feat: v0.7.0 四阶段全量迭代 — 修复面收口 · 安全纵深 · 架构还债 · 能力演进
P1 修复面收口: v0.6.3 截断自愈推全量(Anthropic/Ollama/非流式/引擎兜底); SSE 上游错误帧检测进重试通道; clearMessages 摘要游标根治; truncateResult 内联图片白名单统一; 前端四 bug(确认弹窗锁死/MemoryViewer/ Virtuoso Footer/abort 尾部过滤) + reasoning 缓冲跨迭代污染; 托盘通知过滤与新建会话死链接线 P2 安全纵深: MCP 审批闭环(ConfirmationHook×PolicyEngine 联动+重名拒注册); SSRF 收敛 ssrf-guard 共享模块 (web_fetch 双通道校验+重定向终态复检); Electron 加固(preload CJS 化→sandbox:true/CSP/权限白名单/will-navigate); run_command cmd.exe 白名单通道元字符守门; diff_viewer 10MB 预检; Anthropic thinking 预算下限; Agnes 思考显式关闭 P3 架构还债: OpenAICompatibleAdapter 中间基类收敛四家样板; 错误分类单轨化(删 mapError/getFetchSignal, 超时显式 ETIMEDOUT); PRAGMA user_version 迁移版本化; 死代码清理专项(cn.ts/SHORTCUTS/ContextMenu 分支/ getWindowState/modifiedArgs/sandbox 空壳); i18next 引入; a11y 第一轮; SearXNG 页批量草稿模型统一 P4 能力演进: Ollama pull 可取消/capabilities 探测/num_ctx 实测缓存; UpdateService feed 比对式自动更新 (app:updateCheck IPC + StatusBar 入口); MiMo providerOptions(web_search 服务端工具/strict JSON); web_fetch extract_mode=markdown(turndown); network.proxyUrl 全局代理(Chromium sessions+undici dispatcher) 测试: 264 → 507 用例(Electron ABI 全绿零跳过), 覆盖引擎压缩管线/重试竞速/MEMORY.md 闸门/file_editor 五操作/ filesystem 七工具实体夹具/git 真实仓库/SSE 错误帧/全线截断自愈/Provider 请求形态矩阵/SSRF 表测/钩子分级矩阵/ OutputValidator 全量/SLO 指标/MCP 安全纯函数/task_manager 链路/渲染层纯域/i18n 桥契约
This commit is contained in:
@@ -1,12 +1,18 @@
|
||||
/**
|
||||
* BaseAdapter 单元测试（v0.4.1 测试补齐）
|
||||
* 覆盖：错误映射（mapError）、HTTP 错误识别（throwHttpError）、
|
||||
* ContentFilterError、上下文窗口读取、fetchWithTimeout 超时与清理
|
||||
* BaseAdapter 单元测试（v0.6.4 P3-2 错误分类单轨化后重写）
|
||||
*
|
||||
* 契约变更说明：
|
||||
* - mapError 已删除（生产路径死代码，与 engine.isRetryableError 双轨漂移）。
|
||||
* 错误分类的唯一事实来源是 engine.isRetryableError —— 本文件改为验证
|
||||
* "BaseAdapter 抛出的错误携带可判定字段"的形状契约：
|
||||
* throwHttpError → error.status；fetchWithTimeout 超时 → code='ETIMEDOUT'
|
||||
* + 'timed out' message（命中引擎网络超时分支）。
|
||||
* - 新增：错误体长度截断、外部 abort 与自身超时的区分。
|
||||
*/
|
||||
|
||||
import { describe, it, expect, vi, afterEach } from 'vitest';
|
||||
import { BaseAdapter, ContentFilterError } from '../base-adapter';
|
||||
import { MetonaErrorCode, MetonaStreamEventType } from '../../types';
|
||||
import { MetonaStreamEventType } from '../../types';
|
||||
import type {
|
||||
IMetonaProviderAdapter,
|
||||
AdapterConfig,
|
||||
@@ -30,11 +36,6 @@ class TestAdapter extends BaseAdapter {
|
||||
// 空实现
|
||||
}
|
||||
|
||||
/** 测试辅助: 暴露 protected mapError */
|
||||
mapErrorPublic(error: unknown) {
|
||||
return this.mapError(error);
|
||||
}
|
||||
|
||||
/** 测试辅助: 暴露 protected throwHttpError */
|
||||
async throwHttpErrorPublic(response: Response, context: string) {
|
||||
return this.throwHttpError(response, context);
|
||||
@@ -56,62 +57,42 @@ function makeAdapter(config: Partial<AdapterConfig> = {}): TestAdapter {
|
||||
});
|
||||
}
|
||||
|
||||
describe('BaseAdapter — mapError 错误映射', () => {
|
||||
it('timeout 消息映射为 NETWORK_TIMEOUT 且可重试', () => {
|
||||
const adapter = makeAdapter();
|
||||
const err = adapter.mapErrorPublic(new Error('Request timeout after 30s'));
|
||||
expect(err.code).toBe(MetonaErrorCode.NETWORK_TIMEOUT);
|
||||
expect(err.retryable).toBe(true);
|
||||
expect(err.provider).toBe('test');
|
||||
});
|
||||
// ===== 错误形状契约（engine.isRetryableError 的输入保证） =====
|
||||
|
||||
it('ECONNREFUSED 映射为 NETWORK_ERROR 且可重试', () => {
|
||||
const adapter = makeAdapter();
|
||||
const err = adapter.mapErrorPublic(new Error('fetch failed: ECONNREFUSED 127.0.0.1:11434'));
|
||||
expect(err.code).toBe(MetonaErrorCode.NETWORK_ERROR);
|
||||
expect(err.retryable).toBe(true);
|
||||
});
|
||||
describe('BaseAdapter — 抛出错误的可判定形状（单轨化契约）', () => {
|
||||
/** 与 engine.isRetryableError 相同的判定逻辑（镜像断言用） */
|
||||
const isRetryableShape = (err: unknown): boolean => {
|
||||
const e = err as { status?: number; code?: string; message?: string };
|
||||
if (e.status === 429) return true;
|
||||
if (e.status && e.status >= 500 && e.status < 600) return true;
|
||||
if (e.code === 'ECONNRESET' || e.code === 'ETIMEDOUT' || e.code === 'ENOTFOUND') return true;
|
||||
const msg = e.message?.toLowerCase() ?? '';
|
||||
if (msg.includes('aborted') || msg.includes('socket hang up')) return true;
|
||||
return false;
|
||||
};
|
||||
|
||||
it('HTTP 401 优先按 status code 映射为 AUTH_INVALID 且不可重试', () => {
|
||||
const adapter = makeAdapter();
|
||||
const e = new Error('API error: 401 Unauthorized');
|
||||
(e as Error & { status: number }).status = 401;
|
||||
const err = adapter.mapErrorPublic(e);
|
||||
expect(err.code).toBe(MetonaErrorCode.AUTH_INVALID);
|
||||
expect(err.retryable).toBe(false);
|
||||
});
|
||||
|
||||
it('HTTP 429 映射为 RATE_LIMITED 且可重试', () => {
|
||||
const adapter = makeAdapter();
|
||||
const e = new Error('429 Too Many Requests');
|
||||
(e as Error & { status: number }).status = 429;
|
||||
const err = adapter.mapErrorPublic(e);
|
||||
expect(err.code).toBe(MetonaErrorCode.RATE_LIMITED);
|
||||
expect(err.retryable).toBe(true);
|
||||
expect(err.retryAfterMs).toBe(5000);
|
||||
});
|
||||
|
||||
it('ContentFilterError 优先映射为 CONTENT_FILTERED', () => {
|
||||
const adapter = makeAdapter();
|
||||
const cf = new ContentFilterError('high risk content', 'MiMo');
|
||||
const err = adapter.mapErrorPublic(cf);
|
||||
expect(err.code).toBe(MetonaErrorCode.CONTENT_FILTERED);
|
||||
expect(err.retryable).toBe(false);
|
||||
});
|
||||
|
||||
it('普通 Error 映射为 UNKNOWN 且不可重试', () => {
|
||||
const adapter = makeAdapter();
|
||||
const err = adapter.mapErrorPublic(new Error('whatever'));
|
||||
expect(err.code).toBe(MetonaErrorCode.UNKNOWN);
|
||||
expect(err.retryable).toBe(false);
|
||||
});
|
||||
});
|
||||
|
||||
describe('BaseAdapter — throwHttpError', () => {
|
||||
function makeResponse(status: number, body: string): Response {
|
||||
return new Response(body, { status, statusText: 'Status' });
|
||||
}
|
||||
|
||||
it('throwHttpError 携带 status；429/5xx 形状可重试，4xx 不可', async () => {
|
||||
const adapter = makeAdapter();
|
||||
try {
|
||||
await adapter.throwHttpErrorPublic(makeResponse(429, 'rate limited'), 'T');
|
||||
expect.fail('should throw');
|
||||
} catch (e) {
|
||||
expect((e as Error & { status?: number }).status).toBe(429);
|
||||
expect(isRetryableShape(e)).toBe(true);
|
||||
}
|
||||
try {
|
||||
await adapter.throwHttpErrorPublic(makeResponse(401, ''), 'T');
|
||||
expect.fail('should throw');
|
||||
} catch (e) {
|
||||
expect((e as Error & { status?: number }).status).toBe(401);
|
||||
expect(isRetryableShape(e)).toBe(false);
|
||||
}
|
||||
});
|
||||
|
||||
it('content_filter 错误体抛出 ContentFilterError（含 status）', async () => {
|
||||
const adapter = makeAdapter();
|
||||
const body = JSON.stringify({ error: { code: 'content_filter', message: 'high risk' } });
|
||||
@@ -126,16 +107,20 @@ describe('BaseAdapter — throwHttpError', () => {
|
||||
}
|
||||
});
|
||||
|
||||
it('普通错误体抛出带 status 属性的 Error（供 isRetryableError 判断）', async () => {
|
||||
it('巨大 HTML 错误体在消息中被截断（v0.6.4）', async () => {
|
||||
const adapter = makeAdapter();
|
||||
await expect(
|
||||
adapter.throwHttpErrorPublic(makeResponse(503, 'Service Unavailable'), 'DeepSeek'),
|
||||
).rejects.toThrow('DeepSeek: 503');
|
||||
const bigHtml = `<html>${'x'.repeat(100_000)}</html>`;
|
||||
let caught: unknown;
|
||||
try {
|
||||
await adapter.throwHttpErrorPublic(makeResponse(503, ''), 'DeepSeek');
|
||||
await adapter.throwHttpErrorPublic(makeResponse(502, bigHtml), 'GW');
|
||||
expect.fail('should throw');
|
||||
} catch (e) {
|
||||
expect((e as Error & { status: number }).status).toBe(503);
|
||||
caught = e;
|
||||
}
|
||||
expect((caught as Error).message.length).toBeLessThan(1_000);
|
||||
expect((caught as Error).message).toContain('[truncated');
|
||||
// 截断不影响 status 判定
|
||||
expect((caught as Error & { status?: number }).status).toBe(502);
|
||||
});
|
||||
});
|
||||
|
||||
@@ -145,7 +130,7 @@ describe('BaseAdapter — getContextWindow', () => {
|
||||
expect(adapter.getContextWindow()).toBe(128_000);
|
||||
});
|
||||
|
||||
it('未配置时返回保守默认值 1M', () => {
|
||||
it('未配置时返回兜底默认值 1M（子类应覆盖真实窗口）', () => {
|
||||
const adapter = makeAdapter();
|
||||
expect(adapter.getContextWindow()).toBe(1_000_000);
|
||||
});
|
||||
@@ -164,10 +149,11 @@ describe('BaseAdapter — listModels / healthCheck', () => {
|
||||
});
|
||||
});
|
||||
|
||||
describe('BaseAdapter — fetchWithTimeout', () => {
|
||||
describe('BaseAdapter — fetchWithTimeout（P3-2 超时分类单轨化）', () => {
|
||||
afterEach(() => {
|
||||
vi.unstubAllGlobals();
|
||||
vi.restoreAllMocks();
|
||||
vi.useRealTimers();
|
||||
});
|
||||
|
||||
it('正常请求返回 Response 并清理 timer', async () => {
|
||||
@@ -181,27 +167,32 @@ describe('BaseAdapter — fetchWithTimeout', () => {
|
||||
expect(fetchMock).toHaveBeenCalledOnce();
|
||||
});
|
||||
|
||||
it('超时后 abort 请求（AbortError）', async () => {
|
||||
it('自身超时 → 显式 ETIMEDOUT + timed out 消息（命中引擎网络超时可重试分支）', async () => {
|
||||
const adapter = makeAdapter();
|
||||
vi.useFakeTimers();
|
||||
const fetchMock = vi.fn(
|
||||
(_url: string, init: RequestInit) =>
|
||||
new Promise<Response>((_resolve, reject) => {
|
||||
init.signal?.addEventListener('abort', () =>
|
||||
reject(new DOMException('Aborted', 'AbortError')),
|
||||
reject(new DOMException('This operation was aborted', 'AbortError')),
|
||||
);
|
||||
}),
|
||||
);
|
||||
vi.stubGlobal('fetch', fetchMock);
|
||||
|
||||
const promise = adapter.fetchWithTimeoutPublic('https://api.test.com/v1/chat', {}, 100);
|
||||
const expectation = expect(promise).rejects.toThrow('Aborted');
|
||||
const expectation = expect(promise).rejects.toSatisfy((e: Error & { code?: string }) => {
|
||||
expect(e.code).toBe('ETIMEDOUT');
|
||||
expect(e.message).toContain('timed out after 100ms');
|
||||
// 关键：消息不再是裸 "Aborted" —— 引擎按网络超时（而非碰巧可重试）分类
|
||||
return true;
|
||||
});
|
||||
vi.advanceTimersByTime(150);
|
||||
await expectation;
|
||||
vi.useRealTimers();
|
||||
});
|
||||
|
||||
it('外部 abort 信号触发请求中断', async () => {
|
||||
it('外部 abort（用户中断）→ 原样 AbortError，不被改写为超时', async () => {
|
||||
const adapter = makeAdapter();
|
||||
const controller = new AbortController();
|
||||
adapter.setAbortSignal(controller.signal);
|
||||
@@ -210,16 +201,20 @@ describe('BaseAdapter — fetchWithTimeout', () => {
|
||||
(_url: string, init: RequestInit) =>
|
||||
new Promise<Response>((_resolve, reject) => {
|
||||
init.signal?.addEventListener('abort', () =>
|
||||
reject(new DOMException('Aborted', 'AbortError')),
|
||||
reject(new DOMException('This operation was aborted', 'AbortError')),
|
||||
);
|
||||
}),
|
||||
);
|
||||
vi.stubGlobal('fetch', fetchMock);
|
||||
|
||||
const promise = adapter.fetchWithTimeoutPublic('https://api.test.com/v1/chat', {}, 30_000);
|
||||
const expectation = expect(promise).rejects.toThrow('Aborted');
|
||||
controller.abort();
|
||||
await expectation;
|
||||
await promise.catch((e: Error & { code?: string }) => {
|
||||
expect(e.name).toBe('AbortError');
|
||||
// 未被转译为字符串型 ETIMEDOUT（注意 Node DOMException 自带数字 code=20）
|
||||
expect(e.code).not.toBe('ETIMEDOUT');
|
||||
expect(e.message).not.toContain('timed out after');
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
|
||||
@@ -0,0 +1,338 @@
|
||||
/**
|
||||
* Provider 请求形态测试矩阵（v0.6.4 P2-6）
|
||||
*
|
||||
* 此前 ollama（599 行）/ anthropic（532 行）两个最复杂的适配器零测试 —— 恰好也是
|
||||
* 本轮审计中缺陷密度最高的文件。本文件通过 mock fetch 记录真实请求体，
|
||||
* 锁定以下契约：
|
||||
*
|
||||
* Anthropic:
|
||||
* A1 消息转换（system 顶层 / user-assistant-tool 三角色映射 / 孤立 tool_result 过滤）
|
||||
* A2 max_tokens 按模型钳制（引擎默认 63488 → sonnet 64000 / opus 32000）
|
||||
* A3 thinking 预算下限保护（小 maxTokens 场景 budget≥1024 且 < max_tokens，此前 API 400）
|
||||
* A4 thinking 开启时不传 temperature；关闭时显式传递
|
||||
*
|
||||
* Ollama:
|
||||
* O1 options 映射（num_predict=numTokens、num_ctx=contextLength、stop、top_p）
|
||||
* O2 think 参数 effort 映射（low→"low"、max→true）与未配置时缺省
|
||||
* O3 图片归一化（data URI 剥前缀；无 URL 触发下载分支时零网络请求）
|
||||
*
|
||||
* Agnes:
|
||||
* G1 思考模式对称性 —— thinkingEnabled=false 必须显式发送 enable_thinking:false
|
||||
*/
|
||||
|
||||
import { describe, it, expect, vi } from 'vitest';
|
||||
|
||||
vi.mock('electron-log', () => ({
|
||||
default: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
|
||||
}));
|
||||
|
||||
import { AnthropicAdapter } from '../anthropic.adapter';
|
||||
import { OllamaAdapter } from '../ollama.adapter';
|
||||
import { MimoAdapter } from '../mimo.adapter';
|
||||
import { AgnesAdapter } from '../agnes-ai.adapter';
|
||||
import type { MetonaRequest } from '../../types';
|
||||
|
||||
/** 安装全局 fetch 捕获器：记录每次请求体并返回一个三家协议都能解析的合成响应 */
|
||||
function captureFetch(): { bodies: Array<Record<string, unknown>> } {
|
||||
const bodies: Array<Record<string, unknown>> = [];
|
||||
// 兼容三家的非流式解析所需的最小字段集：
|
||||
// OpenAI 兼容(agnes): choices[].message/finish_reason；Anthropic: content[]/usage/stop_reason；
|
||||
// Ollama: message/done/prompt_eval_count/eval_count
|
||||
const genericBody = {
|
||||
id: 'cmpl-test',
|
||||
object: 'chat.completion',
|
||||
created: Date.now(),
|
||||
model: 'test-model',
|
||||
choices: [{ index: 0, message: { role: 'assistant', content: 'ok' }, finish_reason: 'stop' }],
|
||||
content: [],
|
||||
usage: {
|
||||
prompt_tokens: 3,
|
||||
completion_tokens: 2,
|
||||
total_tokens: 5,
|
||||
input_tokens: 3,
|
||||
output_tokens: 2,
|
||||
prompt_eval_count: 3,
|
||||
eval_count: 2,
|
||||
},
|
||||
stop_reason: 'end_turn',
|
||||
message: { role: 'assistant', content: 'ok' },
|
||||
done: true,
|
||||
};
|
||||
const fetchMock = vi.fn(async (_url: string | URL, init?: RequestInit) => {
|
||||
bodies.push(JSON.parse(String(init?.body ?? '{}')) as Record<string, unknown>);
|
||||
return new Response(JSON.stringify(genericBody), {
|
||||
status: 200,
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
});
|
||||
});
|
||||
vi.stubGlobal('fetch', fetchMock);
|
||||
return { bodies };
|
||||
}
|
||||
|
||||
function makeRequest(overrides?: Partial<MetonaRequest>): MetonaRequest {
|
||||
return {
|
||||
meta: {
|
||||
sessionId: 's1',
|
||||
iteration: 1,
|
||||
requestId: 'r1',
|
||||
timestamp: Date.now(),
|
||||
agentVersion: 'test',
|
||||
},
|
||||
systemPrompt: {
|
||||
roleDefinition: 'You are Metona.',
|
||||
outputConstraints: 'Be concise.',
|
||||
safetyGuidelines: 'Stay safe.',
|
||||
dynamicReminders: '',
|
||||
},
|
||||
messages: [{ role: 'user', content: 'hi', timestamp: Date.now() }],
|
||||
params: { maxTokens: 63_488, temperature: 0, stream: false },
|
||||
...overrides,
|
||||
};
|
||||
}
|
||||
|
||||
// ===== Anthropic =====
|
||||
|
||||
describe('AnthropicAdapter — 请求体契约', () => {
|
||||
it('A1: system 拼为顶层字段；tool 结果映射为 user 角色 tool_result 块', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({
|
||||
messages: [
|
||||
{ role: 'user', content: 'read it', timestamp: Date.now() },
|
||||
{
|
||||
role: 'assistant',
|
||||
content: null,
|
||||
toolCalls: [
|
||||
{ id: 'tc_1', name: 'read_file', args: { path: 'a.txt' }, iteration: 1, timestamp: Date.now() },
|
||||
],
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
{ role: 'tool', content: null, toolResult: { toolCallId: 'tc_1', toolName: 'read_file', result: 'data', success: true, durationMs: 1, timestamp: Date.now() }, timestamp: Date.now() },
|
||||
