1package provider
2
3import (
4 "context"
5 "encoding/json"
6 "errors"
7 "fmt"
8 "io"
9 "log/slog"
10 "regexp"
11 "strconv"
12 "strings"
13 "time"
14
15 "github.com/anthropics/anthropic-sdk-go"
16 "github.com/anthropics/anthropic-sdk-go/bedrock"
17 "github.com/anthropics/anthropic-sdk-go/option"
18 "github.com/anthropics/anthropic-sdk-go/vertex"
19 "github.com/charmbracelet/catwalk/pkg/catwalk"
20 "github.com/charmbracelet/crush/internal/config"
21 "github.com/charmbracelet/crush/internal/llm/tools"
22 "github.com/charmbracelet/crush/internal/log"
23 "github.com/charmbracelet/crush/internal/message"
24)
25
26// Pre-compiled regex for parsing context limit errors.
27var contextLimitRegex = regexp.MustCompile(`input length and ` + "`max_tokens`" + ` exceed context limit: (\d+) \+ (\d+) > (\d+)`)
28
29type anthropicClient struct {
30 providerOptions providerClientOptions
31 tp AnthropicClientType
32 client anthropic.Client
33 adjustedMaxTokens int // Used when context limit is hit
34}
35
36type AnthropicClient ProviderClient
37
38type AnthropicClientType string
39
40const (
41 AnthropicClientTypeNormal AnthropicClientType = "normal"
42 AnthropicClientTypeBedrock AnthropicClientType = "bedrock"
43 AnthropicClientTypeVertex AnthropicClientType = "vertex"
44)
45
46func newAnthropicClient(opts providerClientOptions, tp AnthropicClientType) AnthropicClient {
47 return &anthropicClient{
48 providerOptions: opts,
49 tp: tp,
50 client: createAnthropicClient(opts, tp),
51 }
52}
53
54func createAnthropicClient(opts providerClientOptions, tp AnthropicClientType) anthropic.Client {
55 anthropicClientOptions := []option.RequestOption{}
56
57 // Check if Authorization header is provided in extra headers
58 hasBearerAuth := false
59 if opts.extraHeaders != nil {
60 for key := range opts.extraHeaders {
61 if strings.ToLower(key) == "authorization" {
62 hasBearerAuth = true
63 break
64 }
65 }
66 }
67
68 isBearerToken := strings.HasPrefix(opts.apiKey, "Bearer ")
69
70 if opts.apiKey != "" && !hasBearerAuth {
71 if isBearerToken {
72 slog.Debug("API key starts with 'Bearer ', using as Authorization header")
73 anthropicClientOptions = append(anthropicClientOptions, option.WithHeader("Authorization", opts.apiKey))
74 } else {
75 // Use standard X-Api-Key header
76 anthropicClientOptions = append(anthropicClientOptions, option.WithAPIKey(opts.apiKey))
77 }
78 } else if hasBearerAuth {
79 slog.Debug("Skipping X-Api-Key header because Authorization header is provided")
80 }
81
82 if config.Get().Options.Debug {
83 httpClient := log.NewHTTPClient()
84 anthropicClientOptions = append(anthropicClientOptions, option.WithHTTPClient(httpClient))
85 }
86
87 switch tp {
88 case AnthropicClientTypeBedrock:
89 anthropicClientOptions = append(anthropicClientOptions, bedrock.WithLoadDefaultConfig(context.Background()))
90 case AnthropicClientTypeVertex:
91 project := opts.extraParams["project"]
92 location := opts.extraParams["location"]
93 anthropicClientOptions = append(anthropicClientOptions, vertex.WithGoogleAuth(context.Background(), location, project))
94 }
95 for key, header := range opts.extraHeaders {
96 anthropicClientOptions = append(anthropicClientOptions, option.WithHeaderAdd(key, header))
97 }
98 for key, value := range opts.extraBody {
99 anthropicClientOptions = append(anthropicClientOptions, option.WithJSONSet(key, value))
100 }
101 return anthropic.NewClient(anthropicClientOptions...)
