n\ |
| 2587 | exit_status: {}\n\ |
| 2588 | original_chars: {original_chars}\n\ |
| 2589 | sha256: {sha}\n\ |
| 2590 | exact_detail: unavailable; no session-owned artifact was recorded\n\ |
| 2591 | first_chars:\n\ |
| 2592 | {head}\n\n\ |
| 2593 | [... truncated {omitted} chars from middle ...]\n\n\ |
| 2594 | last_chars:\n\ |
| 2595 | {tail}", |
| 2596 | tool_command_or_query(input), |
| 2597 | tool_exit_status(content) |
| 2598 | ); |
| 2599 | |
| 2600 | WireToolResult { |
| 2601 | sent_chars: compacted.chars().count(), |
| 2602 | content: compacted, |
| 2603 | original_chars, |
| 2604 | truncated: true, |
| 2605 | deduplicated: false, |
| 2606 | } |
| 2607 | } |
| 2608 | |
| 2609 | fn tool_command_or_query(input: &Value) -> String { |
| 2610 | for key in ["command", "cmd", "query", "q", "pattern", "path", "url"] { |
| 2611 | if let Some(value) = input.get(key) { |
| 2612 | return summarize_for_metadata(value, 500); |
| 2613 | } |
| 2614 | } |
| 2615 | summarize_for_metadata(input, 500) |
| 2616 | } |
| 2617 | |
| 2618 | fn tool_exit_status(content: &str) -> String { |
| 2619 | if let Ok(value) = serde_json::from_str::<Value>(content) { |
| 2620 | for key in ["exit_code", "exit_status", "status", "code"] { |
| 2621 | if let Some(value) = value.get(key) { |
| 2622 | return summarize_for_metadata(value, 120); |
| 2623 | } |
| 2624 | } |
| 2625 | } |
| 2626 | |
| 2627 | for line in content.lines().take(20) { |
| 2628 | let trimmed = line.trim(); |
| 2629 | for prefix in ["Exit code:", "exit code:", "Exit status:", "exit status:"] { |
| 2630 | if let Some(value) = trimmed.strip_prefix(prefix) { |
| 2631 | return value.trim().to_string(); |
| 2632 | } |
| 2633 | } |
| 2634 | } |
| 2635 | "unknown".to_string() |
| 2636 | } |
| 2637 | |
| 2638 | fn summarize_for_metadata(value: &Value, max_chars: usize) -> String { |
| 2639 | let raw = value |
| 2640 | .as_str() |
| 2641 | .map(str::to_string) |
| 2642 | .unwrap_or_else(|| value.to_string()); |
| 2643 | let mut summarized = first_chars(&raw.replace('\n', "\\n"), max_chars); |
| 2644 | if raw.chars().count() > max_chars { |
| 2645 | summarized.push_str("..."); |
| 2646 | } |
| 2647 | summarized |
| 2648 | } |
| 2649 | |
| 2650 | fn first_chars(value: &str, count: usize) -> String { |
| 2651 | value.chars().take(count).collect() |
| 2652 | } |
| 2653 | |
| 2654 | fn last_chars(value: &str, count: usize) -> String { |
| 2655 | let mut chars: Vec<char> = value.chars().rev().take(count).collect(); |
| 2656 | chars.reverse(); |
| 2657 | chars.into_iter().collect() |
| 2658 | } |
| 2659 | |
| 2660 | fn merge_adjacent_user_content(previous: Value, current: Value) -> Value { |
| 2661 | match (previous, current) { |
| 2662 | (Value::String(left), Value::String(right)) => json!(format!("{left}\n\n{right}")), |
| 2663 | (left, right) => { |
| 2664 | let mut parts = Vec::new(); |
| 2665 | for content in [left, right] { |
| 2666 | match content { |
| 2667 | Value::Array(items) => parts.extend(items), |
| 2668 | Value::String(text) => parts.push(json!({"type": "text", "text": text})), |
| 2669 | other => parts.push(other), |
| 2670 | } |
| 2671 | } |
| 2672 | Value::Array(parts) |
| 2673 | } |
| 2674 | } |
| 2675 | } |
| 2676 | |
| 2677 | fn build_chat_messages_with_reasoning( |
| 2678 | system: Option<&SystemPrompt>, |
| 2679 | messages: &[Message], |
| 2680 | tool_result_budget: usize, |
| 2681 | include_reasoning: bool, |
| 2682 | include_tool_budget_metadata: bool, |
| 2683 | ) -> Vec<Value> { |
| 2684 | let mut out = Vec::new(); |
| 2685 | let mut pending_tool_calls: HashMap<String, PendingToolCallInfo> = HashMap::new(); |
| 2686 | // Chat Completions requires every result for one assistant tool-call batch |
| 2687 | // to be contiguous. Keep tool-result media aside until the complete batch |
| 2688 | // has been emitted as `role: tool` messages. |
| 2689 | let mut deferred_tool_result_images = Vec::new(); |
| 2690 | let mut seen_tool_results: HashMap<String, SeenToolResult> = HashMap::new(); |
| 2691 | let mut last_full_turn_meta: Option<LastFullTurnMeta> = None; |
| 2692 | |
| 2693 | if let Some(instructions) = system_to_instructions(system.cloned()) |
| 2694 | && !instructions.trim().is_empty() |
| 2695 | { |
| 2696 | out.push(json!({ |
| 2697 | "role": "system", |
| 2698 | "content": instructions, |
| 2699 | })); |
| 2700 | } |
| 2701 | |
| 2702 | // Persisted compaction keeps its summary after the bounded last round. |
| 2703 | // On strict paired chat templates a user message after a tool result is |
| 2704 | // invalid. Reorder only a generated summary immediately after a tool |
| 2705 | // result; its independent provenance block rules out quoted user text. |
| 2706 | // The session log retains every original message and tool ID. |
| 2707 | // Limitation: this normalization applies to Chat Completions only. |
| 2708 | let summary_index = messages |
| 2709 | .iter() |
| 2710 | .enumerate() |
| 2711 | .rev() |
| 2712 | .find_map(|(index, message)| { |
| 2713 | (index > 0 |
| 2714 | && crate::compaction::is_wire_compaction_checkpoint_message(message) |
| 2715 | && messages[index - 1] |
| 2716 | .content |
| 2717 | .iter() |
| 2718 | .any(|block| matches!(block, ContentBlock::ToolResult { .. }))) |
| 2719 | .then_some(index) |
| 2720 | }); |
| 2721 | let summary_target = summary_index.and_then(|summary_index| { |
| 2722 | messages[..summary_index] |
| 2723 | .iter() |
| 2724 | .rposition(|message| { |
| 2725 | crate::runtime_handoff::classify_user_turn_prompt(message) |
| 2726 | != crate::runtime_handoff::UserTurnPromptKind::NotPrompt |
| 2727 | }) |
| 2728 | .or_else(|| { |
| 2729 | messages[..summary_index] |
| 2730 | .iter() |
| 2731 | .position(|message| message.role.is_assistant_like()) |
| 2732 | }) |
| 2733 | }); |
| 2734 | let wire_messages = (0..messages.len()) |
| 2735 | .filter(|index| Some(*index) != summary_index || summary_target.is_none()) |
| 2736 | .flat_map(|index| { |
| 2737 | if Some(index) == summary_target { |
| 2738 | [summary_index, Some(index)] |
| 2739 | .into_iter() |
| 2740 | .flatten() |
| 2741 | .collect::<Vec<_>>() |
| 2742 | } else { |
| 2743 | vec![index] |
| 2744 | } |
| 2745 | }); |
| 2746 | |
| 2747 | for message_index in wire_messages { |
| 2748 | let message = &messages[message_index]; |
| 2749 | // Which wire channel this message belongs in is decided by the shared |
| 2750 | // placement table, not by an `if` chain local to this adapter. |
| 2751 | let placement = role_placement(&message.role, WireDialect::ChatCompletions); |
| 2752 | let mut text_parts = Vec::new(); |
| 2753 | let mut image_parts = Vec::new(); |
| 2754 | let mut thinking_parts = Vec::new(); |
| 2755 | let mut tool_calls = Vec::new(); |
| 2756 | let mut tool_call_infos = Vec::new(); |
| 2757 | let mut tool_results: Vec<(String, String, String, Vec<Value>)> = Vec::new(); |
| 2758 | let mut turn_meta_budget: Option<TurnMetaBudget> = None; |
| 2759 | |
| 2760 | for block in &message.content { |
| 2761 | match block { |
| 2762 | ContentBlock::Text { text, .. } => { |
| 2763 | if is_turn_meta_text(text) { |
| 2764 | let (rendered, budget) = |
| 2765 | render_turn_meta_for_wire(text, &mut last_full_turn_meta); |
| 2766 | text_parts.push(rendered); |
| 2767 | turn_meta_budget = Some(budget); |
| 2768 | } else { |
| 2769 | text_parts.push(text.clone()); |
| 2770 | } |
| 2771 | } |
| 2772 | ContentBlock::ImageUrl { image_url } => { |
| 2773 | image_parts.push(json!({ |
| 2774 | "type": "image_url", |
| 2775 | "image_url": { |
| 2776 | "url": image_url.url.clone(), |
| 2777 | }, |
| 2778 | })); |
| 2779 | } |
| 2780 | ContentBlock::Thinking { thinking, .. } => thinking_parts.push(thinking.clone()), |
| 2781 | ContentBlock::ToolUse { |
| 2782 | id, |
| 2783 | name, |
| 2784 | input, |
| 2785 | caller, |
| 2786 | thought_signature, |
| 2787 | .. |
| 2788 | } => { |
| 2789 | let args = serde_json::to_string(input).unwrap_or_else(|_| input.to_string()); |
| 2790 | let mut call = json!({ |
| 2791 | "id": id, |
| 2792 | "type": "function", |
| 2793 | "function": { |
| 2794 | "name": to_api_tool_name(name), |
| 2795 | "arguments": args, |
| 2796 | } |
| 2797 | }); |
| 2798 | if let Some(signature) = thought_signature { |
| 2799 | call["extra_content"]["google"]["thought_signature"] = json!(signature); |
| 2800 | } |
| 2801 | if let Some(caller) = caller { |
| 2802 | call["caller"] = json!({ |
| 2803 | "type": caller.caller_type, |
| 2804 | "tool_id": caller.tool_id, |
| 2805 | }); |
| 2806 | } |
| 2807 | tool_calls.push(call); |
| 2808 | tool_call_infos.push(( |
| 2809 | id.clone(), |
| 2810 | PendingToolCallInfo { |
| 2811 | tool_name: name.clone(), |
| 2812 | input: input.clone(), |
| 2813 | }, |
| 2814 | )); |
| 2815 | } |
| 2816 | ContentBlock::ToolResult { |
| 2817 | tool_use_id, |
| 2818 | content, |
| 2819 | content_blocks, |
| 2820 | .. |
| 2821 | } => { |
| 2822 | let message_label = format!("Message #{message_index}"); |
| 2823 | tool_results.push(( |
| 2824 | tool_use_id.clone(), |
| 2825 | content.clone(), |
| 2826 | message_label, |
| 2827 | content_blocks.clone().unwrap_or_default(), |
| 2828 | )); |
| 2829 | } |
| 2830 | ContentBlock::ServerToolUse { .. } |
| 2831 | | ContentBlock::ToolSearchToolResult { .. } |
| 2832 | | ContentBlock::CodeExecutionToolResult { .. } => {} |
| 2833 | } |
| 2834 | } |
| 2835 | |
| 2836 | let out_len_before_role_projection = out.len(); |
| 2837 | if placement.is_assistant_channel() { |
| 2838 | let content = if placement == RolePlacement::InterruptedAssistant { |
| 2839 | format!( |
| 2840 | "{}{}", |
| 2841 | codewhale_models::INTERRUPTED_ASSISTANT_CONTEXT_PREFIX, |
| 2842 | text_parts.join("\n") |
| 2843 | ) |
| 2844 | } else { |
| 2845 | text_parts.join("\n") |
| 2846 | }; |
| 2847 | let mut reasoning_content = thinking_parts.join("\n"); |
| 2848 | let has_text = !content.trim().is_empty(); |
| 2849 | let has_tool_calls = !tool_calls.is_empty(); |
| 2850 | // Reasoning replay must be a function of the stored message ONLY, |
| 2851 | // never of later history. DeepSeek's prefix cache hashes the raw |
| 2852 | // bytes of every message; flipping `reasoning_content` on/off |
| 2853 | // depending on whether a follow-up user turn exists rewrites a |
| 2854 | // historical message between turns and busts the cache from that |
| 2855 | // point onwards. Always emit `reasoning_content` when the model |
| 2856 | // requires replay AND the stored message carries thinking text. |
| 2857 | // Tool-call messages with empty thinking still need a placeholder |
| 2858 | // (DeepSeek 400s without it), but text-only assistant messages |
| 2859 | // simply omit the field when there's nothing to replay. |
| 2860 | let mut has_reasoning = include_reasoning && !reasoning_content.trim().is_empty(); |
| 2861 | if include_reasoning && has_tool_calls && !has_reasoning { |
| 2862 | logging::warn( |
| 2863 | "Substituting placeholder reasoning_content for DeepSeek tool-call assistant message", |
| 2864 | ); |
| 2865 | reasoning_content = String::from(REASONING_REPLAY_PLACEHOLDER); |
| 2866 | has_reasoning = true; |
| 2867 | } |
| 2868 | |
| 2869 | // DeepSeek rejects assistant messages where both `content` and |
| 2870 | // `tool_calls` are missing/null. Skip such entries even if they |
| 2871 | // carry reasoning-only metadata unless we can send a non-null |
| 2872 | // placeholder content field. |
| 2873 | if !has_text && !has_tool_calls && !has_reasoning { |
| 2874 | pending_tool_calls.clear(); |
| 2875 | deferred_tool_result_images.clear(); |
| 2876 | continue; |
| 2877 | } |
| 2878 | |
| 2879 | let mut msg = json!({ |
| 2880 | "role": "assistant", |
| 2881 | "content": if has_text { |
| 2882 | json!(content) |
| 2883 | } else if has_reasoning { |
| 2884 | json!("") |
| 2885 | } else { |
| 2886 | Value::Null |
| 2887 | }, |
| 2888 | }); |
| 2889 | if has_reasoning { |
| 2890 | msg["reasoning_content"] = json!(reasoning_content); |
| 2891 | } |
| 2892 | if has_tool_calls { |
| 2893 | msg["tool_calls"] = json!(tool_calls); |
| 2894 | let expected_tool_result_count = tool_call_infos.len(); |
| 2895 | pending_tool_calls = tool_call_infos.into_iter().collect(); |
| 2896 | deferred_tool_result_images.clear(); |
| 2897 | if pending_tool_calls.len() != expected_tool_result_count { |
| 2898 | logging::warn( |
| 2899 | "Rejecting assistant tool-call batch with duplicate tool_call IDs", |
| 2900 | ); |
| 2901 | pending_tool_calls.clear(); |
| 2902 | } |
| 2903 | } else { |
| 2904 | pending_tool_calls.clear(); |
| 2905 | deferred_tool_result_images.clear(); |
| 2906 | } |
| 2907 | out.push(msg); |
| 2908 | } else if matches!(placement, RolePlacement::System | RolePlacement::Developer) { |
| 2909 | let content = text_parts.join("\n"); |
| 2910 | if !content.trim().is_empty() { |
| 2911 | let mut msg = json!({ |
| 2912 | "role": if placement == RolePlacement::Developer { |
| 2913 | "developer" |
| 2914 | } else { |
| 2915 | "system" |
| 2916 | }, |
| 2917 | "content": content, |
| 2918 | }); |
| 2919 | if include_tool_budget_metadata && let Some(turn_meta) = &turn_meta_budget { |
| 2920 | msg["_turn_meta_budget"] = turn_meta_budget_json(turn_meta); |
| 2921 | } |
| 2922 | out.push(msg); |
| 2923 | } |
| 2924 | } else if placement == RolePlacement::User { |
| 2925 | let content = text_parts.join("\n"); |
| 2926 | let has_text = !content.trim().is_empty(); |
| 2927 | let has_images = !image_parts.is_empty(); |
| 2928 | if has_text || has_images { |
| 2929 | let wire_content = if has_images { |
| 2930 | let mut parts = Vec::new(); |
| 2931 | if has_text { |
| 2932 | parts.push(json!({ |
| 2933 | "type": "text", |
| 2934 | "text": content, |
| 2935 | })); |
| 2936 | } |
| 2937 | parts.extend(image_parts); |
| 2938 | json!(parts) |
| 2939 | } else { |
| 2940 | json!(content) |
| 2941 | }; |
| 2942 | let mut msg = json!({ |
| 2943 | "role": "user", |
| 2944 | "content": wire_content, |
| 2945 | }); |
| 2946 | if include_tool_budget_metadata && let Some(turn_meta) = &turn_meta_budget { |
| 2947 | msg["_turn_meta_budget"] = turn_meta_budget_json(turn_meta); |
| 2948 | } |
| 2949 | if (Some(message_index) == summary_index |
| 2950 | || Some(message_index) == summary_target |
| 2951 | || crate::compaction::is_wire_compaction_checkpoint_message(message) |
| 2952 | || crate::runtime_handoff::is_agent_topology_checkpoint(message) |
| 2953 | || crate::runtime_handoff::is_restored_agent_topology_checkpoint(message)) |
| 2954 | && let Some(previous) = out.last_mut() |
| 2955 | && previous.get("role").and_then(Value::as_str) == Some("user") |
| 2956 | { |
| 2957 | let previous_content = previous["content"].take(); |
| 2958 | let current_content = msg["content"].take(); |
| 2959 | previous["content"] = |
| 2960 | merge_adjacent_user_content(previous_content, current_content); |
| 2961 | if previous.get("_turn_meta_budget").is_none() |
| 2962 | && let Some(meta) = msg.get("_turn_meta_budget") |
| 2963 | { |
| 2964 | previous["_turn_meta_budget"] = meta.clone(); |
| 2965 | } |
| 2966 | } else { |
| 2967 | out.push(msg); |
| 2968 | } |
| 2969 | } |
| 2970 | } |
| 2971 | |
| 2972 | // A user/system/developer wire message closes the contiguous run that |
| 2973 | // must follow an assistant tool-call message. If the same stored |
| 2974 | // message also carries a later tool result, reject it here instead of |
| 2975 | // briefly accepting the result and letting any synthesized media |
| 2976 | // escape after the safety pass strips the malformed batch. |
| 2977 | if out.len() > out_len_before_role_projection |
| 2978 | && !placement.is_assistant_channel() |
| 2979 | && !pending_tool_calls.is_empty() |
| 2980 | { |
| 2981 | logging::warn("Dropping tool-call batch interrupted by non-tool content"); |
| 2982 | pending_tool_calls.clear(); |
| 2983 | deferred_tool_result_images.clear(); |
| 2984 | } |
| 2985 | |
| 2986 | if !tool_results.is_empty() { |
| 2987 | if pending_tool_calls.is_empty() { |
| 2988 | logging::warn("Dropping tool results without matching tool_calls"); |
| 2989 | } else { |
| 2990 | for (tool_id, content, message_label, content_blocks) in tool_results { |
| 2991 | if let Some(tool_info) = pending_tool_calls.remove(&tool_id) { |
| 2992 | let (image, omitted) = crate::image_attach::provider_tool_result_image_refs( |
| 2993 | Some(&content_blocks), |
| 2994 | ); |
| 2995 | let content = |
| 2996 | crate::image_attach::tool_result_text_with_omission(&content, omitted); |
| 2997 | let wire_result = compact_tool_result_for_wire( |
| 2998 | &tool_info.tool_name, |
| 2999 | &tool_info.input, |
| 3000 | &content, |
| 3001 | &message_label, |
| 3002 | tool_result_budget, |
| 3003 | &mut seen_tool_results, |
| 3004 | ); |
| 3005 | let mut tool_msg = json!({ |
| 3006 | "role": "tool", |
| 3007 | "tool_call_id": tool_id, |
| 3008 | "content": wire_result.content, |
| 3009 | }); |
| 3010 | if include_tool_budget_metadata { |
| 3011 | tool_msg["_tool_result_budget"] = json!({ |
| 3012 | "original_chars": wire_result.original_chars, |
| 3013 | "sent_chars": wire_result.sent_chars, |
| 3014 | "truncated": wire_result.truncated, |
| 3015 | "deduplicated": wire_result.deduplicated, |
| 3016 | }); |
| 3017 | } |
| 3018 | out.push(tool_msg); |
| 3019 | if let Some((mime_type, data)) = image { |
| 3020 | deferred_tool_result_images.push(json!({ |
| 3021 | "type": "text", |
| 3022 | "text": format!( |
| 3023 | "Image returned by tool `{}` (call `{tool_id}`):", |
| 3024 | tool_info.tool_name, |
| 3025 | ), |
| 3026 | })); |
| 3027 | deferred_tool_result_images.push(json!({ |
| 3028 | "type": "image_url", |
| 3029 | "image_url": { |
| 3030 | "url": format!("data:{mime_type};base64,{data}") |
| 3031 | }, |
| 3032 | })); |
| 3033 | } |
| 3034 | } else { |
| 3035 | logging::warn(format!( |
| 3036 | "Dropping tool result for unknown tool_call_id: {tool_id}" |
| 3037 | )); |
| 3038 | } |
| 3039 | } |
| 3040 | if pending_tool_calls.is_empty() && !deferred_tool_result_images.is_empty() { |
| 3041 | out.push(json!({ |
| 3042 | "role": "user", |
| 3043 | "content": std::mem::take(&mut deferred_tool_result_images), |
| 3044 | })); |
| 3045 | } |
| 3046 | } |
| 3047 | } else if !placement.is_assistant_channel() { |
| 3048 | pending_tool_calls.clear(); |
| 3049 | deferred_tool_result_images.clear(); |
| 3050 | } |
| 3051 | } |
| 3052 | |
| 3053 | // Safety net: after compaction, an assistant message may have tool_calls |
| 3054 | // whose results were summarized away. The API rejects these, so strip |
| 3055 | // the tool_calls (downgrading to a plain assistant message) and remove |
| 3056 | // the now-orphaned tool result messages. |
| 3057 | let mut i = 0; |
| 3058 | while i < out.len() { |
| 3059 | let is_assistant_with_tools = out[i].get("role").and_then(Value::as_str) |
| 3060 | == Some("assistant") |
| 3061 | && out[i].get("tool_calls").is_some(); |
| 3062 | |
| 3063 | if is_assistant_with_tools { |
| 3064 | let expected_ids: Vec<String> = out[i] |
| 3065 | .get("tool_calls") |
| 3066 | .and_then(Value::as_array) |
| 3067 | .map(|calls| { |
| 3068 | calls |
| 3069 | .iter() |
| 3070 | .filter_map(|c| c.get("id").and_then(Value::as_str).map(String::from)) |
| 3071 | .collect() |
| 3072 | }) |
| 3073 | .unwrap_or_default(); |
| 3074 | |
| 3075 | // Collect tool result IDs immediately following this assistant message. |
| 3076 | let mut found_ids = Vec::new(); |
| 3077 | let mut tool_result_end = i + 1; |
| 3078 | while tool_result_end < out.len() { |
| 3079 | if out[tool_result_end].get("role").and_then(Value::as_str) == Some("tool") { |
| 3080 | if let Some(id) = out[tool_result_end] |
| 3081 | .get("tool_call_id") |
| 3082 | .and_then(Value::as_str) |
| 3083 | { |
| 3084 | found_ids.push(id.to_string()); |
| 3085 | } |
| 3086 | tool_result_end += 1; |
| 3087 | } else { |
| 3088 | break; |
| 3089 | } |
| 3090 | } |
| 3091 | |
| 3092 | // Chat Completions accepts only the immediately contiguous tool |
| 3093 | // run after its assistant tool-call message. Do not accept a |
| 3094 | // later tool result after user/system content has intervened. |
| 3095 | let results_match = expected_ids.len() == found_ids.len() |
| 3096 | && expected_ids.iter().all(|id| found_ids.contains(id)); |
| 3097 | if !results_match { |
| 3098 | let missing: Vec<_> = expected_ids |
| 3099 | .iter() |
| 3100 | .filter(|id| !found_ids.contains(*id)) |
| 3101 | .collect(); |
| 3102 | logging::warn(format!( |
| 3103 | "Stripping orphaned tool_calls from assistant message \ |
| 3104 | (expected {} tool results, found {}, missing: {:?})", |
| 3105 | expected_ids.len(), |
| 3106 | found_ids.len(), |
| 3107 | missing |
| 3108 | )); |
| 3109 | if let Some(obj) = out[i].as_object_mut() { |
| 3110 | obj.remove("tool_calls"); |
| 3111 | } |
| 3112 | // If tool_calls were the only assistant content, remove the now-invalid |
| 3113 | // assistant message entirely (DeepSeek requires content or tool_calls). |
| 3114 | let assistant_content_empty = out[i] |
| 3115 | .get("content") |
| 3116 | .is_none_or(|v| v.is_null() || v.as_str().is_some_and(str::is_empty)); |
| 3117 | if assistant_content_empty { |
| 3118 | // Remove orphaned tool results tied to this stripped assistant call set. |
| 3119 | let mut j = out.len(); |
| 3120 | while j > i + 1 { |
| 3121 | j -= 1; |
| 3122 | if out[j].get("role").and_then(Value::as_str) == Some("tool") |
| 3123 | && let Some(id) = out[j].get("tool_call_id").and_then(Value::as_str) |
| 3124 | && expected_ids.iter().any(|expected| expected == id) |
| 3125 | { |
| 3126 | out.remove(j); |
| 3127 | } |
| 3128 | } |
| 3129 | out.remove(i); |
| 3130 | i = i.saturating_sub(1); |
| 3131 | continue; |
| 3132 | } |
| 3133 | // Remove contiguous tool results first |
| 3134 | if tool_result_end > i + 1 { |
| 3135 | out.drain((i + 1)..tool_result_end); |
| 3136 | } |
| 3137 | // Remove any remaining non-contiguous tool results referencing expected_ids |
| 3138 | // (scan backward to avoid index shifting issues) |
| 3139 | let mut j = out.len(); |
| 3140 | while j > i + 1 { |
| 3141 | j -= 1; |
| 3142 | if out[j].get("role").and_then(Value::as_str) == Some("tool") |
| 3143 | && let Some(id) = out[j].get("tool_call_id").and_then(Value::as_str) |
| 3144 | && expected_ids.iter().any(|expected| expected == id) |
| 3145 | { |
| 3146 | out.remove(j); |
| 3147 | } |
| 3148 | } |
| 3149 | } |
| 3150 | } |
| 3151 | i += 1; |
| 3152 | } |
| 3153 | |
| 3154 | out |
| 3155 | } |
| 3156 | |
| 3157 | pub(super) fn tool_to_chat(tool: &Tool) -> Value { |
| 3158 | let mut value = json!({ |
| 3159 | "type": "function", |
| 3160 | "function": { |
| 3161 | "name": to_api_tool_name(&tool.name), |
| 3162 | "description": tool.description, |
| 3163 | "parameters": tool.input_schema, |
| 3164 | } |
| 3165 | }); |
| 3166 | if let Some(strict) = tool.strict |
| 3167 | && let Some(function) = value.get_mut("function") |
| 3168 | { |
| 3169 | function["strict"] = json!(strict); |
| 3170 | } |
| 3171 | value |
| 3172 | } |
| 3173 | |
| 3174 | pub(super) fn tool_to_chat_for_base_url(tool: &Tool, base_url: &str) -> Value { |
| 3175 | let mut value = tool_to_chat(tool); |
| 3176 | if !deepseek_base_url_supports_strict_tools(base_url) |
| 3177 | && let Some(function) = value.get_mut("function") |
| 3178 | && let Some(obj) = function.as_object_mut() |
| 3179 | { |
| 3180 | obj.remove("strict"); |
| 3181 | } |
| 3182 | value |
| 3183 | } |
| 3184 | |
| 3185 | fn deepseek_base_url_supports_strict_tools(base_url: &str) -> bool { |
| 3186 | let trimmed = base_url.trim_end_matches('/').to_ascii_lowercase(); |
| 3187 | let is_deepseek = trimmed == "https://api.deepseek.com" |
| 3188 | || trimmed == "https://api.deepseek.com/v1" |
| 3189 | || trimmed == "https://api.deepseek.com/beta" |
| 3190 | || trimmed == "https://api.deepseeki.com" |
| 3191 | || trimmed == "https://api.deepseeki.com/v1" |
| 3192 | || trimmed == "https://api.deepseeki.com/beta"; |
| 3193 | !is_deepseek || trimmed.ends_with("/beta") |
| 3194 | } |
| 3195 | |
| 3196 | fn map_tool_choice_for_chat(choice: &Value) -> Option<Value> { |
| 3197 | if let Some(choice_str) = choice.as_str() { |
| 3198 | return Some(json!(choice_str)); |
| 3199 | } |
| 3200 | let Some(choice_type) = choice.get("type").and_then(Value::as_str) else { |
| 3201 | return Some(choice.clone()); |
| 3202 | }; |
| 3203 | |
| 3204 | match choice_type { |
| 3205 | "auto" | "none" => Some(json!(choice_type)), |
| 3206 | "any" => Some(json!("auto")), |
| 3207 | "tool" => choice.get("name").and_then(Value::as_str).map(|name| { |
| 3208 | json!({ |
| 3209 | "type": "function", |
| 3210 | "function": { "name": to_api_tool_name(name) } |
| 3211 | }) |
| 3212 | }), |
| 3213 | _ => Some(choice.clone()), |
| 3214 | } |
| 3215 | } |
| 3216 | |
| 3217 | fn should_send_tool_choice_for_chat(provider: ProviderKind, effort: Option<&str>) -> bool { |
| 3218 | if !matches!(provider, ProviderKind::Deepseek) { |
| 3219 | return true; |
| 3220 | } |
| 3221 | !reasoning_effort_enables_thinking(effort) |
| 3222 | } |
| 3223 | |
| 3224 | fn reasoning_effort_enables_thinking(effort: Option<&str>) -> bool { |
| 3225 | let Some(effort) = effort else { |
| 3226 | return false; |
| 3227 | }; |
| 3228 | !matches!( |
| 3229 | effort.trim().to_ascii_lowercase().as_str(), |
| 3230 | "off" | "disabled" | "none" | "false" |
| 3231 | ) |
| 3232 | } |
| 3233 | |
| 3234 | /// Final-pass sanitizer over the outgoing chat-completions JSON payload. |
| 3235 | /// Forces a non-empty `reasoning_content` onto assistant messages that carry |
| 3236 | /// `tool_calls`, when the model + effort combination requires it. DeepSeek's |
| 3237 | /// thinking-mode API rejects such messages with a 400 error; substituting a |
| 3238 | /// placeholder keeps the conversation chain intact. Non-tool assistant |
| 3239 | /// reasoning can stay omitted once a later user text turn begins. |
| 3240 | /// |
| 3241 | /// Also tallies the size of all replayed `reasoning_content` and logs it, so |
| 3242 | /// users on `RUST_LOG=codewhale_tui=debug` can see how much of their input |
| 3243 | /// budget is being spent re-sending prior thinking traces. |
| 3244 | #[cfg(test)] |
| 3245 | pub(super) fn sanitize_thinking_mode_messages( |
| 3246 | body: &mut Value, |
| 3247 | model: &str, |
| 3248 | effort: Option<&str>, |
| 3249 | provider: ProviderKind, |
| 3250 | ) -> Option<u32> { |
| 3251 | sanitize_thinking_mode_messages_for_route(body, model, effort, provider, "") |
| 3252 | } |
| 3253 | |
| 3254 | /// Route-aware variant of `sanitize_thinking_mode_messages`. |
| 3255 | /// |
| 3256 | /// The wrapper above remains intentionally route-agnostic for existing test |
| 3257 | /// helpers and generic callers. Production chat requests call this version so |
| 3258 | /// exact Kimi Code K3 assistant tool turns retain the reasoning trace that |
| 3259 | /// K3 expects on the next request. |
| 3260 | pub(super) fn sanitize_thinking_mode_messages_for_route( |
| 3261 | body: &mut Value, |
| 3262 | model: &str, |
| 3263 | effort: Option<&str>, |
| 3264 | provider: ProviderKind, |
| 3265 | base_url: &str, |
| 3266 | ) -> Option<u32> { |
| 3267 | // Mistral replay is encoded inside polymorphic `content` blocks, not the |
| 3268 | // DeepSeek `reasoning_content` field. Running the DeepSeek placeholder |
| 3269 | // sanitizer after reshaping would add a second, invalid reasoning dialect |
| 3270 | // to assistant tool-call turns. |
| 3271 | if is_exact_mistral_chat_route(provider, base_url) { |
| 3272 | return None; |
| 3273 | } |
| 3274 | if !should_replay_reasoning_content_for_provider_on_route(provider, base_url, model, effort) { |
| 3275 | return None; |
| 3276 | } |
| 3277 | let messages = body.get_mut("messages").and_then(Value::as_array_mut)?; |
| 3278 | let mut substitutions: u32 = 0; |
| 3279 | let mut replay_chars: u64 = 0; |
| 3280 | let mut replay_messages: u32 = 0; |
| 3281 | for (idx, msg) in messages.iter_mut().enumerate() { |
| 3282 | if msg.get("role").and_then(Value::as_str) != Some("assistant") { |
| 3283 | continue; |
| 3284 | } |
| 3285 | let has_tool_calls = msg.get("tool_calls").is_some(); |
| 3286 | let needs_placeholder = msg |
| 3287 | .get("reasoning_content") |
| 3288 | .and_then(Value::as_str) |
| 3289 | .is_none_or(|s| s.trim().is_empty()); |
| 3290 | if has_tool_calls && needs_placeholder { |
| 3291 | msg["reasoning_content"] = json!(REASONING_REPLAY_PLACEHOLDER); |
