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/log"
17 "github.com/charmbracelet/crush/internal/message"
18 "github.com/google/uuid"
19 "github.com/openai/openai-go"
20 "github.com/openai/openai-go/option"
21 "github.com/openai/openai-go/packages/param"
22 "github.com/openai/openai-go/shared"
23)
24
25type openaiClient struct {
26 providerOptions providerClientOptions
27 client openai.Client
28}
29
30type OpenAIClient ProviderClient
31
32func newOpenAIClient(opts providerClientOptions) OpenAIClient {
33 return &openaiClient{
34 providerOptions: opts,
35 client: createOpenAIClient(opts),
36 }
37}
38
39func createOpenAIClient(opts providerClientOptions) openai.Client {
40 openaiClientOptions := []option.RequestOption{}
41 if opts.apiKey != "" {
42 openaiClientOptions = append(openaiClientOptions, option.WithAPIKey(opts.apiKey))
43 }
44 if opts.baseURL != "" {
45 resolvedBaseURL, err := opts.cfg.Resolve(opts.baseURL)
46 if err == nil && resolvedBaseURL != "" {
47 openaiClientOptions = append(openaiClientOptions, option.WithBaseURL(resolvedBaseURL))
48 }
49 }
50
51 if opts.cfg.Options.Debug {
52 httpClient := log.NewHTTPClient()
53 openaiClientOptions = append(openaiClientOptions, option.WithHTTPClient(httpClient))
54 }
55
56 for key, value := range opts.extraHeaders {
57 openaiClientOptions = append(openaiClientOptions, option.WithHeader(key, value))
58 }
59
60 for extraKey, extraValue := range opts.extraBody {
61 openaiClientOptions = append(openaiClientOptions, option.WithJSONSet(extraKey, extraValue))
62 }
63
64 return openai.NewClient(openaiClientOptions...)
65}
66
67func (o *openaiClient) convertMessages(messages []message.Message) (openaiMessages []openai.ChatCompletionMessageParamUnion) {
68 isAnthropicModel := o.providerOptions.config.ID == string(catwalk.InferenceProviderOpenRouter) && strings.HasPrefix(o.Model().ID, "anthropic/")
69 // Add system message first
70 systemMessage := o.providerOptions.systemMessage
71 if o.providerOptions.systemPromptPrefix != "" {
72 systemMessage = o.providerOptions.systemPromptPrefix + "\n" + systemMessage
73 }
74
75 system := openai.SystemMessage(systemMessage)
76 if isAnthropicModel && !o.providerOptions.disableCache {
77 systemTextBlock := openai.ChatCompletionContentPartTextParam{Text: systemMessage}
78 systemTextBlock.SetExtraFields(
79 map[string]any{
80 "cache_control": map[string]string{
81 "type": "ephemeral",
82 },
83 },
84 )
85 var content []openai.ChatCompletionContentPartTextParam
86 content = append(content, systemTextBlock)
87 system = openai.SystemMessage(content)
88 }
89 openaiMessages = append(openaiMessages, system)
90
91 for i, msg := range messages {
92 cache := false
93 if i > len(messages)-3 {
94 cache = true
95 }
96 switch msg.Role {
97 case message.User:
98 var content []openai.ChatCompletionContentPartUnionParam
99
100 textBlock := openai.ChatCompletionContentPartTextParam{Text: msg.Content().String()}
101 content = append(content, openai.ChatCompletionContentPartUnionParam{OfText: &textBlock})
102 hasBinaryContent := false
103 for _, binaryContent := range msg.BinaryContent() {
104 hasBinaryContent = true
105 imageURL := openai.ChatCompletionContentPartImageImageURLParam{URL: binaryContent.String(catwalk.InferenceProviderOpenAI)}
106 imageBlock := openai.ChatCompletionContentPartImageParam{ImageURL: imageURL}
107
108 content = append(content, openai.ChatCompletionContentPartUnionParam{OfImageURL: &imageBlock})
109 }
110 if cache && !o.providerOptions.disableCache && isAnthropicModel {
111 textBlock.SetExtraFields(map[string]any{
112 "cache_control": map[string]string{
113 "type": "ephemeral",
114 },
115 })
116 }
117 if hasBinaryContent || (isAnthropicModel && !o.providerOptions.disableCache) {
118 openaiMessages = append(openaiMessages, openai.UserMessage(content))
119 } else {
120 openaiMessages = append(openaiMessages, openai.UserMessage(msg.Content().String()))
121 }
122
123 case message.Assistant:
124 assistantMsg := openai.ChatCompletionAssistantMessageParam{
125 Role: "assistant",
126 }
127
128 // Only include finished tool calls; interrupted tool calls must not be resent.