102}
103
104func (a *anthropicClient) convertMessages(messages []message.Message) (anthropicMessages []anthropic.MessageParam) {
105 for i, msg := range messages {
106 cache := false
107 if i > len(messages)-3 {
108 cache = true
109 }
110 switch msg.Role {
111 case message.User:
112 content := anthropic.NewTextBlock(msg.Content().String())
113 if cache && !a.providerOptions.disableCache {
114 content.OfText.CacheControl = anthropic.CacheControlEphemeralParam{
115 Type: "ephemeral",
116 }
117 }
118 var contentBlocks []anthropic.ContentBlockParamUnion
119 contentBlocks = append(contentBlocks, content)
120 for _, binaryContent := range msg.BinaryContent() {
121 base64Image := binaryContent.String(catwalk.InferenceProviderAnthropic)
122 imageBlock := anthropic.NewImageBlockBase64(binaryContent.MIMEType, base64Image)
123 contentBlocks = append(contentBlocks, imageBlock)
124 }
125 anthropicMessages = append(anthropicMessages, anthropic.NewUserMessage(contentBlocks...))
126
127 case message.Assistant:
128 blocks := []anthropic.ContentBlockParamUnion{}
129
130 // Add thinking blocks first if present (required when thinking is enabled with tool use)
131 if reasoningContent := msg.ReasoningContent(); reasoningContent.Thinking != "" {
132 thinkingBlock := anthropic.NewThinkingBlock(reasoningContent.Signature, reasoningContent.Thinking)
133 blocks = append(blocks, thinkingBlock)
134 }
135
136 if msg.Content().String() != "" {
137 content := anthropic.NewTextBlock(msg.Content().String())
138 if cache && !a.providerOptions.disableCache {
139 content.OfText.CacheControl = anthropic.CacheControlEphemeralParam{
140 Type: "ephemeral",
141 }
142 }
143 blocks = append(blocks, content)
144 }
145
146 for _, toolCall := range msg.ToolCalls() {
147 var inputMap map[string]any
148 err := json.Unmarshal([]byte(toolCall.Input), &inputMap)
149 if err != nil {
150 continue
151 }
152 blocks = append(blocks, anthropic.NewToolUseBlock(toolCall.ID, inputMap, toolCall.Name))
153 }
154
155 if len(blocks) == 0 {
156 continue
157 }
158 anthropicMessages = append(anthropicMessages, anthropic.NewAssistantMessage(blocks...))
159
160 case message.Tool:
161 results := make([]anthropic.ContentBlockParamUnion, len(msg.ToolResults()))
162 for i, toolResult := range msg.ToolResults() {
163 results[i] = anthropic.NewToolResultBlock(toolResult.ToolCallID, toolResult.Content, toolResult.IsError)
164 }
165 anthropicMessages = append(anthropicMessages, anthropic.NewUserMessage(results...))
166 }
167 }
168 return
169}
170
171func (a *anthropicClient) convertTools(tools []tools.BaseTool) []anthropic.ToolUnionParam {
172 anthropicTools := make([]anthropic.ToolUnionParam, len(tools))
173
174 for i, tool := range tools {
175 info := tool.Info()
176 toolParam := anthropic.ToolParam{
177 Name: info.Name,
178 Description: anthropic.String(info.Description),
179 InputSchema: anthropic.ToolInputSchemaParam{
180 Properties: info.Parameters,
181 // TODO: figure out how we can tell claude the required fields?