// 孤立 tool_result（前面没有对应 tool_use）应被过滤
|
||||
{ role: 'tool', content: null, toolResult: { toolCallId: 'tc_orphan', toolName: 'x', result: '', success: true, durationMs: 1, timestamp: Date.now() }, timestamp: Date.now() },
|
||||
{ role: 'user', content: 'next?', timestamp: Date.now() },
|
||||
],
|
||||
}),
|
||||
);
|
||||
|
||||
const body = bodies[0];
|
||||
expect(body.system).toContain('You are Metona.');
|
||||
expect(Array.isArray(body.messages)).toBe(true);
|
||||
const msgs = body.messages as Array<{ role: string; content: Array<Record<string, unknown>> }>;
|
||||
// tool_use 的 assistant 消息存在且携带 id/name
|
||||
const assistantToolMsg = msgs.find((m) => m.role === 'assistant');
|
||||
expect(assistantToolMsg?.content[0]).toMatchObject({ type: 'tool_use', id: 'tc_1', name: 'read_file' });
|
||||
// tool 结果以 user 角色 tool_result 形态出现且配对 id 正确；孤立者被丢弃
|
||||
const toolResultBlocks = msgs.flatMap((m) =>
|
||||
m.content.filter((c) => c.type === 'tool_result'),
|
||||
);
|
||||
expect(toolResultBlocks).toHaveLength(1);
|
||||
expect(toolResultBlocks[0].tool_use_id).toBe('tc_1');
|
||||
});
|
||||
|
||||
it('A2: max_tokens 按模型上限钳制（63488 → sonnet 64000 / opus 32000）', async () => {
|
||||
const sonnet = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
const opus = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'claude-opus-4-1',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await sonnet.send(makeRequest());
|
||||
await opus.send(makeRequest());
|
||||
// 引擎默认 63488 低于 sonnet 上限 64000 → 原样保留；opus 上限 32000 → 钳制生效
|
||||
expect(bodies[0].max_tokens).toBe(63_488);
|
||||
expect(bodies[1].max_tokens).toBe(32_000);
|
||||
});
|
||||
|
||||
it('A3: 小 maxTokens 时 thinking budget 不跌破协议下限 1024（v0.6.4 边界加固）', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'claude-haiku-4-5',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({ params: { maxTokens: 1500, temperature: 0, stream: false, thinkingEnabled: true, thinkingEffort: 'low' } }),
|
||||
);
|
||||
const body = bodies[0];
|
||||
const thinking = body.thinking as { type: string; budget_tokens: number };
|
||||
// max_tokens 被抬升到安全下限，budget 落在 [1024, max_tokens/2] 区间内
|
||||
expect(body.max_tokens as number).toBeGreaterThanOrEqual(2048);
|
||||
expect(thinking.budget_tokens).toBeGreaterThanOrEqual(1024);
|
||||
expect(thinking.budget_tokens).toBeLessThanOrEqual((body.max_tokens as number) / 2);
|
||||
});
|
||||
|
||||
it('A4: thinking 开启不传 temperature；关闭时显式传递', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://a.test',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({ params: { maxTokens: 4096, temperature: 0.7, stream: false, thinkingEnabled: true } }),
|
||||
);
|
||||
expect(bodies[0].temperature).toBeUndefined();
|
||||
expect(bodies[0].thinking).toBeDefined();
|
||||
|
||||
await adapter.send(
|
||||
makeRequest({ params: { maxTokens: 4096, temperature: 0.7, stream: false, thinkingEnabled: false } }),
|
||||
);
|
||||
expect(bodies[1].temperature).toBe(0.7);
|
||||
expect(bodies[1].thinking).toBeUndefined();
|
||||
});
|
||||
});
|
||||
|
||||
// ===== Ollama =====
|
||||
|
||||
describe('OllamaAdapter — 请求体契约', () => {
|
||||
function makeOllama(): OllamaAdapter {
|
||||
return new OllamaAdapter({
|
||||
provider: 'ollama',
|
||||
baseURL: 'http://localhost:11434',
|
||||
defaultModel: 'qwen3',
|
||||
});
|
||||
}
|
||||
|
||||
it('O1: options 映射 num_predict/num_ctx/stop/top_p/temperature', async () => {
|
||||
const adapter = makeOllama();
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({
|
||||
params: {
|
||||
maxTokens: 8192,
|
||||
temperature: 0.3,
|
||||
topP: 0.9,
|
||||
stream: false,
|
||||
contextLength: 16384,
|
||||
stopSequences: ['STOP'],
|
||||
},
|
||||
}),
|
||||
);
|
||||
const options = bodies[0].options as Record<string, unknown>;
|
||||
expect(options.num_predict).toBe(8192);
|
||||
expect(options.num_ctx).toBe(16384);
|
||||
expect(options.temperature).toBe(0.3);
|
||||
expect(options.top_p).toBe(0.9);
|
||||
expect(options.stop).toEqual(['STOP']);
|
||||
});
|
||||
|
||||
it('O2: think 参数 effort 映射（low→"low"、max→true）；未开启思考时缺省', async () => {
|
||||
const adapter = makeOllama();
|
||||
const { bodies } = captureFetch();
|
||||
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false, thinkingEnabled: true, thinkingEffort: 'low' } }));
|
||||
expect(bodies[0].think).toBe('low');
|
||||
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false, thinkingEnabled: true, thinkingEffort: 'max' } }));
|
||||
expect(bodies[1].think).toBe(true);
|
||||
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false, thinkingEnabled: false } }));
|
||||
expect(bodies[2].think).toBeUndefined();
|
||||
});
|
||||
|
||||
it('O3: data URI 图片剥前缀转纯 base64 数组（无网络下载路径触发）', async () => {
|
||||
const adapter = makeOllama();
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(
|
||||
makeRequest({
|
||||
messages: [
|
||||
{
|
||||
role: 'user',
|
||||
content: '看图',
|
||||
images: [{ url: 'data:image/png;base64,iVBORw0KGgoAAAANSU', detail: 'auto' }],
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
],
|
||||
}),
|
||||
);
|
||||
const messages = bodies[0].messages as Array<Record<string, unknown>>;
|
||||
const userMsg = messages[messages.length - 1];
|
||||
expect(userMsg.images).toEqual(['iVBORw0KGgoAAAANSU']);
|
||||
});
|
||||
});
|
||||
|
||||
// ===== MiMo providerOptions（v0.6.4 P4-3） =====
|
||||
|
||||
describe('MimoAdapter — 服务端能力扩展（providerOptions）', () => {
|
||||
it('enableWebSearch 开启时附加 {type:web_search} 服务端工具', async () => {
|
||||
const adapter = new MimoAdapter({
|
||||
provider: 'mimo',
|
||||
baseURL: 'http://m.test/v1',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'mimo-v2.5',
|
||||
providerOptions: { enableWebSearch: true },
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await adapter.send(makeRequest());
|
||||
const tools = bodies[0].tools as Array<Record<string, unknown>>;
|
||||
expect(tools.some((tc) => (tc as { type?: string }).type === 'web_search')).toBe(true);
|
||||
expect(bodies[0].tool_choice).toBe('auto');
|
||||
});
|
||||
|
||||
it('responseFormatJson 开启时写入 response_format json_object；默认不写', async () => {
|
||||
const on = new MimoAdapter({
|
||||
provider: 'mimo',
|
||||
baseURL: 'http://m.test/v1',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'mimo-v2.5',
|
||||
providerOptions: { responseFormatJson: true },
|
||||
});
|
||||
const off = new MimoAdapter({
|
||||
provider: 'mimo',
|
||||
baseURL: 'http://m.test/v1',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'mimo-v2.5',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
await on.send(makeRequest());
|
||||
await off.send(makeRequest());
|
||||
expect(bodies[0].response_format).toEqual({ type: 'json_object' });
|
||||
expect(bodies[1].response_format).toBeUndefined();
|
||||
});
|
||||
});
|
||||
|
||||
// ===== Agnes =====
|
||||
|
||||
describe('AgnesAdapter — 思考模式对称性（v0.6.4）', () => {
|
||||
it('G1: thinkingEnabled=false 显式发送 enable_thinking:false（此前无法关闭服务端默认思考）', async () => {
|
||||
const adapter = new AgnesAdapter({
|
||||
provider: 'agnes',
|
||||
baseURL: 'http://g.test/v1',
|
||||
apiKey: 'k',
|
||||
defaultModel: 'agnes-2.0-flash',
|
||||
});
|
||||
const { bodies } = captureFetch();
|
||||
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false, thinkingEnabled: true, thinkingEffort: 'high' } }));
|
||||
expect(
|
||||
((bodies[0].chat_template_kwargs as Record<string, unknown>) ?? {}).enable_thinking,
|
||||
).toBe(true);
|
||||
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false, thinkingEnabled: false } }));
|
||||
expect(
|
||||
((bodies[1].chat_template_kwargs as Record<string, unknown>) ?? {}).enable_thinking,
|
||||
).toBe(false);
|
||||
|
||||
// 未配置 thinkingEnabled 同样视为关闭（显式 disabled 保持与服务端默认的确定性）
|
||||
await adapter.send(makeRequest({ params: { maxTokens: 4096, temperature: 0, stream: false } }));
|
||||
expect(
|
||||
((bodies[2].chat_template_kwargs as Record<string, unknown>) ?? {}).enable_thinking,
|
||||
).toBe(false);
|
||||
});
|
||||
});
|
||||
@@ -216,7 +216,7 @@ describe('parseOpenAICompatibleResponse — 非流式响应', () => {
|
||||
expect(result.toolCalls![0].args).toEqual({ cmd: 'ls' });
|
||||
});
|
||||
|
||||
it('损坏的 tool_calls arguments 降级为空对象', () => {
|
||||
it('损坏的 tool_calls arguments 转为 _truncatedArguments 自愈载荷（v0.6.4: 不再静默降级 {}）', () => {
|
||||
const result = parseOpenAICompatibleResponse({
|
||||
choices: [
|
||||
{
|
||||
@@ -229,7 +229,9 @@ describe('parseOpenAICompatibleResponse — 非流式响应', () => {
|
||||
],
|
||||
usage: {},
|
||||
});
|
||||
expect(result.toolCalls![0].args).toEqual({});
|
||||
const args = result.toolCalls![0].args as Record<string, unknown>;
|
||||
expect(args._truncatedArguments).toBe(true);
|
||||
expect(String(args._truncatedReason)).toContain('truncated');
|
||||
});
|
||||
|
||||
it('mapOpenAIFinishReason 覆盖 MiMo repetition_truncation', () => {
|
||||
|
||||
@@ -0,0 +1,540 @@
|
||||
/**
|
||||
* 流式上游错误帧 + 全线截断自愈测试（v0.6.4）
|
||||
*
|
||||
* 背景（v0.6.3 审计遗留）：
|
||||
* 1. 错误帧黑洞 —— OpenAI 兼容网关中途发送的 `{"error":{...}}` 数据帧被解析器
|
||||
* 整帧吞掉（零日志），任何上游错误都伪装成"干净的空回复 + 正常 DONE"，
|
||||
* 且以普通事件而非异常出现，绕过引擎的重试/故障转移通道。
|
||||
* 2. 截断自愈只修了 OpenAI 共享层 —— v0.6.3 的 _truncatedArguments 修复未覆盖：
|
||||
* - Anthropic：content_block_stop 解析失败静默 args={}；断流时未完成块整体蒸发
|
||||
* - Ollama：NDJSON 坏参抛错落入外层 catch，工具调用丢弃且同 chunk USAGE/DONE 被跳过
|
||||
* - 非流式 parseOpenAICompatibleResponse：坏参仍静默 {}
|
||||
* - 引擎兜底缓冲 finalizeToolCallsFromBuffer：坏参静默 {}
|
||||
*
|
||||
* 本文件锁定以下契约：
|
||||
* A. 上游错误帧 → 抛出携带归一化 status 的 SseUpstreamError（可驱动重试判定）
|
||||
* B. finish_reason=content_filter → ContentFilterError（终态、不重试）
|
||||
* C. data:{无空格} 变体正常解析
|
||||
* D. 非流式/Ollama/Anthropic 截断参数统一转 _truncatedArguments 自愈载荷
|
||||
* E. Ollama 坏参不再吞掉同 chunk 的 done/USAGE 处理
|
||||
* F. Anthropic 断流时未完成 tool_use 块 flush 为自愈调用 + DONE
|
||||
*/
|
||||
|
||||
import { describe, it, expect, vi } from 'vitest';
|
||||
|
||||
vi.mock('electron-log', () => ({
|
||||
default: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
|
||||
}));
|
||||
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse, SseUpstreamError } from '../shared/sse-stream';
|
||||
import { ContentFilterError } from '../base-adapter';
|
||||
import { OllamaAdapter } from '../ollama.adapter';
|
||||
import { AnthropicAdapter } from '../anthropic.adapter';
|
||||
import { MetonaErrorCode, MetonaStreamEventType } from '../../types';
|
||||
import type { MetonaRequest } from '../../types';
|
||||
|
||||
const encoder = new TextEncoder();
|
||||
|
||||
function makeStream(lines: string[]): ReadableStream<Uint8Array> {
|
||||
const payload = lines.join('\n') + '\n';
|
||||
return new ReadableStream<Uint8Array>({
|
||||
start(controller) {
|
||||
controller.enqueue(encoder.encode(payload));
|
||||
controller.close();
|
||||
},
|
||||
});
|
||||
}
|
||||
|
||||
async function collectExpectingThrow(stream: ReadableStream<Uint8Array>): Promise<unknown> {
|
||||
try {
|
||||
for await (const _ev of parseSSEStream(stream, 'r_test', 's_test', 1)) {
|
||||
void _ev;
|
||||
}
|
||||
} catch (err) {
|
||||
return err;
|
||||
}
|
||||
throw new Error('expected parseSSEStream to throw but it completed normally');
|
||||
}
|
||||
|
||||
function sseData(json: unknown): string {
|
||||
return `data: ${JSON.stringify(json)}`;
|
||||
}
|
||||
|
||||
// ===== A. 上游错误帧 → 抛出结构化异常 =====
|
||||
|
||||
describe('parseSSEStream — 上游错误帧（v0.6.4 错误帧黑洞根治）', () => {
|
||||
it('顶层 error 帧（含数值 status）→ 抛出携带该 status 的 SseUpstreamError', async () => {
|
||||
const err = await collectExpectingThrow(
|
||||
makeStream([sseData({ error: { message: 'Gateway timeout', status: 504 } })]),
|
||||
);
|
||||
expect(err).toBeInstanceOf(SseUpstreamError);
|
||||
expect((err as SseUpstreamError).status).toBe(504);
|
||||
expect((err as Error).message).toContain('Gateway timeout');
|
||||
});
|
||||
|
||||
it('choices[0].error 变体包装也能检出', async () => {
|
||||
const err = await collectExpectingThrow(
|
||||
makeStream([
|
||||
sseData({
|
||||
choices: [{ error: { message: 'bad gateway', code: 'upstream_failure', status: 502 } }],
|
||||
}),
|
||||
]),
|
||||
);
|
||||
expect(err).toBeInstanceOf(SseUpstreamError);
|
||||
expect((err as SseUpstreamError).status).toBe(502);
|
||||
});
|
||||
|
||||
it('字符串型顶层 error 也能检出', async () => {
|
||||
const err = await collectExpectingThrow(makeStream(['data: {"error":"service unavailable"}']));
|
||||
expect(err).toBeInstanceOf(SseUpstreamError);
|
||||
expect((err as Error).message).toContain('service unavailable');
|
||||
});
|
||||
|
||||
it('providerCode 归一化：rate_limit_exceeded 无数值 status → 映射 429（可重试）', async () => {
|
||||
const err = await collectExpectingThrow(
|
||||
makeStream([
|
||||
sseData({ error: { code: 'rate_limit_exceeded', message: 'too many requests' } }),
|
||||
]),
|
||||
);
|
||||
expect(err).toBeInstanceOf(SseUpstreamError);
|
||||
expect((err as SseUpstreamError).status).toBe(429);
|
||||
expect((err as SseUpstreamError).providerCode).toBe('rate_limit_exceeded');
|
||||
});
|
||||
|
||||
it('insufficient_quota → 402；invalid_api_key → 401（不可重试区间）', async () => {
|
||||
const e1 = await collectExpectingThrow(
|
||||
makeStream([sseData({ error: { code: 'insufficient_quota', message: 'quota exceeded' } })]),
|
||||
);
|
||||
expect((e1 as SseUpstreamError).status).toBe(402);
|
||||
|
||||
const e2 = await collectExpectingThrow(
|
||||
makeStream([
|
||||
sseData({ error: { code: 'invalid_api_key', message: 'Incorrect API key provided' } }),
|
||||
]),
|
||||
);
|
||||
expect((e2 as SseUpstreamError).status).toBe(401);
|
||||
});
|
||||
|
||||
it('含 content_filter 码的错误帧 → ContentFilterError（复用专用类型）', async () => {
|
||||
const err = await collectExpectingThrow(
|
||||
makeStream([
|
||||
sseData({ error: { code: 'content_filter', message: 'rejected by safety policy' } }),
|
||||
]),
|
||||
);
|
||||
expect(err).toBeInstanceOf(ContentFilterError);
|
||||
});
|
||||
|
||||
it('isRetryable 契约对齐：错误带 status 时，engine.isRetryableError 的 429/5xx 判定可直接命中', async () => {
|
||||
// 用与引擎 isRetryableError 相同的判定逻辑验证字段形态
|
||||
const isRetryableShape = (err: unknown): boolean => {
|
||||
const e = err as { status?: number; message?: string };
|
||||
if (e.status === 429) return true;
|
||||
if (e.status && e.status >= 500 && e.status < 600) return true;
|
||||
return false;
|
||||
};
|
||||
const rateLimited = await collectExpectingThrow(
|
||||
makeStream([sseData({ error: { code: 'rate_limit_exceeded', message: 'rl' } })]),
|
||||
);
|
||||
expect(isRetryableShape(rateLimited)).toBe(true);
|
||||
const authFail = await collectExpectingThrow(
|
||||
makeStream([sseData({ error: { code: 'invalid_api_key', message: 'auth' } })]),
|
||||
);
|
||||
expect(isRetryableShape(authFail)).toBe(false);
|
||||
});
|
||||
|
||||
it('正常数据帧不含 error 字段时不受影响（回归）', async () => {
|
||||
// choices[0] 中存在 delta 但无 error → 正常产出文本增量并 DONE 收尾
|
||||
const events: string[] = [];
|
||||
const stream = makeStream([
|
||||
sseData({ choices: [{ delta: { content: 'hello' } }] }),
|
||||
'data: [DONE]',
|
||||
]);
|
||||
for await (const ev of parseSSEStream(stream, 'r', 's', 1)) {
|
||||
events.push(ev.type);
|
||||
}
|
