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