182 },
183 }
184
185 if i == len(tools)-1 && !a.providerOptions.disableCache {
186 toolParam.CacheControl = anthropic.CacheControlEphemeralParam{
187 Type: "ephemeral",
188 }
189 }
190
191 anthropicTools[i] = anthropic.ToolUnionParam{OfTool: &toolParam}
192 }
193
194 return anthropicTools
195}
196
197func (a *anthropicClient) finishReason(reason string) message.FinishReason {
198 switch reason {
199 case "end_turn":
200 return message.FinishReasonEndTurn
201 case "max_tokens":
202 return message.FinishReasonMaxTokens
203 case "tool_use":
204 return message.FinishReasonToolUse
205 case "stop_sequence":
206 return message.FinishReasonEndTurn
207 default:
208 return message.FinishReasonUnknown
209 }
210}
211
212func (a *anthropicClient) isThinkingEnabled() bool {
213 cfg := config.Get()
214 modelConfig := cfg.Models[config.SelectedModelTypeLarge]
215 if a.providerOptions.modelType == config.SelectedModelTypeSmall {
216 modelConfig = cfg.Models[config.SelectedModelTypeSmall]
217 }
218 return a.Model().CanReason && modelConfig.Think
219}
220
221func (a *anthropicClient) preparedMessages(messages []anthropic.MessageParam, tools []anthropic.ToolUnionParam) anthropic.MessageNewParams {
222 model := a.providerOptions.model(a.providerOptions.modelType)
223 var thinkingParam anthropic.ThinkingConfigParamUnion
224 cfg := config.Get()
225 modelConfig := cfg.Models[config.SelectedModelTypeLarge]
226 if a.providerOptions.modelType == config.SelectedModelTypeSmall {
227 modelConfig = cfg.Models[config.SelectedModelTypeSmall]
228 }
229 temperature := anthropic.Float(0)
230
231 maxTokens := model.DefaultMaxTokens
232 if modelConfig.MaxTokens > 0 {
233 maxTokens = modelConfig.MaxTokens
234 }
235 if a.isThinkingEnabled() {
236 thinkingParam = anthropic.ThinkingConfigParamOfEnabled(int64(float64(maxTokens) * 0.8))
237 temperature = anthropic.Float(1)
238 }
239 // Override max tokens if set in provider options
240 if a.providerOptions.maxTokens > 0 {
241 maxTokens = a.providerOptions.maxTokens
242 }
243
244 // Use adjusted max tokens if context limit was hit
245 if a.adjustedMaxTokens > 0 {
246 maxTokens = int64(a.adjustedMaxTokens)
247 }
248
249 systemBlocks := []anthropic.TextBlockParam{}
250
251 // Add custom system prompt prefix if configured
252 if a.providerOptions.systemPromptPrefix != "" {
253 systemBlocks = append(systemBlocks, anthropic.TextBlockParam{
254 Text: a.providerOptions.systemPromptPrefix,
255 })
256 }
257
258 systemBlocks = append(systemBlocks, anthropic.TextBlockParam{
259 Text: a.providerOptions.systemMessage,
260 CacheControl: anthropic.CacheControlEphemeralParam{
261 Type: "ephemeral",
262 },
263 })
264
265 return anthropic.MessageNewParams{
266 Model: anthropic.Model(model.ID),
267 MaxTokens: maxTokens,
268 Temperature: temperature,
269 Messages: messages,
270 Tools: tools,
271 Thinking: thinkingParam,
272 System: systemBlocks,
273 }
274}
275
276func (a *anthropicClient) send(ctx context.Context, messages []message.Message, tools []tools.BaseTool) (response *ProviderResponse, err error) {
277 attempts := 0
278 for {
279 attempts++
280 // Prepare messages on each attempt in case max_tokens was adjusted
281 preparedMessages := a.preparedMessages(a.convertMessages(messages), a.convertTools(tools))
282
283 var opts []option.RequestOption
284 if a.isThinkingEnabled() {
285 opts = append(opts, option.WithHeaderAdd("anthropic-beta", "interleaved-thinking-2025-05-14"))
286 }
287 anthropicResponse, err := a.client.Messages.New(
288 ctx,
289 preparedMessages,
290 opts...,
291 )
292 // If there is an error we are going to see if we can retry the call
293 if err != nil {
294 retry, after, retryErr := a.shouldRetry(attempts, err)
295 if retryErr != nil {
296 return nil, retryErr
297 }
298 if retry {
299 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries, "error", err)
300 select {
301 case <-ctx.Done():
302 return nil, ctx.Err()
303 case <-time.After(time.Duration(after) * time.Millisecond):
304 continue
305 }