| 3292 | substitutions = substitutions.saturating_add(1); |
| 3293 | logging::warn(format!( |
| 3294 | "Final sanitizer: forced reasoning_content placeholder on assistant[{idx}]", |
| 3295 | )); |
| 3296 | } |
| 3297 | if let Some(reasoning) = msg.get("reasoning_content").and_then(Value::as_str) { |
| 3298 | let len = reasoning.len() as u64; |
| 3299 | if len > 0 { |
| 3300 | replay_chars = replay_chars.saturating_add(len); |
| 3301 | replay_messages = replay_messages.saturating_add(1); |
| 3302 | } |
| 3303 | } |
| 3304 | } |
| 3305 | if substitutions > 0 { |
| 3306 | logging::warn(format!( |
| 3307 | "Final sanitizer: {substitutions} assistant message(s) needed reasoning_content placeholder", |
| 3308 | )); |
| 3309 | } |
| 3310 | if replay_messages == 0 { |
| 3311 | return None; |
| 3312 | } |
| 3313 | // ~4 chars/token is the standard rough estimate; DeepSeek tokens skew |
| 3314 | // a touch shorter on Chinese/code but this is order-of-magnitude info. |
| 3315 | let approx_tokens = (replay_chars / 4).min(u64::from(u32::MAX)) as u32; |
| 3316 | logging::info(format!( |
| 3317 | "Reasoning-content replay: {replay_messages} assistant message(s), ~{approx_tokens} input tokens ({replay_chars} chars) being re-sent in this request", |
| 3318 | )); |
| 3319 | Some(approx_tokens) |
| 3320 | } |
| 3321 | |
| 3322 | /// Sums the byte length of `reasoning_content` across all assistant messages in |
| 3323 | /// an outgoing chat-completions body. Used by tests; the production sanitizer |
| 3324 | /// computes the same number inline and logs it. |
| 3325 | #[cfg(test)] |
| 3326 | pub(super) fn count_reasoning_replay_chars(body: &Value) -> u64 { |
| 3327 | let Some(messages) = body.get("messages").and_then(Value::as_array) else { |
| 3328 | return 0; |
| 3329 | }; |
| 3330 | messages |
| 3331 | .iter() |
| 3332 | .filter(|m| m.get("role").and_then(Value::as_str) == Some("assistant")) |
| 3333 | .filter_map(|m| m.get("reasoning_content").and_then(Value::as_str)) |
| 3334 | .map(|s| s.len() as u64) |
| 3335 | .sum() |
| 3336 | } |
| 3337 | |
| 3338 | /// Render the transport-shape headers we care about for #103 diagnostics. |
| 3339 | /// Always returns SOMETHING printable so the decode-error log line is parseable |
| 3340 | /// even when the server stripped a header we expected. |
| 3341 | fn format_stream_headers(headers: &reqwest::header::HeaderMap) -> String { |
| 3342 | const FIELDS: &[&str] = &[ |
| 3343 | "content-encoding", |
| 3344 | "transfer-encoding", |
| 3345 | "connection", |
| 3346 | "server", |
| 3347 | ]; |
| 3348 | let mut parts: Vec<String> = Vec::with_capacity(FIELDS.len()); |
| 3349 | for field in FIELDS { |
| 3350 | let rendered = headers |
| 3351 | .get(*field) |
| 3352 | .and_then(|v| v.to_str().ok()) |
| 3353 | .unwrap_or("(absent)"); |
| 3354 | parts.push(format!("{field}={rendered}")); |
| 3355 | } |
| 3356 | parts.join(", ") |
| 3357 | } |
| 3358 | |
| 3359 | /// Diagnostic logger fired when DeepSeek rejects the request despite the |
| 3360 | /// sanitizer. Walks the body and logs which assistant messages have tool_calls |
| 3361 | /// but no `reasoning_content` — useful to track down a code path that bypasses |
| 3362 | /// the sanitizer entirely. |
| 3363 | fn log_thinking_mode_violations(body: &Value) { |
| 3364 | let Some(messages) = body.get("messages").and_then(Value::as_array) else { |
| 3365 | logging::warn("400-after-sanitizer: body has no `messages` array"); |
| 3366 | return; |
| 3367 | }; |
| 3368 | let mut violations: Vec<String> = Vec::new(); |
| 3369 | for (idx, msg) in messages.iter().enumerate() { |
| 3370 | if msg.get("role").and_then(Value::as_str) != Some("assistant") { |
| 3371 | continue; |
| 3372 | } |
| 3373 | let reasoning = msg |
| 3374 | .get("reasoning_content") |
| 3375 | .and_then(Value::as_str) |
| 3376 | .unwrap_or(""); |
| 3377 | let has_tc = msg.get("tool_calls").is_some(); |
| 3378 | if reasoning.trim().is_empty() { |
| 3379 | violations.push(format!( |
| 3380 | "assistant[{idx}] (reasoning_content missing, tool_calls={has_tc})" |
| 3381 | )); |
| 3382 | } |
| 3383 | } |
| 3384 | if violations.is_empty() { |
| 3385 | logging::warn( |
| 3386 | "400-after-sanitizer: all assistant messages have reasoning_content — DeepSeek rejected for a different reason", |
| 3387 | ); |
| 3388 | } else { |
| 3389 | logging::warn(format!( |
| 3390 | "400-after-sanitizer: {} assistant message(s) lack reasoning_content despite sanitizer: {}", |
| 3391 | violations.len(), |
| 3392 | violations.join(", ") |
| 3393 | )); |
| 3394 | } |
| 3395 | } |
| 3396 | |
| 3397 | fn requires_reasoning_content(model: &str) -> bool { |
| 3398 | let lower = model.to_lowercase(); |
| 3399 | // V4-family direct model IDs. |
| 3400 | lower.contains("deepseek-v4") |
| 3401 | // Public DeepSeek API aliases routed server-side to the V4 family. |
| 3402 | // `deepseek-chat` resolves to `deepseek-v4-flash` and `deepseek-reasoner` |
| 3403 | // resolves to `deepseek-v4-pro`; both have thinking mode enabled by |
| 3404 | // default, so any assistant message carrying tool_calls must replay |
| 3405 | // `reasoning_content` on subsequent turns or the API returns 400. |
| 3406 | || lower.starts_with("deepseek-chat") |
| 3407 | || lower.starts_with("deepseek-reasoner") |
| 3408 | || has_deepseek_r_series_marker(&lower) |
| 3409 | // #6044: the V4.1 official id dropped the version number entirely |
| 3410 | // (`deepseek-flash`), so the literal arms above cannot see it and |
| 3411 | // the decode/replay classifiers depended on a later catalog fallback |
| 3412 | // to catch it. The catalog owns the capability — consult it here so |
| 3413 | // every caller (stream style, wire replay, prompt inspection) agrees |
| 3414 | // without another hardcoded id. |
| 3415 | || (lower.starts_with("deepseek-") && model_supports_reasoning(model)) |
| 3416 | } |
| 3417 | |
| 3418 | fn should_replay_reasoning_content(model: &str, effort: Option<&str>) -> bool { |
| 3419 | if effort |
| 3420 | .map(|value| { |
| 3421 | matches!( |
| 3422 | value.trim().to_ascii_lowercase().as_str(), |
| 3423 | "off" | "disabled" | "none" | "false" |
| 3424 | ) |
| 3425 | }) |
| 3426 | .unwrap_or(false) |
| 3427 | { |
| 3428 | return false; |
| 3429 | } |
| 3430 | |
| 3431 | requires_reasoning_content(model) |
| 3432 | } |
| 3433 | |
| 3434 | #[cfg(test)] |
| 3435 | fn should_replay_reasoning_content_for_provider( |
| 3436 | provider: ProviderKind, |
| 3437 | model: &str, |
| 3438 | effort: Option<&str>, |
| 3439 | ) -> bool { |
| 3440 | should_replay_reasoning_content_for_provider_on_route(provider, "", model, effort) |
| 3441 | } |
| 3442 | |
| 3443 | /// Route-aware reasoning replay policy. |
| 3444 | /// |
| 3445 | /// Keep the bare K3 identifier out of the global model catalog: direct |
| 3446 | /// Moonshot and arbitrary OpenAI-compatible routes can also expose a `k3` |
| 3447 | /// model name, but only Kimi Code's exact membership-plan endpoint has this |
| 3448 | /// replay contract. |
| 3449 | fn should_replay_reasoning_content_for_provider_on_route( |
| 3450 | provider: ProviderKind, |
| 3451 | base_url: &str, |
| 3452 | model: &str, |
| 3453 | effort: Option<&str>, |
| 3454 | ) -> bool { |
| 3455 | // Exact always-thinking routes replay their reasoning trace regardless of |
| 3456 | // a stale caller effort: the API contract requires the assistant |
| 3457 | // reasoning field on later tool turns for multi-turn continuity. |
| 3458 | if is_exact_direct_moonshot_k3_route(provider, base_url, model) |
| 3459 | || is_exact_kimi_code_k3_route(provider, base_url, model) |
| 3460 | || (is_exact_mistral_chat_route(provider, base_url) |
| 3461 | && mistral_model_has_native_reasoning(model)) |
| 3462 | { |
| 3463 | return true; |
| 3464 | } |
| 3465 | |
| 3466 | // The exact Model Studio route policy evaluates BEFORE any generic |
| 3467 | // model-name heuristic. Only models Alibaba documents as accepting |
| 3468 | // `preserve_thinking` replay historical `reasoning_content`, plus the |
| 3469 | // concrete DeepSeek V4 family ids whose own API contract requires the |
| 3470 | // reasoning field on tool turns. A model named `foo-thinking` or |
| 3471 | // `foo-reasoner` proves nothing about DashScope's request dialect and |
| 3472 | // must not gain replay here — replaying stale Thinking blocks feeds the |
| 3473 | // model its own past reasoning and re-triggers it every turn (observed |
| 3474 | // as a repeated handoff loop with the always-thinking qwen3.8 family). |
| 3475 | // Pi does not replay those either. |
| 3476 | if is_exact_modelstudio_chat_route(provider, base_url) { |
| 3477 | if modelstudio_model_supports_preserve_thinking(model) { |
| 3478 | // Thinking-only preserve models (kimi-k2.7-code) are |
| 3479 | // always-thinking and replay even with a stale `off`; hybrid |
| 3480 | // preserve models defer to the effort gate like every other |
| 3481 | // hybrid. |
| 3482 | if is_exact_modelstudio_thinking_only_route(provider, base_url, model) |
| 3483 | || !modelstudio_effort_disables_thinking(effort) |
| 3484 | { |
| 3485 | return true; |
| 3486 | } |
| 3487 | return false; |
| 3488 | } |
| 3489 | let lower = model.trim().to_ascii_lowercase(); |
| 3490 | return lower.contains("deepseek-v4") |
| 3491 | || lower.starts_with("deepseek-chat") |
| 3492 | || lower.starts_with("deepseek-reasoner"); |
| 3493 | } |
| 3494 | |
| 3495 | if effort |
| 3496 | .map(|value| { |
| 3497 | matches!( |
| 3498 | value.trim().to_ascii_lowercase().as_str(), |
| 3499 | "off" | "disabled" | "none" | "false" |
| 3500 | ) |
| 3501 | }) |
| 3502 | .unwrap_or(false) |
| 3503 | { |
| 3504 | return false; |
| 3505 | } |
| 3506 | |
| 3507 | if requires_reasoning_content(model) { |
| 3508 | return true; |
| 3509 | } |
| 3510 | |
| 3511 | // Xiaomi MiMo's API requires the assistant `reasoning_content` back on |
| 3512 | // tool-call turns (omitting it is a 400), for every MiMo chat model, not |
| 3513 | // only the ids the offline catalog marks as reasoning (#6501). A turn |
| 3514 | // that produced no reasoning replays nothing. |
| 3515 | if provider == ProviderKind::XiaomiMimo { |
| 3516 | return true; |
| 3517 | } |
| 3518 | |
| 3519 | if is_exact_mistral_chat_route(provider, base_url) |
| 3520 | && mistral_model_has_adjustable_reasoning(model) |
| 3521 | { |
| 3522 | return true; |
| 3523 | } |
| 3524 | |
| 3525 | if !provider_accepts_reasoning_content(provider) { |
| 3526 | // Generic non-DeepSeek model on a provider that rejects the field: |
| 3527 | // keep stripping it (preserves the #1542 fix). But a known DeepSeek |
| 3528 | // reasoning model pointed at a DeepSeek-compatible endpoint via the |
| 3529 | // generic `openai` provider still requires reasoning_content replay, |
| 3530 | // or the thinking-mode API returns 400 (#1739 / #1694). |
| 3531 | return false; |
| 3532 | } |
| 3533 | |
| 3534 | model_supports_reasoning(model) |
| 3535 | } |
| 3536 | |
| 3537 | /// Test shorthand: does this route surface `reasoning_content` / |
| 3538 | /// `reasoning` / `reasoning_details` deltas as Thinking by default? |
| 3539 | #[cfg(test)] |
| 3540 | fn is_reasoning_model_for_stream(provider: ProviderKind, model: &str) -> bool { |
| 3541 | reasoning_stream_style_for_stream(provider, model, None) == ReasoningStreamStyle::SeparateField |
| 3542 | } |
| 3543 | |
| 3544 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 3545 | pub(super) enum ReasoningStreamStyle { |
| 3546 | SeparateField, |
| 3547 | InlineTags, |
| 3548 | MistralBlocks, |
| 3549 | None, |
| 3550 | } |
| 3551 | |
| 3552 | #[cfg(test)] |
| 3553 | fn reasoning_stream_style_for_stream( |
| 3554 | provider: ProviderKind, |
| 3555 | model: &str, |
| 3556 | configured: Option<&str>, |
| 3557 | ) -> ReasoningStreamStyle { |
| 3558 | reasoning_stream_style_for_route(provider, "", model, configured) |
| 3559 | } |
| 3560 | |
| 3561 | /// Choose stream decoding semantics for a fully resolved provider route. |
| 3562 | fn reasoning_stream_style_for_route( |
| 3563 | provider: ProviderKind, |
| 3564 | base_url: &str, |
| 3565 | model: &str, |
| 3566 | configured: Option<&str>, |
| 3567 | ) -> ReasoningStreamStyle { |
| 3568 | if is_exact_mistral_chat_route(provider, base_url) && mistral_model_supports_reasoning(model) { |
| 3569 | return ReasoningStreamStyle::MistralBlocks; |
| 3570 | } |
| 3571 | if let Some(configured) = configured { |
| 3572 | if let Some(style) = parse_reasoning_stream_style(configured) { |
| 3573 | return style; |
| 3574 | } |
| 3575 | logging::warn(format!( |
| 3576 | "Ignoring unrecognized reasoning_stream_style `{configured}`; expected separate_field, inline_tags, or none" |
| 3577 | )); |
| 3578 | } |
| 3579 | // #6501: the dedicated reasoning fields are reasoning on every |
| 3580 | // Chat Completions route that sends them (DeepSeek, xAI, Xiaomi MiMo, |
| 3581 | // Kimi, GLM, MiniMax, DashScope, vLLM/SGLang reasoning parsers, OpenRouter, |
| 3582 | // Ollama, ...), and the non-streaming parser already treats them that way |
| 3583 | // unconditionally. Gating the stream on a provider allowlist AND a |
| 3584 | // per-model catalog row meant every unlisted provider (xAI, StepFun, |
| 3585 | // Together, Ollama, the Codewhale gateway, ...) and every model id newer |
| 3586 | // than the offline catalog (grok-4.7, a fresh MiMo id) rendered its |
| 3587 | // reasoning as answer prose. Surfacing a field that never arrives costs |
| 3588 | // nothing; a gateway that really puts its answer in `reasoning_content` |
| 3589 | // opts out with `reasoning_stream_style = "none"`. Replay of reasoning in |
| 3590 | // request history stays separately gated by |
| 3591 | // `should_replay_reasoning_content_for_provider_on_route`. |
| 3592 | ReasoningStreamStyle::SeparateField |
| 3593 | } |
| 3594 | |
| 3595 | fn parse_reasoning_stream_style(value: &str) -> Option<ReasoningStreamStyle> { |
| 3596 | match value.trim().to_ascii_lowercase().replace('-', "_").as_str() { |
| 3597 | "separate_field" | "separate" | "field" => Some(ReasoningStreamStyle::SeparateField), |
| 3598 | "inline_tags" | "inline" | "think_tags" | "thinking_tags" => { |
| 3599 | Some(ReasoningStreamStyle::InlineTags) |
| 3600 | } |
| 3601 | "none" | "text" | "disabled" | "off" => Some(ReasoningStreamStyle::None), |
| 3602 | _ => None, |
| 3603 | } |
| 3604 | } |
| 3605 | |
| 3606 | /// Providers whose chat-completions API both returns and accepts a dedicated |
| 3607 | /// `reasoning_content` field on assistant messages. |
| 3608 | /// |
| 3609 | /// Arcee is intentionally included. Trinity-Large-Thinking natively emits |
| 3610 | /// `<think>...</think>` traces, but Arcee's hosted API serves it through vLLM |
| 3611 | /// with `--reasoning-parser deepseek_r1`, which parses those blocks into a |
| 3612 | /// `reasoning_content` field (verified live against `api.arcee.ai`: thinking |
| 3613 | /// streams as `delta.reasoning_content`, the answer as `delta.content`, with no |
| 3614 | /// `<think>` tags on the wire). Arcee's docs require replaying `reasoning_content` |
| 3615 | /// on assistant tool-call turns; dropping it makes the model emit tool calls as |
| 3616 | /// raw XML inside its thinking ("xml_in_reasoning" pitfall). Do not remove Arcee |
| 3617 | /// here without new live evidence — see docs.arcee.ai/capabilities/reasoning-traces. |
| 3618 | fn provider_accepts_reasoning_content(provider: ProviderKind) -> bool { |
| 3619 | matches!( |
| 3620 | provider, |
| 3621 | ProviderKind::Deepseek |
| 3622 | | ProviderKind::NvidiaNim |
| 3623 | | ProviderKind::Openrouter |
| 3624 | | ProviderKind::XiaomiMimo |
| 3625 | | ProviderKind::Novita |
| 3626 | | ProviderKind::Fireworks |
| 3627 | | ProviderKind::Siliconflow |
| 3628 | | ProviderKind::SiliconflowCN |
| 3629 | | ProviderKind::Volcengine |
| 3630 | | ProviderKind::Arcee |
| 3631 | | ProviderKind::Minimax |
| 3632 | | ProviderKind::Sglang |
| 3633 | | ProviderKind::Zai |
| 3634 | | ProviderKind::Moonshot // #3016: Kimi thinking traces use reasoning_content |
| 3635 | ) |
| 3636 | } |
| 3637 | |
| 3638 | fn has_deepseek_r_series_marker(model_lower: &str) -> bool { |
| 3639 | const PREFIX: &str = "deepseek-r"; |
| 3640 | model_lower.match_indices(PREFIX).any(|(idx, _)| { |
| 3641 | model_lower[idx + PREFIX.len()..] |
| 3642 | .chars() |
| 3643 | .next() |
| 3644 | .is_some_and(|ch| ch.is_ascii_digit()) |
| 3645 | }) |
| 3646 | } |
| 3647 | |
| 3648 | /// Transport-only reasoning replay placeholder. DeepSeek-family chat wires |
| 3649 | /// reject assistant tool-call messages with an empty `reasoning_content`, so |
| 3650 | /// the request serializer substitutes this string for replay only. Providers |
| 3651 | /// that mirror assistant history (GLM-5.x) stream the substituted field back |
| 3652 | /// as a live reasoning delta; ingest must drop that exact echo so a wire-only |
| 3653 | /// placeholder never becomes a persisted or displayed thinking block. |
| 3654 | pub(crate) const REASONING_REPLAY_PLACEHOLDER: &str = "(reasoning omitted)"; |
| 3655 | |
| 3656 | #[must_use] |
| 3657 | pub(crate) fn is_reasoning_replay_placeholder(text: &str) -> bool { |
| 3658 | text.trim() == REASONING_REPLAY_PLACEHOLDER |
| 3659 | } |
| 3660 | |
| 3661 | fn reasoning_delta( |
| 3662 | value: &Value, |
| 3663 | choice_index: u32, |
| 3664 | reasoning_detail_buffers: &mut std::collections::HashMap<u32, String>, |
| 3665 | ) -> Option<String> { |
| 3666 | if let Some(reasoning) = value |
| 3667 | .get("reasoning_content") |
| 3668 | .or_else(|| value.get("reasoning")) |
| 3669 | .and_then(Value::as_str) |
| 3670 | { |
| 3671 | if is_reasoning_replay_placeholder(reasoning) { |
| 3672 | return None; |
| 3673 | } |
| 3674 | return Some(reasoning.to_string()); |
| 3675 | } |
| 3676 | |
| 3677 | let details = value.get("reasoning_details").and_then(Value::as_array)?; |
| 3678 | let full_text = details |
| 3679 | .iter() |
| 3680 | .filter_map(|detail| detail.get("text").and_then(Value::as_str)) |
| 3681 | .collect::<String>(); |
| 3682 | if full_text.is_empty() { |
| 3683 | return None; |
| 3684 | } |
| 3685 | |
| 3686 | let previous = reasoning_detail_buffers.entry(choice_index).or_default(); |
| 3687 | let delta = full_text |
| 3688 | .strip_prefix(previous.as_str()) |
| 3689 | .unwrap_or(&full_text) |
| 3690 | .to_string(); |
| 3691 | *previous = full_text; |
| 3692 | Some(delta) |
| 3693 | } |
| 3694 | |
| 3695 | fn reasoning_message_text(value: &Value) -> Option<String> { |
| 3696 | if let Some(reasoning) = value |
| 3697 | .get("reasoning_content") |
| 3698 | .or_else(|| value.get("reasoning")) |
| 3699 | .and_then(Value::as_str) |
| 3700 | { |
| 3701 | if is_reasoning_replay_placeholder(reasoning) { |
| 3702 | return None; |
| 3703 | } |
| 3704 | return Some(reasoning.to_string()); |
| 3705 | } |
| 3706 | value |
| 3707 | .get("reasoning_details") |
| 3708 | .and_then(Value::as_array) |
| 3709 | .map(|details| { |
| 3710 | details |
| 3711 | .iter() |
| 3712 | .filter_map(|detail| detail.get("text").and_then(Value::as_str)) |
| 3713 | .collect::<String>() |
| 3714 | }) |
| 3715 | } |
| 3716 | |
| 3717 | #[cfg(test)] |
| 3718 | pub(super) fn parse_chat_message(payload: &Value) -> Result<MessageResponse> { |
| 3719 | parse_chat_message_for_route(payload, ProviderKind::Openai, "") |
| 3720 | } |
| 3721 | |
| 3722 | fn parse_chat_message_for_route( |
| 3723 | payload: &Value, |
| 3724 | provider: ProviderKind, |
| 3725 | base_url: &str, |
| 3726 | ) -> Result<MessageResponse> { |
| 3727 | let id = payload |
| 3728 | .get("id") |
| 3729 | .and_then(Value::as_str) |
| 3730 | .unwrap_or("chatcmpl") |
| 3731 | .to_string(); |
| 3732 | let model = payload |
| 3733 | .get("model") |
| 3734 | .and_then(Value::as_str) |
| 3735 | .unwrap_or("unknown") |
| 3736 | .to_string(); |
| 3737 | |
| 3738 | let choices = payload |
| 3739 | .get("choices") |
| 3740 | .and_then(Value::as_array) |
| 3741 | .context("Chat API response missing choices")?; |
| 3742 | let choice = choices |
| 3743 | .first() |
| 3744 | .context("Chat API response missing first choice")?; |
| 3745 | let message = choice |
| 3746 | .get("message") |
| 3747 | .context("Chat API response missing message")?; |
| 3748 | |
| 3749 | let mut content_blocks = Vec::new(); |
| 3750 | if let Some(reasoning) = |
| 3751 | reasoning_message_text(message).filter(|reasoning| !reasoning.trim().is_empty()) |
| 3752 | { |
| 3753 | content_blocks.push(ContentBlock::Thinking { |
| 3754 | signature: None, |
| 3755 | state: None, |
| 3756 | thinking: reasoning.to_string(), |
| 3757 | }); |
| 3758 | } |
| 3759 | let (mistral_thinking, mistral_text) = if is_exact_mistral_chat_route(provider, base_url) { |
| 3760 | extract_mistral_polymorphic_content(message) |
| 3761 | } else { |
| 3762 | (None, None) |
| 3763 | }; |
| 3764 | if let Some(thinking) = mistral_thinking.filter(|s| !s.trim().is_empty()) { |
| 3765 | content_blocks.push(ContentBlock::Thinking { |
| 3766 | signature: None, |
| 3767 | state: None, |
| 3768 | thinking, |
| 3769 | }); |
| 3770 | } |
| 3771 | if let Some(text) = mistral_text.filter(|s| !s.trim().is_empty()) { |
| 3772 | content_blocks.push(ContentBlock::Text { |
| 3773 | text, |
| 3774 | cache_control: None, |
| 3775 | }); |
| 3776 | } else if let Some(text) = message.get("content").and_then(Value::as_str) |
| 3777 | && !text.trim().is_empty() |
| 3778 | { |
| 3779 | content_blocks.push(ContentBlock::Text { |
| 3780 | text: text.to_string(), |
| 3781 | cache_control: None, |
| 3782 | }); |
| 3783 | } |
| 3784 | |
| 3785 | if let Some(tool_calls) = message.get("tool_calls").and_then(Value::as_array) { |
| 3786 | for call in tool_calls { |
| 3787 | let id = call |
| 3788 | .get("id") |
| 3789 | .and_then(Value::as_str) |
| 3790 | .unwrap_or("tool_call") |
| 3791 | .to_string(); |
| 3792 | let function = call.get("function"); |
| 3793 | let name = tool_name_or_fallback( |
| 3794 | function.and_then(|f| f.get("name")).and_then(Value::as_str), |
| 3795 | &id, |
| 3796 | "Non-streaming response", |
| 3797 | ); |
| 3798 | let arguments = function |
| 3799 | .and_then(|f| f.get("arguments")) |
| 3800 | .and_then(Value::as_str) |
| 3801 | .map(|raw| serde_json::from_str(raw).unwrap_or(Value::String(raw.to_string()))) |
| 3802 | .unwrap_or(Value::Null); |
| 3803 | let caller = call.get("caller").and_then(|v| { |
| 3804 | v.get("type") |
| 3805 | .and_then(Value::as_str) |
| 3806 | .map(|caller_type| ToolCaller { |
| 3807 | caller_type: caller_type.to_string(), |
| 3808 | tool_id: v |
| 3809 | .get("tool_id") |
| 3810 | .and_then(Value::as_str) |
| 3811 | .map(std::string::ToString::to_string), |
| 3812 | }) |
| 3813 | }); |
| 3814 | |
| 3815 | let thought_signature = call |
| 3816 | .pointer("/extra_content/google/thought_signature") |
| 3817 | .and_then(Value::as_str) |
| 3818 | .map(str::to_string); |
| 3819 | content_blocks.push(ContentBlock::ToolUse { |
| 3820 | execution_id: None, |
| 3821 | id, |
| 3822 | name: from_api_tool_name(&name), |
| 3823 | input: arguments, |
| 3824 | caller, |
| 3825 | thought_signature, |
| 3826 | }); |
| 3827 | } |
| 3828 | } |
| 3829 | |
| 3830 | let usage = parse_usage(payload.get("usage")); |
| 3831 | |
| 3832 | Ok(MessageResponse { |
| 3833 | id, |
| 3834 | r#type: "message".to_string(), |
| 3835 | role: "assistant".to_string(), |
| 3836 | content: content_blocks, |
| 3837 | model, |
| 3838 | stop_reason: choice |
| 3839 | .get("finish_reason") |
| 3840 | .and_then(Value::as_str) |
| 3841 | .map(str::to_string), |
| 3842 | stop_sequence: None, |
| 3843 | container: None, |
| 3844 | usage, |
| 3845 | }) |
| 3846 | } |
| 3847 | |
| 3848 | #[derive(Debug, Default)] |
| 3849 | struct InlineReasoningTagState { |
| 3850 | inside_think: bool, |
| 3851 | pending: String, |
| 3852 | } |
| 3853 | |
| 3854 | #[derive(Debug, PartialEq, Eq)] |
| 3855 | enum ReasoningSegment { |
| 3856 | Text(String), |
| 3857 | Thinking(String), |
| 3858 | } |
| 3859 | |
| 3860 | fn inline_reasoning_segments( |
| 3861 | content: &str, |
| 3862 | state: &mut InlineReasoningTagState, |
| 3863 | flush: bool, |
| 3864 | ) -> Vec<ReasoningSegment> { |
| 3865 | state.pending.push_str(content); |
| 3866 | let mut segments = Vec::new(); |
| 3867 | |
| 3868 | loop { |
| 3869 | if state.pending.is_empty() { |
| 3870 | break; |
| 3871 | } |
| 3872 | |
| 3873 | if state.inside_think { |
| 3874 | if let Some(close_at) = state.pending.find("</think>") { |
| 3875 | push_reasoning_segment( |
| 3876 | &mut segments, |
| 3877 | ReasoningSegment::Thinking(state.pending[..close_at].to_string()), |
| 3878 | ); |
| 3879 | state.pending.drain(..close_at + "</think>".len()); |
| 3880 | state.inside_think = false; |
| 3881 | continue; |
| 3882 | } |
| 3883 | |
| 3884 | let hold_len = if flush { |
| 3885 | 0 |
| 3886 | } else { |
| 3887 | trailing_tag_prefix_len(&state.pending, "</think>") |
| 3888 | }; |
| 3889 | let emit_len = state.pending.len().saturating_sub(hold_len); |
| 3890 | if emit_len > 0 { |
| 3891 | push_reasoning_segment( |
| 3892 | &mut segments, |
| 3893 | ReasoningSegment::Thinking(state.pending[..emit_len].to_string()), |
| 3894 | ); |
| 3895 | state.pending.drain(..emit_len); |
| 3896 | } |
| 3897 | break; |
| 3898 | } |
| 3899 | |
| 3900 | if let Some(open_at) = state.pending.find("<think>") { |
| 3901 | push_reasoning_segment( |
| 3902 | &mut segments, |
| 3903 | ReasoningSegment::Text(state.pending[..open_at].to_string()), |
| 3904 | ); |
| 3905 | state.pending.drain(..open_at + "<think>".len()); |
| 3906 | state.inside_think = true; |
| 3907 | continue; |
| 3908 | } |
| 3909 | |
| 3910 | let hold_len = if flush { |
| 3911 | 0 |
| 3912 | } else { |
| 3913 | trailing_tag_prefix_len(&state.pending, "<think>") |
| 3914 | }; |
| 3915 | let emit_len = state.pending.len().saturating_sub(hold_len); |
| 3916 | if emit_len > 0 { |
| 3917 | push_reasoning_segment( |
| 3918 | &mut segments, |
| 3919 | ReasoningSegment::Text(state.pending[..emit_len].to_string()), |
| 3920 | ); |
| 3921 | state.pending.drain(..emit_len); |
| 3922 | } |
| 3923 | break; |
| 3924 | } |
| 3925 | |
| 3926 | segments |
| 3927 | } |
| 3928 | |
| 3929 | fn trailing_tag_prefix_len(content: &str, tag: &str) -> usize { |
| 3930 | let max_len = tag.len().min(content.len()); |
| 3931 | for len in (1..=max_len).rev() { |
| 3932 | let start = content.len() - len; |
| 3933 | if content.is_char_boundary(start) && tag.starts_with(&content[start..]) { |
| 3934 | return len; |
| 3935 | } |
| 3936 | } |
| 3937 | 0 |
| 3938 | } |
| 3939 | |
| 3940 | fn push_reasoning_segment(segments: &mut Vec<ReasoningSegment>, segment: ReasoningSegment) { |
| 3941 | match &segment { |
| 3942 | ReasoningSegment::Text(text) | ReasoningSegment::Thinking(text) if text.is_empty() => {} |
| 3943 | _ => segments.push(segment), |
| 3944 | } |
| 3945 | } |
| 3946 | |
| 3947 | fn push_text_delta( |
| 3948 | events: &mut Vec<StreamEvent>, |
| 3949 | content_index: &mut u32, |
| 3950 | text_started: &mut bool, |
| 3951 | thinking_started: &mut bool, |
| 3952 | text: String, |
| 3953 | ) { |
| 3954 | if *thinking_started { |
| 3955 | events.push(StreamEvent::ContentBlockStop { |
| 3956 | index: *content_index, |
| 3957 | }); |
| 3958 | *content_index += 1; |
| 3959 | *thinking_started = false; |
| 3960 | } |
| 3961 | if !*text_started { |
| 3962 | events.push(StreamEvent::ContentBlockStart { |
| 3963 | index: *content_index, |
| 3964 | content_block: ContentBlockStart::Text { |
| 3965 | text: String::new(), |
| 3966 | }, |
| 3967 | }); |
| 3968 | *text_started = true; |
| 3969 | } |
| 3970 | events.push(StreamEvent::ContentBlockDelta { |
| 3971 | index: *content_index, |
| 3972 | delta: Delta::TextDelta { text }, |
| 3973 | }); |
| 3974 | } |
| 3975 | |
| 3976 | fn push_thinking_delta( |
| 3977 | events: &mut Vec<StreamEvent>, |
| 3978 | content_index: &mut u32, |
| 3979 | text_started: &mut bool, |
| 3980 | thinking_started: &mut bool, |
| 3981 | thinking: String, |
| 3982 | ) { |
| 3983 | if *text_started { |
| 3984 | events.push(StreamEvent::ContentBlockStop { |
| 3985 | index: *content_index, |
| 3986 | }); |
| 3987 | *content_index += 1; |
| 3988 | *text_started = false; |
| 3989 | } |
| 3990 | if !*thinking_started { |
| 3991 | events.push(StreamEvent::ContentBlockStart { |
| 3992 | index: *content_index, |
| 3993 | content_block: ContentBlockStart::Thinking { |
| 3994 | thinking: String::new(), |
| 3995 | }, |
