@danceiny/gotry 0.0.1-rc.9 → 0.2.0-rc.28

package/dist/src/dsh-llm.js

@@ -1,174 +1,230 @@
+/**
+ * 真 LlmPort:OpenAI 兼容接口的 provider 中立适配器(DeepSeek/MiniMax-M2 实测)。
+ * 零新依赖(node 内建 fetch)。环境变量:LLM_API_KEY/LLM_BASE_URL/LLM_MODEL
+ * (兼容旧 DEEPSEEK_* 别名);GOTRY_LLM_TIMEOUT_MS = 单次请求总预算(默认 300000ms,
+ * 非正整数忽略)。MiniMax-M2 是推理模型:输出带 <think> 块,
+ * 必须先剥离再解析——JSON 藏在 think 里是常见失败模式。
+ * 无 key 时抛出明确错误,replay-real 自动回退 mock(ADR-8)。
+ * 责任铁律不变:本适配器只做翻译/润色/解释,判定与算术在确定性组件。
+ */
 import { parseFlightPackToSpec } from './unified.js';
 import { buildTimeAnchor } from './time-anchor.js';
 import { buildSlotSystem, flagExpiredSlots, normalizeExtraction } from './travel-slots.js';
-const model = ()=>process.env['LLM_MODEL'] ?? process.env['DEEPSEEK_MODEL'] ?? 'MiniMax-M2';
-const base = ()=>(process.env['LLM_BASE_URL'] ?? process.env['DEEPSEEK_BASE_URL'] ?? 'https://api.minimax.io/v1').replace(/\/$/, '');
-async function chat(messages, json) {
+import { mergeProfileWorkWindow } from './flight-pack-adapter.js';
+// 惰性读取(env 在调用时取值):模块顶常量会在 .env 加载前冻结(ESM import 提升),
+// 脚本先 loadEnv 再 import 也救不了——401 错配(key 发往默认端点)的存量隐患由此根除。
+const model = () => process.env['LLM_MODEL'] ?? process.env['DEEPSEEK_MODEL'] ?? 'MiniMax-M2';
+const base = () => (process.env['LLM_BASE_URL'] ?? process.env['DEEPSEEK_BASE_URL'] ?? 'https://api.minimax.io/v1').replace(/\/$/, '');
+function emptyUsage() {
+    return { calls: 0, inputTokens: 0, outputTokens: 0, inputCacheHitTokens: 0, inputCacheMissTokens: 0, responsesMissingUsage: 0 };
+}
+/** 单次请求(连接→响应头→正文读完)的默认总预算:对推理模型足够宽松,只兜「连上了却永不应答」的卡死。 */
+export const DEFAULT_LLM_TIMEOUT_MS = 300_000;
+/**
+ * 超时上限 = 2^31-1 ms(≈24.8 天)。Node 对更大的延迟会溢出并**悄悄改成 1ms**
+ * (TimeoutOverflowWarning),即「配得越大越立刻超时」——所以必须夹住,绝不原样透传。
+ */
+const MAX_LLM_TIMEOUT_MS = 2_147_483_647;
+/** 只认十进制正整数(数值或纯数字串);其余(0/负数/小数/NaN/带单位/空串)一律 undefined——绝不退化成「无超时」。 */
+function parseTimeoutMs(raw) {
+    const n = typeof raw === 'string' ? (/^\d+$/.test(raw.trim()) ? Number(raw.trim()) : Number.NaN) : raw;
+    return typeof n === 'number' && Number.isSafeInteger(n) && n > 0 ? Math.min(n, MAX_LLM_TIMEOUT_MS) : undefined;
+}
+/** 解析顺序:本次调用显式值 > 环境变量 GOTRY_LLM_TIMEOUT_MS(调用时取值,同 base()/model())> 默认值。 */
+export function resolveLlmTimeoutMs(callMs) {
+    return parseTimeoutMs(callMs) ?? parseTimeoutMs(process.env['GOTRY_LLM_TIMEOUT_MS']) ?? DEFAULT_LLM_TIMEOUT_MS;
+}
+/**
+ * chat() 请求被截断(超时 / 调用方取消)的类型化失败——不是挂起,也不会被吞。
+ * 与既有错误形态一致,而非另起一套:
+ *  - message 沿用本模块 `llm <kind>: <detail>` 约定(同 `llm 429: …`),只读 .message 的调用方
+ *    (nightly-evidence / persona-sim / time-eval --real)原样可用;
+ *  - name 取平台既有名 TimeoutError / AbortError(AbortSignal.timeout 与手动 abort 抛的同名),
+ *    persona-sim、channel-probe 等按 name 分类的代码零改动即可识别;
+ *  - cause 保留 fetch 抛出的原始错误(调用方自定义的 abort reason 也在其中)。
+ * HTTP 非 2xx(`llm <status>`)、DNS/连接重置、坏 JSON 仍是原样的普通 Error——只改写「被我们的 signal 终止」的失败。
+ */
+export class LlmRequestError extends Error {
+    code;
+    /** 本次生效的超时预算;仅 code === 'timeout' 有值 */
+    timeoutMs;
+    constructor(code, detail, options = {}) {
+        super(`llm ${code}: ${detail}`, options.cause === undefined ? undefined : { cause: options.cause });
+        this.name = code === 'timeout' ? 'TimeoutError' : 'AbortError';
+        this.code = code;
+        this.timeoutMs = options.timeoutMs;
+    }
+}
+/**
+ * 发一次请求并读完正文。**同一个 signal 贯穿 fetch 与 res.json()/res.text()**,所以超时覆盖整个请求
