feat: add configurable provider RPM limiting

.gitignore

@@ -11,3 +11,4 @@ scripts/claude-proxy.log
 .env
 .env.local
 .env.*.local
+.vscode/

CHANGELOG.md

@@ -94,6 +94,20 @@
 - reduce maxInputTokens for GPT-5.6 models to 272000 (b5febfc, SmallMain)
 - remove redundant Anthropic Messages variant (aa0d17c, RheagalFire)
 
+## v7.12.5 - 2026-07-23
+
+### Fixes
+- return cancelled queued RPM acquisitions immediately instead of waiting behind earlier FIFO entries; cancelled entries still consume no token when they reach the queue head
+
+## v7.12.4 - 2026-07-14
+
+### Fixes
+- make provider RPM token buckets FIFO-safe under concurrent chat requests and show available tokens with two decimal places
+- clarify per-window logical chat request RPM behavior and preserve explicit per-provider `rpm: 0` overrides
+- treat invalid/empty per-provider `rateLimit` (e.g. `{}` or non-integer `rpm`) as unconfigured so the global default applies, while still honoring explicit `rpm: 0`
+- make RPM token acquisition cancellation-aware: aborting a chat request while waiting for a token no longer blocks or consumes a token
+- log the exact post-acquire token snapshot by peeking the bucket without an extra time-based refill
+
 ## v7.12.3 - 2026-07-12
 
 ### Features

l10n/bundle.l10n.json

@@ -89,6 +89,10 @@
   "Extra body parameters must be configured in VS Code settings (JSON).": "Extra body parameters must be configured in VS Code settings (JSON).",
   "{0} properties": "{0} properties",
   "Network Settings": "Network Settings",
+  "Rate Limit (RPM)": "Rate Limit (RPM)",
+  "Maximum requests per minute (0 = disabled)": "Maximum requests per minute (0 = disabled)",
+  "Enter maximum requests per minute (0 = disabled)": "Enter maximum requests per minute (0 = disabled)",
+  "rpm: {0}": "rpm: {0}",
   "default": "default",
   "conn: {0}ms": "conn: {0}ms",
   "resp: {0}ms": "resp: {0}ms",

l10n/bundle.l10n.zh-cn.json

@@ -89,6 +89,10 @@
   "Extra body parameters must be configured in VS Code settings (JSON).": "额外 Body 参数必须在 VS Code 设置（JSON）中配置。",
   "{0} properties": "{0} 个属性",
   "Network Settings": "网络设置",
+  "Rate Limit (RPM)": "RPM 限流",
+  "Maximum requests per minute (0 = disabled)": "每分钟最大请求数（0 = 禁用）",
+  "Enter maximum requests per minute (0 = disabled)": "输入每分钟最大请求数（0 = 禁用）",
+  "rpm: {0}": "RPM: {0}",
   "default": "默认",
   "conn: {0}ms": "连接：{0}ms",
   "resp: {0}ms": "响应：{0}ms",

package.json

@@ -268,6 +268,18 @@
           "scope": "application",
           "description": "%configuration.networkSettings.description%",
           "properties": {
+            "rateLimit": {
+              "type": "object",
+              "description": "%configuration.networkSettings.rateLimit.description%",
+              "properties": {
+                "rpm": {
+                  "type": "integer",
+                  "description": "%configuration.networkSettings.rateLimit.rpm.description%",
+                  "default": 0,
+                  "minimum": 0
+                }
+              }
+            },
             "timeout": {
               "type": "object",
               "description": "%configuration.networkSettings.timeout.description%",
@@ -1003,6 +1015,18 @@
                 "description": "%configuration.endpoints.extraBody.description%",
                 "additionalProperties": true
               },
+              "rateLimit": {
+                "type": "object",
+                "description": "%configuration.endpoints.rateLimit.description%",
+                "properties": {
+                  "rpm": {
+                    "type": "integer",
+                    "description": "%configuration.endpoints.rateLimit.rpm.description%",
+                    "default": 0,
+                    "minimum": 0
+                  }
+                }
+              },
               "timeout": {
                 "type": "object",
                 "description": "%configuration.endpoints.timeout.description%",

