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 c.Name() == "yolo-auto" && effort == "minimum" {
279 // The advertised wire value is minimal, not zot's UI spelling.
280 effort = "minimal"
281 }
282 if effort != "" {
283 out.ReasoningEffort = effort
284 }
285 } else if maxTok > 0 {
286 // Omit max_tokens when unknown (some discovered models don't
287 // advertise an output cap) so the server applies its own default
288 // instead of receiving an invalid max_tokens: 0.
289 out.MaxTokens = &maxTok
290 }
291
292 if req.System != "" {
293 out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})
294 }
295
296 deferredMode := supportsDeferredTools(c.Name()) && isKimiDeferredModel(req.Model)
297 toolByName := make(map[string]Tool, len(req.Tools))
298 for _, t := range req.Tools {
299 toolByName[t.Name] = t
300 }
301 activatedTools := activatedToolNames(req.Messages)
302
303 // DeepSeek's chat-completions API rejects the multimodal content
304 // schema (parts arrays containing image_url). Force every user/tool
305 // message to a plain string and silently drop image blocks for
306 // this provider so historical sessions with screenshots still replay.
307 textOnly := c.name == "deepseek"
308
309 req.Messages = RepairOrphanedToolResults(req.Messages)
310 for msgIndex := 0; msgIndex < len(req.Messages); msgIndex++ {
311 msg := req.Messages[msgIndex]
312 addedToolNames := append([]string(nil), msg.AddedToolNames...)
313 switch msg.Role {
314 case RoleUser:
315 content := buildOAIUserContent(msg.Content, textOnly)
316 out.Messages = append(out.Messages, oaiMessage{Role: "user", Content: content})
317 case RoleAssistant:
318 am := oaiMessage{Role: "assistant"}
319 var text strings.Builder
320 var reasoning strings.Builder
321 for _, b := range msg.Content {
322 switch v := b.(type) {
323 case TextBlock:
324 if strings.TrimSpace(v.Text) == "" {
325 continue
326 }
327 if text.Len() > 0 {
328 text.WriteString("\n")
329 }
330 text.WriteString(v.Text)
331 case ToolCallBlock:
332 if v.Server {
333 continue
334 }
335 args := v.Arguments
336 if len(args) == 0 || !json.Valid(args) {
337 args = json.RawMessage("{}")
338 }
339 am.ToolCalls = append(am.ToolCalls, oaiToolCall{
340 ID: v.ID,
341 Type: "function",
342 Function: oaiToolCallFn{
343 Name: v.Name,
344 Arguments: string(args),
345 },
346 })
347 case ReasoningBlock:
348 if v.Summary != "" {
349 if reasoning.Len() > 0 {
350 reasoning.WriteString("\n")
351 }
352 reasoning.WriteString(v.Summary)
353 }
354 }
355 }
356 if text.Len() > 0 {
357 am.Content = text.String()
358 }
359 if reasoning.Len() > 0 && len(am.ToolCalls) > 0 {
360 am.ReasoningContent = reasoning.String()
361 }
362 // Kimi rejects assistant messages with neither visible text nor
363 // tool calls ("assistant must not be empty"). This can happen when
364 // a previous stream produced only reasoning_content, which zot keeps
365 // internally for provider replay but cannot send back as standalone
366 // assistant content on OpenAI-compatible chat-completions APIs.
367 if am.Content == nil && len(am.ToolCalls) == 0 {
368 continue
369 }
370 out.Messages = append(out.Messages, am)
371 case RoleTool:
372 // Chat Completions tool messages only carry text. Keep every
373 // tool result paired with its call, then deliver any images in
374 // one user message after the complete batch of tool results.
375 // Strict OpenAI-compatible APIs reject image_url blocks on a
376 // role=tool message.
377 var images []Content
378 groupEnd := msgIndex
379 for groupEnd < len(req.Messages) && req.Messages[groupEnd].Role == RoleTool {
380 toolMessage := req.Messages[groupEnd]
381 if groupEnd > msgIndex {
382 addedToolNames = append(addedToolNames, toolMessage.AddedToolNames...)
