定位 'AI 回复 0 字' 时单凭 finish_reason 和 aborted 区分不出 '上游没发任何 data' 和 '上游发了帧但 content 为空'。callModel 在 forEachSSEData 回调首行 累加 sseFrames 并写进返回值(不影响其他字段),streamAssistant 拿到 result 后打印一行 info: [sse] <sessionId> 本轮汇总:帧 N,finish=<reason>,aborted=<bool> 判读: - 帧=0 且 aborted=false:上游未发任何 data(流异常/上游 200 但 body 空) - 帧>0 且 finish=stop 且 content 空:上游发了帧但 content 为空(罕见,模型主动空回) - 帧>0 且 finish=length:上游被截断,需要重试或调大 max_tokens - aborted=true:用户在流中途取消 llm.js 保持纯 IO 职责,只多返回一个计数字段,不引入 logger。
225 lines
9.5 KiB
JavaScript
225 lines
9.5 KiB
JavaScript
// 上游调用:OpenAI 兼容 /chat/completions 流式接口(由 cli-proxy-api 代理)。
|
||
// 累积 content 与 tool_calls 分片;支持 abort。
|
||
const fs = require('fs');
|
||
const path = require('path');
|
||
|
||
const SYSTEM_PROMPT_TEMPLATE = fs.readFileSync(path.join(__dirname, 'templates', 'system-prompt.txt'), 'utf8');
|
||
|
||
// system prompt 完全静态:所有会话共用同一份内容,利于上游前缀缓存。
|
||
// 动态会话信息(日期/任务目录/模型显示名)不再拼入 system prompt,
|
||
// 而是在新建会话时作为首条 user 消息发送,见 engine.js 的 sessionContextContent。
|
||
function buildSystemPrompt() {
|
||
return SYSTEM_PROMPT_TEMPLATE;
|
||
}
|
||
|
||
// 部署级回复语言指令:拼在 system prompt 末尾(尾部对弱模型约束力最强)。
|
||
// 同一进程内恒定,不破坏前缀缓存;改配置需重启生效。
|
||
function systemPromptText(language) {
|
||
const t = String(language || '').trim();
|
||
if (!t) return SYSTEM_PROMPT_TEMPLATE;
|
||
return SYSTEM_PROMPT_TEMPLATE +
|
||
'\n\n## Reply Language\n' +
|
||
'- ALWAYS respond to the user in ' + t + '. Tool descriptions, tool outputs and errors are usually in English — ' +
|
||
'never let them switch your reply language. Tool names, parameters, code, file paths and shell commands stay as-is; ' +
|
||
'every sentence addressed to the user MUST be in ' + t + '.';
|
||
}
|
||
|
||
// 多模态内容块存取:messages.content 列为 TEXT,含图片的 tool 消息
|
||
// 以 {"__blocks__":[content...]} 哨兵 JSON 落库(见需求 §5),
|
||
// 发请求时还原为 OpenAI 内容块数组(text + image_url)。tool/user 消息共用。
|
||
function encodeBlocks(blocks) { return JSON.stringify({ __blocks__: blocks }); }
|
||
function decodeBlocks(content) {
|
||
if (typeof content !== 'string' || !content.startsWith('{"__blocks__":')) return null;
|
||
try {
|
||
const parsed = JSON.parse(content);
|
||
return Array.isArray(parsed.__blocks__) ? parsed.__blocks__ : null;
|
||
} catch (_) { return null; }
|
||
}
|
||
|
||
// 将 DB 消息行转为 OpenAI messages 数组(不含 system,由调用方拼接)。
|
||
// 省 token:image_url 块仅在其位于最后一条 assistant 消息之后(模型尚未对它回复)时保留;
|
||
// 之前的视为已消费,发请求时替换为文字标记。DB 始终存全量:UI 历史照常渲染图片,
|
||
// 模型如需重看可再次 read_file(新消息在尾部,会重新携带图片)。
|
||
function buildMessages(session, language) {
|
||
const messages = [{ role: 'system', content: systemPromptText(language) }];
|
||
const history = session.history;
|
||
let lastAssistant = -1;
|
||
for (let i = history.length - 1; i >= 0; i--) {
|
||
if (history[i].role === 'assistant') { lastAssistant = i; break; }
|
||
}
|
||
history.forEach((row, i) => {
|
||
if (row.role === 'assistant') {
|
||
const m = { role: 'assistant', content: row.content || '' };
|
||
if (row.tool_calls) m.tool_calls = JSON.parse(row.tool_calls);
|
||
if (!m.content) delete m.content;
|
||
messages.push(m);
|
||
return;
|
||
}
