1package provider
2
3import (
4 "context"
5 "encoding/json"
6 "errors"
7 "fmt"
8 "io"
9 "log/slog"
10 "slices"
11 "strings"
12 "time"
13
14 "github.com/charmbracelet/catwalk/pkg/catwalk"
15 "github.com/charmbracelet/crush/internal/config"
16 "github.com/charmbracelet/crush/internal/llm/tools"
17 "github.com/charmbracelet/crush/internal/log"
18 "github.com/charmbracelet/crush/internal/message"
19 "github.com/google/uuid"
20 "github.com/openai/openai-go"
21 "github.com/openai/openai-go/option"
22 "github.com/openai/openai-go/packages/param"
23 "github.com/openai/openai-go/shared"
24)
25
26type openaiClient struct {
27 providerOptions providerClientOptions
28 client openai.Client
29}
30
31type OpenAIClient ProviderClient
32
33func newOpenAIClient(opts providerClientOptions) OpenAIClient {
34 return &openaiClient{
35 providerOptions: opts,
36 client: createOpenAIClient(opts),
37 }
38}
39
40func createOpenAIClient(opts providerClientOptions) openai.Client {
41 openaiClientOptions := []option.RequestOption{}
42 if opts.apiKey != "" {
43 openaiClientOptions = append(openaiClientOptions, option.WithAPIKey(opts.apiKey))
44 }
45 if opts.baseURL != "" {
46 resolvedBaseURL, err := config.Get().Resolve(opts.baseURL)
47 if err == nil {
48 openaiClientOptions = append(openaiClientOptions, option.WithBaseURL(resolvedBaseURL))
49 }
50 }
51
52 if config.Get().Options.Debug {
53 httpClient := log.NewHTTPClient()
54 openaiClientOptions = append(openaiClientOptions, option.WithHTTPClient(httpClient))
55 }
56
57 for key, value := range opts.extraHeaders {
58 openaiClientOptions = append(openaiClientOptions, option.WithHeader(key, value))
59 }
60
61 for extraKey, extraValue := range opts.extraBody {
62 openaiClientOptions = append(openaiClientOptions, option.WithJSONSet(extraKey, extraValue))
63 }
64
65 return openai.NewClient(openaiClientOptions...)
66}
67
68func (o *openaiClient) convertMessages(messages []message.Message) (openaiMessages []openai.ChatCompletionMessageParamUnion) {
69 isAnthropicModel := o.providerOptions.config.ID == string(catwalk.InferenceProviderOpenRouter) && strings.HasPrefix(o.Model().ID, "anthropic/")
70 // Add system message first
71 systemMessage := o.providerOptions.systemMessage
72 if o.providerOptions.systemPromptPrefix != "" {
73 systemMessage = o.providerOptions.systemPromptPrefix + "\n" + systemMessage
74 }
75
76 system := openai.SystemMessage(systemMessage)
77 if isAnthropicModel && !o.providerOptions.disableCache {
78 systemTextBlock := openai.ChatCompletionContentPartTextParam{Text: systemMessage}
79 systemTextBlock.SetExtraFields(
80 map[string]any{
81 "cache_control": map[string]string{
82 "type": "ephemeral",
83 },
84 },
85 )
86 var content []openai.ChatCompletionContentPartTextParam
87 content = append(content, systemTextBlock)
88 system = openai.SystemMessage(content)
89 }
90 openaiMessages = append(openaiMessages, system)
91
92 for i, msg := range messages {
93 cache := false
94 if i > len(messages)-3 {
95 cache = true
96 }
97 switch msg.Role {
98 case message.User:
99 var content []openai.ChatCompletionContentPartUnionParam
100
101 textBlock := openai.ChatCompletionContentPartTextParam{Text: msg.Content().String()}
102 content = append(content, openai.ChatCompletionContentPartUnionParam{OfText: &textBlock})
103 hasBinaryContent := false
104 for _, binaryContent := range msg.BinaryContent() {
105 hasBinaryContent = true
