Files
CLIProxyAPI/internal/runtime/executor/claude_executor_execute.go
sususu a2933c7737 fix(claude): scope Anthropic beta and count_tokens policies
Claude Code builds Anthropic-Beta per request instead of sending a fixed list.
Captured from an isolated 2.1.220 profile pointed at api.anthropic.com through
a local proxy, over two rounds covering 11 model IDs and the [1m] variants:

  constant  claude-code, interleaved-thinking, redact-thinking,
            thinking-token-count, context-management, prompt-caching-scope
  tools     advanced-tool-use-2025-11-20 only when tools are declared
  model     mid-conversation-system-2026-04-07 only on models that accept a
            role=system turn
  [1m]      context-1m-2025-08-07, directly after claude-code-20250219 rather
            than at the end
  trailing  effort-2025-11-24, then server-side-fallback-2026-06-01

claude-sonnet-5 emits mid-conversation-system-2026-04-07, so it accepts a
role=system turn and must not sit in the legacy reminder whitelist.

count_tokens does not reuse the inference fingerprint. Running /context in an
interactive session issues 37 identical calls, which made the endpoint
observable for the first time: four betas only, and 21 headers rather than 22
because X-Stainless-Timeout is absent. The profile is selected from the request
path so no call site has to thread another flag.
2026-08-03 14:47:26 +08:00

