| 1 | // Package responses implements the OpenAI Responses API wire protocol. |
| 2 | // DeepSeek uses it statelessly and requires the complete input history on every |
| 3 | // request; compatible stateful endpoints may opt into previous_response_id. |
| 4 | package responses |
| 5 | |
| 6 | import ( |
| 7 | "bufio" |
| 8 | "bytes" |
| 9 | "context" |
| 10 | "crypto/sha256" |
| 11 | "encoding/hex" |
| 12 | "encoding/json" |
| 13 | "errors" |
| 14 | "fmt" |
| 15 | "io" |
| 16 | "net/http" |
| 17 | "strings" |
| 18 | "sync" |
| 19 | "sync/atomic" |
| 20 | "time" |
| 21 | |
| 22 | "reasonix/internal/netclient" |
| 23 | "reasonix/internal/provider" |
| 24 | ) |
| 25 | |
| 26 | const ( |
| 27 | defaultStreamIdleTimeout = 120 * time.Second |
| 28 | maxReplayableSearchItemBytes = 512 * 1024 |
| 29 | ) |
| 30 | |
| 31 | func init() { |
| 32 | provider.Register("responses", newFromConfig) |
| 33 | provider.Register("dashscope-responses", newFromConfig) |
| 34 | } |
| 35 | |
| 36 | func newFromConfig(cfg provider.Config) (provider.Provider, error) { |
| 37 | effort, _ := cfg.Extra["effort"].(string) |
| 38 | mode, _ := cfg.Extra["mode"].(string) |
| 39 | webSearch, _ := cfg.Extra["web_search"].(bool) |
| 40 | var stateful *bool |
| 41 | switch value := cfg.Extra["stateful"].(type) { |
| 42 | case bool: |
| 43 | stateful = &value |
| 44 | case *bool: |
| 45 | stateful = value |
| 46 | } |
| 47 | proxy, _ := cfg.Extra["proxy_spec"].(netclient.ProxySpec) |
| 48 | keyEnv, _ := cfg.Extra["api_key_env"].(string) |
| 49 | keySource, _ := cfg.Extra["api_key_source"].(string) |
| 50 | maxOutputTokens, _ := cfg.Extra["max_output_tokens"].(int) |
| 51 | return New(Config{ |
| 52 | Name: cfg.Name, APIKey: cfg.APIKey, BaseURL: cfg.BaseURL, Model: cfg.Model, |
| 53 | Effort: effort, Mode: mode, Stateful: stateful, WebSearch: webSearch, Proxy: proxy, |
| 54 | KeyEnv: keyEnv, KeySource: keySource, MaxOutputTokens: maxOutputTokens, |
| 55 | // Extra 原样透传：vision 等能力开关由调用方（boot/CLI）写入 |
| 56 | // cfg.Extra，factory 若丢弃则 New() 读不到（评审 #7234 第 3 点）。 |
| 57 | Extra: cfg.Extra, |
| 58 | }), nil |
| 59 | } |
| 60 | |
| 61 | // Config holds Responses API provider settings. |
| 62 | type Config struct { |
| 63 | Name string |
| 64 | APIKey string |
| 65 | BaseURL string |
| 66 | Model string |
| 67 | Effort string |
| 68 | Mode string // stateful | stateless; empty uses vendor detection. |
| 69 | Stateful *bool // legacy form of Mode; nil preserves vendor detection. |
| 70 | WebSearch bool // expose the provider-executed web_search tool. |
| 71 | Proxy netclient.ProxySpec |
| 72 | KeyEnv string |
| 73 | KeySource string |
| 74 | // MaxOutputTokens is the total provider output budget. Zero enables Reasonix's |
| 75 | // 32K reasoning safety default on official DeepSeek and otherwise omits the |
| 76 | // field; thinking-disabled DeepSeek requests and negative values omit it. |
| 77 | MaxOutputTokens int |
| 78 | // SessionCache controls DashScope's opt-in header. The header is never sent |
| 79 | // to non-DashScope endpoints even when this value is true. |
| 80 | SessionCache *bool |
| 81 | // Extra carries kind-specific options; "vision" (bool) enables embedding |
| 82 | // attached Images as input_image parts on user turns. |
| 83 | Extra map[string]any |
| 84 | } |
| 85 | |
| 86 | func (c Config) mode() string { |
| 87 | mode := strings.ToLower(strings.TrimSpace(c.Mode)) |
| 88 | if mode == "stateful" || mode == "stateless" { |
| 89 | return mode |
| 90 | } |
| 91 | if c.Stateful != nil { |
| 92 | if *c.Stateful { |
| 93 | return "stateful" |
| 94 | } |
| 95 | return "stateless" |
| 96 | } |
| 97 | if capabilitiesFor(DetectVendor(c.BaseURL)).stateless { |
| 98 | return "stateless" |
| 99 | } |
| 100 | return "stateful" |
| 101 | } |
| 102 | |
| 103 | // DetectVendor lives in vendor.go (capabilities table): it covers dashscope/ |
| 104 | // deepseek (incl. eu.deepseek.com) / mimo via exact-host matching. |
| 105 | |
| 106 | type client struct { |
| 107 | name, apiKey, keyEnv, keySource string |
| 108 | baseURL, model, effort string |
| 109 | vendor, mode string |
| 110 | caps vendorCapabilities |
| 111 | sessionCache bool |
| 112 | webSearch bool |
| 113 | maxOutputTokens int |
| 114 | vision bool // model accepts image input; embed Images as input_image parts |
| 115 | http *http.Client |
| 116 | idleTimeout time.Duration |
| 117 | authed atomic.Bool |
| 118 | |
| 119 | mu sync.Mutex |
| 120 | lastResponseID string |
| 121 | expectedPrefixDigest string |
| 122 | } |
| 123 | |
| 124 | // New creates a Responses API provider. |
| 125 | func New(cfg Config) provider.Provider { |
