@danceiny/gotry 0.0.1-rc.14 → 0.0.1-rc.15

package/dist/scripts/booking-saga-tests.js

@@ -0,0 +1,187 @@
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { BOOKING_SAGA_SCHEMA, SAGA_EDGES, SAGA_REJECTIONS, SAGA_STATUSES, SAGA_TRIGGER_EVENT, resolveSagaTrigger, sagaLegalPaths, sagaTraceViolations } from '../src/booking-saga.js';
+import { ensureLedger } from '../src/state-ledger.js';
+let pass = 0;
+let fail = 0;
+function assert(cond, msg) {
+    if (cond) {
+        pass++;
+        console.log(`  ok - ${msg}`);
+    } else {
+        fail++;
+        console.error(`  FAIL - ${msg}`);
+    }
+}
+function parseReceipt(payload) {
+    try {
+        const doc = JSON.parse(payload);
+        return typeof doc.receipt === 'string' && doc.receipt !== '' ? doc.receipt : undefined;
+    } catch  {
+        return undefined;
+    }
+}
+const TRIGGERS = [
+    'propose',
+    'confirm',
+    'compensate'
+];
+const ORIGINS = [
+    'none',
+    ...SAGA_STATUSES
+];
+const COMBOS = ORIGINS.flatMap((from)=>TRIGGERS.map((trigger)=>`${from}:${trigger}`));
+const edgeKeys = new Set(SAGA_EDGES.map((e)=>`${e.from}:${e.trigger}`));
+const rejectKeys = new Set(SAGA_REJECTIONS.map((e)=>`${e.from}:${e.trigger}`));
+assert(BOOKING_SAGA_SCHEMA === 'booking_saga_fsm.v1' && SAGA_STATUSES.join(',') === 'pending,confirmed,compensated', 'schema 词唯一;状态字母表与 pending_writes CHECK 逐字一致');
+assert(COMBOS.length === 12 && new Set(COMBOS).size === 12 && edgeKeys.size === SAGA_EDGES.length && rejectKeys.size === SAGA_REJECTIONS.length, '12 个 (origin,trigger) 组合;边表与拒绝表内部无重复');
+assert(COMBOS.every((k)=>edgeKeys.has(k) !== rejectKeys.has(k)), '每格恰为一条合法边或一条结构化拒绝(全函数,无空洞无第三态)');
+assert(SAGA_EDGES.every((e)=>SAGA_STATUSES.includes(e.to) && e.event === SAGA_TRIGGER_EVENT[e.trigger] && e.guard.length > 0), '合法边:终点在字母表内、事件 kind 与触发器一一对应、守卫语义齐全');
+const allConsistent = ORIGINS.every((from)=>TRIGGERS.every((trigger)=>{
+        const v = resolveSagaTrigger(from, trigger);
+        const edge = SAGA_EDGES.find((e)=>e.from === from && e.trigger === trigger);
+        if (edge) return v.ok && v.to === edge.to && v.event === edge.event;
+        return !v.ok && v.reason === SAGA_REJECTIONS.find((x)=>x.from === from && x.trigger === trigger).reason;
+    }));
+assert(allConsistent, 'resolveSagaTrigger 十二格与边表/拒绝表全一致(解析器无平行第二逻辑)');
+{
+    const vProp = resolveSagaTrigger('none', 'propose');
+    const vConf = resolveSagaTrigger('pending', 'confirm');
+    const vComp = resolveSagaTrigger('confirmed', 'compensate');
+    assert(vProp.ok && vProp.to === 'pending' && vConf.ok && vConf.to === 'confirmed' && vComp.ok && vComp.to === 'compensated', '主路径 ∅→pending→confirmed→compensated 连通(L2 登记只不执行 / L3 确认携 receipt / saga 补偿)');
+    const vDouble = resolveSagaTrigger('confirmed', 'confirm');
+    assert(!vDouble.ok && vDouble.reason === 'already-confirmed', '二次确认结构化拒绝(already-confirmed)');
+    const vIdem = resolveSagaTrigger('compensated', 'propose');
+    const vRevive = resolveSagaTrigger('compensated', 'confirm');
+    const vReComp = resolveSagaTrigger('compensated', 'compensate');
+    assert(!vIdem.ok && vIdem.reason === 'idem-exists' && !vRevive.ok && vRevive.reason === 'absorbed-compensated' && !vReComp.ok && vReComp.reason === 'already-compensated', 'compensated 吸收态:propose=幂等存在,confirm/compensate 拒绝(无复活路径)');
+    assert(sagaLegalPaths().length === 4, '合法路径恰四条(主路径及其真前缀)');
+    const legalClean = sagaLegalPaths().every((path)=>{
+        const events = path.map((trigger, i)=>({
+                seq: i + 1,
+                kind: SAGA_TRIGGER_EVENT[trigger],
+                receipt: trigger === 'confirm' ? 'PNR-X' : undefined
+            }));
+        return sagaTraceViolations(events).length === 0;
+    });
+    assert(legalClean, '四条合法路径的审计链均零违例');
+    const flagOf = (events)=>sagaTraceViolations(events)[0];
+    assert(flagOf([
+        {
+            kind: 'write.confirmed',
+            receipt: 'PNR-A'
+        }
+    ]) !== undefined, '缺 pending 起点的 confirm 被审计链抓出');
+    assert(flagOf([
+        {
+            kind: 'write.pending'
+        },
+        {
+            kind: 'write.confirmed',
+            receipt: 'PNR-A'
+        },
+        {
+            kind: 'write.confirmed',
+            receipt: 'PNR-B'
+        }
+    ]) !== undefined, '双确认被审计链抓出(每键至多确认一次)');
+    assert(flagOf([
+        {
+            kind: 'write.pending'
+        },
+        {
+            kind: 'write.compensated'
+        },