129 if len(msg.ToolCalls()) > 0 {
130 finished := make([]message.ToolCall, 0, len(msg.ToolCalls()))
131 for _, call := range msg.ToolCalls() {
132 if call.Finished {
133 finished = append(finished, call)
134 }
135 }
136 if len(finished) > 0 {
137 assistantMsg.ToolCalls = make([]openai.ChatCompletionMessageToolCallParam, len(finished))
138 for i, call := range finished {
139 assistantMsg.ToolCalls[i] = openai.ChatCompletionMessageToolCallParam{
140 ID: call.ID,
141 Type: "function",
142 Function: openai.ChatCompletionMessageToolCallFunctionParam{
143 Name: call.Name,
144 Arguments: call.Input,
145 },
146 }
147 }
148 }
149 }
150 if msg.Content().String() != "" {
151 assistantMsg.Content = openai.ChatCompletionAssistantMessageParamContentUnion{
152 OfString: param.NewOpt(msg.Content().Text),
153 }
154 }
155
156 if cache && !o.providerOptions.disableCache && isAnthropicModel {
157 assistantMsg.SetExtraFields(map[string]any{
158 "cache_control": map[string]string{
159 "type": "ephemeral",
160 },
161 })
162 }
163 // Skip empty assistant messages (no content and no finished tool calls)
164 if msg.Content().String() == "" && len(assistantMsg.ToolCalls) == 0 {
165 continue
166 }
167
168 openaiMessages = append(openaiMessages, openai.ChatCompletionMessageParamUnion{
169 OfAssistant: &assistantMsg,
170 })
171
172 case message.Tool:
173 for _, result := range msg.ToolResults() {
174 openaiMessages = append(openaiMessages,
175 openai.ToolMessage(result.Content, result.ToolCallID),
176 )
177 }
178 }
179 }
180
181 return openaiMessages
182}
183
184func (o *openaiClient) convertTools(tools []tools.BaseTool) []openai.ChatCompletionToolParam {
185 openaiTools := make([]openai.ChatCompletionToolParam, len(tools))
186
187 for i, tool := range tools {
188 info := tool.Info()
189 openaiTools[i] = openai.ChatCompletionToolParam{
190 Function: openai.FunctionDefinitionParam{
191 Name: info.Name,
192 Description: openai.String(info.Description),
193 Parameters: openai.FunctionParameters{
194 "type": "object",
195 "properties": info.Parameters,
196 "required": info.Required,
197 },
198 },
199 }
200 }
201
202 return openaiTools
203}
204
205func (o *openaiClient) finishReason(reason string) message.FinishReason {
206 switch reason {
207 case "stop":
208 return message.FinishReasonEndTurn
209 case "length":
210 return message.FinishReasonMaxTokens
211 case "tool_calls":
212 return message.FinishReasonToolUse
213 default:
214 return message.FinishReasonUnknown
215 }
216}
217
218func (o *openaiClient) preparedParams(messages []openai.ChatCompletionMessageParamUnion, tools []openai.ChatCompletionToolParam) openai.ChatCompletionNewParams {
219 model := o.providerOptions.model(o.providerOptions.modelType)
220 cfg := o.providerOptions.cfg
221
222 modelConfig := cfg.Models[config.SelectedModelTypeLarge]
223 if o.providerOptions.modelType == config.SelectedModelTypeSmall {
224 modelConfig = cfg.Models[config.SelectedModelTypeSmall]
225 }
226
227 reasoningEffort := modelConfig.ReasoningEffort
228
229 params := openai.ChatCompletionNewParams{
230 Model: openai.ChatModel(model.ID),
231 Messages: messages,
232 Tools: tools,
233 }
234
235 maxTokens := model.DefaultMaxTokens
236 if modelConfig.MaxTokens > 0 {
237 maxTokens = modelConfig.MaxTokens
238 }
239
240 // Override max tokens if set in provider options
241 if o.providerOptions.maxTokens > 0 {