| 126 | vendor := DetectVendor(cfg.BaseURL) |
| 127 | cap := capabilitiesFor(vendor) |
| 128 | maxOutputTokens := cfg.MaxOutputTokens |
| 129 | // 默认输出预算从 vendor 表取（deepseek 32K / mimo 64K）——消除硬编码 |
| 130 | // 常量分叉（review：responses.go 硬编码与 caps.defaultMaxOutputTokens |
| 131 | // 职责重叠）。条件保留：thinking-disabled 的 deepseek 请求不设自动 |
| 132 | // 预算（与 openai.go 一致——服务端默认即可；测试断言该行为）。 |
| 133 | if maxOutputTokens == 0 && cap.defaultMaxOutputTokens > 0 && |
| 134 | !(vendor == "deepseek" && responsesReasoningDisabled(cfg.Effort)) { |
| 135 | maxOutputTokens = cap.defaultMaxOutputTokens |
| 136 | } |
| 137 | sessionCache := cap.sessionCacheHeader |
| 138 | if cfg.SessionCache != nil { |
| 139 | sessionCache = *cfg.SessionCache |
| 140 | } |
| 141 | vision, _ := cfg.Extra["vision"].(bool) |
| 142 | httpClient := &http.Client{Timeout: 300 * time.Second} |
| 143 | if built, err := netclient.NewHTTPClient(cfg.Proxy, netclient.TransportOptions{ |
| 144 | DialTimeout: 30 * time.Second, KeepAlive: 30 * time.Second, |
| 145 | TLSHandshakeTimeout: 15 * time.Second, ResponseHeaderTimeout: 120 * time.Second, |
| 146 | }); err == nil { |
| 147 | httpClient = built |
| 148 | } |
| 149 | return &client{ |
| 150 | name: cfg.Name, apiKey: cfg.APIKey, keyEnv: cfg.KeyEnv, keySource: cfg.KeySource, |
| 151 | baseURL: strings.TrimRight(cfg.BaseURL, "/"), model: cfg.Model, effort: cfg.Effort, |
| 152 | vendor: vendor, caps: cap, mode: cfg.mode(), sessionCache: sessionCache, webSearch: cfg.WebSearch, maxOutputTokens: maxOutputTokens, |
| 153 | vision: vision, |
| 154 | http: httpClient, idleTimeout: defaultStreamIdleTimeout, |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | func responsesReasoningDisabled(effort string) bool { |
| 159 | switch strings.ToLower(strings.TrimSpace(effort)) { |
| 160 | case "none", "disabled", "off": |
| 161 | return true |
| 162 | default: |
| 163 | return false |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | func (c *client) Name() string { return c.name } |
| 168 | |
| 169 | // RequiresToolCallReasoning tells the agent to preserve stateless vendors' |
| 170 | // reasoning on assistant tool-call turns so the follow-up can replay it. |
| 171 | // DeepSeek and MiMo document this requirement for multi-turn tool calls. |
| 172 | func (c *client) RequiresToolCallReasoning() bool { |
| 173 | return c.caps.toolCallReasoning |
| 174 | } |
| 175 | |
| 176 | func (c *client) MissingToolCallReasoningWarningIdentity() string { |
| 177 | if c == nil { |
| 178 | return "" |
| 179 | } |
| 180 | return strings.Join([]string{ |
| 181 | "responses", strings.TrimSpace(c.name), strings.TrimSpace(c.baseURL), |
| 182 | strings.TrimSpace(c.model), strings.TrimSpace(c.vendor), strings.TrimSpace(c.mode), strings.TrimSpace(c.effort), |
| 183 | }, "\x00") |
| 184 | } |
| 185 | |
| 186 | // WarnOnMissingToolCallReasoning reports a tool_calls turn that arrived |
| 187 | // without reasoning only for vendors whose endpoint reliably emits it. |
| 188 | // DeepSeek's official API emits tool-call reasoning for its pro-tier models, |
| 189 | // so a missing chain-of-thought there is a real degradation worth one warning. |
| 190 | // MiMo documents reasoning alongside tool calls but does not guarantee it on |
| 191 | // every round (observed: mimo-v2.5-pro tool-call turn with empty reasoning), |
| 192 | // so a missing chain-of-thought is endpoint-conditional, not a degradation |
| 193 | // signal — silence the warning. Capability-driven (review #7234): |
| 194 | // toolCallReasoning=false vendors (DashScope) never warn — no round-trip |
| 195 | // contract; singleSegmentReasoning=true vendors (MiMo) never warn — their |
| 196 | // tool-call thinking is a single optional segment. Only multi-segment |
| 197 | // thinking vendors that require replay (DeepSeek) warn, scoped to non-flash. |
| 198 | func (c *client) WarnOnMissingToolCallReasoning() bool { |
| 199 | if !c.caps.toolCallReasoning || c.caps.singleSegmentReasoning { |
| 200 | return false |
| 201 | } |
| 202 | model := strings.ToLower(strings.TrimSpace(c.model)) |
| 203 | // Flash-tier DeepSeek models do not emit tool-call reasoning (same carve |
| 204 | // as openai.go expectsDeepSeekToolCallReasoning). |
| 205 | return !strings.Contains(model, "flash") |
| 206 | } |
| 207 | |
| 208 | func (c *client) sendOpts() provider.SendOptions { |
| 209 | return provider.SendOptions{Provider: c.name, KeyEnv: c.keyEnv, KeySource: c.keySource, KeyPresent: c.apiKey != "", RetryAuth: c.authed.Load()} |
| 210 | } |
| 211 | |
| 212 | // ResetContext drops stateful continuation metadata. Full-input stateless mode |