||||
expect(events).toContain(MetonaStreamEventType.TEXT_DELTA);
|
||||
expect(events[events.length - 1]).toBe(MetonaStreamEventType.DONE);
|
||||
});
|
||||
});
|
||||
|
||||
// ===== B/C. content_filter 终止映射 + 无空格 data 变体 =====
|
||||
|
||||
describe('parseSSEStream — content_filter 与行格式兼容', () => {
|
||||
it('finish_reason=content_filter → 抛出 ContentFilterError（不再当普通结束）', async () => {
|
||||
const err = await collectExpectingThrow(
|
||||
makeStream([sseData({ choices: [{ delta: {}, finish_reason: 'content_filter' }] })]),
|
||||
);
|
||||
expect(err).toBeInstanceOf(ContentFilterError);
|
||||
});
|
||||
|
||||
it('data:{}（无空格）变体被正常解析（此前整帧跳过）', async () => {
|
||||
const events: Array<{ type: string }> = [];
|
||||
const stream = new ReadableStream<Uint8Array>({
|
||||
start(controller) {
|
||||
controller.enqueue(encoder.encode('data:{"choices":[{"delta":{"content":"hi"}}]}\n'));
|
||||
controller.enqueue(encoder.encode('data:[DONE]\n'));
|
||||
controller.close();
|
||||
},
|
||||
});
|
||||
for await (const ev of parseSSEStream(stream, 'r', 's', 1)) {
|
||||
events.push({ type: ev.type });
|
||||
}
|
||||
expect(events.some((e) => e.type === MetonaStreamEventType.TEXT_DELTA)).toBe(true);
|
||||
expect(events[events.length - 1].type).toBe(MetonaStreamEventType.DONE);
|
||||
});
|
||||
});
|
||||
|
||||
// ===== D. 非流式截断自愈同步 =====
|
||||
|
||||
describe('parseOpenAICompatibleResponse — 非流式截断自愈（v0.6.4 同步）', () => {
|
||||
it('坏 JSON arguments 不再静默 {}，转为 _truncatedArguments 载荷', () => {
|
||||
const result = parseOpenAICompatibleResponse({
|
||||
choices: [
|
||||
{
|
||||
message: {
|
||||
role: 'assistant',
|
||||
content: null,
|
||||
tool_calls: [
|
||||
{
|
||||
id: 'call_1',
|
||||
function: { name: 'write_file', arguments: '{"file_path": "a.html", "con' },
|
||||
},
|
||||
],
|
||||
},
|
||||
finish_reason: 'tool_calls',
|
||||
},
|
||||
],
|
||||
usage: { prompt_tokens: 10, completion_tokens: 5, total_tokens: 15 },
|
||||
});
|
||||
|
||||
expect(result.toolCalls).toHaveLength(1);
|
||||
const args = result.toolCalls![0].args as Record<string, unknown>;
|
||||
expect(args._truncatedArguments).toBe(true);
|
||||
expect(String(args._truncatedReason)).toContain('truncated');
|
||||
});
|
||||
|
||||
it('合法对象型 arguments 保持原样（回归）', () => {
|
||||
const result = parseOpenAICompatibleResponse({
|
||||
choices: [
|
||||
{
|
||||
message: {
|
||||
role: 'assistant',
|
||||
tool_calls: [{ id: 'c1', function: { name: 'think', arguments: '{"a":1}' } }],
|
||||
},
|
||||
finish_reason: 'tool_calls',
|
||||
},
|
||||
],
|
||||
});
|
||||
expect(result.toolCalls![0].args).toEqual({ a: 1 });
|
||||
});
|
||||
});
|
||||
|
||||
// ===== E/F. Ollama NDJSON 与 Anthropic 事件机 =====
|
||||
|
||||
/** 构造全局 fetch mock：返回给定行的 NDJSON/SSE 流 */
|
||||
function mockFetchWithLines(lines: string[]): ReturnType<typeof vi.fn> {
|
||||
const payload = encoder.encode(lines.join('\n') + '\n');
|
||||
const body = new ReadableStream<Uint8Array>({
|
||||
start(controller) {
|
||||
controller.enqueue(payload);
|
||||
controller.close();
|
||||
},
|
||||
});
|
||||
const fetchMock = vi.fn().mockResolvedValue(
|
||||
new Response(body, {
|
||||
status: 200,
|
||||
headers: { 'Content-Type': 'application/x-ndjson' },
|
||||
}),
|
||||
);
|
||||
vi.stubGlobal('fetch', fetchMock);
|
||||
return fetchMock;
|
||||
}
|
||||
|
||||
const baseRequest: MetonaRequest = {
|
||||
meta: { sessionId: 's1', iteration: 1, requestId: 'r1', timestamp: Date.now(), agentVersion: 'test' },
|
||||
systemPrompt: { roleDefinition: 'rd', outputConstraints: '', safetyGuidelines: '' },
|
||||
messages: [{ role: 'user', content: 'hi', timestamp: Date.now() }],
|
||||
params: { maxTokens: 4096, temperature: 0, stream: true },
|
||||
};
|
||||
|
||||
describe('OllamaAdapter.sendStream — NDJSON 截断自愈（v0.6.4）', () => {
|
||||
it('坏 JSON arguments → _truncatedArguments 工具调用，且后续 done chunk 的 USAGE/DONE 不再被吞掉', async () => {
|
||||
const adapter = new OllamaAdapter({
|
||||
provider: 'ollama',
|
||||
baseURL: 'http://localhost:11434',
|
||||
defaultModel: 'qwen3',
|
||||
});
|
||||
mockFetchWithLines([
|
||||
JSON.stringify({
|
||||
model: 'qwen3',
|
||||
message: {
|
||||
role: 'assistant',
|
||||
content: '',
|
||||
tool_calls: [{ function: { name: 'write_file', arguments: '{"path": "a.txt", "cont' } }],
|
||||
},
|
||||
}),
|
||||
// 关键：同一响应流中随后仍有收尾 chunk（原实现外层 catch 会跳过这些处理）
|
||||
JSON.stringify({
|
||||
model: 'qwen3',
|
||||
message: { role: 'assistant', content: '' },
|
||||
done: true,
|
||||
done_reason: 'stop',
|
||||
prompt_eval_count: 11,
|
||||
eval_count: 7,
|
||||
}),
|
||||
]);
|
||||
|
||||
const events: string[] = [];
|
||||
let usageInputTokens = -1;
|
||||
for await (const ev of adapter.sendStream(baseRequest)) {
|
||||
events.push(ev.type);
|
||||
if (ev.type === MetonaStreamEventType.USAGE) usageInputTokens = ev.usage!.inputTokens ?? 0;
|
||||
}
|
||||
|
||||
// 流不再被坏参打断：usage 与 done 都到达
|
||||
expect(usageInputTokens).toBe(11);
|
||||
expect(events[events.length - 1]).toBe(MetonaStreamEventType.DONE);
|
||||
});
|
||||
|
||||
it('自愈载荷内容正确（_truncatedArguments=true + reason 含 truncated）', async () => {
|
||||
const adapter = new OllamaAdapter({
|
||||
provider: 'ollama',
|
||||
baseURL: 'http://localhost:11434',
|
||||
defaultModel: 'qwen3',
|
||||
});
|
||||
mockFetchWithLines([
|
||||
JSON.stringify({
|
||||
model: 'm',
|
||||
message: {
|
||||
role: 'assistant',
|
||||
tool_calls: [{ function: { name: 'read_file', arguments: '{"file_path": "b.t' } }],
|
||||
},
|
||||
done: false,
|
||||
}),
|
||||
JSON.stringify({ model: 'm', message: { role: 'assistant', content: '' }, done: true }),
|
||||
]);
|
||||
|
||||
let completeArgs: Record<string, unknown> | undefined;
|
||||
for await (const ev of adapter.sendStream(baseRequest)) {
|
||||
if (ev.type === MetonaStreamEventType.TOOL_CALL_COMPLETE) {
|
||||
completeArgs = ev.toolCall!.args as Record<string, unknown>;
|
||||
}
|
||||
}
|
||||
expect(completeArgs).toBeDefined();
|
||||
expect(completeArgs!._truncatedArguments).toBe(true);
|
||||
expect(String(completeArgs!._truncatedReason)).toContain('truncated');
|
||||
});
|
||||
});
|
||||
|
||||
describe('AnthropicAdapter.sendStream — 事件机截断自愈 + 断流 flush（v0.6.4）', () => {
|
||||
it('缺口 A：content_block_stop 时坏 JSON → _truncatedArguments（不再静默 {}）', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://anthropic.test',
|
||||
apiKey: 'sk-test',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
mockFetchWithLines([
|
||||
'event: content_block_start',
|
||||
sseData({
|
||||
type: 'content_block_start',
|
||||
index: 0,
|
||||
content_block: { type: 'tool_use', id: 'toolu_1', name: 'write_file' },
|
||||
}),
|
||||
'event: content_block_delta',
|
||||
sseData({
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: { type: 'input_json_delta', partial_json: '{"file_path": "a.html", "con' },
|
||||
}),
|
||||
'event: content_block_stop',
|
||||
sseData({ type: 'content_block_stop', index: 0 }),
|
||||
'event: message_stop',
|
||||
sseData({ type: 'message_stop' }),
|
||||
]);
|
||||
|
||||
let completeArgs: Record<string, unknown> | undefined;
|
||||
let completeId: string | undefined;
|
||||
for await (const ev of adapter.sendStream(baseRequest)) {
|
||||
if (ev.type === MetonaStreamEventType.TOOL_CALL_COMPLETE) {
|
||||
completeArgs = ev.toolCall!.args as Record<string, unknown>;
|
||||
completeId = ev.toolCall!.id;
|
||||
}
|
||||
}
|
||||
expect(completeArgs).toBeDefined();
|
||||
expect(completeArgs!._truncatedArguments).toBe(true);
|
||||
// 保留上游原始 block id（非 nanoid 重造）
|
||||
expect(completeId).toBe('toolu_1');
|
||||
});
|
||||
|
||||
it('缺口 B：断流未完成 tool_use 块 → flush 为自愈调用 + 补发 DONE（不再整体蒸发）', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://anthropic.test',
|
||||
apiKey: 'sk-test',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
// 有 content_block_start，但流在 content_block_stop/message_stop 之前断开
|
||||
const payload =
|
||||
'event: message_start\ndata: {"type":"message_start","message":{"role":"assistant","usage":{"input_tokens":42}}}\n\n' +
|
||||
'event: content_block_start\ndata: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_X","name":"edit_file"}}\n\n' +
|
||||
'event: content_block_delta\ndata: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\\"pa"}}\n\n';
|
||||
const body = new ReadableStream<Uint8Array>({
|
||||
start(controller) {
|
||||
controller.enqueue(encoder.encode(payload));
|
||||
controller.close();
|
||||
},
|
||||
});
|
||||
vi.stubGlobal(
|
||||
'fetch',
|
||||
vi.fn().mockResolvedValue(new Response(body, { status: 200 })),
|
||||
);
|
||||
|
||||
const events: Array<{ type: string; toolCallId?: string; toolCallName?: string }> = [];
|
||||
for await (const ev of adapter.sendStream(baseRequest)) {
|
||||
events.push({
|
||||
type: ev.type,
|
||||
toolCallId: ev.toolCall?.id,
|
||||
toolCallName: ev.toolCall?.name,
|
||||
});
|
||||
}
|
||||
|
||||
const complete = events.find((e) => e.type === MetonaStreamEventType.TOOL_CALL_COMPLETE);
|
||||
// 核心契约：断流前缓冲中的 block 必须以 TOOL_CALL_COMPLETE 产出（引擎才不会误判空回复完成）
|
||||
expect(complete).toBeDefined();
|
||||
expect(complete!.toolCallId).toBe('toolu_X');
|
||||
expect(complete!.toolCallName).toBe('edit_file');
|
||||
expect(events[events.length - 1].type).toBe(MetonaStreamEventType.DONE);
|
||||
});
|
||||
|
||||
it('error 事件 → 抛出携带归一化 status 的异常（overloaded → 529 可重试语义）', async () => {
|
||||
const adapter = new AnthropicAdapter({
|
||||
provider: 'anthropic',
|
||||
baseURL: 'http://anthropic.test',
|
||||
apiKey: 'sk-test',
|
||||
defaultModel: 'claude-sonnet-4-5',
|
||||
});
|
||||
mockFetchWithLines([
|
||||
'event: error',
|
||||
sseData({ type: 'error', error: { type: 'overloaded_error', message: 'Overloaded' } }),
|
||||
]);
|
||||
|
||||
let caught: unknown;
|
||||
try {
|
||||
for await (const _ev of adapter.sendStream(baseRequest)) {
|
||||
void _ev;
|
||||
}
|
||||
} catch (err) {
|
||||
caught = err;
|
||||
}
|
||||
expect(caught).toBeDefined();
|
||||
expect((caught as Error & { status?: number }).status).toBe(529);
|
||||
expect((caught as Error).message).toContain('Overloaded');
|
||||
});
|
||||
});
|
||||
|
||||
// ===== G. 引擎侧 ERROR 事件保留结构化码 =====
|
||||
|
||||
describe('MetonaErrorCode — CONTENT_FILTERED 枚举契约（finish 映射依赖）', () => {
|
||||
it('code 值稳定为 content_filtered', () => {
|
||||
expect(MetonaErrorCode.CONTENT_FILTERED).toBe('content_filtered');
|
||||
});
|
||||
});
|
||||
|
||||
// ===== H. 引擎集成：错误帧异常进入重试/故障转移通道；ERROR 码映射 CONTENT_FILTERED =====
|
||||
|
||||
import { AgentLoopEngine } from '../../agent-loop/engine';
|
||||
import { TerminationReason } from '../../agent-loop/types';
|
||||
import type { IMetonaProviderAdapter, MetonaResponse, MetonaStreamEvent } from '../../types';
|
||||
|
||||
function textDone(text: string): MetonaStreamEvent[] {
|
||||
return [
|
||||
{ type: MetonaStreamEventType.TEXT_DELTA, requestId: 'r1', sessionId: 's1', iteration: 1, seq: 0, timestamp: Date.now(), delta: text },
|
||||
{ type: MetonaStreamEventType.DONE, requestId: 'r1', sessionId: 's1', iteration: 1, seq: 1, timestamp: Date.now() },
|
||||
];
|
||||
}
|
||||
|
||||
describe('AgentLoopEngine 集成 — v0.6.4 错误通道单轨化', () => {
|
||||
const userMessage = { role: 'user' as const, content: 'hi', timestamp: Date.now() };
|
||||
const systemPrompt = { roleDefinition: '', outputConstraints: '', safetyGuidelines: '' };
|
||||
|
||||
function scriptedAdapter(
|
||||
behaviors: Array<{ throws?: Error; events?: MetonaStreamEvent[] }>,
|
||||
): { adapter: IMetonaProviderAdapter; calls: () => number } {
|
||||
let call = 0;
|
||||
const base: IMetonaProviderAdapter = {
|
||||
providerId: 'mock',
|
||||
supportedModels: ['m'],
|
||||
supportsToolCalling: true,
|
||||
supportsThinking: false,
|
||||
getContextWindow: () => 1_000_000,
|
||||
send: async (): Promise<MetonaResponse> => ({
|
||||
meta: { requestId: 'r', provider: 'mock', model: 'm', latencyMs: 0, timestamp: Date.now() },
|
||||
content: '',
|
||||
usage: { inputTokens: 1, outputTokens: 1, totalTokens: 2 },
|
||||
finishReason: 'stop' as never,
|
||||
}),
|
||||
sendStream: async function* (): AsyncIterable<MetonaStreamEvent> {
|
||||
const b = behaviors[Math.min(call, behaviors.length - 1)];
|
||||
call++;
|
||||
if (b.throws) throw b.throws;
|
||||
for (const ev of b.events ?? []) yield ev;
|
||||
},
|
||||
setAbortSignal: vi.fn(),
|
||||
healthCheck: async () => true,
|
||||
};
|
||||
return { adapter: base, calls: () => call };
|
||||
}
|
||||
|
||||
it('SseUpstreamError(429) 首次失败 → 引擎指数退避重试后成功（不再落入 UNKNOWN 终态）', async () => {
|
||||
vi.useFakeTimers();
|
||||
try {
|
||||
const { adapter } = scriptedAdapter([
|
||||
{ throws: new SseUpstreamError('rate limited', { status: 429 }) },
|
||||
{ events: textDone('recovered answer') },
|
||||
]);
|
||||
const engine = new AgentLoopEngine({ retryCount: 3 }, adapter);
|
||||
// 推进退避定时器（1s/2s/4s + jitter 上限）
|
||||
const runPromise = engine.runStream(userMessage, 's1', [], systemPrompt);
|
||||
await vi.advanceTimersByTimeAsync(10_000);
|
||||
const output = await runPromise;
|
||||
expect(output.terminationReason).toBe(TerminationReason.COMPLETED);
|
||||
expect(output.finalAnswer).toBe('recovered answer');
|
||||
} finally {
|
||||
vi.useRealTimers();
|
||||
}
|
||||
});
|
||||
|
||||
it('引擎收到的流内 ERROR 带 code=content_filtered → finish 发出 CONTENT_FILTERED 错误事件', async () => {
|
||||
const { adapter } = scriptedAdapter([
|
||||
{
|
||||
events: [
|
||||
{
|
||||
type: MetonaStreamEventType.ERROR,
|
||||
requestId: 'r1',
|
||||
sessionId: 's1',
|
||||
iteration: 1,
|
||||
seq: 0,
|
||||
timestamp: Date.now(),
|
||||
error: {
|
||||
code: MetonaErrorCode.CONTENT_FILTERED,
|
||||
message: '内容被安全审核拦截',
|
||||
retryable: false,
|
||||
},
|
||||
},
|
||||
{ type: MetonaStreamEventType.DONE, requestId: 'r1', sessionId: 's1', iteration: 1, seq: 1, timestamp: Date.now() },
|
||||
],
|
||||
},
|
||||
]);
|
||||
const engine = new AgentLoopEngine({ retryCount: 0 }, adapter);
|
||||
const errorEvents: Array<{ code?: string; message?: string }> = [];
|
||||
engine.on('streamEvent', (ev: MetonaStreamEvent) => {
|
||||
if (ev.type === MetonaStreamEventType.ERROR) {
|
||||
errorEvents.push({ code: ev.error?.code, message: ev.error?.message });
|
||||
}
|
||||
});
|
||||
const output = await engine.runStream(userMessage, 's1', [], systemPrompt);
|
||||
expect(output.terminationReason).toBe(TerminationReason.ERROR);
|
||||
expect(errorEvents[errorEvents.length - 1]?.code).toBe(MetonaErrorCode.CONTENT_FILTERED);
|
||||
});
|
||||
});
|
||||
@@ -3,26 +3,21 @@
|
||||
*
|
||||
* OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、多模态（图片 — URL + Base64）。
|
||||
*
|
||||
* 独立继承 BaseAdapter，通过 shared/openai-format 和 shared/sse-stream 复用
|
||||
* OpenAI 兼容格式构建和 SSE 流式解析逻辑。不与其他 Provider Adapter 耦合。
|
||||
*
|
||||
* 与 DeepSeek 的差异：
|
||||
* - Thinking 模式使用 chat_template_kwargs（非 thinking 字段）
|
||||
* - 默认 max_tokens 更大（65536 vs 8192）
|
||||
* v0.6.4 P3-1: 继承 OpenAICompatibleAdapter —— 传输/组装/回退链收敛到共享基类，
|
||||
* 本文件只保留 Agnes 差异点：chat_template_kwargs 思考开关（v0.6.4 对称性修复）、
|
||||
* 无条件 includeImages、非流式默认超时 300s。
|
||||
* 注：Agnes API 未提供 /models 端点，listModels 使用基类默认实现。
