1package provider
2
3import (
4 "bytes"
5 "context"
6 "encoding/base64"
7 "encoding/json"
8 "fmt"
9 "io"
10 "net/http"
11 "strings"
12 "time"
13)
14
15const openaiDefaultBaseURL = "https://api.openai.com"
16
17type openaiClient struct {
18 apiKey string
19 baseURL string
20 name string
21 oauth bool // when true, apiKey actually holds an OAuth access token
22 headers map[string]string
23 http *http.Client
24}
25
26// NewOpenAI creates an OpenAI client using an API key. baseURL may be empty.
27func NewOpenAI(apiKey, baseURL string) Client {
28 if baseURL == "" {
29 baseURL = openaiDefaultBaseURL
30 }
31 return &openaiClient{
32 apiKey: apiKey,
33 baseURL: strings.TrimRight(baseURL, "/"),
34 name: "openai",
35 http: &http.Client{Timeout: 0},
36 }
37}
38
39// NewOpenAIOAuth creates an OpenAI client using a subscription OAuth access token.
40// The token is sent as an HTTP Bearer credential on the standard chat/completions endpoint.
41func NewOpenAIOAuth(accessToken, baseURL string) Client {
42 if baseURL == "" {
43 baseURL = openaiDefaultBaseURL
44 }
45 return &openaiClient{
46 apiKey: accessToken,
47 baseURL: strings.TrimRight(baseURL, "/"),
48 name: "openai",
49 oauth: true,
50 http: &http.Client{Timeout: 0},
51 }
52}
53
54// NewKimi creates a Kimi/Moonshot client. Kimi's chat API is OpenAI-compatible.
55func NewKimi(apiKey, baseURL string) Client {
56 return NewKimiWithHeaders(apiKey, baseURL, nil)
57}
58
59// NewDeepSeek creates a DeepSeek client. DeepSeek's chat API is
60// OpenAI-compatible at https://api.deepseek.com/v1.
61func NewDeepSeek(apiKey, baseURL string) Client {
62 if baseURL == "" {
63 baseURL = "https://api.deepseek.com/v1"
64 }
65 return &openaiClient{
66 apiKey: apiKey,
67 baseURL: strings.TrimRight(baseURL, "/"),
68 name: "deepseek",
69 http: &http.Client{Timeout: 0},
70 }
71}
72
73// NewKimiWithHeaders creates a Kimi/Moonshot client with extra headers.
74// Subscription tokens from Kimi Code need the official CLI's X-Msh-* headers.
75func NewKimiWithHeaders(apiKey, baseURL string, headers map[string]string) Client {
76 if baseURL == "" {
77 baseURL = "https://api.kimi.com/coding/v1"
78 }
79 return &openaiClient{
80 apiKey: apiKey,
81 baseURL: strings.TrimRight(baseURL, "/"),
82 name: "kimi",
83 headers: headers,
84 http: &http.Client{Timeout: 0},
85 }
86}
87
88func (c *openaiClient) Name() string {
89 if c.name != "" {
90 return c.name
91 }
92 return "openai"
93}
94
95// ---- wire types ----
96
97type oaiContentText struct {
98 Type string `json:"type"` // "text"
99 Text string `json:"text"`
100}
101
102type oaiContentImage struct {
103 Type string `json:"type"` // "image_url"
104 ImageURL struct {
105 URL string `json:"url"`
106 } `json:"image_url"`
107}
108
109type oaiToolCallFn struct {
110 Name string `json:"name"`
111 Arguments string `json:"arguments"` // JSON string
112}
113
114type oaiToolCall struct {
115 ID string `json:"id"`
116 Type string `json:"type"` // "function"
117 Function oaiToolCallFn `json:"function"`
118}
119
120type oaiMessage struct {
121 Role string `json:"role"`
122 Content interface{} `json:"content,omitempty"` // string or []block
123 Name string `json:"name,omitempty"`
124 ToolCalls []oaiToolCall `json:"tool_calls,omitempty"`
125 ToolCallID string `json:"tool_call_id,omitempty"`
126 // ReasoningContent carries the model's chain-of-thought summary
127 // alongside an assistant tool-call message. Required by Kimi's
128 // chat completions endpoint when thinking is enabled and the
129 // assistant message contains a tool call; OpenAI ignores it.