| 213 | // is unaffected. |
| 214 | func (c *client) ResetContext() { |
| 215 | c.mu.Lock() |
| 216 | c.lastResponseID = "" |
| 217 | c.expectedPrefixDigest = "" |
| 218 | c.mu.Unlock() |
| 219 | } |
| 220 | |
| 221 | func (c *client) Stream(ctx context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 222 | requestCtx := provider.WithRequestAttemptCounter(ctx) |
| 223 | body, usedPrevious, wireMessages := c.buildRequestBody(req) |
| 224 | resp, err := c.send(requestCtx, body) |
| 225 | if err != nil && usedPrevious && isStalePreviousResponseError(err) { |
| 226 | // A stateful response ID may expire server-side. Retrying once with full |
| 227 | // history is safe because no response body has started streaming. |
| 228 | c.ResetContext() |
| 229 | body, _, wireMessages = c.buildRequestBody(req) |
| 230 | resp, err = c.send(requestCtx, body) |
| 231 | } |
| 232 | if err != nil { |
| 233 | return nil, err |
| 234 | } |
| 235 | c.authed.Store(true) |
| 236 | out := make(chan provider.Chunk, 64) |
| 237 | go c.readStream(requestCtx, resp, out, wireMessages) |
| 238 | return out, nil |
| 239 | } |
| 240 | |
| 241 | func (c *client) send(ctx context.Context, body map[string]any) (*http.Response, error) { |
| 242 | payload, err := json.Marshal(body) |
| 243 | if err != nil { |
| 244 | return nil, fmt.Errorf("responses: marshal request: %w", err) |
| 245 | } |
| 246 | newRequest := func(ctx context.Context) (*http.Request, error) { |
| 247 | req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/responses", bytes.NewReader(payload)) |
| 248 | if err != nil { |
| 249 | return nil, err |
| 250 | } |
| 251 | req.Header.Set("Content-Type", "application/json") |
| 252 | req.Header.Set("Authorization", "Bearer "+c.apiKey) |
| 253 | if c.caps.sessionCacheHeader && c.sessionCache { |
| 254 | req.Header.Set("x-dashscope-session-cache", "enable") |
| 255 | } |
| 256 | return req, nil |
| 257 | } |
| 258 | return provider.SendWithRetry(ctx, c.http, c.sendOpts(), newRequest) |
| 259 | } |
| 260 | |
| 261 | func isStalePreviousResponseError(err error) bool { |
| 262 | var apiErr *provider.APIError |
| 263 | if !errors.As(err, &apiErr) || apiErr.Status != http.StatusBadRequest { |
| 264 | return false |
| 265 | } |
| 266 | body := strings.ToLower(apiErr.Body) |
| 267 | mentionsID := strings.Contains(body, "previous_response_id") || strings.Contains(body, "previous response") || strings.Contains(body, "response id") |
| 268 | return mentionsID && |
| 269 | (strings.Contains(body, "not found") || strings.Contains(body, "invalid") || strings.Contains(body, "expired")) |
| 270 | } |
| 271 | |
| 272 | func (c *client) buildRequestBody(req provider.Request) (map[string]any, bool, []provider.Message) { |
| 273 | messages := provider.SanitizeToolPairing(provider.ModelMessages(req.Messages)) |
| 274 | body := map[string]any{"model": c.model, "stream": true} |
| 275 | |
| 276 | effort := strings.ToLower(strings.TrimSpace(c.effort)) |
| 277 | switch effort { |
| 278 | case "auto": |
| 279 | effort = "" |
| 280 | case "disabled", "off": |
| 281 | effort = "none" |
| 282 | } |
| 283 | if effort != "" { |
| 284 | body["reasoning"] = map[string]any{"effort": effort} |
| 285 | } |
| 286 | maxOutputTokens := req.MaxTokens |
| 287 | if maxOutputTokens == 0 { |
| 288 | maxOutputTokens = c.maxOutputTokens |
| 289 | } |
| 290 | if maxOutputTokens == 0 && c.caps.defaultMaxOutputTokens > 0 { |
| 291 | // 与 New() 构造期默认同条件：thinking-disabled 的 deepseek 请求 |
| 292 | // 不设自动预算（服务端默认即可——测试断言该行为）。 |
| 293 | if !(c.vendor == "deepseek" && responsesReasoningDisabled(c.effort)) { |
| 294 | maxOutputTokens = c.caps.defaultMaxOutputTokens |
| 295 | } |
| 296 | } |
| 297 | if maxOutputTokens > 0 { |
| 298 | body["max_output_tokens"] = maxOutputTokens |
| 299 | } |
| 300 | if req.ResponseFormat != nil && req.ResponseFormat.Type != "" { |
| 301 | // Structured output: Responses text.format. MiMo/DashScope/OpenAI |
| 302 | // all accept {"text":{"format":{"type":"json_object"}}}. The model |
| 303 | // only emits JSON when the instructions also demand it. |
| 304 | body["text"] = map[string]any{ |
| 305 | "format": map[string]any{"type": req.ResponseFormat.Type}, |
| 306 | } |
| 307 | } |
| 308 | if req.Temperature != nil && !c.caps.ignoresTemperature { |
| 309 | body["temperature"] = *req.Temperature |
| 310 | } |
| 311 | if c.webSearch || len(req.Tools) > 0 { |
| 312 | tools := make([]map[string]any, 0, len(req.Tools)+1) |
| 313 | // Keep the server tool first and stable across turns. DeepSeek executes |
| 314 | // this tool itself; ordinary Reasonix tools remain function entries. |