|
||||
*
|
||||
* @see apis/agnes-ai-api-docs-20260625.html
|
||||
*/
|
||||
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
import { MetonaFinishReason } from '../types';
|
||||
import log from 'electron-log';
|
||||
import type { MetonaRequest } from '../types';
|
||||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||||
import log from 'electron-log';
|
||||
import { OpenAICompatibleAdapter } from './shared/openai-compatible-base';
|
||||
|
||||
export class AgnesAdapter extends BaseAdapter {
|
||||
// H-2 修复: provider → providerId（规范要求）
|
||||
export class AgnesAdapter extends OpenAICompatibleAdapter {
|
||||
override readonly providerId: string = 'agnes';
|
||||
readonly supportedModels = ['agnes-2.0-flash'];
|
||||
readonly supportsToolCalling = true;
|
||||
@@ -41,114 +36,27 @@ export class AgnesAdapter extends BaseAdapter {
|
||||
},
|
||||
};
|
||||
|
||||
// ===== POST /chat/completions (非流式) =====
|
||||
// ===== 共享基类差异声明 =====
|
||||
|
||||
// H-2 修复: chat → send（规范要求）
|
||||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||||
const body = this.toNativeRequest(request, false);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
|
||||
if (!response.ok) {
|
||||
await this.throwHttpError(response, 'Agnes AI API error');
|
||||
}
|
||||
|
||||
const data = (await response.json()) as Record<string, unknown>;
|
||||
const parsed = parseOpenAICompatibleResponse(data);
|
||||
|
||||
return {
|
||||
meta: {
|
||||
requestId: request.meta.requestId,
|
||||
provider: this.providerId,
|
||||
model: (data.model as string) ?? this.config.defaultModel,
|
||||
latencyMs: 0,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
content: parsed.content,
|
||||
reasoningContent: parsed.reasoningContent,
|
||||
toolCalls: parsed.toolCalls,
|
||||
usage: parsed.usage,
|
||||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||||
};
|
||||
protected override chatCompletionsUrl(): string {
|
||||
return `${this.config.baseURL}/chat/completions`;
|
||||
}
|
||||
|
||||
// ===== POST /chat/completions (流式) =====
|
||||
|
||||
// H-2 修复: chatStream → sendStream（规范要求）
|
||||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||||
const body = this.toNativeRequest(request, true);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
|
||||
if (!response.ok || !response.body) {
|
||||
await this.throwHttpError(response, 'Agnes AI stream error');
|
||||
}
|
||||
|
||||
yield* parseSSEStream(
|
||||
// 非空断言：上方 if 已确保 response.body 不为 null
|
||||
response.body!,
|
||||
request.meta.requestId,
|
||||
request.meta.sessionId,
|
||||
request.meta.iteration,
|
||||
);
|
||||
protected override sendTimeoutMs(): number {
|
||||
return 300_000;
|
||||
}
|
||||
|
||||
/**
|
||||
* H-2 修复: 获取上下文窗口大小（规范要求）
|
||||
*
|
||||
* v0.3.1: 优先使用配置注入的 contextWindow，回退到 MODEL_INFO 默认值。
|
||||
* Agnes OpenAI 兼容 API 不支持 context_window 参数，此值仅用于
|
||||
* Engine 压缩判断和前端 UI 显示。
|
||||
* 注意：Agnes API 未提供 /models 端点，listModels 使用基类默认实现。
|
||||
*/
|
||||
override getContextWindow(): number {
|
||||
// v0.3.1: 优先使用配置注入的 contextWindow
|
||||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||||
return this.config.contextWindow;
|
||||
}
|
||||
// 回退到 MODEL_INFO
|
||||
const modelInfo = AgnesAdapter.MODEL_INFO[this.config.defaultModel];
|
||||
return modelInfo?.contextWindow ?? 1_000_000;
|
||||
protected override modelInfoTable(): Record<string, MetonaModelInfo> {
|
||||
return AgnesAdapter.MODEL_INFO;
|
||||
}
|
||||
|
||||
// ========== 私有方法 ==========
|
||||
protected override providerLabel(): string {
|
||||
return 'Agnes AI';
|
||||
}
|
||||
|
||||
/**
|
||||
* 构建 Agnes AI 原生请求体
|
||||
*
|
||||
* Agnes AI 特有参数：
|
||||
* - 多模态图片：user 消息的 images[] → OpenAI content 数组 [{type:"text"}, {type:"image_url"}]
|
||||
* 支持 HTTPS URL 或 base64 Data URI（与 MiMo 一致）
|
||||
* - chat_template_kwargs: { enable_thinking: true } — 启用思考模式（非 thinking 字段）
|
||||
* - 默认 max_tokens: 65536（1M 上下文，65.5K 最大输出）
|
||||
*/
|
||||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
// ========== 协议参数映射（Agnes 差异点） ==========
|
||||
|
||||
protected override toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
// v0.6.2: images 处理收敛至共享层（原索引对齐循环在孤立 tool 过滤后会错位）
|
||||
const messages = buildOpenAICompatibleMessages(request, true);
|
||||
const tools = buildOpenAICompatibleTools(request.tools);
|
||||
@@ -182,12 +90,18 @@ export class AgnesAdapter extends BaseAdapter {
|
||||
body.tools = tools;
|
||||
}
|
||||
|
||||
// C-3 修复: Thinking 模式 — Agnes 使用 chat_template_kwargs 而非 thinking
|
||||
// C-3 修复 + v0.6.4 对称性修复: Thinking 模式 — Agnes 使用 chat_template_kwargs
|
||||
// Agnes API 仅支持 enable_thinking: true/false，不支持 effort 级别
|
||||
// thinkingEffort === 'low' 时映射为 false（不启用深度思考），其他级别映射为 true
|
||||
if (request.params.thinkingEnabled) {
|
||||
// thinkingEffort === 'low' 时映射为 false；thinkingEnabled 为 false 或未配置时
|
||||
// 显式发送 enable_thinking:false —— 原实现只在 thinkingEnabled===true 时写该字段，
|
||||
// 若服务端默认开启思考，客户端没有任何路径把它关掉（DeepSeek/MiMo 均显式发送
|
||||
// disabled 保持对称，唯独此处漏了）。
|
||||
{
|
||||
const effort = request.params.thinkingEffort ?? 'high';
|
||||
body.chat_template_kwargs = { enable_thinking: effort !== 'low' };
|
||||
// 未配置 thinkingEnabled 一律显式关闭 —— 与 DeepSeek/MiMo 的"服务端默认开启，
|
||||
// 必须显式发送 disabled"口径对齐，让行为确定性不依赖服务端隐式默认。
|
||||
const wantThinking = request.params.thinkingEnabled === true && effort !== 'low';
|
||||
body.chat_template_kwargs = { enable_thinking: wantThinking };
|
||||
}
|
||||
|
||||
// 停止序列
|
||||
|
||||
@@ -16,7 +16,8 @@
|
||||
* @see https://docs.anthropic.com/en/api/messages
|
||||
*/
|
||||
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import { BaseAdapter, ContentFilterError } from './base-adapter';
|
||||
import { truncatedArgumentsPayload } from './shared/sse-stream';
|
||||
import log from 'electron-log';
|
||||
import { nanoid } from 'nanoid';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
@@ -119,6 +120,12 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
// 工具调用缓冲：content block index → { id, name, argsBuffer }
|
||||
const toolBlocks = new Map<number, { id: string; name: string; argsBuffer: string }>();
|
||||
|
||||
// v0.6.4 竞态修复: message_start 捕获的 input_tokens 改为本次调用的局部闭包变量。
|
||||
// 原实现放在实例字段（this.lastInputTokens）—— fallback adapter 是跨引擎共享的
|
||||
// 单例（agent-engine-manager 把同一实例注入所有引擎），故障转移后多个并发会话
|
||||
// 共用该 Anthropic 实例时 input_tokens 会互相串号。局部化后天然隔离。
|
||||
let messageStartInputTokens = 0;
|
||||
|
||||
const base = () => ({
|
||||
requestId: request.meta.requestId,
|
||||
sessionId: request.meta.sessionId,
|
||||
@@ -127,6 +134,37 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
timestamp: Date.now(),
|
||||
});
|
||||
|
||||
/**
|
||||
* v0.6.4 错误事件单轨化: Anthropic `error` SSE 事件不再以普通 ERROR 流事件转发
|
||||
* （引擎对 ERROR 事件的旧处理是 throw 普通 Error，最终落入 UNKNOWN 且完全绕过
|
||||
* chatStreamWithRetry 的重试/故障转移）。改为抛出携带归一化 status 的异常，
|
||||
* 与 HTTP 层 throwHttpError 同轨：overloaded/rate_limit 走重试、authentication/
|
||||
* invalid_request 不重试并可触发 fallback。
|
||||
*/
|
||||
const anthropicErrorCodeToStatus = (code: string): number => {
|
||||
switch (code) {
|
||||
case 'overloaded_error':
|
||||
return 529;
|
||||
case 'rate_limit_error':
|
||||
return 429;
|
||||
case 'api_error':
|
||||
return 500;
|
||||
case 'timeout_error':
|
||||
return 504;
|
||||
case 'authentication_error':
|
||||
return 401;
|
||||
case 'permission_error':
|
||||
return 403;
|
||||
case 'not_found_error':
|
||||
return 404;
|
||||
case 'request_too_large':
|
||||
case 'invalid_request_error':
|
||||
return 400;
|
||||
default:
|
||||
return 500;
|
||||
}
|
||||
};
|
||||
|
||||
const processEvent = (name: string, data: Record<string, unknown>): MetonaStreamEvent[] => {
|
||||
const events: MetonaStreamEvent[] = [];
|
||||
switch (name) {
|
||||
@@ -173,8 +211,17 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
let args: Record<string, unknown> = {};
|
||||
try {
|
||||
args = block.argsBuffer ? JSON.parse(block.argsBuffer) : {};
|
||||
} catch {
|
||||
args = {};
|
||||
} catch (err) {
|
||||
// v0.6.4 缺口 A 修复: content_block_stop 时 argsBuffer 解析失败（流截断致
|
||||
// JSON 半截）—— 原实现静默降级 args={}，与 v0.6.3 已修复的 OpenAI 共享层
|
||||
// 行为完全相同：工具以"缺少必要参数"泛化失败，模型无从得知发生了截断，
|
||||
// 长文件写入场景直接导致"空回复 → 会话无声终止"。现统一转为
|
||||
// _truncatedArguments 错误参数触发模型自愈（与共享层同源、同文案契约）。
|
||||
const sample = block.argsBuffer.slice(-120);
|
||||
log.warn(
|
||||
`[Anthropic] Tool call args truncated at content_block_stop (unparseable JSON, ${(err as Error).message}). Tail: ...${sample}`,
|
||||
);
|
||||
args = truncatedArgumentsPayload((err as Error).message, sample);
|
||||
}
|
||||
events.push({
|
||||
type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
|
||||
@@ -199,10 +246,14 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
type: MetonaStreamEventType.USAGE,
|
||||
...base(),
|
||||
usage: {
|
||||
inputTokens: (this.lastInputTokens as number) ?? 0,
|
||||
inputTokens: messageStartInputTokens,
|
||||
outputTokens: (usage.output_tokens as number) ?? 0,
|
||||
totalTokens:
|
||||
((this.lastInputTokens as number) ?? 0) + ((usage.output_tokens as number) ?? 0),
|
||||
messageStartInputTokens + ((usage.output_tokens as number) ?? 0),
|
||||
// v0.6.4: 补采 Anthropic 自己的缓存字段（其他 provider 均已采集，
|
||||
// cache_read/creation_input_tokens 与 output_tokens 同在 usage 内）
|
||||
cacheHitTokens: (usage.cache_read_input_tokens as number) ?? undefined,
|
||||
cacheMissTokens: (usage.cache_creation_input_tokens as number) ?? undefined,
|
||||
},
|
||||
});
|
||||
}
|
||||
@@ -215,16 +266,18 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
}
|
||||
case 'error': {
|
||||
const err = data.error as Record<string, unknown> | undefined;
|
||||
events.push({
|
||||
type: MetonaStreamEventType.ERROR,
|
||||
...base(),
|
||||
error: {
|
||||
code: 'unknown' as never,
|
||||
message: (err?.message as string) ?? 'Anthropic stream error',
|
||||
retryable: false,
|
||||
},
|
||||
});
|
||||
break;
|
||||
const code = (err?.type as string) ?? 'api_error';
|
||||
const message = (err?.message as string) ?? 'Anthropic stream error';
|
||||
const status = anthropicErrorCodeToStatus(code);
|
||||
log.warn(
|
||||
`[Anthropic] Upstream error event: ${code} (normalized status=${status}) — throwing for retry/failover handling`,
|
||||
);
|
||||
if (code === 'content_filter_error') {
|
||||
throw new ContentFilterError(message, 'Anthropic SSE error event');
|
||||
}
|
||||
const throwable = new Error(`anthropic_stream_error (${code}): ${message}`);
|
||||
(throwable as Error & { status: number }).status = status;
|
||||
throw throwable;
|
||||
}
|
||||
}
|
||||
return events;
|
||||
@@ -256,13 +309,15 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
if (eventName === 'message_start') {
|
||||
const msg = data.message as Record<string, unknown> | undefined;
|
||||
const usage = msg?.usage as Record<string, unknown> | undefined;
|
||||
this.lastInputTokens = (usage?.input_tokens as number) ?? 0;
|
||||
messageStartInputTokens = (usage?.input_tokens as number) ?? 0;
|
||||
continue;
|
||||
}
|
||||
for (const ev of processEvent(eventName, data)) {
|
||||
yield ev;
|
||||
}
|
||||
} catch (parseErr) {
|
||||
// ContentFilterError / 带 status 的上游错误由 processEvent 抛出，需原样透传
|
||||
if (parseErr instanceof Error && parseErr.name !== 'SyntaxError') throw parseErr;
|
||||
log.warn(
|
||||
`[Anthropic] Failed to parse SSE line: ${(parseErr as Error).message}`,
|
||||
trimmed.slice(0, 200),
|
||||
@@ -271,15 +326,48 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
}
|
||||
}
|
||||
|
||||
// 流中断（连接断开等）补发 DONE，防止 Agent Loop 挂起（与 Ollama 行为一致）
|
||||
// v0.6.4 缺口 B 修复: 流中断时不再让缓冲中的 tool_use 整体蒸发。
|
||||
// 原实现在 content_block_start 与 content_block_stop 之间断连时，toolBlocks 里
|
||||
// 未完成的 block 既不产生 TOOL_CALL_COMPLETE、也不 flush —— 引擎看到"零工具调用
|
||||
// + 零文本"→ 误判 COMPLETED 空回复 → 会话无声停止（正是 v0.6.3 宣称根治、但在
|
||||
// Anthropic 流上仍然存活的场景）。现于补发 DONE 之前，将所有未完成块按截断契约
|
||||
// 转为 _truncatedArguments 自愈 tool call（解析成功的则正常产出）。
|
||||
if (!streamEndedNormally) {
|
||||
const unfinished = [...toolBlocks.entries()];
|
||||
if (unfinished.length > 0) {
|
||||
log.warn(
|
||||
`[Anthropic] Stream ended without message_stop with ${unfinished.length} unfinished tool block(s) — flushing as truncated/self-healing tool calls`,
|
||||
);
|
||||
for (const [, block] of unfinished) {
|
||||
let args: Record<string, unknown> = {};
|
||||
try {
|
||||
args = block.argsBuffer ? JSON.parse(block.argsBuffer) : {};
|
||||
} catch (err) {
|
||||
args = truncatedArgumentsPayload(
|
||||
(err as Error).message,
|
||||
block.argsBuffer.slice(-120),
|
||||
);
|
||||
}
|
||||
yield {
|
||||
type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
|
||||
...base(),
|
||||
toolCall: {
|
||||
id: block.id,
|
||||
name: block.name,
|
||||
args,
|
||||
iteration: request.meta.iteration,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
};
|
||||
}
|
||||
} else {
|
||||
log.warn('[Anthropic] Stream ended without message_stop (connection likely dropped)');
|
||||
}
|
||||
toolBlocks.clear();
|
||||
yield { type: MetonaStreamEventType.DONE, ...base() };
|
||||
}
|
||||
}
|
||||
|
||||
/** message_start 捕获的 input_tokens（供 message_delta 汇总 usage） */
|
||||
private lastInputTokens = 0;
|
||||
|
||||
// ===== 模型与上下文窗口 =====
|
||||
|
||||
override async listModels(): Promise<MetonaModelInfo[]> {
|
||||
@@ -311,6 +399,8 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
request: MetonaRequest,
|
||||
stream: boolean,
|
||||
): Promise<Record<string, unknown>> {
|
||||
const thinkingRequested = Boolean(request.params.thinkingEnabled);
|
||||
|
||||
// System Prompt 拼接（Anthropic 使用顶层 system 字段）
|
||||
const system = [
|
||||
request.systemPrompt.roleDefinition,
|
||||
@@ -405,9 +495,17 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
const anthropicMaxOutput =
|
||||
AnthropicAdapter.MODEL_INFO[this.config.defaultModel]?.maxOutputTokens ?? 64_000;
|
||||
|
||||
// v0.6.4 边界加固: thinking 开启时保证 max_tokens ≥ 2048 —— 协议要求
|
||||
// budget_tokens >= 1024 且 < max_tokens。原实现当用户配置极小 maxTokens
|
||||
// （如 1500）时 Math.floor(1500/2)=750 < 1024 直接 API 400。
|
||||
const requestedMaxTokens = request.params.maxTokens ?? 8192;
|
||||
const maxTokensForRequest = thinkingRequested
|
||||
? Math.max(2048, Math.min(requestedMaxTokens, anthropicMaxOutput))
|
||||
: Math.min(requestedMaxTokens, anthropicMaxOutput);
|
||||
|
||||
const body: Record<string, unknown> = {
|
||||
model: this.config.defaultModel,
|
||||
max_tokens: Math.min(request.params.maxTokens ?? 8192, anthropicMaxOutput),
|
||||
max_tokens: maxTokensForRequest,
|
||||
system,
|
||||
messages: merged,
|
||||
stream,
|
||||
@@ -423,17 +521,15 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
}
|
||||
|
||||
// Thinking 模式：budget_tokens（必须小于 max_tokens，此处钳制到一半）
|
||||
if (request.params.thinkingEnabled) {
|
||||
if (thinkingRequested) {
|
||||
const budgetMap: Record<string, number> = {
|
||||
low: 1024,
|
||||
medium: 4096,
|
||||
high: 16384,
|
||||
max: 32768,
|
||||
};
|
||||
const budget = Math.min(
|
||||
budgetMap[request.params.thinkingEffort ?? 'high'] ?? 16384,
|
||||
Math.floor((body.max_tokens as number) / 2),
|
||||
);
|
||||