246 lines
9.6 KiB
Go

package executor
import (
"bytes"
"context"
"fmt"
"io"
"net/http"
"github.com/router-for-me/CLIProxyAPI/v7/internal/runtime/executor/helps"
"github.com/router-for-me/CLIProxyAPI/v7/internal/thinking"
cliproxyauth "github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/auth"
cliproxyexecutor "github.com/router-for-me/CLIProxyAPI/v7/sdk/cliproxy/executor"
sdktranslator "github.com/router-for-me/CLIProxyAPI/v7/sdk/translator"
log "github.com/sirupsen/logrus"
)
func (e *ClaudeExecutor) Execute(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (resp cliproxyexecutor.Response, err error) {
if opts.Alt == "responses/compact" {
return resp, statusErr{code: http.StatusNotImplemented, msg: "/responses/compact not supported"}
}
baseModel := thinking.ParseSuffix(req.Model).ModelName
upstreamModel := e.upstreamModel(baseModel)
apiKey, baseURL := claudeCreds(auth)
if baseURL == "" {
baseURL = "https://api.anthropic.com"
}
url := fmt.Sprintf("%s/v1/messages?beta=true", baseURL)
oauthToken := isClaudeOAuthToken(apiKey)
cchSigning := claudeCCHSigningEnabled(apiKey, claudeCCHUpstreamAnthropic, url)
reporter := helps.NewExecutorUsageReporter(ctx, e, baseModel, auth)
defer reporter.TrackFailure(ctx, &err)
from := opts.SourceFormat
responseFormat := cliproxyexecutor.ResponseFormatOrSource(opts)
to := sdktranslator.FromString("claude")
// Use an upstream stream whenever the downstream response needs translation
// from Claude events. Native Claude responses use the JSON response path.
upstreamStream := responseFormat != to
originalPayloadSource := req.Payload
if len(opts.OriginalRequest) > 0 {
originalPayloadSource = opts.OriginalRequest
}
originalPayload := originalPayloadSource
incomingHeaders, claudeCodeDetection := detectIncomingClaudeCodeRequest(ctx, opts.Headers, originalPayload, false)
confirmedClaudeCode := claudeCodeDetection.Confirmed
claudeSessionID := ""
if oauthToken {
claudeSessionID = helps.ClaudeAgentSessionUUID(incomingHeaders, originalPayload, req.Payload, opts.Metadata, req.Metadata)
}
originalTranslated := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, originalPayload, upstreamStream)
body := helps.TranslateRequestWithCodexMultiAgentV2(ctx, opts.Headers, e.cfg, from, to, baseModel, req.Payload, upstreamStream)
body = helps.SetStringIfDifferent(body, "model", upstreamModel)
body, err = helps.ApplyRequestThinking(body, req, opts, from.String(), to.String(), e.Identifier())
if err != nil {
return resp, err
}
if rebuildMidSystemMessageEnabled(e.cfg, auth) {
body = rebuildMidSystemMessagesToTopLevel(body)
}
// Apply cloaking (system prompt injection, fake user ID, sensitive word obfuscation)
// based on client type and configuration.
var cloaked bool
body, cloaked, err = applyCloaking(
ctx,
e.cfg,
auth,
body,
apiKey,
confirmedClaudeCode,
cchSigning,
)
if err != nil {
return resp, err
}
// Only the Messages endpoint on Anthropic itself was captured; count_tokens
// keeps its own shape and other gateways never see this field.
if cloaked && isAnthropicUpstreamBase(baseURL) {
body = injectClaudeCodeContextManagement(body)
}
requestedModel := helps.PayloadRequestedModel(opts, req.Model)
requestPath := helps.PayloadRequestPath(opts)
body = helps.ApplyPayloadConfigWithRequest(e.cfg, baseModel, to.String(), from.String(), "", body, originalTranslated, requestedModel, requestPath, opts.Headers)
body = ensureModelMaxTokens(body, baseModel)
// Disable thinking if tool_choice forces tool use (Anthropic API constraint)
body = disableThinkingIfToolChoiceForced(body)
body = normalizeClaudeSamplingForUpstream(body)
// Auto-inject cache_control if missing (optimization for ClawdBot/clients without caching support)
if countCacheControls(body) == 0 {
body = ensureCacheControl(body)
}
// Enforce Anthropic's cache_control block limit (max 4 breakpoints per request).
// Cloaking and ensureCacheControl may push the total over 4 when the client
// already sends multiple cache_control blocks.
body = enforceCacheControlLimit(body, 4)
// Normalize TTL values to prevent ordering violations under prompt-caching-scope-2026-01-05.
// A 1h-TTL block must not appear after a 5m-TTL block in evaluation order (tools→system→messages).
body = normalizeCacheControlTTL(body)
// Payload rules and other request processing may rewrite stream. Keep the
// upstream body, transport headers, and response parser on one authority.
body = helps.SetBoolIfDifferent(body, "stream", upstreamStream)
// Extract betas from body and convert to header
var extraBetas []string
extraBetas, body = extractAndRemoveBetas(body)
bodyForTranslation := body
bodyForUpstream := body
var oauthToolNamesReverseMap map[string]string