| 3996 | }); |
| 3997 | *thinking_started = true; |
| 3998 | } |
| 3999 | events.push(StreamEvent::ContentBlockDelta { |
| 4000 | index: *content_index, |
| 4001 | delta: Delta::ThinkingDelta { thinking }, |
| 4002 | }); |
| 4003 | } |
| 4004 | |
| 4005 | // === SSE Chunk Parser === |
| 4006 | |
| 4007 | enum SseDataFrame { |
| 4008 | Done, |
| 4009 | Events(Vec<StreamEvent>), |
| 4010 | } |
| 4011 | |
| 4012 | // The six `&mut` streaming-state fields plus the style flag are a deliberate, |
| 4013 | // shared parser-state set (mirrored by `parse_sse_chunk*`); bundling them into a |
| 4014 | // struct would only add reborrow noise on this hot SSE path. |
| 4015 | #[allow(clippy::too_many_arguments)] |
| 4016 | fn parse_sse_data_frame( |
| 4017 | data: &str, |
| 4018 | content_index: &mut u32, |
| 4019 | text_started: &mut bool, |
| 4020 | thinking_started: &mut bool, |
| 4021 | tool_indices: &mut std::collections::HashMap<u32, u32>, |
| 4022 | reasoning_detail_buffers: &mut std::collections::HashMap<u32, String>, |
| 4023 | inline_reasoning_tags: &mut InlineReasoningTagState, |
| 4024 | reasoning_stream_style: ReasoningStreamStyle, |
| 4025 | ) -> SseDataFrame { |
| 4026 | if data.trim() == "[DONE]" { |
| 4027 | return SseDataFrame::Done; |
| 4028 | } |
| 4029 | let events = serde_json::from_str::<Value>(data).map_or_else( |
| 4030 | |_| Vec::new(), |
| 4031 | |chunk_json| { |
| 4032 | parse_sse_chunk_with_reasoning_style( |
| 4033 | &chunk_json, |
| 4034 | content_index, |
| 4035 | text_started, |
| 4036 | thinking_started, |
| 4037 | tool_indices, |
| 4038 | reasoning_detail_buffers, |
| 4039 | inline_reasoning_tags, |
| 4040 | reasoning_stream_style, |
| 4041 | ) |
| 4042 | }, |
| 4043 | ); |
| 4044 | SseDataFrame::Events(events) |
| 4045 | } |
| 4046 | |
| 4047 | /// Parse a single SSE chunk from the Chat Completions streaming API into |
| 4048 | /// our internal `StreamEvent` representation. |
| 4049 | #[cfg(test)] |
| 4050 | pub(super) fn parse_sse_chunk( |
| 4051 | chunk: &Value, |
| 4052 | content_index: &mut u32, |
| 4053 | text_started: &mut bool, |
| 4054 | thinking_started: &mut bool, |
| 4055 | tool_indices: &mut std::collections::HashMap<u32, u32>, |
| 4056 | reasoning_detail_buffers: &mut std::collections::HashMap<u32, String>, |
| 4057 | is_reasoning_model: bool, |
| 4058 | ) -> Vec<StreamEvent> { |
| 4059 | let mut inline_reasoning_tags = InlineReasoningTagState::default(); |
| 4060 | let reasoning_stream_style = if is_reasoning_model { |
| 4061 | ReasoningStreamStyle::SeparateField |
| 4062 | } else { |
| 4063 | ReasoningStreamStyle::None |
| 4064 | }; |
| 4065 | parse_sse_chunk_with_reasoning_style( |
| 4066 | chunk, |
| 4067 | content_index, |
| 4068 | text_started, |
| 4069 | thinking_started, |
| 4070 | tool_indices, |
| 4071 | reasoning_detail_buffers, |
| 4072 | &mut inline_reasoning_tags, |
| 4073 | reasoning_stream_style, |
| 4074 | ) |
| 4075 | } |
| 4076 | |
| 4077 | // Same deliberate shared parser-state set as `parse_sse_data_frame`. |
| 4078 | #[allow(clippy::too_many_arguments)] |
| 4079 | fn parse_sse_chunk_with_reasoning_style( |
| 4080 | chunk: &Value, |
| 4081 | content_index: &mut u32, |
| 4082 | text_started: &mut bool, |
| 4083 | thinking_started: &mut bool, |
| 4084 | tool_indices: &mut std::collections::HashMap<u32, u32>, |
| 4085 | reasoning_detail_buffers: &mut std::collections::HashMap<u32, String>, |
| 4086 | inline_reasoning_tags: &mut InlineReasoningTagState, |
| 4087 | reasoning_stream_style: ReasoningStreamStyle, |
| 4088 | ) -> Vec<StreamEvent> { |
| 4089 | let mut events = Vec::new(); |
| 4090 | |
| 4091 | // OpenAI-compatible providers surface mid-stream failures as a chunk-level |
| 4092 | // `error` object (sometimes with `type: "error"`), delivered before |
| 4093 | // `[DONE]`. Silently dropping it turned rate-limit / context-length / |
| 4094 | // server errors into a truncated turn that looked successful — the frame |
| 4095 | // is now surfaced through the same `StreamEvent::Error` contract the |
| 4096 | // Anthropic path uses (#3014, ops R3). |
| 4097 | if let Some(error) = chunk.get("error") { |
| 4098 | let error = match error { |
| 4099 | Value::Object(_) => error.clone(), |
| 4100 | Value::String(message) => serde_json::json!({ "message": message }), |
| 4101 | _ => serde_json::json!({ "message": "provider stream error" }), |
| 4102 | }; |
| 4103 | events.push(StreamEvent::Error { error }); |
| 4104 | return events; |
| 4105 | } |
| 4106 | |
| 4107 | let Some(choices) = chunk.get("choices").and_then(Value::as_array) else { |
| 4108 | // Usage-only chunk (sent at end with stream_options) |
| 4109 | if let Some(usage_val) = chunk.get("usage") { |
| 4110 | let usage = parse_usage(Some(usage_val)); |
| 4111 | events.push(StreamEvent::MessageDelta { |
| 4112 | delta: MessageDelta { |
| 4113 | stop_reason: None, |
| 4114 | stop_sequence: None, |
| 4115 | }, |
| 4116 | usage: Some(usage), |
| 4117 | }); |
| 4118 | } |
| 4119 | return events; |
| 4120 | }; |
| 4121 | |
| 4122 | if choices.is_empty() { |
| 4123 | if let Some(usage_val) = chunk.get("usage") { |
| 4124 | let usage = parse_usage(Some(usage_val)); |
| 4125 | events.push(StreamEvent::MessageDelta { |
| 4126 | delta: MessageDelta { |
| 4127 | stop_reason: None, |
| 4128 | stop_sequence: None, |
| 4129 | }, |
| 4130 | usage: Some(usage), |
| 4131 | }); |
| 4132 | } |
| 4133 | return events; |
| 4134 | } |
| 4135 | |
| 4136 | for choice in choices { |
| 4137 | let choice_index = choice.get("index").and_then(Value::as_u64).unwrap_or(0) as u32; |
| 4138 | let delta = choice.get("delta"); |
| 4139 | let finish_reason = choice |
| 4140 | .get("finish_reason") |
| 4141 | .and_then(Value::as_str) |
| 4142 | .map(str::to_string); |
| 4143 | |
| 4144 | if let Some(delta) = delta { |
| 4145 | let reasoning_text = reasoning_delta(delta, choice_index, reasoning_detail_buffers) |
| 4146 | .filter(|s| !s.is_empty()); |
| 4147 | // Mistral la Plateforme streams reasoning as a polymorphic |
| 4148 | // `delta.content` value: an array of typed {type: thinking|text} |
| 4149 | // blocks while thinking, then a plain string once the final |
| 4150 | // answer starts. Flatten thinking sub-blocks into a single |
| 4151 | // reasoning delta and treat text sub-blocks as normal content |
| 4152 | // before the shared string fallback below. |
| 4153 | let (mistral_thinking, mistral_text) = |
| 4154 | if reasoning_stream_style == ReasoningStreamStyle::MistralBlocks { |
| 4155 | extract_mistral_polymorphic_content(delta) |
| 4156 | } else { |
| 4157 | (None, None) |
| 4158 | }; |
| 4159 | if let Some(reasoning) = mistral_thinking.as_deref() { |
| 4160 | push_thinking_delta( |
| 4161 | &mut events, |
| 4162 | content_index, |
| 4163 | text_started, |
| 4164 | thinking_started, |
| 4165 | reasoning.to_string(), |
| 4166 | ); |
| 4167 | } |
| 4168 | let content_text = mistral_text.or_else(|| { |
| 4169 | delta |
| 4170 | .get("content") |
| 4171 | .and_then(Value::as_str) |
| 4172 | .filter(|s| !s.is_empty()) |
| 4173 | .map(str::to_string) |
| 4174 | }); |
| 4175 | |
| 4176 | // Handle reasoning_content / reasoning thinking deltas. |
| 4177 | if reasoning_stream_style == ReasoningStreamStyle::SeparateField |
| 4178 | && let Some(reasoning) = reasoning_text.as_deref() |
| 4179 | { |
| 4180 | push_thinking_delta( |
| 4181 | &mut events, |
| 4182 | content_index, |
| 4183 | text_started, |
| 4184 | thinking_started, |
| 4185 | reasoning.to_string(), |
| 4186 | ); |
| 4187 | } |
| 4188 | |
| 4189 | // Generic OpenAI-compatible proxies sometimes stream answer text |
| 4190 | // in `reasoning_content`. If this route is configured with no |
| 4191 | // reasoning semantics, render that field as normal text when no |
| 4192 | // `content` delta is present. |
| 4193 | match (content_text, reasoning_stream_style) { |
| 4194 | (Some(content), ReasoningStreamStyle::InlineTags) => { |
| 4195 | for segment in inline_reasoning_segments(&content, inline_reasoning_tags, false) |
| 4196 | { |
| 4197 | match segment { |
| 4198 | ReasoningSegment::Text(text) => push_text_delta( |
| 4199 | &mut events, |
| 4200 | content_index, |
| 4201 | text_started, |
| 4202 | thinking_started, |
| 4203 | text, |
| 4204 | ), |
| 4205 | ReasoningSegment::Thinking(thinking) => push_thinking_delta( |
| 4206 | &mut events, |
| 4207 | content_index, |
| 4208 | text_started, |
| 4209 | thinking_started, |
| 4210 | thinking, |
| 4211 | ), |
| 4212 | } |
| 4213 | } |
| 4214 | } |
| 4215 | (Some(content), _) => push_text_delta( |
| 4216 | &mut events, |
| 4217 | content_index, |
| 4218 | text_started, |
| 4219 | thinking_started, |
| 4220 | content, |
| 4221 | ), |
| 4222 | (None, ReasoningStreamStyle::None) => { |
| 4223 | if let Some(content) = reasoning_text { |
| 4224 | push_text_delta( |
| 4225 | &mut events, |
| 4226 | content_index, |
| 4227 | text_started, |
| 4228 | thinking_started, |
| 4229 | content, |
| 4230 | ); |
| 4231 | } |
| 4232 | } |
| 4233 | (None, _) => {} |
| 4234 | } |
| 4235 | |
| 4236 | // Handle tool calls |
| 4237 | if let Some(tool_calls) = delta.get("tool_calls").and_then(Value::as_array) { |
| 4238 | for tc in tool_calls { |
| 4239 | let tc_index = tc.get("index").and_then(Value::as_u64).unwrap_or(0) as u32; |
| 4240 | let tool_block_index = match tool_indices.entry(tc_index) { |
| 4241 | std::collections::hash_map::Entry::Occupied(entry) => *entry.get(), |
| 4242 | std::collections::hash_map::Entry::Vacant(entry) => { |
| 4243 | // Close text block if transitioning to tool use |
| 4244 | if *text_started { |
| 4245 | events.push(StreamEvent::ContentBlockStop { |
| 4246 | index: *content_index, |
| 4247 | }); |
| 4248 | *content_index += 1; |
| 4249 | *text_started = false; |
| 4250 | } |
| 4251 | if *thinking_started { |
| 4252 | events.push(StreamEvent::ContentBlockStop { |
| 4253 | index: *content_index, |
| 4254 | }); |
| 4255 | *content_index += 1; |
| 4256 | *thinking_started = false; |
| 4257 | } |
| 4258 | |
| 4259 | let block_index = *content_index; |
| 4260 | let id = tc |
| 4261 | .get("id") |
| 4262 | .and_then(Value::as_str) |
| 4263 | .map(str::to_string) |
| 4264 | // Some upstream gateways (and the responses-API |
| 4265 | // bridge) elide the `id` on the first chunk of a |
| 4266 | // tool call. Falling back to a constant string |
| 4267 | // collides when the model emits parallel tool |
| 4268 | // calls in the same delta — every call ended up |
| 4269 | // with the same id and downstream tool-result |
| 4270 | // routing matched the first one twice. Index by |
| 4271 | // the content-block position to keep the |
| 4272 | // fallback unique within the response. |
| 4273 | .unwrap_or_else(|| format!("call_{block_index}")); |
| 4274 | let name = tc |
| 4275 | .get("function") |
| 4276 | .and_then(|f| f.get("name")) |
| 4277 | .and_then(Value::as_str); |
| 4278 | let name = tool_name_or_fallback(name, &id, "Streaming response chunk"); |
| 4279 | let caller = tc.get("caller").and_then(|v| { |
| 4280 | v.get("type").and_then(Value::as_str).map(|caller_type| { |
| 4281 | ToolCaller { |
| 4282 | caller_type: caller_type.to_string(), |
| 4283 | tool_id: v |
| 4284 | .get("tool_id") |
| 4285 | .and_then(Value::as_str) |
| 4286 | .map(std::string::ToString::to_string), |
| 4287 | } |
| 4288 | }) |
| 4289 | }); |
| 4290 | |
| 4291 | let thought_signature = tc |
| 4292 | .pointer("/extra_content/google/thought_signature") |
| 4293 | .and_then(Value::as_str) |
| 4294 | .map(str::to_string); |
| 4295 | events.push(StreamEvent::ContentBlockStart { |
| 4296 | index: block_index, |
| 4297 | content_block: ContentBlockStart::ToolUse { |
| 4298 | id, |
| 4299 | name: from_api_tool_name(&name), |
| 4300 | input: json!({}), |
| 4301 | caller, |
| 4302 | thought_signature, |
| 4303 | }, |
| 4304 | }); |
| 4305 | *content_index = (*content_index).saturating_add(1); |
| 4306 | entry.insert(block_index); |
| 4307 | block_index |
| 4308 | } |
| 4309 | }; |
| 4310 | |
| 4311 | // Stream tool call arguments |
| 4312 | if let Some(args) = tc |
| 4313 | .get("function") |
| 4314 | .and_then(|f| f.get("arguments")) |
| 4315 | .and_then(Value::as_str) |
| 4316 | && !args.is_empty() |
| 4317 | { |
| 4318 | events.push(StreamEvent::ContentBlockDelta { |
| 4319 | index: tool_block_index, |
| 4320 | delta: Delta::InputJsonDelta { |
| 4321 | partial_json: args.to_string(), |
| 4322 | }, |
| 4323 | }); |
| 4324 | } |
| 4325 | } |
| 4326 | } |
| 4327 | } |
| 4328 | |
| 4329 | // Handle finish reason |
| 4330 | if let Some(reason) = finish_reason { |
| 4331 | if reasoning_stream_style == ReasoningStreamStyle::InlineTags { |
| 4332 | for segment in inline_reasoning_segments("", inline_reasoning_tags, true) { |
| 4333 | match segment { |
| 4334 | ReasoningSegment::Text(text) => push_text_delta( |
| 4335 | &mut events, |
| 4336 | content_index, |
| 4337 | text_started, |
| 4338 | thinking_started, |
| 4339 | text, |
| 4340 | ), |
| 4341 | ReasoningSegment::Thinking(thinking) => push_thinking_delta( |
| 4342 | &mut events, |
| 4343 | content_index, |
| 4344 | text_started, |
| 4345 | thinking_started, |
| 4346 | thinking, |
| 4347 | ), |
| 4348 | } |
| 4349 | } |
| 4350 | } |
| 4351 | // Close any open blocks |
| 4352 | if *text_started { |
| 4353 | events.push(StreamEvent::ContentBlockStop { |
| 4354 | index: *content_index, |
| 4355 | }); |
| 4356 | *text_started = false; |
| 4357 | } |
| 4358 | if *thinking_started { |
| 4359 | events.push(StreamEvent::ContentBlockStop { |
| 4360 | index: *content_index, |
| 4361 | }); |
| 4362 | *thinking_started = false; |
| 4363 | } |
| 4364 | // Close tool blocks |
| 4365 | let mut open_tool_indices: Vec<u32> = |
| 4366 | tool_indices.drain().map(|(_, idx)| idx).collect(); |
| 4367 | open_tool_indices.sort_unstable(); |
| 4368 | for tool_block_index in open_tool_indices { |
| 4369 | events.push(StreamEvent::ContentBlockStop { |
| 4370 | index: tool_block_index, |
| 4371 | }); |
| 4372 | } |
| 4373 | |
| 4374 | // Emit usage from the chunk if available |
| 4375 | let chunk_usage = chunk.get("usage").map(|u| parse_usage(Some(u))); |
| 4376 | events.push(StreamEvent::MessageDelta { |
| 4377 | delta: MessageDelta { |
| 4378 | stop_reason: Some(reason), |
| 4379 | stop_sequence: None, |
| 4380 | }, |
| 4381 | usage: chunk_usage, |
| 4382 | }); |
| 4383 | } |
| 4384 | } |
| 4385 | |
| 4386 | events |
| 4387 | } |
| 4388 | |
| 4389 | fn tool_name_or_fallback(name: Option<&str>, id: &str, source: &str) -> String { |
| 4390 | let trimmed = name.unwrap_or("").trim(); |
| 4391 | if trimmed.is_empty() { |
| 4392 | logging::warn(format!( |
| 4393 | "{source} returned an empty tool name for call {id}; using unknown_tool" |
| 4394 | )); |
| 4395 | "unknown_tool".to_string() |
| 4396 | } else { |
| 4397 | trimmed.to_string() |
| 4398 | } |
| 4399 | } |
| 4400 | |
| 4401 | // === #103 Phase 1: stream-decode diagnostics =================================== |
| 4402 | |
| 4403 | #[cfg(test)] |
| 4404 | mod stream_diagnostics_tests { |
| 4405 | use super::*; |
| 4406 | use reqwest::header::{HeaderMap, HeaderValue}; |
| 4407 | |
| 4408 | #[test] |
| 4409 | fn stream_idle_timeout_reports_progress_and_timing() { |
| 4410 | let message = super::super::stream_entry::idle_timeout_message( |
| 4411 | Duration::from_secs(240), |
| 4412 | 8192, |
| 4413 | Duration::from_millis(73_500), |
| 4414 | Duration::from_millis(41_250), |
| 4415 | ); |
| 4416 | |
| 4417 | assert_eq!( |
| 4418 | message, |
| 4419 | "SSE stream idle timeout after 240s — no data received \ |
| 4420 | (bytes_received=8192, stream_age_ms=73500, ms_since_last_chunk=41250)" |
| 4421 | ); |
| 4422 | } |
| 4423 | |
| 4424 | #[test] |
| 4425 | fn chat_completions_error_frames_surface_as_stream_events() { |
| 4426 | let mut content_index = 0u32; |
| 4427 | let mut text_started = false; |
| 4428 | let mut thinking_started = false; |
| 4429 | let mut tool_indices = std::collections::HashMap::new(); |
| 4430 | let mut reasoning_buffers = std::collections::HashMap::new(); |
| 4431 | for chunk in [ |
| 4432 | json!({ "error": { "message": "rate limit exceeded", "type": "rate_limit_error" } }), |
| 4433 | json!({ "type": "error", "error": { "message": "context length exceeded" } }), |
| 4434 | json!({ "error": "server error" }), |
| 4435 | ] { |
| 4436 | let events = parse_sse_chunk( |
| 4437 | &chunk, |
| 4438 | &mut content_index, |
| 4439 | &mut text_started, |
| 4440 | &mut thinking_started, |
| 4441 | &mut tool_indices, |
| 4442 | &mut reasoning_buffers, |
| 4443 | false, |
| 4444 | ); |
| 4445 | assert_eq!( |
| 4446 | events.len(), |
| 4447 | 1, |
| 4448 | "a chunk-level error frame must not be swallowed ({chunk})" |
| 4449 | ); |
| 4450 | match &events[0] { |
| 4451 | StreamEvent::Error { error } => assert!( |
| 4452 | !error.is_null() |
| 4453 | && (error.get("message").and_then(Value::as_str).is_some() |
| 4454 | || error.is_string()), |
| 4455 | "the provider error message must survive parsing: {error}" |
| 4456 | ), |
| 4457 | other => panic!("expected StreamEvent::Error, got {other:?}"), |
| 4458 | } |
| 4459 | } |
| 4460 | // A normal content chunk still parses as a delta after the error path. |
| 4461 | let deltas = parse_sse_chunk( |
| 4462 | &json!({"choices": [{"index": 0, "delta": {"content": "ok"}}]}), |
| 4463 | &mut content_index, |
| 4464 | &mut text_started, |
| 4465 | &mut thinking_started, |
| 4466 | &mut tool_indices, |
| 4467 | &mut reasoning_buffers, |
| 4468 | false, |
| 4469 | ); |
| 4470 | assert!( |
| 4471 | deltas.iter().any(|event| { |
| 4472 | matches!( |
| 4473 | event, |
| 4474 | StreamEvent::ContentBlockDelta { |
| 4475 | delta: Delta::TextDelta { text }, .. |
| 4476 | } if text == "ok" |
| 4477 | ) |
| 4478 | }), |
| 4479 | "content deltas still parse: {deltas:?}" |
| 4480 | ); |
| 4481 | } |
| 4482 | |
| 4483 | #[test] |
| 4484 | fn deepseek_thinking_omits_tool_choice() { |
| 4485 | for effort in [Some("high"), Some("max"), Some("medium"), Some("")] { |
| 4486 | assert!( |
| 4487 | !should_send_tool_choice_for_chat(ProviderKind::Deepseek, effort), |
| 4488 | "DeepSeek thinking rejects explicit tool_choice for {effort:?}" |
| 4489 | ); |
| 4490 | assert!( |
| 4491 | !should_send_tool_choice_for_chat(ProviderKind::Deepseek, effort), |
| 4492 | "DeepSeek CN thinking rejects explicit tool_choice for {effort:?}" |
| 4493 | ); |
| 4494 | } |
| 4495 | |
| 4496 | for effort in [ |
| 4497 | None, |
| 4498 | Some("off"), |
| 4499 | Some("disabled"), |
| 4500 | Some("none"), |
| 4501 | Some("false"), |
| 4502 | ] { |
| 4503 | assert!(should_send_tool_choice_for_chat( |
| 4504 | ProviderKind::Deepseek, |
| 4505 | effort |
| 4506 | )); |
| 4507 | } |
| 4508 | assert!(should_send_tool_choice_for_chat( |
| 4509 | ProviderKind::Openrouter, |
| 4510 | Some("high") |
| 4511 | )); |
| 4512 | } |
| 4513 | |
| 4514 | #[test] |
| 4515 | fn format_stream_headers_renders_all_fields_when_present() { |
| 4516 | let mut headers = HeaderMap::new(); |
| 4517 | headers.insert("content-encoding", HeaderValue::from_static("gzip")); |
| 4518 | headers.insert("transfer-encoding", HeaderValue::from_static("chunked")); |
| 4519 | headers.insert("connection", HeaderValue::from_static("keep-alive")); |
| 4520 | headers.insert("server", HeaderValue::from_static("openresty/1.25.3.1")); |
| 4521 | |
| 4522 | let rendered = format_stream_headers(&headers); |
| 4523 | // Order is fixed by FIELDS in the helper; assert each field appears. |
| 4524 | assert!( |
| 4525 | rendered.contains("content-encoding=gzip"), |
| 4526 | "got: {rendered}" |
| 4527 | ); |
| 4528 | assert!( |
| 4529 | rendered.contains("transfer-encoding=chunked"), |
| 4530 | "got: {rendered}" |
| 4531 | ); |
| 4532 | assert!( |
| 4533 | rendered.contains("connection=keep-alive"), |
| 4534 | "got: {rendered}" |
| 4535 | ); |
| 4536 | assert!( |
| 4537 | rendered.contains("server=openresty/1.25.3.1"), |
| 4538 | "got: {rendered}" |
| 4539 | ); |
| 4540 | } |
| 4541 | |
| 4542 | #[test] |
| 4543 | fn format_stream_headers_marks_missing_fields_as_absent() { |
| 4544 | // DeepSeek frequently omits content-encoding when not compressing. |
| 4545 | // The diagnostic must still produce a parseable line so log scrapers |
| 4546 | // don't lose the slot. |
| 4547 | let headers = HeaderMap::new(); |
| 4548 | let rendered = format_stream_headers(&headers); |
| 4549 | assert!( |
| 4550 | rendered.contains("content-encoding=(absent)"), |
| 4551 | "missing field must be explicitly marked; got: {rendered}" |
| 4552 | ); |
| 4553 | assert!( |
| 4554 | rendered.contains("transfer-encoding=(absent)"), |
| 4555 | "missing field must be explicitly marked; got: {rendered}" |
| 4556 | ); |
| 4557 | } |
| 4558 | |
| 4559 | #[test] |
| 4560 | fn format_stream_headers_handles_non_ascii_value_gracefully() { |
| 4561 | // If a header value isn't UTF-8, `.to_str()` fails — we must not panic |
| 4562 | // and should still produce a parseable line. |
| 4563 | let mut headers = HeaderMap::new(); |
| 4564 | // 0xFF is a valid byte but invalid UTF-8 start byte. |
| 4565 | headers.insert( |
| 4566 | "server", |
| 4567 | HeaderValue::from_bytes(b"\xff\xfemystery").expect("header value"), |
| 4568 | ); |
| 4569 | let rendered = format_stream_headers(&headers); |
| 4570 | assert!( |
| 4571 | rendered.contains("server=(absent)"), |
| 4572 | "non-UTF8 header values fall back to (absent); got: {rendered}" |
| 4573 | ); |
| 4574 | } |
| 4575 | } |
| 4576 | |
| 4577 | #[cfg(test)] |
| 4578 | mod arcee_waf_message_encoding_tests { |
| 4579 | use super::build_chat_messages_for_request_and_provider; |
| 4580 | use crate::config::ProviderKind; |
| 4581 | use codewhale_models::{MessageRequest, SystemPrompt}; |
| 4582 | use serde_json::Value; |
| 4583 | |
| 4584 | fn request_with_system(system: &str) -> MessageRequest { |
| 4585 | MessageRequest { |
| 4586 | model: "trinity-large-thinking".to_string(), |
| 4587 | messages: Vec::new(), |
| 4588 | max_tokens: 16, |
| 4589 | system: Some(SystemPrompt::Text(system.to_string())), |
| 4590 | tools: None, |
| 4591 | tool_choice: None, |
| 4592 | metadata: None, |
| 4593 | thinking: None, |
| 4594 | reasoning_effort: None, |
| 4595 | stream: None, |
| 4596 | temperature: None, |
| 4597 | top_p: None, |
| 4598 | } |
| 4599 | } |
| 4600 | |
| 4601 | fn decoded_content(content: &Value) -> String { |
| 4602 | if let Some(text) = content.as_str() { |
| 4603 | return text.to_string(); |
| 4604 | } |
| 4605 | content |
| 4606 | .as_array() |
| 4607 | .expect("content parts") |
| 4608 | .iter() |
| 4609 | .map(|part| part.get("text").and_then(Value::as_str).expect("text part")) |
| 4610 | .collect() |
| 4611 | } |
| 4612 | |
| 4613 | #[test] |
| 4614 | fn arcee_splits_waf_trigger_without_changing_decoded_system_prompt() { |
| 4615 | let system = "Run calculations with `python -c 'print(1)'` when a tool is available."; |
| 4616 | let request = request_with_system(system); |
| 4617 | |
| 4618 | let messages = build_chat_messages_for_request_and_provider(&request, ProviderKind::Arcee); |
| 4619 | let content = &messages[0]["content"]; |
| 4620 | |
| 4621 | assert!( |
| 4622 | content.is_array(), |
| 4623 | "Arcee system content with a WAF trigger should be encoded as text parts" |
| 4624 | ); |
| 4625 | assert_eq!(decoded_content(content), system); |
| 4626 | let serialized = serde_json::to_string(&messages).expect("serialize messages"); |
| 4627 | assert!( |
| 4628 | !serialized.contains("python -c"), |
| 4629 | "wire JSON should not contain the Cloudflare trigger contiguously: {serialized}" |
| 4630 | ); |
| 4631 | } |
| 4632 | |
| 4633 | #[test] |
| 4634 | fn non_arcee_providers_keep_system_prompt_as_string() { |
| 4635 | let system = "Run calculations with `python -c 'print(1)'` when a tool is available."; |
| 4636 | let request = request_with_system(system); |
| 4637 | |
| 4638 | let messages = build_chat_messages_for_request_and_provider(&request, ProviderKind::Openai); |
| 4639 | |
| 4640 | assert_eq!(messages[0]["content"].as_str(), Some(system)); |
| 4641 | } |
| 4642 | |
| 4643 | #[test] |
| 4644 | fn arcee_keeps_non_triggering_system_prompt_as_string() { |
| 4645 | let system = "Use read-only tools to inspect files before reporting results."; |
| 4646 | let request = request_with_system(system); |
| 4647 | |
| 4648 | let messages = build_chat_messages_for_request_and_provider(&request, ProviderKind::Arcee); |
| 4649 | |
| 4650 | assert_eq!(messages[0]["content"].as_str(), Some(system)); |
| 4651 | } |
| 4652 | } |
| 4653 | |
| 4654 | #[cfg(test)] |
| 4655 | mod minimax_reasoning_replay_tests { |
| 4656 | use super::{ |
| 4657 | build_chat_messages_for_request_and_provider, |
| 4658 | build_chat_messages_for_request_and_provider_and_route, |
| 4659 | }; |
| 4660 | use crate::config::{ |
| 4661 | DEFAULT_KIMI_CODE_BASE_URL, DEFAULT_MINIMAX_MODEL, DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 4662 | DEFAULT_MOONSHOT_BASE_URL, KIMI_CODE_K3_MODEL, ProviderKind, |
| 4663 | }; |
| 4664 | use codewhale_models::Role; |
| 4665 | use codewhale_models::{ContentBlock, Message, MessageRequest}; |
| 4666 | |
| 4667 | fn request_with_assistant_thinking() -> MessageRequest { |
| 4668 | MessageRequest { |
| 4669 | model: DEFAULT_MINIMAX_MODEL.to_string(), |
| 4670 | messages: vec![Message { |
| 4671 | role: Role::Assistant, |
| 4672 | content: vec![ |
| 4673 | ContentBlock::Thinking { |
| 4674 | thinking: "Inspect tool state".to_string(), |
| 4675 | signature: None, |
| 4676 | state: None, |
| 4677 | }, |
| 4678 | ContentBlock::Text { |
| 4679 | text: "Done.".to_string(), |
| 4680 | cache_control: None, |
| 4681 | }, |
| 4682 | ], |
| 4683 | }], |
| 4684 | max_tokens: 16, |
| 4685 | system: None, |
| 4686 | tools: None, |
| 4687 | tool_choice: None, |
| 4688 | metadata: None, |
| 4689 | thinking: None, |
| 4690 | reasoning_effort: None, |
| 4691 | stream: None, |
| 4692 | temperature: None, |
| 4693 | top_p: None, |
| 4694 | } |
| 4695 | } |
| 4696 | |
| 4697 | #[test] |
| 4698 | fn minimax_history_replays_thinking_as_reasoning_details() { |
| 4699 | let request = request_with_assistant_thinking(); |
| 4700 | |
| 4701 | let messages = |
| 4702 | build_chat_messages_for_request_and_provider(&request, ProviderKind::Minimax); |
| 4703 | let assistant = &messages[0]; |
| 4704 | |
| 4705 | assert_eq!( |
| 4706 | assistant |
| 4707 | .get("reasoning_content") |
| 4708 | .and_then(|value| value.as_str()), |
| 4709 | Some("Inspect tool state") |
| 4710 | ); |
| 4711 | assert_eq!( |
| 4712 | assistant |
| 4713 | .pointer("/reasoning_details/0/type") |
| 4714 | .and_then(|value| value.as_str()), |
| 4715 | Some("text") |
| 4716 | ); |
| 4717 | assert_eq!( |
| 4718 | assistant |
| 4719 | .pointer("/reasoning_details/0/text") |
| 4720 | .and_then(|value| value.as_str()), |
| 4721 | Some("Inspect tool state") |
| 4722 | ); |
| 4723 | } |
| 4724 | |
| 4725 | #[test] |
| 4726 | fn kimi_code_k3_replays_thinking_only_on_the_exact_membership_route() { |
| 4727 | let mut request = request_with_assistant_thinking(); |
| 4728 | request.model = KIMI_CODE_K3_MODEL.to_string(); |
| 4729 | |
| 4730 | let exact = build_chat_messages_for_request_and_provider_and_route( |
| 4731 | &request, |
| 4732 | ProviderKind::Moonshot, |
| 4733 | DEFAULT_KIMI_CODE_BASE_URL, |
| 4734 | ); |
| 4735 | assert_eq!( |
| 4736 | exact[0] |
| 4737 | .get("reasoning_content") |
| 4738 | .and_then(serde_json::Value::as_str), |
| 4739 | Some("Inspect tool state") |
| 4740 | ); |
| 4741 | |
| 4742 | let neighbor = build_chat_messages_for_request_and_provider_and_route( |
| 4743 | &request, |
| 4744 | ProviderKind::Moonshot, |
| 4745 | DEFAULT_MOONSHOT_BASE_URL, |
| 4746 | ); |
| 4747 | assert!( |
| 4748 | neighbor[0].get("reasoning_content").is_none(), |
| 4749 | "a generic Moonshot k3 identifier must not inherit Kimi Code replay" |
| 4750 | ); |
| 4751 | } |
| 4752 | |
| 4753 | #[test] |
| 4754 | fn modelstudio_qwen38_wire_body_replays_no_historical_reasoning_across_tool_loop() { |
| 4755 | // One user handoff, one assistant Thinking + Text + ToolUse turn, and |
| 4756 | // its matching ToolResult. On the exact Model Studio route the |
| 4757 | // always-thinking qwen3.8 family must not receive historical |
| 4758 | // `reasoning_content`, while the handoff occurs exactly once and the |
| 4759 | // tool call id, arguments, and result stay intact. |
| 4760 | let mut request = request_with_assistant_thinking(); |