130 ReasoningContent string `json:"reasoning_content,omitempty"`
131}
132
133type oaiTool struct {
134 Type string `json:"type"` // "function"
135 Function struct {
136 Name string `json:"name"`
137 Description string `json:"description,omitempty"`
138 Parameters json.RawMessage `json:"parameters"`
139 } `json:"function"`
140}
141
142type oaiStreamOptions struct {
143 IncludeUsage bool `json:"include_usage"`
144}
145
146type oaiRequest struct {
147 Model string `json:"model"`
148 Messages []oaiMessage `json:"messages"`
149 Tools []oaiTool `json:"tools,omitempty"`
150 ToolChoice string `json:"tool_choice,omitempty"`
151 Stream bool `json:"stream"`
152 StreamOptions *oaiStreamOptions `json:"stream_options,omitempty"`
153 Temperature *float32 `json:"temperature,omitempty"`
154 MaxTokens *int `json:"max_tokens,omitempty"`
155 MaxCompletionTok *int `json:"max_completion_tokens,omitempty"`
156 ReasoningEffort string `json:"reasoning_effort,omitempty"`
157}
158
159// ---- request building ----
160
161func (c *openaiClient) buildRequest(req Request) (*oaiRequest, error) {
162 // Look up the model by id across all providers (not just openai)
163 // because the OpenAI client is also used for ollama and other
164 // OpenAI-compatible backends.
165 m, err := FindModel("", req.Model)
166 if err != nil {
167 // Unknown model: use sensible defaults so local/custom
168 // models still work without a catalog entry.
169 m = Model{
170 ID: req.Model,
171 ContextWindow: 32768,
172 MaxOutput: 8192,
173 }
174 }
175 out := &oaiRequest{
176 Model: req.Model,
177 Stream: true,
178 StreamOptions: &oaiStreamOptions{IncludeUsage: true},
179 Temperature: req.Temperature,
180 }
181
182 maxTok := req.MaxTokens
183 if maxTok <= 0 {
184 maxTok = m.MaxOutput
185 }
186 if m.Reasoning {
187 out.MaxCompletionTok = &maxTok
188 effort := OpenAIReasoningEffort(req.Reasoning)
189 if usesAdaptiveThinking(m) {
190 // Some gateways expose adaptive-thinking Anthropic models through
191 // the OpenAI-compatible chat-completions wire. They accept the
192 // same reasoning_effort knob, including the top "xhigh" tier;
193 // don't clamp zot's "maximum" to "high" for those models.
194 effort = OpenAICompatAnthropicEffort(req.Reasoning)
195 }
196 if effort != "" {
197 out.ReasoningEffort = effort
198 }
199 } else {
200 out.MaxTokens = &maxTok
201 }
202
203 if req.System != "" {
204 out.Messages = append(out.Messages, oaiMessage{Role: "system", Content: req.System})
205 }
206
207 // DeepSeek's chat-completions API rejects the multimodal content
208 // schema (parts arrays containing image_url). Force every user/tool
209 // message to a plain string and silently drop image blocks for
210 // this provider so historical sessions with screenshots still replay.
