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 case "minimal":
259 params.ReasoningEffort = shared.ReasoningEffort("minimal")
260 default:
261 params.ReasoningEffort = shared.ReasoningEffort(reasoningEffort)
262 }
263 } else {
264 params.MaxTokens = openai.Int(maxTokens)
265 }
266
267 return params
268}
269
270func (o *openaiClient) send(ctx context.Context, messages []message.Message, tools []tools.BaseTool) (response *ProviderResponse, err error) {
271 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
272 attempts := 0
273 for {
274 attempts++
275 openaiResponse, err := o.client.Chat.Completions.New(
276 ctx,
277 params,
278 )
279 // If there is an error we are going to see if we can retry the call
280 if err != nil {
281 retry, after, retryErr := o.shouldRetry(attempts, err)
282 if retryErr != nil {
283 return nil, retryErr
284 }
285 if retry {
286 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries)
287 select {
288 case <-ctx.Done():
289 return nil, ctx.Err()
290 case <-time.After(time.Duration(after) * time.Millisecond):
291 continue
292 }
293 }
294 return nil, retryErr
295 }
296
297 if len(openaiResponse.Choices) == 0 {
298 return nil, fmt.Errorf("received empty response from OpenAI API - check endpoint configuration")
299 }
300
301 content := ""
302 if openaiResponse.Choices[0].Message.Content != "" {
303 content = openaiResponse.Choices[0].Message.Content
304 }
305
306 toolCalls := o.toolCalls(*openaiResponse)
307 finishReason := o.finishReason(string(openaiResponse.Choices[0].FinishReason))
308
309 if len(toolCalls) > 0 {
310 finishReason = message.FinishReasonToolUse
311 }
312
313 return &ProviderResponse{
314 Content: content,
315 ToolCalls: toolCalls,
316 Usage: o.usage(*openaiResponse),
317 FinishReason: finishReason,
318 }, nil
319 }
320}
321
322func (o *openaiClient) stream(ctx context.Context, messages []message.Message, tools []tools.BaseTool) <-chan ProviderEvent {
323 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
324 params.StreamOptions = openai.ChatCompletionStreamOptionsParam{
325 IncludeUsage: openai.Bool(true),
326 }
327
328 attempts := 0
329 eventChan := make(chan ProviderEvent)
330
331 go func() {
332 for {
333 attempts++
334 // Kujtim: fixes an issue with anthropig models on openrouter
335 if len(params.Tools) == 0 {
336 params.Tools = nil
337 }
338 openaiStream := o.client.Chat.Completions.NewStreaming(
339 ctx,
340 params,
341 )
342
343 acc := openai.ChatCompletionAccumulator{}
344 currentContent := ""
345 toolCalls := make([]message.ToolCall, 0)
346 var msgToolCalls []openai.ChatCompletionMessageToolCall
347 for openaiStream.Next() {
348 chunk := openaiStream.Current()
349 // Kujtim: this is an issue with openrouter qwen, its sending -1 for the tool index
350 if len(chunk.Choices) > 0 && len(chunk.Choices[0].Delta.ToolCalls) > 0 && chunk.Choices[0].Delta.ToolCalls[0].Index == -1 {
351 chunk.Choices[0].Delta.ToolCalls[0].Index = 0
352 }
353 acc.AddChunk(chunk)
354 for i, choice := range chunk.Choices {
355 reasoning, ok := choice.Delta.JSON.ExtraFields["reasoning"]
356 if ok && reasoning.Raw() != "" {
357 reasoningStr := ""
358 json.Unmarshal([]byte(reasoning.Raw()), &reasoningStr)
359 if reasoningStr != "" {
360 eventChan <- ProviderEvent{
361 Type: EventThinkingDelta,
362 Thinking: reasoningStr,
363 }
364 }
365 }
366 if choice.Delta.Content != "" {
367 eventChan <- ProviderEvent{
368 Type: EventContentDelta,
369 Content: choice.Delta.Content,
370 }
371 currentContent += choice.Delta.Content
372 } else if len(choice.Delta.ToolCalls) > 0 {
373 toolCall := choice.Delta.ToolCalls[0]
374 newToolCall := false
375 if len(msgToolCalls)-1 >= int(toolCall.Index) { // tool call exists
376 existingToolCall := msgToolCalls[toolCall.Index]
377 if toolCall.ID != "" && toolCall.ID != existingToolCall.ID {
378 found := false
379 // try to find the tool based on the ID
380 for i, tool := range msgToolCalls {
381 if tool.ID == toolCall.ID {
382 msgToolCalls[i].Function.Arguments += toolCall.Function.Arguments
383 found = true
384 }
385 }
386 if !found {
387 newToolCall = true
388 }
389 } else {
390 msgToolCalls[toolCall.Index].Function.Arguments += toolCall.Function.Arguments
391 }
392 } else {
393 newToolCall = true
394 }
395 if newToolCall { // new tool call
396 if toolCall.ID == "" {
397 toolCall.ID = uuid.NewString()
398 }
399 eventChan <- ProviderEvent{
400 Type: EventToolUseStart,
401 ToolCall: &message.ToolCall{
402 ID: toolCall.ID,
403 Name: toolCall.Function.Name,
404 Finished: false,
405 },
406 }
407 msgToolCalls = append(msgToolCalls, openai.ChatCompletionMessageToolCall{
408 ID: toolCall.ID,
409 Type: "function",
410 Function: openai.ChatCompletionMessageToolCallFunction{
411 Name: toolCall.Function.Name,
412 Arguments: toolCall.Function.Arguments,
413 },
414 })
415 }
416 }
417 acc.Choices[i].Message.ToolCalls = slices.Clone(msgToolCalls)
418 }
419 }
420
421 err := openaiStream.Err()
422 if err == nil || errors.Is(err, io.EOF) {
423 if len(acc.Choices) == 0 {
424 eventChan <- ProviderEvent{
425 Type: EventError,
426 Error: fmt.Errorf("received empty streaming response from OpenAI API - check endpoint configuration"),
427 }
428 return
429 }
430
431 resultFinishReason := acc.Choices[0].FinishReason
432 if resultFinishReason == "" {
433 // If the finish reason is empty, we assume it was a successful completion
434 // INFO: this is happening for openrouter for some reason
435 resultFinishReason = "stop"
436 }
437 // Stream completed successfully
438 finishReason := o.finishReason(resultFinishReason)
439 if len(acc.Choices[0].Message.ToolCalls) > 0 {
440 toolCalls = append(toolCalls, o.toolCalls(acc.ChatCompletion)...)
