| 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 | "maps" |
| 17 | "net/http" |
| 18 | "strings" |
| 19 | "sync" |
| 20 | "sync/atomic" |
| 21 | "time" |
| 22 | |
| 23 | "reasonix/internal/netclient" |
| 24 | "reasonix/internal/provider" |
| 25 | "reasonix/internal/provider/openai" |
| 26 | ) |
| 27 | |
| 28 | const ( |
| 29 | defaultStreamIdleTimeout = 300 * time.Second |
| 30 | maxReplayableSearchItemBytes = 512 * 1024 |
| 31 | ) |
| 32 | |
| 33 | func init() { |
| 34 | provider.RegisterReasoning("responses", ReasoningForConfig) |
| 35 | provider.RegisterReasoning("dashscope-responses", ReasoningForConfig) |
| 36 | provider.Register("responses", newFromConfig) |
| 37 | provider.Register("dashscope-responses", newFromConfig) |
| 38 | } |
| 39 | |
| 40 | // Config holds Responses API provider settings. |
| 41 | type Config struct { |
| 42 | HTTPClient *http.Client |
| 43 | Name string |
| 44 | DisplayName string |
| 45 | Protocol string |
| 46 | APIKey string |
| 47 | BaseURL string |
| 48 | Model string |
| 49 | ModelInfo *provider.ModelInfo |
| 50 | Effort string |
| 51 | Mode string // stateful | stateless; empty uses vendor detection. |
| 52 | Stateful *bool // legacy form of Mode; nil preserves vendor detection. |
| 53 | WebSearch bool // expose the provider-executed web_search tool. |
| 54 | Proxy netclient.ProxySpec |
| 55 | KeyEnv string |
| 56 | KeySource string |
| 57 | RequestURL string // optional exact Responses request URL; empty derives from BaseURL |
| 58 | // MaxOutputTokens is the total provider output budget. Zero omits the field |
| 59 | // on official DeepSeek (server 384K ceiling) and unknown endpoints; MiMo |
| 60 | // still applies its 16K/32K ladder. Negative values omit it. |
| 61 | MaxOutputTokens int |
| 62 | // SessionCache controls DashScope's opt-in header. The header is never sent |
| 63 | // to non-DashScope endpoints even when this value is true. |
| 64 | SessionCache *bool |
| 65 | // Extra carries kind-specific options; "vision" (bool) enables embedding |
| 66 | // attached Images as input_image parts on user turns. |
| 67 | Extra map[string]any |
| 68 | } |
| 69 | |
| 70 | func (c Config) mode() string { |
| 71 | mode := strings.ToLower(strings.TrimSpace(c.Mode)) |
| 72 | if mode == "stateful" || mode == "stateless" { |
| 73 | return mode |
| 74 | } |
| 75 | if c.Stateful != nil { |
| 76 | if *c.Stateful { |
| 77 | return "stateful" |
| 78 | } |
| 79 | return "stateless" |
| 80 | } |
| 81 | if capabilitiesFor(DetectVendor(c.BaseURL)).stateless { |
| 82 | return "stateless" |
| 83 | } |
| 84 | return "stateful" |
| 85 | } |
| 86 | |
| 87 | // DetectVendor lives in vendor.go (capabilities table): it covers dashscope/ |
| 88 | // deepseek (incl. eu.deepseek.com) / mimo via exact-host matching. |
| 89 | |
| 90 | type client struct { |
| 91 | identityHeaders http.Header |
| 92 | reasoning provider.ReasoningCapability |
| 93 | name string |
| 94 | identity provider.RequestIdentity |
| 95 | apiKey, keyEnv, keySource string |
| 96 | baseURL, requestURL, model, effort string |
| 97 | vendor, mode string |
| 98 | caps vendorCapabilities |
| 99 | sessionCache bool |
| 100 | search provider.SearchPolicy |
| 101 | maxOutputTokens int |
| 102 | vision bool // model accepts image input; embed Images as input_image parts |
| 103 | modelInfo provider.ModelInfo |
| 104 | http *http.Client |
| 105 | idleTimeout time.Duration |
| 106 | authed atomic.Bool |
| 107 | |
| 108 | mu sync.Mutex |
| 109 | lastResponseID string |
| 110 | expectedPrefixDigest string |
| 111 | } |
| 112 | |
| 113 | // New creates a Responses API provider. |
| 114 | func New(cfg Config) provider.Provider { |
| 115 | cfg.Extra = maps.Clone(cfg.Extra) |
| 116 | if cfg.Extra == nil { |
| 117 | cfg.Extra = map[string]any{} |
| 118 | } |
| 119 | if cfg.RequestURL != "" { |
| 120 | cfg.Extra["request_url"] = cfg.RequestURL |
| 121 | } |
| 122 | resolved := provider.ApplyOpenCodeGoContract("responses", provider.Config{BaseURL: cfg.BaseURL, Model: cfg.Model, Extra: cfg.Extra}) |
| 123 | cfg.Extra = resolved.Extra |
| 124 | vendor := DetectVendor(cfg.BaseURL) |