| 4761 | request.model = "qwen3.8-max".to_string(); |
| 4762 | request.messages = vec![ |
| 4763 | Message { |
| 4764 | role: Role::User, |
| 4765 | content: vec![ContentBlock::Text { |
| 4766 | text: "HANDOFF-SENTINEL: fix the widget.".to_string(), |
| 4767 | cache_control: None, |
| 4768 | }], |
| 4769 | }, |
| 4770 | Message { |
| 4771 | role: Role::Assistant, |
| 4772 | content: vec![ |
| 4773 | ContentBlock::Thinking { |
| 4774 | thinking: "stale thinking from the prior turn".to_string(), |
| 4775 | signature: None, |
| 4776 | state: None, |
| 4777 | }, |
| 4778 | ContentBlock::Text { |
| 4779 | text: "I'll read the widget first.".to_string(), |
| 4780 | cache_control: None, |
| 4781 | }, |
| 4782 | ContentBlock::ToolUse { |
| 4783 | execution_id: None, |
| 4784 | id: "call_qwen38_001".to_string(), |
| 4785 | name: "read".to_string(), |
| 4786 | input: serde_json::json!({ "path": "widget.rs" }), |
| 4787 | caller: None, |
| 4788 | thought_signature: None, |
| 4789 | }, |
| 4790 | ], |
| 4791 | }, |
| 4792 | Message { |
| 4793 | role: Role::User, |
| 4794 | content: vec![ContentBlock::ToolResult { |
| 4795 | execution_id: None, |
| 4796 | tool_use_id: "call_qwen38_001".to_string(), |
| 4797 | content: "widget.rs: struct Widget { .. }".to_string(), |
| 4798 | is_error: None, |
| 4799 | content_blocks: None, |
| 4800 | }], |
| 4801 | }, |
| 4802 | ]; |
| 4803 | |
| 4804 | let messages = build_chat_messages_for_request_and_provider_and_route( |
| 4805 | &request, |
| 4806 | ProviderKind::ModelstudioTokenPlan, |
| 4807 | DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 4808 | ); |
| 4809 | |
| 4810 | // The handoff occurs exactly once, as a plain user text message. |
| 4811 | let handoff_carriers: Vec<&serde_json::Value> = messages |
| 4812 | .iter() |
| 4813 | .filter(|message| { |
| 4814 | message.get("role").and_then(serde_json::Value::as_str) == Some("user") |
| 4815 | && message |
| 4816 | .get("content") |
| 4817 | .and_then(serde_json::Value::as_str) |
| 4818 | .is_some_and(|content| content.contains("HANDOFF-SENTINEL")) |
| 4819 | }) |
| 4820 | .collect(); |
| 4821 | assert_eq!( |
| 4822 | handoff_carriers.len(), |
| 4823 | 1, |
| 4824 | "the handoff must appear in exactly one user message: {messages:?}" |
| 4825 | ); |
| 4826 | |
| 4827 | // The assistant turn keeps its text and tool call, but no historical |
| 4828 | // reasoning_content for qwen3.8. |
| 4829 | let assistant = messages |
| 4830 | .iter() |
| 4831 | .find(|message| { |
| 4832 | message.get("role").and_then(serde_json::Value::as_str) == Some("assistant") |
| 4833 | }) |
| 4834 | .expect("assistant message"); |
| 4835 | assert_eq!( |
| 4836 | assistant.get("content").and_then(serde_json::Value::as_str), |
| 4837 | Some("I'll read the widget first.") |
| 4838 | ); |
| 4839 | assert!( |
| 4840 | assistant.get("reasoning_content").is_none(), |
| 4841 | "qwen3.8 must not receive historical reasoning_content: {assistant:?}" |
| 4842 | ); |
| 4843 | let tool_calls = assistant |
| 4844 | .get("tool_calls") |
| 4845 | .and_then(serde_json::Value::as_array) |
| 4846 | .expect("tool_calls array"); |
| 4847 | assert_eq!(tool_calls.len(), 1); |
| 4848 | assert_eq!(tool_calls[0]["id"], serde_json::json!("call_qwen38_001")); |
| 4849 | assert_eq!( |
| 4850 | tool_calls[0].pointer("/function/name"), |
| 4851 | Some(&serde_json::json!("read")) |
| 4852 | ); |
| 4853 | assert_eq!( |
| 4854 | tool_calls[0].pointer("/function/arguments"), |
| 4855 | Some(&serde_json::json!(r#"{"path":"widget.rs"}"#)) |
| 4856 | ); |
| 4857 | |
| 4858 | // The matching tool result rides along under its original id. |
| 4859 | let tool_result = messages |
| 4860 | .iter() |
| 4861 | .find(|message| message.get("role").and_then(serde_json::Value::as_str) == Some("tool")) |
| 4862 | .expect("tool result message"); |
| 4863 | assert_eq!( |
| 4864 | tool_result.get("tool_call_id"), |
| 4865 | Some(&serde_json::json!("call_qwen38_001")) |
| 4866 | ); |
| 4867 | assert_eq!( |
| 4868 | tool_result |
| 4869 | .get("content") |
| 4870 | .and_then(serde_json::Value::as_str), |
| 4871 | Some("widget.rs: struct Widget { .. }") |
| 4872 | ); |
| 4873 | |
| 4874 | // Control: the same handoff/tool loop with a documented |
| 4875 | // preserve-thinking model replays its historical reasoning, proving |
| 4876 | // the strip above is model-gated, not route-gated. |
| 4877 | let mut preserve_request = request; |
| 4878 | preserve_request.model = "qwen3.7-plus".to_string(); |
| 4879 | let preserve_messages = build_chat_messages_for_request_and_provider_and_route( |
| 4880 | &preserve_request, |
| 4881 | ProviderKind::ModelstudioTokenPlan, |
| 4882 | DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 4883 | ); |
| 4884 | let preserve_assistant = preserve_messages |
| 4885 | .iter() |
| 4886 | .find(|message| { |
| 4887 | message.get("role").and_then(serde_json::Value::as_str) == Some("assistant") |
| 4888 | }) |
| 4889 | .expect("assistant message"); |
| 4890 | assert_eq!( |
| 4891 | preserve_assistant |
| 4892 | .get("reasoning_content") |
| 4893 | .and_then(serde_json::Value::as_str), |
| 4894 | Some("stale thinking from the prior turn"), |
| 4895 | "documented preserve-thinking models keep replaying history" |
| 4896 | ); |
| 4897 | } |
| 4898 | } |
| 4899 | |
| 4900 | // === #103 Phase 4: SSE decoder behavior on canned chunk sequences ============ |
| 4901 | |
| 4902 | #[cfg(test)] |
| 4903 | #[path = "chat/tests/stream_decoder.rs"] |
| 4904 | mod stream_decoder_tests; |
| 4905 | |
| 4906 | #[cfg(test)] |
| 4907 | mod alias_thinking_detection_tests { |
| 4908 | //! Regression coverage for the DeepSeek public model aliases. |
| 4909 | //! |
| 4910 | //! `deepseek-chat` and `deepseek-reasoner` are the canonical alias names |
| 4911 | //! published in DeepSeek's API docs. Server-side they resolve to V4-flash |
| 4912 | //! and V4-pro respectively, both of which have thinking mode enabled by |
| 4913 | //! default. If the TUI does not classify those aliases as reasoning |
| 4914 | //! models, the sanitizer skips replaying `reasoning_content` on tool-call |
| 4915 | //! assistant messages and DeepSeek returns a 400 ("the `reasoning_content` |
| 4916 | //! in the thinking mode must be passed back to the API") on the second |
| 4917 | //! turn. See upstream API docs: |
| 4918 | //! <https://api-docs.deepseek.com/guides/thinking_mode> |
| 4919 | use super::{ |
| 4920 | ReasoningStreamStyle, apply_direct_moonshot_k3_fixed_sampling, |
| 4921 | apply_inkling_reasoning_effort, apply_kimi_code_fixed_sampling, |
| 4922 | apply_kimi_code_k3_reasoning_effort, apply_openai_reasoning_effort, |
| 4923 | apply_provider_token_limit, apply_route_reasoning_controls, is_reasoning_model_for_stream, |
| 4924 | provider_accepts_reasoning_content, reasoning_stream_style_for_route, |
| 4925 | requires_reasoning_content, should_replay_reasoning_content, |
| 4926 | should_replay_reasoning_content_for_provider, |
| 4927 | should_replay_reasoning_content_for_provider_on_route, |
| 4928 | }; |
| 4929 | use crate::config::ProviderKind; |
| 4930 | use serde_json::json; |
| 4931 | |
| 4932 | #[test] |
| 4933 | fn aliases_routed_to_v4_require_reasoning_content() { |
| 4934 | // Documented public aliases. |
| 4935 | assert!(requires_reasoning_content("deepseek-chat")); |
| 4936 | assert!(requires_reasoning_content("deepseek-reasoner")); |
| 4937 | // Case-insensitive: users sometimes copy/paste with capitalisation. |
| 4938 | assert!(requires_reasoning_content("DeepSeek-Chat")); |
| 4939 | assert!(requires_reasoning_content("DEEPSEEK-REASONER")); |
| 4940 | } |
| 4941 | |
| 4942 | #[test] |
| 4943 | fn explicit_v4_ids_still_require_reasoning_content() { |
| 4944 | // Direct V4 IDs continue to match (regression guard for the existing |
| 4945 | // `lower.contains("deepseek-v4")` branch). |
| 4946 | assert!(requires_reasoning_content("deepseek-v4-flash")); |
| 4947 | assert!(requires_reasoning_content("deepseek-v4-pro")); |
| 4948 | } |
| 4949 | |
| 4950 | #[test] |
| 4951 | fn non_thinking_aliases_remain_excluded() { |
| 4952 | // Legacy non-thinking IDs and unrelated provider models must not be |
| 4953 | // misclassified, otherwise we would force a placeholder |
| 4954 | // `reasoning_content` on providers that reject the field. |
| 4955 | assert!(!requires_reasoning_content("deepseek-v3")); |
| 4956 | assert!(!requires_reasoning_content("deepseek-coder")); |
| 4957 | assert!(!requires_reasoning_content("qwen3-coder")); |
| 4958 | assert!(!requires_reasoning_content("claude-sonnet-4-6")); |
| 4959 | } |
| 4960 | |
| 4961 | #[test] |
| 4962 | fn alias_prefix_handles_suffixed_variants() { |
| 4963 | // OpenRouter / proxy deployments occasionally suffix the canonical |
| 4964 | // alias (e.g. `deepseek-chat:free`). Those routes still hit V4 |
| 4965 | // server-side, so they must continue to require reasoning_content. |
| 4966 | assert!(requires_reasoning_content("deepseek-chat:free")); |
| 4967 | assert!(requires_reasoning_content("deepseek-reasoner-2025-05")); |
| 4968 | } |
| 4969 | |
| 4970 | #[test] |
| 4971 | fn explicit_reasoning_off_overrides_alias_detection() { |
| 4972 | // `reasoning_effort = "off"` is the documented escape hatch: even when |
| 4973 | // the model is in the thinking family, the user can opt out and the |
| 4974 | // sanitizer must respect that choice. |
| 4975 | assert!(!should_replay_reasoning_content( |
| 4976 | "deepseek-chat", |
| 4977 | Some("off") |
| 4978 | )); |
| 4979 | assert!(!should_replay_reasoning_content( |
| 4980 | "deepseek-reasoner", |
| 4981 | Some("disabled") |
| 4982 | )); |
| 4983 | // Without an explicit override, alias models still trigger replay. |
| 4984 | assert!(should_replay_reasoning_content("deepseek-chat", None)); |
| 4985 | assert!(should_replay_reasoning_content( |
| 4986 | "deepseek-reasoner", |
| 4987 | Some("medium") |
| 4988 | )); |
| 4989 | } |
| 4990 | |
| 4991 | #[test] |
| 4992 | fn generic_openai_provider_does_not_accept_reasoning_content_semantics() { |
| 4993 | assert!(!provider_accepts_reasoning_content(ProviderKind::Openai)); |
| 4994 | assert!(provider_accepts_reasoning_content(ProviderKind::Deepseek)); |
| 4995 | assert!(provider_accepts_reasoning_content(ProviderKind::NvidiaNim)); |
| 4996 | assert!(provider_accepts_reasoning_content(ProviderKind::XiaomiMimo)); |
| 4997 | assert!(provider_accepts_reasoning_content(ProviderKind::Arcee)); |
| 4998 | assert!(provider_accepts_reasoning_content(ProviderKind::Minimax)); |
| 4999 | assert!(provider_accepts_reasoning_content(ProviderKind::Zai)); |
| 5000 | // #3016: Moonshot's native endpoint streams Kimi thinking as |
| 5001 | // reasoning_content. |
| 5002 | assert!(provider_accepts_reasoning_content(ProviderKind::Moonshot)); |
| 5003 | } |
| 5004 | |
| 5005 | /// Alibaba's classic pay-as-you-go DashScope endpoints are genuine |
| 5006 | /// Alibaba Chat Completions hosts serving the same models; before |
| 5007 | /// 2026-08-04 they were missing from the verifier allowlist, so every |
| 5008 | /// reasoning control was silently stripped there (fail-closed feature |
| 5009 | /// loss, not a leak). The intl spelling matches provider_defaults. |
| 5010 | #[test] |
| 5011 | fn classic_dashscope_hosts_are_verified_modelstudio_chat_routes() { |
| 5012 | for base_url in [ |
| 5013 | "https://dashscope.aliyuncs.com/compatible-mode/v1", |
| 5014 | "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", |
| 5015 | "https://dashscope-intl.aliyuncs.com/compatible-mode/v1/", |
| 5016 | ] { |
| 5017 | assert!( |
| 5018 | super::is_exact_modelstudio_chat_route( |
| 5019 | ProviderKind::ModelstudioTokenPlan, |
| 5020 | base_url |
| 5021 | ), |
| 5022 | "{base_url}" |
| 5023 | ); |
| 5024 | let mut body = json!({}); |
| 5025 | apply_route_reasoning_controls( |
| 5026 | &mut body, |
| 5027 | ProviderKind::ModelstudioTokenPlan, |
| 5028 | base_url, |
| 5029 | "qwen3.7-plus", |
| 5030 | Some("off"), |
| 5031 | ); |
| 5032 | assert_eq!(body["enable_thinking"], json!(false), "{base_url}: {body}"); |
| 5033 | } |
| 5034 | // Lookalike hosts stay unverified — fail closed. |
| 5035 | for base_url in [ |
| 5036 | "https://dashscope.aliyuncs.com.evil.example/compatible-mode/v1", |
| 5037 | "https://notdashscope.aliyuncs.com/compatible-mode/v1", |
| 5038 | "https://dashscope.aliyuncs.com/other-path/v1", |
| 5039 | ] { |
| 5040 | assert!( |
| 5041 | !super::is_exact_modelstudio_chat_route( |
| 5042 | ProviderKind::ModelstudioTokenPlan, |
| 5043 | base_url |
| 5044 | ), |
| 5045 | "{base_url}" |
| 5046 | ); |
| 5047 | } |
| 5048 | } |
| 5049 | |
| 5050 | #[test] |
| 5051 | fn modelstudio_hybrid_routes_send_documented_thinking_controls() { |
| 5052 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5053 | for (effort, enabled) in [ |
| 5054 | (None, true), |
| 5055 | (Some("low"), true), |
| 5056 | (Some("high"), true), |
| 5057 | (Some("xhigh"), true), |
| 5058 | (Some("off"), false), |
| 5059 | ] { |
| 5060 | let mut body = json!({}); |
| 5061 | apply_route_reasoning_controls( |
| 5062 | &mut body, |
| 5063 | ProviderKind::ModelstudioTokenPlan, |
| 5064 | base_url, |
| 5065 | "qwen3.7-plus", |
| 5066 | effort, |
| 5067 | ); |
| 5068 | |
| 5069 | assert_eq!(body["enable_thinking"], json!(enabled), "{effort:?}"); |
| 5070 | assert_eq!(body["preserve_thinking"], json!(enabled), "{effort:?}"); |
| 5071 | assert!(body.get("thinking").is_none(), "{effort:?}: {body}"); |
| 5072 | assert!(body.get("reasoning_effort").is_none(), "{effort:?}: {body}"); |
| 5073 | } |
| 5074 | } |
| 5075 | |
| 5076 | #[test] |
| 5077 | fn modelstudio_deepseek_v4_maps_effort_to_documented_values() { |
| 5078 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5079 | for (requested, expected) in [("low", "high"), ("high", "high"), ("xhigh", "max")] { |
| 5080 | let mut body = json!({}); |
| 5081 | apply_route_reasoning_controls( |
| 5082 | &mut body, |
| 5083 | ProviderKind::ModelstudioTokenPlan, |
| 5084 | base_url, |
| 5085 | "deepseek-v4-pro", |
| 5086 | Some(requested), |
| 5087 | ); |
| 5088 | |
| 5089 | assert_eq!(body["enable_thinking"], json!(true), "{requested}"); |
| 5090 | assert_eq!(body["reasoning_effort"], json!(expected), "{requested}"); |
| 5091 | } |
| 5092 | } |
| 5093 | |
| 5094 | #[test] |
| 5095 | fn modelstudio_reasoning_controls_fail_closed_on_custom_gateways() { |
| 5096 | let mut body = json!({ |
| 5097 | "enable_thinking": true, |
| 5098 | "preserve_thinking": true, |
| 5099 | "reasoning_effort": "high", |
| 5100 | }); |
| 5101 | apply_route_reasoning_controls( |
| 5102 | &mut body, |
| 5103 | ProviderKind::ModelstudioTokenPlan, |
| 5104 | "https://proxy.example/v1", |
| 5105 | "qwen3.7-plus", |
| 5106 | Some("high"), |
| 5107 | ); |
| 5108 | |
| 5109 | assert!(body.get("enable_thinking").is_none()); |
| 5110 | assert!(body.get("preserve_thinking").is_none()); |
| 5111 | assert!(body.get("reasoning_effort").is_none()); |
| 5112 | } |
| 5113 | |
| 5114 | #[test] |
| 5115 | fn modelstudio_anthropic_identities_write_nothing_on_the_chat_path() { |
| 5116 | // The Messages adapter owns these two. If `wire = "openai"` ever routes |
| 5117 | // them through Chat Completions, the shaper must strip rather than |
| 5118 | // inherit the OpenAI-dialect fields — there is no provider-enum writer |
| 5119 | // left to re-add them. |
| 5120 | for provider in [ |
| 5121 | ProviderKind::ModelstudioTokenPlanAnthropic, |
| 5122 | ProviderKind::ModelstudioCodingPlanAnthropic, |
| 5123 | ] { |
| 5124 | for base_url in [ |
| 5125 | crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL, |
| 5126 | crate::config::MODELSTUDIO_TOKEN_PLAN_ANTHROPIC_BASE_URL, |
| 5127 | ] { |
| 5128 | let mut body = json!({ "enable_thinking": true }); |
| 5129 | apply_route_reasoning_controls( |
| 5130 | &mut body, |
| 5131 | provider, |
| 5132 | base_url, |
| 5133 | "qwen3.7-plus", |
| 5134 | Some("high"), |
| 5135 | ); |
| 5136 | assert_eq!(body, json!({}), "{provider:?} {base_url}"); |
| 5137 | } |
| 5138 | } |
| 5139 | } |
| 5140 | |
| 5141 | #[test] |
| 5142 | fn modelstudio_qwen38_route_streams_reasoning_without_replaying_history() { |
| 5143 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5144 | for model in ["qwen3.8-max", "qwen3.8-max-preview"] { |
| 5145 | // qwen3.8 is thinking-only. Effort selection must never hide its |
| 5146 | // separate reasoning stream, including the stale `off` state |
| 5147 | // that can arrive before route normalization. |
| 5148 | assert_eq!( |
| 5149 | reasoning_stream_style_for_route( |
| 5150 | ProviderKind::ModelstudioTokenPlan, |
| 5151 | base_url, |
| 5152 | model, |
| 5153 | None, |
| 5154 | ), |
| 5155 | ReasoningStreamStyle::SeparateField, |
| 5156 | "{model}" |
| 5157 | ); |
| 5158 | // ...but Alibaba does not document `preserve_thinking` for the |
| 5159 | // qwen3.8 family, so no historical `reasoning_content` may be |
| 5160 | // replayed — even when a stale effort claims thinking is off. |
| 5161 | // Replaying it feeds the model its own past Thinking blocks and |
| 5162 | // re-triggers them every turn (observed handoff loop). |
| 5163 | for effort in [None, Some("off"), Some("high"), Some("xhigh")] { |
| 5164 | assert!( |
| 5165 | !should_replay_reasoning_content_for_provider_on_route( |
| 5166 | ProviderKind::ModelstudioTokenPlan, |
| 5167 | base_url, |
| 5168 | model, |
| 5169 | effort, |
| 5170 | ), |
| 5171 | "{model} {effort:?}" |
| 5172 | ); |
| 5173 | } |
| 5174 | // ...and no enable/disable or preserve switch is ever sent for |
| 5175 | // them. |
| 5176 | for effort in [None, Some("off"), Some("high")] { |
| 5177 | let mut body = json!({}); |
| 5178 | apply_route_reasoning_controls( |
| 5179 | &mut body, |
| 5180 | ProviderKind::ModelstudioTokenPlan, |
| 5181 | base_url, |
| 5182 | model, |
| 5183 | effort, |
| 5184 | ); |
| 5185 | assert!(body.get("enable_thinking").is_none(), "{model}: {body}"); |
| 5186 | assert!( |
| 5187 | body.get("preserve_thinking").is_none(), |
| 5188 | "{model} {effort:?}: {body}" |
| 5189 | ); |
| 5190 | assert!(body.get("reasoning_effort").is_none(), "{model}: {body}"); |
| 5191 | } |
| 5192 | } |
| 5193 | } |
| 5194 | |
| 5195 | #[test] |
| 5196 | fn modelstudio_hybrid_route_classifies_reasoning_and_replays_history() { |
| 5197 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5198 | assert_eq!( |
| 5199 | reasoning_stream_style_for_route( |
| 5200 | ProviderKind::ModelstudioTokenPlan, |
| 5201 | base_url, |
| 5202 | "qwen3.7-plus", |
| 5203 | None, |
| 5204 | ), |
| 5205 | ReasoningStreamStyle::SeparateField, |
| 5206 | ); |
| 5207 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5208 | ProviderKind::ModelstudioTokenPlan, |
| 5209 | base_url, |
| 5210 | "qwen3.7-plus", |
| 5211 | None, |
| 5212 | )); |
| 5213 | assert!(!should_replay_reasoning_content_for_provider_on_route( |
| 5214 | ProviderKind::ModelstudioTokenPlan, |
| 5215 | base_url, |
| 5216 | "qwen3.7-plus", |
| 5217 | Some("off"), |
| 5218 | )); |
| 5219 | } |
| 5220 | |
| 5221 | #[test] |
| 5222 | fn modelstudio_replay_stays_narrow_until_a_live_key_confirms_it() { |
| 5223 | // Deliberately narrower than PR #5233: only `preserve_thinking` models |
| 5224 | // replay. GLM and DeepSeek-V3.x on Model Studio stay stripped until |
| 5225 | // someone with a key confirms DashScope accepts `reasoning_content` in |
| 5226 | // input messages. deepseek-v4* is unaffected — it replays through |
| 5227 | // `requires_reasoning_content` on every provider. |
| 5228 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5229 | for model in ["glm-5.2", "deepseek-v3.2", "deepseek-v3.1"] { |
| 5230 | assert!( |
| 5231 | !should_replay_reasoning_content_for_provider_on_route( |
| 5232 | ProviderKind::ModelstudioTokenPlan, |
| 5233 | base_url, |
| 5234 | model, |
| 5235 | None, |
| 5236 | ), |
| 5237 | "{model}" |
| 5238 | ); |
| 5239 | } |
| 5240 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5241 | ProviderKind::ModelstudioTokenPlan, |
| 5242 | base_url, |
| 5243 | "deepseek-v4-pro", |
| 5244 | None, |
| 5245 | )); |
| 5246 | } |
| 5247 | |
| 5248 | #[test] |
| 5249 | fn modelstudio_coding_plan_chat_route_is_classified_for_all_supported_identities() { |
| 5250 | // The picker represents Coding Plan as mode = "coding-plan" under |
| 5251 | // the primary provider id, so the chat client receives |
| 5252 | // ModelstudioTokenPlan with the Coding Plan URL. Direct configuration |
| 5253 | // also retains the legacy ModelstudioCodingPlan identity. |
| 5254 | let base_url = crate::config::DEFAULT_MODELSTUDIO_CODING_PLAN_BASE_URL; |
| 5255 | for provider in [ |
| 5256 | ProviderKind::ModelstudioTokenPlan, |
| 5257 | ProviderKind::ModelstudioCodingPlan, |
| 5258 | ] { |
| 5259 | let mut body = json!({}); |
| 5260 | apply_route_reasoning_controls( |
| 5261 | &mut body, |
| 5262 | provider, |
| 5263 | base_url, |
| 5264 | "qwen3.7-plus", |
| 5265 | Some("high"), |
| 5266 | ); |
| 5267 | |
| 5268 | assert_eq!(body["enable_thinking"], json!(true), "{provider:?}"); |
| 5269 | assert_eq!(body["preserve_thinking"], json!(true), "{provider:?}"); |
| 5270 | assert_eq!( |
| 5271 | reasoning_stream_style_for_route(provider, base_url, "qwen3.7-plus", None), |
| 5272 | ReasoningStreamStyle::SeparateField, |
| 5273 | "{provider:?}", |
| 5274 | ); |
| 5275 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5276 | provider, |
| 5277 | base_url, |
| 5278 | "qwen3.7-plus", |
| 5279 | None, |
| 5280 | )); |
| 5281 | } |
| 5282 | } |
| 5283 | |
| 5284 | #[test] |
| 5285 | fn modelstudio_workspace_scoped_token_plan_route_is_recognized() { |
| 5286 | let workspace_url = |
| 5287 | "https://workspace-123.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1"; |
| 5288 | // Stream classification works on the workspace-scoped host... |
| 5289 | assert_eq!( |
| 5290 | reasoning_stream_style_for_route( |
| 5291 | ProviderKind::ModelstudioTokenPlan, |
| 5292 | workspace_url, |
| 5293 | "qwen3.8-max", |
| 5294 | None, |
| 5295 | ), |
| 5296 | ReasoningStreamStyle::SeparateField, |
| 5297 | ); |
| 5298 | assert_eq!( |
| 5299 | reasoning_stream_style_for_route( |
| 5300 | ProviderKind::ModelstudioTokenPlan, |
| 5301 | workspace_url, |
| 5302 | "qwen3.7-plus", |
| 5303 | None, |
| 5304 | ), |
| 5305 | ReasoningStreamStyle::SeparateField, |
| 5306 | ); |
| 5307 | // ...and the route gate decides replay the same way it does on the |
| 5308 | // default host: qwen3.8 never replays, documented preserve models do. |
| 5309 | assert!(!should_replay_reasoning_content_for_provider_on_route( |
| 5310 | ProviderKind::ModelstudioTokenPlan, |
| 5311 | workspace_url, |
| 5312 | "qwen3.8-max", |
| 5313 | None, |
| 5314 | )); |
| 5315 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5316 | ProviderKind::ModelstudioTokenPlan, |
| 5317 | workspace_url, |
| 5318 | "qwen3.7-plus", |
| 5319 | None, |
| 5320 | )); |
| 5321 | let mut body = json!({}); |
| 5322 | apply_route_reasoning_controls( |
| 5323 | &mut body, |
| 5324 | ProviderKind::ModelstudioTokenPlan, |
| 5325 | workspace_url, |
| 5326 | "qwen3.7-plus", |
| 5327 | Some("high"), |
| 5328 | ); |
| 5329 | assert_eq!(body["preserve_thinking"], json!(true), "{body}"); |
| 5330 | } |
| 5331 | |
| 5332 | #[test] |
| 5333 | fn modelstudio_thinking_named_unknown_model_cannot_bypass_the_route_gate() { |
| 5334 | // An unknown Model Studio model whose *name* suggests reasoning must |
| 5335 | // not gain replay: only the documented `preserve_thinking` list (plus |
| 5336 | // the concrete DeepSeek V4 family ids) authorizes historical |
| 5337 | // `reasoning_content` on exact Model Studio routes. The generic |
| 5338 | // `-thinking`/`reasoner` heuristics prove nothing about DashScope's |
| 5339 | // request dialect. |
| 5340 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5341 | for model in [ |
| 5342 | "foo-thinking", |
| 5343 | "foo-reasoner", |
| 5344 | "new-model-reasoning", |
| 5345 | "acme-reasoner-v2", |
| 5346 | ] { |
| 5347 | assert!( |
| 5348 | !should_replay_reasoning_content_for_provider_on_route( |
| 5349 | ProviderKind::ModelstudioTokenPlan, |
| 5350 | base_url, |
| 5351 | model, |
| 5352 | None, |
| 5353 | ), |
| 5354 | "{model}" |
| 5355 | ); |
| 5356 | // The shaper writes no reasoning controls for it either — fail |
| 5357 | // closed on the wire. |
| 5358 | let mut body = json!({ "enable_thinking": true, "preserve_thinking": true }); |
| 5359 | apply_route_reasoning_controls( |
| 5360 | &mut body, |
| 5361 | ProviderKind::ModelstudioTokenPlan, |
| 5362 | base_url, |
| 5363 | model, |
| 5364 | Some("high"), |
| 5365 | ); |
| 5366 | assert!(body.get("enable_thinking").is_none(), "{model}: {body}"); |
| 5367 | assert!(body.get("preserve_thinking").is_none(), "{model}: {body}"); |
| 5368 | } |
| 5369 | // A suggestive name is not a replay contract on any route. |
| 5370 | assert!(!requires_reasoning_content("foo-thinking")); |
| 5371 | assert!(!requires_reasoning_content("foo-reasoner")); |
| 5372 | } |
| 5373 | |
| 5374 | #[test] |
| 5375 | fn modelstudio_qwen36_flash_preserves_thinking_per_current_documentation() { |
| 5376 | // Current Alibaba documentation explicitly includes qwen3.6-flash |
| 5377 | // (and its documented snapshots) among the `preserve_thinking` |
| 5378 | // models. Keep positive coverage so any future narrowing of the |
| 5379 | // preserve list has to remove it deliberately. |
| 5380 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5381 | for model in ["qwen3.6-flash", "qwen3.6-flash-2026-04-16"] { |
| 5382 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5383 | ProviderKind::ModelstudioTokenPlan, |
| 5384 | base_url, |
| 5385 | model, |
| 5386 | None, |
| 5387 | )); |
| 5388 | let mut body = json!({}); |
| 5389 | apply_route_reasoning_controls( |
| 5390 | &mut body, |
| 5391 | ProviderKind::ModelstudioTokenPlan, |
| 5392 | base_url, |
| 5393 | model, |
| 5394 | Some("high"), |
| 5395 | ); |
| 5396 | assert_eq!(body["enable_thinking"], json!(true), "{model}"); |
| 5397 | assert_eq!(body["preserve_thinking"], json!(true), "{model}: {body}"); |
| 5398 | // Hybrid semantics: an explicit off disables both, and replay |
| 5399 | // follows the effort gate. |
| 5400 | assert!(!should_replay_reasoning_content_for_provider_on_route( |
| 5401 | ProviderKind::ModelstudioTokenPlan, |
| 5402 | base_url, |
| 5403 | model, |
| 5404 | Some("off"), |
| 5405 | )); |
| 5406 | let mut off_body = json!({}); |
| 5407 | apply_route_reasoning_controls( |
| 5408 | &mut off_body, |
| 5409 | ProviderKind::ModelstudioTokenPlan, |
| 5410 | base_url, |
| 5411 | model, |
| 5412 | Some("off"), |
| 5413 | ); |
| 5414 | assert_eq!(off_body["enable_thinking"], json!(false), "{model}"); |
| 5415 | assert_eq!(off_body["preserve_thinking"], json!(false), "{model}"); |
| 5416 | } |
| 5417 | |
| 5418 | for invented in [ |
| 5419 | "qwen3.6-flash-future", |
| 5420 | "qwen3.7-plus-proxy", |
| 5421 | "kimi-k2.7-code-unverified", |
| 5422 | ] { |
| 5423 | assert!( |
| 5424 | !should_replay_reasoning_content_for_provider_on_route( |
| 5425 | ProviderKind::ModelstudioTokenPlan, |
| 5426 | base_url, |
| 5427 | invented, |
| 5428 | None, |
| 5429 | ), |
| 5430 | "prefix lookalikes must fail closed: {invented}" |
| 5431 | ); |
| 5432 | } |
| 5433 | } |
| 5434 | |
| 5435 | #[test] |
| 5436 | fn modelstudio_kimi_k27_code_is_thinking_only_and_preserves_trace() { |
| 5437 | // NOTE: unlike the qwen3.8 pair, this classification is asserted by |
| 5438 | // PR #5233 rather than corroborated by models_dev.bundled.json, which |
| 5439 | // lists kimi-k2.7-code with `reasoning: true` and no `always_on`. |
| 5440 | let base_url = crate::config::DEFAULT_MODELSTUDIO_TOKEN_PLAN_BASE_URL; |
| 5441 | for model in [ |
| 5442 | "kimi-k2.7-code", |
| 5443 | "kimi/kimi-k2.7-code", |
| 5444 | "kimi/kimi-k2.7-code-highspeed", |
| 5445 | ] { |
| 5446 | let mut body = json!({}); |
| 5447 | apply_route_reasoning_controls( |
| 5448 | &mut body, |
| 5449 | ProviderKind::ModelstudioTokenPlan, |
| 5450 | base_url, |
| 5451 | model, |
| 5452 | Some("off"), |
| 5453 | ); |
| 5454 | |
| 5455 | assert!(body.get("enable_thinking").is_none(), "{model}: {body}"); |
| 5456 | assert_eq!(body["preserve_thinking"], json!(true), "{model}"); |
| 5457 | assert_eq!( |
| 5458 | reasoning_stream_style_for_route( |
| 5459 | ProviderKind::ModelstudioTokenPlan, |
| 5460 | base_url, |
| 5461 | model, |
| 5462 | None, |