|
||
// tool/user 行:含图哨兵 JSON 还原为内容块;已消费的图剥掉 base64,只留文字描述 + 标记
|
||
const blocks = decodeBlocks(row.content);
|
||
let content = row.content || '';
|
||
if (blocks) {
|
||
if (i < lastAssistant) {
|
||
content = blocks.filter(b => b && b.type === 'text').map(b => b.text).join('\n') +
|
||
'\n[图片数据已省略:模型此前已识别过本图;如需重新查看请再次调用 read_file]';
|
||
} else {
|
||
content = blocks;
|
||
}
|
||
}
|
||
if (row.role === 'tool') {
|
||
messages.push({ role: 'tool', tool_call_id: row.tool_call_id, name: row.name, content });
|
||
} else {
|
||
messages.push({ role: row.role, content });
|
||
}
|
||
});
|
||
return messages;
|
||
}
|
||
|
||
// 解析 SSE 行流,对每个 data: JSON 调用 onEvent。
|
||
// 可选回调:onMeta(line) 逐行接收非 data 字段(event:/id:/retry:/注释/未知字段),
|
||
// onBadData(raw) 在 data 行 JSON.parse 失败时被调用(原始字符串)。
|
||
// 两个回调都可选,传入则启用;流式行为不变(data 仍逐帧下传到 onEvent)。
|
||
async function forEachSSEData(response, onEvent, onMeta, onBadData) {
|
||
const reader = response.body.getReader();
|
||
const decoder = new TextDecoder();
|
||
let buf = '';
|
||
while (true) {
|
||
const { done, value } = await reader.read();
|
||
if (done) break;
|
||
buf += decoder.decode(value, { stream: true });
|
||
let idx;
|
||
while ((idx = buf.indexOf('\n\n')) >= 0) {
|
||
const block = buf.slice(0, idx);
|
||
buf = buf.slice(idx + 2);
|
||
const dataPieces = [];
|
||
for (const line of block.split('\n')) {
|
||
const t = line.trim();
|
||
if (t.startsWith('data:')) {
|
||
dataPieces.push(t.slice(5).trim());
|
||
} else if (t.length && onMeta) {
|
||
// SSE 注释行以 ':' 开头(如 ":keep-alive"),这里原样交给 onMeta
|
||
onMeta(t);
|
||
}
|
||
}
|
||
if (!dataPieces.length) continue;
|
||
const data = dataPieces.join('\n');
|
||
if (data === '[DONE]') return;
|
||
try {
|
||
onEvent(JSON.parse(data));
|
||
} catch (_) {
|
||
if (onBadData) onBadData(data);
|
||
}
|
||
}
|
||
}
|
||
}
|
||
|
||
// 深度合并 src 到 dst(就地改 dst):同名键两边都是普通对象则递归;其余 src 覆盖 dst。
|
||
// 用于将模型的 extra_params / thinking_params 合入上游请求体。
|
||
function deepMergeInto(dst, src) {
|
||
for (const k of Object.keys(src || {})) {
|
||
const sv = src[k];
|
||
if (sv && typeof sv === 'object' && !Array.isArray(sv) && dst[k] && typeof dst[k] === 'object' && !Array.isArray(dst[k])) {
|
||
deepMergeInto(dst[k], sv);
|
||
} else {
|
||
dst[k] = sv;
|
||
}
|
||
}
|
||
return dst;
|
||
}
|
||
|
||
// 调模型一次。onText(textDelta) 回调流式文本。
|
||
// 返回 { content, tool_calls: [{id,name,arguments(string)}], finish_reason, usage, aborted }
|
||
// options.deepThink: true 时将 modelCfg.thinking_params 合并进 body。
|
||
async function callModel(cfg, modelId, messages, tools, options) {
|
||
const opts = options || {};
|
||
const controller = new AbortController();
|
||
if (opts.signal) {
|
||
if (opts.signal.aborted) controller.abort();
|
||
else opts.signal.addEventListener('abort', () => controller.abort());
|
||
}
|
||
// modelId = 内部唯一键「平台名/模型id」;发给上游的模型名用 mid(原始 id)。
|
||
const modelCfg = (cfg.models || []).find(m => m.id === modelId);
|
||
if (!modelCfg || !modelCfg.url) throw new Error('模型未在配置中定义或缺少 url: ' + modelId);
|
||
|
||
const body = {
|
||
