1package provider
2
3import (
4 "bytes"
5 "context"
6 "encoding/base64"
7 "encoding/json"
8 "fmt"
9 "io"
10 "net/http"
11 "regexp"
12 "strings"
13 "time"
14)
15
16const openaiDefaultBaseURL = "https://api.openai.com"
17
18// versionSegmentSuffix matches a trailing API version segment such as
19// "/v1" or Z.AI's "/v4".
20var versionSegmentSuffix = regexp.MustCompile(`/v\d+$`)
21
22// chatCompletionsURL builds the chat-completions endpoint for an
23// OpenAI-compatible base URL. A base that already carries an API
24// version segment gets "/chat/completions" appended directly; a bare
25// host (e.g. api.openai.com) gets the conventional "/v1/chat/completions".
26//
27// Matching any "/vN" segment (not just "/v1") keeps Z.AI's coding-plan
28// base, which ends in "/paas/v4", from getting a spurious "/v1" that
29// yields ".../paas/v4/v1/chat/completions" and a 404.
30func chatCompletionsURL(baseURL string) string {
31 if versionSegmentSuffix.MatchString(baseURL) {
32 return baseURL + "/chat/completions"
33 }
34 return baseURL + "/v1/chat/completions"
35}
36
37type openaiClient struct {
38 apiKey string
39 baseURL string
40 chatCompletionsPath string
41 name string
42 oauth bool // when true, apiKey actually holds an OAuth access token
43 headers map[string]string
44 http *http.Client
45}
46
47// NewOpenAI creates an OpenAI client using an API key. baseURL may be empty.
48func NewOpenAI(apiKey, baseURL string) Client {
49 if baseURL == "" {
50 baseURL = openaiDefaultBaseURL
51 }
52 return &openaiClient{
53 apiKey: apiKey,
54 baseURL: strings.TrimRight(baseURL, "/"),
55 name: "openai",
56 http: &http.Client{Timeout: 0},
57 }
58}
59
60// NewOpenAIOAuth creates an OpenAI client using a subscription OAuth access token.
61// The token is sent as an HTTP Bearer credential on the standard chat/completions endpoint.
62func NewOpenAIOAuth(accessToken, baseURL string) Client {
63 if baseURL == "" {
64 baseURL = openaiDefaultBaseURL
65 }
66 return &openaiClient{
67 apiKey: accessToken,
68 baseURL: strings.TrimRight(baseURL, "/"),
69 name: "openai",
70 oauth: true,
71 http: &http.Client{Timeout: 0},
72 }
73}
74
75// NewKimi creates a Kimi/Moonshot client. Kimi's chat API is OpenAI-compatible.
76func NewKimi(apiKey, baseURL string) Client {
77 return NewKimiWithHeaders(apiKey, baseURL, nil)
78}
79
80// NewDeepSeek creates a DeepSeek client. DeepSeek's chat API is
81// OpenAI-compatible at https://api.deepseek.com/v1.
82func NewDeepSeek(apiKey, baseURL string) Client {
83 if baseURL == "" {
84 baseURL = "https://api.deepseek.com/v1"
85 }
86 return &openaiClient{
87 apiKey: apiKey,
88 baseURL: strings.TrimRight(baseURL, "/"),
89 name: "deepseek",
90 http: &http.Client{Timeout: 0},
91 }
92}
93
94// NewKimiWithHeaders creates a Kimi/Moonshot client with extra headers.
95// Subscription tokens from Kimi Code need the official CLI's X-Msh-* headers.
96func NewKimiWithHeaders(apiKey, baseURL string, headers map[string]string) Client {
97 if baseURL == "" {
98 baseURL = "https://api.kimi.com/coding/v1"
99 }
100 return &openaiClient{
101 apiKey: apiKey,
102 baseURL: strings.TrimRight(baseURL, "/"),
103 name: "kimi",
104 headers: headers,
105 http: &http.Client{Timeout: 0},
106 }
107}
108
109func (c *openaiClient) Name() string {
110 if c.name != "" {
111 return c.name
112 }
113 return "openai"