| 5463 | ), |
| 5464 | ReasoningStreamStyle::SeparateField, |
| 5465 | "{model}", |
| 5466 | ); |
| 5467 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5468 | ProviderKind::ModelstudioTokenPlan, |
| 5469 | base_url, |
| 5470 | model, |
| 5471 | Some("off"), |
| 5472 | )); |
| 5473 | } |
| 5474 | } |
| 5475 | |
| 5476 | #[test] |
| 5477 | fn stream_classifies_moonshot_kimi_as_reasoning() { |
| 5478 | // #3016: without this, Kimi thinking leaked into answer text. |
| 5479 | assert!(is_reasoning_model_for_stream( |
| 5480 | ProviderKind::Moonshot, |
| 5481 | "kimi-k2.6" |
| 5482 | )); |
| 5483 | assert!( |
| 5484 | is_reasoning_model_for_stream(ProviderKind::Moonshot, "kimi-for-coding"), |
| 5485 | "Kimi Code's stable model id now maps to K2.7 Code and streams reasoning_content" |
| 5486 | ); |
| 5487 | } |
| 5488 | |
| 5489 | #[test] |
| 5490 | fn moonshot_and_minimax_replay_reasoning_content_for_supported_models() { |
| 5491 | assert!(should_replay_reasoning_content_for_provider( |
| 5492 | ProviderKind::Moonshot, |
| 5493 | "kimi-k2.7-code", |
| 5494 | None, |
| 5495 | )); |
| 5496 | assert!(should_replay_reasoning_content_for_provider( |
| 5497 | ProviderKind::Moonshot, |
| 5498 | "kimi-for-coding", |
| 5499 | None, |
| 5500 | )); |
| 5501 | assert!(should_replay_reasoning_content_for_provider( |
| 5502 | ProviderKind::Minimax, |
| 5503 | "MiniMax-M3", |
| 5504 | None, |
| 5505 | )); |
| 5506 | assert!(should_replay_reasoning_content_for_provider( |
| 5507 | ProviderKind::Zai, |
| 5508 | "GLM-5.2", |
| 5509 | None, |
| 5510 | )); |
| 5511 | assert!(should_replay_reasoning_content_for_provider( |
| 5512 | ProviderKind::Zai, |
| 5513 | "GLM-5.3", |
| 5514 | None, |
| 5515 | )); |
| 5516 | assert!(!should_replay_reasoning_content_for_provider( |
| 5517 | ProviderKind::Moonshot, |
| 5518 | "kimi-for-coding", |
| 5519 | Some("off"), |
| 5520 | )); |
| 5521 | } |
| 5522 | |
| 5523 | #[test] |
| 5524 | fn bare_k3_reasoning_semantics_are_scoped_to_exact_kimi_code_route() { |
| 5525 | let kimi_code = crate::config::DEFAULT_KIMI_CODE_BASE_URL; |
| 5526 | let direct_moonshot = crate::config::DEFAULT_MOONSHOT_BASE_URL; |
| 5527 | |
| 5528 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5529 | ProviderKind::Moonshot, |
| 5530 | kimi_code, |
| 5531 | crate::config::KIMI_CODE_K3_MODEL, |
| 5532 | Some("high"), |
| 5533 | )); |
| 5534 | assert_eq!( |
| 5535 | reasoning_stream_style_for_route( |
| 5536 | ProviderKind::Moonshot, |
| 5537 | kimi_code, |
| 5538 | crate::config::KIMI_CODE_K3_MODEL, |
| 5539 | None, |
| 5540 | ), |
| 5541 | ReasoningStreamStyle::SeparateField |
| 5542 | ); |
| 5543 | |
| 5544 | assert!(!should_replay_reasoning_content_for_provider_on_route( |
| 5545 | ProviderKind::Moonshot, |
| 5546 | direct_moonshot, |
| 5547 | crate::config::KIMI_CODE_K3_MODEL, |
| 5548 | Some("high"), |
| 5549 | )); |
| 5550 | // Display is not replay: a reasoning field that arrives on the |
| 5551 | // neighboring route still renders as Thinking (#6501), while the |
| 5552 | // replay contract above stays scoped to the exact membership route. |
| 5553 | assert_eq!( |
| 5554 | reasoning_stream_style_for_route( |
| 5555 | ProviderKind::Moonshot, |
| 5556 | direct_moonshot, |
| 5557 | crate::config::KIMI_CODE_K3_MODEL, |
| 5558 | None, |
| 5559 | ), |
| 5560 | ReasoningStreamStyle::SeparateField |
| 5561 | ); |
| 5562 | assert!( |
| 5563 | should_replay_reasoning_content_for_provider_on_route( |
| 5564 | ProviderKind::Moonshot, |
| 5565 | kimi_code, |
| 5566 | crate::config::KIMI_CODE_K3_MODEL, |
| 5567 | Some("off"), |
| 5568 | ), |
| 5569 | "exact membership K3 stays always-thinking even for a stale raw Off caller" |
| 5570 | ); |
| 5571 | } |
| 5572 | |
| 5573 | #[test] |
| 5574 | fn direct_moonshot_k3_is_always_thinking_and_replays_reasoning() { |
| 5575 | let direct = crate::config::DEFAULT_MOONSHOT_BASE_URL; |
| 5576 | let model = crate::config::MOONSHOT_KIMI_K3_MODEL; |
| 5577 | |
| 5578 | for effort in [Some("off"), Some("low"), Some("high"), Some("max"), None] { |
| 5579 | assert!(should_replay_reasoning_content_for_provider_on_route( |
| 5580 | ProviderKind::Moonshot, |
| 5581 | direct, |
| 5582 | model, |
| 5583 | effort, |
| 5584 | )); |
| 5585 | } |
| 5586 | assert_eq!( |
| 5587 | reasoning_stream_style_for_route(ProviderKind::Moonshot, direct, model, None), |
| 5588 | ReasoningStreamStyle::SeparateField |
| 5589 | ); |
| 5590 | } |
| 5591 | |
| 5592 | #[test] |
| 5593 | fn xiaomi_mimo_uses_max_completion_tokens_payload_key() { |
| 5594 | let mut body = json!({ |
| 5595 | "model": "mimo-v2.5-pro", |
| 5596 | "messages": [], |
| 5597 | "max_tokens": 8192, |
| 5598 | }); |
| 5599 | |
| 5600 | apply_provider_token_limit( |
| 5601 | &mut body, |
| 5602 | ProviderKind::XiaomiMimo, |
| 5603 | "https://api.xiaomimimo.com/v1", |
| 5604 | "mimo-v2.5-pro", |
| 5605 | 8192, |
| 5606 | ); |
| 5607 | |
| 5608 | assert!(body.get("max_tokens").is_none()); |
| 5609 | assert_eq!( |
| 5610 | body.get("max_completion_tokens") |
| 5611 | .and_then(serde_json::Value::as_u64), |
| 5612 | Some(8192) |
| 5613 | ); |
| 5614 | } |
| 5615 | |
| 5616 | #[test] |
| 5617 | fn openai_reasoning_model_uses_completion_token_limit_and_effort_field() { |
| 5618 | let mut body = json!({ |
| 5619 | "model": "gpt-5.5", |
| 5620 | "messages": [], |
| 5621 | "max_tokens": 4096, |
| 5622 | }); |
| 5623 | |
| 5624 | apply_provider_token_limit( |
| 5625 | &mut body, |
| 5626 | ProviderKind::Openai, |
| 5627 | "https://api.openai.com/v1", |
| 5628 | "gpt-5.5", |
| 5629 | 4096, |
| 5630 | ); |
| 5631 | apply_openai_reasoning_effort(&mut body, ProviderKind::Openai, "gpt-5.5", Some("high")); |
| 5632 | |
| 5633 | assert!(body.get("max_tokens").is_none()); |
| 5634 | assert_eq!( |
| 5635 | body.get("max_completion_tokens") |
| 5636 | .and_then(serde_json::Value::as_u64), |
| 5637 | Some(4096) |
| 5638 | ); |
| 5639 | assert_eq!( |
| 5640 | body.get("reasoning_effort") |
| 5641 | .and_then(serde_json::Value::as_str), |
| 5642 | Some("high") |
| 5643 | ); |
| 5644 | } |
| 5645 | |
| 5646 | #[test] |
| 5647 | fn gpt_56_uses_documented_max_reasoning_effort() { |
| 5648 | let mut body = json!({ |
| 5649 | "model": "gpt-5.6-sol", |
| 5650 | "messages": [], |
| 5651 | "max_tokens": 8192, |
| 5652 | }); |
| 5653 | |
| 5654 | apply_provider_token_limit( |
| 5655 | &mut body, |
| 5656 | ProviderKind::Openai, |
| 5657 | "https://api.openai.com/v1", |
| 5658 | "gpt-5.6-sol", |
| 5659 | 8192, |
| 5660 | ); |
| 5661 | apply_openai_reasoning_effort(&mut body, ProviderKind::Openai, "gpt-5.6-sol", Some("max")); |
| 5662 | |
| 5663 | assert!(body.get("max_tokens").is_none()); |
| 5664 | assert_eq!(body["max_completion_tokens"], json!(8192)); |
| 5665 | assert_eq!(body["reasoning_effort"], json!("max")); |
| 5666 | } |
| 5667 | |
| 5668 | #[test] |
| 5669 | fn grok_47_and_46_use_exact_first_party_reasoning_effort_ladder() { |
| 5670 | // Both catalog rows document low/medium/high/xhigh; reasoning cannot |
| 5671 | // be disabled, so `off` is the documented default `high`. |
| 5672 | for model in [ |
| 5673 | crate::config::XAI_GROK_4_7_MODEL, |
| 5674 | crate::config::XAI_GROK_4_6_MODEL, |
| 5675 | ] { |
| 5676 | for (requested, expected) in [ |
| 5677 | ("off", "high"), |
| 5678 | ("low", "low"), |
| 5679 | ("medium", "medium"), |
| 5680 | ("high", "high"), |
| 5681 | ("xhigh", "xhigh"), |
| 5682 | ("max", "xhigh"), |
| 5683 | ] { |
| 5684 | let mut body = json!({}); |
| 5685 | apply_route_reasoning_controls( |
| 5686 | &mut body, |
| 5687 | ProviderKind::Xai, |
| 5688 | crate::config::DEFAULT_XAI_BASE_URL, |
| 5689 | model, |
| 5690 | Some(requested), |
| 5691 | ); |
| 5692 | assert_eq!( |
| 5693 | body, |
| 5694 | json!({ "reasoning_effort": expected }), |
| 5695 | "{model} {requested}" |
| 5696 | ); |
| 5697 | } |
| 5698 | } |
| 5699 | |
| 5700 | // A row without a documented effort option (grok-4.3) and an id the |
| 5701 | // catalog does not know send nothing. |
| 5702 | for model in [crate::config::XAI_GROK_4_3_MODEL, "grok-9-unknown"] { |
| 5703 | let mut body = json!({}); |
| 5704 | apply_route_reasoning_controls( |
| 5705 | &mut body, |
| 5706 | ProviderKind::Xai, |
| 5707 | crate::config::DEFAULT_XAI_BASE_URL, |
| 5708 | model, |
| 5709 | Some("medium"), |
| 5710 | ); |
| 5711 | assert_eq!(body, json!({}), "{model}"); |
| 5712 | } |
| 5713 | |
| 5714 | let mut provider_default = json!({}); |
| 5715 | apply_route_reasoning_controls( |
| 5716 | &mut provider_default, |
| 5717 | ProviderKind::Xai, |
| 5718 | crate::config::DEFAULT_XAI_BASE_URL, |
| 5719 | crate::config::XAI_GROK_4_6_MODEL, |
| 5720 | Some("auto"), |
| 5721 | ); |
| 5722 | assert_eq!(provider_default, json!({})); |
| 5723 | |
| 5724 | let mut custom = json!({}); |
| 5725 | apply_route_reasoning_controls( |
| 5726 | &mut custom, |
| 5727 | ProviderKind::Xai, |
| 5728 | "https://gateway.example/v1", |
| 5729 | crate::config::XAI_GROK_4_6_MODEL, |
| 5730 | Some("medium"), |
| 5731 | ); |
| 5732 | assert_eq!(custom, json!({})); |
| 5733 | } |
| 5734 | |
| 5735 | #[test] |
| 5736 | fn grok_45_uses_first_party_ladder_and_maps_xhigh_to_high() { |
| 5737 | for (requested, expected) in [ |
| 5738 | ("off", "high"), |
| 5739 | ("low", "low"), |
| 5740 | ("medium", "medium"), |
| 5741 | ("high", "high"), |
| 5742 | ("xhigh", "high"), |
| 5743 | ("max", "high"), |
| 5744 | ] { |
| 5745 | let mut body = json!({}); |
| 5746 | apply_route_reasoning_controls( |
| 5747 | &mut body, |
| 5748 | ProviderKind::Xai, |
| 5749 | crate::config::DEFAULT_XAI_BASE_URL, |
| 5750 | crate::config::XAI_GROK_4_5_MODEL, |
| 5751 | Some(requested), |
| 5752 | ); |
| 5753 | assert_eq!(body, json!({ "reasoning_effort": expected }), "{requested}"); |
| 5754 | } |
| 5755 | } |
| 5756 | |
| 5757 | #[test] |
| 5758 | fn inkling_uses_its_exact_reasoning_vocabulary_without_thinking_extension() { |
| 5759 | for (requested, expected) in [ |
| 5760 | ("off", "none"), |
| 5761 | ("minimal", "minimal"), |
| 5762 | ("low", "low"), |
| 5763 | ("medium", "medium"), |
| 5764 | ("high", "high"), |
| 5765 | ("max", "max"), |
| 5766 | ("xhigh", "max"), |
| 5767 | ] { |
| 5768 | let mut body = json!({ |
| 5769 | "thinking": { "type": "enabled" }, |
| 5770 | "reasoning_effort": "xhigh", |
| 5771 | }); |
| 5772 | |
| 5773 | apply_inkling_reasoning_effort( |
| 5774 | &mut body, |
| 5775 | ProviderKind::Together, |
| 5776 | "thinkingmachines/inkling", |
| 5777 | Some(requested), |
| 5778 | ); |
| 5779 | |
| 5780 | assert_eq!(body["reasoning_effort"], json!(expected)); |
| 5781 | assert!(body.get("thinking").is_none()); |
| 5782 | } |
| 5783 | } |
| 5784 | |
| 5785 | #[test] |
| 5786 | fn inkling_reasoning_override_is_scoped_to_the_exact_together_route() { |
| 5787 | let mut other_model = json!({ "thinking": { "type": "enabled" } }); |
| 5788 | apply_inkling_reasoning_effort( |
| 5789 | &mut other_model, |
| 5790 | ProviderKind::Together, |
| 5791 | "deepseek-ai/DeepSeek-V4-Pro", |
| 5792 | Some("max"), |
| 5793 | ); |
| 5794 | assert_eq!(other_model["thinking"]["type"], json!("enabled")); |
| 5795 | assert!(other_model.get("reasoning_effort").is_none()); |
| 5796 | |
| 5797 | let mut other_provider = json!({ "thinking": { "type": "enabled" } }); |
| 5798 | apply_inkling_reasoning_effort( |
| 5799 | &mut other_provider, |
| 5800 | ProviderKind::Openrouter, |
| 5801 | "thinkingmachines/inkling", |
| 5802 | Some("max"), |
| 5803 | ); |
| 5804 | assert_eq!(other_provider["thinking"]["type"], json!("enabled")); |
| 5805 | assert!(other_provider.get("reasoning_effort").is_none()); |
| 5806 | } |
| 5807 | |
| 5808 | #[test] |
| 5809 | fn kimi_code_k3_uses_documented_nested_thinking_effort() { |
| 5810 | for (requested, expected) in [ |
| 5811 | ("low", json!({ "type": "enabled", "effort": "low" })), |
| 5812 | ("minimum", json!({ "type": "enabled", "effort": "low" })), |
| 5813 | ("light", json!({ "type": "enabled", "effort": "low" })), |
| 5814 | ("medium", json!({ "type": "enabled", "effort": "high" })), |
| 5815 | ("high", json!({ "type": "enabled", "effort": "high" })), |
| 5816 | ("xhigh", json!({ "type": "enabled", "effort": "max" })), |
| 5817 | ("ultra", json!({ "type": "enabled", "effort": "max" })), |
| 5818 | ("max", json!({ "type": "enabled", "effort": "max" })), |
| 5819 | ("none", json!({ "type": "enabled", "effort": "low" })), |
| 5820 | ("off", json!({ "type": "enabled", "effort": "low" })), |
| 5821 | ] { |
| 5822 | let mut body = json!({ "reasoning_effort": "stale" }); |
| 5823 | apply_kimi_code_k3_reasoning_effort( |
| 5824 | &mut body, |
| 5825 | ProviderKind::Moonshot, |
| 5826 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5827 | crate::config::KIMI_CODE_K3_MODEL, |
| 5828 | Some(requested), |
| 5829 | ); |
| 5830 | |
| 5831 | assert_eq!(body["thinking"], expected, "requested {requested}"); |
| 5832 | assert!(body.get("reasoning_effort").is_none()); |
| 5833 | } |
| 5834 | } |
| 5835 | |
| 5836 | #[test] |
| 5837 | fn kimi_code_k3_256k_uses_k3_reasoning_and_membership_sampling_contracts() { |
| 5838 | let mut body = json!({ |
| 5839 | "reasoning_effort": "stale", |
| 5840 | "temperature": 0.3, |
| 5841 | "top_p": 0.8, |
| 5842 | }); |
| 5843 | apply_kimi_code_k3_reasoning_effort( |
| 5844 | &mut body, |
| 5845 | ProviderKind::Moonshot, |
| 5846 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5847 | crate::config::KIMI_CODE_K3_256K_MODEL, |
| 5848 | Some("max"), |
| 5849 | ); |
| 5850 | apply_kimi_code_fixed_sampling( |
| 5851 | &mut body, |
| 5852 | ProviderKind::Moonshot, |
| 5853 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5854 | crate::config::KIMI_CODE_K3_256K_MODEL, |
| 5855 | ); |
| 5856 | |
| 5857 | assert_eq!( |
| 5858 | body["thinking"], |
| 5859 | json!({ "type": "enabled", "effort": "max" }) |
| 5860 | ); |
| 5861 | assert!(body.get("reasoning_effort").is_none()); |
| 5862 | assert!(body.get("temperature").is_none()); |
| 5863 | assert!(body.get("top_p").is_none()); |
| 5864 | } |
| 5865 | |
| 5866 | #[test] |
| 5867 | fn kimi_code_fixed_sampling_does_not_leak_to_neighbor_routes() { |
| 5868 | for (provider, base_url, model) in [ |
| 5869 | ( |
| 5870 | ProviderKind::Moonshot, |
| 5871 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 5872 | crate::config::KIMI_CODE_K3_256K_MODEL, |
| 5873 | ), |
| 5874 | ( |
| 5875 | ProviderKind::Openrouter, |
| 5876 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5877 | crate::config::KIMI_CODE_K3_256K_MODEL, |
| 5878 | ), |
| 5879 | ( |
| 5880 | ProviderKind::Moonshot, |
| 5881 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5882 | "k3-256k-preview", |
| 5883 | ), |
| 5884 | ] { |
| 5885 | let mut body = json!({ "temperature": 0.3, "top_p": 0.8 }); |
| 5886 | apply_kimi_code_fixed_sampling(&mut body, provider, base_url, model); |
| 5887 | assert_eq!(body["temperature"], json!(0.3)); |
| 5888 | assert_eq!(body["top_p"], json!(0.8)); |
| 5889 | } |
| 5890 | } |
| 5891 | |
| 5892 | #[test] |
| 5893 | fn direct_moonshot_k3_uses_top_level_effort_and_never_disables_thinking() { |
| 5894 | for (requested, expected) in [ |
| 5895 | ("off", "low"), |
| 5896 | ("none", "low"), |
| 5897 | ("low", "low"), |
| 5898 | ("medium", "high"), |
| 5899 | ("high", "high"), |
| 5900 | ("xhigh", "max"), |
| 5901 | ("max", "max"), |
| 5902 | ] { |
| 5903 | let mut body = json!({ |
| 5904 | "model": crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5905 | "thinking": { "type": "disabled" }, |
| 5906 | }); |
| 5907 | apply_route_reasoning_controls( |
| 5908 | &mut body, |
| 5909 | ProviderKind::Moonshot, |
| 5910 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 5911 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5912 | Some(requested), |
| 5913 | ); |
| 5914 | |
| 5915 | assert_eq!(body["reasoning_effort"], json!(expected), "{requested}"); |
| 5916 | assert!(body.get("thinking").is_none(), "{requested}: {body}"); |
| 5917 | } |
| 5918 | |
| 5919 | let mut provider_default = json!({ "thinking": { "type": "enabled" } }); |
| 5920 | apply_route_reasoning_controls( |
| 5921 | &mut provider_default, |
| 5922 | ProviderKind::Moonshot, |
| 5923 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 5924 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5925 | Some("auto"), |
| 5926 | ); |
| 5927 | assert!(provider_default.get("thinking").is_none()); |
| 5928 | assert!(provider_default.get("reasoning_effort").is_none()); |
| 5929 | } |
| 5930 | |
| 5931 | #[test] |
| 5932 | fn direct_moonshot_k3_uses_modern_token_field_and_fixed_sampling_only_on_exact_route() { |
| 5933 | let mut direct = json!({ |
| 5934 | "max_tokens": 64, |
| 5935 | "temperature": 0.2, |
| 5936 | "top_p": 0.9, |
| 5937 | }); |
| 5938 | apply_provider_token_limit( |
| 5939 | &mut direct, |
| 5940 | ProviderKind::Moonshot, |
| 5941 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 5942 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5943 | 64, |
| 5944 | ); |
| 5945 | apply_direct_moonshot_k3_fixed_sampling( |
| 5946 | &mut direct, |
| 5947 | ProviderKind::Moonshot, |
| 5948 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 5949 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5950 | ); |
| 5951 | assert_eq!(direct["max_completion_tokens"], json!(64)); |
| 5952 | assert!(direct.get("max_tokens").is_none()); |
| 5953 | assert!(direct.get("temperature").is_none()); |
| 5954 | assert!(direct.get("top_p").is_none()); |
| 5955 | |
| 5956 | let mut neighbor = json!({ |
| 5957 | "max_tokens": 64, |
| 5958 | "temperature": 0.2, |
| 5959 | "top_p": 0.9, |
| 5960 | }); |
| 5961 | apply_provider_token_limit( |
| 5962 | &mut neighbor, |
| 5963 | ProviderKind::Moonshot, |
| 5964 | "https://proxy.example/v1", |
| 5965 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5966 | 64, |
| 5967 | ); |
| 5968 | apply_direct_moonshot_k3_fixed_sampling( |
| 5969 | &mut neighbor, |
| 5970 | ProviderKind::Moonshot, |
| 5971 | "https://proxy.example/v1", |
| 5972 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 5973 | ); |
| 5974 | assert_eq!(neighbor["max_tokens"], json!(64)); |
| 5975 | assert!(neighbor.get("max_completion_tokens").is_none()); |
| 5976 | assert_eq!(neighbor["temperature"], json!(0.2)); |
| 5977 | assert_eq!(neighbor["top_p"], json!(0.9)); |
| 5978 | } |
| 5979 | |
| 5980 | #[test] |
| 5981 | fn direct_and_membership_k3_reasoning_dialects_do_not_cross_routes() { |
| 5982 | let mut membership = json!({}); |
| 5983 | apply_route_reasoning_controls( |
| 5984 | &mut membership, |
| 5985 | ProviderKind::Moonshot, |
| 5986 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5987 | crate::config::KIMI_CODE_K3_MODEL, |
| 5988 | Some("max"), |
| 5989 | ); |
| 5990 | assert_eq!( |
| 5991 | membership["thinking"], |
| 5992 | json!({ "type": "enabled", "effort": "max" }) |
| 5993 | ); |
| 5994 | assert!(membership.get("reasoning_effort").is_none()); |
| 5995 | |
| 5996 | for (base_url, model) in [ |
| 5997 | ( |
| 5998 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 5999 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 6000 | ), |
| 6001 | ( |
| 6002 | crate::config::DEFAULT_MOONSHOT_BASE_URL, |
| 6003 | crate::config::KIMI_CODE_K3_MODEL, |
| 6004 | ), |
| 6005 | ( |
| 6006 | "https://proxy.example/v1", |
| 6007 | crate::config::MOONSHOT_KIMI_K3_MODEL, |
| 6008 | ), |
| 6009 | ] { |
| 6010 | let mut neighbor = json!({}); |
| 6011 | apply_route_reasoning_controls( |
| 6012 | &mut neighbor, |
| 6013 | ProviderKind::Moonshot, |
| 6014 | base_url, |
| 6015 | model, |
| 6016 | Some("max"), |
| 6017 | ); |
| 6018 | assert_eq!( |
| 6019 | neighbor["thinking"], |
| 6020 | json!({ "type": "enabled" }), |
| 6021 | "{base_url} / {model}" |
| 6022 | ); |
| 6023 | assert!(neighbor.get("reasoning_effort").is_none()); |
| 6024 | assert!(neighbor.pointer("/thinking/effort").is_none()); |
| 6025 | } |
| 6026 | } |
| 6027 | |
| 6028 | #[test] |
| 6029 | fn kimi_code_k3_effort_override_never_leaks_to_neighbor_routes() { |
| 6030 | for (base_url, model) in [ |
| 6031 | (crate::config::DEFAULT_KIMI_CODE_BASE_URL, "kimi-k3"), |
| 6032 | ( |
| 6033 | crate::config::DEFAULT_KIMI_CODE_BASE_URL, |
| 6034 | crate::config::DEFAULT_KIMI_CODE_MODEL, |
| 6035 | ), |
| 6036 | (crate::config::DEFAULT_MOONSHOT_BASE_URL, "k3"), |
| 6037 | ] { |
| 6038 | let mut body = json!({ "thinking": { "type": "enabled" } }); |
| 6039 | apply_kimi_code_k3_reasoning_effort( |
| 6040 | &mut body, |
| 6041 | ProviderKind::Moonshot, |
| 6042 | base_url, |
| 6043 | model, |
| 6044 | Some("max"), |
| 6045 | ); |
| 6046 | |
| 6047 | assert_eq!(body["thinking"], json!({ "type": "enabled" })); |
| 6048 | assert!( |
| 6049 | body.pointer("/thinking/effort").is_none(), |
| 6050 | "{base_url} / {model}" |
| 6051 | ); |
| 6052 | assert!(body.get("reasoning_effort").is_none()); |
| 6053 | } |
| 6054 | } |
| 6055 | |
| 6056 | #[test] |
| 6057 | fn muse_spark_uses_meta_reasoning_effort_without_openai_token_rewrite() { |
| 6058 | let mut body = json!({ |
| 6059 | "model": "muse-spark-1.1", |
| 6060 | "messages": [], |
| 6061 | "max_tokens": 8192, |
| 6062 | }); |
| 6063 | |
| 6064 | apply_provider_token_limit( |
| 6065 | &mut body, |
| 6066 | ProviderKind::Meta, |
| 6067 | "https://api.meta.ai/v1", |
| 6068 | "muse-spark-1.1", |
| 6069 | 8192, |
| 6070 | ); |
| 6071 | apply_openai_reasoning_effort(&mut body, ProviderKind::Meta, "muse-spark-1.1", Some("max")); |
| 6072 | |
| 6073 | assert_eq!(body["max_tokens"], json!(8192)); |
| 6074 | assert!(body.get("max_completion_tokens").is_none()); |
| 6075 | assert_eq!(body["reasoning_effort"], json!("xhigh")); |
| 6076 | } |
| 6077 | |
| 6078 | #[test] |
| 6079 | fn provider_regression_5853_muse_family_preserves_reasoning_effort() { |
| 6080 | for model in [ |
| 6081 | "muse-spark-1.2", |
| 6082 | "muse-spark-1.3", |
| 6083 | "muse-spark-1.3-contributor", |
| 6084 | " MUSE-SPARK-1.3 ", |
| 6085 | ] { |
| 6086 | for (effort, wire) in [ |
| 6087 | ("low", "low"), |
| 6088 | ("medium", "medium"), |
| 6089 | ("high", "high"), |
| 6090 | ("xhigh", "xhigh"), |
| 6091 | ("max", "xhigh"), |
| 6092 | ("ultra", "xhigh"), |
| 6093 | ] { |
| 6094 | let mut body = json!({"model": model}); |
| 6095 | apply_openai_reasoning_effort(&mut body, ProviderKind::Meta, model, Some(effort)); |
| 6096 | assert_eq!(body["reasoning_effort"], wire, "{model}, effort={effort}"); |
| 6097 | } |
| 6098 | } |
| 6099 | for (provider, model, effort) in [ |
| 6100 | (ProviderKind::Openai, "muse-spark-1.3", Some("high")), |
| 6101 | (ProviderKind::Meta, "muse-glimmer-fixture", Some("high")), |
| 6102 | (ProviderKind::Meta, "muse-sparks-fixture", Some("high")), |
| 6103 | (ProviderKind::Meta, "muse-spark-1.3", None), |
| 6104 | ] { |
| 6105 | let mut body = json!({"model": model}); |
| 6106 | apply_openai_reasoning_effort(&mut body, provider, model, effort); |
| 6107 | assert!( |
| 6108 | body.get("reasoning_effort").is_none(), |
| 6109 | "{provider:?}, {model}" |
| 6110 | ); |
| 6111 | } |
| 6112 | } |
| 6113 | |
| 6114 | #[test] |
| 6115 | fn openai_non_reasoning_model_omits_reasoning_only_fields() { |
| 6116 | let mut body = json!({ |
| 6117 | "model": "gpt-4o", |
| 6118 | "messages": [], |
| 6119 | "max_tokens": 4096, |
| 6120 | }); |
| 6121 | |
| 6122 | apply_provider_token_limit( |
| 6123 | &mut body, |
| 6124 | ProviderKind::Openai, |
| 6125 | "https://api.openai.com/v1", |
| 6126 | "gpt-4o", |
| 6127 | 4096, |
| 6128 | ); |
| 6129 | apply_openai_reasoning_effort(&mut body, ProviderKind::Openai, "gpt-4o", Some("high")); |
| 6130 | |
| 6131 | assert_eq!( |
| 6132 | body.get("max_tokens").and_then(serde_json::Value::as_u64), |
| 6133 | Some(4096) |
| 6134 | ); |
| 6135 | assert!(body.get("max_completion_tokens").is_none()); |
| 6136 | assert!(body.get("reasoning_effort").is_none()); |
| 6137 | } |
| 6138 | |
| 6139 | #[test] |
| 6140 | fn openai_provider_deepseek_compatible_model_keeps_chat_token_field() { |
| 6141 | let mut body = json!({ |
| 6142 | "model": "deepseek-v4-pro", |
| 6143 | "messages": [], |
| 6144 | "max_tokens": 4096, |
| 6145 | }); |
| 6146 | |
| 6147 | apply_provider_token_limit( |
| 6148 | &mut body, |
| 6149 | ProviderKind::Openai, |
| 6150 | "https://api.openai.com/v1", |
| 6151 | "deepseek-v4-pro", |
| 6152 | 4096, |
| 6153 | ); |
| 6154 | apply_openai_reasoning_effort( |
| 6155 | &mut body, |
| 6156 | ProviderKind::Openai, |
| 6157 | "deepseek-v4-pro", |
| 6158 | Some("high"), |
| 6159 | ); |
| 6160 | |
| 6161 | assert_eq!( |
| 6162 | body.get("max_tokens").and_then(serde_json::Value::as_u64), |
| 6163 | Some(4096) |
| 6164 | ); |
| 6165 | assert!(body.get("max_completion_tokens").is_none()); |
| 6166 | assert!(body.get("reasoning_effort").is_none()); |
| 6167 | } |
| 6168 | |
| 6169 | #[test] |
| 6170 | fn deepseek_model_on_openai_provider_still_replays_reasoning_content() { |
| 6171 | // #1739 / #1694: a DeepSeek thinking model pointed at a |
| 6172 | // DeepSeek-compatible endpoint via the generic `openai` provider must |
| 6173 | // still replay reasoning_content, even though the provider itself does |
| 6174 | // not accept the field. Otherwise the thinking-mode API returns 400. |
| 6175 | assert!(should_replay_reasoning_content_for_provider( |
| 6176 | ProviderKind::Openai, |
| 6177 | "deepseek-v4-flash", |
| 6178 | None, |
| 6179 | )); |
| 6180 | assert!(should_replay_reasoning_content_for_provider( |
| 6181 | ProviderKind::Openai, |
| 6182 | "deepseek-v4-pro", |
| 6183 | None, |
| 6184 | )); |
| 6185 | assert!(should_replay_reasoning_content_for_provider( |
| 6186 | ProviderKind::Openai, |
| 6187 | "deepseek-reasoner", |
| 6188 | Some("medium"), |
| 6189 | )); |
| 6190 | // The documented escape hatch still wins over model detection. |
| 6191 | assert!(!should_replay_reasoning_content_for_provider( |
| 6192 | ProviderKind::Openai, |
| 6193 | "deepseek-v4-flash", |
| 6194 | Some("off"), |
| 6195 | )); |
| 6196 | } |
| 6197 | |
| 6198 | #[test] |
| 6199 | fn generic_model_on_openai_provider_still_strips_reasoning_content() { |
| 6200 | // #1542 no-regression guard: a genuine non-DeepSeek model on the |
| 6201 | // openai provider must continue to have reasoning_content stripped. |
| 6202 | assert!(!should_replay_reasoning_content_for_provider( |
| 6203 | ProviderKind::Openai, |
| 6204 | "qwen3-coder", |
| 6205 | None, |
| 6206 | )); |
| 6207 | assert!(!should_replay_reasoning_content_for_provider( |
| 6208 | ProviderKind::Openai, |
| 6209 | "claude-sonnet-4-6", |
| 6210 | None, |
| 6211 | )); |
| 6212 | } |
| 6213 | |
| 6214 | #[test] |
| 6215 | fn suggestive_unknown_model_names_never_authorize_reasoning_replay() { |
| 6216 | for provider in [ |
| 6217 | ProviderKind::Openai, |
| 6218 | ProviderKind::Deepseek, |
| 6219 | ProviderKind::Openrouter, |
| 6220 | ProviderKind::Moonshot, |
| 6221 | ProviderKind::Zai, |
| 6222 | ] { |
| 6223 | for model in [ |
| 6224 | "foo-thinking", |
| 6225 | "foo-reasoner", |
| 6226 | "acme-reasoning", |
| 6227 | "future-reasoner-v9", |
| 6228 | ] { |
| 6229 | assert!( |
| 6230 | !should_replay_reasoning_content_for_provider(provider, model, None), |
| 6231 | "{provider:?} {model}" |
| 6232 | ); |
| 6233 | } |
| 6234 | } |
| 6235 | } |
| 6236 | |
| 6237 | #[test] |
| 6238 | fn stream_classifies_deepseek_model_on_openai_provider_as_reasoning() { |
| 6239 | // #1739: the SSE parser must treat a DeepSeek thinking model on the |
| 6240 | // generic `openai` provider (DeepSeek-compatible endpoint) as a |
| 6241 | // reasoning model, or incoming `reasoning_content` tokens are stored |
| 6242 | // as answer text and the subsequent replay still 400s. |
| 6243 | assert!(is_reasoning_model_for_stream( |
| 6244 | ProviderKind::Openai, |
| 6245 | "deepseek-v4-flash" |
| 6246 | )); |
| 6247 | assert!(is_reasoning_model_for_stream( |
| 6248 | ProviderKind::Openai, |
| 6249 | "deepseek-v4-pro" |
| 6250 | )); |
| 6251 | assert!(is_reasoning_model_for_stream( |
| 6252 | ProviderKind::Openai, |
| 6253 | "deepseek-reasoner" |
| 6254 | )); |
| 6255 | // Native DeepSeek provider was already correct; stays correct. |
| 6256 | assert!(is_reasoning_model_for_stream( |
| 6257 | ProviderKind::Deepseek, |
| 6258 | "deepseek-v4-pro" |
| 6259 | )); |
| 6260 | } |
| 6261 | |
| 6262 | #[test] |
| 6263 | fn zai_tiered_effort_applies_to_glm_5_2_and_glm_5_3_but_not_5_1() { |
| 6264 | let zai = crate::config::DEFAULT_ZAI_BASE_URL; |
| 6265 | // GLM-5.3 and GLM-5.3-Flash inherit GLM-5.2's reasoning_options |
