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