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