+ * (含「响应头已到、正文卡住」与「非 2xx 的错误正文卡住」),调用方取消同理。
+ * 只有「被我们的 signal 终止」才改写成 LlmRequestError;其余错误原样上抛。
+ */
+async function requestCompletion(url, init, request) {
+    const timeoutMs = resolveLlmTimeoutMs(request.timeoutMs);
+    const timeout = AbortSignal.timeout(timeoutMs);
+    const signal = request.signal ? AbortSignal.any([request.signal, timeout]) : timeout;
+    try {
+        const res = await fetch(url, { method: 'POST', headers: init.headers, body: init.body, signal });
+        if (!res.ok)
+            return { ok: false, status: res.status, text: await res.text() };
+        return { ok: true, data: await res.json() };
+    }
+    catch (error) {
+        if (!signal.aborted)
+            throw error;
+        // AbortSignal.any 的 reason 是最先触发的那个源的 reason:等于超时信号的 reason 即超时先到,否则是调用方取消
+        if (timeout.aborted && signal.reason === timeout.reason) {
+            throw new LlmRequestError('timeout', `no complete response within ${timeoutMs}ms (GOTRY_LLM_TIMEOUT_MS)`, { cause: error, timeoutMs });
+        }
+        throw new LlmRequestError('aborted', "request cancelled by the caller's AbortSignal", { cause: error });
+    }
+}
+async function chat(messages, json, usage, request = {}) {
     const key = process.env['LLM_API_KEY'] ?? process.env['DEEPSEEK_API_KEY'];
-    if (!key) throw new Error('LLM_API_KEY 未设置(兼容 DEEPSEEK_API_KEY 别名)——真 LLM 路径不可用,请回退 mock(ADR-8)');
-    const res = await fetch(`${base()}/chat/completions`, {
-        method: 'POST',
-        headers: {
-            'Content-Type': 'application/json',
-            Authorization: `Bearer ${key}`
-        },
+    if (!key)
+        throw new Error('LLM_API_KEY 未设置(兼容 DEEPSEEK_API_KEY 别名)——真 LLM 路径不可用,请回退 mock(ADR-8)');
+    const reply = await requestCompletion(`${base()}/chat/completions`, {
+        headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${key}` },
         body: JSON.stringify({
             model: model(),
             messages,
-            ...json ? {
-                response_format: {
-                    type: 'json_object'
-                }
-            } : {},
-            temperature: json ? 0 : 0.7
-        })
-    });
-    if (!res.ok) throw new Error(`llm ${res.status}: ${(await res.text()).slice(0, 300)}`);
-    const data = await res.json();
+            ...(json ? { response_format: { type: 'json_object' } } : {}),
+            temperature: json ? 0 : 0.7,
+        }),
+    }, request);
+    if (!reply.ok)
+        throw new Error(`llm ${reply.status}: ${reply.text.slice(0, 300)}`);
+    const data = reply.data;
+    if (usage) {
+        usage.calls += 1;
+        const u = data.usage;
+        if (u && Number.isFinite(u.prompt_tokens) && Number.isFinite(u.completion_tokens)) {
+            usage.inputTokens += u.prompt_tokens ?? 0;
+            usage.outputTokens += u.completion_tokens ?? 0;
+            usage.inputCacheHitTokens += u.prompt_cache_hit_tokens ?? 0;
+            // 无缓存明细的 provider:全部输入按 miss 记(随后按 peak miss 价换算,只高不低)
+            usage.inputCacheMissTokens += u.prompt_cache_miss_tokens ?? (u.prompt_tokens ?? 0);
+        }
+        else {
+            usage.responsesMissingUsage += 1;
+        }
+    }
     const raw = data.choices[0]?.message?.content ?? '';
+    // 推理模型(MiniMax-M2 等):剥 <think> 块,只留正文;未闭合时留全文由上层容错
     return raw.replace(/<think>[\s\S]*?<\/think>/g, '').trim() || raw;
 }
+/** 从模型输出稳健地抠出 JSON 对象(容忍围栏/前后文) */