211 textOnly := c.name == "deepseek"
212
213 req.Messages = RepairOrphanedToolResults(req.Messages)
214 for _, msg := range req.Messages {
215 switch msg.Role {
216 case RoleUser:
217 content := buildOAIUserContent(msg.Content, textOnly)
218 out.Messages = append(out.Messages, oaiMessage{Role: "user", Content: content})
219 case RoleAssistant:
220 am := oaiMessage{Role: "assistant"}
221 var text strings.Builder
222 var reasoning strings.Builder
223 for _, b := range msg.Content {
224 switch v := b.(type) {
225 case TextBlock:
226 if strings.TrimSpace(v.Text) == "" {
227 continue
228 }
229 if text.Len() > 0 {
230 text.WriteString("\n")
231 }
232 text.WriteString(v.Text)
233 case ToolCallBlock:
234 args := v.Arguments
235 if len(args) == 0 || !json.Valid(args) {
236 args = json.RawMessage("{}")
237 }
238 am.ToolCalls = append(am.ToolCalls, oaiToolCall{
239 ID: v.ID,
240 Type: "function",
241 Function: oaiToolCallFn{
242 Name: v.Name,
243 Arguments: string(args),
244 },
245 })
246 case ReasoningBlock:
247 if v.Summary != "" {
248 if reasoning.Len() > 0 {
249 reasoning.WriteString("\n")
250 }
251 reasoning.WriteString(v.Summary)
252 }
253 }
254 }
255 if text.Len() > 0 {
256 am.Content = text.String()
257 }
258 if reasoning.Len() > 0 && len(am.ToolCalls) > 0 {
259 am.ReasoningContent = reasoning.String()
260 }
261 // Kimi rejects assistant messages with neither visible text nor
262 // tool calls ("assistant must not be empty"). This can happen when
263 // a previous stream produced only reasoning_content, which zot keeps
264 // internally for provider replay but cannot send back as standalone
265 // assistant content on OpenAI-compatible chat-completions APIs.
266 if am.Content == nil && len(am.ToolCalls) == 0 {
267 continue
268 }
269 out.Messages = append(out.Messages, am)
270 case RoleTool:
271 // Each ToolResultBlock becomes its own tool message. Preserve
272 // image blocks for vision-capable OpenAI models instead of
273 // flattening the tool output to plain text.
274 for _, b := range msg.Content {
275 if tr, ok := b.(ToolResultBlock); ok {
276 content := buildOAIToolContent(tr.Content, tr.IsError, textOnly)
277 out.Messages = append(out.Messages, oaiMessage{
278 Role: "tool",
279 ToolCallID: tr.CallID,
280 Content: content,
281 })
282 }
283 }
284 }
285 }
286
287 for _, t := range req.Tools {
288 var tool oaiTool
289 tool.Type = "function"
290 tool.Function.Name = t.Name
291 tool.Function.Description = t.Description
292 tool.Function.Parameters = t.Schema
293 out.Tools = append(out.Tools, tool)
294 }
295 if len(out.Tools) > 0 {
296 out.ToolChoice = "auto"
297 }
298
299 return out, nil
300}
301
302func buildOAIUserContent(blocks []Content, textOnly bool) interface{} {
303 hasImage := false
304 for _, b := range blocks {
305 if _, ok := b.(ImageBlock); ok {
306 hasImage = true
307 break
308 }
309 }
310 if textOnly || !hasImage {
311 var sb strings.Builder
312 for _, b := range blocks {
313 if tb, ok := b.(TextBlock); ok {
314 if sb.Len() > 0 {
315 sb.WriteString("\n")
316 }
317 sb.WriteString(tb.Text)
318 }
319 }
320 return sb.String()
321 }
322 return buildOAIContentBlocks(blocks, false)
323}
324
325func buildOAIToolContent(blocks []Content, isError, textOnly bool) interface{} {
326 hasImage := false
327 for _, b := range blocks {
328 if _, ok := b.(ImageBlock); ok {
329 hasImage = true
330 break
331 }
332 }
333 if textOnly || !hasImage {
334 var sb strings.Builder
335 for _, b := range blocks {
336 if tb, ok := b.(TextBlock); ok {
337 if sb.Len() > 0 {
338 sb.WriteString("\n")
339 }
340 sb.WriteString(tb.Text)
341 }
342 }
343 if isError && sb.Len() > 0 {
344 sb.WriteString(" [error]")
345 }
346 return sb.String()