242 maxTokens = o.providerOptions.maxTokens
243 }
244 if model.CanReason {
245 params.MaxCompletionTokens = openai.Int(maxTokens)
246 switch reasoningEffort {
247 case "low":
248 params.ReasoningEffort = shared.ReasoningEffortLow
249 case "medium":
250 params.ReasoningEffort = shared.ReasoningEffortMedium
251 case "high":
252 params.ReasoningEffort = shared.ReasoningEffortHigh
253 case "minimal":
254 params.ReasoningEffort = shared.ReasoningEffort("minimal")
255 default:
256 params.ReasoningEffort = shared.ReasoningEffort(reasoningEffort)
257 }
258 } else {
259 params.MaxTokens = openai.Int(maxTokens)
260 }
261
262 return params
263}
264
265func (o *openaiClient) send(ctx context.Context, messages []message.Message, tools []tools.BaseTool) (response *ProviderResponse, err error) {
266 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
267 attempts := 0
268 for {
269 attempts++
270 openaiResponse, err := o.client.Chat.Completions.New(
271 ctx,
272 params,
273 )
274 // If there is an error we are going to see if we can retry the call
275 if err != nil {
276 retry, after, retryErr := o.shouldRetry(attempts, err)
277 if retryErr != nil {
278 return nil, retryErr
279 }
280 if retry {
281 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries, "error", err)
282 select {
283 case <-ctx.Done():
284 return nil, ctx.Err()
285 case <-time.After(time.Duration(after) * time.Millisecond):
286 continue
287 }
288 }
289 return nil, retryErr
290 }
291
292 if len(openaiResponse.Choices) == 0 {
293 return nil, fmt.Errorf("received empty response from OpenAI API - check endpoint configuration")
294 }
295
296 content := ""
297 if openaiResponse.Choices[0].Message.Content != "" {
298 content = openaiResponse.Choices[0].Message.Content
299 }
300
301 toolCalls := o.toolCalls(*openaiResponse)
302 finishReason := o.finishReason(string(openaiResponse.Choices[0].FinishReason))
303
304 if len(toolCalls) > 0 {
305 finishReason = message.FinishReasonToolUse
306 }
307
308 return &ProviderResponse{
309 Content: content,
310 ToolCalls: toolCalls,
311 Usage: o.usage(*openaiResponse),
312 FinishReason: finishReason,
313 }, nil
314 }
315}
316
317func (o *openaiClient) stream(ctx context.Context, messages []message.Message, tools []tools.BaseTool) <-chan ProviderEvent {
318 params := o.preparedParams(o.convertMessages(messages), o.convertTools(tools))
319 params.StreamOptions = openai.ChatCompletionStreamOptionsParam{
320 IncludeUsage: openai.Bool(true),
321 }
322
323 attempts := 0
324 eventChan := make(chan ProviderEvent)
325
326 go func() {
327 for {
328 attempts++
329 // Kujtim: fixes an issue with anthropig models on openrouter
330 if len(params.Tools) == 0 {
331 params.Tools = nil
332 }
333 openaiStream := o.client.Chat.Completions.NewStreaming(
334 ctx,
335 params,
336 )
337
338 acc := openai.ChatCompletionAccumulator{}
339 currentContent := ""
340 toolCalls := make([]message.ToolCall, 0)
341 msgToolCalls := make(map[int64]openai.ChatCompletionMessageToolCall)
342 toolMap := make(map[string]openai.ChatCompletionMessageToolCall)
343 toolCallIDMap := make(map[string]string)
344 for openaiStream.Next() {
345 chunk := openaiStream.Current()
346 // Kujtim: this is an issue with openrouter qwen, its sending -1 for the tool index
347 if len(chunk.Choices) != 0 && len(chunk.Choices[0].Delta.ToolCalls) > 0 && chunk.Choices[0].Delta.ToolCalls[0].Index == -1 {