package.nls.json

@@ -44,7 +44,7 @@
   "configuration.balanceWarning.timeThresholdDays.description": "Time threshold in days for expiration warnings (supports decimals). Default: 1.",
   "configuration.balanceWarning.amountThreshold.description": "Balance amount threshold for warnings (unitless; currency ignored). Default: 1.",
   "configuration.balanceWarning.tokenThresholdMillions.description": "Token remaining threshold in millions for warnings. Default: 1.",
-  "configuration.networkSettings.description": "Global network settings. Timeout and retry affect chat requests; proxy affects provider HTTP requests.",
+  "configuration.networkSettings.description": "Global network settings. Rate limits, timeouts, and retries affect logical chat requests; proxy affects provider HTTP requests.",
   "configuration.networkSettings.timeout.description": "Timeout settings for chat requests (milliseconds).",
   "configuration.networkSettings.timeout.connection.description": "TCP connection timeout in milliseconds (chat requests).",
   "configuration.networkSettings.timeout.response.description": "Maximum time between SSE data chunks in milliseconds (chat requests).",
@@ -56,6 +56,8 @@
   "configuration.networkSettings.retry.jitterFactor.description": "Jitter factor (0-1) to randomize retry delay (chat requests).",
   "configuration.networkSettings.retry.statusCodes.description": "HTTP status codes that should trigger retries for chat requests. When set, this overrides the default retryable status codes.",
   "configuration.networkSettings.proxy.description": "Global proxy settings for provider requests.",
+  "configuration.networkSettings.rateLimit.description": "Default rate-limit settings for each provider in this VS Code window. Provider-level values override these settings.",
+  "configuration.networkSettings.rateLimit.rpm.description": "Maximum logical chat requests per minute for each provider in this VS Code window. Set to 0 (default) to disable rate limiting.",
   "configuration.proxy.type.description": "Proxy mode.",
   "configuration.proxy.type.enumDescriptions.0": "Use VS Code HTTP proxy settings.",
   "configuration.proxy.type.enumDescriptions.1": "Connect directly without using a proxy.",
@@ -150,6 +152,8 @@
   "configuration.endpoints.retry.jitterFactor.description": "Jitter factor (0-1) to randomize retry delay (chat requests).",
   "configuration.endpoints.proxy.description": "Proxy settings for this provider. Overrides global proxy settings.",
   "configuration.endpoints.autoFetchOfficialModels.description": "Automatically fetch and sync official models from the provider API.",
+  "configuration.endpoints.rateLimit.description": "Rate-limit settings for logical chat requests in this VS Code window.",
+  "configuration.endpoints.rateLimit.rpm.description": "Maximum logical chat requests per minute for this provider in this VS Code window. Set to 0 (default) to disable rate limiting.",
   "configuration.endpoints.models.description": "List of available models.",
   "configuration.endpoints.models.string.description": "Model ID",
   "configuration.endpoints.models.id.description": "Model ID (e.g., claude-sonnet-4-20250514).",