| 315 | if c.webSearch { |
| 316 | tools = append(tools, map[string]any{"type": "web_search"}) |
| 317 | } |
| 318 | for _, tool := range req.Tools { |
| 319 | parameters := tool.Parameters |
| 320 | if len(parameters) == 0 { |
| 321 | parameters = provider.CanonicalizeSchema(nil) |
| 322 | } |
| 323 | tools = append(tools, map[string]any{ |
| 324 | "type": "function", "name": tool.Name, "description": tool.Description, |
| 325 | "parameters": json.RawMessage(parameters), |
| 326 | }) |
| 327 | } |
| 328 | body["tools"] = tools |
| 329 | } |
| 330 | instructions, rest := splitInstructions(messages) |
| 331 | if instructions != "" { |
| 332 | body["instructions"] = instructions |
| 333 | } |
| 334 | |
| 335 | c.mu.Lock() |
| 336 | previousID, expectedDigest := c.lastResponseID, c.expectedPrefixDigest |
| 337 | c.mu.Unlock() |
| 338 | if c.mode == "stateful" && previousID != "" && len(messages) > 0 && |
| 339 | messages[len(messages)-1].Role == provider.RoleUser && |
| 340 | c.conversationDigest(messages[:len(messages)-1]) == expectedDigest { |
| 341 | body["input"] = messages[len(messages)-1].Content |
| 342 | body["previous_response_id"] = previousID |
| 343 | return body, true, messages |
| 344 | } |
| 345 | |
| 346 | body["input"] = messagesToInput(rest, c.vision, c.vendor == "deepseek", c.caps.summaryRequired) |
| 347 | return body, false, messages |
| 348 | } |
| 349 | |
| 350 | func splitInstructions(messages []provider.Message) (string, []provider.Message) { |
| 351 | if len(messages) == 0 || messages[0].Role != provider.RoleSystem { |
| 352 | return "", messages |
| 353 | } |
| 354 | return messages[0].Content, messages[1:] |
| 355 | } |
| 356 | |
| 357 | func messagesToInput(messages []provider.Message, vision, replayDeepSeekItems, summary bool) []map[string]any { |
| 358 | input := make([]map[string]any, 0, len(messages)*2) |
| 359 | for _, message := range messages { |
| 360 | switch message.Role { |
| 361 | case provider.RoleSystem, provider.RoleUser: |
| 362 | // Text-only turns keep the documented TextInput string shape. |
| 363 | // Vision-capable user turns with attached images switch to the |
| 364 | // InputItemList array form ({type:input_text} + {type:input_image}) |
| 365 | // so the text and every image ride the same message, matching the |
| 366 | // MiMo/DashScope multimodal example. The system message is always |
| 367 | // plain text: images only attach to user turns. |
| 368 | if vision && message.Role == provider.RoleUser && len(message.Images) > 0 { |
| 369 | parts := make([]map[string]string, 0, len(message.Images)+1) |
| 370 | if message.Content != "" { |
| 371 | parts = append(parts, map[string]string{"type": "input_text", "text": message.Content}) |
| 372 | } |
| 373 | for _, url := range message.Images { |
| 374 | parts = append(parts, map[string]string{"type": "input_image", "image_url": url}) |
| 375 | } |
| 376 | input = append(input, map[string]any{"role": "user", "content": parts}) |
| 377 | } else { |
| 378 | input = append(input, map[string]any{"role": string(message.Role), "content": message.Content}) |
| 379 | } |
| 380 | case provider.RoleAssistant: |
| 381 | if message.ReasoningContent != "" { |
| 382 | // Reasoning items: the OpenAI base format only needs |
| 383 | // `content`. DashScope additionally requires a `summary` |
| 384 | // list ("Invalid 'summary': summary is required and must be |
| 385 | // a list for reasoning."). Other vendors (MiMo) do not |
| 386 | // define summary in their schema; sending it leaks the |
| 387 | // reasoning text into an extra field the server may echo |
| 388 | // back into the model context, doubling chain-of-thought |
| 389 | // each turn — so only send it where the wire demands it. |
| 390 | item := map[string]any{ |
| 391 | "type": "reasoning", |
| 392 | "content": []map[string]string{{"type": "reasoning_text", "text": message.ReasoningContent}}, |
| 393 | } |
| 394 | if message.ReasoningID != "" { |
| 395 | // OpenAI Responses schema marks Reasoning.id required; |
| 396 | // round-trip the provider-issued id when we captured one. |
| 397 | item["id"] = message.ReasoningID |
| 398 | } |
| 399 | if message.ReasoningStatus != "" { |
| 400 | item["status"] = message.ReasoningStatus |
| 401 | } |
| 402 | if summary { |
| 403 | item["summary"] = []map[string]string{{"type": "summary_text", "text": message.ReasoningContent}} |
| 404 | } |
| 405 | input = append(input, item) |
| 406 | } |
| 407 | if replayDeepSeekItems { |
| 408 | for _, raw := range message.ResponsesItems { |
| 409 | if item, ok := decodeReplayableWebSearchItem(raw); ok { |