| 125 | cap := capabilitiesFor(vendor) |
| 126 | // Explicit replay contracts apply to compatible gateways as well as exact |
| 127 | // vendor hosts. Do not inherit endpoint defaults, headers, or output limits. |
| 128 | if protocol, _ := cfg.Extra["reasoning_protocol"].(string); strings.EqualFold(strings.TrimSpace(protocol), "deepseek") || strings.EqualFold(strings.TrimSpace(protocol), "mimo") { |
| 129 | cap.toolCallReasoning = true |
| 130 | } |
| 131 | maxOutputTokens := cfg.MaxOutputTokens |
| 132 | // Official DeepSeek omits max_output_tokens (server 384K). MiMo still uses |
| 133 | // the 16K/32K effort ladder. Compact_ratio is independent. |
| 134 | if maxOutputTokens == 0 && vendor == "mimo" { |
| 135 | maxOutputTokens = responsesAutoOutputBudget(vendor, cfg.Effort) |
| 136 | } else if maxOutputTokens == 0 && vendor != "deepseek" && cap.defaultMaxOutputTokens > 0 { |
| 137 | maxOutputTokens = cap.defaultMaxOutputTokens |
| 138 | } |
| 139 | sessionCache := cap.sessionCacheHeader |
| 140 | if cfg.SessionCache != nil { |
| 141 | sessionCache = *cfg.SessionCache |
| 142 | } |
| 143 | vision, _ := cfg.Extra["vision"].(bool) |
| 144 | if cfg.ModelInfo != nil { |
| 145 | vision = cfg.ModelInfo.SupportsInput(provider.ModalityImage) |
| 146 | } |
| 147 | // Official DeepSeek image input is pinned to one SKU. Ignore metadata or |
| 148 | // Extra["vision"] for Flash/Pro. |
| 149 | vision = openai.DeepSeekImageInputAllowed(vendor == "deepseek", cfg.RequestURL, cfg.Model, cfg.ModelInfo != nil, vision) |
| 150 | httpClient := &http.Client{} |
| 151 | if built, err := netclient.NewHTTPClient(cfg.Proxy, netclient.TransportOptions{ |
| 152 | DialTimeout: 30 * time.Second, KeepAlive: 30 * time.Second, |
| 153 | TLSHandshakeTimeout: 15 * time.Second, ResponseHeaderTimeout: 300 * time.Second, |
| 154 | }); err == nil { |
| 155 | httpClient = built |
| 156 | } |
| 157 | if cfg.HTTPClient != nil { |
| 158 | httpClient = cfg.HTTPClient |
| 159 | } |
| 160 | baseURL := strings.TrimRight(strings.TrimSpace(cfg.BaseURL), "/") |
| 161 | requestURL := strings.TrimSpace(cfg.RequestURL) |
| 162 | if requestURL == "" { |
| 163 | requestURL = baseURL + "/responses" |
| 164 | } |
| 165 | modelInfo := provider.ModelInfo{ID: cfg.Model, InputModalities: []provider.ModelModality{provider.ModalityText}} |
| 166 | if cfg.ModelInfo != nil { |
| 167 | modelInfo = *cfg.ModelInfo |
| 168 | modelInfo.ID = cfg.Model |
| 169 | } |
| 170 | clientWebSearch, _ := cfg.Extra["client_web_search"].(bool) |
| 171 | if reject, _ := cfg.Extra["reject_redirects"].(bool); reject { |
| 172 | httpClient.CheckRedirect = func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse } |
| 173 | } |
| 174 | if vision { |
| 175 | modelInfo.InputModalities = []provider.ModelModality{provider.ModalityText, provider.ModalityImage} |
| 176 | } else if modelInfo.SupportsInput(provider.ModalityImage) { |
| 177 | modelInfo.InputModalities = []provider.ModelModality{provider.ModalityText} |
| 178 | } |
| 179 | return &client{ |
| 180 | identityHeaders: provider.NewClientIdentityHeaders(), |
| 181 | name: cfg.Name, |
| 182 | identity: provider.RequestIdentity{Provider: cfg.Name, DisplayName: cfg.DisplayName, Protocol: cfg.Protocol}, |
| 183 | apiKey: cfg.APIKey, keyEnv: cfg.KeyEnv, keySource: cfg.KeySource, |
| 184 | reasoning: ReasoningForConfig(provider.Config{BaseURL: cfg.BaseURL, Model: cfg.Model, Extra: cfg.Extra}), |
| 185 | baseURL: baseURL, requestURL: requestURL, model: cfg.Model, effort: cfg.Effort, |
| 186 | vendor: vendor, caps: cap, mode: cfg.mode(), sessionCache: sessionCache, search: provider.SearchPolicy{NativeEnabled: cfg.WebSearch, ClientEnabled: clientWebSearch}, maxOutputTokens: maxOutputTokens, |
| 187 | vision: vision, |
| 188 | modelInfo: modelInfo, |
| 189 | http: httpClient, idleTimeout: defaultStreamIdleTimeout, |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | func (c *client) ModelInfo() provider.ModelInfo { |
| 194 | if c == nil { |
| 195 | return provider.ModelInfo{} |
| 196 | } |
| 197 | info := c.modelInfo |
| 198 | info.InputModalities = append([]provider.ModelModality(nil), info.InputModalities...) |