114}
115
116// ---- wire types ----
117
118type oaiContentText struct {
119 Type string `json:"type"` // "text"
120 Text string `json:"text"`
121}
122
123type oaiContentImage struct {
124 Type string `json:"type"` // "image_url"
125 ImageURL struct {
126 URL string `json:"url"`
127 } `json:"image_url"`
128}
129
130type oaiToolCallFn struct {
131 Name string `json:"name"`
132 Arguments string `json:"arguments"` // JSON string
133}
134
135type oaiToolCall struct {
136 ID string `json:"id"`
137 Type string `json:"type"` // "function" or "openrouter:..."
138 Function oaiToolCallFn `json:"function,omitempty"`
139 Name string `json:"name,omitempty"`
140 Arguments json.RawMessage `json:"arguments,omitempty"`
141}
142
143type oaiMessage struct {
144 Role string `json:"role"`
145 Content interface{} `json:"content,omitempty"` // string or []block
146 Name string `json:"name,omitempty"`
147 ToolCalls []oaiToolCall `json:"tool_calls,omitempty"`
148 ToolCallID string `json:"tool_call_id,omitempty"`
149 Tools []oaiTool `json:"tools,omitempty"`
150 // ReasoningContent carries the model's chain-of-thought summary
151 // alongside an assistant tool-call message. Required by Kimi's
152 // chat completions endpoint when thinking is enabled and the
153 // assistant message contains a tool call; OpenAI ignores it.
154 ReasoningContent string `json:"reasoning_content,omitempty"`
155}
156
157type oaiTool struct {
158 Type string `json:"type"` // "function" or "openrouter:..."
159 Parameters json.RawMessage `json:"parameters,omitempty"`
160 Function *struct {
161 Name string `json:"name"`
162 Description string `json:"description,omitempty"`
163 Parameters json.RawMessage `json:"parameters"`
164 } `json:"function,omitempty"`
165}
166
167type oaiStreamOptions struct {
168 IncludeUsage bool `json:"include_usage"`
169}
170
171type oaiRequest struct {
172 Model string `json:"model"`
173 Messages []oaiMessage `json:"messages"`
174 Tools []oaiTool `json:"tools,omitempty"`
175 ToolChoice string `json:"tool_choice,omitempty"`
176 Stream bool `json:"stream"`
177 StreamOptions *oaiStreamOptions `json:"stream_options,omitempty"`
178 Temperature *float32 `json:"temperature,omitempty"`
179 MaxTokens *int `json:"max_tokens,omitempty"`
180 MaxCompletionTok *int `json:"max_completion_tokens,omitempty"`
181 ReasoningEffort string `json:"reasoning_effort,omitempty"`
182 SessionID string `json:"session_id,omitempty"`
183 MaxToolCalls int `json:"max_tool_calls,omitempty"`
184}
185
186// ---- request building ----
187
188func (c *openaiClient) buildRequest(req Request) (*oaiRequest, error) {
189 // The OpenAI wire client is shared by many providers, so prefer its own
190 // provider namespace before falling back to a provider-agnostic lookup.
191 m, err := FindModel(c.Name(), req.Model)
192 if err != nil {
193 m, err = FindModel("", req.Model)
194 }
195 if err != nil {
196 // Unknown model: use sensible defaults so local/custom
197 // models still work without a catalog entry.
198 m = Model{
199 ID: req.Model,
200 ContextWindow: 32768,
201 MaxOutput: 8192,
202 }
203 }
204 reasoning := ClampReasoningForModel(m, req.Reasoning)
205 out := &oaiRequest{
206 Model: req.Model,
207 Stream: true,
208 StreamOptions: &oaiStreamOptions{IncludeUsage: true},
209 Temperature: req.Temperature,
210 }
211 if c.Name() == "openrouter" {
212 if req.SessionID != "" {
213 out.SessionID = req.SessionID
214 }
215 if req.MaxToolCalls > 0 {
216 out.MaxToolCalls = req.MaxToolCalls
217 }
218 }
219
220 maxTok := req.MaxTokens
221 if maxTok <= 0 {
222 maxTok = m.MaxOutput
223 }
224 // Clamp max_tokens so output plus a minimum input reservation fits
225 // within the context window. Some providers (OpenRouter) enforce
226 // input + max_output <= context_length and reject requests where the
227 // total exceeds it. Reserving headroom guarantees the system prompt,
228 // first message, and tool definitions have room.