347 }
348 return buildOAIContentBlocks(blocks, isError)
349}
350
351func buildOAIContentBlocks(blocks []Content, isError bool) []interface{} {
352 var arr []interface{}
353 for _, b := range blocks {
354 switch v := b.(type) {
355 case TextBlock:
356 arr = append(arr, oaiContentText{Type: "text", Text: v.Text})
357 case ImageBlock:
358 var img oaiContentImage
359 img.Type = "image_url"
360 img.ImageURL.URL = "data:" + v.MimeType + ";base64," + base64.StdEncoding.EncodeToString(v.Data)
361 arr = append(arr, img)
362 }
363 }
364 if isError {
365 arr = append(arr, oaiContentText{Type: "text", Text: "[error]"})
366 }
367 return arr
368}
369
370// ---- streaming ----
371
372func (c *openaiClient) Stream(ctx context.Context, req Request) (<-chan Event, error) {
373 apiPath := "/v1/chat/completions"
374 if strings.HasSuffix(c.baseURL, "/v1") {
375 apiPath = "/chat/completions"
376 }
377 wire, err := c.buildRequest(req)
378 if err != nil {
379 return nil, err
380 }
381 body, err := json.Marshal(wire)
382 if err != nil {
383 return nil, err
384 }
385 newReq := func() (*http.Request, error) {
386 httpReq, err := http.NewRequestWithContext(ctx, "POST", c.baseURL+apiPath, bytes.NewReader(body))
387 if err != nil {
388 return nil, err
389 }
390 httpReq.Header.Set("content-type", "application/json")
391 httpReq.Header.Set("accept", "text/event-stream")
392 httpReq.Header.Set("authorization", "Bearer "+c.apiKey)
393 for k, v := range c.headers {
394 httpReq.Header.Set(k, v)
395 }
396 return httpReq, nil
397 }
398
399 resp, err := doStreamWithRetry(ctx, c.http, newReq)
400 if err != nil {
401 return nil, fmt.Errorf("%s: %w", c.Name(), err)
402 }
403 if resp.StatusCode != http.StatusOK {
404 b, _ := io.ReadAll(resp.Body)
405 resp.Body.Close()
406 return nil, fmt.Errorf("%s: http %d: %s", c.Name(), resp.StatusCode, strings.TrimSpace(string(b)))
407 }
408
409 out := make(chan Event, 16)
410 go c.runStream(ctx, resp, req, out)
411 return out, nil
412}
413
414func (c *openaiClient) runStream(ctx context.Context, resp *http.Response, req Request, out chan<- Event) {
415 defer close(out)
416 defer resp.Body.Close()
417
418 model, _ := FindModel("", req.Model)
419 out <- EventStart{Model: req.Model, Provider: c.Name()}
420
421 raw := make(chan sseEvent, 16)
422 go readSSE(resp.Body, raw)
423
424 // Interleaved block tracking: text and tool_calls preserve their
425 // emission order so the assistant message renders in the same order
426 // the model produced it. The builder fires one kind of block at a
427 // time — incoming text deltas after a tool_call split into a fresh
428 // text block; subsequent tool_calls each get their own slot.
429 type blockEntry struct {
430 kind string // "text" | "tool_use"
431 textBuf strings.Builder
432 toolID string
433 toolName string
434 toolArgs strings.Builder
435 announced bool
436 }
437 var (
438 blocks []*blockEntry
439 currentText *blockEntry // most-recent text block, nil if none
440 toolByIdx = map[int]*blockEntry{} // openai tool_call index -> block
441 reasoningBuf strings.Builder
442 usage Usage
443 stop StopReason = StopEnd
444 finalErr error
445 )
446
447 appendText := func(delta string) {
448 if currentText == nil {
449 currentText = &blockEntry{kind: "text"}
450 blocks = append(blocks, currentText)
451 }
452 currentText.textBuf.WriteString(delta)
453 }
454
455 getOrCreateTool := func(idx int) *blockEntry {
456 if t, ok := toolByIdx[idx]; ok {
457 return t
458 }
459 t := &blockEntry{kind: "tool_use"}
460 toolByIdx[idx] = t
461 blocks = append(blocks, t)
462 // A new tool block breaks the current text block. Subsequent text
463 // deltas will start a fresh text block after this tool.