306 }
307 return nil, retryErr
308 }
309
310 content := ""
311 for _, block := range anthropicResponse.Content {
312 if text, ok := block.AsAny().(anthropic.TextBlock); ok {
313 content += text.Text
314 }
315 }
316
317 return &ProviderResponse{
318 Content: content,
319 ToolCalls: a.toolCalls(*anthropicResponse),
320 Usage: a.usage(*anthropicResponse),
321 }, nil
322 }
323}
324
325func (a *anthropicClient) stream(ctx context.Context, messages []message.Message, tools []tools.BaseTool) <-chan ProviderEvent {
326 attempts := 0
327 eventChan := make(chan ProviderEvent)
328 go func() {
329 for {
330 attempts++
331 // Prepare messages on each attempt in case max_tokens was adjusted
332 preparedMessages := a.preparedMessages(a.convertMessages(messages), a.convertTools(tools))
333
334 var opts []option.RequestOption
335 if a.isThinkingEnabled() {
336 opts = append(opts, option.WithHeaderAdd("anthropic-beta", "interleaved-thinking-2025-05-14"))
337 }
338
339 anthropicStream := a.client.Messages.NewStreaming(
340 ctx,
341 preparedMessages,
342 opts...,
343 )
344 accumulatedMessage := anthropic.Message{}
345
346 currentToolCallID := ""
347 for anthropicStream.Next() {
348 event := anthropicStream.Current()
349 err := accumulatedMessage.Accumulate(event)
350 if err != nil {
351 slog.Warn("Error accumulating message", "error", err)
352 continue
353 }
354
355 switch event := event.AsAny().(type) {
356 case anthropic.ContentBlockStartEvent:
357 switch event.ContentBlock.Type {
358 case "text":
359 eventChan <- ProviderEvent{Type: EventContentStart}
360 case "tool_use":
361 currentToolCallID = event.ContentBlock.ID
362 eventChan <- ProviderEvent{
363 Type: EventToolUseStart,
364 ToolCall: &message.ToolCall{
365 ID: event.ContentBlock.ID,
366 Name: event.ContentBlock.Name,
367 Finished: false,
368 },
369 }
370 }
371
372 case anthropic.ContentBlockDeltaEvent:
373 if event.Delta.Type == "thinking_delta" && event.Delta.Thinking != "" {
374 eventChan <- ProviderEvent{
375 Type: EventThinkingDelta,
376 Thinking: event.Delta.Thinking,
377 }
378 } else if event.Delta.Type == "signature_delta" && event.Delta.Signature != "" {
379 eventChan <- ProviderEvent{
380 Type: EventSignatureDelta,
381 Signature: event.Delta.Signature,
382 }
383 } else if event.Delta.Type == "text_delta" && event.Delta.Text != "" {
384 eventChan <- ProviderEvent{
385 Type: EventContentDelta,
386 Content: event.Delta.Text,
387 }
388 } else if event.Delta.Type == "input_json_delta" {
389 if currentToolCallID != "" {
390 eventChan <- ProviderEvent{
391 Type: EventToolUseDelta,
392 ToolCall: &message.ToolCall{
393 ID: currentToolCallID,
394 Finished: false,
395 Input: event.Delta.PartialJSON,
396 },
397 }
398 }
399 }
400 case anthropic.ContentBlockStopEvent:
401 if currentToolCallID != "" {
402 eventChan <- ProviderEvent{
403 Type: EventToolUseStop,
404 ToolCall: &message.ToolCall{
405 ID: currentToolCallID,
406 },
407 }
408 currentToolCallID = ""
409 } else {
410 eventChan <- ProviderEvent{Type: EventContentStop}
411 }
412
413 case anthropic.MessageStopEvent:
414 content := ""
415 for _, block := range accumulatedMessage.Content {
416 if text, ok := block.AsAny().(anthropic.TextBlock); ok {
417 content += text.Text
418 }
419 }
420
421 eventChan <- ProviderEvent{
422 Type: EventComplete,
423 Response: &ProviderResponse{
424 Content: content,
425 ToolCalls: a.toolCalls(accumulatedMessage),
426 Usage: a.usage(accumulatedMessage),
427 FinishReason: a.finishReason(string(accumulatedMessage.StopReason)),
428 },
429 Content: content,
430 }
431 }
432 }
433
434 err := anthropicStream.Err()
435 if err == nil || errors.Is(err, io.EOF) {
436 close(eventChan)
437 return
438 }
439
440 // If there is an error we are going to see if we can retry the call
441 retry, after, retryErr := a.shouldRetry(attempts, err)
442 if retryErr != nil {
443 eventChan <- ProviderEvent{Type: EventError, Error: retryErr}
444 close(eventChan)
445 return
446 }
447 if retry {
