Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。
TypeScript agentai-agentai-assistantandroidclaude-codecodexelectroniosllmmacosreact-nativetypescriptwindows
import type {
HookCallback,
HookCallbackMatcher,
HookEvent,
PreToolUseHookInput,
} from '@anthropic-ai/claude-agent-sdk';
export type ClaudeSubagentModelAccessStatus = 'allowed' | 'denied' | 'unknown';
export interface ClaudeSubagentModelAccessResult {
status: ClaudeSubagentModelAccessStatus;
/** 可选的 host 诊断;只有 denied 会展示给用户。 */
reason?: string;
}
export type ResolveClaudeSubagentModelAccess = (
model: string,
) => ClaudeSubagentModelAccessResult | Promise<ClaudeSubagentModelAccessResult>;
export type ResolveClaudeSubagentModelContextWindow = (
model: string,
) => number | undefined;
export function normalizeClaudeSubagentModel(model: string): string {
const normalized = model.trim().toLowerCase();
return normalized.endsWith('[1m]')
? normalized.slice(0, -'[1m]'.length)
: normalized;
}
/** Add or remove Claude Code's explicit 1M context wire suffix for a known model. */
export function claudeSubagentModelWithContextWindow(
model: string,
contextWindow: number | undefined,
): string {
const trimmed = model.trim();
if (
!trimmed
|| typeof contextWindow !== 'number'
|| !Number.isFinite(contextWindow)
|| contextWindow <= 0
) {
return trimmed;
}
const bare = trimmed.replace(/\[1m\]$/i, '');
return contextWindow >= 1_000_000 ? `${bare}[1m]` : bare;
}
/**
* 返回 Agent/Task 实际会用的显式模型。平台 env 覆写优先于 tool input;
* 缺省和 inherit 都交给 Claude 自己解析,不对无法证明的隐式默认值做拦截。
*/
export function effectiveClaudeSubagentModel(
forcedModel: string | undefined,
toolName: string,
toolInput: unknown,
): string | undefined {
if (toolName !== 'Agent' && toolName !== 'Task') return undefined;
const input = typeof toolInput === 'object' && toolInput !== null
? toolInput as Record<string, unknown>
: {};
const requested = typeof input.model === 'string' ? input.model : '';
const effective = normalizeClaudeSubagentModel(forcedModel ?? requested);
return !effective || effective === 'inherit' ? undefined : effective;
}
export function claudeSubagentModelDenialReason(model: string, detail?: string): string {
return detail?.trim()
|| `Subagent model "${model}" is not available from the current account and provider. Choose an available model, or remove the unavailable override from the Agent call or Subagent Model setting.`;
}
/**
* PreToolUse 先于权限模式执行,因此 Full access 也无法绕过。resolver 每次调用都
* 现场读取 host 的当前账号/路由状态;缺失、异常或 unknown 一律放行,防止静态目录
* 或旧快照被误当成权限拒绝。
*/
export function buildClaudeSubagentModelGuardHooks(
resolveAccess: ResolveClaudeSubagentModelAccess | undefined,
forcedModel?: string,
onDeny?: (model: string) => void,
resolveContextWindow?: ResolveClaudeSubagentModelContextWindow,
): Partial<Record<HookEvent, HookCallbackMatcher[]>> {
if (!resolveAccess && !resolveContextWindow) return {};
const guard: HookCallback = async (input) => {
if (input.hook_event_name !== 'PreToolUse') return { continue: true };
const pre = input as PreToolUseHookInput;
const model = effectiveClaudeSubagentModel(forcedModel, pre.tool_name, pre.tool_input);
const toolInput = typeof pre.tool_input === 'object' && pre.tool_input !== null
? pre.tool_input as Record<string, unknown>
: undefined;
const requestedModel = typeof toolInput?.model === 'string' ? toolInput.model : undefined;
const modelToRewrite = requestedModel ?? forcedModel;
let wireModel: string | undefined;
if (modelToRewrite && resolveContextWindow) {
try {
wireModel = claudeSubagentModelWithContextWindow(
modelToRewrite,
resolveContextWindow(normalizeClaudeSubagentModel(modelToRewrite)),
);
} catch {
// Context metadata is advisory; an unavailable live resolver must not
// block an otherwise valid Agent/Task invocation.
wireModel = undefined;
}
}
const updatedInput = requestedModel && wireModel && wireModel !== requestedModel
? { ...toolInput, model: wireModel }
: undefined;
const continueWithUpdatedInput = updatedInput
? {
continue: true as const,
hookSpecificOutput: {
hookEventName: 'PreToolUse' as const,
updatedInput,
},
}
: { continue: true as const };
if (!model) return continueWithUpdatedInput;
if (resolveAccess) {
let access: ClaudeSubagentModelAccessResult;
try {
access = await resolveAccess(model);
} catch {
return continueWithUpdatedInput;
}
if (access.status === 'denied') {
onDeny?.(model);
return {
continue: true,
hookSpecificOutput: {
hookEventName: 'PreToolUse',
permissionDecision: 'deny',
permissionDecisionReason: claudeSubagentModelDenialReason(model, access.reason),
},
};
}
}
return continueWithUpdatedInput;
};
return {
PreToolUse: [
{ matcher: 'Agent', hooks: [guard] },
{ matcher: 'Task', hooks: [guard] },
],
};
}