| 410 | input = append(input, item) |
| 411 | } |
| 412 | } |
| 413 | } |
| 414 | if message.Content != "" || len(message.ToolCalls) == 0 { |
| 415 | input = append(input, map[string]any{"role": "assistant", "content": message.Content}) |
| 416 | } |
| 417 | for _, call := range message.ToolCalls { |
| 418 | input = append(input, map[string]any{ |
| 419 | "type": "function_call", "call_id": call.ID, |
| 420 | "name": call.Name, "arguments": call.Arguments, |
| 421 | }) |
| 422 | } |
| 423 | case provider.RoleTool: |
| 424 | input = append(input, map[string]any{ |
| 425 | "type": "function_call_output", "call_id": message.ToolCallID, "output": message.Content, |
| 426 | }) |
| 427 | } |
| 428 | } |
| 429 | return input |
| 430 | } |
| 431 | |
| 432 | func decodeReplayableWebSearchItem(raw json.RawMessage) (map[string]any, bool) { |
| 433 | if len(raw) == 0 || len(raw) > maxReplayableSearchItemBytes || !json.Valid(raw) { |
| 434 | return nil, false |
| 435 | } |
| 436 | var item map[string]any |
| 437 | if err := json.Unmarshal(raw, &item); err != nil || item["type"] != "web_search_call" { |
| 438 | return nil, false |
| 439 | } |
| 440 | id, _ := item["id"].(string) |
| 441 | status, _ := item["status"].(string) |
| 442 | if strings.TrimSpace(id) == "" || status != "completed" { |
| 443 | return nil, false |
| 444 | } |
| 445 | return item, true |
| 446 | } |
| 447 | |
| 448 | func (c *client) conversationDigest(messages []provider.Message) string { |
| 449 | instructions, rest := splitInstructions(messages) |
| 450 | // Digest must mirror the wire exactly: the stateful fast path compares |
| 451 | // this against the previous request's input, so a mismatch would skip |
| 452 | // previous_response_id and force a full replay (cache-hit loss). Use the |
| 453 | // same vision/summary knobs as buildRequestBody. |
| 454 | payload, _ := json.Marshal(struct { |
| 455 | Instructions string `json:"instructions,omitempty"` |
| 456 | Input []map[string]any `json:"input"` |
| 457 | }{Instructions: instructions, Input: messagesToInput(rest, c.vision, c.vendor == "deepseek", c.caps.summaryRequired)}) |
| 458 | sum := sha256.Sum256(payload) |
| 459 | return hex.EncodeToString(sum[:]) |
| 460 | } |
| 461 | |
| 462 | type streamedCall struct { |
| 463 | id, name, arguments string |
| 464 | argChars int |
| 465 | completed bool |
| 466 | } |
| 467 | |
| 468 | func (c *client) readStream(ctx context.Context, resp *http.Response, out chan<- provider.Chunk, requestMessages []provider.Message) { |
| 469 | defer resp.Body.Close() |
| 470 | defer close(out) |
| 471 | |
| 472 | scanner := bufio.NewScanner(resp.Body) |
| 473 | scanner.Buffer(make([]byte, 64*1024), 4*1024*1024) |
| 474 | idle := c.idleTimeout |
| 475 | if idle <= 0 { |
| 476 | idle = defaultStreamIdleTimeout |
| 477 | } |
| 478 | watchDone := make(chan struct{}) |
| 479 | activity := make(chan struct{}, 1) |
| 480 | var stalled atomic.Bool |
| 481 | go func() { |
| 482 | timer := time.NewTimer(idle) |
| 483 | defer timer.Stop() |
| 484 | for { |
| 485 | select { |
| 486 | case <-ctx.Done(): |
| 487 | _ = resp.Body.Close() |
| 488 | return |
| 489 | case <-watchDone: |
| 490 | return |
| 491 | case <-activity: |
| 492 | if !timer.Stop() { |
| 493 | select { |
| 494 | case <-timer.C: |
| 495 | default: |
| 496 | } |
| 497 | } |
| 498 | timer.Reset(idle) |
| 499 | case <-timer.C: |
| 500 | stalled.Store(true) |
| 501 | _ = resp.Body.Close() |
| 502 | return |
| 503 | } |
| 504 | } |
| 505 | }() |
| 506 | defer close(watchDone) |
| 507 | |
| 508 | calls := make(map[string]*streamedCall) |
| 509 | callOrder := make([]string, 0) |
| 510 | callForItem := func(itemID string) *streamedCall { |
| 511 | if call := calls[itemID]; call != nil { |
| 512 | return call |
| 513 | } |
| 514 | call := &streamedCall{id: itemID} |
| 515 | calls[itemID] = call |
| 516 | callOrder = append(callOrder, itemID) |
| 517 | return call |
| 518 | } |
| 519 | textDeltas := make(map[string]bool) |
| 520 | reasoningDeltas := make(map[string]bool) |
| 521 | seenSearchItems := make(map[string]struct{}) |
| 522 | var responsesItems []json.RawMessage |
| 523 | var text, reasoning strings.Builder |
| 524 | reasoningID := "" |
| 525 | reasoningStatus := "" |
| 526 | terminal := false |
| 527 | failed := false |
| 528 | completedResponseID := "" |
| 529 | |
| 530 | for scanner.Scan() { |
| 531 | select { |
| 532 | case activity <- struct{}{}: |
| 533 | default: |
| 534 | } |
| 535 | line := scanner.Text() |
| 536 | if !strings.HasPrefix(line, "data:") { |
| 537 | continue |
| 538 | } |