106 imageURL := openai.ChatCompletionContentPartImageImageURLParam{URL: binaryContent.String(catwalk.InferenceProviderOpenAI)}
107 imageBlock := openai.ChatCompletionContentPartImageParam{ImageURL: imageURL}
108
109 content = append(content, openai.ChatCompletionContentPartUnionParam{OfImageURL: &imageBlock})
110 }
111 if cache && !o.providerOptions.disableCache && isAnthropicModel {
112 textBlock.SetExtraFields(map[string]any{
113 "cache_control": map[string]string{
114 "type": "ephemeral",
115 },
116 })
117 }
118 if hasBinaryContent || (isAnthropicModel && !o.providerOptions.disableCache) {
119 openaiMessages = append(openaiMessages, openai.UserMessage(content))
120 } else {
121 openaiMessages = append(openaiMessages, openai.UserMessage(msg.Content().String()))
122 }
123
124 case message.Assistant:
125 assistantMsg := openai.ChatCompletionAssistantMessageParam{
126 Role: "assistant",
127 }
128
129 hasContent := false
130 if msg.Content().String() != "" {
131 hasContent = true
132 textBlock := openai.ChatCompletionContentPartTextParam{Text: msg.Content().String()}
133 if cache && !o.providerOptions.disableCache && isAnthropicModel {
134 textBlock.SetExtraFields(map[string]any{
135 "cache_control": map[string]string{
136 "type": "ephemeral",
137 },
138 })
139 }
140 assistantMsg.Content = openai.ChatCompletionAssistantMessageParamContentUnion{
141 OfArrayOfContentParts: []openai.ChatCompletionAssistantMessageParamContentArrayOfContentPartUnion{
142 {
143 OfText: &textBlock,
144 },
145 },
146 }
147 if !isAnthropicModel {
148 assistantMsg.Content = openai.ChatCompletionAssistantMessageParamContentUnion{
149 OfString: param.NewOpt(msg.Content().String()),
150 }
151 }
152 }
153
154 if len(msg.ToolCalls()) > 0 {
155 hasContent = true
156 assistantMsg.ToolCalls = make([]openai.ChatCompletionMessageToolCallParam, len(msg.ToolCalls()))
157 for i, call := range msg.ToolCalls() {
158 assistantMsg.ToolCalls[i] = openai.ChatCompletionMessageToolCallParam{
159 ID: call.ID,
160 Type: "function",
161 Function: openai.ChatCompletionMessageToolCallFunctionParam{
162 Name: call.Name,
163 Arguments: call.Input,
164 },
165 }
166 }
167 }
168 if !hasContent {
169 slog.Warn("There is a message without content, investigate, this should not happen")
170 continue
171 }
172
173 openaiMessages = append(openaiMessages, openai.ChatCompletionMessageParamUnion{
174 OfAssistant: &assistantMsg,
175 })
176
177 case message.Tool:
178 for _, result := range msg.ToolResults() {
179 openaiMessages = append(openaiMessages,
180 openai.ToolMessage(result.Content, result.ToolCallID),
181 )
182 }
183 }
184 }
185
186 return
187}
188
189func (o *openaiClient) convertTools(tools []tools.BaseTool) []openai.ChatCompletionToolParam {
190 openaiTools := make([]openai.ChatCompletionToolParam, len(tools))
191
192 for i, tool := range tools {
193 info := tool.Info()
194 openaiTools[i] = openai.ChatCompletionToolParam{
195 Function: openai.FunctionDefinitionParam{
196 Name: info.Name,
197 Description: openai.String(info.Description),
198 Parameters: openai.FunctionParameters{
199 "type": "object",
200 "properties": info.Parameters,
201 "required": info.Required,
202 },
203 },
204 }
205 }
206
207 return openaiTools
208}
209
210func (o *openaiClient) finishReason(reason string) message.FinishReason {
211 switch reason {
212 case "stop":
213 return message.FinishReasonEndTurn
214 case "length":
215 return message.FinishReasonMaxTokens
216 case "tool_calls":