| 199 | return info |
| 200 | } |
| 201 | |
| 202 | func responsesReasoningDisabled(effort string) bool { |
| 203 | switch strings.ToLower(strings.TrimSpace(effort)) { |
| 204 | case "none", "disabled", "off": |
| 205 | return true |
| 206 | default: |
| 207 | return false |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | // responsesAutoOutputBudget is the MiMo (and similar) 16K/32K ladder. |
| 212 | // Official DeepSeek must not call this; it omits max_output_tokens instead. |
| 213 | func responsesAutoOutputBudget(vendor, effort string) int { |
| 214 | if responsesReasoningDisabled(effort) { |
| 215 | return provider.AutoOutputBudget(false, effort) |
| 216 | } |
| 217 | e := strings.ToLower(strings.TrimSpace(effort)) |
| 218 | if vendor == "deepseek" && (e == "" || e == "auto") { |
| 219 | e = "high" |
| 220 | } |
| 221 | return provider.AutoOutputBudget(true, e) |
| 222 | } |
| 223 | |
| 224 | func (c *client) Name() string { return c.name } |
| 225 | |
| 226 | func (c *client) NativeToolSearchAvailable() bool { |
| 227 | return c != nil && provider.IsFirstPartyOpenAI(c.baseURL) && nativeToolSearchModel(c.model) |
| 228 | } |
| 229 | |
| 230 | func nativeToolSearchModel(model string) bool { |
| 231 | model = strings.ToLower(strings.TrimSpace(model)) |
| 232 | return strings.HasPrefix(model, "gpt-5.4") || strings.HasPrefix(model, "gpt-5.5") || strings.HasPrefix(model, "gpt-5.6") |
| 233 | } |
| 234 | |
| 235 | func (c *client) sendOpts() provider.SendOptions { |
| 236 | return provider.SendOptions{Provider: c.name, ProviderDisplayName: c.identity.DisplayName, Protocol: c.identity.Protocol, KeyEnv: c.keyEnv, KeySource: c.keySource, KeyPresent: c.apiKey != "", RetryAuth: c.authed.Load()} |
| 237 | } |
| 238 | |
| 239 | // ResetContext drops stateful continuation metadata. Full-input stateless mode |
| 240 | // is unaffected. |
| 241 | func (c *client) ResetContext() { |
| 242 | c.mu.Lock() |
| 243 | c.lastResponseID = "" |
| 244 | c.expectedPrefixDigest = "" |
| 245 | c.mu.Unlock() |
| 246 | } |
| 247 | |
| 248 | func (c *client) Stream(ctx context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 249 | if c.effort != "auto" && c.effort != "off" { |
| 250 | if err := c.reasoning.Validate(c.model, c.effort); err != nil { |
| 251 | return nil, err |
| 252 | } |
| 253 | } |
| 254 | if err := c.reasoning.Validate(c.model, req.EffortOverride); err != nil { |
| 255 | return nil, err |
| 256 | } |
| 257 | requestCtx := provider.WithRequestAttemptCounter(ctx) |
| 258 | body, usedPrevious, wireMessages := c.buildRequestBody(req) |
| 259 | resp, err := c.send(requestCtx, body) |
| 260 | if err != nil && usedPrevious && isStalePreviousResponseError(err) { |
| 261 | // A stateful response ID may expire server-side. Retrying once with full |
| 262 | // history is safe because no response body has started streaming. |
| 263 | c.ResetContext() |
| 264 | body, _, wireMessages = c.buildRequestBody(req) |
| 265 | resp, err = c.send(requestCtx, body) |
| 266 | } |
| 267 | if err != nil { |
| 268 | return nil, err |
| 269 | } |
| 270 | c.authed.Store(true) |
| 271 | out := make(chan provider.Chunk, 64) |
| 272 | go c.readStream(requestCtx, resp, out, wireMessages) |
| 273 | return out, nil |
| 274 | } |
| 275 | |
| 276 | func (c *client) send(ctx context.Context, body map[string]any) (*http.Response, error) { |
| 277 | payload, err := json.Marshal(body) |
| 278 | if err != nil { |
| 279 | return nil, fmt.Errorf("responses: marshal request: %w", err) |
| 280 | } |
| 281 | newRequest := func(ctx context.Context) (*http.Request, error) { |
| 282 | req, err := http.NewRequestWithContext(ctx, http.MethodPost, c.requestURL, bytes.NewReader(payload)) |
| 283 | if err != nil { |
| 284 | return nil, err |
| 285 | } |
| 286 | req.Header.Set("Content-Type", "application/json") |
| 287 | req.Header.Set("Authorization", "Bearer "+c.apiKey) |
| 288 | provider.ApplyOpenCodeGoHeaders(req, c.baseURL, c.identityHeaders) |
| 289 | if c.caps.sessionCacheHeader && c.sessionCache { |
| 290 | req.Header.Set("x-dashscope-session-cache", "enable") |
| 291 | } |
| 292 | return req, nil |