| 6266 | // (effort high/max), so they must take the same tiered wire path. |
| 6267 | for model in [ |
| 6268 | crate::config::ZAI_GLM_5_2_MODEL, |
| 6269 | crate::config::ZAI_GLM_5_3_MODEL, |
| 6270 | crate::config::ZAI_GLM_5_3_FLASH_MODEL, |
| 6271 | ] { |
| 6272 | let mut body = json!({}); |
| 6273 | apply_route_reasoning_controls(&mut body, ProviderKind::Zai, zai, model, Some("max")); |
| 6274 | assert_eq!(body["reasoning_effort"], json!("max"), "{model} at max"); |
| 6275 | |
| 6276 | let mut body = json!({}); |
| 6277 | apply_route_reasoning_controls(&mut body, ProviderKind::Zai, zai, model, Some("high")); |
| 6278 | assert_eq!(body["reasoning_effort"], json!("high"), "{model} at high"); |
| 6279 | } |
| 6280 | |
| 6281 | // GLM-5.1 and GLM-5-Turbo keep only the generic thinking control. |
| 6282 | for model in [ |
| 6283 | crate::config::ZAI_GLM_5_1_MODEL, |
| 6284 | crate::config::ZAI_GLM_5_TURBO_MODEL, |
| 6285 | ] { |
| 6286 | let mut body = json!({}); |
| 6287 | apply_route_reasoning_controls(&mut body, ProviderKind::Zai, zai, model, Some("max")); |
| 6288 | assert!( |
| 6289 | body.get("reasoning_effort").is_none(), |
| 6290 | "{model} must not receive tiered effort" |
| 6291 | ); |
| 6292 | } |
| 6293 | |
| 6294 | // A compatible gateway is not evidence of the Z.ai dialect, for 5.3 |
| 6295 | // exactly as for 5.2. |
| 6296 | let mut body = json!({"thinking": {"type": "enabled"}}); |
| 6297 | apply_route_reasoning_controls( |
| 6298 | &mut body, |
| 6299 | ProviderKind::Zai, |
| 6300 | "https://gateway.example.com/v1", |
| 6301 | crate::config::ZAI_GLM_5_3_MODEL, |
| 6302 | Some("max"), |
| 6303 | ); |
| 6304 | assert!(body.get("reasoning_effort").is_none()); |
| 6305 | assert!(body.get("thinking").is_none()); |
| 6306 | } |
| 6307 | |
| 6308 | #[test] |
| 6309 | fn zai_forced_thinking_models_never_send_thinking_disabled() { |
| 6310 | // BigModel and Z.ai document GLM-5.3 / GLM-5.3-Flash as forced-thinking: |
| 6311 | // `thinking.type: "disabled"` errors, effort accepts only low/high/max, |
| 6312 | // and the migration note for a former `disabled` payload is |
| 6313 | // `enabled` + `reasoning_effort: "low"`. Both hosts of the first-party |
| 6314 | // open platform — api.z.ai and open.bigmodel.cn — get the rewrite. |
| 6315 | for zai in [ |
| 6316 | crate::config::DEFAULT_ZAI_BASE_URL, |
| 6317 | "https://open.bigmodel.cn/api/paas/v4", |
| 6318 | ] { |
| 6319 | for model in [ |
| 6320 | crate::config::ZAI_GLM_5_3_MODEL, |
| 6321 | crate::config::ZAI_GLM_5_3_FLASH_MODEL, |
| 6322 | ] { |
| 6323 | let mut body = json!({}); |
| 6324 | apply_route_reasoning_controls( |
| 6325 | &mut body, |
| 6326 | ProviderKind::Zai, |
| 6327 | zai, |
| 6328 | model, |
| 6329 | Some("off"), |
| 6330 | ); |
| 6331 | assert_eq!( |
| 6332 | body["thinking"]["type"], |
| 6333 | json!("enabled"), |
| 6334 | "{model} must not send the rejected disabled toggle" |
| 6335 | ); |
| 6336 | assert_eq!( |
| 6337 | body["reasoning_effort"], |
| 6338 | json!("low"), |
| 6339 | "{model} off becomes low" |
| 6340 | ); |
| 6341 | |
| 6342 | let mut body = json!({}); |
| 6343 | apply_route_reasoning_controls( |
| 6344 | &mut body, |
| 6345 | ProviderKind::Zai, |
| 6346 | zai, |
| 6347 | model, |
| 6348 | Some("low"), |
| 6349 | ); |
| 6350 | assert_eq!( |
| 6351 | body["reasoning_effort"], |
| 6352 | json!("low"), |
| 6353 | "{model} low is native" |
| 6354 | ); |
| 6355 | |
| 6356 | let mut body = json!({}); |
| 6357 | apply_route_reasoning_controls( |
| 6358 | &mut body, |
| 6359 | ProviderKind::Zai, |
| 6360 | zai, |
| 6361 | model, |
| 6362 | Some("medium"), |
| 6363 | ); |
| 6364 | assert_eq!( |
| 6365 | body["reasoning_effort"], |
| 6366 | json!("high"), |
| 6367 | "{model} medium maps to high" |
| 6368 | ); |
| 6369 | |
| 6370 | let mut body = json!({}); |
| 6371 | apply_route_reasoning_controls( |
| 6372 | &mut body, |
| 6373 | ProviderKind::Zai, |
| 6374 | zai, |
| 6375 | model, |
| 6376 | Some("max"), |
| 6377 | ); |
| 6378 | assert_eq!( |
| 6379 | body["reasoning_effort"], |
| 6380 | json!("max"), |
| 6381 | "{model} max stays max" |
| 6382 | ); |
| 6383 | |
| 6384 | // Unknown legacy values leave the field omitted so the API keeps |
| 6385 | // its documented default; nothing may reintroduce `disabled`. |
| 6386 | let mut body = json!({}); |
| 6387 | apply_route_reasoning_controls( |
| 6388 | &mut body, |
| 6389 | ProviderKind::Zai, |
| 6390 | zai, |
| 6391 | model, |
| 6392 | Some("auto"), |
| 6393 | ); |
| 6394 | assert!( |
| 6395 | body.get("reasoning_effort").is_none(), |
| 6396 | "{model} auto stays omitted" |
| 6397 | ); |
| 6398 | assert_ne!(body["thinking"]["type"], json!("disabled")); |
| 6399 | } |
| 6400 | |
| 6401 | // GLM-5.2 honours the generic disabled toggle on both hosts. On |
| 6402 | // BigModel that toggle now reaches the API (its docs still list |
| 6403 | // `disabled` for GLM-5.2) instead of being stripped by the |
| 6404 | // fail-closed gateway path. |
| 6405 | let mut body = json!({}); |
| 6406 | apply_route_reasoning_controls( |
| 6407 | &mut body, |
| 6408 | ProviderKind::Zai, |
| 6409 | zai, |
| 6410 | crate::config::ZAI_GLM_5_2_MODEL, |
| 6411 | Some("off"), |
| 6412 | ); |
| 6413 | assert_eq!(body["thinking"]["type"], json!("disabled")); |
| 6414 | assert!(body.get("reasoning_effort").is_none()); |
| 6415 | } |
| 6416 | } |
| 6417 | |
| 6418 | #[test] |
| 6419 | fn zai_bigmodel_adjacent_routes_stay_fail_closed() { |
| 6420 | // BigModel's `/preview` product and a plain-http neighbor are not the |
| 6421 | // documented Chat dialect, so they keep the gateway treatment: no |
| 6422 | // Z.ai reasoning fields, including on a forced-thinking model. |
| 6423 | for neighboring_route in [ |
| 6424 | "https://open.bigmodel.cn/api/paas/v4/preview", |
| 6425 | "http://open.bigmodel.cn/api/paas/v4", |
| 6426 | ] { |
| 6427 | let mut body = json!({"thinking": {"type": "enabled"}}); |
| 6428 | apply_route_reasoning_controls( |
| 6429 | &mut body, |
| 6430 | ProviderKind::Zai, |
| 6431 | neighboring_route, |
| 6432 | crate::config::ZAI_GLM_5_3_MODEL, |
| 6433 | Some("max"), |
| 6434 | ); |
| 6435 | assert!( |
| 6436 | body.get("reasoning_effort").is_none(), |
| 6437 | "{neighboring_route} must not gain tiered effort" |
| 6438 | ); |
| 6439 | assert!( |
| 6440 | body.get("thinking").is_none(), |
| 6441 | "{neighboring_route} must not keep the Z.ai thinking object" |
| 6442 | ); |
| 6443 | } |
| 6444 | } |
| 6445 | |
| 6446 | #[test] |
| 6447 | fn stream_classifies_known_large_reasoning_models_as_reasoning() { |
| 6448 | // Xiaomi MiMo and OpenRouter/Qwen/Trinity can stream private reasoning through a |
| 6449 | // `reasoning` delta without using a DeepSeek-looking model name. The |
| 6450 | // renderer must still route that field into Thinking cells instead |
| 6451 | // of plain assistant prose. |
| 6452 | assert!( |
| 6453 | is_reasoning_model_for_stream(ProviderKind::XiaomiMimo, "mimo-v2.5-pro"), |
| 6454 | "mimo-v2.5-pro should stream reasoning as thinking on Xiaomi MiMo" |
| 6455 | ); |
| 6456 | assert!( |
| 6457 | is_reasoning_model_for_stream(ProviderKind::Arcee, "trinity-large-thinking"), |
| 6458 | "trinity-large-thinking should stream reasoning as thinking on direct Arcee" |
| 6459 | ); |
| 6460 | assert!( |
| 6461 | is_reasoning_model_for_stream(ProviderKind::Zai, "GLM-5.2"), |
| 6462 | "GLM-5.2 should stream reasoning_content as thinking on direct Z.ai" |
| 6463 | ); |
| 6464 | assert!( |
| 6465 | is_reasoning_model_for_stream(ProviderKind::Zai, "GLM-5.3"), |
| 6466 | "GLM-5.3 inherits GLM-5.2's reasoning capability on direct Z.ai" |
| 6467 | ); |
| 6468 | for model in [ |
| 6469 | "arcee-ai/trinity-large-thinking", |
| 6470 | "minimax/minimax-m3", |
| 6471 | "xiaomi/mimo-v2.5-pro", |
| 6472 | ] { |
| 6473 | assert!( |
| 6474 | is_reasoning_model_for_stream(ProviderKind::Openrouter, model), |
| 6475 | "{model} should stream reasoning as thinking on OpenRouter" |
| 6476 | ); |
| 6477 | } |
| 6478 | } |
| 6479 | |
| 6480 | #[test] |
| 6481 | fn stream_surfaces_reasoning_fields_on_every_unconfigured_route() { |
| 6482 | // #6501: grok-4.7 on xAI and a MiMo id newer than the offline catalog |
| 6483 | // streamed their reasoning as answer prose because the stream gate |
| 6484 | // required a provider allowlist AND a catalog row. The reasoning |
| 6485 | // fields are reasoning on every route that sends them; only an |
| 6486 | // explicit `reasoning_stream_style = "none"` restores pass-through. |
| 6487 | for (provider, model) in [ |
| 6488 | (ProviderKind::Xai, "grok-4.7"), |
| 6489 | (ProviderKind::Xai, "grok-4.6"), |
| 6490 | (ProviderKind::XiaomiMimo, "mimo-v2.7-pro-unreleased"), |
| 6491 | (ProviderKind::XiaomiMimo, "mimo-v2.6-pro"), |
| 6492 | (ProviderKind::Openai, "qwen3-coder"), |
| 6493 | (ProviderKind::Deepseek, "qwen3-coder"), |
| 6494 | (ProviderKind::Stepfun, "step-3.5"), |
| 6495 | (ProviderKind::Together, "any/model"), |
| 6496 | (ProviderKind::Ollama, "gpt-oss:20b"), |
| 6497 | (ProviderKind::Codewhale, "grok-4.7"), |
| 6498 | ] { |
| 6499 | assert!( |
| 6500 | is_reasoning_model_for_stream(provider, model), |
| 6501 | "{provider:?} {model} must surface reasoning fields as Thinking" |
| 6502 | ); |
| 6503 | } |
| 6504 | assert_eq!( |
| 6505 | super::reasoning_stream_style_for_stream(ProviderKind::Xai, "grok-4.7", Some("none")), |
| 6506 | ReasoningStreamStyle::None, |
| 6507 | "an explicit route override still wins" |
| 6508 | ); |
| 6509 | } |
| 6510 | |
| 6511 | #[test] |
| 6512 | fn xiaomi_mimo_replays_reasoning_for_every_model_id() { |
| 6513 | // #6501: MiMo returns 400 when a tool-call turn omits the assistant |
| 6514 | // `reasoning_content`, including ids newer than the offline catalog. |
| 6515 | for model in ["mimo-v2.5-pro", "mimo-v2.6-pro", "mimo-v2.7-pro-unreleased"] { |
| 6516 | assert!( |
| 6517 | should_replay_reasoning_content_for_provider(ProviderKind::XiaomiMimo, model, None), |
| 6518 | "{model}" |
| 6519 | ); |
| 6520 | } |
| 6521 | assert!(!should_replay_reasoning_content_for_provider( |
| 6522 | ProviderKind::XiaomiMimo, |
| 6523 | "mimo-v2.6-pro", |
| 6524 | Some("off"), |
| 6525 | )); |
| 6526 | } |
| 6527 | |
| 6528 | #[test] |
| 6529 | fn stream_display_does_not_authorize_reasoning_replay() { |
| 6530 | // #1542 stays fixed: rendering a reasoning delta as Thinking must not |
| 6531 | // make a provider that rejects `reasoning_content` receive it back. |
| 6532 | for (provider, model) in [ |
| 6533 | (ProviderKind::Openai, "qwen3-coder"), |
| 6534 | (ProviderKind::Openai, "claude-sonnet-4-6"), |
| 6535 | (ProviderKind::Xai, "grok-4.7"), |
| 6536 | ] { |
| 6537 | assert!(is_reasoning_model_for_stream(provider, model)); |
| 6538 | assert!( |
| 6539 | !should_replay_reasoning_content_for_provider(provider, model, None), |
| 6540 | "{provider:?} {model}" |
| 6541 | ); |
| 6542 | } |
| 6543 | } |
| 6544 | } |
| 6545 | |
| 6546 | #[cfg(test)] |
| 6547 | mod image_block_wire_tests { |
| 6548 | //! The OpenAI-compatible projection of [`ContentBlock::ImageUrl`]. |
| 6549 | //! |
| 6550 | //! Chat Completions is the wire format behind the large majority of |
| 6551 | //! CodeWhale's provider routes, so a regression here is a regression for |
| 6552 | //! most of them at once. The shape is fixed by OpenAI's spec: a `user` |
| 6553 | //! message whose `content` is an array of parts, with the image as |
| 6554 | //! `{"type":"image_url","image_url":{"url":…}}`. |
| 6555 | use super::{ProviderKind, build_chat_messages, build_chat_wire_body}; |
| 6556 | use codewhale_models::Role; |
| 6557 | use codewhale_models::{ContentBlock, ImageUrlContent, Message, MessageRequest}; |
| 6558 | |
| 6559 | const DATA_URL: &str = "data:image/png;base64,QUJD"; |
| 6560 | |
| 6561 | #[test] |
| 6562 | fn compaction_checkpoint_keeps_complete_tool_round_on_wire() { |
| 6563 | let mut messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6564 | {"role":"user","content":[{"type":"text","text":"Analyze the data"}]}, |
| 6565 | {"role":"assistant","content":[{"type":"tool_use","id":"call_1","name":"read","input":{"path":"a.txt"}}]}, |
| 6566 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"ready"}]}, |
| 6567 | {"role":"assistant","content":[{"type":"tool_use","id":"call_2","name":"read","input":{"path":"b.txt"}}]}, |
| 6568 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"call_2","content":"done"}]} |
| 6569 | ])).unwrap(); |
| 6570 | let summary = codewhale_models::SystemPrompt::Text( |
| 6571 | crate::compaction::build_compaction_summary_block_text("Compacted summary", ""), |
| 6572 | ); |
| 6573 | messages.push(crate::compaction::compaction_checkpoint_message(&summary)); |
| 6574 | crate::runtime_handoff::replace_agent_topology_checkpoint(&mut messages, &[]); |
| 6575 | let stored = messages.clone(); |
| 6576 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6577 | assert_eq!( |
| 6578 | messages, stored, |
| 6579 | "request construction must not alter saved history" |
| 6580 | ); |
| 6581 | let roles: Vec<&str> = wire |
| 6582 | .iter() |
| 6583 | .map(|message| message["role"].as_str().unwrap()) |
| 6584 | .collect(); |
| 6585 | assert_eq!(roles, ["user", "assistant", "tool", "assistant", "tool"]); |
| 6586 | let prompt = wire[0]["content"].as_str().unwrap(); |
| 6587 | assert!(prompt.contains("Compacted summary")); |
| 6588 | assert!(prompt.contains("Analyze the data")); |
| 6589 | assert!(prompt.contains("codewhale.agent_topology.v1")); |
| 6590 | assert_eq!(wire[1]["tool_calls"][0]["id"], "call_1"); |
| 6591 | assert_eq!(wire[2]["tool_call_id"], "call_1"); |
| 6592 | assert_eq!(wire[3]["tool_calls"][0]["id"], "call_2"); |
| 6593 | assert_eq!(wire[4]["tool_call_id"], "call_2"); |
| 6594 | |
| 6595 | messages.push( |
| 6596 | serde_json::from_value(serde_json::json!({ |
| 6597 | "role":"assistant","content":[{"type":"text","text":"Analysis complete"}] |
| 6598 | })) |
| 6599 | .unwrap(), |
| 6600 | ); |
| 6601 | messages.push( |
| 6602 | serde_json::from_value(serde_json::json!({ |
| 6603 | "role":"user","content":[{"type":"text","text":"What happened next?"}] |
| 6604 | })) |
| 6605 | .unwrap(), |
| 6606 | ); |
| 6607 | let later_wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6608 | let later_roles: Vec<&str> = later_wire |
| 6609 | .iter() |
| 6610 | .map(|message| message["role"].as_str().unwrap()) |
| 6611 | .collect(); |
| 6612 | assert_eq!( |
| 6613 | later_roles, |
| 6614 | [ |
| 6615 | "user", |
| 6616 | "assistant", |
| 6617 | "tool", |
| 6618 | "assistant", |
| 6619 | "tool", |
| 6620 | "assistant", |
| 6621 | "user" |
| 6622 | ] |
| 6623 | ); |
| 6624 | assert_eq!(later_wire[6]["content"], "What happened next?"); |
| 6625 | |
| 6626 | let restored = crate::compaction::restore_compaction_checkpoint( |
| 6627 | crate::runtime_handoff::project_owned_messages_for_restore(messages.clone()), |
| 6628 | Some(&summary), |
| 6629 | ); |
| 6630 | let restored_wire = build_chat_messages(None, &restored, "gpt-4o"); |
| 6631 | let restored_roles: Vec<&str> = restored_wire |
| 6632 | .iter() |
| 6633 | .map(|message| message["role"].as_str().unwrap()) |
| 6634 | .collect(); |
| 6635 | assert_eq!(restored_roles, later_roles); |
| 6636 | assert!( |
| 6637 | restored_wire[0]["content"] |
| 6638 | .as_str() |
| 6639 | .unwrap() |
| 6640 | .contains("restored Agent topology checkpoint") |
| 6641 | ); |
| 6642 | assert_eq!(restored_wire[6]["content"], "What happened next?"); |
| 6643 | } |
| 6644 | |
| 6645 | #[test] |
| 6646 | fn quoted_compaction_marker_does_not_reorder_user_wire_messages() { |
| 6647 | // The current marker, and the legacy one older sessions still carry. |
| 6648 | for marker in [ |
| 6649 | crate::compaction::COMPACTION_SUMMARY_MARKER, |
| 6650 | crate::compaction::LEGACY_V2_COMPACTION_SUMMARY_MARKER, |
| 6651 | ] { |
| 6652 | let quote = format!("Please explain: {marker}"); |
| 6653 | let messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6654 | {"role":"user","content":[{"type":"text","text":"First question"}]}, |
| 6655 | {"role":"assistant","content":[{"type":"text","text":"First answer"}]}, |
| 6656 | {"role":"user","content":[{"type":"text","text":quote}]}, |
| 6657 | {"role":"assistant","content":[{"type":"text","text":"It introduces a summary."}]}, |
| 6658 | {"role":"user","content":[{"type":"text","text":"Follow-up question"}]} |
| 6659 | ])) |
| 6660 | .unwrap(); |
| 6661 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6662 | let roles: Vec<&str> = wire |
| 6663 | .iter() |
| 6664 | .map(|message| message["role"].as_str().unwrap()) |
| 6665 | .collect(); |
| 6666 | assert_eq!(roles, ["user", "assistant", "user", "assistant", "user"]); |
| 6667 | assert_eq!(wire[0]["content"], "First question"); |
| 6668 | assert_eq!(wire[2]["content"], quote); |
| 6669 | assert_eq!(wire[4]["content"], "Follow-up question"); |
| 6670 | } |
| 6671 | let messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6672 | {"role":"user","content":[{"type":"text","text":"First question"}]}, |
| 6673 | {"role":"assistant","content":[{"type":"text","text":"First answer"}]}, |
| 6674 | {"role":"user","content":[{"type":"text","text":"placeholder"}]}, |
| 6675 | {"role":"assistant","content":[{"type":"text","text":"It introduces a summary."}]}, |
| 6676 | {"role":"user","content":[{"type":"text","text":"Follow-up question"}]} |
| 6677 | ])) |
| 6678 | .unwrap(); |
| 6679 | let roles = ["user", "assistant", "user", "assistant", "user"]; |
| 6680 | |
| 6681 | let quoted_exact_header = crate::compaction::build_compaction_summary_block_text( |
| 6682 | "This text was pasted by a user", |
| 6683 | "", |
| 6684 | ); |
| 6685 | let mut with_exact_quote = messages.clone(); |
| 6686 | with_exact_quote[2] = Message { |
| 6687 | role: Role::User, |
| 6688 | content: vec![ContentBlock::Text { |
| 6689 | text: quoted_exact_header.clone(), |
| 6690 | cache_control: None, |
| 6691 | }], |
| 6692 | }; |
| 6693 | let exact_wire = build_chat_messages(None, &with_exact_quote, "gpt-4o"); |
| 6694 | let exact_roles: Vec<&str> = exact_wire |
| 6695 | .iter() |
| 6696 | .map(|message| message["role"].as_str().unwrap()) |
| 6697 | .collect(); |
| 6698 | assert_eq!(exact_roles, roles); |
| 6699 | assert_eq!(exact_wire[2]["content"], quoted_exact_header); |
| 6700 | |
| 6701 | let mut after_tool: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6702 | {"role":"user","content":[{"type":"text","text":"Read first"}]}, |
| 6703 | {"role":"assistant","content":[{"type":"tool_use","id":"call_1","name":"read","input":{"path":"a.txt"}}]}, |
| 6704 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"contents"}]}, |
| 6705 | {"role":"user","content":[{"type":"text","text":"placeholder"}]}, |
| 6706 | {"role":"assistant","content":[{"type":"text","text":"Answer"}]} |
| 6707 | ])).unwrap(); |
| 6708 | after_tool[3] = Message { |
| 6709 | role: Role::User, |
| 6710 | content: vec![ContentBlock::Text { |
| 6711 | text: quoted_exact_header.clone(), |
| 6712 | cache_control: None, |
| 6713 | }], |
| 6714 | }; |
| 6715 | let after_tool_wire = build_chat_messages(None, &after_tool, "gpt-4o"); |
| 6716 | let after_tool_roles: Vec<&str> = after_tool_wire |
| 6717 | .iter() |
| 6718 | .map(|message| message["role"].as_str().unwrap()) |
| 6719 | .collect(); |
| 6720 | assert_eq!( |
| 6721 | after_tool_roles, |
| 6722 | ["user", "assistant", "tool", "user", "assistant"] |
| 6723 | ); |
| 6724 | assert_eq!(after_tool_wire[3]["content"], quoted_exact_header); |
| 6725 | } |
| 6726 | |
| 6727 | #[test] |
| 6728 | fn topology_checkpoint_after_tool_result_keeps_wire_tool_pair() { |
| 6729 | let mut messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6730 | {"role":"user","content":[{"type":"text","text":"Read the file"}]}, |
| 6731 | {"role":"assistant","content":[{"type":"tool_use","id":"call_1","name":"read","input":{"path":"a.txt"}}]}, |
| 6732 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"contents"}]} |
| 6733 | ])).unwrap(); |
| 6734 | crate::runtime_handoff::replace_agent_topology_checkpoint(&mut messages, &[]); |
| 6735 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6736 | let roles: Vec<&str> = wire |
| 6737 | .iter() |
| 6738 | .map(|message| message["role"].as_str().unwrap()) |
| 6739 | .collect(); |
| 6740 | assert_eq!(roles, ["user", "assistant", "tool"]); |
| 6741 | assert!( |
| 6742 | wire[0]["content"] |
| 6743 | .as_str() |
| 6744 | .unwrap() |
| 6745 | .contains("agent_topology_v1") |
| 6746 | ); |
| 6747 | assert_eq!(wire[1]["tool_calls"][0]["id"], "call_1"); |
| 6748 | assert_eq!(wire[2]["tool_call_id"], "call_1"); |
| 6749 | } |
| 6750 | |
| 6751 | #[test] |
| 6752 | fn compaction_without_retained_user_has_one_wire_user_before_tools() { |
| 6753 | let mut messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6754 | {"role":"assistant","content":[{"type":"tool_use","id":"call_1","name":"read","input":{"path":"a.txt"}}]}, |
| 6755 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"contents"}]} |
| 6756 | ])).unwrap(); |
| 6757 | let summary = codewhale_models::SystemPrompt::Text( |
| 6758 | crate::compaction::build_compaction_summary_block_text("Summary", ""), |
| 6759 | ); |
| 6760 | messages.push(crate::compaction::compaction_checkpoint_message(&summary)); |
| 6761 | crate::runtime_handoff::replace_agent_topology_checkpoint(&mut messages, &[]); |
| 6762 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6763 | let roles: Vec<&str> = wire |
| 6764 | .iter() |
| 6765 | .map(|message| message["role"].as_str().unwrap()) |
| 6766 | .collect(); |
| 6767 | assert_eq!(roles, ["user", "assistant", "tool"]); |
| 6768 | let prompt = wire[0]["content"].as_str().unwrap(); |
| 6769 | assert!(prompt.contains("Summary")); |
| 6770 | assert!(prompt.contains("agent_topology_v1")); |
| 6771 | assert_eq!(wire[2]["tool_call_id"], "call_1"); |
| 6772 | } |
| 6773 | |
| 6774 | #[test] |
| 6775 | fn compaction_after_unanswered_user_prompt_has_one_wire_user() { |
| 6776 | let mut messages: Vec<Message> = serde_json::from_value(serde_json::json!([ |
| 6777 | {"role":"user","content":[{"type":"text","text":"Please continue"}]} |
| 6778 | ])) |
| 6779 | .unwrap(); |
| 6780 | let summary = codewhale_models::SystemPrompt::Text( |
| 6781 | crate::compaction::build_compaction_summary_block_text("Earlier work", ""), |
| 6782 | ); |
| 6783 | messages.push(crate::compaction::compaction_checkpoint_message(&summary)); |
| 6784 | crate::runtime_handoff::replace_agent_topology_checkpoint(&mut messages, &[]); |
| 6785 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 6786 | assert_eq!(wire.len(), 1); |
| 6787 | assert_eq!(wire[0]["role"], "user"); |
| 6788 | let prompt = wire[0]["content"].as_str().unwrap(); |
| 6789 | assert!(prompt.contains("Please continue")); |
| 6790 | assert!(prompt.contains("Earlier work")); |
| 6791 | assert!(prompt.contains("agent_topology_v1")); |
| 6792 | } |
| 6793 | |
| 6794 | fn fixture_tool_use(id: &str) -> ContentBlock { |
| 6795 | ContentBlock::ToolUse { |
| 6796 | execution_id: None, |
| 6797 | id: id.to_string(), |
| 6798 | name: "read".to_string(), |
| 6799 | input: serde_json::json!({"path": format!("{id}.txt")}), |
| 6800 | caller: None, |
| 6801 | thought_signature: None, |
| 6802 | } |
| 6803 | } |
| 6804 | |
| 6805 | fn fixture_tool_result(id: &str) -> Message { |
| 6806 | Message { |
| 6807 | role: Role::User, |
| 6808 | content: vec![ContentBlock::ToolResult { |
| 6809 | execution_id: None, |
| 6810 | tool_use_id: id.to_string(), |
| 6811 | content: format!("result for {id}"), |
| 6812 | is_error: Some(false), |
| 6813 | content_blocks: None, |
| 6814 | }], |
| 6815 | } |
| 6816 | } |
| 6817 | |
| 6818 | fn request_with_image() -> MessageRequest { |
| 6819 | MessageRequest { |
| 6820 | model: "gpt-4o".to_string(), |
| 6821 | messages: vec![Message { |
| 6822 | role: Role::User, |
| 6823 | content: vec![ |
| 6824 | ContentBlock::Text { |
| 6825 | text: "what is in this screenshot?".to_string(), |
| 6826 | cache_control: None, |
| 6827 | }, |
| 6828 | ContentBlock::ImageUrl { |
| 6829 | image_url: ImageUrlContent { |
| 6830 | url: DATA_URL.to_string(), |
| 6831 | }, |
| 6832 | }, |
| 6833 | ], |
| 6834 | }], |
| 6835 | max_tokens: 128, |
| 6836 | system: None, |
| 6837 | tools: None, |
| 6838 | tool_choice: None, |
| 6839 | metadata: None, |
| 6840 | thinking: None, |
| 6841 | reasoning_effort: None, |
| 6842 | stream: None, |
| 6843 | temperature: None, |
| 6844 | top_p: None, |
| 6845 | } |
| 6846 | } |
| 6847 | |
| 6848 | #[test] |
| 6849 | fn user_image_becomes_a_multimodal_parts_array() { |
| 6850 | let body = build_chat_wire_body( |
| 6851 | &request_with_image(), |
| 6852 | ProviderKind::Openai, |
| 6853 | "https://api.openai.com/v1", |
| 6854 | false, |
| 6855 | None, |
| 6856 | ) |
| 6857 | .expect("wire body"); |
| 6858 | |
| 6859 | let messages = body.body["messages"].as_array().expect("messages"); |
| 6860 | let user = messages |
| 6861 | .iter() |
| 6862 | .find(|message| message["role"] == "user") |
| 6863 | .expect("a user message"); |
| 6864 | let parts = user["content"] |
| 6865 | .as_array() |
| 6866 | .expect("content must be a parts array once an image is present, not a bare string"); |
| 6867 | |
| 6868 | let image = parts |
| 6869 | .iter() |
| 6870 | .find(|part| part["type"] == "image_url") |
| 6871 | .expect("an image_url part"); |
| 6872 | assert_eq!(image["image_url"]["url"], DATA_URL); |
| 6873 | |
| 6874 | let text = parts |
| 6875 | .iter() |
| 6876 | .find(|part| part["type"] == "text") |
| 6877 | .expect("the accompanying text part"); |
| 6878 | assert!( |
| 6879 | text["text"] |
| 6880 | .as_str() |
| 6881 | .expect("text") |
| 6882 | .contains("what is in this screenshot?"), |
| 6883 | "the question must survive alongside the image: {user}" |
| 6884 | ); |
| 6885 | } |
| 6886 | |
| 6887 | #[test] |
| 6888 | fn deepseek_vision_exp_uses_chat_image_url_request_shape() { |
| 6889 | let mut request = request_with_image(); |
| 6890 | request.model = "deepseek-v4-flash-vision-exp".to_string(); |
| 6891 | |
| 6892 | let body = build_chat_wire_body( |
| 6893 | &request, |
| 6894 | ProviderKind::Deepseek, |
| 6895 | "https://api.deepseek.com/beta", |
| 6896 | false, |
| 6897 | None, |
| 6898 | ) |
| 6899 | .expect("DeepSeek vision wire body"); |
| 6900 | |
| 6901 | assert_eq!(body.body["model"], "deepseek-v4-flash-vision-exp"); |
| 6902 | let messages = body.body["messages"].as_array().expect("messages"); |
| 6903 | let user = messages |
| 6904 | .iter() |
| 6905 | .find(|message| message["role"] == "user") |
| 6906 | .expect("a user message"); |
| 6907 | let parts = user["content"] |
| 6908 | .as_array() |
| 6909 | .expect("DeepSeek vision content must use multimodal parts"); |
| 6910 | |
| 6911 | assert!(parts.iter().any(|part| { |
| 6912 | part["type"] == "text" && part["text"] == "what is in this screenshot?" |
| 6913 | })); |
| 6914 | assert!( |
| 6915 | parts.iter().any(|part| { |
| 6916 | part["type"] == "image_url" && part["image_url"]["url"] == DATA_URL |
| 6917 | }) |
| 6918 | ); |
| 6919 | } |
| 6920 | |
| 6921 | #[test] |
| 6922 | fn a_message_with_no_image_keeps_its_plain_string_content() { |
| 6923 | // Promoting every user turn to a parts array would change the request |
| 6924 | // bytes for every text-only route, and with them the prompt-cache |
| 6925 | // prefix. Images must be the only thing that triggers the array form. |
| 6926 | let mut request = request_with_image(); |
| 6927 | request.messages[0] |
| 6928 | .content |
| 6929 | .retain(|block| !matches!(block, ContentBlock::ImageUrl { .. })); |
| 6930 | |
| 6931 | let body = build_chat_wire_body( |
| 6932 | &request, |
| 6933 | ProviderKind::Openai, |
| 6934 | "https://api.openai.com/v1", |
| 6935 | false, |
| 6936 | None, |
| 6937 | ) |
| 6938 | .expect("wire body"); |
| 6939 | |
| 6940 | let messages = body.body["messages"].as_array().expect("messages"); |
| 6941 | let user = messages |
| 6942 | .iter() |
| 6943 | .find(|message| message["role"] == "user") |
| 6944 | .expect("a user message"); |
| 6945 | assert!( |
| 6946 | user["content"].is_string(), |
| 6947 | "text-only turns must stay a plain string: {user}" |
| 6948 | ); |
| 6949 | } |
| 6950 | |
| 6951 | #[test] |
| 6952 | fn tool_result_image_follows_its_tool_message_as_multimodal_user_content() { |
| 6953 | let mut request = request_with_image(); |
| 6954 | request.messages = vec![ |