383 }
384 for _, b := range toolMessage.Content {
385 if tr, ok := b.(ToolResultBlock); ok {
386 content := buildOAIToolContent(tr.Content, tr.IsError, textOnly)
387 out.Messages = append(out.Messages, oaiMessage{
388 Role: "tool",
389 ToolCallID: tr.CallID,
390 Content: content,
391 })
392 if !textOnly {
393 for _, inner := range tr.Content {
394 if image, ok := inner.(ImageBlock); ok {
395 images = append(images, image)
396 }
397 }
398 }
399 }
400 }
401 groupEnd++
402 }
403 // The built-in OpenAI client also persists this image mirror so
404 // Responses-routed models can consume it. Reuse that next message
405 // instead of sending the image twice.
406 hasPersistedMirror := groupEnd < len(req.Messages) && isOpenAIToolImageMirror(req.Messages[groupEnd])
407 if len(images) > 0 && !hasPersistedMirror {
408 content := append([]Content{TextBlock{Text: "Tool output included the following image content:"}}, images...)
409 out.Messages = append(out.Messages, oaiMessage{
410 Role: "user",
411 Content: buildOAIContentBlocks(content, false),
412 })
413 }
414 msgIndex = groupEnd - 1
415 }
416 if deferredMode {
417 var loaded []oaiTool
418 for _, name := range addedToolNames {
419 if t, ok := toolByName[name]; ok && t.Deferred {
420 loaded = append(loaded, makeOAITool(t))
421 }
422 }
423 if len(loaded) > 0 {
424 out.Messages = append(out.Messages, oaiMessage{Role: "system", Tools: loaded})
425 }
426 }
427 }
428
429 for _, t := range req.Tools {
430 if t.Deferred && (deferredMode || !activatedTools[t.Name]) {
431 continue
432 }
433 out.Tools = append(out.Tools, makeOAITool(t))
434 }
435 if len(out.Tools) > 0 {
436 out.ToolChoice = "auto"
437 }
438
439 return out, nil
440}
441
442func isKimiDeferredModel(model string) bool {
443 model = strings.ToLower(model)
444 return model == "kimi-k3" || strings.HasSuffix(model, "/kimi-k3")
445}
446
447// supportsDeferredTools reports whether a provider speaks Moonshot's
448// deferred-tools handshake, where a tool that becomes available mid-conversation
449// is delivered on a system message carrying a `tools` array.
450//
451// That message is a Moonshot API extension, not OpenAI wire format: it has no
452// content, which a strict chat-completions schema rejects outright. Gondola,
453// which serves kimi-k3 on top of Venice, answers "messages.N: Invalid input"
454// and the conversation cannot recover, because the injection is rebuilt from
455// message history on every later turn. Providers outside this list take the
456// standard path instead, where an activated deferred tool joins the top-level
457// tools array.