const effortBudget = budgetMap[request.params.thinkingEffort ?? 'high'] ?? 16384;
|
||||
const budget = Math.min(effortBudget, Math.floor(maxTokensForRequest / 2));
|
||||
body.thinking = { type: 'enabled', budget_tokens: budget };
|
||||
} else {
|
||||
body.temperature = request.params.temperature;
|
||||
@@ -485,9 +581,13 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
|
||||
for (const block of contentBlocks) {
|
||||
if (block.type === 'text') text += (block.text as string) ?? '';
|
||||
else if (block.type === 'thinking')
|
||||
reasoningContent = (block.thinking as string) ?? undefined;
|
||||
else if (block.type === 'tool_use') {
|
||||
else if (block.type === 'thinking') {
|
||||
// v0.6.4 修复: 多个 thinking 块应为累加（原实现后者覆盖前者，长推理链丢内容）
|
||||
const thinking = (block.thinking as string) ?? '';
|
||||
if (thinking) {
|
||||
reasoningContent = reasoningContent ? `${reasoningContent}\n\n${thinking}` : thinking;
|
||||
}
|
||||
} else if (block.type === 'tool_use') {
|
||||
let args: Record<string, unknown> = {};
|
||||
const rawInput = block.input;
|
||||
if (rawInput && typeof rawInput === 'object') args = rawInput as Record<string, unknown>;
|
||||
@@ -503,12 +603,16 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
|
||||
const usage = (data.usage as Record<string, number>) ?? {};
|
||||
const stopReason = (data.stop_reason as string) ?? 'end_turn';
|
||||
// v0.6.4: refusal / content_filter 不再折叠为 STOP —— 语义丢失会让上层把
|
||||
// "被拒绝的回答"当正常回复展示；统一映射为 CONTENT_FILTERED 走友好提示链路
|
||||
const finishReason: MetonaFinishReason =
|
||||
stopReason === 'tool_use'
|
||||
? MetonaFinishReason.TOOL_CALLS
|
||||
: stopReason === 'max_tokens'
|
||||
? MetonaFinishReason.LENGTH
|
||||
: MetonaFinishReason.STOP;
|
||||
: stopReason === 'refusal' || stopReason === 'content_filter'
|
||||
? MetonaFinishReason.CONTENT_FILTER
|
||||
: MetonaFinishReason.STOP;
|
||||
|
||||
return {
|
||||
meta: {
|
||||
@@ -525,6 +629,9 @@ export class AnthropicAdapter extends BaseAdapter {
|
||||
inputTokens: usage.input_tokens ?? 0,
|
||||
outputTokens: usage.output_tokens ?? 0,
|
||||
totalTokens: (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0),
|
||||
// v0.6.4: 补采缓存字段（与非流式调用方对齐其他 provider 的口径）
|
||||
cacheHitTokens: usage.cache_read_input_tokens,
|
||||
cacheMissTokens: usage.cache_creation_input_tokens,
|
||||
},
|
||||
finishReason,
|
||||
};
|
||||
|
||||
@@ -2,9 +2,19 @@
|
||||
* Provider Adapter — 基类
|
||||
*
|
||||
* 所有 Provider 适配器共享的基类逻辑：
|
||||
* - 请求超时处理
|
||||
* - 错误映射到 MetonaError
|
||||
* - 流式事件标准化
|
||||
* - 请求超时处理（含显式的网络超时错误分类）
|
||||
* - 内容审核错误类型
|
||||
*
|
||||
* v0.6.4 P3-2 错误分类单轨化：
|
||||
* 此前本类存在两套互相漂移的错误分类器 —— `mapError`（protected，生产路径
|
||||
* 无任何调用方，仅测试引用）与 engine.isRetryableError（真正生效）。运行时
|
||||
* 行为由后者单独决定，导致 v0.4.1/#23 的映射修复只体现在测试里。
|
||||
* 现已删除 mapError 与废弃的 getFetchSignal：错误分类的唯一事实来源是
|
||||
* engine.isRetryableError（读取 error.status / error.code / message），
|
||||
* 本层负责保证抛出的错误携带可判定的结构化字段：
|
||||
* - HTTP 非 2xx → throwHttpError 挂 status
|
||||
* - 本方法超时 → code='ETIMEDOUT' + 'timed out' message
|
||||
* - content_filter → ContentFilterError 实例
|
||||
*/
|
||||
|
||||
import type {
|
||||
@@ -13,9 +23,7 @@ import type {
|
||||
MetonaRequest,
|
||||
MetonaResponse,
|
||||
MetonaStreamEvent,
|
||||
MetonaError,
|
||||
} from '../types';
|
||||
import { MetonaErrorCode } from '../types';
|
||||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||||
|
||||
/**
|
||||
@@ -70,7 +78,7 @@ export abstract class BaseAdapter implements IMetonaProviderAdapter {
|
||||
// 优先使用 AdapterConfig.contextWindow（如果存在）
|
||||
const ctx = (this.config as AdapterConfig & { contextWindow?: number }).contextWindow;
|
||||
if (typeof ctx === 'number' && ctx > 0) return ctx;
|
||||
// 默认 1M（保守值，子类应覆盖）
|
||||
// 默认值仅是兜底 —— 子类应声明真实的模型级窗口，避免压缩阈值计算失真
|
||||
return 1_000_000;
|
||||
}
|
||||
|
||||
@@ -82,43 +90,18 @@ export abstract class BaseAdapter implements IMetonaProviderAdapter {
|
||||
this.externalAbortSignal = signal;
|
||||
}
|
||||
|
||||
/**
|
||||
* C-2 修复: 合并外部 abort signal 和 timeout signal
|
||||
*
|
||||
* 使用 AbortSignal.any() 合并两个信号，任一触发都会中断 fetch：
|
||||
* - timeout signal：防止请求挂起
|
||||
* - external abort signal：用户主动中断
|
||||
*
|
||||
* @param timeoutMs 超时时间（毫秒）
|
||||
* @returns 合并后的 AbortSignal
|
||||
* @deprecated 审查修复 M20: 使用 fetchWithTimeout 替代。
|
||||
* getFetchSignal 内部 AbortSignal.timeout() 创建的 timer 在请求成功完成后仍会存活到超时，
|
||||
* 高频调用下 timer 句柄累积；fetchWithTimeout 用 setTimeout + clearTimeout 已解决此问题。
|
||||
*/
|
||||
protected getFetchSignal(timeoutMs: number): AbortSignal {
|
||||
const timeoutSignal = AbortSignal.timeout(timeoutMs);
|
||||
|
||||
// 如果没有外部信号，直接使用 timeout signal
|
||||
if (!this.externalAbortSignal) {
|
||||
return timeoutSignal;
|
||||
}
|
||||
|
||||
// 如果外部信号已经 abort，直接返回它
|
||||
if (this.externalAbortSignal.aborted) {
|
||||
return this.externalAbortSignal;
|
||||
}
|
||||
|
||||
// 合并两个信号 — 任一触发都会 abort
|
||||
// Node.js 20+ / Electron 35+ 支持 AbortSignal.any()
|
||||
return AbortSignal.any([timeoutSignal, this.externalAbortSignal]);
|
||||
}
|
||||
|
||||
/**
|
||||
* #24 修复: 封装 fetch + 超时控制，在 finally 中 clearTimeout，避免 timer 泄漏
|
||||
*
|
||||
* getFetchSignal 使用 AbortSignal.timeout() 内部创建的 timer 在请求成功完成后
|
||||
* 仍会存活到超时，高频调用下 timer 句柄累积。本方法使用 setTimeout + clearTimeout
|
||||
* 确保 fetch 完成（无论成功/失败/abort）后立即清理 timer。
|
||||
* v0.6.4 P3-2（错误分类单轨化）: 本方法自身触发的超时不再以裸 DOMException
|
||||
* （消息不含 timeout 字样、被引擎误归 UNKNOWN 后仅因含 "aborted" 碰巧可重试）
|
||||
* 冒泡 —— 显式转译为带 ETIMEDOUT code 的 Error，使其进入 engine.isRetryableError
|
||||
* 的网络超时判定分支，与其他网络错误同轨。用户主动中断（外部信号）则原样抛出
|
||||
* AbortError —— 引擎 chatStreamWithRetry 入口由 this.aborted 拦截，不会误触发重试。
|
||||
*
|
||||
* 已知边界（设计取舍，注明而非隐藏）：响应头返回后 clearTimeout，后续 SSE 流体
|
||||
* 不再受本超时约束；长挂流由引擎 totalTimeoutMs 兜底。中止时通过 removeEventListener
|
||||
* 解除外部信号监听 —— 流式消费阶段外部 abort 不再打断底层连接（消费方停止拉取即终结）。
|
||||
*
|
||||
* @param url 请求 URL
|
||||
* @param init fetch init（不含 signal，由本方法内部管理）
|
||||
@@ -130,7 +113,11 @@ export abstract class BaseAdapter implements IMetonaProviderAdapter {
|
||||
timeoutMs: number,
|
||||
): Promise<Response> {
|
||||
const controller = new AbortController();
|
||||
const timer = setTimeout(() => controller.abort(), timeoutMs);
|
||||
let timedOut = false;
|
||||
const timer = setTimeout(() => {
|
||||
timedOut = true;
|
||||
controller.abort();
|
||||
}, timeoutMs);
|
||||
|
||||
// 审查修复 M20: 保存 listener 引用，finally 中 removeEventListener 清理，避免 listener 泄漏
|
||||
const onExternalAbort = () => controller.abort();
|
||||
@@ -146,11 +133,24 @@ export abstract class BaseAdapter implements IMetonaProviderAdapter {
|
||||
}
|
||||
|
||||
return await fetch(url, { ...init, signal: controller.signal });
|
||||
} catch (err) {
|
||||
// 区分中止来源：
|
||||
// a) 本方法超时且非外部中断 → 归类为可重试的网络超时（ETIMEDOUT）
|
||||
// b) 外部信号 abort（用户中断）→ 原样抛 AbortError
|
||||
// c) 底层网络错误 → 原样抛出
|
||||
const externalAborted = this.externalAbortSignal?.aborted === true;
|
||||
if (timedOut && !externalAborted) {
|
||||
const timeoutError = new Error(
|
||||
`Request timed out after ${timeoutMs}ms (url=${String(url).slice(0, 120)})`,
|
||||
);
|
||||
(timeoutError as Error & { code: string }).code = 'ETIMEDOUT';
|
||||
throw timeoutError;
|
||||
}
|
||||
throw err;
|
||||
} finally {
|
||||
// #24 修复: 关键 — 无论请求成功、失败还是 abort，都清理 timer
|
||||
clearTimeout(timer);
|
||||
// 审查修复 M20: 清理 externalAbortSignal 上注册的 listener
|
||||
// （即使 { once: true }，请求正常完成时 listener 仍挂在 signal 上直到 abort 或 GC，需显式移除）
|
||||
if (this.externalAbortSignal) {
|
||||
this.externalAbortSignal.removeEventListener('abort', onExternalAbort);
|
||||
}
|
||||
@@ -212,84 +212,12 @@ export abstract class BaseAdapter implements IMetonaProviderAdapter {
|
||||
}
|
||||
}
|
||||
|
||||
const error = new Error(
|
||||
`${context}: ${response.status} ${response.statusText}${errorBody ? ` - ${errorBody}` : ''}`,
|
||||
);
|
||||
// v0.6.4: 巨大 HTML 错误页整体拼进消息会造成日志/事件载荷爆炸 —— 截断到合理长度
|
||||
const safeBody =
|
||||
errorBody.length > 500 ? `${errorBody.slice(0, 500)}…[truncated ${errorBody.length} chars]` : errorBody;
|
||||
|
||||
const error = new Error(`${context}: ${response.status} ${response.statusText}${safeBody ? ` - ${safeBody}` : ''}`);
|
||||
(error as Error & { status: number }).status = response.status;
|
||||
throw error;
|
||||
}
|
||||
|
||||
/**
|
||||
* 将原生错误映射为 MetonaError
|
||||
*/
|
||||
protected mapError(error: unknown): MetonaError {
|
||||
if (error instanceof Error) {
|
||||
// v0.3.17: 优先识别 ContentFilterError
|
||||
if (error instanceof ContentFilterError) {
|
||||
return {
|
||||
code: MetonaErrorCode.CONTENT_FILTERED,
|
||||
message: '内容被 Provider 安全审核拦截，请修改图片或文本后重试',
|
||||
provider: this.providerId,
|
||||
retryable: false,
|
||||
};
|
||||
}
|
||||
|
||||
const msg = error.message.toLowerCase();
|
||||
|
||||
// v0.4.1 修复: msg 已 toLowerCase，网络错误码常量必须用小写比较
|
||||
//（原 'ETIMEDOUT'/'ECONNREFUSED' 等大写常量在小写消息上永不匹配，
|
||||
// 导致网络错误全部落入 UNKNOWN，无法触发引擎的重试逻辑）
|
||||
if (msg.includes('timeout') || msg.includes('etimedout')) {
|
||||
return {
|
||||
code: MetonaErrorCode.NETWORK_TIMEOUT,
|
||||
message: error.message,
|
||||
provider: this.providerId,
|
||||
retryable: true,
|
||||
retryAfterMs: 3000,
|
||||
};
|
||||
}
|
||||
|
||||
if (msg.includes('econnrefused') || msg.includes('enotfound') || msg.includes('econnreset')) {
|
||||
return {
|
||||
code: MetonaErrorCode.NETWORK_ERROR,
|
||||
message: error.message,
|
||||
provider: this.providerId,
|
||||
retryable: true,
|
||||
retryAfterMs: 3000,
|
||||
};
|
||||
}
|
||||
|
||||
// #23 修复: 优先基于 HTTP status code 判断 401/429，避免字符串 includes 误匹配 URL 端口等数字
|
||||
// throwHttpError 已将 response.status 挂到 error.status，优先读取此字段
|
||||
const httpStatus = (error as Error & { status?: number }).status;
|
||||
|
||||
// #23 修复: 401 认证失败 — 优先用 status code，'unauthorized' 是单词不会误匹配
|
||||
if (httpStatus === 401 || msg.includes('unauthorized')) {
|
||||
return {
|
||||
code: MetonaErrorCode.AUTH_INVALID,
|
||||
message: 'API key 无效或已过期',
|
||||
provider: this.providerId,
|
||||
retryable: false,
|
||||
};
|
||||
}
|
||||
|
||||
// #23 修复: 429 限流 — 优先用 status code，'rate limit' 是单词不会误匹配
|
||||
if (httpStatus === 429 || msg.includes('rate limit')) {
|
||||
return {
|
||||
code: MetonaErrorCode.RATE_LIMITED,
|
||||
message: '请求过于频繁，请稍后重试',
|
||||
provider: this.providerId,
|
||||
retryable: true,
|
||||
retryAfterMs: 5000,
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
return {
|
||||
code: MetonaErrorCode.UNKNOWN,
|
||||
message: error instanceof Error ? error.message : 'Unknown error',
|
||||
provider: this.providerId,
|
||||
retryable: false,
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
@@ -4,22 +4,21 @@
|
||||
* 基于 OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、流式输出。
|
||||
* 模型: deepseek-v4-flash / deepseek-v4-pro（1M 上下文，384K 最大输出）
|
||||
*
|
||||
* 独立继承 BaseAdapter，通过 shared/openai-format 和 shared/sse-stream 复用
|
||||
* OpenAI 兼容格式构建和 SSE 流式解析逻辑。不与其他 Provider Adapter 耦合。
|
||||
* v0.6.4 P3-1: 继承 OpenAICompatibleAdapter —— send/sendStream/响应组装/
|
||||
* 认证头/上下文窗口回退链全部收敛到共享基类，本文件只保留 DeepSeek 差异点：
|
||||
* vision 模型判定、/models 合并、/user/balance、thinking+reasoning_effort 映射。
|
||||
*
|
||||
* @see apis/deepseek-api-docs-20260518.html
|
||||
*/
|
||||
|
||||
import log from 'electron-log';
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
import { MetonaFinishReason } from '../types';
|
||||
import type { MetonaRequest } from '../types';
|
||||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||||
import { OpenAICompatibleAdapter } from './shared/openai-compatible-base';
|
||||
|
||||
export class DeepSeekAdapter extends BaseAdapter {
|
||||
// H-2 修复: provider → providerId（规范要求）
|
||||
export class DeepSeekAdapter extends OpenAICompatibleAdapter {
|
||||
// H-2 修复: providerId（规范要求）
|
||||
override readonly providerId: string = 'deepseek';
|
||||
readonly supportedModels = [
|
||||
'deepseek-v4-pro',
|
||||
@@ -61,6 +60,24 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
},
|
||||
};
|
||||
|
||||
// ===== 共享基类差异声明 =====
|
||||
|
||||
protected override chatCompletionsUrl(): string {
|
||||
return `${this.config.baseURL}/chat/completions`;
|
||||
}
|
||||
|
||||
protected override sendTimeoutMs(): number {
|
||||
return 120_000;
|
||||
}
|
||||
|
||||
protected override modelInfoTable(): Record<string, MetonaModelInfo> {
|
||||
return DeepSeekAdapter.MODEL_INFO;
|
||||
}
|
||||
|
||||
protected override providerLabel(): string {
|
||||
return 'DeepSeek';
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.5.4: 当前模型是否支持多模态图片输入
|
||||
*
|
||||
@@ -72,94 +89,13 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
return this.config.defaultModel.includes('vision');
|
||||
}
|
||||
|
||||
// ===== POST /chat/completions (非流式) =====
|
||||
|
||||
// H-2 修复: chat → send（规范要求）
|
||||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||||
const body = this.toNativeRequest(request, false);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 120_000,
|
||||
);
|
||||
|
||||
if (!response.ok) {
|
||||
await this.throwHttpError(response, 'DeepSeek API error');
|
||||
}
|
||||
|
||||
const data = (await response.json()) as Record<string, unknown>;
|
||||
const parsed = parseOpenAICompatibleResponse(data);
|
||||
|
||||
return {
|
||||
meta: {
|
||||
requestId: request.meta.requestId,
|
||||
provider: this.providerId,
|
||||
model: (data.model as string) ?? this.config.defaultModel,
|
||||
latencyMs: 0,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
content: parsed.content,
|
||||
reasoningContent: parsed.reasoningContent,
|
||||
toolCalls: parsed.toolCalls,
|
||||
usage: parsed.usage,
|
||||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||||
};
|
||||
}
|
||||
|
||||
// ===== POST /chat/completions (流式) =====
|
||||
|
||||
// H-2 修复: chatStream → sendStream（规范要求）
|
||||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||||
const body = this.toNativeRequest(request, true);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
|
||||
if (!response.ok || !response.body) {
|
||||
await this.throwHttpError(response, 'DeepSeek stream error');
|
||||
}
|
||||
|
||||
yield* parseSSEStream(
|
||||
// 非空断言：上方 if 已确保 response.body 不为 null
|
||||
// TypeScript 无法通过 await Promise<never> 正确收窄，需显式断言
|
||||
response.body!,
|
||||
request.meta.requestId,
|
||||
request.meta.sessionId,
|
||||
request.meta.iteration,
|
||||
);
|
||||
}
|
||||
|
||||
// ===== GET /models =====
|
||||
|
||||
/**
|
||||
* H-2 修复: 返回 MetonaModelInfo[]（规范要求）
|
||||
*
|
||||
* 优先尝试从 API 获取实时模型列表，并合并本地 MODEL_INFO 元数据。
|
||||
* API 不可用时回退到 supportedModels。
|
||||
*/
|
||||
async listModels(): Promise<MetonaModelInfo[]> {
|
||||
override async listModels(): Promise<MetonaModelInfo[]> {
|
||||
try {
|
||||
const response = await fetch(`${this.config.baseURL}/models`, {
|
||||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||||
@@ -179,36 +115,13 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
return this.supportedModels.map((id) => DeepSeekAdapter.MODEL_INFO[id] ?? { id });
|
||||
}
|
||||
|
||||
/**
|
||||
* H-2 修复: 获取上下文窗口大小（规范要求）
|
||||
*
|
||||
* v0.3.1: 优先使用配置注入的 contextWindow，回退到 MODEL_INFO 默认值。
|
||||
* DeepSeek OpenAI 兼容 API 不支持 context_window 参数，此值仅用于
|