229 //
230 // The cap is derived from the context window, never from MaxOutput:
231 // MaxOutput is already the output ceiling, so subtracting from it
232 // would shrink every model's budget even when its output comfortably
233 // fits the window. We only lower maxTok when ContextWindow - reserve
234 // is actually tighter than the requested budget, which is exactly the
235 // pathological case (e.g. a model whose MaxOutput equals its window).
236 //
237 // The reserve is proportional (window/8, capped at 4096) rather than a
238 // flat 4096 so small-window models aren't over-penalized: a flat 4096
239 // would halve gpt-4's 8192 budget, while window/8 reserves a sensible
240 // 1024 there and still tops out at 4096 for large contexts.
241 //
242 // Some providers (OpenRouter) report inflated model-level context
243 // windows (e.g. 1000000) while the serving provider enforces a much
244 // tighter limit (e.g. 262144). Discovery already prefers the serving
245 // provider's smaller context_length, so m.ContextWindow is the real
246 // limit by the time we get here.
247 if m.ContextWindow > 0 {
248 reserve := m.ContextWindow / 8
249 const maxReserve = 4096
250 if reserve > maxReserve {
251 reserve = maxReserve
252 }
253 clamped := m.ContextWindow - reserve
254 if clamped < 1 {
255 clamped = 1
256 }
257 if maxTok > clamped {
258 maxTok = clamped
259 }
260 }
261 if m.Reasoning {
262 if maxTok > 0 {
263 out.MaxCompletionTok = &maxTok
264 }
265 effort := OpenAIReasoningEffort(reasoning)
266 if usesAdaptiveThinking(m) {
267 // Some gateways expose adaptive-thinking Anthropic models through
268 // the OpenAI-compatible chat-completions wire. They accept the
269 // same reasoning_effort knob, including the top "xhigh" tier;
270 // don't clamp zot's "maximum" to "high" for those models.
271 effort = OpenAICompatAnthropicEffort(reasoning)
272 }
273 if hasReasoningLevelOverride(m, req.Reasoning) {
274 // Explicit model mappings describe the endpoint's actual effort
275 // values and take precedence over conservative protocol defaults.
276 effort = reasoning
277 }
278 if effort != "" {
279 out.ReasoningEffort = effort
280 }
281 } else if maxTok > 0 {
282 // Omit max_tokens when unknown (some discovered models don't
283 // advertise an output cap) so the server applies its own default
284 // instead of receiving an invalid max_tokens: 0.
285 out.MaxTokens = &maxTok
286 }
287
288 if req.System != "" {
289 out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})
290 }
291
292 deferredMode := supportsDeferredTools(c.Name()) && isKimiDeferredModel(req.Model)
293 toolByName := make(map[string]Tool, len(req.Tools))
294 for _, t := range req.Tools {
295 toolByName[t.Name] = t
296 }
297 activatedTools := activatedToolNames(req.Messages)
298
299 // DeepSeek's chat-completions API rejects the multimodal content
300 // schema (parts arrays containing image_url). Force every user/tool
301 // message to a plain string and silently drop image blocks for
302 // this provider so historical sessions with screenshots still replay.
303 textOnly := c.name == "deepseek"
304
305 req.Messages = RepairOrphanedToolResults(req.Messages)
306 for msgIndex := 0; msgIndex < len(req.Messages); msgIndex++ {
307 msg := req.Messages[msgIndex]
308 addedToolNames := append([]string(nil), msg.AddedToolNames...)
309 switch msg.Role {
310 case RoleUser:
311 content := buildOAIUserContent(msg.Content, textOnly)
312 out.Messages = append(out.Messages, oaiMessage{Role: "user", Content: content})
313 case RoleAssistant:
314 am := oaiMessage{Role: "assistant"}
315 var text strings.Builder
316 var reasoning strings.Builder
317 for _, b := range msg.Content {
318 switch v := b.(type) {
319 case TextBlock:
320 if strings.TrimSpace(v.Text) == "" {
321 continue
322 }
323 if text.Len() > 0 {
324 text.WriteString("\n")
325 }
326 text.WriteString(v.Text)
327 case ToolCallBlock:
328 if v.Server {
329 continue
330 }
331 args := v.Arguments
332 if len(args) == 0 || !json.Valid(args) {
333 args = json.RawMessage("{}")
334 }
335 am.ToolCalls = append(am.ToolCalls, oaiToolCall{
336 ID: v.ID,
337 Type: "function",
338 Function: oaiToolCallFn{
339 Name: v.Name,
340 Arguments: string(args),
341 },
342 })
343 case ReasoningBlock:
344 if v.Summary != "" {
345 if reasoning.Len() > 0 {
346 reasoning.WriteString("\n")
347 }
348 reasoning.WriteString(v.Summary)
349 }
350 }
351 }
352 if text.Len() > 0 {
353 am.Content = text.String()
354 }
355 if reasoning.Len() > 0 && len(am.ToolCalls) > 0 {
356 am.ReasoningContent = reasoning.String()
357 }
358 // Kimi rejects assistant messages with neither visible text nor
359 // tool calls ("assistant must not be empty"). This can happen when
360 // a previous stream produced only reasoning_content, which zot keeps
361 // internally for provider replay but cannot send back as standalone
362 // assistant content on OpenAI-compatible chat-completions APIs.