458func supportsDeferredTools(provider string) bool {
459 return provider == "moonshotai" || provider == "moonshotai-cn"
460}
461
462func isOpenRouterServerToolType(typ string) bool {
463 return strings.HasPrefix(typ, "openrouter:")
464}
465
466func makeOAITool(t Tool) oaiTool {
467 if isOpenRouterServerToolType(t.Name) {
468 return oaiTool{Type: t.Name, Parameters: t.Schema}
469 }
470 var tool oaiTool
471 tool.Type = "function"
472 tool.Function = &struct {
473 Name string `json:"name"`
474 Description string `json:"description,omitempty"`
475 Parameters json.RawMessage `json:"parameters"`
476 }{
477 Name: t.Name,
478 Description: t.Description,
479 Parameters: t.Schema,
480 }
481 return tool
482}
483
484func buildOAIUserContent(blocks []Content, textOnly bool) interface{} {
485 hasImage := false
486 for _, b := range blocks {
487 if _, ok := b.(ImageBlock); ok {
488 hasImage = true
489 break
490 }
491 }
492 if textOnly || !hasImage {
493 var sb strings.Builder
494 for _, b := range blocks {
495 if tb, ok := b.(TextBlock); ok {
496 if sb.Len() > 0 {
497 sb.WriteString("\n")
498 }
499 sb.WriteString(tb.Text)
500 }
501 }
502 return sb.String()
503 }
504 return buildOAIContentBlocks(blocks, false)
505}
506
507func buildOAIToolContent(blocks []Content, isError, textOnly bool) string {
508 var sb strings.Builder
509 hasImage := false
510 for _, b := range blocks {
511 switch v := b.(type) {
512 case TextBlock:
513 if sb.Len() > 0 {
514 sb.WriteString("\n")
515 }
516 sb.WriteString(v.Text)
517 case ImageBlock:
518 hasImage = true
519 }
520 }
521 if sb.Len() == 0 && hasImage && !textOnly {
522 sb.WriteString("(see attached image)")
523 }
524 if isError && sb.Len() > 0 {
525 sb.WriteString(" [error]")
526 }
527 return sb.String()
528}
529
530func isOpenAIToolImageMirror(msg Message) bool {
531 if msg.Role != RoleUser || len(msg.Content) < 2 {
532 return false
533 }
534 text, ok := msg.Content[0].(TextBlock)
535 if !ok || text.Text != "Tool output included the following image content:" {
536 return false
537 }
538 for _, content := range msg.Content[1:] {
539 if _, ok := content.(ImageBlock); ok {
540 return true
541 }
542 }
543 return false
544}
545
546func buildOAIContentBlocks(blocks []Content, isError bool) []interface{} {
547 var arr []interface{}
548 for _, b := range blocks {
549 switch v := b.(type) {
550 case TextBlock:
551 arr = append(arr, oaiContentText{Type: "text", Text: v.Text})
552 case ImageBlock:
553 var img oaiContentImage
554 img.Type = "image_url"
555 img.ImageURL.URL = "data:" + v.MimeType + ";base64," + base64.StdEncoding.EncodeToString(v.Data)
556 arr = append(arr, img)
557 }
558 }
559 if isError {
560 arr = append(arr, oaiContentText{Type: "text", Text: "[error]"})
561 }
562 return arr
563}
564
565// ---- streaming ----
566
567func (c *openaiClient) chatCompletionsURL() string {
568 if c.chatCompletionsPath != "" {
569 return strings.TrimRight(c.baseURL, "/") + "/" + strings.TrimLeft(c.chatCompletionsPath, "/")
570 }
571 return chatCompletionsURL(c.baseURL)
572}
573
574func (c *openaiClient) Stream(ctx context.Context, req Request) (<-chan Event, error) {
575 endpoint := c.chatCompletionsURL()
576 wire, err := c.buildRequest(req)
577 if err != nil {
578 return nil, err
579 }
580 body, err := json.Marshal(wire)
581 if err != nil {
582 return nil, err
583 }
584 newReq := func() (*http.Request, error) {
585 httpReq, err := http.NewRequestWithContext(ctx, "POST", endpoint, bytes.NewReader(body))
586 if err != nil {
587 return nil, err
588 }
589 httpReq.Header.Set("content-type", "application/json")
590 httpReq.Header.Set("accept", "text/event-stream")
591 httpReq.Header.Set("authorization", "Bearer "+c.apiKey)
592 for k, v := range c.headers {
593 httpReq.Header.Set(k, v)
594 }
595 return httpReq, nil
596 }
597
598 resp, err := doStreamWithRetry(ctx, c.http, newReq)
599 if err != nil {
600 return nil, fmt.Errorf("%s: %w", c.Name(), err)
601 }
602 if resp.StatusCode != http.StatusOK {
603 b, _ := io.ReadAll(resp.Body)
604 resp.Body.Close()
605 return nil, fmt.Errorf("%s: http %d: %s", c.Name(), resp.StatusCode, strings.TrimSpace(string(b)))
606 }
607
608 out := make(chan Event, 16)
609 go c.runStream(ctx, resp, req, out)
610 return out, nil
611}
612
613func (c *openaiClient) runStream(ctx context.Context, resp *http.Response, req Request, out chan<- Event) {
614 defer close(out)
615 defer resp.Body.Close()
616
617 model, _ := FindModel("", req.Model)
618 out <- EventStart{Model: req.Model, Provider: c.Name()}
619
620 raw := make(chan sseEvent, 16)
621 go readSSE(resp.Body, raw)
622
623 // Interleaved block tracking: text and tool_calls preserve their
624 // emission order so the assistant message renders in the same order
625 // the model produced it. The builder fires one kind of block at a
626 // time — incoming text deltas after a tool_call split into a fresh
627 // text block; subsequent tool_calls each get their own slot.