 function parseJsonBlock(text) {
     const m = text.match(/\{[\s\S]*\}/);
-    if (!m) return null;
+    if (!m)
+        return null;
     try {
         return JSON.parse(m[0]);
-    } catch  {
+    }
+    catch {
         return null;
     }
 }
 const FACTS_SYSTEM = `你是旅行规划的事实抽取器。从对话中抽取两类事实并以 JSON 返回:
 {"calendar": {"year": 数字, "assertedWeekdays": {"YYYY-MM-DD": "mon|tue|wed|thu|fri|sat|sun"}},
- "profile": {"workWindow": {"homeTzOffsetMin": 数字, "startMin": 数字, "endMin": 数字, "workdays": [0,1,2,3,4], "evidence": "用户原话"},
+ "profile": {"workWindow": {"homeTzOffsetMin": 数字(仅 legacy v1;用户未明确时可省略), "startMin": 数字, "endMin": 数字, "workdays": [0,1,2,3,4], "evidence": "用户原话"},
               "companions": ["..."], "budgetTier": "economy|comfort|convenience",
               "bookedResources": [{"kind": "flight|hotel", "ref": "...", "window": "..."}]}}
-只放用户明确说过的事实;没有的字段省略。分钟数从 HH:MM 换算;UTC+4 → homeTzOffsetMin=240。
+只放用户明确说过的事实;没有的字段省略。分钟数从 HH:MM 换算;用户未明确说 numeric offset 时不要猜、不要输出 homeTzOffsetMin;v2 pack 的 IANA home zone 属于数据层权威。
 **休假语义(关键)**:用户说「请假/年假/不用办公/休假」→ workWindow 输出 {"vacation": true}(不是省略!省略会触发重复追问);只有用户明确给了工作时间才输出完整 workWindow 对象。只输出 JSON。`;
 const SKELETON_SYSTEM = `你是行程骨架抽取器。从对话中抽取行程的**骨架**——段(移动)与锚点,不包含任何班次数据(班次来自数据层,你不要编造时刻/价格/航班号)。
 输出 JSON:{"scenario":"erhai|workation|yunnan|generic","segments":[{"id":"f1","role":"choice|fixed","route":"HKG->HKT","dateHint":"2026-07-18","anchors":{"arriveByMin":885}}]}
 规则:每个跨城移动一段;锚点只放用户明说或必然的(如"当天到"→arriveByMin 23:59=1439);时刻用当日分钟。
 scenario 判定:「洱海/大理/千岛湖/太湖+选目的地」→erhai(候选集);「普吉/workation/远程办公+多城链」→workation(五段链);「云南/大理丽江」→yunnan;不确定→generic。只输出 JSON。`;
-export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
+/**
+ * request:本端口每次 chat() 共用的请求级选项(timeoutMs / signal)。signal 是端口级取消——
+ * 调用方可借它在会话/批次截止时真正掐断在途请求,而不是放着它继续挂在 provider 上。
+ */
+export function createOpenAICompatLlm(flightPackPath, clock = () => new Date(), request = {}) {
     const pack = flightPackPath;
-    const historyText = (h)=>h.map((t)=>`${t.role === 'user' ? '用户' : '助手'}: ${t.text}`).join('\n');
-    const anchorContext = ()=>`时间锚点卡:\n${buildTimeAnchor(clock()).card}`;
+    const usage = emptyUsage();
+    const historyText = (h) => h.map(t => `${t.role === 'user' ? '用户' : '助手'}: ${t.text}`).join('\n');
+    // 时间锚点卡:每条抽取链路都带上「今天」——legacy 路径此前无时间注入,
+    // 过期/相对日期语义全靠它(算术在 time-anchor 层,LLM 只查卡)。
+    const anchorContext = () => `时间锚点卡:\n${buildTimeAnchor(clock()).card}`;
     return {
-        async extractFacts (history) {
-            const out = await chat([
-                {
-                    role: 'system',
-                    content: FACTS_SYSTEM
-                },
-                {
-                    role: 'user',
-                    content: `${anchorContext()}\n\n${historyText(history)}`
-                }
-            ], true);
+        async extractFacts(history) {
+            const out = await chat([{ role: 'system', content: FACTS_SYSTEM }, { role: 'user', content: `${anchorContext()}\n\n${historyText(history)}` }], true, usage, request);
             const obj = parseJsonBlock(out);
-            if (!obj) return {
-                assumptions: []
-            };
+            if (!obj)
+                return { assumptions: [] };
             const calendar = obj['calendar'];
             const profile = obj['profile'];
-            const assumptions = Object.keys(profile ?? {}).map((f)=>({