217 return message.FinishReasonToolUse
218 default:
219 return message.FinishReasonUnknown
220 }
221}
222
223func (o *openaiClient) preparedParams(messages []openai.ChatCompletionMessageParamUnion, tools []openai.ChatCompletionToolParam) openai.ChatCompletionNewParams {
224 model := o.providerOptions.model(o.providerOptions.modelType)
225 cfg := config.Get()
226
227 modelConfig := cfg.Models[config.SelectedModelTypeLarge]
228 if o.providerOptions.modelType == config.SelectedModelTypeSmall {
229 modelConfig = cfg.Models[config.SelectedModelTypeSmall]
230 }
231
232 reasoningEffort := modelConfig.ReasoningEffort
233
234 params := openai.ChatCompletionNewParams{
235 Model: openai.ChatModel(model.ID),
236 Messages: messages,
237 Tools: tools,
238 }
239
240 maxTokens := model.DefaultMaxTokens
241 if modelConfig.MaxTokens > 0 {
242 maxTokens = modelConfig.MaxTokens
243 }
244
245 // Override max tokens if set in provider options
246 if o.providerOptions.maxTokens > 0 {
247 maxTokens = o.providerOptions.maxTokens
248 }
249 if model.CanReason {
250 params.MaxCompletionTokens = openai.Int(maxTokens)
251 switch reasoningEffort {
252 case "low":
253 params.ReasoningEffort = shared.ReasoningEffortLow
254 case "medium":
255 params.ReasoningEffort = shared.ReasoningEffortMedium
256 case "high":
257 params.ReasoningEffort = shared.ReasoningEffortHigh
258 default:
259 params.ReasoningEffort = shared.ReasoningEffort(reasoningEffort)
260 }
261 } else {
262 params.MaxTokens = openai.Int(maxTokens)
263 }
264
265 return params
266}
267
268func (o *openaiClient) send(ctx context.Context, messages []message.Message, tools []tools.BaseTool) (response *ProviderResponse, err error) {
269 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
270 attempts := 0
271 for {
272 attempts++
273 openaiResponse, err := o.client.Chat.Completions.New(
274 ctx,
275 params,
276 )
277 // If there is an error we are going to see if we can retry the call
278 if err != nil {
279 retry, after, retryErr := o.shouldRetry(attempts, err)
280 if retryErr != nil {
281 return nil, retryErr
282 }
283 if retry {
284 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries)
285 select {
286 case <-ctx.Done():
287 return nil, ctx.Err()
288 case <-time.After(time.Duration(after) * time.Millisecond):
289 continue
290 }
291 }
292 return nil, retryErr
293 }
294
295 if len(openaiResponse.Choices) == 0 {
296 return nil, fmt.Errorf("received empty response from OpenAI API - check endpoint configuration")
297 }
298
299 content := ""
300 if openaiResponse.Choices[0].Message.Content != "" {
301 content = openaiResponse.Choices[0].Message.Content
302 }
303
304 toolCalls := o.toolCalls(*openaiResponse)
305 finishReason := o.finishReason(string(openaiResponse.Choices[0].FinishReason))
306
307 if len(toolCalls) > 0 {
308 finishReason = message.FinishReasonToolUse
309 }
310
311 return &ProviderResponse{
312 Content: content,
313 ToolCalls: toolCalls,
314 Usage: o.usage(*openaiResponse),
315 FinishReason: finishReason,
316 }, nil
317 }
318}
319
320func (o *openaiClient) stream(ctx context.Context, messages []message.Message, tools []tools.BaseTool) <-chan ProviderEvent {
321 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
322 params.StreamOptions = openai.ChatCompletionStreamOptionsParam{
323 IncludeUsage: openai.Bool(true),
324 }
325
326 attempts := 0
327 eventChan := make(chan ProviderEvent)
328
329 go func() {
330 for {
331 attempts++