628 type blockEntry struct {
629 kind string // "text" | "tool_use"
630 textBuf strings.Builder
631 toolID string
632 toolName string
633 toolArgs strings.Builder
634 announced bool
635 server bool
636 }
637 var (
638 blocks []*blockEntry
639 currentText *blockEntry // most-recent text block, nil if none
640 toolByIdx = map[int]*blockEntry{} // openai tool_call index -> block
641 reasoningBuf strings.Builder
642 usage Usage
643 stop StopReason = StopEnd
644 finalErr error
645 )
646
647 appendText := func(delta string) {
648 if currentText == nil {
649 currentText = &blockEntry{kind: "text"}
650 blocks = append(blocks, currentText)
651 }
652 currentText.textBuf.WriteString(delta)
653 }
654
655 getOrCreateTool := func(idx int) *blockEntry {
656 if t, ok := toolByIdx[idx]; ok {
657 return t
658 }
659 t := &blockEntry{kind: "tool_use"}
660 toolByIdx[idx] = t
661 blocks = append(blocks, t)
662 // A new tool block breaks the current text block. Subsequent text
663 // deltas will start a fresh text block after this tool.
664 currentText = nil
665 return t
666 }
667
668 assembleMsg := func() Message {
669 content := []Content{}
670 for _, b := range blocks {
671 switch b.kind {
672 case "text":
673 if b.textBuf.Len() > 0 {
674 content = append(content, TextBlock{Text: b.textBuf.String()})
675 }
676 case "tool_use":
677 args := b.toolArgs.String()
678 if args == "" || !json.Valid([]byte(args)) {
679 args = "{}"
680 }
681 content = append(content, ToolCallBlock{
682 ID: b.toolID, Name: b.toolName, Arguments: json.RawMessage(args), Server: b.server,
683 })
684 }
685 }
686 if reasoningBuf.Len() > 0 {
687 content = append(content, ReasoningBlock{Summary: reasoningBuf.String()})
688 }
689 return Message{Role: RoleAssistant, Content: content, Time: time.Now()}
690 }
691
692 sendDone := func() {
693 usage.CostUSD = ComputeCost(model, usage)
694 out <- EventUsage{Usage: usage}
695 out <- EventDone{Stop: stop, Err: finalErr, Message: assembleMsg()}
696 }
697
698 for {
699 select {
700 case <-ctx.Done():
701 stop = StopAborted
702 finalErr = ctx.Err()
703 sendDone()
704 return
705 case ev, ok := <-raw:
706 if !ok {
707 sendDone()
708 return
709 }
710 if ev.Data == "[DONE]" {
711 sendDone()
712 return
713 }
714 var chunk struct {
715 Choices []struct {
716 Index int `json:"index"`
717 Delta struct {
718 Content string `json:"content"`
719 ReasoningContent string `json:"reasoning_content"`
720 ToolCalls []struct {
721 Index int `json:"index"`
722 ID string `json:"id"`
723 Type string `json:"type"`
724 Name string `json:"name"`
725 Arguments json.RawMessage `json:"arguments"`
726 Function struct {
727 Name string `json:"name"`
728 Arguments string `json:"arguments"`
729 } `json:"function"`
730 } `json:"tool_calls"`
731 } `json:"delta"`
732 FinishReason string `json:"finish_reason"`
733 } `json:"choices"`
734 Usage *struct {
735 PromptTokens int `json:"prompt_tokens"`