| 293 | } |
| 294 | return provider.SendWithRetry(ctx, c.http, c.sendOpts(), newRequest) |
| 295 | } |
| 296 | |
| 297 | func isStalePreviousResponseError(err error) bool { |
| 298 | var apiErr *provider.APIError |
| 299 | if !errors.As(err, &apiErr) || apiErr.Status != http.StatusBadRequest { |
| 300 | return false |
| 301 | } |
| 302 | body := strings.ToLower(apiErr.Body) |
| 303 | mentionsID := strings.Contains(body, "previous_response_id") || strings.Contains(body, "previous response") || strings.Contains(body, "response id") |
| 304 | return mentionsID && |
| 305 | (strings.Contains(body, "not found") || strings.Contains(body, "invalid") || strings.Contains(body, "expired")) |
| 306 | } |
| 307 | |
| 308 | func (c *client) buildRequestBody(req provider.Request) (map[string]any, bool, []provider.Message) { |
| 309 | messages := provider.SanitizeToolPairing(provider.ModelMessages(req.Messages)) |
| 310 | body := map[string]any{"model": c.model, "stream": true} |
| 311 | |
| 312 | effort := c.effort |
| 313 | if req.EffortOverride != "" { |
| 314 | effort = req.EffortOverride |
| 315 | } |
| 316 | |
| 317 | switch effort { |
| 318 | case "auto": |
| 319 | effort = "" |
| 320 | case "disabled", "off": |
| 321 | effort = "none" |
| 322 | } |
| 323 | if effort != "" { |
| 324 | body["reasoning"] = map[string]any{"effort": effort} |
| 325 | } |
| 326 | maxOutputTokens := req.MaxTokens |
| 327 | if maxOutputTokens == 0 { |
| 328 | maxOutputTokens = c.maxOutputTokens |
| 329 | } |
| 330 | if maxOutputTokens == 0 && c.vendor == "mimo" { |
| 331 | maxOutputTokens = responsesAutoOutputBudget(c.vendor, c.effort) |
| 332 | } else if maxOutputTokens == 0 && c.vendor != "deepseek" && c.caps.defaultMaxOutputTokens > 0 { |
| 333 | maxOutputTokens = c.caps.defaultMaxOutputTokens |
| 334 | } |
| 335 | if maxOutputTokens > 0 { |
| 336 | body["max_output_tokens"] = maxOutputTokens |
| 337 | } |
| 338 | if req.ResponseFormat != nil && req.ResponseFormat.Type != "" { |
| 339 | // Structured output: Responses text.format. MiMo/DashScope/OpenAI |
| 340 | // all accept {"text":{"format":{"type":"json_object"}}}. The model |
| 341 | // only emits JSON when the instructions also demand it. |
| 342 | body["text"] = map[string]any{ |
| 343 | "format": map[string]any{"type": req.ResponseFormat.Type}, |
| 344 | } |
| 345 | } |
| 346 | if req.Temperature != nil && !c.caps.ignoresTemperature { |
| 347 | body["temperature"] = *req.Temperature |
| 348 | } |
| 349 | if c.search.NativeEnabled || len(req.Tools) > 0 { |
| 350 | body["tools"] = encodeResponsesTools(c, req) |
| 351 | } |
| 352 | instructions, rest := splitInstructions(messages) |
| 353 | if instructions != "" { |
| 354 | body["instructions"] = instructions |
| 355 | } |
| 356 | |
| 357 | c.mu.Lock() |
| 358 | previousID, expectedDigest := c.lastResponseID, c.expectedPrefixDigest |
| 359 | c.mu.Unlock() |
| 360 | if c.canUseStatefulContinuation(messages, previousID, expectedDigest) { |
| 361 | body["input"] = messages[len(messages)-1].Content |
| 362 | body["previous_response_id"] = previousID |
| 363 | return body, true, messages |
| 364 | } |
| 365 | |
| 366 | body["input"] = messagesToInput(rest, c.vision, c.search.NativeEnabled, c.caps.summaryRequired) |
| 367 | return body, false, messages |
| 368 | } |
| 369 | |
| 370 | func inputImagePart(ref string) map[string]string { |
| 371 | switch provider.ClassifyImage(ref) { |
| 372 | case provider.ImageFileID: |
| 373 | return map[string]string{"type": "input_image", "file_id": ref} |
| 374 | case provider.ImageDataURL, provider.ImageHTTPURL: |
| 375 | return map[string]string{"type": "input_image", "image_url": ref} |
| 376 | default: |
| 377 | return nil |
| 378 | } |
| 379 | } |
| 380 | |
| 381 | func splitInstructions(messages []provider.Message) (string, []provider.Message) { |
| 382 | if len(messages) == 0 || messages[0].Role != provider.RoleSystem { |
| 383 | return "", messages |
| 384 | } |
| 385 | return messages[0].Content, messages[1:] |
| 386 | } |
| 387 | |
| 388 | func decodeReplayableWebSearchItem(raw json.RawMessage) (map[string]any, bool) { |