package.nls.zh-cn.json

@@ -44,7 +44,7 @@
   "configuration.balanceWarning.timeThresholdDays.description": "时间阈值（天），用于到期提醒（支持小数）。默认值：1。",
   "configuration.balanceWarning.amountThreshold.description": "余额阈值（无单位，忽略货币）。默认值：1。",
   "configuration.balanceWarning.tokenThresholdMillions.description": "令牌数阈值（单位：百万），用于剩余令牌提醒。默认值：1。",
-  "configuration.networkSettings.description": "全局网络设置。超时与重试影响聊天请求；代理影响供应商 HTTP 请求。",
+  "configuration.networkSettings.description": "全局网络设置。限流、超时与重试影响逻辑聊天请求；代理影响供应商 HTTP 请求。",
   "configuration.networkSettings.timeout.description": "聊天请求的超时设置（毫秒）。",
   "configuration.networkSettings.timeout.connection.description": "TCP 连接超时（毫秒）（聊天请求）。",
   "configuration.networkSettings.timeout.response.description": "SSE 流式传输期间，两次数据分片之间的最大等待时间（毫秒）（聊天请求）。",
@@ -56,6 +56,8 @@
   "configuration.networkSettings.retry.jitterFactor.description": "聊天请求的抖动因子（0-1），用于随机化重试延迟。",
   "configuration.networkSettings.retry.statusCodes.description": "聊天请求中触发重试的 HTTP 状态码。设置后会覆盖默认可重试状态码。",
   "configuration.networkSettings.proxy.description": "供应商请求的全局代理设置。",
+  "configuration.networkSettings.rateLimit.description": "此 VS Code 窗口内每个供应商的默认主动限流设置；供应商级设置可覆盖。",
+  "configuration.networkSettings.rateLimit.rpm.description": "此 VS Code 窗口内每个供应商的逻辑聊天请求每分钟最大数量。设为 0（默认）表示不限流。",
   "configuration.proxy.type.description": "代理模式。",
   "configuration.proxy.type.enumDescriptions.0": "使用 VS Code HTTP 代理设置。",
   "configuration.proxy.type.enumDescriptions.1": "直连，不使用代理。",
@@ -150,6 +152,8 @@
   "configuration.endpoints.retry.jitterFactor.description": "聊天请求的抖动因子（0-1），用于随机化重试延迟。",
   "configuration.endpoints.proxy.description": "此供应商的代理设置。会覆盖全局代理设置。",
   "configuration.endpoints.autoFetchOfficialModels.description": "自动从供应商 API 拉取并同步官方模型。",
+  "configuration.endpoints.rateLimit.description": "此 VS Code 窗口内逻辑聊天请求的主动限流设置。",
+  "configuration.endpoints.rateLimit.rpm.description": "此 VS Code 窗口内该供应商的逻辑聊天请求每分钟最大数量。设为 0（默认）表示不限流。",
   "configuration.endpoints.models.description": "可用模型列表。",
   "configuration.endpoints.models.string.description": "模型 ID",
   "configuration.endpoints.models.id.description": "模型 ID（例如：claude-sonnet-4-20250514）。",

scripts/chat-lib-diff-patches.ts

@@ -395,29 +395,30 @@ export function patchChatLibDiffSource(
   source: string,
 ): string {
   const resolved = resolve(filePath);
+  const normalizedSource = source.replace(/\r\n?/g, '\n');
   if (resolved === arraysPath) {
-    return patchArrays(filePath, source);
+    return patchArrays(filePath, normalizedSource);
   }
   if (resolved === assertPath) {
-    return patchAssert(filePath, source);
+    return patchAssert(filePath, normalizedSource);
   }
   if (resolved === abstractTextPath) {
-    return patchAbstractText(filePath, source);
+    return patchAbstractText(filePath, normalizedSource);
   }
   if (resolved === defaultLinesDiffComputerPath) {
-    return patchDefaultLinesDiffComputer(filePath, source);
+    return patchDefaultLinesDiffComputer(filePath, normalizedSource);
   }
   if (resolved === heuristicSequenceOptimizationsPath) {
-    return patchHeuristicSequenceOptimizations(filePath, source);
+    return patchHeuristicSequenceOptimizations(filePath, normalizedSource);
   }
   if (resolved === positionToOffsetPath) {
-    return patchPositionToOffset(filePath, source);
+    return patchPositionToOffset(filePath, normalizedSource);
   }
   if (resolved === positionToOffsetImplPath) {
-    return patchPositionToOffsetImpl(filePath, source);
+    return patchPositionToOffsetImpl(filePath, normalizedSource);
   }
   if (resolved === rangeMappingPath) {
-    return patchRangeMapping(filePath, source);
+    return patchRangeMapping(filePath, normalizedSource);
   }
   return source;
 }