if oauthToken && cloaked {
mcpAliases := resolveClaudeMCPAliasOptions(ctx)
bodyForUpstream, oauthToolNamesReverseMap = prepareClaudeOAuthToolNamesForUpstream(bodyForUpstream, mcpAliases)
}
bodyForUpstream = sanitizeClaudeMessagesForClaudeUpstreamWithDebug(ctx, bodyForUpstream, baseModel)
if oauthToken {
bodyForUpstream, _, err = helps.ApplyClaudeCredentialMetadata(bodyForUpstream, auth, claudeSessionID)
if err != nil {
return resp, fmt.Errorf("apply Claude credential metadata: %w", err)
}
}
if cchSigning {
fallbackBilling := claudeCCHFallbackBillingHeader(ctx, e.cfg, bodyForUpstream, claudeCodeDetection.Entrypoint)
bodyForUpstream, err = finalizeAnthropicMessagesBodyCCH(bodyForUpstream, fallbackBilling)
if err != nil {
return resp, fmt.Errorf("finalize Claude CCH: %w", err)
}
}
reporter.SetTranslatedReasoningEffort(bodyForUpstream, to.String())
httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(bodyForUpstream))
if err != nil {
return resp, err
}
if errHeaders := applyClaudeHeaders(httpReq, auth, apiKey, upstreamStream, extraBetas, bodyForUpstream, e.cfg, incomingHeaders, confirmedClaudeCode && !cloaked, claudeSessionID); errHeaders != nil {
return resp, errHeaders
}
var authID, authLabel, authType, authValue string
if auth != nil {
authID = auth.ID
authLabel = auth.Label
authType, authValue = auth.AccountInfo()
}
helps.RecordAPIRequest(ctx, e.cfg, helps.UpstreamRequestLog{
URL: url,
Method: http.MethodPost,
Headers: httpReq.Header.Clone(),
Body: bodyForUpstream,
Provider: e.upstreamRequestLogProvider(),
AuthID: authID,
AuthLabel: authLabel,
AuthType: authType,
AuthValue: authValue,
})
httpClient := helps.NewUtlsHTTPClient(ctx, e.cfg, auth, 0)
httpClient = reporter.TrackHTTPClient(httpClient)
httpResp, err := httpClient.Do(httpReq)
if err != nil {
helps.RecordAPIResponseError(ctx, e.cfg, err)
return resp, err
}
helps.RecordAPIResponseMetadata(ctx, e.cfg, httpResp.StatusCode, httpResp.Header.Clone())
if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
// Decompress error responses — pass the Content-Encoding value (may be empty)
// and let decodeResponseBody handle both header-declared and magic-byte-detected
// compression. This keeps error-path behaviour consistent with the success path.
errBody, decErr := decodeResponseBody(httpResp.Body, httpResp.Header.Get("Content-Encoding"))
if decErr != nil {
helps.RecordAPIResponseError(ctx, e.cfg, decErr)
msg := fmt.Sprintf("failed to decode error response body: %v", decErr)
helps.LogWithRequestID(ctx).Warn(msg)
return resp, statusErr{code: httpResp.StatusCode, msg: msg}
}
b, readErr := io.ReadAll(errBody)
if readErr != nil {
helps.RecordAPIResponseError(ctx, e.cfg, readErr)
msg := fmt.Sprintf("failed to read error response body: %v", readErr)
helps.LogWithRequestID(ctx).Warn(msg)
b = []byte(msg)
}
helps.AppendAPIResponseChunk(ctx, e.cfg, b)
helps.LogWithRequestID(ctx).Debugf("request error, error status: %d, error message: %s", httpResp.StatusCode, helps.SummarizeErrorBody(httpResp.Header.Get("Content-Type"), b))
err = statusErr{code: httpResp.StatusCode, msg: string(b)}
if errClose := errBody.Close(); errClose != nil {
log.Errorf("response body close error: %v", errClose)
}
return resp, err
}
decodedBody, err := decodeResponseBody(httpResp.Body, httpResp.Header.Get("Content-Encoding"))
if err != nil {
helps.RecordAPIResponseError(ctx, e.cfg, err)
if errClose := httpResp.Body.Close(); errClose != nil {
log.Errorf("response body close error: %v", errClose)
}
return resp, err
}
defer func() {
if errClose := decodedBody.Close(); errClose != nil {
log.Errorf("response body close error: %v", errClose)
}
}()
data, err := io.ReadAll(decodedBody)
if err != nil {
helps.RecordAPIResponseError(ctx, e.cfg, err)
return resp, err
}
helps.AppendAPIResponseChunk(ctx, e.cfg, data)
if upstreamStream {
if errValidate := validateClaudeStreamingResponse(data); errValidate != nil {
helps.RecordAPIResponseError(ctx, e.cfg, errValidate)
return resp, errValidate
}
lines := bytes.Split(data, []byte("\n"))
for i, line := range lines {
if detail, ok := helps.ParseClaudeStreamUsage(line); ok {
reporter.Publish(ctx, detail)
}
lines[i] = restoreClaudeOAuthToolNamesFromStreamLine(line, oauthToolNamesReverseMap)
}
data = bytes.Join(lines, []byte("\n"))
} else {
reporter.Publish(ctx, helps.ParseClaudeUsage(data))
data = restoreClaudeOAuthToolNamesFromResponse(data, oauthToolNamesReverseMap)
}
data = e.restoreResponseModel(data, req.Model)
var param any
out := sdktranslator.TranslateNonStream(
ctx,
to,
responseFormat,
req.Model,
opts.OriginalRequest,
bodyForTranslation,
data,
&param,
)
resp = cliproxyexecutor.Response{Payload: out, Headers: httpResp.Header.Clone()}
return resp, nil
}