348 chunk.Choices[0].Delta.ToolCalls[0].Index = 0
349 }
350 acc.AddChunk(chunk)
351 for i, choice := range chunk.Choices {
352 reasoning, ok := choice.Delta.JSON.ExtraFields["reasoning"]
353 if ok && reasoning.Raw() != "" {
354 reasoningStr := ""
355 json.Unmarshal([]byte(reasoning.Raw()), &reasoningStr)
356 if reasoningStr != "" {
357 eventChan <- ProviderEvent{
358 Type: EventThinkingDelta,
359 Thinking: reasoningStr,
360 }
361 }
362 }
363 if choice.Delta.Content != "" {
364 eventChan <- ProviderEvent{
365 Type: EventContentDelta,
366 Content: choice.Delta.Content,
367 }
368 currentContent += choice.Delta.Content
369 } else if len(choice.Delta.ToolCalls) > 0 {
370 toolCall := choice.Delta.ToolCalls[0]
371 if strings.HasPrefix(toolCall.ID, "functions.") {
372 exID, ok := toolCallIDMap[toolCall.ID]
373 if !ok {
374 newID := uuid.NewString()
375 toolCallIDMap[toolCall.ID] = newID
376 toolCall.ID = newID
377 } else {
378 toolCall.ID = exID
379 }
380 }
381 newToolCall := false
382 if existingToolCall, ok := msgToolCalls[toolCall.Index]; ok { // tool call exists
383 if toolCall.ID != "" && toolCall.ID != existingToolCall.ID {
384 found := false
385 // try to find the tool based on the ID
386 for _, tool := range msgToolCalls {
387 if tool.ID == toolCall.ID {
388 existingToolCall.Function.Arguments += toolCall.Function.Arguments
389 msgToolCalls[toolCall.Index] = existingToolCall
390 toolMap[existingToolCall.ID] = existingToolCall
391 found = true
392 }
393 }
394 if !found {
395 newToolCall = true
396 }
397 } else {
398 existingToolCall.Function.Arguments += toolCall.Function.Arguments
399 msgToolCalls[toolCall.Index] = existingToolCall
400 toolMap[existingToolCall.ID] = existingToolCall
401 }
402 } else {
403 newToolCall = true
404 }
405 if newToolCall { // new tool call
406 if toolCall.ID == "" {
407 toolCall.ID = uuid.NewString()
408 }
409 eventChan <- ProviderEvent{
410 Type: EventToolUseStart,
411 ToolCall: &message.ToolCall{
412 ID: toolCall.ID,
413 Name: toolCall.Function.Name,
414 Finished: false,
415 },
416 }
417 msgToolCalls[toolCall.Index] = openai.ChatCompletionMessageToolCall{
418 ID: toolCall.ID,
419 Type: "function",
420 Function: openai.ChatCompletionMessageToolCallFunction{
421 Name: toolCall.Function.Name,
422 Arguments: toolCall.Function.Arguments,
423 },
424 }
425 toolMap[toolCall.ID] = msgToolCalls[toolCall.Index]
426 }
427 toolCalls := []openai.ChatCompletionMessageToolCall{}
428 for _, tc := range toolMap {
429 toolCalls = append(toolCalls, tc)
430 }
431 acc.Choices[i].Message.ToolCalls = toolCalls
432 }
433 }
434 }
435
436 err := openaiStream.Err()
437 if err == nil || errors.Is(err, io.EOF) {
438 if len(acc.Choices) == 0 {
439 eventChan <- ProviderEvent{
440 Type: EventError,
441 Error: fmt.Errorf("received empty streaming response from OpenAI API - check endpoint configuration"),
442 }
443 return
444 }
445
446 resultFinishReason := acc.Choices[0].FinishReason
447 if resultFinishReason == "" {
448 // If the finish reason is empty, we assume it was a successful completion
449 // INFO: this is happening for openrouter for some reason
450 resultFinishReason = "stop"
451 }
452 // Stream completed successfully
453 finishReason := o.finishReason(resultFinishReason)
454 if len(acc.Choices[0].Message.ToolCalls) > 0 {
455 toolCalls = append(toolCalls, o.toolCalls(acc.ChatCompletion)...)
