diff --git a/CHANGELOG.md b/CHANGELOG.md index 5ef2a61..972b14c 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -13,6 +13,13 @@ for reviewers. Add user-visible changes here before running `bun run release:prepare `. +### Removed + +- **The "Extended Context (1M)" toggle is gone.** It opted Anthropic requests + into the 1M-token context beta, which bills over-200K requests at higher + rates and fails outright on lower-tier API keys. Models now always run with + the standard 200K context window. + ## [0.13.0] - 2026-09-13 ### Added diff --git a/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx b/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx index f03be52..053b763 100644 --- a/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx +++ b/apps/electron/src/renderer/pages/settings/AiSettingsPage.tsx @@ -639,7 +639,6 @@ export default function AiSettingsPage() { // Default settings state (app-level) const [defaultThinking, setDefaultThinking] = useState(DEFAULT_THINKING_LEVEL) const [extendedPromptCache, setExtendedPromptCache] = useState(false) - const [enable1MContext, setEnable1MContext] = useState(false) const [rtkEnabled, setRtkEnabled] = useState(false) const [rtkStatus, setRtkStatus] = useState<{ installed: boolean; path: string | null; version: string | null } | null>(null) const [rtkRechecking, setRtkRechecking] = useState(false) @@ -673,9 +672,6 @@ export default function AiSettingsPage() { const extendedCache = await window.electronAPI.getExtendedPromptCache() setExtendedPromptCache(extendedCache) - const enable1M = await window.electronAPI.getEnable1MContext() - setEnable1MContext(enable1M) - const rtkOn = await window.electronAPI.getRtkEnabled() setRtkEnabled(rtkOn) @@ -979,11 +975,6 @@ export default function AiSettingsPage() { await window.electronAPI?.setExtendedPromptCache(enabled) }, []) - const handleEnable1MContextChange = useCallback(async (enabled: boolean) => { - setEnable1MContext(enabled) - await window.electronAPI?.setEnable1MContext(enabled) - }, []) - const handleRtkToggle = useCallback(async (enabled: boolean) => { setRtkEnabled(enabled) await window.electronAPI?.setRtkEnabled(enabled) @@ -1125,12 +1116,6 @@ export default function AiSettingsPage() { {/* Performance */} - setRichToolDescriptions(enabled: boolean): Promise - // Prompt caching & context + // Prompt caching getExtendedPromptCache(): Promise setExtendedPromptCache(enabled: boolean): Promise - getEnable1MContext(): Promise - setEnable1MContext(enabled: boolean): Promise // RTK token optimization getRtkEnabled(): Promise diff --git a/apps/electron/src/transport/channel-map.ts b/apps/electron/src/transport/channel-map.ts index 21c9142..c4fc618 100644 --- a/apps/electron/src/transport/channel-map.ts +++ b/apps/electron/src/transport/channel-map.ts @@ -203,8 +203,6 @@ export const CHANNEL_MAP = { setRichToolDescriptions: invoke(RPC_CHANNELS.appearance.SET_RICH_TOOL_DESCRIPTIONS), getExtendedPromptCache: invoke(RPC_CHANNELS.caching.GET_EXTENDED_PROMPT_CACHE), setExtendedPromptCache: invoke(RPC_CHANNELS.caching.SET_EXTENDED_PROMPT_CACHE), - getEnable1MContext: invoke(RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT), - setEnable1MContext: invoke(RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT), getRtkEnabled: invoke(RPC_CHANNELS.rtk.GET_ENABLED), setRtkEnabled: invoke(RPC_CHANNELS.rtk.SET_ENABLED), getRtkStatus: invoke(RPC_CHANNELS.rtk.GET_STATUS), diff --git a/packages/server-core/src/handlers/rpc/settings.ts b/packages/server-core/src/handlers/rpc/settings.ts index d77445c..24392b6 100644 --- a/packages/server-core/src/handlers/rpc/settings.ts +++ b/packages/server-core/src/handlers/rpc/settings.ts @@ -31,8 +31,6 @@ export const HANDLED_CHANNELS = [ RPC_CHANNELS.appearance.SET_RICH_TOOL_DESCRIPTIONS, RPC_CHANNELS.caching.GET_EXTENDED_PROMPT_CACHE, RPC_CHANNELS.caching.SET_EXTENDED_PROMPT_CACHE, - RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT, - RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT, RPC_CHANNELS.sessions.GET_MODEL, RPC_CHANNELS.sessions.SET_MODEL, RPC_CHANNELS.settings.GET_DEFAULT_THINKING_LEVEL, @@ -296,18 +294,6 @@ export function registerSettingsHandlers(server: RpcServer, deps: HandlerDeps): setExtendedPromptCache(enabled) }) - // Get 1M context window setting - server.handle(RPC_CHANNELS.caching.GET_ENABLE_1M_CONTEXT, async () => { - const { getEnable1MContext } = await import('@bitlab/shared/config/storage') - return getEnable1MContext() - }) - - // Set 1M context window setting - server.handle(RPC_CHANNELS.caching.SET_ENABLE_1M_CONTEXT, async (_ctx, enabled: boolean) => { - const { setEnable1MContext } = await import('@bitlab/shared/config/storage') - setEnable1MContext(enabled) - }) - // ============================================================ // RTK Token-Optimization Settings // ============================================================ diff --git a/packages/shared/src/agent/backend/types.ts b/packages/shared/src/agent/backend/types.ts index 0d9234a..272c56d 100644 --- a/packages/shared/src/agent/backend/types.ts +++ b/packages/shared/src/agent/backend/types.ts @@ -203,9 +203,6 @@ export interface CoreBackendConfig { */ onImageResize?: (filePath: string, maxSizeBytes: number) => Promise; - /** Enable 1M context window for current Opus models. Default: true. Set false to use 200K and conserve usage limits. */ - enable1MContext?: boolean; - } // ============================================================ diff --git a/packages/shared/src/config/storage.ts b/packages/shared/src/config/storage.ts index 5bb4aca..fd3c2b1 100644 --- a/packages/shared/src/config/storage.ts +++ b/packages/shared/src/config/storage.ts @@ -93,9 +93,8 @@ export interface StoredConfig { /** Whether skills may pre-approve tools for their turn. Absent means enabled. */ skillToolApproval?: boolean; mcpSettings?: BitlabMcpSettings; - // Prompt caching & context + // Prompt caching extendedPromptCache?: boolean; // Use 1h prompt cache TTL instead of 5m (default: false) - enable1MContext?: boolean; // Enable 1M context window for supported models (default: false — opt-in; requires Anthropic Tier 4+) // Token optimization rtkEnabled?: boolean; // Route Bash commands through rtk to compress tool output (default: false). https://github.com/rtk-ai/rtk // Network proxy @@ -588,28 +587,6 @@ export function setMcpSettings(settings: BitlabMcpSettings): void { saveConfig(config); } -/** - * Get whether 1M context window is enabled. - * When disabled, models use 200K context and the interceptor strips the context-1m beta header. - * Defaults to false — the 1M beta requires Anthropic Tier 4+, and enabling it by default - * causes 400 "Invalid Request" for lower-tier API keys on large contexts (issue #567). - * Users opt in via AI Settings → Performance → Extended Context (1M). - */ -export function getEnable1MContext(): boolean { - const config = loadStoredConfig(); - return config?.enable1MContext === true; -} - -/** - * Set whether 1M context window is enabled. - */ -export function setEnable1MContext(enabled: boolean): void { - const config = loadStoredConfig(); - if (!config) return; - config.enable1MContext = enabled; - saveConfig(config); -} - /** * Get whether rtk Bash-output compression is enabled. * When enabled, the PreToolUse pipeline rewrites Bash commands to their `rtk` equivalents diff --git a/packages/shared/src/i18n/locales/en.json b/packages/shared/src/i18n/locales/en.json index 838c8af..d935f40 100644 --- a/packages/shared/src/i18n/locales/en.json +++ b/packages/shared/src/i18n/locales/en.json @@ -584,8 +584,6 @@ "settings.ai.defaultSectionDesc": "Settings for new chats when no workspace override is set.", "settings.ai.description": "Model, thinking, connections", "settings.ai.enterConnectionName": "Enter connection name...", - "settings.ai.extendedContext": "Extended Context (1M)", - "settings.ai.extendedContextDesc": "Use 1M token context window for current Opus models. Disable to use 200K and conserve usage limits.", "settings.ai.extendedPromptCache": "Extended prompt cache (1 