| 389 | if len(raw) == 0 || len(raw) > maxReplayableSearchItemBytes || !json.Valid(raw) { |
| 390 | return nil, false |
| 391 | } |
| 392 | var item map[string]any |
| 393 | if err := json.Unmarshal(raw, &item); err != nil || item["type"] != "web_search_call" { |
| 394 | return nil, false |
| 395 | } |
| 396 | id, _ := item["id"].(string) |
| 397 | status, _ := item["status"].(string) |
| 398 | if strings.TrimSpace(id) == "" || status != "completed" { |
| 399 | return nil, false |
| 400 | } |
| 401 | return item, true |
| 402 | } |
| 403 | |
| 404 | func (c *client) conversationDigest(messages []provider.Message) string { |
| 405 | instructions, rest := splitInstructions(messages) |
| 406 | // Digest must mirror the wire exactly: the stateful fast path compares |
| 407 | // this against the previous request's input, so a mismatch would skip |
| 408 | // previous_response_id and force a full replay (cache-hit loss). Use the |
| 409 | // same vision/summary knobs as buildRequestBody. |
| 410 | payload, _ := json.Marshal(struct { |
| 411 | Instructions string `json:"instructions,omitempty"` |
| 412 | Input []map[string]any `json:"input"` |
| 413 | }{Instructions: instructions, Input: messagesToInput(rest, c.vision, c.search.NativeEnabled, c.caps.summaryRequired)}) |
| 414 | sum := sha256.Sum256(payload) |
| 415 | return hex.EncodeToString(sum[:]) |
| 416 | } |
| 417 | |
| 418 | type streamedCall struct { |
| 419 | id, name, arguments string |
| 420 | argChars int |
| 421 | completed bool |
| 422 | } |
| 423 | |
| 424 | func (c *client) readStream(ctx context.Context, resp *http.Response, out chan<- provider.Chunk, requestMessages []provider.Message) { |
| 425 | defer resp.Body.Close() |
| 426 | defer close(out) |
| 427 | |
| 428 | scanner := bufio.NewScanner(resp.Body) |
| 429 | scanner.Buffer(make([]byte, 64*1024), 4*1024*1024) |
| 430 | idle := c.idleTimeout |
| 431 | if idle <= 0 { |
| 432 | idle = defaultStreamIdleTimeout |
| 433 | } |
| 434 | watchDone := make(chan struct{}) |
| 435 | activity := make(chan struct{}, 1) |
| 436 | var reasoningSnapshots responseReasoningSnapshots |
| 437 | var stalled atomic.Bool |
| 438 | go func() { |
| 439 | timer := time.NewTimer(idle) |
| 440 | defer timer.Stop() |
| 441 | for { |
| 442 | select { |
| 443 | case <-ctx.Done(): |
| 444 | _ = resp.Body.Close() |
| 445 | return |
| 446 | case <-watchDone: |
| 447 | return |
| 448 | case <-activity: |
| 449 | if !timer.Stop() { |
| 450 | select { |
| 451 | case <-timer.C: |
| 452 | default: |
| 453 | } |
| 454 | } |
| 455 | timer.Reset(idle) |
| 456 | case <-timer.C: |
| 457 | stalled.Store(true) |
| 458 | _ = resp.Body.Close() |
| 459 | return |
| 460 | } |
| 461 | } |
| 462 | }() |
| 463 | defer close(watchDone) |
| 464 | |
| 465 | calls := make(map[string]*streamedCall) |
| 466 | callOrder := make([]string, 0) |
| 467 | callForItem := func(itemID string) *streamedCall { |
| 468 | if call := calls[itemID]; call != nil { |
| 469 | return call |
| 470 | } |
| 471 | call := &streamedCall{id: itemID} |
| 472 | calls[itemID] = call |
| 473 | callOrder = append(callOrder, itemID) |
| 474 | return call |
| 475 | } |
| 476 | textDeltas := make(map[string]bool) |
| 477 | reasoningDeltas := make(map[string]bool) |
| 478 | seenSearchItems := make(map[string]struct{}) |
| 479 | var responsesItems []json.RawMessage |
| 480 | var text, reasoning strings.Builder |
| 481 | reasoningID := "" |
| 482 | reasoningStatus := "" |
| 483 | terminal := false |
| 484 | failed := false |
| 485 | completedResponseID := "" |
| 486 | |
| 487 | for scanner.Scan() { |
| 488 | select { |
| 489 | case activity <- struct{}{}: |
| 490 | default: |
| 491 | } |
| 492 | line := scanner.Text() |
| 493 | if !strings.HasPrefix(line, "data:") { |
| 494 | continue |
| 495 | } |
| 496 | data := strings.TrimSpace(strings.TrimPrefix(line, "data:")) |
| 497 | if data == "[DONE]" { |
| 498 | terminal = true |
| 499 | break |
| 500 | } |
| 501 | var event sseEvent |
| 502 | if json.Unmarshal([]byte(data), &event) != nil { |
| 503 | continue |
| 504 | } |
| 505 | key := fmt.Sprintf("%s:%d", event.ItemID, event.ContentIndex) |
| 506 | reasoningSnapshots.capture(event) |