363 if am.Content == nil && len(am.ToolCalls) == 0 {
364 continue
365 }
366 out.Messages = append(out.Messages, am)
367 case RoleTool:
368 // Chat Completions tool messages only carry text. Keep every
369 // tool result paired with its call, then deliver any images in
370 // one user message after the complete batch of tool results.
371 // Strict OpenAI-compatible APIs reject image_url blocks on a
372 // role=tool message.
373 var images []Content
374 groupEnd := msgIndex
375 for groupEnd < len(req.Messages) && req.Messages[groupEnd].Role == RoleTool {
376 toolMessage := req.Messages[groupEnd]
377 if groupEnd > msgIndex {
378 addedToolNames = append(addedToolNames, toolMessage.AddedToolNames...)
379 }
380 for _, b := range toolMessage.Content {
381 if tr, ok := b.(ToolResultBlock); ok {
382 content := buildOAIToolContent(tr.Content, tr.IsError, textOnly)
383 out.Messages = append(out.Messages, oaiMessage{
384 Role: "tool",
385 ToolCallID: tr.CallID,
386 Content: content,
387 })
388 if !textOnly {
389 for _, inner := range tr.Content {
390 if image, ok := inner.(ImageBlock); ok {
391 images = append(images, image)
392 }
393 }
394 }
395 }
396 }
397 groupEnd++
398 }
399 // The built-in OpenAI client also persists this image mirror so
400 // Responses-routed models can consume it. Reuse that next message
401 // instead of sending the image twice.
402 hasPersistedMirror := groupEnd < len(req.Messages) && isOpenAIToolImageMirror(req.Messages[groupEnd])
403 if len(images) > 0 && !hasPersistedMirror {
404 content := append([]Content{TextBlock{Text: "Tool output included the following image content:"}}, images...)
405 out.Messages = append(out.Messages, oaiMessage{
406 Role: "user",
407 Content: buildOAIContentBlocks(content, false),
408 })
409 }
410 msgIndex = groupEnd - 1
411 }
412 if deferredMode {
413 var loaded []oaiTool
414 for _, name := range addedToolNames {
415 if t, ok := toolByName[name]; ok && t.Deferred {
416 loaded = append(loaded, makeOAITool(t))
417 }
418 }
419 if len(loaded) > 0 {
420 out.Messages = append(out.Messages, oaiMessage{Role: "system", Tools: loaded})
421 }
422 }
423 }
424
425 for _, t := range req.Tools {
426 if t.Deferred && (deferredMode || !activatedTools[t.Name]) {
427 continue
428 }
429 out.Tools = append(out.Tools, makeOAITool(t))
430 }
431 if len(out.Tools) > 0 {
432 out.ToolChoice = "auto"
433 }
434
435 return out, nil
436}
437
438func isKimiDeferredModel(model string) bool {
439 model = strings.ToLower(model)
440 return model == "kimi-k3" || strings.HasSuffix(model, "/kimi-k3")
441}
442
443// supportsDeferredTools reports whether a provider speaks Moonshot's
444// deferred-tools handshake, where a tool that becomes available mid-conversation
445// is delivered on a system message carrying a `tools` array.
446//
447// That message is a Moonshot API extension, not OpenAI wire format: it has no
448// content, which a strict chat-completions schema rejects outright. Gondola,
449// which serves kimi-k3 on top of Venice, answers "messages.N: Invalid input"
450// and the conversation cannot recover, because the injection is rebuilt from
451// message history on every later turn. Providers outside this list take the
452// standard path instead, where an activated deferred tool joins the top-level
453// tools array.