+        {
+            kind: 'write.confirmed',
+            receipt: 'X'
+        }
+    ]) !== undefined, 'compensated 之后不得再有 saga 事件(吸收态)');
+    assert((flagOf([
+        {
+            kind: 'write.pending'
+        },
+        {
+            kind: 'write.confirmed'
+        }
+    ]) ?? '').includes('receipt'), 'write.confirmed 空 receipt 被抓(L3 必携外部回执)');
+}const root = mkdtempSync(join(tmpdir(), 'gotry-booking-saga-'));
+try {
+    const ledger = ensureLedger(root);
+    const sagaEventsOf = (l, key)=>{
+        const rows = l.db.prepare(`SELECT seq, kind, payload FROM events WHERE tenant_id = ? AND subject_id = ? AND kind IN ('write.pending','write.confirmed','write.compensated') ORDER BY seq`).all(l.tenant, key);
+        return rows.map((r)=>({
+                seq: r.seq,
+                kind: r.kind,
+                receipt: parseReceipt(r.payload)
+            }));
+    };
+    const K1 = 'booking:K1';
+    const proposed = ledger.requestPendingWrite({
+        idemKey: K1,
+        seam: 'flight-order-confirm',
+        payload: {
+            flight: 'MU123'
+        }
+    });
+    assert(proposed.created === true && proposed.status === 'pending' && resolveSagaTrigger('none', 'propose').to === 'pending', '物理提议 ∅→pending 与状态机同归宿(L2 只登记不执行)');
+    const dup = ledger.requestPendingWrite({
+        idemKey: K1,
+        seam: 'flight-order-confirm',
+        payload: {
+            flight: 'MU123'
+        }
+    });
+    assert(dup.created === false && dup.status === 'pending' && resolveSagaTrigger('pending', 'propose').reason === 'idem-exists', '同幂等键重复提议:账本 no-op = 状态机 idem-exists');
+    const confirmed = ledger.confirmPendingWrite(K1, 'PNR-A');
+    assert(confirmed.ok === true && confirmed.status === 'confirmed' && resolveSagaTrigger('pending', 'confirm').to === 'confirmed', 'L3 具名确认 pending→confirmed(携 receipt)');
+    const reConfirmed = ledger.confirmPendingWrite(K1, 'PNR-SHOULD-IGNORE');
+    assert(reConfirmed.ok === false && (ledger.listPendingWrites().find((w)=>w.idem_key === K1)?.receipt ?? null) === 'PNR-A' && resolveSagaTrigger('confirmed', 'confirm').reason === 'already-confirmed', '二次确认被账本守卫拒绝且 receipt 不可变(仍为 PNR-A)');
+    const compensated = ledger.compensatePendingWrite(K1, '用户改签退款');
+    assert(compensated.ok === true && ledger.listPendingWrites().find((w)=>w.idem_key === K1)?.status === 'compensated', '已确认副作用可 saga 补偿(pending|confirmed → compensated)');
+    const afterConfirm = ledger.confirmPendingWrite(K1, 'PNR-POST');
+    const afterCompensate = ledger.compensatePendingWrite(K1, '重复补偿');
+    const afterPropose = ledger.requestPendingWrite({
+        idemKey: K1,
+        seam: 'flight-order-confirm',
+        payload: {}
+    });
+    assert(afterConfirm.ok === false && afterCompensate.ok === false && afterPropose.created === false, '吸收态:compensated 后 confirm/compensate/propose 全拒,零复活');
+    const audit = sagaEventsOf(ledger, K1);
+    assert(audit.map((e)=>e.kind).join(',') === 'write.pending,write.confirmed,write.compensated', '审计事件链与边表路径同构(被拒调用零事件)');
+    assert(sagaTraceViolations(audit).length === 0, '真实账本审计链过词汇层校验零违例');
+    const K3 = 'booking:K3';
+    ledger.requestPendingWrite({
+        idemKey: K3,
+        seam: 'hotel-order-confirm',
+        payload: {}
+    });
+    const emptyReceipt = ledger.confirmPendingWrite(K3, '');
+    assert(emptyReceipt.ok === true && sagaTraceViolations(sagaEventsOf(ledger, K3)).some((v)=>v.includes('receipt')), '已知边界:账本物理层暂不挡空 receipt,词汇层审计链抓出(物理 CHECK 随 M5 拍板入 schema)');
+    const other = ensureLedger(root, 'tenant-b');
+    const crossProp = other.requestPendingWrite({
+        idemKey: K1,
+        seam: 'flight-order-confirm',
+        payload: {}
+    });
+    assert(crossProp.created === true, '同 idem_key 跨租户互不可见:tenant-b 可建自己的 K1(tenant_id 一等字段,ADR-16)');
+    const LOCAL_ONLY = 'booking:local-only';
+    ledger.requestPendingWrite({
+        idemKey: LOCAL_ONLY,
+        seam: 'flight-order-confirm',
+        payload: {}
+    });
+    const crossMissing = other.confirmPendingWrite(LOCAL_ONLY, 'PNR-CROSS');
+    assert(crossMissing.ok === false && crossMissing.status === 'missing', 'tenant-b 看不到 local 的 saga 键(拒绝 missing)');
+} finally{
+    rmSync(root, {
+        recursive: true,
+        force: true
+    });
+}
+console.log(`\nBOOKING SAGA TESTS: ${pass} ok, ${fail} fail`);
+if (fail > 0) process.exit(1);
+
+
+//# sourceURL=/Users/bytedance/work/gotry/ts/scripts/booking-saga-tests.ts