| 507 | switch event.Type { |
| 508 | case "response.output_text.delta": |
| 509 | textDeltas[key] = true |
| 510 | text.WriteString(event.Delta) |
| 511 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkText, Text: event.Delta}) { |
| 512 | return |
| 513 | } |
| 514 | case "response.output_text.done": |
| 515 | if event.Text != "" && !textDeltas[key] { |
| 516 | text.WriteString(event.Text) |
| 517 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkText, Text: event.Text}) { |
| 518 | return |
| 519 | } |
| 520 | } |
| 521 | case "response.reasoning_text.delta", "response.reasoning_summary_text.delta": |
| 522 | reasoningDeltas[key] = true |
| 523 | reasoning.WriteString(event.Delta) |
| 524 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, Text: event.Delta}) { |
| 525 | return |
| 526 | } |
| 527 | case "response.reasoning_text.done", "response.reasoning_summary_text.done": |
| 528 | if event.Text != "" && !reasoningDeltas[key] { |
| 529 | reasoning.WriteString(event.Text) |
| 530 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, Text: event.Text}) { |
| 531 | return |
| 532 | } |
| 533 | } |
| 534 | case "response.output_item.added": |
| 535 | if event.Item != nil { |
| 536 | switch event.Item.Type { |
| 537 | case "function_call": |
| 538 | call := callForItem(event.Item.ID) |
| 539 | call.id = event.Item.CallID |
| 540 | call.name = event.Item.Name |
| 541 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCallStart, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name}}) { |
| 542 | return |
| 543 | } |
| 544 | case "reasoning": |
| 545 | // Capture the provider-issued reasoning item id so the |
| 546 | // next turn's input reasoning item can carry it (the |
| 547 | // OpenAI Responses schema marks Reasoning.id required). |
| 548 | if event.Item.ID != "" { |
| 549 | // 多段推理（DeepSeek 长思考分多段）时末段 id 覆盖：round-trip |
| 550 | // 合并为一个 reasoning item 只带末段 id（服务端接受）。 |
| 551 | reasoningID = event.Item.ID |
| 552 | } |
| 553 | } |
| 554 | } |
| 555 | case "response.function_call_arguments.delta": |
| 556 | call := callForItem(event.ItemID) |
| 557 | call.arguments += event.Delta |
| 558 | call.argChars += len(event.Delta) |
| 559 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCallArgsDelta, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name}, ArgChars: call.argChars}) { |
| 560 | return |
| 561 | } |
| 562 | case "response.function_call_arguments.done": |
| 563 | call := callForItem(event.ItemID) |
| 564 | if event.Arguments != "" { |
| 565 | call.arguments = event.Arguments |
| 566 | } |
| 567 | if !call.completed { |
| 568 | call.completed = true |
| 569 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}}) { |
| 570 | return |
| 571 | } |
| 572 | } |
| 573 | case "response.output_item.done": |
| 574 | if event.Item != nil && event.Item.Type == "web_search_call" && c.search.NativeEnabled { |
| 575 | if _, ok := decodeReplayableWebSearchItem(event.Item.Raw); ok { |
| 576 | key := event.Item.ID |
| 577 | if key == "" { |
| 578 | key = string(event.Item.Raw) |
| 579 | } |
| 580 | if _, seen := seenSearchItems[key]; !seen { |
| 581 | seenSearchItems[key] = struct{}{} |
| 582 | raw := append(json.RawMessage(nil), event.Item.Raw...) |
| 583 | responsesItems = append(responsesItems, raw) |
| 584 | if !emitSearchReplay(ctx, out, raw) { |
| 585 | return |
| 586 | } |
| 587 | } |
| 588 | } |
| 589 | } |
| 590 | if event.Item != nil { |
| 591 | switch event.Item.Type { |
| 592 | case "function_call": |
| 593 | call := callForItem(event.Item.ID) |
| 594 | if event.Item.CallID != "" { |
| 595 | call.id = event.Item.CallID |
| 596 | } |
| 597 | if event.Item.Name != "" { |
| 598 | call.name = event.Item.Name |
| 599 | } |
| 600 | if event.Item.Arguments != "" { |
| 601 | call.arguments = event.Item.Arguments |
| 602 | } |
| 603 | if !call.completed { |
| 604 | call.completed = true |
| 605 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}}) { |
| 606 | return |
| 607 | } |
| 608 | } |
| 609 | case "reasoning": |
| 610 | // The done event carries the final item status |
| 611 | // ("completed" after the thinking stream finishes); |