456 }
457 if len(toolCalls) > 0 {
458 finishReason = message.FinishReasonToolUse
459 }
460
461 eventChan <- ProviderEvent{
462 Type: EventComplete,
463 Response: &ProviderResponse{
464 Content: currentContent,
465 ToolCalls: toolCalls,
466 Usage: o.usage(acc.ChatCompletion),
467 FinishReason: finishReason,
468 },
469 }
470 close(eventChan)
471 return
472 }
473
474 // If there is an error we are going to see if we can retry the call
475 retry, after, retryErr := o.shouldRetry(attempts, err)
476 if retryErr != nil {
477 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
478 close(eventChan)
479 return
480 }
481 if retry {
482 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries, "error", err)
483 select {
484 case <-ctx.Done():
485 // context cancelled
486 if ctx.Err() != nil {
487 eventChan <- ProviderEvent{Type: EventError, Error: ctx.Err()}
488 }
489 close(eventChan)
490 return
491 case <-time.After(time.Duration(after) * time.Millisecond):
492 continue
493 }
494 }
495 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
496 close(eventChan)
497 return
498 }
499 }()
500
501 return eventChan
502}
503
504func (o *openaiClient) shouldRetry(attempts int, err error) (bool, int64, error) {
505 if attempts > maxRetries {
506 return false, 0, fmt.Errorf("maximum retry attempts reached for rate limit: %d retries", maxRetries)
507 }
508 if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
509 return false, 0, err
510 }
511 var apiErr *openai.Error
512 retryMs := 0
513 retryAfterValues := []string{}
514 if errors.As(err, &apiErr) {
515 // Check for token expiration (401 Unauthorized)
516 if apiErr.StatusCode == 401 {
517 o.providerOptions.apiKey, err = o.providerOptions.cfg.Resolve(o.providerOptions.config.APIKey)
518 if err != nil {
519 return false, 0, fmt.Errorf("failed to resolve API key: %w", err)
520 }
521 o.client = createOpenAIClient(o.providerOptions)
522 return true, 0, nil
523 }
524
525 if apiErr.StatusCode != 429 && apiErr.StatusCode != 500 {
526 return false, 0, err
527 }
528
529 retryAfterValues = apiErr.Response.Header.Values("Retry-After")
530 }
531
532 if apiErr != nil {
533 slog.Warn("OpenAI API error", "status_code", apiErr.StatusCode, "message", apiErr.Message, "type", apiErr.Type)
534 if len(retryAfterValues) > 0 {
535 slog.Warn("Retry-After header", "values", retryAfterValues)
536 }
537 } else {
538 slog.Error("OpenAI API error", "error", err.Error(), "attempt", attempts, "max_retries", maxRetries)
539 }
540
541 backoffMs := 2000 * (1 << (attempts - 1))
542 jitterMs := int(float64(backoffMs) * 0.2)
543 retryMs = backoffMs + jitterMs
544 if len(retryAfterValues) > 0 {
545 if _, err := fmt.Sscanf(retryAfterValues[0], "%d", &retryMs); err == nil {
546 retryMs = retryMs * 1000
547 }
548 }
549 return true, int64(retryMs), nil
550}
551
552func (o *openaiClient) toolCalls(completion openai.ChatCompletion) []message.ToolCall {
553 var toolCalls []message.ToolCall
554
555 if len(completion.Choices) > 0 && len(completion.Choices[0].Message.ToolCalls) > 0 {
556 for _, call := range completion.Choices[0].Message.ToolCalls {
557 // accumulator for some reason does this.
558 if call.Function.Name == "" {
559 continue
560 }
561 toolCall := message.ToolCall{
562 ID: call.ID,
563 Name: call.Function.Name,
564 Input: call.Function.Arguments,
565 Type: "function",
566 Finished: true,
567 }
568 toolCalls = append(toolCalls, toolCall)
569 }
570 }
571
572 return toolCalls
573}
574
575func (o *openaiClient) usage(completion openai.ChatCompletion) TokenUsage {
576 cachedTokens := completion.Usage.PromptTokensDetails.CachedTokens
577 inputTokens := completion.Usage.PromptTokens - cachedTokens
578
579 return TokenUsage{
580 InputTokens: inputTokens,
581 OutputTokens: completion.Usage.CompletionTokens,
582 CacheCreationTokens: 0, // OpenAI doesn't provide this directly
583 CacheReadTokens: cachedTokens,
584 }
585}
586
587func (o *openaiClient) Model() catwalk.Model {
588 return o.providerOptions.model(o.providerOptions.modelType)
589}