454func supportsDeferredTools(provider string) bool {
455 return provider == "moonshotai" || provider == "moonshotai-cn"
456}
457
458func isOpenRouterServerToolType(typ string) bool {
459 return strings.HasPrefix(typ, "openrouter:")
460}
461
462func makeOAITool(t Tool) oaiTool {
463 if isOpenRouterServerToolType(t.Name) {
464 return oaiTool{Type: t.Name, Parameters: t.Schema}
465 }
466 var tool oaiTool
467 tool.Type = "function"
468 tool.Function = &struct {
469 Name string `json:"name"`
470 Description string `json:"description,omitempty"`
471 Parameters json.RawMessage `json:"parameters"`
472 }{
473 Name: t.Name,
474 Description: t.Description,
475 Parameters: t.Schema,
476 }
477 return tool
478}
479
480func buildOAIUserContent(blocks []Content, textOnly bool) interface{} {
481 hasImage := false
482 for _, b := range blocks {
483 if _, ok := b.(ImageBlock); ok {
484 hasImage = true
485 break
486 }
487 }
488 if textOnly || !hasImage {
489 var sb strings.Builder
490 for _, b := range blocks {
491 if tb, ok := b.(TextBlock); ok {
492 if sb.Len() > 0 {
493 sb.WriteString("\n")
494 }
495 sb.WriteString(tb.Text)
496 }
497 }
498 return sb.String()
499 }
500 return buildOAIContentBlocks(blocks, false)
501}
502
503func buildOAIToolContent(blocks []Content, isError, textOnly bool) string {
504 var sb strings.Builder
505 hasImage := false
506 for _, b := range blocks {
507 switch v := b.(type) {
508 case TextBlock:
509 if sb.Len() > 0 {
510 sb.WriteString("\n")
511 }
512 sb.WriteString(v.Text)
513 case ImageBlock:
514 hasImage = true
515 }
516 }
517 if sb.Len() == 0 && hasImage && !textOnly {
518 sb.WriteString("(see attached image)")
519 }
520 if isError && sb.Len() > 0 {
521 sb.WriteString(" [error]")
522 }
523 return sb.String()
524}
525
526func isOpenAIToolImageMirror(msg Message) bool {
527 if msg.Role != RoleUser || len(msg.Content) < 2 {
528 return false
529 }
530 text, ok := msg.Content[0].(TextBlock)
531 if !ok || text.Text != "Tool output included the following image content:" {
532 return false
533 }
534 for _, content := range msg.Content[1:] {
535 if _, ok := content.(ImageBlock); ok {
536 return true
537 }
538 }
539 return false
540}
541
542func buildOAIContentBlocks(blocks []Content, isError bool) []interface{} {
543 var arr []interface{}
544 for _, b := range blocks {
545 switch v := b.(type) {
546 case TextBlock:
547 arr = append(arr, oaiContentText{Type: "text", Text: v.Text})
548 case ImageBlock:
549 var img oaiContentImage
550 img.Type = "image_url"
551 img.ImageURL.URL = "data:" + v.MimeType + ";base64," + base64.StdEncoding.EncodeToString(v.Data)
552 arr = append(arr, img)
553 }
554 }
555 if isError {
556 arr = append(arr, oaiContentText{Type: "text", Text: "[error]"})
557 }
558 return arr
559}
560
561// ---- streaming ----
562
563func (c *openaiClient) chatCompletionsURL() string {
564 if c.chatCompletionsPath != "" {
565 return strings.TrimRight(c.baseURL, "/") + "/" + strings.TrimLeft(c.chatCompletionsPath, "/")
566 }
567 return chatCompletionsURL(c.baseURL)
568}
569
570func (c *openaiClient) Stream(ctx context.Context, req Request) (<-chan Event, error) {
571 endpoint := c.chatCompletionsURL()
572 wire, err := c.buildRequest(req)
573 if err != nil {
574 return nil, err
575 }
576 body, err := json.Marshal(wire)
577 if err != nil {
578 return nil, err
579 }
580 newReq := func() (*http.Request, error) {
581 httpReq, err := http.NewRequestWithContext(ctx, "POST", endpoint, bytes.NewReader(body))
582 if err != nil {
583 return nil, err
584 }
585 httpReq.Header.Set("content-type", "application/json")