332 // Kujtim: fixes an issue with anthropig models on openrouter
333 if len(params.Tools) == 0 {
334 params.Tools = nil
335 }
336 openaiStream := o.client.Chat.Completions.NewStreaming(
337 ctx,
338 params,
339 )
340
341 acc := openai.ChatCompletionAccumulator{}
342 currentContent := ""
343 toolCalls := make([]message.ToolCall, 0)
344 var msgToolCalls []openai.ChatCompletionMessageToolCall
345 for openaiStream.Next() {
346 chunk := openaiStream.Current()
347 // Kujtim: this is an issue with openrouter qwen, its sending -1 for the tool index
348 if len(chunk.Choices) > 0 && len(chunk.Choices[0].Delta.ToolCalls) > 0 && chunk.Choices[0].Delta.ToolCalls[0].Index == -1 {
349 chunk.Choices[0].Delta.ToolCalls[0].Index = 0
350 }
351 acc.AddChunk(chunk)
352 for i, choice := range chunk.Choices {
353 reasoning, ok := choice.Delta.JSON.ExtraFields["reasoning"]
354 if ok && reasoning.Raw() != "" {
355 reasoningStr := ""
356 json.Unmarshal([]byte(reasoning.Raw()), &reasoningStr)
357 if reasoningStr != "" {
358 eventChan <- ProviderEvent{
359 Type: EventThinkingDelta,
360 Thinking: reasoningStr,
361 }
362 }
363 }
364 if choice.Delta.Content != "" {
365 eventChan <- ProviderEvent{
366 Type: EventContentDelta,
367 Content: choice.Delta.Content,
368 }
369 currentContent += choice.Delta.Content
370 } else if len(choice.Delta.ToolCalls) > 0 {
371 toolCall := choice.Delta.ToolCalls[0]
372 newToolCall := false
373 if len(msgToolCalls)-1 >= int(toolCall.Index) { // tool call exists
374 existingToolCall := msgToolCalls[toolCall.Index]
375 if toolCall.ID != "" && toolCall.ID != existingToolCall.ID {
376 found := false
377 // try to find the tool based on the ID
378 for i, tool := range msgToolCalls {
379 if tool.ID == toolCall.ID {
380 msgToolCalls[i].Function.Arguments += toolCall.Function.Arguments
381 found = true
382 }
383 }
384 if !found {
385 newToolCall = true
386 }
387 } else {
388 msgToolCalls[toolCall.Index].Function.Arguments += toolCall.Function.Arguments
389 }
390 } else {
391 newToolCall = true
392 }
393 if newToolCall { // new tool call
394 if toolCall.ID == "" {
395 toolCall.ID = uuid.NewString()
396 }
397 eventChan <- ProviderEvent{
398 Type: EventToolUseStart,
399 ToolCall: &message.ToolCall{
400 ID: toolCall.ID,
401 Name: toolCall.Function.Name,
402 Finished: false,
403 },
404 }
405 msgToolCalls = append(msgToolCalls, openai.ChatCompletionMessageToolCall{
406 ID: toolCall.ID,
407 Type: "function",
408 Function: openai.ChatCompletionMessageToolCallFunction{
409 Name: toolCall.Function.Name,
410 Arguments: toolCall.Function.Arguments,
411 },
412 })
413 }
414 }
415 acc.Choices[i].Message.ToolCalls = slices.Clone(msgToolCalls)
416 }
417 }
418
419 err := openaiStream.Err()
420 if err == nil || errors.Is(err, io.EOF) {
421 if len(acc.Choices) == 0 {
422 eventChan <- ProviderEvent{
423 Type: EventError,
424 Error: fmt.Errorf("received empty streaming response from OpenAI API - check endpoint configuration"),
425 }
426 return
427 }
428
429 resultFinishReason := acc.Choices[0].FinishReason
430 if resultFinishReason == "" {
431 // If the finish reason is empty, we assume it was a successful completion
432 // INFO: this is happening for openrouter for some reason
433 resultFinishReason = "stop"
434 }
435 // Stream completed successfully
436 finishReason := o.finishReason(resultFinishReason)
437 if len(acc.Choices[0].Message.ToolCalls) > 0 {
438 toolCalls = append(toolCalls, o.toolCalls(acc.ChatCompletion)...)