||||
* Engine 压缩判断和前端 UI 显示。
|
||||
*/
|
||||
override getContextWindow(): number {
|
||||
// v0.3.1: 优先使用配置注入的 contextWindow
|
||||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||||
return this.config.contextWindow;
|
||||
}
|
||||
// 回退到 MODEL_INFO
|
||||
const modelInfo = DeepSeekAdapter.MODEL_INFO[this.config.defaultModel];
|
||||
return modelInfo?.contextWindow ?? 1_000_000;
|
||||
}
|
||||
|
||||
// ===== GET /user/balance =====
|
||||
|
||||
/**
|
||||
* 查询账户余额
|
||||
*
|
||||
* v0.5.2 修复: DeepSeek 官方 API 实际返回 `balance_infos` 数组格式：
|
||||
* { "is_available": true, "balance_infos": [{ "currency": "CNY",
|
||||
* "total_balance": "110.00", "granted_balance": "10.00", "topped_up_balance": "100.00" }] }
|
||||
* 此前按扁平字段解析（data.total_balance）→ 永远取到 undefined → 恒显示 0。
|
||||
* 现优先取 balance_infos[0]，回退扁平格式（兼容网关/代理的简化响应）。
|
||||
*
|
||||
* URL 规范化: 余额端点为 {root}/user/balance（无 /v1 前缀）。用户配置的
|
||||
* baseURL 可能带 /v1 或尾斜杠（chat 端点两种写法都合法），此处剥离后拼接。
|
||||
* v0.5.2 修复: 官方 API 返回 balance_infos 数组格式（此前按扁平字段解析恒为 0）。
|
||||
* URL 规范化: 余额端点为 {root}/user/balance（无 /v1 前缀），需剥离配置中的尾斜杠与 /v1。
|
||||
*/
|
||||
async getBalance(): Promise<{
|
||||
currency: string;
|
||||
@@ -217,7 +130,6 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
toppedUpBalance: string;
|
||||
} | null> {
|
||||
try {
|
||||
// 规范化 baseURL：去尾斜杠、去尾 /v1（余额端点在根路径下）
|
||||
const root = this.config.baseURL.replace(/\/+$/, '').replace(/\/v1$/, '');
|
||||
const response = await fetch(`${root}/user/balance`, {
|
||||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||||
@@ -238,7 +150,6 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
granted_balance?: string;
|
||||
topped_up_balance?: string;
|
||||
};
|
||||
// 优先官方 balance_infos 数组，回退扁平格式
|
||||
const info = data.balance_infos?.[0] ?? data;
|
||||
return {
|
||||
currency: info.currency ?? 'CNY',
|
||||
@@ -251,17 +162,9 @@ export class DeepSeekAdapter extends BaseAdapter {
|
||||
}
|
||||
}
|
||||
|
||||
// ========== 私有方法 ==========
|
||||
// ========== 协议参数映射（DeepSeek 差异点） ==========
|
||||
|
||||
/**
|
||||
* 构建 DeepSeek 原生请求体
|
||||
*
|
||||
* DeepSeek 特有参数：
|
||||
* - thinking: { type: "enabled" } — 启用思考模式
|
||||
* - reasoning_effort — 思考强度映射
|
||||
* - stream_options: { include_usage: true } — 流式返回 usage
|
||||
*/
|
||||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
protected override toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
// v0.6.2: images 处理收敛至共享层（includeImages = vision 模型才转换，
|
||||
// 非 vision 静默丢弃——正确行为，见 openai-format.ts #27 记录）
|
||||
const messages = buildOpenAICompatibleMessages(request, this.isVisionModel());
|
||||
|
||||
@@ -4,34 +4,26 @@
|
||||
* 基于 OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、流式输出。
|
||||
* 模型: mimo-v2.5-pro（1M 上下文 / 131072 max_tokens）/ mimo-v2.5（1M 上下文 / 32768 max_tokens）
|
||||
*
|
||||
* 独立继承 BaseAdapter，通过 shared/openai-format 和 shared/sse-stream 复用
|
||||
* OpenAI 兼容格式构建和 SSE 流式解析逻辑。不与其他 Provider Adapter 耦合。
|
||||
*
|
||||
* 与 DeepSeek 适配器的关键差异：
|
||||
* - 使用 max_completion_tokens（非 max_tokens）
|
||||
* - thinking 参数结构与 DeepSeek 一致（thinking.type: "enabled"/"disabled"）
|
||||
* - 不提供 /models 端点（listModels 回退到本地元数据）
|
||||
* - 不提供 /user/balance 端点
|
||||
* - tool_choice 仅支持 "auto"
|
||||
* v0.6.4 P3-1: 继承 OpenAICompatibleAdapter —— 传输/组装/回退链收敛到共享基类，
|
||||
* 本文件只保留 MiMo 差异点：max_completion_tokens 字段名、tool_choice 强制 "auto"、
|
||||
* 思考模式与 temperature/top_p 互斥、无 /models 端点（本地元数据列表）。
|
||||
*
|
||||
* @see apis/mimo-api-docs-20260715.html
|
||||
*/
|
||||
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
import { MetonaFinishReason } from '../types';
|
||||
import type { MetonaRequest } from '../types';
|
||||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||||
import { OpenAICompatibleAdapter } from './shared/openai-compatible-base';
|
||||
|
||||
export class MimoAdapter extends BaseAdapter {
|
||||
export class MimoAdapter extends OpenAICompatibleAdapter {
|
||||
override readonly providerId: string = 'mimo';
|
||||
readonly supportedModels = ['mimo-v2.5-pro', 'mimo-v2.5'];
|
||||
readonly supportsToolCalling = true;
|
||||
readonly supportsThinking = true;
|
||||
|
||||
// MiMo 模型元信息
|
||||
// mimo-v2.5-pro: 1M 上下文（与 DeepSeek 一致）/ 131072 max_tokens；mimo-v2.5: 1M 上下文 / 32768 max_tokens
|
||||
// mimo-v2.5-pro: 1M 上下文 / 131072 max_tokens；mimo-v2.5: 1M 上下文 / 32768 max_tokens
|
||||
private static readonly MODEL_INFO: Record<string, MetonaModelInfo> = {
|
||||
'mimo-v2.5-pro': {
|
||||
id: 'mimo-v2.5-pro',
|
||||
@@ -53,83 +45,23 @@ export class MimoAdapter extends BaseAdapter {
|
||||
},
|
||||
};
|
||||
|
||||
// ===== POST /chat/completions (非流式) =====
|
||||
// ===== 共享基类差异声明 =====
|
||||
|
||||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||||
const body = this.toNativeRequest(request, false);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 120_000,
|
||||
);
|
||||
|
||||
if (!response.ok) {
|
||||
await this.throwHttpError(response, 'MiMo API error');
|
||||
}
|
||||
|
||||
const data = (await response.json()) as Record<string, unknown>;
|
||||
const parsed = parseOpenAICompatibleResponse(data);
|
||||
|
||||
return {
|
||||
meta: {
|
||||
requestId: request.meta.requestId,
|
||||
provider: this.providerId,
|
||||
model: (data.model as string) ?? this.config.defaultModel,
|
||||
latencyMs: 0,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
content: parsed.content,
|
||||
reasoningContent: parsed.reasoningContent,
|
||||
toolCalls: parsed.toolCalls,
|
||||
usage: parsed.usage,
|
||||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||||
};
|
||||
protected override chatCompletionsUrl(): string {
|
||||
return `${this.config.baseURL}/chat/completions`;
|
||||
}
|
||||
|
||||
// ===== POST /chat/completions (流式) =====
|
||||
|
||||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||||
const body = this.toNativeRequest(request, true);
|
||||
|
||||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch，确保 timer 清理
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
|
||||
if (!response.ok || !response.body) {
|
||||
await this.throwHttpError(response, 'MiMo stream error');
|
||||
}
|
||||
|
||||
yield* parseSSEStream(
|
||||
// 非空断言：上方 if 已确保 response.body 不为 null
|
||||
response.body!,
|
||||
request.meta.requestId,
|
||||
request.meta.sessionId,
|
||||
request.meta.iteration,
|
||||
);
|
||||
protected override sendTimeoutMs(): number {
|
||||
return 120_000;
|
||||
}
|
||||
|
||||
// ===== 模型列表 =====
|
||||
protected override modelInfoTable(): Record<string, MetonaModelInfo> {
|
||||
return MimoAdapter.MODEL_INFO;
|
||||
}
|
||||
|
||||
protected override providerLabel(): string {
|
||||
return 'MiMo';
|
||||
}
|
||||
|
||||
/**
|
||||
* MiMo 官方未提供 /models 端点，直接返回本地元数据。
|
||||
@@ -138,37 +70,9 @@ export class MimoAdapter extends BaseAdapter {
|
||||
return this.supportedModels.map((id) => MimoAdapter.MODEL_INFO[id] ?? { id });
|
||||
}
|
||||
|
||||
/**
|
||||
* 获取上下文窗口大小
|
||||
*
|
||||
* v0.3.1: 优先使用配置注入的 contextWindow，回退到 MODEL_INFO 默认值。
|
||||
* MiMo OpenAI 兼容 API 不支持 context_window 参数，此值仅用于
|
||||
* Engine 压缩判断和前端 UI 显示。
|
||||
*/
|
||||
override getContextWindow(): number {
|
||||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||||
return this.config.contextWindow;
|
||||
}
|
||||
const modelInfo = MimoAdapter.MODEL_INFO[this.config.defaultModel];
|
||||
return modelInfo?.contextWindow ?? 1_000_000;
|
||||
}
|
||||
// ========== 协议参数映射（MiMo 差异点） ==========
|
||||
|
||||
// ========== 私有方法 ==========
|
||||
|
||||
/**
|
||||
* 构建 MiMo 原生请求体
|
||||
*
|
||||
* MiMo 特有参数：
|
||||
* - 多模态图片：user 消息的 images[] → OpenAI content 数组 [{type:"text"}, {type:"image_url"}]
|
||||
* 支持 HTTPS URL 或 base64 Data URI
|
||||
* - thinking: { type: "enabled" / "disabled" } — 与 DeepSeek 一致
|
||||
* - max_completion_tokens — 非 max_tokens（MiMo 使用新字段名）
|
||||
* - stream_options: { include_usage: true } — 流式返回 usage
|
||||
* - tool_choice: "auto" — MiMo 仅支持 auto
|
||||
*
|
||||
* 思考模式下 temperature/top_p 会被 API 强制覆盖，因此不传这两个参数。
|
||||
*/
|
||||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
protected override toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
// v0.6.2: images 处理收敛至共享层（原索引对齐循环在孤立 tool 过滤后会错位）
|
||||
const messages = buildOpenAICompatibleMessages(request, true);
|
||||
const tools = buildOpenAICompatibleTools(request.tools);
|
||||
@@ -198,6 +102,24 @@ export class MimoAdapter extends BaseAdapter {
|
||||
body.tool_choice = 'auto';
|
||||
}
|
||||
|
||||
// v0.6.4 P4-3: MiMo 服务端内置工具透出 —— config.providerOptions.enableWebSearch
|
||||
// 开启后附加 {type:'web_search'} 服务端搜索工具（annotations 引用随响应返回，
|
||||
// 由上层归并为文本内容展示）。与客户端 tools 定义互不影响。
|
||||
const providerOptions = this.config.providerOptions as Record<string, unknown> | undefined;
|
||||
if (providerOptions?.['enableWebSearch'] === true) {
|
||||
const serverTools = body.tools
|
||||
? [...(body.tools as Array<Record<string, unknown>>), { type: 'web_search' }]
|
||||
: [{ type: 'web_search' }];
|
||||
body.tools = serverTools;
|
||||
if (!body.tool_choice) body.tool_choice = 'auto';
|
||||
}
|
||||
|
||||
// v0.6.4 P4-3: strict JSON 响应格式开关（response_format: json_object）——
|
||||
// 供结构化抽取类任务使用；与流式模式兼容性由服务端保证（文档标注支持子集）
|
||||
if (providerOptions?.['responseFormatJson'] === true) {
|
||||
body.response_format = { type: 'json_object' };
|
||||
}
|
||||
|
||||
// Thinking 模式（与 DeepSeek 参数结构一致）
|
||||
// MiMo API 默认 thinking.type = "enabled"，必须显式发送 disabled 才能关闭
|
||||
if (request.params.thinkingEnabled === false) {
|
||||
|
||||
@@ -23,6 +23,7 @@
|
||||
*/
|
||||
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import { truncatedArgumentsPayload } from './shared/sse-stream';
|
||||
import log from 'electron-log';
|
||||
import { nanoid } from 'nanoid';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
@@ -44,6 +45,10 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
constructor(config: ConstructorParameters<typeof BaseAdapter>[0]) {
|
||||
super(config);
|
||||
this.baseURL = config.baseURL || 'http://localhost:11434';
|
||||
// v0.6.4 P4-1: 每个适配器实例（= 每会话独立引擎）启动时做一次 /api/show 探测，
|
||||
// 把 num_ctx 实测值填充进 getContextWindow 缓存。fire-and-forget：失败静默，
|
||||
// 不阻塞/不影响首个请求；此后压缩预算基于实测窗口而非保守默认 4096。
|
||||
this.refreshContextWindow();
|
||||
}
|
||||
|
||||
// ===== POST /api/chat =====
|
||||
@@ -136,7 +141,25 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
if (chunk.message?.tool_calls) {
|
||||
for (const tc of chunk.message.tool_calls) {
|
||||
const args = tc.function?.arguments;
|
||||
const parsedArgs = typeof args === 'string' ? JSON.parse(args) : (args ?? {});
|
||||
// v0.6.4 缺口修复: NDJSON 路径的截断自愈 —— 原实现 JSON.parse 抛错会
|
||||
// 落入外层 catch：该 tool call 整体静默丢弃，且同一行剩余处理
|
||||
// （含 done/USAGE 检查）一并被跳过，与 v0.6.3 已根治的 OpenAI 共享层
|
||||
// 旧行为完全相同。现独立捕获并转为 _truncatedArguments 自愈载荷，
|
||||
// 同时保证本 chunk 的后续分支照常执行。
|
||||
let parsedArgs: Record<string, unknown>;
|
||||
if (typeof args === 'string') {
|
||||
try {
|
||||
parsedArgs = JSON.parse(args);
|
||||
} catch (parseErr) {
|
||||
const sample = args.slice(-120);
|
||||
log.warn(
|
||||
`[Ollama] Tool call args truncated (unparseable JSON, ${(parseErr as Error).message}). Tail: ...${sample}`,
|
||||
);
|
||||
parsedArgs = truncatedArgumentsPayload((parseErr as Error).message, sample);
|
||||
}
|
||||
} else {
|
||||
parsedArgs = (args as Record<string, unknown>) ?? {};
|
||||
}
|
||||
yield {
|
||||
type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
|
||||
requestId: request.meta.requestId,
|
||||
@@ -284,17 +307,25 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
}>;
|
||||
};
|
||||
if (data.models?.length) {
|
||||
return data.models.map((m) => ({
|
||||
id: m.name,
|
||||
name: m.name,
|
||||
// Ollama 模型上下文窗口由 options.num_ctx 决定，此处给保守值
|
||||
contextWindow: OllamaAdapter.DEFAULT_CONTEXT_WINDOW,
|
||||
supportsToolCalling: true, // Ollama 多数模型支持，具体能力需通过 /api/show 查询
|
||||
supportsThinking: true,
|
||||
description: m.details
|
||||
? `${m.details.family ?? 'unknown'} / ${m.details.parameter_size ?? '?'} / ${m.details.quantization_level ?? '?'}`
|
||||
: undefined,
|
||||
}));
|
||||
// v0.6.4 P4-1: 能力标志改为逐模型 /api/show 实测探测；单个探测失败
|
||||
// 该模型回退保守 true（不可用时行为与旧实现一致，fail-open 保可用性）
|
||||
const enriched = await Promise.all(
|
||||
data.models.map(async (m) => {
|
||||
const caps = await this.probeCapabilities(m.name);
|
||||
return {
|
||||
id: m.name,
|
||||
name: m.name,
|
||||
// Ollama 模型上下文窗口由 options.num_ctx 决定，此处给保守值
|
||||
contextWindow: OllamaAdapter.DEFAULT_CONTEXT_WINDOW,
|
||||
supportsToolCalling: caps ? caps.supportsTools : true,
|
||||
supportsThinking: caps ? caps.supportsThinking : true,
|
||||
description: m.details
|
||||
? `${m.details.family ?? 'unknown'} / ${m.details.parameter_size ?? '?'} / ${m.details.quantization_level ?? '?'}`
|
||||
: undefined,
|
||||
};
|
||||
}),
|
||||
);
|
||||
return enriched;
|
||||
}
|
||||
}
|
||||
} catch {
|
||||
@@ -312,7 +343,62 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
* 此处返回默认值，供 Engine 在未指定时参考。
|
||||
*/
|
||||
override getContextWindow(): number {
|
||||
return OllamaAdapter.DEFAULT_CONTEXT_WINDOW;
|
||||
return this.cachedContextWindow ?? OllamaAdapter.DEFAULT_CONTEXT_WINDOW;
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.6.4 P4-1: 从 /api/show 的 parameters 区解析 num_ctx 真值。
|
||||
*
|
||||
* 契约约束：IMetonaProviderAdapter.getContextWindow 是同步接口（引擎压缩判定
|
||||
* 依赖同步取值），无法在内部 await。因此采用"机会主义缓存"策略：
|
||||
* send/sendStream 启动时 fire-and-forget 刷新缓存；首次请求前返回默认 4096，
|
||||
* 之后永远返回实测值。压缩预算的准确性随使用逐渐收敛到真值。
|
||||
*/
|
||||
private cachedContextWindow: number | null = null;
|
||||
private refreshingContextWindow = false;
|
||||
|
||||
private refreshContextWindow(): void {
|
||||
if (this.refreshingContextWindow) return;
|
||||
this.refreshingContextWindow = true;
|
||||
void this.showModel(this.config.defaultModel)
|
||||
.then((info) => {
|
||||
if (!info?.parameters) return;
|
||||
const match = /^num_ctx\s+(\d+)\s*$/m.exec(info.parameters);
|
||||
if (match) {
|
||||
const value = Number(match[1]);
|
||||
if (Number.isFinite(value) && value > 0) {
|
||||
this.cachedContextWindow = value;
|
||||
log.info(`[Ollama] Context window (num_ctx) detected: ${value}`);
|
||||
}
|
||||
}
|
||||
})
|
||||
.catch(() => {
|
||||
/* 模型探测失败不阻塞对话 */
|
||||
})
|
||||
.finally(() => {
|
||||
this.refreshingContextWindow = false;
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* v0.6.4 P4-1: 通过 /api/show 的 capabilities[] 动态探测模型真实能力。
|
||||
* 此前 listModels 对所有本地模型硬编码 supportsToolCalling/supportsThinking:true
|
||||
* （注释自知不准）—— 语言模型不支持 tools 时引擎仍下发工具定义，
|
||||
* 造成"模型口头说调工具实际不调"的回归温床。探测失败返回 null 由调用方回退保守值。
|
||||
*/
|
||||
async probeCapabilities(model: string): Promise<{
|
||||
supportsTools: boolean;
|
||||
supportsVision: boolean;
|
||||
supportsThinking: boolean;
|
||||
} | null> {
|
||||
const info = await this.showModel(model);
|
||||
if (!info || !Array.isArray(info.capabilities)) return null;
|
||||
const caps = new Set(info.capabilities.map((c) => String(c)));
|
||||
return {
|
||||
supportsTools: caps.has('tools'),
|
||||
supportsVision: caps.has('vision'),
|
||||
supportsThinking: caps.has('thinking'),
|
||||
};
|
||||
}
|
||||
|
||||
// ===== POST /api/show =====