scripts/chat-lib-parser-patches.ts

@@ -30,9 +30,10 @@ export function patchChatLibParserSource(
   source: string,
 ): string {
   const resolved = resolve(filePath);
+  const normalizedSource = source.replace(/\r\n?/g, '\n');
   if (resolved === blockTrimmerPath) {
     return replaceOnce(
-      source,
+      normalizedSource,
       "import { IPosition, TextDocumentContents } from '../textDocument';",
       [
         'interface IPosition {',
@@ -51,7 +52,7 @@ export function patchChatLibParserSource(
   }
   if (resolved === parseBlockPath) {
     return replaceOnce(
-      source,
+      normalizedSource,
       [
         '\t\tif (endIndex < solution.length) {',
         '\t\t\t// descendant block is finished, stop at end of block',

src/client/anthropic/client.ts

@@ -76,6 +76,7 @@ import {
   setUserAgentHeader,
 } from '../utils';
 import type { AuthTokenInfo, AuthTokenRefresh } from '../../auth/types';
+import { createRateLimiter } from '../../rate-limit';
 
 type AnthropicMarkerData = {
   raw: BetaMessage;
@@ -116,8 +117,10 @@ const ANTHROPIC_ABNORMAL_STOP_REASONS: ReadonlySet<string> = new Set([
 
 export class AnthropicProvider implements ApiProvider {
   private readonly baseUrl: string;
+  private readonly rateLimiter: ReturnType<typeof createRateLimiter>;
 
   constructor(protected readonly config: ProviderConfig) {
+    this.rateLimiter = createRateLimiter(config.rateLimit);
     this.baseUrl = buildBaseUrl(config.baseUrl, {
       stripPattern: /\/v1$/i,
       useRawBaseUrl: isRawBaseUrlEnabled(config),
@@ -136,6 +139,14 @@ export class AnthropicProvider implements ApiProvider {
    * Create an Anthropic client with custom fetch for retry support.
    * A new client is created per request to enable per-request logging.
    */
+  getRateLimitStatus(): { available: number; capacity: number } | undefined {
+    return this.rateLimiter?.getAvailableTokens();
+  }
+
+  async acquireRateLimitToken(signal?: AbortSignal): Promise<void> {
+    await this.rateLimiter?.acquire(signal);
+  }
+
   protected createClient(
     logger: ProviderHttpLogger | undefined,
     stream: boolean,

src/client/github-copilot/client.ts

@@ -23,6 +23,7 @@ import {
 } from '../../utils';
 import type { ApiProvider } from '../interface';
 import { buildBaseUrl, createCustomFetch, getToken } from '../utils';
+import { createRateLimiter } from '../../rate-limit';
 import { OpenAIChatCompletionProvider } from '../openai/chat-completion-client';
 import { OpenAIResponsesProvider } from '../openai/responses-client';
 import type { ChatCompletionChunk } from 'openai/resources/chat/completions';
@@ -669,6 +670,7 @@ export class GitHubCopilotProvider implements ApiProvider {
   private readonly chatProvider: GitHubCopilotChatCompletionProvider;
   private readonly responsesProvider: GitHubCopilotResponsesProvider;
   private readonly messagesProvider: GitHubCopilotMessagesProvider;
+  private readonly rateLimiter: ReturnType<typeof createRateLimiter>;
 
   private assertCopilotAuth(): void {
     if (this.providerConfig.auth?.method !== 'github-copilot') {
@@ -679,6 +681,7 @@ export class GitHubCopilotProvider implements ApiProvider {
   }
 
   constructor(config: ProviderConfig) {
+    this.rateLimiter = createRateLimiter(config.rateLimit);
     const extraBody = config.extraBody ?? {};
     const hasStore = Object.prototype.hasOwnProperty.call(extraBody, 'store');
     const configWithDefaults: ProviderConfig = hasStore
@@ -699,6 +702,14 @@ export class GitHubCopilotProvider implements ApiProvider {
     });
   }
 