439 }
440 if len(toolCalls) > 0 {
441 finishReason = message.FinishReasonToolUse
442 }
443
444 eventChan <- ProviderEvent{
445 Type: EventComplete,
446 Response: &ProviderResponse{
447 Content: currentContent,
448 ToolCalls: toolCalls,
449 Usage: o.usage(acc.ChatCompletion),
450 FinishReason: finishReason,
451 },
452 }
453 close(eventChan)
454 return
455 }
456
457 // If there is an error we are going to see if we can retry the call
458 retry, after, retryErr := o.shouldRetry(attempts, err)
459 if retryErr != nil {
460 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
461 close(eventChan)
462 return
463 }
464 if retry {
465 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries)
466 select {
467 case <-ctx.Done():
468 // context cancelled
469 if ctx.Err() == nil {
470 eventChan <- ProviderEvent{Type: EventError, Error: ctx.Err()}
471 }
472 close(eventChan)
473 return
474 case <-time.After(time.Duration(after) * time.Millisecond):
475 continue
476 }
477 }
478 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
479 close(eventChan)
480 return
481 }
482 }()
483
484 return eventChan
485}
486
487func (o *openaiClient) shouldRetry(attempts int, err error) (bool, int64, error) {
488 if attempts > maxRetries {
489 return false, 0, fmt.Errorf("maximum retry attempts reached for rate limit: %d retries", maxRetries)
490 }
491 if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
492 return false, 0, err
493 }
494 var apiErr *openai.Error
495 retryMs := 0
496 retryAfterValues := []string{}
497 if errors.As(err, &apiErr) {
498 // Check for token expiration (401 Unauthorized)
499 if apiErr.StatusCode == 401 {
500 o.providerOptions.apiKey, err = config.Get().Resolve(o.providerOptions.config.APIKey)
501 if err != nil {
502 return false, 0, fmt.Errorf("failed to resolve API key: %w", err)
503 }
504 o.client = createOpenAIClient(o.providerOptions)
505 return true, 0, nil
506 }
507
508 if apiErr.StatusCode != 429 && apiErr.StatusCode != 500 {
509 return false, 0, err
510 }
511
512 retryAfterValues = apiErr.Response.Header.Values("Retry-After")
513 }
514
515 if apiErr != nil {
516 slog.Warn("OpenAI API error", "status_code", apiErr.StatusCode, "message", apiErr.Message, "type", apiErr.Type)
517 if len(retryAfterValues) > 0 {
518 slog.Warn("Retry-After header", "values", retryAfterValues)
519 }
520 } else {
521 slog.Error("OpenAI API error", "error", err.Error(), "attempt", attempts, "max_retries", maxRetries)
522 }
523
524 backoffMs := 2000 * (1 << (attempts - 1))
525 jitterMs := int(float64(backoffMs) * 0.2)
526 retryMs = backoffMs + jitterMs
527 if len(retryAfterValues) > 0 {
528 if _, err := fmt.Sscanf(retryAfterValues[0], "%d", &retryMs); err == nil {
529 retryMs = retryMs * 1000
530 }
531 }
532 return true, int64(retryMs), nil
533}
534
535func (o *openaiClient) toolCalls(completion openai.ChatCompletion) []message.ToolCall {
536 var toolCalls []message.ToolCall
537
538 if len(completion.Choices) > 0 && len(completion.Choices[0].Message.ToolCalls) > 0 {
539 for _, call := range completion.Choices[0].Message.ToolCalls {
540 toolCall := message.ToolCall{
541 ID: call.ID,
542 Name: call.Function.Name,
543 Input: call.Function.Arguments,
544 Type: "function",
545 Finished: true,
546 }
547 toolCalls = append(toolCalls, toolCall)
548 }
549 }
550
551 return toolCalls
552}
553
554func (o *openaiClient) usage(completion openai.ChatCompletion) TokenUsage {
555 cachedTokens := completion.Usage.PromptTokensDetails.CachedTokens
556 inputTokens := completion.Usage.PromptTokens - cachedTokens
557
558 return TokenUsage{
559 InputTokens: inputTokens,
560 OutputTokens: completion.Usage.CompletionTokens,
561 CacheCreationTokens: 0, // OpenAI doesn't provide this directly
562 CacheReadTokens: cachedTokens,
563 }
564}
565
566func (o *openaiClient) Model() catwalk.Model {
567 return o.providerOptions.model(o.providerOptions.modelType)
568}