| 612 | // round-trip it with the reasoning item so the input |
| 613 | // matches the wire schema. |
| 614 | if event.Item.Status != "" { |
| 615 | reasoningStatus = event.Item.Status |
| 616 | } |
| 617 | } |
| 618 | } |
| 619 | case "response.completed", "response.incomplete", "response.failed": |
| 620 | terminal = true |
| 621 | if event.Type == "response.incomplete" { |
| 622 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, ReasoningState: provider.ReasoningIncomplete}) { |
| 623 | return |
| 624 | } |
| 625 | } |
| 626 | completedResponseID = terminalResponseID(event) |
| 627 | if !emitTerminalResponseUsage(ctx, out, event) { |
| 628 | return |
| 629 | } |
| 630 | if event.Type == "response.failed" { |
| 631 | failed = true |
| 632 | err := fmt.Errorf("responses: response failed") |
| 633 | if event.Response != nil && event.Response.Error != nil { |
| 634 | if authErr := authErrorFromResponse(c, event.Response.Error); authErr != nil { |
| 635 | err = authErr |
| 636 | } else { |
| 637 | err = fmt.Errorf("responses: %s", event.Response.Error.Message) |
| 638 | } |
| 639 | } |
| 640 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: err}) { |
| 641 | return |
| 642 | } |
| 643 | } |
| 644 | } |
| 645 | if terminal { |
| 646 | break |
| 647 | } |
| 648 | } |
| 649 | |
| 650 | if ctx.Err() != nil { |
| 651 | return |
| 652 | } |
| 653 | if err := scanner.Err(); err != nil { |
| 654 | var reason string |
| 655 | if stalled.Load() { |
| 656 | err = fmt.Errorf("responses: stream idle timeout after %s", idle) |
| 657 | reason = provider.StreamInterruptIdleTimeout |
| 658 | } else { |
| 659 | reason = provider.ClassifyStreamInterrupt(err) |
| 660 | } |
| 661 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: provider.StreamInterrupt(err, reason)}) |
| 662 | return |
| 663 | } |
| 664 | // Protocol-defined terminal response events are required. Connection close |
| 665 | // before a terminal event leaves the attempt uncommitted — including any |
| 666 | // complete tool calls already forwarded as speculative output. |
| 667 | if !terminal { |
| 668 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkError, Err: provider.StreamInterrupt(io.ErrUnexpectedEOF, provider.StreamInterruptPrematureEOF)}) |
| 669 | return |
| 670 | } |
| 671 | if !reasoningSnapshots.emit(ctx, out) { |
| 672 | return |
| 673 | } |
| 674 | responsesItems = append(responsesItems, reasoningSnapshots.items...) |
| 675 | if len(reasoningSnapshots.items) > 0 { |
| 676 | reasoningID, reasoningStatus = reasoningSnapshots.metadata() |
| 677 | } |
| 678 | if completedResponseID != "" { |
| 679 | assistant := provider.Message{Role: provider.RoleAssistant, Content: text.String(), ReasoningContent: reasoning.String(), ReasoningID: reasoningID, ReasoningStatus: reasoningStatus, ResponsesItems: responsesItems} |
| 680 | for _, itemID := range callOrder { |
| 681 | call := calls[itemID] |
| 682 | if call.completed { |
| 683 | assistant.ToolCalls = append(assistant.ToolCalls, provider.ToolCall{ID: call.id, Name: call.name, Arguments: call.arguments}) |
| 684 | } |
| 685 | } |
| 686 | expected := append(append([]provider.Message(nil), requestMessages...), assistant) |
| 687 | c.mu.Lock() |
| 688 | c.lastResponseID = completedResponseID |
| 689 | c.expectedPrefixDigest = c.conversationDigest(expected) |
| 690 | c.mu.Unlock() |
| 691 | } else { |
| 692 | c.ResetContext() |
| 693 | } |
| 694 | if !failed { |
| 695 | // 把 reasoning item 的 id/status 作为元数据 chunk 流给 Agent |
| 696 | // （空 Text，随 ChunkReasoning 语义）——Agent 持久化进 session， |
| 697 | // 下一轮 input reasoning item 回传 id/status（评审 #7234 第 1 点）。 |
| 698 | if reasoningID != "" || reasoningStatus != "" { |
| 699 | if !sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkReasoning, ReasoningID: reasoningID, ReasoningStatus: reasoningStatus}) { |
| 700 | return |
| 701 | } |
| 702 | } |
| 703 | _ = sendChunk(ctx, out, provider.Chunk{Type: provider.ChunkDone}) |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | func sendChunk(ctx context.Context, out chan<- provider.Chunk, chunk provider.Chunk) bool { |
| 708 | select { |