+  getRateLimitStatus(): { available: number; capacity: number } | undefined {
+    return this.rateLimiter?.getAvailableTokens();
+  }
+
+  async acquireRateLimitToken(signal?: AbortSignal): Promise<void> {
+    await this.rateLimiter?.acquire(signal);
+  }
+
   async *streamChat(
     encodedModelId: string,
     model: ModelConfig,

src/client/google/ai-studio-client.ts

@@ -53,6 +53,7 @@ import {
   type RetryConfig,
   withIdleTimeout,
 } from '../../utils';
+import { createRateLimiter } from '../../rate-limit';
 import { getBaseModelId } from '../../model-id-utils';
 import { ThinkingBlockMetadata } from '../types';
 import {
@@ -157,8 +158,18 @@ function getThoughtSignature(part: Part): string | undefined {
 export class GoogleAIStudioProvider implements ApiProvider {
   protected readonly baseUrl: string;
   protected readonly apiVersion: string;
+  protected readonly rateLimiter: ReturnType<typeof createRateLimiter>;
+
+  getRateLimitStatus(): { available: number; capacity: number } | undefined {
+    return this.rateLimiter?.getAvailableTokens();
+  }
+
+  async acquireRateLimitToken(signal?: AbortSignal): Promise<void> {
+    await this.rateLimiter?.acquire(signal);
+  }
 
   constructor(protected readonly config: ProviderConfig) {
+    this.rateLimiter = createRateLimiter(config.rateLimit);
     if (isRawBaseUrlEnabled(config)) {
       this.baseUrl = normalizeRawBaseUrlInput(config.baseUrl);
       this.apiVersion = 'v1beta';

src/client/interface.ts

@@ -54,4 +54,20 @@ export interface ApiProvider {
     refreshCredential?: AuthTokenRefresh,
     signal?: AbortSignal,
   ): Promise<ModelConfig[]>;
+
+  /**
+   * Get the current rate-limit status (token bucket snapshot).
+   * Returns undefined when no rate limiting is configured for this provider.
+   */
+  getRateLimitStatus?(): { available: number; capacity: number } | undefined;
+
+  /**
+   * Acquire a token for one logical chat request, waiting if necessary.
+   * Called by the service before transport selection so HTTP, SSE, and
+   * WebSocket requests share the same limiter.
+   *
+   * If `signal` is aborted while waiting, rejects (without consuming a token)
+   * so a cancelled chat request doesn't waste a rate-limit slot.
+   */
+  acquireRateLimitToken?(signal?: AbortSignal): Promise<void>;
 }

src/client/ollama/client.ts

@@ -62,6 +62,7 @@ import {
   normalizeToolInputSchema,
   processUsage as sharedProcessUsage,
 } from '../utils';
+import { createRateLimiter } from '../../rate-limit';
 
 const TOOL_CALL_ID_PREFIX = 'ollama-tool:';
 
@@ -104,8 +105,18 @@ const OLLAMA_ABNORMAL_DONE_REASONS: ReadonlySet<string> = new Set([
 
 export class OllamaProvider implements ApiProvider {
   private readonly baseUrl: string;
+  private readonly rateLimiter: ReturnType<typeof createRateLimiter>;
+
+  getRateLimitStatus(): { available: number; capacity: number } | undefined {
+    return this.rateLimiter?.getAvailableTokens();
+  }
+
+  async acquireRateLimitToken(signal?: AbortSignal): Promise<void> {
+    await this.rateLimiter?.acquire(signal);
+  }
 
   constructor(private readonly config: ProviderConfig) {
+    this.rateLimiter = createRateLimiter(config.rateLimit);
     this.baseUrl = buildBaseUrl(config.baseUrl, {
       stripPattern: /\/api$/i,
       useRawBaseUrl: isRawBaseUrlEnabled(config),