hour)", "settings.ai.extendedPromptCacheDesc": "Cache prompts for 1 hour instead of 5 minutes. Only applies to Claude models via Anthropic API. Reduces cost for long sessions but increases cache write cost.", "settings.ai.inheritFromApp": "Inherit from app settings", diff --git a/packages/shared/src/i18n/locales/zh-Hans.json b/packages/shared/src/i18n/locales/zh-Hans.json index 5ed73d9..8eb339c 100644 --- a/packages/shared/src/i18n/locales/zh-Hans.json +++ b/packages/shared/src/i18n/locales/zh-Hans.json @@ -584,8 +584,6 @@ "settings.ai.defaultSectionDesc": "未设置 Workspace 覆盖时新聊天的设置。", "settings.ai.description": "模型、思考、连接", "settings.ai.enterConnectionName": "输入连接名称...", - "settings.ai.extendedContext": "扩展上下文 (1M)", - "settings.ai.extendedContextDesc": "为 Opus 4.7 使用 1M token 上下文窗口。禁用则使用 200K 以节省用量配额。", "settings.ai.extendedPromptCache": "扩展提示缓存 (1 小时)", "settings.ai.extendedPromptCacheDesc": "将提示缓存从 5 分钟延长到 1 小时。仅适用于通过 Anthropic API 的 Claude 模型。可降低长会话成本，但会增加缓存写入成本。", "settings.ai.inheritFromApp": "继承应用设置", diff --git a/packages/shared/src/interceptor-common.ts b/packages/shared/src/interceptor-common.ts index b551e48..226eff1 100644 --- a/packages/shared/src/interceptor-common.ts +++ b/packages/shared/src/interceptor-common.ts @@ -138,18 +138,6 @@ export function isExtendedPromptCacheEnabled(): boolean { return config?.extendedPromptCache === true; } -/** - * Check if 1M context window is enabled. - * When disabled, the interceptor strips the context-1m beta header. - * Defaults to false — the 1M beta requires Anthropic Tier 4+, so it's opt-in - * to avoid 400 "Invalid Request" on lower-tier API keys (issue #567). - * Must stay in sync with getEnable1MContext() in config/storage.ts. - */ -export function is1MContextEnabled(): boolean { - const config = getInterceptorConfig(); - return config?.enable1MContext === true; -} - // ============================================================================ // LAST API ERROR // ============================================================================ diff --git a/packages/shared/src/protocol/channels.ts b/packages/shared/src/protocol/channels.ts index dc95fa8..61bb20f 100644 --- a/packages/shared/src/protocol/channels.ts +++ b/packages/shared/src/protocol/channels.ts @@ -284,8 +284,6 @@ export const RPC_CHANNELS = { caching: { GET_EXTENDED_PROMPT_CACHE: 'caching:getExtendedPromptCache', SET_EXTENDED_PROMPT_CACHE: 'caching:setExtendedPromptCache', - GET_ENABLE_1M_CONTEXT: 'caching:getEnable1MContext', - SET_ENABLE_1M_CONTEXT: 'caching:setEnable1MContext', }, rtk: { GET_ENABLED: 'rtk:getEnabled', diff --git a/packages/shared/src/unified-network-interceptor.ts b/packages/shared/src/unified-network-interceptor.ts index 7a3475a..3154af5 100644 --- a/packages/shared/src/unified-network-interceptor.ts +++ b/packages/shared/src/unified-network-interceptor.ts @@ -24,7 +24,6 @@ import { debugLog, isRichToolDescriptionsEnabled, isExtendedPromptCacheEnabled, - is1MContextEnabled, setStoredError, toolMetadataStore, displayNameSchema, @@ -790,16 +789,14 @@ const anthropicAdapter: ApiAdapter = { sanitizeEmptyTextCacheControl(body); upgradePromptCacheTtl(body); - // Strip SDK-injected 1M context beta when setting disables it. - // The SDK adds this header automatically for Opus/Sonnet 4.6 models, - // but the user may want 200K context to conserve usage limits. - if (!is1MContextEnabled()) { - debugLog('[Anthropic] Stripping context-1m beta header (enable1MContext=false)'); - init = { - ...init, - headers: stripBetaHeader(init?.headers as HeadersInitType | undefined, 'context-1m-2025-08-07'), - }; - } + // Always strip the SDK-injected 1M context beta. The SDK adds this header + // automatically for Opus/Sonnet 4.6 models, but it 400s on lower-tier + // Anthropic API keys (issue #567) and burns usage limits faster, so all + // requests stay on the 200K context window. + init = { + ...init, + headers: stripBetaHeader(init?.headers as HeadersInitType | undefined, 'context-1m-2025-08-07'), + }; const fastMode = shouldEnableFastMode(body.model); if (fastMode) {