| 539 | data := strings.TrimSpace(strings.TrimPrefix(line, "data:")) |
| 540 | if data == "[DONE]" { |
| 541 | terminal = true |
| 542 | break |
| 543 | } |
| 544 | var event sseEvent |
| 545 | if json.Unmarshal([]byte(data), &event) != nil { |
| 546 | continue |
| 547 | } |
| 548 | key := fmt.Sprintf("%s:%d", event.ItemID, event.ContentIndex) |
| 549 | switch event.Type { |
| 550 | case "response.output_text.delta": |
| 551 | textDeltas[key] = true |
| 552 | text.WriteString(event.Delta) |
| 553 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkText, Text: event.Delta}) { |
| 554 | return |
| 555 | } |
| 556 | case "response.output_text.done": |
| 557 | if event.Text != "" && !textDeltas[key] { |
| 558 | text.WriteString(event.Text) |
| 559 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkText, Text: event.Text}) { |
| 560 | return |
| 561 | } |
| 562 | } |
| 563 | case "response.reasoning_text.delta", "response.reasoning_summary_text.delta": |
| 564 | reasoningDeltas[key] = true |
| 565 | reasoning.WriteString(event.Delta) |
| 566 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, Text: event.Delta}) { |
| 567 | return |
| 568 | } |
| 569 | case "response.reasoning_text.done", "response.reasoning_summary_text.done": |
| 570 | if event.Text != "" && !reasoningDeltas[key] { |
| 571 | reasoning.WriteString(event.Text) |
| 572 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, Text: event.Text}) { |
| 573 | return |
| 574 | } |
| 575 | } |
| 576 | case "response.output_item.added": |
| 577 | if event.Item != nil { |
| 578 | switch event.Item.Type { |
| 579 | case "function_call": |
| 580 | call := callForItem(event.Item.ID) |
| 581 | call.id = event.Item.CallID |
| 582 | call.name = event.Item.Name |
| 583 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCallStart, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name}}) { |
| 584 | return |
| 585 | } |
| 586 | case "reasoning": |
| 587 | // Capture the provider-issued reasoning item id so the |
| 588 | // next turn's input reasoning item can carry it (the |
| 589 | // OpenAI Responses schema marks Reasoning.id required). |
| 590 | if event.Item.ID != "" { |
| 591 | // 多段推理（DeepSeek 长思考分多段）时末段 id 覆盖：round-trip |
| 592 | // 合并为一个 reasoning item 只带末段 id（服务端接受）。 |
| 593 | reasoningID = event.Item.ID |
| 594 | } |
| 595 | } |
| 596 | } |
| 597 | case "response.function_call_arguments.delta": |
| 598 | call := callForItem(event.ItemID) |
| 599 | call.arguments += event.Delta |
| 600 | call.argChars += len(event.Delta) |
| 601 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCallArgsDelta, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name}, ArgChars: call.argChars}) { |
| 602 | return |
| 603 | } |
| 604 | case "response.function_call_arguments.done": |
| 605 | call := callForItem(event.ItemID) |
| 606 | if event.Arguments != "" { |
| 607 | call.arguments = event.Arguments |
| 608 | } |
| 609 | if !call.completed { |
| 610 | call.completed = true |
| 611 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}}) { |
| 612 | return |
| 613 | } |
| 614 | } |
| 615 | case "response.output_item.done": |
| 616 | if event.Item != nil && event.Item.Type == "web_search_call" && c.vendor == "deepseek" { |
| 617 | if _, ok := decodeReplayableWebSearchItem(event.Item.Raw); ok { |
| 618 | key := event.Item.ID |
| 619 | if key == "" { |
| 620 | key = string(event.Item.Raw) |
| 621 | } |
| 622 | if _, seen := seenSearchItems[key]; !seen { |
| 623 | seenSearchItems[key] = struct{}{} |
| 624 | raw := append(json.RawMessage(nil), event.Item.Raw...) |
| 625 | responsesItems = append(responsesItems, raw) |
| 626 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkResponsesItem, ResponsesItem: raw}) { |
| 627 | return |
| 628 | } |
| 629 | } |
| 630 | } |
| 631 | } |
| 632 | if event.Item != nil { |
| 633 | switch event.Item.Type { |
| 634 | case "function_call": |
| 635 | call := callForItem(event.Item.ID) |
| 636 | if event.Item.CallID != "" { |
| 637 | call.id = event.Item.CallID |
| 638 | } |
| 639 | if event.Item.Name != "" { |
| 640 | call.name = event.Item.Name |
| 641 | } |
| 642 | if event.Item.Arguments != "" { |
| 643 | call.arguments = event.Item.Arguments |
| 644 | } |
| 645 | if !call.completed { |
| 646 | call.completed = true |
| 647 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}}) { |
| 648 | return |
| 649 | } |
| 650 | } |
| 651 | case "reasoning": |
| 652 | // The done event carries the final item status |
| 653 | // ("completed" after the thinking stream finishes); |
| 654 | // round-trip it with the reasoning item so the input |
| 655 | // matches the wire schema. |