736 CompletionTokens int `json:"completion_tokens"`
737 PromptTokensDetails struct {
738 CachedTokens int `json:"cached_tokens"`
739 } `json:"prompt_tokens_details"`
740 CompletionTokensDetails *struct {
741 ReasoningTokens int `json:"reasoning_tokens"`
742 } `json:"completion_tokens_details"`
743 } `json:"usage"`
744 Error *struct {
745 Message string `json:"message"`
746 Type string `json:"type"`
747 } `json:"error"`
748 }
749 if err := json.Unmarshal([]byte(ev.Data), &chunk); err != nil {
750 continue
751 }
752 if chunk.Error != nil {
753 stop = StopError
754 finalErr = fmt.Errorf("openai: %s", chunk.Error.Message)
755 sendDone()
756 return
757 }
758 if chunk.Usage != nil {
759 usage.InputTokens = chunk.Usage.PromptTokens - chunk.Usage.PromptTokensDetails.CachedTokens
760 if usage.InputTokens < 0 {
761 usage.InputTokens = chunk.Usage.PromptTokens
762 }
763 usage.OutputTokens = chunk.Usage.CompletionTokens
764 usage.CacheReadTokens = chunk.Usage.PromptTokensDetails.CachedTokens
765 if details := chunk.Usage.CompletionTokensDetails; details != nil {
766 usage.ReasoningTokens = details.ReasoningTokens
767 usage.ReasoningTokensKnown = true
768 }
769 }
770 for _, ch := range chunk.Choices {
771 if ch.Delta.ReasoningContent != "" {
772 reasoningBuf.WriteString(ch.Delta.ReasoningContent)
773 }
774 if ch.Delta.Content != "" {
775 appendText(ch.Delta.Content)
776 out <- EventTextDelta{Delta: ch.Delta.Content}
777 }
778 for _, tc := range ch.Delta.ToolCalls {
779 t := getOrCreateTool(tc.Index)
780 if tc.ID != "" {
781 t.toolID = tc.ID
782 }
783 if isOpenRouterServerToolType(tc.Type) {
784 t.server = true
785 if tc.Name != "" {
786 t.toolName = tc.Name
787 } else if t.toolName == "" {
788 t.toolName = tc.Type
789 }
790 if len(tc.Arguments) > 0 {
791 t.toolArgs.Write(tc.Arguments)
792 }
793 }
794 if tc.Function.Name != "" {
795 t.toolName = tc.Function.Name
796 if isOpenRouterServerToolType(tc.Function.Name) {
797 t.server = true
798 }
799 }
800 if !t.announced && t.toolID != "" && t.toolName != "" {
801 t.announced = true
802 out <- EventToolStart{ID: t.toolID, Name: t.toolName}
803 }
804 if tc.Function.Arguments != "" {
805 t.toolArgs.WriteString(tc.Function.Arguments)
806 if t.announced {
807 out <- EventToolArgs{ID: t.toolID, Delta: tc.Function.Arguments}
808 }
809 }
810 }
811 switch ch.FinishReason {
812 case "stop":
813 stop = StopEnd
814 case "length":
815 stop = StopLength
816 case "tool_calls", "function_call":
817 stop = StopToolUse
818 hasClientTool := false
819 for _, b := range blocks {
820 if b.kind == "tool_use" && b.announced {
821 out <- EventToolEnd{ID: b.toolID}
822 if !b.server {
823 hasClientTool = true
824 }
825 }
826 }
827 if !hasClientTool {
828 // OpenRouter executed every tool server-side; the
829 // client must not send tool_result messages.
830 stop = StopEnd
831 }
832 }
833 }
834 }
835 }
836}