| 709 | case out <- chunk: |
| 710 | return true |
| 711 | default: |
| 712 | } |
| 713 | notifySendChunkEnterBlocking() |
| 714 | select { |
| 715 | case out <- chunk: |
| 716 | return true |
| 717 | case <-ctx.Done(): |
| 718 | return false |
| 719 | } |
| 720 | } |
| 721 | |
| 722 | func usageFromResponse(response *sseResponse) *provider.Usage { |
| 723 | usage := &provider.Usage{} |
| 724 | if response == nil || response.Usage == nil { |
| 725 | return usage |
| 726 | } |
| 727 | u := response.Usage |
| 728 | cached, reasoning := 0, 0 |
| 729 | if u.InputTokensDetails != nil { |
| 730 | cached = u.InputTokensDetails.CachedTokens |
| 731 | } |
| 732 | if u.OutputTokensDetails != nil { |
| 733 | reasoning = u.OutputTokensDetails.ReasoningTokens |
| 734 | } |
| 735 | miss := max(u.InputTokens-cached, 0) |
| 736 | total := u.TotalTokens |
| 737 | if total == 0 { |
| 738 | total = u.InputTokens + u.OutputTokens |
| 739 | } |
| 740 | return &provider.Usage{PromptTokens: u.InputTokens, CompletionTokens: u.OutputTokens, TotalTokens: total, CacheHitTokens: cached, CacheMissTokens: miss, ReasoningTokens: reasoning} |
| 741 | } |
| 742 | |
| 743 | func authErrorFromResponse(c *client, responseError *sseError) error { |
| 744 | if responseError == nil { |
| 745 | return nil |
| 746 | } |
| 747 | value := strings.ToLower(responseError.Code + " " + responseError.Message) |
| 748 | if !strings.Contains(value, "auth") && !strings.Contains(value, "api key") && !strings.Contains(value, "unauthorized") && !strings.Contains(value, "forbidden") && !strings.Contains(value, "permission") { |
| 749 | return nil |
| 750 | } |
| 751 | status := http.StatusUnauthorized |
| 752 | if strings.Contains(value, "forbidden") || strings.Contains(value, "permission") { |
| 753 | status = http.StatusForbidden |
| 754 | } |
| 755 | return &provider.AuthError{Provider: c.name, ProviderDisplayName: c.identity.DisplayName, Protocol: c.identity.Protocol, KeyEnv: c.keyEnv, KeySource: c.keySource, Status: status, HasKey: c.apiKey != "", Body: responseError.Message} |
| 756 | } |
| 757 | |
| 758 | type sseEvent struct { |
| 759 | Type string `json:"type"` |
| 760 | Delta string `json:"delta"` |
| 761 | Text string `json:"text"` |
| 762 | Arguments string `json:"arguments"` |
| 763 | ItemID string `json:"item_id"` |
| 764 | ContentIndex int `json:"content_index"` |
| 765 | Item *sseItem `json:"item"` |
| 766 | Response *sseResponse `json:"response"` |
| 767 | } |
| 768 | |
| 769 | type sseItem struct { |
| 770 | ID, Type, CallID, Name, Arguments, Status string |
| 771 | Raw json.RawMessage |
| 772 | } |
| 773 | |
| 774 | func (i *sseItem) UnmarshalJSON(data []byte) error { |
| 775 | var wire struct { |
| 776 | ID string `json:"id"` |
| 777 | Type string `json:"type"` |
| 778 | CallID string `json:"call_id"` |
| 779 | Name string `json:"name"` |
| 780 | Arguments string `json:"arguments"` |
| 781 | Status string `json:"status"` |
| 782 | } |
| 783 | if err := json.Unmarshal(data, &wire); err != nil { |
| 784 | return err |
| 785 | } |
| 786 | *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...)} |
| 787 | return nil |
| 788 | } |
| 789 | |
| 790 | type sseResponse struct { |
| 791 | Output []sseItem `json:"output"` |
| 792 | ID string `json:"id"` |
| 793 | Usage *sseUsage `json:"usage"` |
| 794 | Error *sseError `json:"error"` |
| 795 | IncompleteDetails incompleteDetails `json:"incomplete_details"` |
| 796 | } |
| 797 | |
| 798 | type incompleteDetails struct { |
| 799 | Reason string `json:"reason"` |
| 800 | } |
| 801 | type sseError struct { |
| 802 | Message string `json:"message"` |
| 803 | Code string `json:"code"` |
| 804 | } |
| 805 | type sseUsage struct { |
| 806 | InputTokens int `json:"input_tokens"` |
| 807 | OutputTokens int `json:"output_tokens"` |
| 808 | TotalTokens int `json:"total_tokens"` |
| 809 | InputTokensDetails *struct { |
| 810 | CachedTokens int `json:"cached_tokens"` |
| 811 | } `json:"input_tokens_details"` |
| 812 | OutputTokensDetails *struct { |
| 813 | ReasoningTokens int `json:"reasoning_tokens"` |
| 814 | } `json:"output_tokens_details"` |
| 815 | } |
| 816 |