586 httpReq.Header.Set("accept", "text/event-stream")
587 httpReq.Header.Set("authorization", "Bearer "+c.apiKey)
588 for k, v := range c.headers {
589 httpReq.Header.Set(k, v)
590 }
591 return httpReq, nil
592 }
593
594 resp, err := doStreamWithRetry(ctx, c.http, newReq)
595 if err != nil {
596 return nil, fmt.Errorf("%s: %w", c.Name(), err)
597 }
598 if resp.StatusCode != http.StatusOK {
599 b, _ := io.ReadAll(resp.Body)
600 resp.Body.Close()
601 return nil, fmt.Errorf("%s: http %d: %s", c.Name(), resp.StatusCode, strings.TrimSpace(string(b)))
602 }
603
604 out := make(chan Event, 16)
605 go c.runStream(ctx, resp, req, out)
606 return out, nil
607}
608
609func (c *openaiClient) runStream(ctx context.Context, resp *http.Response, req Request, out chan<- Event) {
610 defer close(out)
611 defer resp.Body.Close()
612
613 model, _ := FindModel("", req.Model)
614 out <- EventStart{Model: req.Model, Provider: c.Name()}
615
616 raw := make(chan sseEvent, 16)
617 go readSSE(resp.Body, raw)
618
619 // Interleaved block tracking: text and tool_calls preserve their
620 // emission order so the assistant message renders in the same order
621 // the model produced it. The builder fires one kind of block at a
622 // time — incoming text deltas after a tool_call split into a fresh
623 // text block; subsequent tool_calls each get their own slot.
624 type blockEntry struct {
625 kind string // "text" | "tool_use"
626 textBuf strings.Builder
627 toolID string
628 toolName string
629 toolArgs strings.Builder
630 announced bool
631 server bool
632 }
633 var (
634 blocks []*blockEntry
635 currentText *blockEntry // most-recent text block, nil if none
636 toolByIdx = map[int]*blockEntry{} // openai tool_call index -> block
637 reasoningBuf strings.Builder
638 usage Usage
639 stop StopReason = StopEnd
640 finalErr error
641 )
642
643 appendText := func(delta string) {
644 if currentText == nil {
645 currentText = &blockEntry{kind: "text"}
646 blocks = append(blocks, currentText)
647 }
648 currentText.textBuf.WriteString(delta)
649 }
650
651 getOrCreateTool := func(idx int) *blockEntry {
652 if t, ok := toolByIdx[idx]; ok {
653 return t
654 }
655 t := &blockEntry{kind: "tool_use"}
656 toolByIdx[idx] = t
657 blocks = append(blocks, t)
658 // A new tool block breaks the current text block. Subsequent text
659 // deltas will start a fresh text block after this tool.
660 currentText = nil
661 return t
662 }
663
664 assembleMsg := func() Message {
665 content := []Content{}
666 for _, b := range blocks {
667 switch b.kind {
668 case "text":
669 if b.textBuf.Len() > 0 {
670 content = append(content, TextBlock{Text: b.textBuf.String()})
671 }
672 case "tool_use":
673 args := b.toolArgs.String()
674 if args == "" || !json.Valid([]byte(args)) {
675 args = "{}"
676 }
677 content = append(content, ToolCallBlock{
678 ID: b.toolID, Name: b.toolName, Arguments: json.RawMessage(args), Server: b.server,
679 })
680 }
681 }
682 if reasoningBuf.Len() > 0 {
683 content = append(content, ReasoningBlock{Summary: reasoningBuf.String()})
684 }
685 return Message{Role: RoleAssistant, Content: content, Time: time.Now()}
686 }
687
688 sendDone := func() {
689 usage.CostUSD = ComputeCost(model, usage)
690 out <- EventUsage{Usage: usage}
691 out <- EventDone{Stop: stop, Err: finalErr, Message: assembleMsg()}
692 }
693
694 for {
695 select {
696 case <-ctx.Done():
697 stop = StopAborted
698 finalErr = ctx.Err()
699 sendDone()
700 return
701 case ev, ok := <-raw:
702 if !ok {
703 sendDone()
704 return
705 }
706 if ev.Data == "[DONE]" {