package/dist/scripts/bootstrap-tests.js

@@ -35,7 +35,8 @@ assert.equal(c1.code, 0, `--check-only 应 exit 0,实际 ${c1.code}\n${c1.out}`)
 assert.ok(c1.out.includes('hbcli'), '报告应含 hbcli 节');
 assert.ok(c1.out.includes('agent-reach'), '报告应含 agent-reach 节');
 assert.ok(c1.out.includes('flyai'), '报告应含 flyai(无需安装)节');
-console.log('1. --check-only 探测报告 exit 0(hbcli/agent-reach/flyai 三节齐)OK');
+assert.ok(c1.out.includes('dsh-better-sidebar'), '报告应含 dsh-better-sidebar 节(issue #25 产物查看面)');
+console.log('1. --check-only 探测报告 exit 0(hbcli/agent-reach/flyai/dsh-better-sidebar 四节齐)OK');
 const c2 = runBootstrap([
     '--auto'
 ], {
@@ -50,7 +51,16 @@ const c3 = runBootstrap([], {
 assert.equal(c3.code, 0, '显式模式跳过态应 exit 0');
 assert.ok(c3.out.includes('跳过'), '应输出跳过说明');
 console.log('3. 显式模式 + GOTRY_SETUP_SKIP=1 → exit 0 OK');
-console.log('BOOTSTRAP TESTS: 3/3 OK(探测报告/跳过开关/postinstall 非致命)');
+const c4 = runBootstrap([
+    '--check-only'
+], {
+    GOTRY_SETUP_SIDEBAR: '0'
+});
+assert.equal(c4.code, 0, '单项跳过态应 exit 0');
+assert.ok(c4.out.includes('GOTRY_SETUP_SIDEBAR=0 跳过'), '应输出侧栏单项跳过说明');
+assert.ok(c4.out.includes('hbcli'), '其余节不受单项开关影响');
+console.log('4. GOTRY_SETUP_SIDEBAR=0 单项跳过(其余节照常)OK');
+console.log('BOOTSTRAP TESTS: 4/4 OK(探测报告/跳过开关/单项开关/postinstall 非致命)');
 
