feat: 添加 Bedrock API 客户端及 API 层增强

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/services/api/claude.ts

@@ -101,6 +101,8 @@ import {
   extractQuotaStatusFromHeaders,
 } from '../claudeAiLimits.js'
 import { getAPIContextManagement } from '../compact/apiMicrocompact.js'
+import { bedrockAdapter } from '../providerUsage/adapters/bedrock.js'
+import { updateProviderBuckets } from '../providerUsage/store.js'
 
 /* eslint-disable @typescript-eslint/no-require-imports */
 const autoModeStateModule = feature('TRANSCRIPT_CLASSIFIER')
@@ -541,13 +543,12 @@ export async function verifyApiKey(
           }),
         async anthropic => {
           const messages: MessageParam[] = [{ role: 'user', content: 'test' }]
-          // biome-ignore lint/plugin: API key verification is intentionally a minimal direct call
           await anthropic.beta.messages.create({
             model,
             max_tokens: 1,
             messages,
             temperature: 1,
-            ...(betas.length > 0 && { betas }),
+            ...(betas.length > 0 && { betas: betas.filter(Boolean) }),
             metadata: getAPIMetadata(),
             ...getExtraBodyParams(),
           })
@@ -878,7 +879,6 @@ export async function* executeNonStreamingRequest(
       )
 
       try {
-        // biome-ignore lint/plugin: non-streaming API call
         return await anthropic.beta.messages.create(
           {
             ...adjustedParams,
@@ -1215,10 +1215,15 @@ async function* queryModel(
     cacheEditingBetaHeader = betas.CACHE_EDITING_BETA_HEADER
     const featureEnabled = isCachedMicrocompactEnabled()
     const modelSupported = isModelSupportedForCacheEditing(options.model)
-    cachedMCEnabled = featureEnabled && modelSupported
+    // cachedMC requires a non-empty beta header; the CACHE_EDITING_BETA_HEADER
+    // constant is '' in this fork (upstream hasn't published the real value).
+    // Without it, cache_reference and cache_edits in the request body cause
+    // API 400: "tool_result.cache_reference: Extra inputs are not permitted".
+    const headerAvailable = !!cacheEditingBetaHeader
+    cachedMCEnabled = featureEnabled && modelSupported && headerAvailable
     const config = getCachedMCConfig()
     logForDebugging(
-      `Cached MC gate: enabled=${featureEnabled} modelSupported=${modelSupported} model=${options.model} supportedModels=${jsonStringify((config as any).supportedModels)}`,
+      `Cached MC gate: enabled=${featureEnabled} modelSupported=${modelSupported} headerAvailable=${headerAvailable} model=${options.model} supportedModels=${jsonStringify((config as Record<string, unknown>).supportedModels)}`,
     )
   }
 
@@ -1724,6 +1729,7 @@ async function* queryModel(
       options.querySource === 'repl_main_thread'
     if (
       cacheEditingHeaderLatched &&
+      cacheEditingBetaHeader &&
       getAPIProvider() === 'firstParty' &&
       options.querySource === 'repl_main_thread' &&
       !betasParams.includes(cacheEditingBetaHeader)
@@ -1740,7 +1746,12 @@ async function* queryModel(
       ? (options.temperatureOverride ?? 1)
       : undefined
 
-    lastRequestBetas = betasParams
+    // Filter out any empty-string beta headers before sending.
+    // Constants like CACHE_EDITING_BETA_HEADER or AFK_MODE_BETA_HEADER
+    // can be '' when their feature gate is off; an empty string in the
+    // betas array produces an invalid anthropic-beta header (400 error).
+    const filteredBetas = betasParams.filter(Boolean)
+    lastRequestBetas = filteredBetas
 
     return {
       model: normalizeModelStringForAPI(options.model),
@@ -1756,7 +1767,7 @@ async function* queryModel(
       system,
       tools: allTools,
       tool_choice: options.toolChoice,
-      ...(useBetas && { betas: betasParams }),
+      ...(useBetas && { betas: filteredBetas }),
       metadata: getAPIMetadata(),
       max_tokens: maxOutputTokens,
       thinking,
@@ -1864,7 +1875,6 @@ async function* queryModel(
         // Use raw stream instead of BetaMessageStream to avoid O(n²) partial JSON parsing
         // BetaMessageStream calls partialParse() on every input_json_delta, which we don't need
         // since we handle tool input accumulation ourselves
-        // biome-ignore lint/plugin: main conversation loop handles attribution separately
         const result = await anthropic.beta.messages
           .create(
             { ...params, stream: true },
@@ -2445,6 +2455,16 @@ async function* queryModel(
       const resp = streamResponse as unknown as Response | undefined
       if (resp) {
         extractQuotaStatusFromHeaders(resp.headers)
+        // Non-Anthropic providers that flow through this same client path
+        // (Bedrock) expose their own throttle headers — let their adapter
+        // overwrite the store with its bucket(s). Anthropic's adapter runs
+        // inside extractQuotaStatusFromHeaders.
+        if (getAPIProvider() === 'bedrock') {
+          updateProviderBuckets(
+            'bedrock',
+            bedrockAdapter.parseHeaders(resp.headers),
+          )
+        }
         // Store headers for gateway detection
         responseHeaders = resp.headers
       }
@@ -3229,6 +3249,7 @@ export function addCacheBreakpoints(
 
   // Add cache_reference to tool_result blocks that are within the cached prefix.
   // Must be done AFTER cache_edits insertion since that modifies content arrays.
+  // Note: this code only runs when useCachedMC=true (early return at line ~3202).
   if (enablePromptCaching) {
     // Find the last message containing a cache_control marker
     let lastCCMsg = -1