| 6955 | Message { |
| 6956 | role: Role::Assistant, |
| 6957 | content: vec![ContentBlock::ToolUse { |
| 6958 | execution_id: None, |
| 6959 | id: "call_image_1".to_string(), |
| 6960 | name: "read".to_string(), |
| 6961 | input: serde_json::json!({"path": "shot.png"}), |
| 6962 | caller: None, |
| 6963 | thought_signature: None, |
| 6964 | }], |
| 6965 | }, |
| 6966 | Message { |
| 6967 | role: Role::User, |
| 6968 | content: vec![ContentBlock::ToolResult { |
| 6969 | execution_id: None, |
| 6970 | tool_use_id: "call_image_1".to_string(), |
| 6971 | content: "screenshot captured".to_string(), |
| 6972 | is_error: Some(false), |
| 6973 | content_blocks: Some(vec![serde_json::json!({ |
| 6974 | "type": "image", |
| 6975 | "mime_type": "image/png", |
| 6976 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==", |
| 6977 | })]), |
| 6978 | }], |
| 6979 | }, |
| 6980 | ]; |
| 6981 | |
| 6982 | let body = build_chat_wire_body( |
| 6983 | &request, |
| 6984 | ProviderKind::Openai, |
| 6985 | "https://api.openai.com/v1", |
| 6986 | false, |
| 6987 | None, |
| 6988 | ) |
| 6989 | .expect("wire body"); |
| 6990 | let messages = body.body["messages"].as_array().expect("messages"); |
| 6991 | let tool_index = messages |
| 6992 | .iter() |
| 6993 | .position(|message| message["role"] == "tool") |
| 6994 | .expect("tool result"); |
| 6995 | assert_eq!(messages[tool_index]["tool_call_id"], "call_image_1"); |
| 6996 | assert_eq!(messages[tool_index]["content"], "screenshot captured"); |
| 6997 | |
| 6998 | let image_message = &messages[tool_index + 1]; |
| 6999 | assert_eq!(image_message["role"], "user"); |
| 7000 | let parts = image_message["content"].as_array().expect("image parts"); |
| 7001 | assert_eq!(parts[1]["type"], "image_url"); |
| 7002 | assert_eq!( |
| 7003 | parts[1]["image_url"]["url"], |
| 7004 | "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==" |
| 7005 | ); |
| 7006 | assert!( |
| 7007 | parts[0]["text"] |
| 7008 | .as_str() |
| 7009 | .is_some_and(|text| text.contains("read") && text.contains("call_image_1")) |
| 7010 | ); |
| 7011 | } |
| 7012 | |
| 7013 | #[test] |
| 7014 | fn tool_result_images_follow_the_entire_tool_call_batch() { |
| 7015 | let mut request = request_with_image(); |
| 7016 | request.messages = vec![ |
| 7017 | Message { |
| 7018 | role: Role::Assistant, |
| 7019 | content: vec![ |
| 7020 | ContentBlock::ToolUse { |
| 7021 | execution_id: None, |
| 7022 | id: "call_image_1".to_string(), |
| 7023 | name: "read".to_string(), |
| 7024 | input: serde_json::json!({"path": "first.png"}), |
| 7025 | caller: None, |
| 7026 | thought_signature: None, |
| 7027 | }, |
| 7028 | ContentBlock::ToolUse { |
| 7029 | execution_id: None, |
| 7030 | id: "call_image_2".to_string(), |
| 7031 | name: "read".to_string(), |
| 7032 | input: serde_json::json!({"path": "second.png"}), |
| 7033 | caller: None, |
| 7034 | thought_signature: None, |
| 7035 | }, |
| 7036 | ], |
| 7037 | }, |
| 7038 | Message { |
| 7039 | role: Role::User, |
| 7040 | content: vec![ContentBlock::ToolResult { |
| 7041 | execution_id: None, |
| 7042 | tool_use_id: "call_image_1".to_string(), |
| 7043 | content: "first screenshot captured".to_string(), |
| 7044 | is_error: Some(false), |
| 7045 | content_blocks: Some(vec![serde_json::json!({ |
| 7046 | "type": "image", |
| 7047 | "mime_type": "image/png", |
| 7048 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==", |
| 7049 | })]), |
| 7050 | }], |
| 7051 | }, |
| 7052 | Message { |
| 7053 | role: Role::User, |
| 7054 | content: vec![ContentBlock::ToolResult { |
| 7055 | execution_id: None, |
| 7056 | tool_use_id: "call_image_2".to_string(), |
| 7057 | content: "second screenshot captured".to_string(), |
| 7058 | is_error: Some(false), |
| 7059 | content_blocks: Some(vec![serde_json::json!({ |
| 7060 | "type": "image", |
| 7061 | "mime_type": "image/png", |
| 7062 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGNgYPj/HwADAgH/5ncLrgAAAABJRU5ErkJggg==", |
| 7063 | })]), |
| 7064 | }], |
| 7065 | }, |
| 7066 | ]; |
| 7067 | |
| 7068 | let body = build_chat_wire_body( |
| 7069 | &request, |
| 7070 | ProviderKind::Openai, |
| 7071 | "https://api.openai.com/v1", |
| 7072 | false, |
| 7073 | None, |
| 7074 | ) |
| 7075 | .expect("wire body"); |
| 7076 | let messages = body.body["messages"].as_array().expect("messages"); |
| 7077 | let roles: Vec<_> = messages |
| 7078 | .iter() |
| 7079 | .map(|message| message["role"].as_str().expect("role")) |
| 7080 | .collect(); |
| 7081 | assert_eq!(roles, ["assistant", "tool", "tool", "user"]); |
| 7082 | assert_eq!(messages[1]["tool_call_id"], "call_image_1"); |
| 7083 | assert_eq!(messages[2]["tool_call_id"], "call_image_2"); |
| 7084 | |
| 7085 | let image_parts = messages[3]["content"].as_array().expect("image parts"); |
| 7086 | assert_eq!(image_parts.len(), 4); |
| 7087 | assert_eq!( |
| 7088 | image_parts[1]["image_url"]["url"], |
| 7089 | "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==" |
| 7090 | ); |
| 7091 | assert_eq!( |
| 7092 | image_parts[3]["image_url"]["url"], |
| 7093 | "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGNgYPj/HwADAgH/5ncLrgAAAABJRU5ErkJggg==" |
| 7094 | ); |
| 7095 | } |
| 7096 | |
| 7097 | #[test] |
| 7098 | fn out_of_order_tool_results_remain_a_contiguous_complete_batch() { |
| 7099 | let messages = vec![ |
| 7100 | Message { |
| 7101 | role: Role::Assistant, |
| 7102 | content: vec![fixture_tool_use("call_one"), fixture_tool_use("call_two")], |
| 7103 | }, |
| 7104 | fixture_tool_result("call_two"), |
| 7105 | fixture_tool_result("call_one"), |
| 7106 | ]; |
| 7107 | |
| 7108 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 7109 | let roles: Vec<_> = wire |
| 7110 | .iter() |
| 7111 | .map(|message| message["role"].as_str().expect("role")) |
| 7112 | .collect(); |
| 7113 | assert_eq!(roles, ["assistant", "tool", "tool"]); |
| 7114 | assert_eq!(wire[1]["tool_call_id"], "call_two"); |
| 7115 | assert_eq!(wire[2]["tool_call_id"], "call_one"); |
| 7116 | } |
| 7117 | |
| 7118 | #[test] |
| 7119 | fn incomplete_tool_result_batch_is_downgraded_before_serialization() { |
| 7120 | let messages = vec![ |
| 7121 | Message { |
| 7122 | role: Role::Assistant, |
| 7123 | content: vec![fixture_tool_use("call_one"), fixture_tool_use("call_two")], |
| 7124 | }, |
| 7125 | fixture_tool_result("call_one"), |
| 7126 | ]; |
| 7127 | |
| 7128 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 7129 | assert!( |
| 7130 | !wire |
| 7131 | .iter() |
| 7132 | .any(|message| message.get("tool_calls").is_some()), |
| 7133 | "an incomplete tool batch must not reach the provider: {wire:?}" |
| 7134 | ); |
| 7135 | assert!( |
| 7136 | !wire |
| 7137 | .iter() |
| 7138 | .any(|message| message["role"].as_str() == Some("tool")), |
| 7139 | "the partial result must be removed with its incomplete call batch: {wire:?}" |
| 7140 | ); |
| 7141 | } |
| 7142 | |
| 7143 | #[test] |
| 7144 | fn interleaved_tool_result_batch_is_downgraded_before_serialization() { |
| 7145 | let messages = vec![ |
| 7146 | Message { |
| 7147 | role: Role::Assistant, |
| 7148 | content: vec![ |
| 7149 | ContentBlock::ToolUse { |
| 7150 | execution_id: None, |
| 7151 | id: "call_one".to_string(), |
| 7152 | name: "read".to_string(), |
| 7153 | input: serde_json::json!({"path": "first.png"}), |
| 7154 | caller: None, |
| 7155 | thought_signature: None, |
| 7156 | }, |
| 7157 | ContentBlock::ToolUse { |
| 7158 | execution_id: None, |
| 7159 | id: "call_two".to_string(), |
| 7160 | name: "read".to_string(), |
| 7161 | input: serde_json::json!({"path": "second.png"}), |
| 7162 | caller: None, |
| 7163 | thought_signature: None, |
| 7164 | }, |
| 7165 | ], |
| 7166 | }, |
| 7167 | Message { |
| 7168 | role: Role::User, |
| 7169 | content: vec![ContentBlock::ToolResult { |
| 7170 | execution_id: None, |
| 7171 | tool_use_id: "call_one".to_string(), |
| 7172 | content: "first screenshot captured".to_string(), |
| 7173 | is_error: Some(false), |
| 7174 | content_blocks: Some(vec![serde_json::json!({ |
| 7175 | "type": "image", |
| 7176 | "mime_type": "image/png", |
| 7177 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==", |
| 7178 | })]), |
| 7179 | }], |
| 7180 | }, |
| 7181 | Message { |
| 7182 | role: Role::User, |
| 7183 | content: vec![ |
| 7184 | ContentBlock::Text { |
| 7185 | text: "interloper".to_string(), |
| 7186 | cache_control: None, |
| 7187 | }, |
| 7188 | ContentBlock::ToolResult { |
| 7189 | execution_id: None, |
| 7190 | tool_use_id: "call_two".to_string(), |
| 7191 | content: "second screenshot captured".to_string(), |
| 7192 | is_error: Some(false), |
| 7193 | content_blocks: Some(vec![serde_json::json!({ |
| 7194 | "type": "image", |
| 7195 | "mime_type": "image/png", |
| 7196 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==", |
| 7197 | })]), |
| 7198 | }, |
| 7199 | ], |
| 7200 | }, |
| 7201 | ]; |
| 7202 | |
| 7203 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 7204 | assert!( |
| 7205 | !wire |
| 7206 | .iter() |
| 7207 | .any(|message| message.get("tool_calls").is_some()), |
| 7208 | "a non-contiguous tool batch must not reach the provider: {wire:?}" |
| 7209 | ); |
| 7210 | assert!( |
| 7211 | !wire |
| 7212 | .iter() |
| 7213 | .any(|message| message["role"].as_str() == Some("tool")), |
| 7214 | "orphaned tool replies must be removed with their stripped call batch: {wire:?}" |
| 7215 | ); |
| 7216 | assert!( |
| 7217 | !wire.iter().any(|message| { |
| 7218 | message["content"].as_array().is_some_and(|parts| { |
| 7219 | parts |
| 7220 | .iter() |
| 7221 | .any(|part| part["type"].as_str() == Some("image_url")) |
| 7222 | }) |
| 7223 | }), |
| 7224 | "images from a stripped tool batch must not survive as user input: {wire:?}" |
| 7225 | ); |
| 7226 | } |
| 7227 | |
| 7228 | #[test] |
| 7229 | fn duplicate_tool_call_ids_are_downgraded_before_serialization() { |
| 7230 | let messages = vec![ |
| 7231 | Message { |
| 7232 | role: Role::Assistant, |
| 7233 | content: vec![ |
| 7234 | ContentBlock::ToolUse { |
| 7235 | execution_id: None, |
| 7236 | id: "duplicate".to_string(), |
| 7237 | name: "read".to_string(), |
| 7238 | input: serde_json::json!({"path": "first.png"}), |
| 7239 | caller: None, |
| 7240 | thought_signature: None, |
| 7241 | }, |
| 7242 | ContentBlock::ToolUse { |
| 7243 | execution_id: None, |
| 7244 | id: "duplicate".to_string(), |
| 7245 | name: "read".to_string(), |
| 7246 | input: serde_json::json!({"path": "second.png"}), |
| 7247 | caller: None, |
| 7248 | thought_signature: None, |
| 7249 | }, |
| 7250 | ], |
| 7251 | }, |
| 7252 | Message { |
| 7253 | role: Role::User, |
| 7254 | content: vec![ContentBlock::ToolResult { |
| 7255 | execution_id: None, |
| 7256 | tool_use_id: "duplicate".to_string(), |
| 7257 | content: "one result for two calls".to_string(), |
| 7258 | is_error: Some(false), |
| 7259 | content_blocks: Some(vec![serde_json::json!({ |
| 7260 | "type": "image", |
| 7261 | "mime_type": "image/png", |
| 7262 | "data": "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR4nGP4z8DwHwAFAAH/iZk9HQAAAABJRU5ErkJggg==", |
| 7263 | })]), |
| 7264 | }], |
| 7265 | }, |
| 7266 | ]; |
| 7267 | |
| 7268 | let wire = build_chat_messages(None, &messages, "gpt-4o"); |
| 7269 | assert!( |
| 7270 | !wire |
| 7271 | .iter() |
| 7272 | .any(|message| message.get("tool_calls").is_some()), |
| 7273 | "duplicate call IDs cannot satisfy two tool calls: {wire:?}" |
| 7274 | ); |
| 7275 | assert!( |
| 7276 | !wire |
| 7277 | .iter() |
| 7278 | .any(|message| message["role"].as_str() == Some("tool")), |
| 7279 | "the ambiguous tool result must be removed with the stripped batch: {wire:?}" |
| 7280 | ); |
| 7281 | assert!( |
| 7282 | !wire.iter().any(|message| { |
| 7283 | message["content"].as_array().is_some_and(|parts| { |
| 7284 | parts |
| 7285 | .iter() |
| 7286 | .any(|part| part["type"].as_str() == Some("image_url")) |
| 7287 | }) |
| 7288 | }), |
| 7289 | "media from an ambiguous duplicate-ID batch must not survive: {wire:?}" |
| 7290 | ); |
| 7291 | } |
| 7292 | } |
| 7293 | |
| 7294 | #[cfg(test)] |
| 7295 | mod mistral_reasoning_tests { |
| 7296 | use super::*; |
| 7297 | |
| 7298 | fn request_with_assistant_thinking_and_tool() -> MessageRequest { |
| 7299 | MessageRequest { |
| 7300 | model: "mistral-medium-latest".to_string(), |
| 7301 | messages: vec![ |
| 7302 | Message { |
| 7303 | role: Role::Assistant, |
| 7304 | content: vec![ |
| 7305 | ContentBlock::Thinking { |
| 7306 | thinking: "Inspect the current state before calling the tool." |
| 7307 | .to_string(), |
| 7308 | signature: None, |
| 7309 | state: None, |
| 7310 | }, |
| 7311 | ContentBlock::Text { |
| 7312 | text: "I will inspect it now.".to_string(), |
| 7313 | cache_control: None, |
| 7314 | }, |
| 7315 | ContentBlock::ToolUse { |
| 7316 | execution_id: None, |
| 7317 | id: "call-1".to_string(), |
| 7318 | name: "read_file".to_string(), |
| 7319 | input: json!({"path": "README.md"}), |
| 7320 | caller: None, |
| 7321 | thought_signature: None, |
| 7322 | }, |
| 7323 | ], |
| 7324 | }, |
| 7325 | Message { |
| 7326 | role: Role::User, |
| 7327 | content: vec![ContentBlock::ToolResult { |
| 7328 | execution_id: None, |
| 7329 | tool_use_id: "call-1".to_string(), |
| 7330 | content: "contents".to_string(), |
| 7331 | is_error: None, |
| 7332 | content_blocks: None, |
| 7333 | }], |
| 7334 | }, |
| 7335 | ], |
| 7336 | max_tokens: 64, |
| 7337 | system: None, |
| 7338 | tools: None, |
| 7339 | tool_choice: None, |
| 7340 | metadata: None, |
| 7341 | thinking: None, |
| 7342 | reasoning_effort: Some("high".to_string()), |
| 7343 | stream: None, |
| 7344 | temperature: None, |
| 7345 | top_p: None, |
| 7346 | } |
| 7347 | } |
| 7348 | |
| 7349 | #[test] |
| 7350 | fn mistral_effort_wire_value_covers_codewhale_tiers() { |
| 7351 | assert_eq!(mistral_reasoning_effort_wire_value("off"), Some("none")); |
| 7352 | assert_eq!( |
| 7353 | mistral_reasoning_effort_wire_value("disabled"), |
| 7354 | Some("none") |
| 7355 | ); |
| 7356 | assert_eq!(mistral_reasoning_effort_wire_value("none"), Some("none")); |
| 7357 | assert_eq!(mistral_reasoning_effort_wire_value("false"), Some("none")); |
| 7358 | assert_eq!(mistral_reasoning_effort_wire_value("high"), Some("high")); |
| 7359 | assert_eq!(mistral_reasoning_effort_wire_value("xhigh"), Some("high")); |
| 7360 | assert_eq!(mistral_reasoning_effort_wire_value("max"), Some("high")); |
| 7361 | assert_eq!(mistral_reasoning_effort_wire_value("ultra"), Some("high")); |
| 7362 | assert_eq!( |
| 7363 | mistral_reasoning_effort_wire_value("ultracode"), |
| 7364 | Some("high") |
| 7365 | ); |
| 7366 | // Intermediate tiers must be omitted so the request falls back to |
| 7367 | // Mistral's own default rather than 400 code 3051 on unsupported |
| 7368 | // values like "low"/"medium" that the server does not accept today. |
| 7369 | assert_eq!(mistral_reasoning_effort_wire_value("low"), None); |
| 7370 | assert_eq!(mistral_reasoning_effort_wire_value("medium"), None); |
| 7371 | assert_eq!(mistral_reasoning_effort_wire_value("mid"), None); |
| 7372 | assert_eq!(mistral_reasoning_effort_wire_value("minimal"), None); |
| 7373 | } |
| 7374 | |
| 7375 | #[test] |
| 7376 | fn mistral_model_gate_only_matches_reasoning_capable_families() { |
| 7377 | assert!(mistral_model_supports_reasoning("mistral-medium-latest")); |
| 7378 | assert!(mistral_model_supports_reasoning("mistral-medium-3-5")); |
| 7379 | assert!(mistral_model_supports_reasoning("mistral-small-latest")); |
| 7380 | assert!(mistral_model_supports_reasoning("mistral-small-2603")); |
| 7381 | assert!(mistral_model_supports_reasoning("magistral-small-latest")); |
| 7382 | assert!(mistral_model_supports_reasoning("MISTRAL-MEDIUM-LATEST")); |
| 7383 | assert!(!mistral_model_supports_reasoning("mistral-code-latest")); |
| 7384 | assert!(!mistral_model_supports_reasoning("codestral-latest")); |
| 7385 | assert!(!mistral_model_supports_reasoning("mistral-large-latest")); |
| 7386 | assert!(!mistral_model_supports_reasoning("mistral-nemo-2407")); |
| 7387 | } |
| 7388 | |
| 7389 | #[test] |
| 7390 | fn mistral_route_shaper_writes_reasoning_only_for_supported_models() { |
| 7391 | let mut body = json!({"model": "mistral-medium-latest"}); |
| 7392 | apply_mistral_route_reasoning_controls( |
| 7393 | &mut body, |
| 7394 | ProviderKind::Mistral, |
| 7395 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7396 | "mistral-medium-latest", |
| 7397 | Some("high"), |
| 7398 | ); |
| 7399 | assert_eq!(body["reasoning_effort"], json!("high")); |
| 7400 | |
| 7401 | let mut body = json!({"model": "mistral-code-latest"}); |
| 7402 | apply_mistral_route_reasoning_controls( |
| 7403 | &mut body, |
| 7404 | ProviderKind::Mistral, |
| 7405 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7406 | "mistral-code-latest", |
| 7407 | Some("high"), |
| 7408 | ); |
| 7409 | assert!( |
| 7410 | body.get("reasoning_effort").is_none(), |
| 7411 | "non-reasoning models must never see reasoning_effort (Mistral 400s on 3051): {body}" |
| 7412 | ); |
| 7413 | |
| 7414 | let mut body = json!({"model": "mistral-medium-latest", "reasoning_effort": "stale"}); |
| 7415 | apply_mistral_route_reasoning_controls( |
| 7416 | &mut body, |
| 7417 | ProviderKind::Mistral, |
| 7418 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7419 | "mistral-medium-latest", |
| 7420 | Some("low"), |
| 7421 | ); |
| 7422 | assert!( |
| 7423 | body.get("reasoning_effort").is_none(), |
| 7424 | "intermediate tiers must be stripped rather than sent unsupported: {body}" |
| 7425 | ); |
| 7426 | |
| 7427 | // Non-Mistral providers must not be touched by this shaper. |
| 7428 | let mut body = json!({"model": "deepseek-v4-pro", "reasoning_effort": "high"}); |
| 7429 | apply_mistral_route_reasoning_controls( |
| 7430 | &mut body, |
| 7431 | ProviderKind::Deepseek, |
| 7432 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7433 | "deepseek-v4-pro", |
| 7434 | Some("high"), |
| 7435 | ); |
| 7436 | assert_eq!(body["reasoning_effort"], json!("high")); |
| 7437 | |
| 7438 | let mut body = json!({ |
| 7439 | "model": "mistral-medium-latest", |
| 7440 | "thinking": {"type": "enabled"}, |
| 7441 | "reasoning_effort": "stale", |
| 7442 | }); |
| 7443 | apply_mistral_route_reasoning_controls( |
| 7444 | &mut body, |
| 7445 | ProviderKind::Mistral, |
| 7446 | "https://gateway.example.test/v1", |
| 7447 | "mistral-medium-latest", |
| 7448 | Some("high"), |
| 7449 | ); |
| 7450 | assert!(body.get("thinking").is_none()); |
| 7451 | assert!(body.get("reasoning_effort").is_none()); |
| 7452 | |
| 7453 | let mut native = json!({"model": "magistral-small-latest"}); |
| 7454 | apply_mistral_route_reasoning_controls( |
| 7455 | &mut native, |
| 7456 | ProviderKind::Mistral, |
| 7457 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7458 | "magistral-small-latest", |
| 7459 | Some("off"), |
| 7460 | ); |
| 7461 | assert!( |
| 7462 | native.get("reasoning_effort").is_none(), |
| 7463 | "legacy native Magistral is always-reasoning and does not use the adjustable effort field" |
| 7464 | ); |
| 7465 | } |
| 7466 | |
| 7467 | #[test] |
| 7468 | fn mistral_wire_dialect_is_limited_to_exact_first_party_routes() { |
| 7469 | for official in [ |
| 7470 | "https://api.mistral.ai/v1", |
| 7471 | "https://api.eu.mistral.ai/v1/", |
| 7472 | "https://api.us.mistral.ai/v1", |
| 7473 | ] { |
| 7474 | assert!(is_exact_mistral_chat_route(ProviderKind::Mistral, official)); |
| 7475 | } |
| 7476 | for neighbor in [ |
| 7477 | "http://api.mistral.ai/v1", |
| 7478 | "https://api.mistral.ai/v2", |
| 7479 | "https://proxy.example.test/v1", |
| 7480 | "https://api.mistral.ai.evil.test/v1", |
| 7481 | ] { |
| 7482 | assert!(!is_exact_mistral_chat_route( |
| 7483 | ProviderKind::Mistral, |
| 7484 | neighbor |
| 7485 | )); |
| 7486 | } |
| 7487 | assert!(!is_exact_mistral_chat_route( |
| 7488 | ProviderKind::Openai, |
| 7489 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7490 | )); |
| 7491 | assert_eq!( |
| 7492 | reasoning_stream_style_for_route( |
| 7493 | ProviderKind::Mistral, |
| 7494 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7495 | "mistral-medium-latest", |
| 7496 | None, |
| 7497 | ), |
| 7498 | ReasoningStreamStyle::MistralBlocks |
| 7499 | ); |
| 7500 | assert_eq!( |
| 7501 | reasoning_stream_style_for_route( |
| 7502 | ProviderKind::Mistral, |
| 7503 | "https://gateway.example.test/v1", |
| 7504 | "mistral-medium-latest", |
| 7505 | None, |
| 7506 | ), |
| 7507 | ReasoningStreamStyle::SeparateField |
| 7508 | ); |
| 7509 | } |
| 7510 | |
| 7511 | #[test] |
| 7512 | fn extract_mistral_polymorphic_content_flattens_thinking_and_text() { |
| 7513 | // Non-reasoning response: plain string content. Extractor returns |
| 7514 | // (None, None) so the shared string fallback still runs. |
| 7515 | let plain = json!({"content": "Hello world"}); |
| 7516 | assert_eq!(extract_mistral_polymorphic_content(&plain), (None, None)); |
| 7517 | |
| 7518 | // Missing content: no panic, returns (None, None). |
| 7519 | let empty = json!({}); |
| 7520 | assert_eq!(extract_mistral_polymorphic_content(&empty), (None, None)); |
| 7521 | |
| 7522 | // Reasoning response: nested thinking array + text block. |
| 7523 | let reasoning = json!({"content": [ |
| 7524 | {"type": "thinking", "thinking": [ |
| 7525 | {"type": "text", "text": "First "}, |
| 7526 | {"type": "text", "text": "second."}, |
| 7527 | ], "closed": true}, |
| 7528 | {"type": "text", "text": "Final answer."}, |
| 7529 | ]}); |
| 7530 | let (thinking, text) = extract_mistral_polymorphic_content(&reasoning); |
| 7531 | assert_eq!(thinking.as_deref(), Some("First second.")); |
| 7532 | assert_eq!(text.as_deref(), Some("Final answer.")); |
| 7533 | |
| 7534 | // Thinking-only chunk (mid-stream) with no closing text yet. |
| 7535 | let thinking_only = json!({"content": [ |
| 7536 | {"type": "thinking", "thinking": [{"type": "text", "text": "still thinking"}]}, |
| 7537 | ]}); |
| 7538 | let (thinking, text) = extract_mistral_polymorphic_content(&thinking_only); |
| 7539 | assert_eq!(thinking.as_deref(), Some("still thinking")); |
| 7540 | assert_eq!(text, None); |
| 7541 | } |
| 7542 | |
| 7543 | #[test] |
| 7544 | fn reshape_mistral_messages_reconstructs_polymorphic_shape_for_assistant_replay() { |
| 7545 | // Assistant message with stored reasoning_content is reshaped into |
| 7546 | // Mistral's polymorphic content-as-array shape. |
| 7547 | let mut messages = vec![ |
| 7548 | json!({"role": "user", "content": "compute 3+4"}), |
| 7549 | json!({ |
| 7550 | "role": "assistant", |
| 7551 | "content": "The answer is 7.", |
| 7552 | "reasoning_content": "Let me add 3 and 4 to get 7.", |
| 7553 | }), |
| 7554 | json!({"role": "user", "content": "now multiply by 2"}), |
| 7555 | ]; |
| 7556 | reshape_mistral_messages_for_reasoning_replay(&mut messages); |
| 7557 | |
| 7558 | assert_eq!(messages[0]["role"], "user"); |
| 7559 | assert!( |
| 7560 | messages[0]["content"].is_string(), |
| 7561 | "user turns are left untouched: {}", |
| 7562 | messages[0] |
| 7563 | ); |
| 7564 | |
| 7565 | assert_eq!(messages[1]["role"], "assistant"); |
| 7566 | assert!( |
| 7567 | messages[1].get("reasoning_content").is_none(), |
| 7568 | "reasoning_content field must be removed after reshape: {}", |
| 7569 | messages[1] |
| 7570 | ); |
| 7571 | let content = messages[1]["content"] |
| 7572 | .as_array() |
| 7573 | .expect("assistant content is now an array"); |
| 7574 | assert_eq!(content.len(), 2); |
| 7575 | assert_eq!(content[0]["type"], "thinking"); |
| 7576 | assert_eq!(content[0]["closed"], true); |
| 7577 | assert_eq!(content[0]["thinking"][0]["type"], "text"); |
| 7578 | assert_eq!( |
| 7579 | content[0]["thinking"][0]["text"], |
| 7580 | "Let me add 3 and 4 to get 7." |
| 7581 | ); |
| 7582 | assert_eq!(content[1]["type"], "text"); |
| 7583 | assert_eq!(content[1]["text"], "The answer is 7."); |
| 7584 | |
| 7585 | // Assistant with no reasoning stays as-is (plain string content). |
| 7586 | let mut plain = vec![json!({"role": "assistant", "content": "hi"})]; |
| 7587 | reshape_mistral_messages_for_reasoning_replay(&mut plain); |
| 7588 | assert_eq!(plain[0]["content"], "hi"); |
| 7589 | |
| 7590 | // Empty reasoning is treated as absent — no reshape. |
| 7591 | let mut empty = vec![json!({ |
| 7592 | "role": "assistant", |
| 7593 | "content": "hi", |
| 7594 | "reasoning_content": " ", |
| 7595 | })]; |
| 7596 | reshape_mistral_messages_for_reasoning_replay(&mut empty); |
| 7597 | assert_eq!(empty[0]["content"], "hi"); |
| 7598 | } |
| 7599 | |
| 7600 | #[test] |
| 7601 | fn mistral_prompt_builder_replays_stored_thinking_as_polymorphic_content() { |
| 7602 | let request = request_with_assistant_thinking_and_tool(); |
| 7603 | let exact = build_chat_messages_for_request_and_provider_and_route( |
| 7604 | &request, |
| 7605 | ProviderKind::Mistral, |
| 7606 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7607 | ); |
| 7608 | let assistant = &exact[0]; |
| 7609 | assert!(assistant.get("reasoning_content").is_none()); |
| 7610 | assert!(assistant.get("tool_calls").is_some()); |
| 7611 | let content = assistant["content"] |
| 7612 | .as_array() |
| 7613 | .expect("exact Mistral history uses polymorphic content"); |
| 7614 | assert_eq!(content[0]["type"], "thinking"); |
| 7615 | assert_eq!( |
| 7616 | content[0]["thinking"][0]["text"], |
| 7617 | "Inspect the current state before calling the tool." |
| 7618 | ); |
| 7619 | assert_eq!(content[1]["type"], "text"); |
| 7620 | assert_eq!(content[1]["text"], "I will inspect it now."); |
| 7621 | |
| 7622 | for (provider, base_url) in [ |
| 7623 | (ProviderKind::Mistral, "https://gateway.example.test/v1"), |
| 7624 | ( |
| 7625 | ProviderKind::Openai, |
| 7626 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7627 | ), |
| 7628 | ] { |
| 7629 | let neighbor = build_chat_messages_for_request_and_provider_and_route( |
| 7630 | &request, provider, base_url, |
| 7631 | ); |
| 7632 | assert!(neighbor[0].get("reasoning_content").is_none()); |
| 7633 | assert!( |
| 7634 | neighbor[0]["content"].is_string(), |
| 7635 | "unproven routes must not inherit Mistral's polymorphic dialect: {}", |
| 7636 | neighbor[0] |
| 7637 | ); |
| 7638 | } |
| 7639 | } |
| 7640 | |
| 7641 | #[test] |
| 7642 | fn mistral_stream_tool_call_replay_does_not_gain_reasoning_content() { |
| 7643 | let request = request_with_assistant_thinking_and_tool(); |
| 7644 | let wire = build_chat_wire_body( |
| 7645 | &request, |
| 7646 | ProviderKind::Mistral, |
| 7647 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7648 | true, |
| 7649 | None, |
| 7650 | ) |
| 7651 | .expect("Mistral stream wire body"); |
| 7652 | let assistant = &wire.body["messages"][0]; |
| 7653 | assert!(assistant["content"].is_array()); |
| 7654 | assert!(assistant.get("reasoning_content").is_none()); |
| 7655 | assert!(assistant.get("tool_calls").is_some()); |
| 7656 | assert_eq!(wire.body["reasoning_effort"], "high"); |
| 7657 | assert_eq!(wire.replay_input_tokens, None); |
| 7658 | } |
| 7659 | |
| 7660 | #[test] |
| 7661 | fn mistral_nonstream_parser_is_route_isolated() { |
| 7662 | let payload = json!({ |
| 7663 | "id": "chatcmpl-mistral", |
| 7664 | "model": "mistral-medium-latest", |
| 7665 | "choices": [{ |
| 7666 | "finish_reason": "stop", |
| 7667 | "message": {"role": "assistant", "content": [ |
| 7668 | {"type": "thinking", "thinking": [ |
| 7669 | {"type": "text", "text": "private trace"} |
| 7670 | ], "closed": true}, |
| 7671 | {"type": "text", "text": "public answer"} |
| 7672 | ]} |
| 7673 | }], |
| 7674 | "usage": {"prompt_tokens": 5, "completion_tokens": 2} |
| 7675 | }); |
| 7676 | let mistral = parse_chat_message_for_route( |
| 7677 | &payload, |
| 7678 | ProviderKind::Mistral, |
| 7679 | crate::config::DEFAULT_MISTRAL_BASE_URL, |
| 7680 | ) |
| 7681 | .expect("Mistral payload parses"); |
| 7682 | assert!(matches!( |
| 7683 | &mistral.content[0], |
| 7684 | ContentBlock::Thinking { thinking, .. } if thinking == "private trace" |
| 7685 | )); |
| 7686 | assert!(matches!( |
| 7687 | &mistral.content[1], |
| 7688 | ContentBlock::Text { text, .. } if text == "public answer" |
| 7689 | )); |
| 7690 | |
| 7691 | let generic = parse_chat_message(&payload).expect("generic payload parses"); |
| 7692 | assert!( |
| 7693 | !generic |
| 7694 | .content |
| 7695 | .iter() |