|
||||
@@ -339,12 +425,22 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
|
||||
// ===== POST /api/pull =====
|
||||
|
||||
async pullModel(model: string, onProgress?: (progress: { status: string; completed?: number; total?: number }) => void): Promise<void> {
|
||||
/**
|
||||
* v0.6.4 P4-1 重构：pull 支持外部取消信号 —— 原实现固定 600s 超时会掐死
|
||||
* 大模型下载（进度不能续命、无取消通道），大仓/慢网络场景必然失败。
|
||||
* 现契约：调用方通过 AbortSignal 控制生命周期（UI 取消按钮即可触发）；
|
||||
* 超时语义交给用户取消或服务端断流（读循环结束即完成），不再人为设上限。
|
||||
*/
|
||||
async pullModel(
|
||||
model: string,
|
||||
onProgress?: (progress: { status: string; completed?: number; total?: number }) => void,
|
||||
signal?: AbortSignal,
|
||||
): Promise<void> {
|
||||
const response = await fetch(`${this.baseURL}/api/pull`, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({ model, stream: true }),
|
||||
signal: AbortSignal.timeout(600_000), // 模型下载可能较慢，10 分钟超时
|
||||
signal,
|
||||
});
|
||||
|
||||
if (!response.ok || !response.body) throw new Error(`Ollama pull error: ${response.status}`);
|
||||
@@ -557,8 +653,15 @@ export class OllamaAdapter extends BaseAdapter {
|
||||
let args: Record<string, unknown> = {};
|
||||
try {
|
||||
args = typeof rawArgs === 'string' ? JSON.parse(rawArgs) : (rawArgs as Record<string, unknown>) ?? {};
|
||||
} catch {
|
||||
args = {};
|
||||
} catch (parseErr) {
|
||||
// v0.6.4: 非流式路径截断自愈对齐 —— 原 catch 静默降级 {}，与流式修复后的
|
||||
// 行为不一致。统一转为 _truncatedArguments 错误参数。
|
||||
const sample =
|
||||
typeof rawArgs === 'string' ? rawArgs.slice(-120) : String(rawArgs).slice(-120);
|
||||
log.warn(
|
||||
`[Ollama] Non-stream tool call args truncated (unparseable JSON, ${(parseErr as Error).message}). Tail: ...${sample}`,
|
||||
);
|
||||
args = truncatedArgumentsPayload((parseErr as Error).message, sample);
|
||||
}
|
||||
return {
|
||||
// L-9 修复（审计补充）: 非流式路径统一使用 nanoid，与流式路径（sendStream）保持一致
|
||||
|
||||
@@ -4,22 +4,23 @@
|
||||
* OpenAI Chat Completions API（/v1/chat/completions），支持 Tool Calling、
|
||||
* 流式输出、多模态图片、o 系列推理模型的 reasoning_effort 参数。
|
||||
*
|
||||
* 与 DeepSeek 适配器的关键差异：
|
||||
* - o 系列 / gpt-5 系列模型使用 max_completion_tokens（非 max_tokens）
|
||||
* - Thinking 模式通过顶层 reasoning_effort 参数（o 系列模型）
|
||||
* - 模型列表从 /v1/models 动态获取
|
||||
* v0.6.4 P3-1: 继承 OpenAICompatibleAdapter —— 传输/组装/回退链收敛到共享基类，
|
||||
* 本文件只保留 OpenAI 差异点：o 系列/gpt-5 的字段名路由与 reasoning_effort、
|
||||
* 推理模型拒图的前置拦截（升级为 ModelCapabilityError）、动态 /models 列表、
|
||||
* 非推理模型 temperature 控制。
|
||||
*
|
||||
* @see apis 官方文档 https://platform.openai.com/docs/api-reference/chat
|
||||
* @see https://platform.openai.com/docs/api-reference/chat
|
||||
*/
|
||||
|
||||
import { BaseAdapter } from './base-adapter';
|
||||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||||
import { MetonaFinishReason } from '../types';
|
||||
import type { MetonaRequest } from '../types';
|
||||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||||
import {
|
||||
ModelCapabilityError,
|
||||
OpenAICompatibleAdapter,
|
||||
} from './shared/openai-compatible-base';
|
||||
|
||||
export class OpenAIAdapter extends BaseAdapter {
|
||||
export class OpenAIAdapter extends OpenAICompatibleAdapter {
|
||||
override readonly providerId: string = 'openai';
|
||||
readonly supportedModels = ['gpt-4o', 'gpt-4o-mini', 'gpt-4.1', 'o3-mini'];
|
||||
readonly supportsToolCalling = true;
|
||||
@@ -64,78 +65,27 @@ export class OpenAIAdapter extends BaseAdapter {
|
||||
},
|
||||
};
|
||||
|
||||
// ===== POST /v1/chat/completions（非流式） =====
|
||||
// ===== 共享基类差异声明 =====
|
||||
|
||||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||||
const body = this.toNativeRequest(request, false);
|
||||
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 120_000,
|
||||
);
|
||||
|
||||
if (!response.ok) {
|
||||
await this.throwHttpError(response, 'OpenAI API error');
|
||||
}
|
||||
|
||||
const data = (await response.json()) as Record<string, unknown>;
|
||||
const parsed = parseOpenAICompatibleResponse(data);
|
||||
|
||||
return {
|
||||
meta: {
|
||||
requestId: request.meta.requestId,
|
||||
provider: this.providerId,
|
||||
model: (data.model as string) ?? this.config.defaultModel,
|
||||
latencyMs: 0,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
content: parsed.content,
|
||||
reasoningContent: parsed.reasoningContent,
|
||||
toolCalls: parsed.toolCalls,
|
||||
usage: parsed.usage,
|
||||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||||
};
|
||||
protected override chatCompletionsUrl(): string {
|
||||
return `${this.config.baseURL}/chat/completions`;
|
||||
}
|
||||
|
||||
// ===== POST /v1/chat/completions（流式） =====
|
||||
protected override sendTimeoutMs(): number {
|
||||
return 120_000;
|
||||
}
|
||||
|
||||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||||
const body = this.toNativeRequest(request, true);
|
||||
protected override modelInfoTable(): Record<string, MetonaModelInfo> {
|
||||
return OpenAIAdapter.MODEL_INFO;
|
||||
}
|
||||
|
||||
const response = await this.fetchWithTimeout(
|
||||
`${this.config.baseURL}/chat/completions`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
},
|
||||
body: JSON.stringify(body),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
protected override providerLabel(): string {
|
||||
return 'OpenAI';
|
||||
}
|
||||
|
||||
if (!response.ok || !response.body) {
|
||||
await this.throwHttpError(response, 'OpenAI stream error');
|
||||
}
|
||||
|
||||
yield* parseSSEStream(
|
||||
// 非空断言：上方 if 已确保 response.body 不为 null
|
||||
response.body!,
|
||||
request.meta.requestId,
|
||||
request.meta.sessionId,
|
||||
request.meta.iteration,
|
||||
);
|
||||
// v0.6.4: OpenAI 家族兜底窗口为 128K（其余 OpenAI 兼容 Provider 为 1M）
|
||||
protected override defaultContextWindowFallback(): number {
|
||||
return 128_000;
|
||||
}
|
||||
|
||||
// ===== GET /v1/models =====
|
||||
@@ -158,34 +108,20 @@ export class OpenAIAdapter extends BaseAdapter {
|
||||
return this.supportedModels.map((id) => OpenAIAdapter.MODEL_INFO[id] ?? { id });
|
||||
}
|
||||
|
||||
override getContextWindow(): number {
|
||||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||||
return this.config.contextWindow;
|
||||
}
|
||||
const modelInfo = OpenAIAdapter.MODEL_INFO[this.config.defaultModel];
|
||||
return modelInfo?.contextWindow ?? 128_000;
|
||||
}
|
||||
// ========== 协议参数映射（OpenAI 差异点） ==========
|
||||
|
||||
// ========== 私有方法 ==========
|
||||
|
||||
/**
|
||||
* 构建 OpenAI 原生请求体
|
||||
*
|
||||
* OpenAI 特有处理：
|
||||
* - 多模态图片：user 消息 images[] → content 数组
|
||||
* - o 系列（o1/o3/o4）与 gpt-5 系列使用 max_completion_tokens + reasoning_effort
|
||||
* - 思考模式下 temperature 被部分推理模型拒绝，不传
|
||||
*/
|
||||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
protected override toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||||
// 推理模型检测（o 系列使用新参数名）
|
||||
const model = this.config.defaultModel;
|
||||
const isReasoningModel = /^(o\d|gpt-5)/.test(model);
|
||||
|
||||
// 推理模型不支持图片输入 — 前置校验（转换在共享层，此处仅拦截）
|
||||
// 推理模型不支持图片输入 — 前置校验
|
||||
// v0.6.4 升级: 原实现抛裸 Error 落入 UNKNOWN 错误码；现在抛 ModelCapabilityError
|
||||
// （携带 status=400），引擎按"不可重试请求级错误"处理，UI 可区分能力限制与一般故障。
|
||||
if (isReasoningModel) {
|
||||
const hasImages = request.messages.some((m) => m.images?.length);
|
||||
if (hasImages) {
|
||||
throw new Error(`Model "${model}" does not support image inputs`);
|
||||
throw new ModelCapabilityError(model, 'image inputs');
|
||||
}
|
||||
}
|
||||
|
||||
@@ -238,6 +174,8 @@ export class OpenAIAdapter extends BaseAdapter {
|
||||
}
|
||||
|
||||
// 停止序列
|
||||
// 已知边界（协议限制，待上游放开后移除此注释）：o 系列不支持 stop 参数，
|
||||
// 当前仍透传 —— 若推理模型 + stop 组合触发 400 属上游约束而非本层缺陷。
|
||||
if (request.params.stopSequences?.length) {
|
||||
body.stop = request.params.stopSequences;
|
||||
}
|
||||
|
||||
@@ -0,0 +1,189 @@
|
||||
/**
|
||||
* OpenAI 兼容 Provider 中间基类（v0.6.4 P3-1）
|
||||
*
|
||||
* 背景：deepseek / agnes-ai / mimo / openai 四家适配器各自复制了几乎逐字相同的
|
||||
* ~60 行传输样板 —— send/sendStream 的 fetchWithTimeout 调用、Bearer 头构建、
|
||||
* HTTP 错误桥接、非流式 JSON → MetonaResponse 的字段组装、SSE 流接入、以及
|
||||
* "config.contextWindow → MODEL_INFO → 兜底" 的上下文窗口回退链。
|
||||
* 任何行为修复都要改四处，是历史缺陷（如超时字段不一致）的直接来源。
|
||||
*
|
||||
* 收敛后职责划分：
|
||||
* - 本基类拥有：send / sendStream / buildHeaders / 响应组装 / finishReason 映射 /
|
||||
* getContextWindow 回退链；
|
||||
* - 子类只声明差异：chatCompletionsUrl、toNativeRequest（协议参数映射）、
|
||||
* sendTimeoutMs（个别 Provider 历史超时不同）、modelInfoTable。
|
||||
*
|
||||
* 外部类型穿透铁律不变：OpenAI 原生类型止步于本文件，向上只产出 Metona IR。
|
||||
*/
|
||||
|
||||
import type {
|
||||
MetonaRequest,
|
||||
MetonaResponse,
|
||||
MetonaStreamEvent,
|
||||
} from '../../types';
|
||||
import { MetonaFinishReason } from '../../types';
|
||||
import type { MetonaModelInfo } from '../../types/metona-adapter';
|
||||
import { parseSSEStream, parseOpenAICompatibleResponse } from './sse-stream';
|
||||
import { BaseAdapter } from '../base-adapter';
|
||||
|
||||
export abstract class OpenAICompatibleAdapter extends BaseAdapter {
|
||||
/**
|
||||
* POST /chat/completions 的完整端点。
|
||||
* 绝大多数 Provider 为 `${baseURL}/chat/completions`；少数代理需要自定义。
|
||||
*/
|
||||
protected abstract chatCompletionsUrl(): string;
|
||||
|
||||
/**
|
||||
* 子类特有的请求体参数映射（messages/tools/thinking/max_tokens 等差异点）。
|
||||
* 返回不含 stream 字段的 body —— stream 由本基类统一注入。
|
||||
*/
|
||||
protected abstract toNativeRequest(
|
||||
request: MetonaRequest,
|
||||
stream: boolean,
|
||||
): Record<string, unknown> | Promise<Record<string, unknown>>;
|
||||
|
||||
/** 非流式 send 的默认超时。DeepSeek/MiMo/OpenAI=120s；Agnes 历史 300s，保留其值。 */
|
||||
protected abstract sendTimeoutMs(): number;
|
||||
|
||||
/** 模型元信息表（子类持有；用于 getContextWindow 回退链与钳制） */
|
||||
protected abstract modelInfoTable(): Record<string, MetonaModelInfo>;
|
||||
|
||||
/** getContextWindow 的最终兜底窗口（未配置且模型未知时使用） */
|
||||
protected defaultContextWindowFallback(): number {
|
||||
return 1_000_000;
|
||||
}
|
||||
|
||||
// ===== 认证头 =====
|
||||
|
||||
protected buildHeaders(): Record<string, string> {
|
||||
return {
|
||||
'Content-Type': 'application/json',
|
||||
Authorization: `Bearer ${this.config.apiKey}`,
|
||||
...this.config.headers,
|
||||
};
|
||||
}
|
||||
|
||||
// ===== POST {chatCompletionsUrl} （非流式） =====
|
||||
|
||||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||||
const nativeRequest = await this.toNativeRequest(request, false);
|
||||
|
||||
const response = await this.fetchWithTimeout(
|
||||
this.chatCompletionsUrl(),
|
||||
{
|
||||
method: 'POST',
|
||||
headers: this.buildHeaders(),
|
||||
body: JSON.stringify({ ...nativeRequest, stream: false }),
|
||||
},
|
||||
this.config.timeoutMs ?? this.sendTimeoutMs(),
|
||||
);
|
||||
|
||||
if (!response.ok) {
|
||||
await this.throwHttpError(response, `${this.providerLabel()} API error`);
|
||||
}
|
||||
|
||||
const data = (await response.json()) as Record<string, unknown>;
|
||||
return this.toMetonaResponseFromOpenAI(data, request.meta.requestId);
|
||||
}
|
||||
|
||||
// ===== POST {chatCompletionsUrl} （流式） =====
|
||||
|
||||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||||
const nativeRequest = await this.toNativeRequest(request, true);
|
||||
|
||||
const response = await this.fetchWithTimeout(
|
||||
this.chatCompletionsUrl(),
|
||||
{
|
||||
method: 'POST',
|
||||
headers: this.buildHeaders(),
|
||||
body: JSON.stringify({ ...nativeRequest, stream: true }),
|
||||
},
|
||||
this.config.timeoutMs ?? 300_000,
|
||||
);
|
||||
|
||||
if (!response.ok || !response.body) {
|
||||
await this.throwHttpError(response, `${this.providerLabel()} stream error`);
|
||||
}
|
||||
|
||||
yield* parseSSEStream(
|
||||
// 非空断言：上方 if 已确保 response.body 不为 null
|
||||
response.body!,
|
||||
request.meta.requestId,
|
||||
request.meta.sessionId,
|
||||
request.meta.iteration,
|
||||
);
|
||||
}
|
||||
|
||||
// ===== 共享装配 =====
|
||||
|
||||
/** Provider 展示名（错误上下文用）：默认取 providerId，子类可覆盖 */
|
||||
protected providerLabel(): string {
|
||||
return this.providerId;
|
||||
}
|
||||
|
||||
/**
|
||||
* 非流式响应组装 —— OpenAI 原生结构到 MetonaResponse 的唯一映射点
|
||||
* （此前在四个子类各有一份逐字拷贝）
|
||||
*/
|
||||
private toMetonaResponseFromOpenAI(
|
||||
data: Record<string, unknown>,
|
||||
requestId: string,
|
||||
): MetonaResponse {
|
||||
const parsed = parseOpenAICompatibleResponse(data);
|
||||
return {
|
||||
meta: {
|
||||
requestId,
|
||||
provider: this.providerId,
|
||||
model: (data.model as string) ?? this.config.defaultModel,
|
||||
latencyMs: 0,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
content: parsed.content,
|
||||
reasoningContent: parsed.reasoningContent,
|
||||
toolCalls: parsed.toolCalls,
|
||||
usage: parsed.usage,
|
||||
finishReason: this.mapFinishReasonToMetona(parsed.finishReason),
|
||||
};
|
||||
}
|
||||
|
||||
/** parseOpenAIFinishReason 输出 → MetonaFinishReason 枚举（显式映射替代裸 as 断言） */
|
||||
private mapFinishReasonToMetona(reason: string): MetonaFinishReason {
|
||||
switch (reason) {
|
||||
case 'length':
|
||||
return MetonaFinishReason.LENGTH;
|
||||
case 'tool_calls':
|
||||
return MetonaFinishReason.TOOL_CALLS;
|
||||
case 'content_filter':
|
||||
return MetonaFinishReason.CONTENT_FILTER;
|
||||
case 'error':
|
||||
return MetonaFinishReason.ERROR;
|
||||
default:
|
||||
return MetonaFinishReason.STOP;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 上下文窗口回退链（v0.6.3 一致化后的统一实现）：
|
||||
* config.contextWindow（用户显式配置）→ 模型元信息 → Provider 兜底。
|
||||
*/
|
||||
override getContextWindow(): number {
|
||||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||||
return this.config.contextWindow;
|
||||
}
|
||||
const modelInfo = this.modelInfoTable()[this.config.defaultModel];
|
||||
return modelInfo?.contextWindow ?? this.defaultContextWindowFallback();
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 模型能力限制类错误（v0.6.4 升级：此前 OpenAI 推理模型拒图抛裸 Error，
|
||||
* 引擎分类落到 UNKNOWN，UI 无法区分"该模型不支持图"与一般故障）。
|
||||
* 携带 status=400 使引擎按"不可重试请求级错误"处理并直接展示原因。
|
||||
*/
|
||||
export class ModelCapabilityError extends Error {
|
||||
readonly status = 400;
|
||||
constructor(model: string, capability: string) {
|
||||
super(`Model "${model}" does not support ${capability}`);
|
||||
this.name = 'ModelCapabilityError';
|
||||
}
|
||||
}
|
||||
@@ -12,6 +12,158 @@ import { nanoid } from 'nanoid';
|
||||
import log from 'electron-log';
|
||||
import type { MetonaStreamEvent, MetonaTokenUsage } from '../../types';
|
||||
import { MetonaStreamEventType } from '../../types';
|
||||
import { ContentFilterError } from '../base-adapter';
|
||||
|
||||
/**
|
||||
* v0.6.4 根治「上游错误帧黑洞」:
|
||||
*
|
||||
* OpenAI 兼容网关常在中途发送 `data: {"error":{...}}` 数据帧（网关超时、限流、
|
||||
* 配额耗尽、鉴权失效等）。旧解析器整条处理链只从 `chunk.choices?.[0]?.delta` 取数，
|
||||
* 这类帧两个分支都不命中、零日志 —— 结果任何上游错误都伪装成"干净的空回复 +
|
||||
* 正常 DONE"，且异常以普通事件而非异常形式出现，绕过了引擎 chatStreamWithRetry
|
||||
* 的重试/故障转移通道。
|
||||
*
|
||||
* 现契约：上游错误帧在解析层**直接抛出**携带结构化 status/providerCode 的
|
||||
* SseUpstreamError —— 异常沿 async generator 传播进 chatStreamWithRetry 的 catch，
|
||||
* 使 429/5xx 自动走指数退避重试、401 等不可重试错误走 fallback 故障转移，
|
||||
* 与 HTTP 状态码路径的行为完全对齐（错误分类单轨化的流式半边）。
|
||||
*/
|
||||
export class SseUpstreamError extends Error {
|
||||
/** 归一化后的 HTTP status（当帧内无数值 status 时按 providerCode 推断） */
|
||||
readonly status?: number;