| 656 | if event.Item.Status != "" { |
| 657 | reasoningStatus = event.Item.Status |
| 658 | } |
| 659 | } |
| 660 | } |
| 661 | case "response.completed", "response.incomplete", "response.failed": |
| 662 | terminal = true |
| 663 | if event.Response != nil { |
| 664 | if event.Type == "response.completed" { |
| 665 | completedResponseID = event.Response.ID |
| 666 | } |
| 667 | usage := usageFromResponse(event.Response) |
| 668 | provider.ApplyRequestAttemptCount(ctx, usage) |
| 669 | if event.Type == "response.incomplete" { |
| 670 | switch event.Response.IncompleteDetails.Reason { |
| 671 | case "max_output_tokens": |
| 672 | usage.FinishReason = "length" |
| 673 | case "content_filter": |
| 674 | usage.FinishReason = "content_filter" |
| 675 | default: |
| 676 | usage.FinishReason = "incomplete" |
| 677 | } |
| 678 | } else if event.Type == "response.completed" && usage.FinishReason == "" { |
| 679 | // A completed response finished normally (stop). Preserve any |
| 680 | // vendor-specific reason already set by usageFromResponse. |
| 681 | usage.FinishReason = "stop" |
| 682 | } |
| 683 | // DashScope occasionally reports a completed event whose usage |
| 684 | // object exists but is all zeros (server-side reporting gap; the |
| 685 | // tokens were actually billed). Emitting that as ChunkUsage |
| 686 | // would corrupt cache-ratio and cost accounting with a spurious |
| 687 | // zero record. 但完成语义必须保留：全零+stop 也发送——计费层 |
| 688 | // （Pricing.Cost）对全零记录天然返回 0 成本，不污染统计；而 |
| 689 | // agent 侧 reasoningOnlyFinishHonoured 依赖收到 usage 对象 |
| 690 | // （FinishReason=stop）才能确认 reasoning-only 完成（#7168 |
| 691 | // 评审"完成语义保留"的完整实现——此前 stop 被抑制时该语义 |
| 692 | // 失效，空回复被误判触发重试）。异常终止 reason |
| 693 | // （length/content_filter/...）始终上报。 |
| 694 | if usage.TotalTokens > 0 || usage.FinishReason != "" { |
| 695 | |
| 696 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkUsage, Usage: usage}) { |
| 697 | return |
| 698 | } |
| 699 | } |
| 700 | } |
| 701 | if event.Type == "response.failed" { |
| 702 | failed = true |
| 703 | err := fmt.Errorf("responses: response failed") |
| 704 | if event.Response != nil && event.Response.Error != nil { |
| 705 | if authErr := authErrorFromResponse(c, event.Response.Error); authErr != nil { |
| 706 | err = authErr |
| 707 | } else { |
| 708 | err = fmt.Errorf("responses: %s", event.Response.Error.Message) |
| 709 | } |
| 710 | } |
| 711 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: err}) { |
| 712 | return |
| 713 | } |
| 714 | } |
| 715 | } |
| 716 | if terminal { |
| 717 | break |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | if ctx.Err() != nil { |
| 722 | return |
| 723 | } |
| 724 | if err := scanner.Err(); err != nil { |
| 725 | var reason string |
| 726 | if stalled.Load() { |
| 727 | err = fmt.Errorf("responses: stream idle timeout after %s", idle) |
| 728 | reason = provider.StreamInterruptIdleTimeout |
| 729 | } else { |
| 730 | reason = provider.ClassifyStreamInterrupt(err) |
| 731 | } |
| 732 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: provider.StreamInterrupt(err, reason)}) |
| 733 | return |
| 734 | } |
| 735 | // Protocol-defined terminal response events are required. Connection close |
| 736 | // before a terminal event leaves the attempt uncommitted — including any |
| 737 | // complete tool calls already forwarded as speculative output. |
| 738 | if !terminal { |
| 739 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: provider.StreamInterrupt(io.ErrUnexpectedEOF, provider.StreamInterruptPrematureEOF)}) |
| 740 | return |
| 741 | } |
| 742 | if completedResponseID != "" { |
| 743 | assistant := provider.Message{Role: provider.RoleAssistant, Content: text.String(), ReasoningContent: reasoning.String(), ReasoningID: reasoningID, ReasoningStatus: reasoningStatus, ResponsesItems: responsesItems} |
| 744 | for _, itemID := range callOrder { |
| 745 | call := calls[itemID] |
| 746 | if call.completed { |
| 747 | assistant.ToolCalls = append(assistant.ToolCalls, provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}) |
| 748 | } |
| 749 | } |
| 750 | expected := append(append([]provider.Message(nil), requestMessages...), assistant) |
| 751 | c.mu.Lock() |
| 752 | c.lastResponseID = completedResponseID |
| 753 | c.expectedPrefixDigest = c.conversationDigest(expected) |
| 754 | c.mu.Unlock() |
| 755 | } else { |
| 756 | c.ResetContext() |
| 757 | } |
| 758 | if !failed { |
| 759 | // 把 reasoning item 的 id/status 作为元数据 chunk 流给 Agent |
| 760 | // （空 Text，随 ChunkReasoning 语义）——Agent 持久化进 session， |
| 761 | // 下一轮 input reasoning item 回传 id/status（评审 #7234 第 1 点）。 |