| 7696 | .any(|block| matches!(block, ContentBlock::Thinking { .. })), |
| 7697 | "typed arrays from another provider must not be reinterpreted as Mistral thinking" |
| 7698 | ); |
| 7699 | } |
| 7700 | |
| 7701 | #[test] |
| 7702 | fn mistral_shared_capability_matches_wire_contract() { |
| 7703 | for model in [ |
| 7704 | "mistral-medium-latest", |
| 7705 | "mistral-small-latest", |
| 7706 | "magistral-small-latest", |
| 7707 | ] { |
| 7708 | assert!(codewhale_models::model_supports_reasoning(model), "{model}"); |
| 7709 | } |
| 7710 | for model in ["mistral-code-latest", "mistral-large-latest"] { |
| 7711 | assert!( |
| 7712 | !codewhale_models::model_supports_reasoning(model), |
| 7713 | "{model}" |
| 7714 | ); |
| 7715 | } |
| 7716 | } |
| 7717 | } |
| 7718 | |
| 7719 | #[cfg(test)] |
| 7720 | mod google_thought_signature_tests { |
| 7721 | use super::*; |
| 7722 | |
| 7723 | // ── Google thought signatures (#v0.9.8 Google backend) ────────────── |
| 7724 | use crate::config::{DEFAULT_GOOGLE_BASE_URL, DEFAULT_OPENAI_BASE_URL}; |
| 7725 | |
| 7726 | /// One signed tool turn. `paired` false is the restart shape: the process |
| 7727 | /// died between the tool call and its result, so the terminal |
| 7728 | /// `tool_result` never reached durable history. |
| 7729 | fn signed_history(signature: Option<&str>, paired: bool) -> Vec<Message> { |
| 7730 | let mut messages = vec![ |
| 7731 | Message { |
| 7732 | role: Role::User, |
| 7733 | content: vec![ContentBlock::Text { |
| 7734 | text: "Read the config.".to_string(), |
| 7735 | cache_control: None, |
| 7736 | }], |
| 7737 | }, |
| 7738 | Message { |
| 7739 | role: Role::Assistant, |
| 7740 | content: vec![ |
| 7741 | ContentBlock::Text { |
| 7742 | text: "Reading now.".to_string(), |
| 7743 | cache_control: None, |
| 7744 | }, |
| 7745 | ContentBlock::ToolUse { |
| 7746 | execution_id: None, |
| 7747 | id: "call-g-1".to_string(), |
| 7748 | name: "read".to_string(), |
| 7749 | input: json!({"path": "config.toml"}), |
| 7750 | caller: None, |
| 7751 | thought_signature: signature.map(str::to_string), |
| 7752 | }, |
| 7753 | ], |
| 7754 | }, |
| 7755 | ]; |
| 7756 | if paired { |
| 7757 | messages.push(Message { |
| 7758 | role: Role::User, |
| 7759 | content: vec![ContentBlock::ToolResult { |
| 7760 | execution_id: None, |
| 7761 | tool_use_id: "call-g-1".to_string(), |
| 7762 | content: "key = \"value\"".to_string(), |
| 7763 | is_error: None, |
| 7764 | content_blocks: None, |
| 7765 | }], |
| 7766 | }); |
| 7767 | } |
| 7768 | messages |
| 7769 | } |
| 7770 | |
| 7771 | fn request_from(messages: Vec<Message>) -> MessageRequest { |
| 7772 | MessageRequest { |
| 7773 | model: "gemini-3.1-pro-preview".to_string(), |
| 7774 | messages, |
| 7775 | max_tokens: 64, |
| 7776 | system: None, |
| 7777 | tools: None, |
| 7778 | tool_choice: None, |
| 7779 | metadata: None, |
| 7780 | thinking: None, |
| 7781 | reasoning_effort: Some("high".to_string()), |
| 7782 | stream: None, |
| 7783 | temperature: None, |
| 7784 | top_p: None, |
| 7785 | } |
| 7786 | } |
| 7787 | |
| 7788 | fn google_request_with_signed_tool(signature: Option<&str>) -> MessageRequest { |
| 7789 | request_from(signed_history(signature, true)) |
| 7790 | } |
| 7791 | |
| 7792 | #[tokio::test] |
| 7793 | async fn gateway_thought_signature_rejection_explains_recovery_after_transport() { |
| 7794 | use crate::llm_client::LlmClient; |
| 7795 | use wiremock::matchers::{method, path}; |
| 7796 | use wiremock::{Mock, MockServer, ResponseTemplate}; |
| 7797 | |
| 7798 | let _ = rustls::crypto::ring::default_provider().install_default(); |
| 7799 | // An unsigned replay must reach the gateway: it may manage Google's |
| 7800 | // signatures itself. Only an actual rejection warrants recovery advice. |
| 7801 | for streaming in [false, true] { |
| 7802 | for status in [200, 400] { |
| 7803 | let server = MockServer::start().await; |
| 7804 | let response = if status == 400 { |
| 7805 | ResponseTemplate::new(status).set_body_json(json!({ |
| 7806 | "error": { |
| 7807 | "code": 400, |
| 7808 | "message": "Function call is missing a thought_signature in functionCall parts." |
| 7809 | } |
| 7810 | })) |
| 7811 | } else if streaming { |
| 7812 | ResponseTemplate::new(status) |
| 7813 | .insert_header("content-type", "text/event-stream") |
| 7814 | .set_body_string("data: [DONE]\n\n") |
| 7815 | } else { |
| 7816 | ResponseTemplate::new(status).set_body_json(json!({ |
| 7817 | "id": "gateway-replay", |
| 7818 | "model": "gemini-3.1-pro-preview", |
| 7819 | "choices": [{ |
| 7820 | "message": {"role": "assistant", "content": "Done."}, |
| 7821 | "finish_reason": "stop" |
| 7822 | }] |
| 7823 | })) |
| 7824 | }; |
| 7825 | Mock::given(method("POST")) |
| 7826 | .and(path("/v1/chat/completions")) |
| 7827 | .respond_with(response) |
| 7828 | .expect(1) |
| 7829 | .mount(&server) |
| 7830 | .await; |
| 7831 | |
| 7832 | let mut client = CodewhaleClient::new(&crate::config::Config { |
| 7833 | provider: Some("openai".to_string()), |
| 7834 | providers: Some(crate::config::ProvidersConfig { |
| 7835 | openai: crate::config::ProviderConfig { |
| 7836 | api_key: Some("gateway-test-key".to_string()), |
| 7837 | base_url: Some(format!("{}/v1", server.uri())), |
| 7838 | model: Some("gemini-3.1-pro-preview".to_string()), |
| 7839 | ..crate::config::ProviderConfig::default() |
| 7840 | }, |
| 7841 | ..crate::config::ProvidersConfig::default() |
| 7842 | }), |
| 7843 | ..crate::config::Config::default() |
| 7844 | }) |
| 7845 | .expect("gateway client"); |
| 7846 | client.isolated_request_state = true; |
| 7847 | let request = google_request_with_signed_tool(None); |
| 7848 | let result = if streaming { |
| 7849 | client.create_message_stream(request).await.map(|_| ()) |
| 7850 | } else { |
| 7851 | client |
| 7852 | .create_message_without_response_cache(request) |
| 7853 | .await |
| 7854 | .map(|_| ()) |
| 7855 | }; |
| 7856 | if status == 400 { |
| 7857 | let error = result.expect_err("gateway rejects unsigned replay"); |
| 7858 | let message = error.to_string(); |
| 7859 | assert!(message.contains("built-in `google` provider"), "{message}"); |
| 7860 | assert!(message.contains("start a new session"), "{message}"); |
| 7861 | assert!(matches!( |
| 7862 | error.downcast_ref::<crate::llm_client::LlmError>(), |
| 7863 | Some(crate::llm_client::LlmError::InvalidRequest { status: 400, .. }) |
| 7864 | )); |
| 7865 | } else { |
| 7866 | result.expect("gateway-managed signatures must still work"); |
| 7867 | } |
| 7868 | server.verify().await; |
| 7869 | } |
| 7870 | } |
| 7871 | } |
| 7872 | |
| 7873 | #[test] |
| 7874 | fn google_route_round_trips_thought_signatures_on_replayed_tool_calls() { |
| 7875 | let request = google_request_with_signed_tool(Some("SIG-abc123")); |
| 7876 | let messages = build_chat_messages_for_request_and_provider_and_route( |
| 7877 | &request, |
| 7878 | ProviderKind::Google, |
| 7879 | DEFAULT_GOOGLE_BASE_URL, |
| 7880 | ); |
| 7881 | let assistant = messages |
| 7882 | .iter() |
| 7883 | .find(|m| m.get("role") == Some(&json!("assistant"))) |
| 7884 | .expect("assistant replay message"); |
| 7885 | let signature = assistant |
| 7886 | .pointer("/tool_calls/0/extra_content/google/thought_signature") |
| 7887 | .and_then(serde_json::Value::as_str); |
| 7888 | assert_eq!(signature, Some("SIG-abc123")); |
| 7889 | } |
| 7890 | |
| 7891 | #[test] |
| 7892 | fn google_route_fails_closed_when_replayed_signature_is_missing() { |
| 7893 | let request = google_request_with_signed_tool(None); |
| 7894 | let error = build_chat_wire_body( |
| 7895 | &request, |
| 7896 | ProviderKind::Google, |
| 7897 | DEFAULT_GOOGLE_BASE_URL, |
| 7898 | false, |
| 7899 | None, |
| 7900 | ) |
| 7901 | .err() |
| 7902 | .expect("missing signature must fail closed before transport"); |
| 7903 | assert!( |
| 7904 | error.to_string().contains("thought signature"), |
| 7905 | "error must name the missing signature: {error}" |
| 7906 | ); |
| 7907 | } |
| 7908 | |
| 7909 | /// Google names the same model `gemini-3-pro` and `models/gemini-3-pro` on |
| 7910 | /// this endpoint. The prefixed spelling used to match none of the thinking |
| 7911 | /// families, so the model that most needs a signature was treated as one |
| 7912 | /// that needs none and the replay reached Google unsigned (#6018). |
| 7913 | #[test] |
| 7914 | fn google_route_fails_closed_for_a_models_prefixed_thinking_id() { |
| 7915 | let mut request = google_request_with_signed_tool(None); |
| 7916 | request.model = "models/gemini-3-pro-preview".to_string(); |
| 7917 | let error = build_chat_wire_body( |
| 7918 | &request, |
| 7919 | ProviderKind::Google, |
| 7920 | DEFAULT_GOOGLE_BASE_URL, |
| 7921 | false, |
| 7922 | None, |
| 7923 | ) |
| 7924 | .err() |
| 7925 | .expect("a models/-prefixed thinking id must fail closed like its bare spelling"); |
| 7926 | assert!( |
| 7927 | error.to_string().contains("thought signature"), |
| 7928 | "error must name the missing signature: {error}" |
| 7929 | ); |
| 7930 | } |
| 7931 | |
| 7932 | #[test] |
| 7933 | fn google_missing_signature_is_a_warning_not_an_error_for_flash_lite() { |
| 7934 | // 2.5 Flash-Lite ships thinking off; Google may legitimately omit |
| 7935 | // signatures there, so replay proceeds. |
| 7936 | let mut request = google_request_with_signed_tool(None); |
| 7937 | request.model = "gemini-2.5-flash-lite".to_string(); |
| 7938 | build_chat_wire_body( |
| 7939 | &request, |
| 7940 | ProviderKind::Google, |
| 7941 | DEFAULT_GOOGLE_BASE_URL, |
| 7942 | false, |
| 7943 | None, |
| 7944 | ) |
| 7945 | .expect("flash-lite replay must not require a signature"); |
| 7946 | } |
| 7947 | |
| 7948 | #[test] |
| 7949 | fn non_google_routes_never_see_google_extra_content() { |
| 7950 | let request = google_request_with_signed_tool(Some("SIG-abc123")); |
| 7951 | let messages = build_chat_messages_for_request_and_provider_and_route( |
| 7952 | &request, |
| 7953 | ProviderKind::Openai, |
| 7954 | DEFAULT_OPENAI_BASE_URL, |
| 7955 | ); |
| 7956 | for message in &messages { |
| 7957 | if let Some(tool_calls) = message.get("tool_calls").and_then(|v| v.as_array()) { |
| 7958 | for call in tool_calls { |
| 7959 | assert!( |
| 7960 | call.get("extra_content").is_none(), |
| 7961 | "Google-only fields must not leak to other providers" |
| 7962 | ); |
| 7963 | } |
| 7964 | } |
| 7965 | } |
| 7966 | } |
| 7967 | |
| 7968 | #[test] |
| 7969 | fn google_neighbor_base_url_does_not_get_google_dialect() { |
| 7970 | // A Google provider row pointed at some other gateway must not |
| 7971 | // carry signatures or fail closed: the dialect binds to the exact |
| 7972 | // official route, not to provider identity alone. |
| 7973 | let request = google_request_with_signed_tool(None); |
| 7974 | build_chat_wire_body( |
| 7975 | &request, |
| 7976 | ProviderKind::Google, |
| 7977 | "https://gateway.example.com/v1", |
| 7978 | false, |
| 7979 | None, |
| 7980 | ) |
| 7981 | .expect("non-official Google base URL must not require signatures"); |
| 7982 | let messages = build_chat_messages_for_request_and_provider_and_route( |
| 7983 | &google_request_with_signed_tool(Some("SIG")), |
| 7984 | ProviderKind::Google, |
| 7985 | "https://gateway.example.com/v1", |
| 7986 | ); |
| 7987 | assert!( |
| 7988 | messages |
| 7989 | .iter() |
| 7990 | .all(|m| m.pointer("/tool_calls/0/extra_content").is_none()), |
| 7991 | "signatures must not be sent to a non-Google endpoint" |
| 7992 | ); |
| 7993 | } |
| 7994 | |
| 7995 | /// The manually configured OpenAI-compatible row (#1519, |
| 7996 | /// `ProviderKind::Custom`) pointed at Google's OpenAI-compat endpoint is |
| 7997 | /// byte-for-byte the same endpoint as the built-in `google` row. C22: it |
| 7998 | /// used to fail the `provider == Google` half of the route gate, so its |
| 7999 | /// signatures were stripped on replay with no warning and later signed |
| 8000 | /// tool turns failed. The gate now binds to the endpoint. |
| 8001 | #[test] |
| 8002 | fn manually_configured_openai_compatible_google_endpoint_replays_signatures() { |
| 8003 | let request = google_request_with_signed_tool(Some("SIG-abc123")); |
| 8004 | for base_url in [ |
| 8005 | DEFAULT_GOOGLE_BASE_URL, |
| 8006 | "https://generativelanguage.googleapis.com/v1beta/openai", |
| 8007 | "https://GenerativeLanguage.googleapis.com/v1beta/openai/", |
| 8008 | ] { |
| 8009 | let messages = build_chat_messages_for_request_and_provider_and_route( |
| 8010 | &request, |
| 8011 | ProviderKind::Custom, |
| 8012 | base_url, |
| 8013 | ); |
| 8014 | let assistant = messages |
| 8015 | .iter() |
| 8016 | .find(|m| m.get("role") == Some(&json!("assistant"))) |
| 8017 | .expect("assistant replay message"); |
| 8018 | assert_eq!( |
| 8019 | assistant |
| 8020 | .pointer("/tool_calls/0/extra_content/google/thought_signature") |
| 8021 | .and_then(serde_json::Value::as_str), |
| 8022 | Some("SIG-abc123"), |
| 8023 | "custom row at Google's endpoint must replay the signature ({base_url})" |
| 8024 | ); |
| 8025 | } |
| 8026 | } |
| 8027 | |
| 8028 | /// Signature preservation is a property of the endpoint, never of the |
| 8029 | /// reasoning setting: an operator who turns reasoning off (or whose |
| 8030 | /// effort is simply absent) still has to replay signed tool history. |
| 8031 | #[test] |
| 8032 | fn signatures_survive_replay_regardless_of_the_reasoning_setting() { |
| 8033 | for effort in [None, Some("off"), Some("low"), Some("high")] { |
| 8034 | for (provider, base_url) in [ |
| 8035 | (ProviderKind::Google, DEFAULT_GOOGLE_BASE_URL), |
| 8036 | ( |
| 8037 | ProviderKind::Custom, |
| 8038 | "https://generativelanguage.googleapis.com/v1beta/openai", |
| 8039 | ), |
| 8040 | ] { |
| 8041 | let mut request = google_request_with_signed_tool(Some("SIG-abc123")); |
| 8042 | request.reasoning_effort = effort.map(str::to_string); |
| 8043 | let body = build_chat_wire_body(&request, provider, base_url, true, None) |
| 8044 | .expect("signed replay builds on a signature-bearing route"); |
| 8045 | let assistant = body.body["messages"] |
| 8046 | .as_array() |
| 8047 | .expect("messages") |
| 8048 | .iter() |
| 8049 | .find(|m| m.get("role") == Some(&json!("assistant"))) |
| 8050 | .expect("assistant replay message"); |
| 8051 | assert_eq!( |
| 8052 | assistant |
| 8053 | .pointer("/tool_calls/0/extra_content/google/thought_signature") |
| 8054 | .and_then(serde_json::Value::as_str), |
| 8055 | Some("SIG-abc123"), |
| 8056 | "reasoning={effort:?} must not govern signature replay ({base_url})" |
| 8057 | ); |
| 8058 | } |
| 8059 | } |
| 8060 | } |
| 8061 | |
| 8062 | /// Fail closed on the manually configured row too — the missing-signature |
| 8063 | /// error is the useful feedback that replaces a silent strip. |
| 8064 | #[test] |
| 8065 | fn custom_row_at_google_endpoint_fails_closed_without_a_signature() { |
| 8066 | let request = google_request_with_signed_tool(None); |
| 8067 | let error = build_chat_wire_body( |
| 8068 | &request, |
| 8069 | ProviderKind::Custom, |
| 8070 | "https://generativelanguage.googleapis.com/v1beta/openai", |
| 8071 | false, |
| 8072 | None, |
| 8073 | ) |
| 8074 | .err() |
| 8075 | .expect("missing signature must fail closed before transport"); |
| 8076 | let rendered = error.to_string(); |
| 8077 | assert!( |
| 8078 | rendered.contains("thought signature") && rendered.contains("call-g-1"), |
| 8079 | "error must name the missing signature and the tool call: {rendered}" |
| 8080 | ); |
| 8081 | } |
| 8082 | |
| 8083 | /// A custom row pointed somewhere else is still a foreign gateway: the |
| 8084 | /// signature is stripped, and the strip reports how much it removed so |
| 8085 | /// the drop is never silent. |
| 8086 | #[test] |
| 8087 | fn custom_row_off_google_endpoint_strips_and_reports_signatures() { |
| 8088 | let request = google_request_with_signed_tool(Some("SIG-abc123")); |
| 8089 | let mut messages = build_chat_messages_for_request_and_provider_and_route( |
| 8090 | &request, |
| 8091 | ProviderKind::Google, |
| 8092 | DEFAULT_GOOGLE_BASE_URL, |
| 8093 | ); |
| 8094 | assert_eq!( |
| 8095 | strip_google_tool_call_extra_content(&mut messages), |
| 8096 | 1, |
| 8097 | "the strip must report the signatures it dropped" |
| 8098 | ); |
| 8099 | assert!( |
| 8100 | messages |
| 8101 | .iter() |
| 8102 | .all(|m| m.pointer("/tool_calls/0/extra_content").is_none()), |
| 8103 | "stripped history must carry no Google-only fields" |
| 8104 | ); |
| 8105 | let via_route = build_chat_messages_for_request_and_provider_and_route( |
| 8106 | &request, |
| 8107 | ProviderKind::Custom, |
| 8108 | "https://gateway.example.com/v1", |
| 8109 | ); |
| 8110 | assert!( |
| 8111 | via_route |
| 8112 | .iter() |
| 8113 | .all(|m| m.pointer("/tool_calls/0/extra_content").is_none()), |
| 8114 | "signatures must not reach a non-Google endpoint" |
| 8115 | ); |
| 8116 | } |
| 8117 | |
| 8118 | #[test] |
| 8119 | fn google_reasoning_uses_compatible_effort_without_rejected_native_fields() { |
| 8120 | // Wire examples and model limits from Google's compatibility docs. |
| 8121 | // Cover the actual request builder, both transport modes and both |
| 8122 | // ways a fresh install can configure the official endpoint (#6018). |
| 8123 | for (model, effort, expected) in [ |
| 8124 | ("gemini-3.1-pro-preview", Some("low"), Some("low")), |
| 8125 | ("gemini-3.1-pro-preview", Some("medium"), Some("medium")), |
| 8126 | ("gemini-3.1-pro-preview", Some("high"), Some("high")), |
| 8127 | ("gemini-3.1-pro-preview", Some("max"), Some("high")), |
| 8128 | ("gemini-3.5-flash-lite", Some("off"), Some("minimal")), |
| 8129 | ("gemini-2.5-flash", Some("off"), Some("none")), |
| 8130 | ("models/gemini-2.5-flash-lite", Some("off"), Some("none")), |
| 8131 | ("gemini-2.5-pro", Some("off"), Some("minimal")), |
| 8132 | ("gemini-3.1-pro-preview", None, None), |
| 8133 | ] { |
| 8134 | for provider in [ProviderKind::Google, ProviderKind::Custom] { |
| 8135 | for streaming in [false, true] { |
| 8136 | let mut request = google_request_with_signed_tool(Some("SIG")); |
| 8137 | request.model = model.to_string(); |
| 8138 | request.reasoning_effort = effort.map(str::to_string); |
| 8139 | let wire = build_chat_wire_body( |
| 8140 | &request, |
| 8141 | provider, |
| 8142 | DEFAULT_GOOGLE_BASE_URL, |
| 8143 | streaming, |
| 8144 | None, |
| 8145 | ) |
| 8146 | .expect("valid signed Google request"); |
| 8147 | assert_eq!( |
| 8148 | wire.body.get("reasoning_effort").and_then(Value::as_str), |
| 8149 | expected, |
| 8150 | "{model}: {effort:?}, {provider:?}, streaming={streaming}" |
| 8151 | ); |
| 8152 | assert!(wire.body.get("google").is_none()); |
| 8153 | assert!(wire.body.get("extra_body").is_none()); |
| 8154 | assert!(wire.body.get("thinking").is_none()); |
| 8155 | } |
| 8156 | } |
| 8157 | } |
| 8158 | } |
| 8159 | |
| 8160 | #[test] |
| 8161 | fn google_reasoning_control_does_not_rewrite_other_endpoints() { |
| 8162 | for provider in [ProviderKind::Google, ProviderKind::Custom] { |
| 8163 | let request = google_request_with_signed_tool(Some("SIG")); |
| 8164 | let wire = build_chat_wire_body( |
| 8165 | &request, |
| 8166 | provider, |
| 8167 | "https://gateway.example.com/v1", |
| 8168 | false, |
| 8169 | None, |
| 8170 | ) |
| 8171 | .expect("valid gateway request"); |
| 8172 | assert!(wire.body.get("reasoning_effort").is_none()); |
| 8173 | assert!(wire.body.get("google").is_none()); |
| 8174 | assert!(wire.body.get("extra_body").is_none()); |
| 8175 | } |
| 8176 | } |
| 8177 | |
| 8178 | #[test] |
| 8179 | fn google_signature_captured_from_non_streaming_tool_call() { |
| 8180 | let payload = json!({ |
| 8181 | "id": "resp-1", |
| 8182 | "model": "gemini-3.1-pro-preview", |
| 8183 | "choices": [{ |
| 8184 | "index": 0, |
| 8185 | "finish_reason": "tool_calls", |
| 8186 | "message": { |
| 8187 | "role": "assistant", |
| 8188 | "content": null, |
| 8189 | "tool_calls": [{ |
| 8190 | "id": "call-g-9", |
| 8191 | "type": "function", |
| 8192 | "function": { |
| 8193 | "name": "read", |
| 8194 | "arguments": "{\"path\":\"x\"}" |
| 8195 | }, |
| 8196 | "extra_content": { |
| 8197 | "google": { "thought_signature": "SIG-stream" } |
| 8198 | } |
| 8199 | }] |
| 8200 | } |
| 8201 | }], |
| 8202 | "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2} |
| 8203 | }); |
| 8204 | let response = parse_chat_message(&payload).expect("parses"); |
| 8205 | let signature = response.content.iter().find_map(|block| match block { |
| 8206 | ContentBlock::ToolUse { |
| 8207 | thought_signature, .. |
| 8208 | } => thought_signature.clone(), |
| 8209 | _ => None, |
| 8210 | }); |
| 8211 | assert_eq!(signature.as_deref(), Some("SIG-stream")); |
| 8212 | } |
| 8213 | |
| 8214 | #[test] |
| 8215 | fn google_signature_captured_from_streaming_first_chunk() { |
| 8216 | let chunk = json!({ |
| 8217 | "choices": [{ |
| 8218 | "index": 0, |
| 8219 | "delta": { |
| 8220 | "tool_calls": [{ |
| 8221 | "index": 0, |
| 8222 | "id": "call-g-7", |
| 8223 | "type": "function", |
| 8224 | "function": { "name": "read", "arguments": "{}" }, |
| 8225 | "extra_content": { |
| 8226 | "google": { "thought_signature": "SIG-delta" } |
| 8227 | } |
| 8228 | }] |
| 8229 | } |
| 8230 | }] |
| 8231 | }); |
| 8232 | let mut content_index = 0u32; |
| 8233 | let mut text_started = false; |
| 8234 | let mut thinking_started = false; |
| 8235 | let mut tool_indices = std::collections::HashMap::new(); |
| 8236 | let mut reasoning_buffers = std::collections::HashMap::new(); |
| 8237 | let mut inline_tags = InlineReasoningTagState::default(); |
| 8238 | let events = parse_sse_chunk_with_reasoning_style( |
| 8239 | &chunk, |
| 8240 | &mut content_index, |
| 8241 | &mut text_started, |
| 8242 | &mut thinking_started, |
| 8243 | &mut tool_indices, |
| 8244 | &mut reasoning_buffers, |
| 8245 | &mut inline_tags, |
| 8246 | ReasoningStreamStyle::None, |
| 8247 | ); |
| 8248 | let signature = events.iter().find_map(|event| match event { |
| 8249 | StreamEvent::ContentBlockStart { |
| 8250 | content_block: |
| 8251 | ContentBlockStart::ToolUse { |
| 8252 | thought_signature, .. |
| 8253 | }, |
| 8254 | .. |
| 8255 | } => thought_signature.clone(), |
| 8256 | _ => None, |
| 8257 | }); |
| 8258 | assert_eq!(signature.as_deref(), Some("SIG-delta")); |
| 8259 | } |
| 8260 | |
| 8261 | /// Run the production restart/resume chain over a message history: |
| 8262 | /// persist to disk, reload, repair crashed tool pairs, then project for |
| 8263 | /// restore exactly as `apply.rs` does before assigning `api_messages`. |
| 8264 | /// Returns the recovery receipt, the restored messages, and the raw |
| 8265 | /// session JSON as it actually sits on disk. |
| 8266 | fn resumed( |
| 8267 | messages: &[Message], |
| 8268 | ) -> ( |
| 8269 | crate::session_manager::SessionRecovery, |
| 8270 | Vec<Message>, |
| 8271 | String, |
| 8272 | ) { |
| 8273 | let dir = tempfile::tempdir().expect("tempdir"); |
| 8274 | let manager = crate::session_manager::SessionManager::new(dir.path().join("sessions")) |
| 8275 | .expect("session manager"); |
| 8276 | let session = crate::session_manager::create_saved_session( |
| 8277 | messages, |
| 8278 | "gemini-3.1-pro-preview", |
| 8279 | dir.path(), |
| 8280 | 0, |
| 8281 | None, |
| 8282 | ); |
| 8283 | let id = session.metadata.id.clone(); |
| 8284 | let path = manager.save_session(&session).expect("save session"); |
| 8285 | let on_disk = std::fs::read_to_string(&path).expect("read persisted session"); |
| 8286 | let recovery = manager |
| 8287 | .recover_session_for_resume(&id) |
| 8288 | .expect("recover session for resume"); |
| 8289 | let restored = crate::runtime_handoff::project_owned_messages_for_restore( |
| 8290 | recovery.session.messages.clone(), |
| 8291 | ); |
| 8292 | (recovery, restored, on_disk) |
| 8293 | } |
| 8294 | |
| 8295 | fn replayed_signature(body: &serde_json::Value) -> Option<String> { |
| 8296 | body["messages"] |
| 8297 | .as_array() |
| 8298 | .expect("wire messages") |
| 8299 | .iter() |
| 8300 | // The crash repair appends a trailing assistant text receipt, so |
| 8301 | // find the tool-call message by shape, never by index. |
| 8302 | .find(|message| message.get("tool_calls").is_some()) |
| 8303 | .expect("assistant tool-call message") |
| 8304 | .pointer("/tool_calls/0/extra_content/google/thought_signature") |
| 8305 | .and_then(serde_json::Value::as_str) |
| 8306 | .map(str::to_string) |
| 8307 | } |
| 8308 | |
| 8309 | /// C22 done-evidence item 3: restarting and resuming must continue the |
| 8310 | /// signed history. This walks the whole persistence chain — durable JSON, |
| 8311 | /// reload, restore projection, wire body — because the signature can be |
| 8312 | /// lost at any of them, and a serde round-trip alone would prove none of |
| 8313 | /// it. Local unit evidence only: it does not exercise a real Gemini call. |
| 8314 | #[test] |
| 8315 | fn resumed_google_session_replays_signed_tool_calls() { |
| 8316 | let original = signed_history(Some("SIG-abc123"), true); |
| 8317 | let (recovery, restored, on_disk) = resumed(&original); |
| 8318 | |
| 8319 | assert!( |
| 8320 | !recovery.changed, |
| 8321 | "a fully paired history needs no repair on resume" |
| 8322 | ); |
| 8323 | assert_eq!( |
| 8324 | recovery.session.messages, original, |
| 8325 | "reload must return the signed history unchanged" |
| 8326 | ); |
| 8327 | assert_eq!( |
| 8328 | restored, original, |
| 8329 | "the restore projection must not touch signed tool history" |
| 8330 | ); |
| 8331 | assert!( |
| 8332 | on_disk.contains("\"thought_signature\"") && on_disk.contains("SIG-abc123"), |
| 8333 | "the signature must reach durable storage, not just live memory" |
| 8334 | ); |
| 8335 | |
| 8336 | for (provider, base_url) in [ |
| 8337 | (ProviderKind::Google, DEFAULT_GOOGLE_BASE_URL), |
| 8338 | ( |
| 8339 | ProviderKind::Custom, |
| 8340 | "https://generativelanguage.googleapis.com/v1beta/openai", |
| 8341 | ), |
| 8342 | ] { |
| 8343 | let body = build_chat_wire_body( |
| 8344 | &request_from(restored.clone()), |
| 8345 | provider, |
| 8346 | base_url, |
| 8347 | true, |
| 8348 | None, |
| 8349 | ) |
| 8350 | .expect("a resumed signed history must build, not fail closed"); |
| 8351 | assert_eq!( |
| 8352 | replayed_signature(&body.body).as_deref(), |
| 8353 | Some("SIG-abc123"), |
| 8354 | "resumed history must still replay the signature ({base_url})" |
| 8355 | ); |
| 8356 | } |
| 8357 | } |
| 8358 | |
| 8359 | /// The real restart shape: the process died between the tool call and its |
| 8360 | /// result, so resume runs `repair_tool_call_pairs`, which rebuilds every |
| 8361 | /// message. That rebuild must not strip `ToolUse` fields — if it did, the |
| 8362 | /// repaired history would fail closed on Gemini 3 forever after. |
| 8363 | #[test] |
| 8364 | fn crash_repaired_resume_keeps_the_signature_on_the_repaired_tool_call() { |
| 8365 | let (recovery, restored, on_disk) = resumed(&signed_history(Some("SIG-abc123"), false)); |
| 8366 | |
| 8367 | assert!(recovery.changed, "a dangling tool call must be repaired"); |
| 8368 | assert_eq!(recovery.repaired_call_count, 1); |
| 8369 | assert_eq!(recovery.duplicate_result_count, 0); |
| 8370 | assert_eq!(recovery.orphan_result_count, 0); |
| 8371 | assert!( |
| 8372 | on_disk.contains("\"thought_signature\"") && on_disk.contains("SIG-abc123"), |
| 8373 | "the signature must reach durable storage, not just live memory" |
| 8374 | ); |
| 8375 | assert!( |
| 8376 | restored |
| 8377 | .iter() |
| 8378 | .any(|message| message.content.iter().any(|block| matches!( |
| 8379 | block, |
| 8380 | ContentBlock::ToolResult { tool_use_id, content, .. } |
| 8381 | if tool_use_id == "call-g-1" && content.contains("crashed_and_repaired") |
| 8382 | ))), |
| 8383 | "the repair must pair the dangling call with a terminal result" |
| 8384 | ); |
| 8385 | |
| 8386 | let body = build_chat_wire_body( |
| 8387 | &request_from(restored), |
| 8388 | ProviderKind::Google, |
| 8389 | DEFAULT_GOOGLE_BASE_URL, |
| 8390 | true, |
| 8391 | None, |
| 8392 | ) |
| 8393 | .expect("a crash-repaired signed history must build, not fail closed"); |
| 8394 | assert_eq!( |
| 8395 | replayed_signature(&body.body).as_deref(), |
| 8396 | Some("SIG-abc123"), |
| 8397 | "crash repair must not strip the thought signature" |
| 8398 | ); |
| 8399 | } |
| 8400 | } |
| 8401 |