448 slog.Warn("Retrying due to rate limit", "attempt", attempts, "max_retries", maxRetries, "error", err)
449 select {
450 case <-ctx.Done():
451 // context cancelled
452 if ctx.Err() != nil {
453 eventChan <- ProviderEvent{Type: EventError, Error: ctx.Err()}
454 }
455 close(eventChan)
456 return
457 case <-time.After(time.Duration(after) * time.Millisecond):
458 continue
459 }
460 }
461 if ctx.Err() != nil {
462 eventChan <- ProviderEvent{Type: EventError, Error: ctx.Err()}
463 }
464
465 close(eventChan)
466 return
467 }
468 }()
469 return eventChan
470}
471
472func (a *anthropicClient) shouldRetry(attempts int, err error) (bool, int64, error) {
473 var apiErr *anthropic.Error
474 if !errors.As(err, &apiErr) {
475 return false, 0, err
476 }
477
478 if attempts > maxRetries {
479 return false, 0, fmt.Errorf("maximum retry attempts reached for rate limit: %d retries", maxRetries)
480 }
481
482 if apiErr.StatusCode == 401 {
483 a.providerOptions.apiKey, err = config.Get().Resolve(a.providerOptions.config.APIKey)
484 if err != nil {
485 return false, 0, fmt.Errorf("failed to resolve API key: %w", err)
486 }
487 a.client = createAnthropicClient(a.providerOptions, a.tp)
488 return true, 0, nil
489 }
490
491 // Handle context limit exceeded error (400 Bad Request)
492 if apiErr.StatusCode == 400 {
493 if adjusted, ok := a.handleContextLimitError(apiErr); ok {
494 a.adjustedMaxTokens = adjusted
495 slog.Debug("Adjusted max_tokens due to context limit", "new_max_tokens", adjusted)
496 return true, 0, nil
497 }
498 }
499
500 isOverloaded := strings.Contains(apiErr.Error(), "overloaded") || strings.Contains(apiErr.Error(), "rate limit exceeded")
501 if apiErr.StatusCode != 429 && apiErr.StatusCode != 529 && !isOverloaded {
502 return false, 0, err
503 }
504
505 retryMs := 0
506 retryAfterValues := apiErr.Response.Header.Values("Retry-After")
507
508 backoffMs := 2000 * (1 << (attempts - 1))
509 jitterMs := int(float64(backoffMs) * 0.2)
510 retryMs = backoffMs + jitterMs
511 if len(retryAfterValues) > 0 {
512 if _, err := fmt.Sscanf(retryAfterValues[0], "%d", &retryMs); err == nil {
513 retryMs = retryMs * 1000
514 }
515 }
516 return true, int64(retryMs), nil
517}
518
519// handleContextLimitError parses context limit error and returns adjusted max_tokens
520func (a *anthropicClient) handleContextLimitError(apiErr *anthropic.Error) (int, bool) {
521 // Parse error message like: "input length and max_tokens exceed context limit: 154978 + 50000 > 200000"
522 errorMsg := apiErr.Error()
523
524 matches := contextLimitRegex.FindStringSubmatch(errorMsg)
525
526 if len(matches) != 4 {
527 return 0, false
528 }
529
530 inputTokens, err1 := strconv.Atoi(matches[1])
531 contextLimit, err2 := strconv.Atoi(matches[3])
532
533 if err1 != nil || err2 != nil {
534 return 0, false
535 }
536
537 // Calculate safe max_tokens with a buffer of 1000 tokens
538 safeMaxTokens := contextLimit - inputTokens - 1000
539
540 // Ensure we don't go below a minimum threshold
541 safeMaxTokens = max(safeMaxTokens, 1000)
542
543 return safeMaxTokens, true
544}
545
546func (a *anthropicClient) toolCalls(msg anthropic.Message) []message.ToolCall {
547 var toolCalls []message.ToolCall
548
549 for _, block := range msg.Content {
550 switch variant := block.AsAny().(type) {
551 case anthropic.ToolUseBlock:
552 toolCall := message.ToolCall{
553 ID: variant.ID,
554 Name: variant.Name,
555 Input: string(variant.Input),
556 Type: string(variant.Type),
557 Finished: true,
558 }
559 toolCalls = append(toolCalls, toolCall)
560 }
561 }
562
563 return toolCalls
564}
565
566func (a *anthropicClient) usage(msg anthropic.Message) TokenUsage {
567 return TokenUsage{
568 InputTokens: msg.Usage.InputTokens,
569 OutputTokens: msg.Usage.OutputTokens,
570 CacheCreationTokens: msg.Usage.CacheCreationInputTokens,
571 CacheReadTokens: msg.Usage.CacheReadInputTokens,
572 }
573}
574
575func (a *anthropicClient) Model() catwalk.Model {
576 return a.providerOptions.model(a.providerOptions.modelType)
577}