model: modelCfg.mid || modelCfg.id,
|
||
messages,
|
||
stream: true,
|
||
stream_options: { include_usage: true },
|
||
temperature: cfg.agent.temperature,
|
||
};
|
||
if (tools && tools.length) body.tools = tools;
|
||
// 模型特有请求参数:与内置字段同名时额外优先(同时思考参数可覆盖)
|
||
if (modelCfg.extra_params && Object.keys(modelCfg.extra_params).length) {
|
||
deepMergeInto(body, modelCfg.extra_params);
|
||
}
|
||
if (opts.deepThink && modelCfg.thinking_params && Object.keys(modelCfg.thinking_params).length) {
|
||
deepMergeInto(body, modelCfg.thinking_params);
|
||
}
|
||
|
||
const headers = { 'Content-Type': 'application/json' };
|
||
if (modelCfg.api_key) headers.Authorization = 'Bearer ' + modelCfg.api_key;
|
||
|
||
const res = await fetch(modelCfg.url + '/chat/completions', {
|
||
method: 'POST',
|
||
headers,
|
||
body: JSON.stringify(body),
|
||
signal: controller.signal,
|
||
});
|
||
if (!res.ok) {
|
||
const text = await res.text().catch(() => '');
|
||
const err = new Error('上游返回 ' + res.status + ': ' + text.slice(0, 500));
|
||
err.upstream_status = res.status;
|
||
throw err;
|
||
}
|
||
|
||
let content = '';
|
||
let reasoning = '';
|
||
let finishReason = null;
|
||
let usage = null;
|
||
let aborted = false;
|
||
let sseFrames = 0; // 收到的上游 data 帧数(仅计数成功下传到 onEvent 的)
|
||
const toolCalls = new Map(); // index -> {id, name, arguments}
|
||
|
||
try {
|
||
await forEachSSEData(res, evt => {
|
||
sseFrames++;
|
||
if (opts.shouldCancel && opts.shouldCancel()) { aborted = true; controller.abort(); return; }
|
||
if (evt.usage) usage = evt.usage;
|
||
const choice = evt.choices && evt.choices[0];
|
||
if (!choice) return;
|
||
if (choice.finish_reason) finishReason = choice.finish_reason;
|
||
const delta = choice.delta || {};
|
||
if (typeof delta.content === 'string' && delta.content.length) {
|
||
content += delta.content;
|
||
if (opts.onText) opts.onText(delta.content);
|
||
}
|
||
// 思考/推理字段:各家上游不一致,同时支持 reasoning 和 reasoning_content;只走一个,优先级 reasoning
|
||
const rPiece = (typeof delta.reasoning === 'string' && delta.reasoning)
|
||
|| (typeof delta.reasoning_content === 'string' && delta.reasoning_content)
|
||
|| '';
|
||
if (rPiece) {
|
||
reasoning += rPiece;
|
||
if (opts.onReasoning) opts.onReasoning(rPiece);
|
||
}
|
||
for (const tc of delta.tool_calls || []) {
|
||
const idx = tc.index || 0;
|
||
if (!toolCalls.has(idx)) toolCalls.set(idx, { id: '', name: '', arguments: '' });
|
||
const acc = toolCalls.get(idx);
|
||
if (tc.id) acc.id = tc.id;
|
||
if (tc.function) {
|
||
if (tc.function.name) acc.name += tc.function.name;
|
||
if (tc.function.arguments) acc.arguments += tc.function.arguments;
|
||
}
|
||
}
|
||
});
|
||
} catch (e) {
|
||
if (e && e.name === 'AbortError') aborted = true; else throw e;
|
||
}
|
||
if (opts.shouldCancel && opts.shouldCancel()) aborted = true;
|
||
|
||
return { content, reasoning, tool_calls: Array.from(toolCalls.values()), finish_reason: finishReason, usage, aborted, sseFrames };
|
||
}
|
||
|
||
module.exports = { callModel, buildMessages, buildSystemPrompt, systemPromptText, encodeBlocks, decodeBlocks };
|