| 762 | if reasoningID != "" || reasoningStatus != "" { |
| 763 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, ReasoningID: reasoningID, ReasoningStatus: reasoningStatus}) { |
| 764 | return |
| 765 | } |
| 766 | } |
| 767 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkDone}) |
| 768 | } |
| 769 | } |
| 770 | |
| 771 | func sendChunk(ctx context.Context, out chan<- provider.Chunk, chunk provider.Chunk) bool { |
| 772 | select { |
| 773 | case out <- chunk: |
| 774 | return true |
| 775 | default: |
| 776 | } |
| 777 | select { |
| 778 | case out <- chunk: |
| 779 | return true |
| 780 | case <-ctx.Done(): |
| 781 | return false |
| 782 | } |
| 783 | } |
| 784 | |
| 785 | func usageFromResponse(response *sseResponse) *provider.Usage { |
| 786 | usage := &provider.Usage{} |
| 787 | if response == nil || response.Usage == nil { |
| 788 | return usage |
| 789 | } |
| 790 | u := response.Usage |
| 791 | cached, reasoning := 0, 0 |
| 792 | if u.InputTokensDetails != nil { |
| 793 | cached = u.InputTokensDetails.CachedTokens |
| 794 | } |
| 795 | if u.OutputTokensDetails != nil { |
| 796 | reasoning = u.OutputTokensDetails.ReasoningTokens |
| 797 | } |
| 798 | miss := u.InputTokens - cached |
| 799 | if miss < 0 { |
| 800 | miss = 0 |
| 801 | } |
| 802 | total := u.TotalTokens |
| 803 | if total == 0 { |
| 804 | total = u.InputTokens + u.OutputTokens |
| 805 | } |
| 806 | return &provider.Usage{PromptTokens: u.InputTokens, CompletionTokens: u.OutputTokens, TotalTokens: total, CacheHitTokens: cached, CacheMissTokens: miss, ReasoningTokens: reasoning} |
| 807 | } |
| 808 | |
| 809 | func authErrorFromResponse(c *client, responseError *sseError) error { |
| 810 | if responseError == nil { |
| 811 | return nil |
| 812 | } |
| 813 | value := strings.ToLower(responseError.Code + " " + responseError.Message) |
| 814 | if !strings.Contains(value, "auth") && !strings.Contains(value, "api key") && !strings.Contains(value, "unauthorized") && !strings.Contains(value, "forbidden") && !strings.Contains(value, "permission") { |
| 815 | return nil |
| 816 | } |
| 817 | status := http.StatusUnauthorized |
| 818 | if strings.Contains(value, "forbidden") || strings.Contains(value, "permission") { |
| 819 | status = http.StatusForbidden |
| 820 | } |
| 821 | return &provider.AuthError{Provider: c.name, KeyEnv: c.keyEnv, KeySource: c.keySource, Status: status, HasKey: c.apiKey != "", Body: responseError.Message} |
| 822 | } |
| 823 | |
| 824 | type sseEvent struct { |
| 825 | Type string `json:"type"` |
| 826 | Delta string `json:"delta"` |
| 827 | Text string `json:"text"` |
| 828 | Arguments string `json:"arguments"` |
| 829 | ItemID string `json:"item_id"` |
| 830 | ContentIndex int `json:"content_index"` |
| 831 | Item *sseItem `json:"item"` |
| 832 | Response *sseResponse `json:"response"` |
| 833 | } |
| 834 | |
| 835 | type sseItem struct { |
| 836 | ID, Type, CallID, Name, Arguments, Status string |
| 837 | Raw json.RawMessage |
| 838 | } |
| 839 | |
| 840 | func (i *sseItem) UnmarshalJSON(data []byte) error { |
| 841 | var wire struct { |
| 842 | ID string `json:"id"` |
| 843 | Type string `json:"type"` |
| 844 | CallID string `json:"call_id"` |
| 845 | Name string `json:"name"` |
| 846 | Arguments string `json:"arguments"` |
| 847 | Status string `json:"status"` |
| 848 | } |
| 849 | if err := json.Unmarshal(data, &wire); err != nil { |
| 850 | return err |
| 851 | } |
| 852 | *i = sseItem{ID: wire.ID, Type: wire.Type, CallID: wire.CallID, Name: wire.Name, Arguments: wire.Arguments, Status: wire.Status, Raw: append(json.RawMessage(nil), data...)} |
| 853 | return nil |
| 854 | } |
| 855 | |
| 856 | type sseResponse struct { |
| 857 | ID string `json:"id"` |
| 858 | Usage *sseUsage `json:"usage"` |
| 859 | Error *sseError `json:"error"` |
| 860 | IncompleteDetails incompleteDetails `json:"incomplete_details"` |
| 861 | } |
| 862 | |
| 863 | type incompleteDetails struct { |
| 864 | Reason string `json:"reason"` |
| 865 | } |
| 866 | type sseError struct { |
| 867 | Message string `json:"message"` |
| 868 | Code string `json:"code"` |
| 869 | } |
| 870 | type sseUsage struct { |
| 871 | InputTokens int `json:"input_tokens"` |
| 872 | OutputTokens int `json:"output_tokens"` |
| 873 | TotalTokens int `json:"total_tokens"` |
| 874 | InputTokensDetails *struct { |
| 875 | CachedTokens int `json:"cached_tokens"` |
| 876 | } `json:"input_tokens_details"` |
| 877 | OutputTokensDetails *struct { |
| 878 | ReasoningTokens int `json:"reasoning_tokens"` |
| 879 | } `json:"output_tokens_details"` |
| 880 | } |
| 881 |