441 }
442 if len(toolCalls) > 0 {
443 finishReason = message.FinishReasonToolUse
444 }
445
446 eventChan <- ProviderEvent{
447 Type: EventComplete,
448 Response: &ProviderResponse{
449 Content: currentContent,
450 ToolCalls: toolCalls,
451 Usage: o.usage(acc.ChatCompletion),
452 FinishReason: finishReason,
453 },
454 }
455 close(eventChan)
456 return
457 }
458
459 // If there is an error we are going to see if we can retry the call
460 retry, after, retryErr := o.shouldRetry(attempts, err)
461 if retryErr != nil {
462 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
463 close(eventChan)
464 return
465 }
466 if retry {
467 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries)
468 select {
469 case <-ctx.Done():
470 // context cancelled
471 if ctx.Err() == nil {
472 eventChan <- ProviderEvent{Type: EventError, Error: ctx.Err()}
473 }
474 close(eventChan)
475 return
476 case <-time.After(time.Duration(after) * time.Millisecond):
477 continue
478 }
479 }
480 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
481 close(eventChan)
482 return
483 }
484 }()
485
486 return eventChan
487}
488
489func (o *openaiClient) shouldRetry(attempts int, err error) (bool, int64, error) {
490 if attempts > maxRetries {
491 return false, 0, fmt.Errorf("maximum retry attempts reached for rate limit: %d retries", maxRetries)
492 }
493 if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
494 return false, 0, err
495 }
496 var apiErr *openai.Error
497 retryMs := 0
498 retryAfterValues := []string{}
499 if errors.As(err, &apiErr) {
500 // Check for token expiration (401 Unauthorized)
501 if apiErr.StatusCode == 401 {
502 o.providerOptions.apiKey, err = config.Get().Resolve(o.providerOptions.config.APIKey)
503 if err != nil {
504 return false, 0, fmt.Errorf("failed to resolve API key: %w", err)
505 }
506 o.client = createOpenAIClient(o.providerOptions)
507 return true, 0, nil
508 }
509
510 if apiErr.StatusCode != 429 && apiErr.StatusCode != 500 {
511 return false, 0, err
512 }
513
514 retryAfterValues = apiErr.Response.Header.Values("Retry-After")
515 }
516
517 if apiErr != nil {
518 slog.Warn("OpenAI API error", "status_code", apiErr.StatusCode, "message", apiErr.Message, "type", apiErr.Type)
519 if len(retryAfterValues) > 0 {
520 slog.Warn("Retry-After header", "values", retryAfterValues)
521 }
522 } else {
523 slog.Error("OpenAI API error", "error", err.Error(), "attempt", attempts, "max_retries", maxRetries)
524 }
525
526 backoffMs := 2000 * (1 << (attempts - 1))
527 jitterMs := int(float64(backoffMs) * 0.2)
528 retryMs = backoffMs + jitterMs
529 if len(retryAfterValues) > 0 {
530 if _, err := fmt.Sscanf(retryAfterValues[0], "%d", &retryMs); err == nil {
531 retryMs = retryMs * 1000
532 }
533 }
534 return true, int64(retryMs), nil
535}
536
537func (o *openaiClient) toolCalls(completion openai.ChatCompletion) []message.ToolCall {
538 var toolCalls []message.ToolCall
539
540 if len(completion.Choices) > 0 && len(completion.Choices[0].Message.ToolCalls) > 0 {
541 for _, call := range completion.Choices[0].Message.ToolCalls {
542 toolCall := message.ToolCall{
543 ID: call.ID,
544 Name: call.Function.Name,
545 Input: call.Function.Arguments,
546 Type: "function",
547 Finished: true,
548 }
549 toolCalls = append(toolCalls, toolCall)
550 }
551 }
552
553 return toolCalls
554}
555
556func (o *openaiClient) usage(completion openai.ChatCompletion) TokenUsage {
557 cachedTokens := completion.Usage.PromptTokensDetails.CachedTokens
558 inputTokens := completion.Usage.PromptTokens - cachedTokens
559
560 return TokenUsage{
561 InputTokens: inputTokens,
562 OutputTokens: completion.Usage.CompletionTokens,
563 CacheCreationTokens: 0, // OpenAI doesn't provide this directly
564 CacheReadTokens: cachedTokens,
565 }
566}
567
568func (o *openaiClient) Model() catwalk.Model {
569 return o.providerOptions.model(o.providerOptions.modelType)
570}