707 sendDone()
708 return
709 }
710 var chunk struct {
711 Choices []struct {
712 Index int `json:"index"`
713 Delta struct {
714 Content string `json:"content"`
715 ReasoningContent string `json:"reasoning_content"`
716 ToolCalls []struct {
717 Index int `json:"index"`
718 ID string `json:"id"`
719 Type string `json:"type"`
720 Name string `json:"name"`
721 Arguments json.RawMessage `json:"arguments"`
722 Function struct {
723 Name string `json:"name"`
724 Arguments string `json:"arguments"`
725 } `json:"function"`
726 } `json:"tool_calls"`
727 } `json:"delta"`
728 FinishReason string `json:"finish_reason"`
729 } `json:"choices"`
730 Usage *struct {
731 PromptTokens int `json:"prompt_tokens"`
732 CompletionTokens int `json:"completion_tokens"`
733 PromptTokensDetails struct {
734 CachedTokens int `json:"cached_tokens"`
735 } `json:"prompt_tokens_details"`
736 CompletionTokensDetails *struct {
737 ReasoningTokens int `json:"reasoning_tokens"`
738 } `json:"completion_tokens_details"`
739 } `json:"usage"`
740 Error *struct {
741 Message string `json:"message"`
742 Type string `json:"type"`
743 } `json:"error"`
744 }
745 if err := json.Unmarshal([]byte(ev.Data), &chunk); err != nil {
746 continue
747 }
748 if chunk.Error != nil {
749 stop = StopError
750 finalErr = fmt.Errorf("openai: %s", chunk.Error.Message)
751 sendDone()
752 return
753 }
754 if chunk.Usage != nil {
755 usage.InputTokens = chunk.Usage.PromptTokens - chunk.Usage.PromptTokensDetails.CachedTokens
756 if usage.InputTokens < 0 {
757 usage.InputTokens = chunk.Usage.PromptTokens
758 }
759 usage.OutputTokens = chunk.Usage.CompletionTokens
760 usage.CacheReadTokens = chunk.Usage.PromptTokensDetails.CachedTokens
761 if details := chunk.Usage.CompletionTokensDetails; details != nil {
762 usage.ReasoningTokens = details.ReasoningTokens
763 usage.ReasoningTokensKnown = true
764 }
765 }
766 for _, ch := range chunk.Choices {
767 if ch.Delta.ReasoningContent != "" {
768 reasoningBuf.WriteString(ch.Delta.ReasoningContent)
769 }
770 if ch.Delta.Content != "" {
771 appendText(ch.Delta.Content)
772 out <- EventTextDelta{Delta: ch.Delta.Content}
773 }
774 for _, tc := range ch.Delta.ToolCalls {
775 t := getOrCreateTool(tc.Index)
776 if tc.ID != "" {
777 t.toolID = tc.ID
778 }
779 if isOpenRouterServerToolType(tc.Type) {
780 t.server = true
781 if tc.Name != "" {
782 t.toolName = tc.Name
783 } else if t.toolName == "" {
784 t.toolName = tc.Type
785 }
786 if len(tc.Arguments) > 0 {
787 t.toolArgs.Write(tc.Arguments)
788 }
789 }
790 if tc.Function.Name != "" {
791 t.toolName = tc.Function.Name
792 if isOpenRouterServerToolType(tc.Function.Name) {
793 t.server = true
794 }
795 }
796 if !t.announced && t.toolID != "" && t.toolName != "" {
797 t.announced = true
798 out <- EventToolStart{ID: t.toolID, Name: t.toolName}
799 }
800 if tc.Function.Arguments != "" {
801 t.toolArgs.WriteString(tc.Function.Arguments)
802 if t.announced {
803 out <- EventToolArgs{ID: t.toolID, Delta: tc.Function.Arguments}
804 }
805 }
806 }
807 switch ch.FinishReason {
808 case "stop":
809 stop = StopEnd
810 case "length":
811 stop = StopLength
812 case "tool_calls", "function_call":
813 stop = StopToolUse
814 hasClientTool := false
815 for _, b := range blocks {
816 if b.kind == "tool_use" && b.announced {
817 out <- EventToolEnd{ID: b.toolID}
818 if !b.server {
819 hasClientTool = true
820 }
821 }
822 }
823 if !hasClientTool {
824 // OpenRouter executed every tool server-side; the
825 // client must not send tool_result messages.
826 stop = StopEnd
827 }
828 }
829 }
830 }
831 }
832}