-                    field: f,
-                    source: 'user-verbatim'
-                }));
-            return {
-                calendar,
-                profile,
-                assumptions
-            };
+            const assumptions = Object.keys(profile ?? {}).map(f => ({ field: f, source: 'user-verbatim' }));
+            return { calendar, profile, assumptions };
         },
-        async extractSpec (history, state) {
+        async extractSpec(history, state) {
+            // 架构(ADR-10):LLM 只产骨架与锚点;班次数据永远来自能力层(数据包/未来实时API)
             const context = `${anchorContext()}\n已断言日历:${JSON.stringify(state.calendar.assertedWeekdays)}\nprofile:${JSON.stringify(state.profile)}`;
-            const out = await chat([
-                {
-                    role: 'system',
-                    content: SKELETON_SYSTEM
-                },
-                {
-                    role: 'user',
-                    content: `${context}\n\n${historyText(history)}`
-                }
-            ], true);
+            const out = await chat([{ role: 'system', content: SKELETON_SYSTEM }, { role: 'user', content: `${context}\n\n${historyText(history)}` }], true, usage, request);
             const skeleton = parseJsonBlock(out);
-            if (!skeleton || !Array.isArray(skeleton['segments']) || skeleton['segments'].length === 0) return null;
-            if (!pack) return null;
+            if (!skeleton || !Array.isArray(skeleton['segments']) || skeleton['segments'].length === 0)
+                return null;
+            // 能力层装数据:航班包提供 services;骨架按段 id 合并锚点
+            if (!pack)
+                return null;
             const { readFile } = await import('node:fs/promises');
             const scenario = String(skeleton['scenario'] ?? 'generic');
-            if (scenario === 'generic') return null;
+            // 场景→数据包路由(薄壳段3:意图决定装哪个包,而非永远装通用包)
+            // generic 不装包——意图不明确时不进求解,让循环继续访谈(ADR-10:翻译不造数)
+            if (scenario === 'generic')
+                return null;
             const packByScenario = {
                 erhai: pack.replace('flights_2026.json', 'golden_erhai.json'),
-                workation: pack,
-                yunnan: pack.replace('flights_2026.json', 'yunnan-pack.json')
+                workation: pack, // 五段链
+                yunnan: pack.replace('flights_2026.json', 'yunnan-pack.json'),
             };
             const packPath = packByScenario[scenario];
-            if (!packPath) return null;
+            if (!packPath)
+                return null;
             let packSpec;
             try {
                 if (scenario === 'erhai') {
-                    return {
-                        segments: [],
-                        note: 'erhai-candidates',
-                        budgetCny: 3000
-                    };
+                    // 洱海 = 候选集场景:不装五段链,返回洱海候选 spec 的轻量标记(由引擎的候选求解处理;
+                    // 循环层看到 scenario=erhai 时走 solveChoiceSegment 而非 solveUnified)
+                    return { segments: [], note: 'erhai-candidates', budgetCny: 3000 };
                 }
                 packSpec = parseFlightPackToSpec(JSON.parse(await readFile(packPath, 'utf-8')));
-            } catch  {
+            }
+            catch (e) {
+                // issue #620:数据包被 parse 边界拒收时,原因必须留痕——否则用户只看到
+                // 「骨架还不完整」,运维看不到是哪个包的哪个字段把求解面挡住了。
+                const err = e;