464 currentText = nil
465 return t
466 }
467
468 assembleMsg := func() Message {
469 content := []Content{}
470 for _, b := range blocks {
471 switch b.kind {
472 case "text":
473 if b.textBuf.Len() > 0 {
474 content = append(content, TextBlock{Text: b.textBuf.String()})
475 }
476 case "tool_use":
477 args := b.toolArgs.String()
478 if args == "" || !json.Valid([]byte(args)) {
479 args = "{}"
480 }
481 content = append(content, ToolCallBlock{
482 ID: b.toolID, Name: b.toolName, Arguments: json.RawMessage(args),
483 })
484 }
485 }
486 if reasoningBuf.Len() > 0 {
487 content = append(content, ReasoningBlock{Summary: reasoningBuf.String()})
488 }
489 return Message{Role: RoleAssistant, Content: content, Time: time.Now()}
490 }
491
492 sendDone := func() {
493 usage.CostUSD = ComputeCost(model, usage)
494 out <- EventUsage{Usage: usage}
495 out <- EventDone{Stop: stop, Err: finalErr, Message: assembleMsg()}
496 }
497
498 for {
499 select {
500 case <-ctx.Done():
501 stop = StopAborted
502 finalErr = ctx.Err()
503 sendDone()
504 return
505 case ev, ok := <-raw:
506 if !ok {
507 sendDone()
508 return
509 }
510 if ev.Data == "[DONE]" {
511 sendDone()
512 return
513 }
514 var chunk struct {
515 Choices []struct {
516 Index int `json:"index"`
517 Delta struct {
518 Content string `json:"content"`
519 ReasoningContent string `json:"reasoning_content"`
520 ToolCalls []struct {
521 Index int `json:"index"`
522 ID string `json:"id"`
523 Type string `json:"type"`
524 Function struct {
525 Name string `json:"name"`
526 Arguments string `json:"arguments"`
527 } `json:"function"`
528 } `json:"tool_calls"`
529 } `json:"delta"`
530 FinishReason string `json:"finish_reason"`
531 } `json:"choices"`
532 Usage *struct {
533 PromptTokens int `json:"prompt_tokens"`
534 CompletionTokens int `json:"completion_tokens"`
535 PromptTokensDetails struct {
536 CachedTokens int `json:"cached_tokens"`
537 } `json:"prompt_tokens_details"`
538 } `json:"usage"`
539 Error *struct {
540 Message string `json:"message"`
541 Type string `json:"type"`
542 } `json:"error"`
543 }
544 if err := json.Unmarshal([]byte(ev.Data), &chunk); err != nil {
545 continue
546 }
547 if chunk.Error != nil {
548 stop = StopError
549 finalErr = fmt.Errorf("openai: %s", chunk.Error.Message)
550 sendDone()
551 return
552 }
553 if chunk.Usage != nil {
554 usage.InputTokens = chunk.Usage.PromptTokens - chunk.Usage.PromptTokensDetails.CachedTokens
555 if usage.InputTokens < 0 {
556 usage.InputTokens = chunk.Usage.PromptTokens
557 }
558 usage.OutputTokens = chunk.Usage.CompletionTokens
559 usage.CacheReadTokens = chunk.Usage.PromptTokensDetails.CachedTokens
560 }
561 for _, ch := range chunk.Choices {
562 if ch.Delta.ReasoningContent != "" {
563 reasoningBuf.WriteString(ch.Delta.ReasoningContent)
564 }
565 if ch.Delta.Content != "" {
566 appendText(ch.Delta.Content)
567 out <- EventTextDelta{Delta: ch.Delta.Content}
568 }
569 for _, tc := range ch.Delta.ToolCalls {
570 t := getOrCreateTool(tc.Index)
571 if tc.ID != "" {
572 t.toolID = tc.ID
573 }
574 if tc.Function.Name != "" {
575 t.toolName = tc.Function.Name
576 }
577 if !t.announced && t.toolID != "" && t.toolName != "" {
578 t.announced = true
579 out <- EventToolStart{ID: t.toolID, Name: t.toolName}
580 }
581 if tc.Function.Arguments != "" {
582 t.toolArgs.WriteString(tc.Function.Arguments)
583 if t.announced {
584 out <- EventToolArgs{ID: t.toolID, Delta: tc.Function.Arguments}
585 }
586 }
587 }
588 switch ch.FinishReason {
589 case "stop":
590 stop = StopEnd
591 case "length":
592 stop = StopLength
593 case "tool_calls", "function_call":
594 stop = StopToolUse
595 for _, b := range blocks {
596 if b.kind == "tool_use" && b.announced {
597 out <- EventToolEnd{ID: b.toolID}
598 }
599 }
600 }
601 }
602 }
603 }
604}