|
||||
/** 上游原始错误码（如 "rate_limit_exceeded" / "insufficient_quota"） */
|
||||
readonly providerCode?: string;
|
||||
|
||||
constructor(message: string, options?: { status?: number; providerCode?: string }) {
|
||||
super(message);
|
||||
this.name = 'SseUpstreamError';
|
||||
this.status = options?.status;
|
||||
this.providerCode = options?.providerCode;
|
||||
}
|
||||
}
|
||||
|
||||
/** v0.6.4: OpenAI 兼容流帧的最小结构化类型（仅承载本解析器实际消费的字段） */
|
||||
interface SseStreamFrame {
|
||||
choices?: Array<{
|
||||
delta?: {
|
||||
content?: string;
|
||||
reasoning_content?: string;
|
||||
tool_calls?: Array<{
|
||||
index?: number;
|
||||
function?: { name?: string; arguments?: string };
|
||||
}>;
|
||||
};
|
||||
finish_reason?: string;
|
||||
}>;
|
||||
usage?: {
|
||||
prompt_tokens?: number;
|
||||
completion_tokens?: number;
|
||||
total_tokens?: number;
|
||||
prompt_cache_hit_tokens?: number;
|
||||
prompt_cache_miss_tokens?: number;
|
||||
completion_tokens_details?: { reasoning_tokens?: number };
|
||||
prompt_tokens_details?: { cached_tokens?: number };
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* 从一条已 JSON.parse 的 SSE 数据帧中提取上游错误信息。
|
||||
* 兼容三种形态：
|
||||
* 1. 顶层 `{error:{message,status}}` — OpenAI 兼容网关最常见
|
||||
* 2. `{choices:[{error:{...}}]}` — 少数代理的变体包装
|
||||
* 3. `{error:"plain string"}` — 极简实现
|
||||
* 返回 null 表示该帧不含错误（正常数据帧）。
|
||||
*/
|
||||
function extractUpstreamErrorFrame(
|
||||
chunk: unknown,
|
||||
): { message: string; status?: number; code?: string } | null {
|
||||
if (!chunk || typeof chunk !== 'object') return null;
|
||||
const c = chunk as Record<string, unknown>;
|
||||
let errObj: unknown = c.error;
|
||||
if (
|
||||
(!errObj || typeof errObj !== 'object') &&
|
||||
Array.isArray(c.choices) &&
|
||||
c.choices.length > 0
|
||||
) {
|
||||
errObj = (c.choices[0] as Record<string, unknown> | undefined)?.error;
|
||||
}
|
||||
|
||||
if (typeof errObj === 'string') {
|
||||
return errObj.trim() ? { message: errObj } : null;
|
||||
}
|
||||
if (!errObj || typeof errObj !== 'object') return null;
|
||||
|
||||
const e = errObj as Record<string, unknown>;
|
||||
const rawStatus =
|
||||
typeof e.status === 'number'
|
||||
? e.status
|
||||
: typeof e.status_code === 'number'
|
||||
? e.status_code
|
||||
: undefined;
|
||||
const rawCode = typeof e.code === 'string' ? e.code : typeof e.type === 'string' ? e.type : '';
|
||||
const message =
|
||||
typeof e.message === 'string'
|
||||
? e.message
|
||||
: typeof e.msg === 'string'
|
||||
? e.msg
|
||||
: JSON.stringify(errObj);
|
||||
|
||||
// 无消息且无状态码的无害空对象不算错误（防御性）
|
||||
if (!message && rawStatus === undefined && !rawCode) return null;
|
||||
return { message: message || `upstream error (${rawCode || rawStatus})`, status: rawStatus, code: rawCode || undefined };
|
||||
}
|
||||
|
||||
/** 上游字符串错误码 → 归一化 HTTP status（用于帧内缺失数值 status 时仍能驱动重试判定） */
|
||||
function providerCodeToStatus(code: string): number | undefined {
|
||||
const c = code.toLowerCase();
|
||||
if (/rate_limit|too_many/.test(c)) return 429;
|
||||
if (/quota|insufficient|billing|exceeded_balance/.test(c)) return 402;
|
||||
if (/invalid_api_key|api_key_invalid|unauthorized|authentication/.test(c)) return 401;
|
||||
if (/forbidden|permission/.test(c)) return 403;
|
||||
if (/model_not_found|no_such_model/.test(c)) return 404;
|
||||
if (/overloaded|capacity|unavailable/.test(c)) return 503;
|
||||
if (/server_error|internal_error|internal/.test(c)) return 500;
|
||||
// 请求级 400 家族（无效参数/上下文超限）— 不映射到可重试区间，保持非重试语义
|
||||
return undefined;
|
||||
}
|
||||
|
||||
/**
|
||||
* 由提取出的错误信息构造待抛出的 SseUpstreamError /
|
||||
* ContentFilterError（content_filter 类直接复用既有专用类型）。
|
||||
*/
|
||||
function makeUpstreamThrowable(info: { message: string; status?: number; code?: string }): Error {
|
||||
if (info.code && info.code.toLowerCase().includes('content_filter')) {
|
||||
return new ContentFilterError(info.message, 'SSE 流中收到上游安全审核错误');
|
||||
}
|
||||
const status = info.status ?? (info.code ? providerCodeToStatus(info.code) : undefined);
|
||||
return new SseUpstreamError(`upstream_error${info.code ? ` (${info.code})` : ''}: ${info.message}`, {
|
||||
status,
|
||||
providerCode: info.code,
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* 截断参数自愈载荷的唯一构造点（v0.6.4: 非流式/Ollama NDJSON/Anthropic 全线共用）。
|
||||
*
|
||||
* 原 v0.6.3 只覆盖了 OpenAI 共享 SSE 层；此处抽出为共享函数后，所有协议路径的
|
||||
* 截断工具调用统一转为显式错误参数 —— 工具执行失败 → 错误结果回传模型 → 模型
|
||||
* 重试/分块写入（ReAct 自愈闭环），彻底消灭"静默丢弃 → 空回复 → 会话无声终止"。
|
||||
*/
|
||||
export function truncatedArgumentsPayload(
|
||||
parseErrorMessage: string,
|
||||
rawTailSample: string,
|
||||
): Record<string, unknown> {
|
||||
return {
|
||||
_truncatedArguments: true,
|
||||
_truncatedReason:
|
||||
'The tool-call arguments JSON was truncated before completion ' +
|
||||
'(likely max_tokens output limit reached while generating this tool call). ' +
|
||||
'The original arguments are lost and cannot be recovered. Please retry with a ' +
|
||||
'smaller output (e.g. write the file in smaller chunks) — do NOT reuse or repeat ' +
|
||||
'the previous oversized arguments.' +
|
||||
` [parser: ${parseErrorMessage}; tail sample: ...${rawTailSample}]`,
|
||||
};
|
||||
}
|
||||
|
||||
/**
|
||||
* L-4 修复: 提取 flushToolCallBuffer 辅助函数，消除 [DONE] 分支和 finish_reason='tool_calls' 分支的重复代码
|
||||
@@ -62,8 +214,9 @@ function* flushToolCallBuffer(
|
||||
},
|
||||
};
|
||||
} catch (err) {
|
||||
// v0.6.3: 截断的工具调用转显式错误参数（不丢弃）— 工具执行失败后
|
||||
// 错误结果回传模型，触发重试/分块写入，替代"静默丢弃→空回复终止会话"
|
||||
// v0.6.3 → v0.6.4: 截断的工具调用转显式错误参数（不丢弃）— 工具执行失败后
|
||||
// 错误结果回传模型，触发重试/分块写入。载荷构造已收敛到共享的
|
||||
// truncatedArgumentsPayload（Ollama NDJSON / Anthropic 事件机 / 非流式同步复用）。
|
||||
const rawTail = buf.argsBuffer.slice(-120);
|
||||
log.warn(
|
||||
`[SSE] Tool call args truncated (unparseable JSON, ${(err as Error).message}). ` +
|
||||
@@ -79,14 +232,7 @@ function* flushToolCallBuffer(
|
||||
toolCall: {
|
||||
id: `tc_${nanoid(8)}`,
|
||||
name: buf.name,
|
||||
args: {
|
||||
_truncatedArguments: true,
|
||||
_truncatedReason:
|
||||
'The streamed arguments JSON was truncated before completion ' +
|
||||
'(likely max_tokens output limit reached while generating this tool call). ' +
|
||||
'The original arguments are lost. Please retry with smaller output ' +
|
||||
'(e.g. write the file in smaller chunks) — do NOT reuse the previous oversized arguments.',
|
||||
},
|
||||
args: truncatedArgumentsPayload((err as Error).message, rawTail),
|
||||
iteration,
|
||||
timestamp: Date.now(),
|
||||
},
|
||||
@@ -149,8 +295,11 @@ export async function* parseSSEStream(
|
||||
|
||||
for (const line of lines) {
|
||||
const trimmed = line.trim();
|
||||
if (!trimmed || !trimmed.startsWith('data: ')) continue;
|
||||
const data = trimmed.slice(6);
|
||||
// v0.6.4: 兼容 `data:{...}`（无空格）变体 — 部分代理网关不带空格，
|
||||
// 原实现的 startsWith('data: ') 会将其整帧跳过
|
||||
if (!trimmed || !trimmed.startsWith('data:')) continue;
|
||||
const data = trimmed.slice(5).trim();
|
||||
if (!data) continue;
|
||||
|
||||
// 流结束
|
||||
if (data === '[DONE]') {
|
||||
@@ -169,109 +318,135 @@ export async function* parseSSEStream(
|
||||
return;
|
||||
}
|
||||
|
||||
// v0.6.4: 结构化类型承载帧内容（替代 JSON.parse 的隐式 any，杜绝字段漂移）
|
||||
let chunk: SseStreamFrame;
|
||||
try {
|
||||
const chunk = JSON.parse(data);
|
||||
const delta = chunk.choices?.[0]?.delta;
|
||||
|
||||
// 文本内容增量
|
||||
if (delta?.content) {
|
||||
yield {
|
||||
type: MetonaStreamEventType.TEXT_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
delta: delta.content,
|
||||
};
|
||||
}
|
||||
|
||||
// 推理内容增量（Thinking 模式）
|
||||
if (delta?.reasoning_content) {
|
||||
yield {
|
||||
type: MetonaStreamEventType.REASONING_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
delta: delta.reasoning_content,
|
||||
};
|
||||
}
|
||||
|
||||
// 工具调用增量 — 缓冲拼接
|
||||
if (delta?.tool_calls) {
|
||||
for (const tc of delta.tool_calls) {
|
||||
const idx = tc.index ?? 0;
|
||||
if (!toolCallsBuffer.has(idx)) {
|
||||
toolCallsBuffer.set(idx, { name: tc.function?.name ?? '', argsBuffer: '' });
|
||||
}
|
||||
const buf = toolCallsBuffer.get(idx)!;
|
||||
if (tc.function?.name) buf.name = tc.function.name;
|
||||
if (tc.function?.arguments) buf.argsBuffer += tc.function.arguments;
|
||||
|
||||
yield {
|
||||
type: MetonaStreamEventType.TOOL_CALL_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
toolCallDelta: {
|
||||
index: idx,
|
||||
name: tc.function?.name,
|
||||
argsDelta: tc.function?.arguments,
|
||||
},
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
// Token 使用统计 / finish_reason
|
||||
if (chunk.usage) {
|
||||
const usage: MetonaTokenUsage = {
|
||||
inputTokens: chunk.usage.prompt_tokens ?? 0,
|
||||
outputTokens: chunk.usage.completion_tokens ?? 0,
|
||||
totalTokens: chunk.usage.total_tokens ?? 0,
|
||||
reasoningTokens: chunk.usage.completion_tokens_details?.reasoning_tokens,
|
||||
// DeepSeek: prompt_cache_hit_tokens / prompt_cache_miss_tokens
|
||||
// MiMo: prompt_tokens_details.cached_tokens
|
||||
cacheHitTokens:
|
||||
chunk.usage.prompt_cache_hit_tokens ??
|
||||
chunk.usage.prompt_tokens_details?.cached_tokens,
|
||||
cacheMissTokens: chunk.usage.prompt_cache_miss_tokens,
|
||||
};
|
||||
|
||||
yield {
|
||||
type: MetonaStreamEventType.USAGE,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
usage,
|
||||
};
|
||||
}
|
||||
|
||||
// 非 [DONE] 但 finish_reason 为 tool_calls 时提前 flush 缓冲区
|
||||
const finishReason = chunk.choices?.[0]?.finish_reason as string | undefined;
|
||||
if (finishReason === 'tool_calls') {
|
||||
// L-4 修复: 使用 flushToolCallBuffer 替代重复的遍历代码
|
||||
yield* flushToolCallBuffer(toolCallsBuffer, requestId, sessionId, iteration, seqRef);
|
||||
}
|
||||
// v0.6.3 归因: 输出 token 上限截断（长工具参数/长文本的常见根因）显式落日志
|
||||
if (finishReason === 'length') {
|
||||
log.warn(
|
||||
`[SSE] finish_reason=length — output truncated by max_tokens limit ` +
|
||||
`(accumulated argsBuffer: ${[...toolCallsBuffer.values()].reduce((n, b) => n + b.argsBuffer.length, 0)} chars, ` +
|
||||
`model may retry with smaller output)`,
|
||||
);
|
||||
}
|
||||
chunk = JSON.parse(data) as SseStreamFrame;
|
||||
} catch (parseErr) {
|
||||
// P2-8 修复: 不再静默跳过，记录 warning 便于排查 SSE 数据损坏
|
||||
log.warn(
|
||||
`[SSE] Failed to parse stream line: ${(parseErr as Error).message}`,
|
||||
line.slice(0, 200),
|
||||
);
|
||||
continue;
|
||||
}
|
||||
|
||||
// v0.6.4: 上游错误帧检测（根治"错误帧黑洞"）。解析失败直接 throw，
|
||||
// 异常沿 chatStreamWithRetry 的 catch 走重试/故障转移通道；
|
||||
// 工具调用缓冲不 flush —— 重试会从零重建整个响应。
|
||||
const upstreamError = extractUpstreamErrorFrame(chunk);
|
||||
if (upstreamError) {
|
||||
log.warn(
|
||||
`[SSE] Upstream error frame received: code=${upstreamError.code ?? 'n/a'} status=${upstreamError.status ?? 'n/a'} message=${upstreamError.message.slice(0, 300)} — throwing for retry/failover handling`,
|
||||
);
|
||||
throw makeUpstreamThrowable(upstreamError);
|
||||
}
|
||||
|
||||
const delta = chunk.choices?.[0]?.delta;
|
||||
|
||||
// 文本内容增量
|
||||
if (delta?.content) {
|
||||
yield {
|
||||
type: MetonaStreamEventType.TEXT_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
delta: delta.content,
|
||||
};
|
||||
}
|
||||
|
||||
// 推理内容增量（Thinking 模式）
|
||||
if (delta?.reasoning_content) {
|
||||
yield {
|
||||
type: MetonaStreamEventType.REASONING_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
delta: delta.reasoning_content,
|
||||
};
|
||||
}
|
||||
|
||||
// 工具调用增量 — 缓冲拼接
|
||||
if (delta?.tool_calls) {
|
||||
for (const tc of delta.tool_calls) {
|
||||
const idx = tc.index ?? 0;
|
||||
if (!toolCallsBuffer.has(idx)) {
|
||||
toolCallsBuffer.set(idx, { name: tc.function?.name ?? '', argsBuffer: '' });
|
||||
}
|
||||
const buf = toolCallsBuffer.get(idx)!;
|
||||
if (tc.function?.name) buf.name = tc.function.name;
|
||||
if (tc.function?.arguments) buf.argsBuffer += tc.function.arguments;
|
||||
|
||||
yield {
|
||||
type: MetonaStreamEventType.TOOL_CALL_DELTA,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
toolCallDelta: {
|
||||
index: idx,
|
||||
name: tc.function?.name,
|
||||
argsDelta: tc.function?.arguments,
|
||||
},
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
// Token 使用统计 / finish_reason
|
||||
const usageRaw = chunk.usage;
|
||||
if (usageRaw) {
|
||||
const usage: MetonaTokenUsage = {
|
||||
inputTokens: usageRaw.prompt_tokens ?? 0,
|
||||
outputTokens: usageRaw.completion_tokens ?? 0,
|
||||
totalTokens: usageRaw.total_tokens ?? 0,
|
||||
reasoningTokens: usageRaw.completion_tokens_details?.reasoning_tokens,
|
||||
// DeepSeek: prompt_cache_hit_tokens / prompt_cache_miss_tokens
|
||||
// MiMo: prompt_tokens_details.cached_tokens
|
||||
cacheHitTokens:
|
||||
usageRaw.prompt_cache_hit_tokens ?? usageRaw.prompt_tokens_details?.cached_tokens,
|
||||
cacheMissTokens: usageRaw.prompt_cache_miss_tokens,
|
||||
};
|
||||
|
||||
yield {
|
||||
type: MetonaStreamEventType.USAGE,
|
||||
requestId,
|
||||
sessionId,
|
||||
iteration,
|
||||
seq: seqRef.seq++,
|
||||
timestamp: Date.now(),
|
||||
usage,
|
||||
};
|
||||
}
|
||||
|
||||
// 非 [DONE] 但 finish_reason 为 tool_calls 时提前 flush 缓冲区
|
||||
const finishReason = chunk.choices?.[0]?.finish_reason as string | undefined;
|
||||
if (finishReason === 'tool_calls') {
|
||||
// L-4 修复: 使用 flushToolCallBuffer 替代重复的遍历代码
|
||||
yield* flushToolCallBuffer(toolCallsBuffer, requestId, sessionId, iteration, seqRef);
|
||||
}
|
||||
// v0.6.3 归因: 输出 token 上限截断（长工具参数/长文本的常见根因）显式落日志
|
||||
if (finishReason === 'length') {
|
||||
log.warn(
|
||||
`[SSE] finish_reason=length — output truncated by max_tokens limit ` +
|
||||
`(accumulated argsBuffer: ${[...toolCallsBuffer.values()].reduce((n, b) => n + b.argsBuffer.length, 0)} chars, ` +
|
||||
`model may retry with smaller output)`,
|
||||
);
|
||||
}
|
||||
// v0.6.4: 流式 content_filter 终止映射 —— 非流式路径早已支持
|
||||
// （throwHttpError → ContentFilterError），流式此前既不映射也不打日志，
|
||||
// 引擎拿到普通结束、用户看不到拦截原因。抛专用类型使 finish() 映射为
|
||||
// CONTENT_FILTERED 错误码 + 友好提示，且不会被重试逻辑反复重放。
|
||||
if (finishReason === 'content_filter') {
|
||||
log.warn('[SSE] finish_reason=content_filter — provider safety filter terminated the response');
|
||||
throw new ContentFilterError(
|
||||
'流式响应被 Provider 安全审核终止',
|
||||
'SSE stream (finish_reason=content_filter)',
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -327,8 +502,16 @@ export function parseOpenAICompatibleResponse(data: Record<string, unknown>): {
|
||||
if (typeof rawArgs === 'string') {
|
||||
try {
|
||||
args = JSON.parse(rawArgs);
|
||||
} catch {
|
||||
args = {};
|
||||
} catch (err) {
|
||||
// v0.6.4: 非流式路径与流式截断自愈对齐 —— 此前坏参静默降级 {} 与流式
|
||||
// 的显式自愈行为不一致（v0.6.3 只修了流式半边）。空参数会让工具以
|
||||
// "缺少必要参数"泛化失败，模型无法得知发生了截断；现在统一转为
|
||||
// _truncatedArguments 错误参数，触发模型分块重试。
|
||||
const sample = rawArgs.slice(-120);
|
||||
log.warn(
|
||||
`[SSE] Non-stream tool call args truncated (unparseable JSON, ${(err as Error).message}). Tail: ...${sample}`,
|
||||
);
|
||||
args = truncatedArgumentsPayload((err as Error).message, sample);
|
||||
}
|
||||
} else if (rawArgs && typeof rawArgs === 'object') {
|
||||
args = rawArgs as Record<string, unknown>;
|
||||
|
||||
Reference in New Issue
Block a user