+                console.error(`[gotry] extractSpec: 数据包不可用 ${packPath}(${err.code ?? 'parse_failed'}):`, (err.message ?? '').slice(0, 200));
                 return null;
             }
-            const anchorsById = new Map(skeleton['segments'].map((s)=>[
-                    String(s['id'] ?? ''),
-                    s['anchors'] ?? {}
-                ]));
-            for (const seg of packSpec.segments){
+            const anchorsById = new Map(skeleton['segments'].map(s => [String(s['id'] ?? ''), (s['anchors'] ?? {})]));
+            for (const seg of packSpec.segments) {
                 const a = anchorsById.get(seg.id);
-                if (a?.arriveByMin !== undefined) seg.anchors = {
-                    arriveByMin: a.arriveByMin
-                };
+                if (a?.arriveByMin !== undefined)
+                    seg.anchors = { arriveByMin: a.arriveByMin };
             }
-            packSpec.workWindow = state.profile.workWindow ? {
-                homeTzOffsetMin: state.profile.workWindow.homeTzOffsetMin,
-                startMin: state.profile.workWindow.startMin,
-                endMin: state.profile.workWindow.endMin,
-                workdays: state.profile.workWindow.workdays
-            } : undefined;
+            packSpec = mergeProfileWorkWindow(packSpec, state.profile.workWindow);
             packSpec.budgetCny = 9000;
             return packSpec;
         },
-        async extractSlots (history, now) {
+        async extractSlots(history, now) {
+            // now 可注入(评测固定锚点);过期判定在 flagExpiredSlots(代码层),模型只管逐字抽取
             const anchor = buildTimeAnchor(now ?? clock());
-            const out = await chat([
-                {
-                    role: 'system',
-                    content: buildSlotSystem(anchor)
-                },
-                {
-                    role: 'user',
-                    content: historyText(history)
-                }
-            ], true);
+            const out = await chat([{ role: 'system', content: buildSlotSystem(anchor) }, { role: 'user', content: historyText(history) }], true, usage, request);
             const obj = parseJsonBlock(out);
-            if (!obj) return null;
-            const ext = normalizeExtraction(obj, history.map((t)=>t.text).join('\n'));
+            if (!obj)
+                return null;
+            // language 判定归代码层(detectLanguage),模型输出仅供参考;过期判定同在代码层
+            const ext = normalizeExtraction(obj, history.map(t => t.text).join('\n'));
             return ext ? flagExpiredSlots(ext, anchor) : null;
         },
-        async polishQuestion (q) {
-            const out = await chat([
-                {
-                    role: 'system',
-                    content: '把旅行规划的追问润色得更自然,保留全部信息(含为什么问),一两句话,不要加表情。'
-                },
-                {
-                    role: 'user',
-                    content: `【${q.key}】${q.text}(为什么问:${q.why})`
-                }
-            ], false);
+        async polishQuestion(q) {
+            const out = await chat([{ role: 'system', content: '把旅行规划的追问润色得更自然,保留全部信息(含为什么问),一两句话,不要加表情。' },
+                { role: 'user', content: `【${q.key}】${q.text}(为什么问:${q.why})` }], false, usage, request);
             return out.trim() || `【${q.key}】${q.text}`;
-        }
+        },
+        usage,
     };
 }
-
-
-//# sourceURL=/Users/bytedance/work/gotry/ts/src/dsh-llm.ts