 
 //# sourceURL=/Users/bytedance/work/gotry/ts/scripts/bootstrap-tests.ts

package/dist/scripts/nightly-evidence-tests.js

@@ -0,0 +1,123 @@
+import assert from 'node:assert/strict';
+import { existsSync, mkdtempSync, rmSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
+import { buildNightlyRunRecord, loadPriceTable, loadPromptSet, priceRunCost, runNightlyEvidence } from './nightly-evidence.js';
+import { parseNightlyRun } from './product-metrics.js';
+let passed = 0;
+function ok(label) {
+    passed += 1;
+    console.log(`  ok ${label}`);
+}
+const priceTable = loadPriceTable('data/llm-price-table.json');
+const promptSet = loadPromptSet('data/m3-nightly-prompts.json');
+assert.equal(priceTable.schema_version, 'gotry_llm_price_table_v1');
+assert.equal(promptSet.schema_version, 'gotry_m3_nightly_prompt_set_v1');
+assert.ok(promptSet.turns.length >= 1);
+ok('封存数据文件 schema 与轮次');
+for (const [name, entry] of Object.entries(priceTable.prices)){
+    assert.ok(entry.input_cache_hit_usd_per_1m_peak <= entry.input_cache_miss_usd_per_1m_peak, `${name} hit<=miss`);
+    assert.ok(entry.output_usd_per_1m_offpeak <= entry.output_usd_per_1m_peak, `${name} offpeak<=peak`);
+}
+ok('价目保守上界不变量(hit<=miss, offpeak<=peak)');
+const baseUsage = {
+    calls: 3,
+    inputTokens: 1_000_000,
+    outputTokens: 1_000_000,
+    inputCacheHitTokens: 0,
+    inputCacheMissTokens: 1_000_000,
+    responsesMissingUsage: 0
+};
+const cost = priceRunCost('deepseek-chat', baseUsage, priceTable);
+assert.equal(cost, 1.76);
+ok('peak miss 换算 0.44+1.32=1.76(经 alias)');
+const hitUsage = {
+    ...baseUsage,
+    inputCacheHitTokens: 1_000_000,
+    inputCacheMissTokens: 0
+};
+assert.equal(priceRunCost('deepseek-chat', hitUsage, priceTable), 1.334);
+ok('peak hit 换算 0.014+1.32=1.334');
+assert.throws(()=>priceRunCost('unknown-model-xyz', baseUsage, priceTable), /no entry for model/);
+const dangling = {
+    ...priceTable,
+    aliases: {
+        'deepseek-chat': 'not-a-model'
+    }
+};
+assert.throws(()=>priceRunCost('deepseek-chat', baseUsage, dangling), /has no entry/);
+assert.throws(()=>priceRunCost('deepseek-chat', {
+        ...baseUsage,
+        responsesMissingUsage: 1
+    }, priceTable), /cost unprovable/);
+ok('未知模型/悬空alias/usage缺失 fail-closed');
+const executeAt = '2026-08-29T12:00:00.000Z';
+const recordInput = {
+    schema_version: 'gotry_m3_nightly_run_v1',
+    executed_at: executeAt,
+    real_llm: true,
+    prompt_set_sha256: 'a'.repeat(64),
+    output_sha256: 'b'.repeat(64),
+    cost_usd: 0.25
+};
+const rec1 = buildNightlyRunRecord(recordInput);
+const rec2 = buildNightlyRunRecord({
+    ...recordInput
+});
+assert.equal(rec1.run_key, rec2.run_key);
+assert.match(rec1.run_key, /^hmac-sha256:[0-9a-f]{64}$/);
+parseNightlyRun(JSON.parse(JSON.stringify(rec1)), 0);
+const rec3 = buildNightlyRunRecord({
+    ...recordInput,
+    cost_usd: 0.26
+});
+assert.notEqual(rec3.run_key, rec1.run_key);
+ok('run_key 确定性 + 已过消费方 parseNightlyRun');
+const prevKey = process.env['LLM_API_KEY'];
+const prevDeepseekKey = process.env['DEEPSEEK_API_KEY'];
+const waitParent = mkdtempSync(join(tmpdir(), 'nightly-wait-'));
+const waitRoot = join(waitParent, 'evidence', 'm3');
+try {
+    delete process.env['LLM_API_KEY'];
+    delete process.env['DEEPSEEK_API_KEY'];
+    const waiting = await runNightlyEvidence({
+        evidenceRoot: waitRoot,
+        dryRun: false,
+        clock: ()=>new Date(executeAt)
+    });
+    assert.equal(waiting.state, 'waiting_external_evidence');
+    assert.equal(waiting.record, null);
+    assert.ok(waiting.reason?.includes('backoff'));
+    assert.equal(existsSync(waitRoot), false, 'waiting 不得创建任何证据目录');
+} finally{
+    if (prevKey !== undefined) process.env['LLM_API_KEY'] = prevKey;
+    if (prevDeepseekKey !== undefined) process.env['DEEPSEEK_API_KEY'] = prevDeepseekKey;
+    rmSync(waitParent, {
+        recursive: true,
+        force: true
+    });
+}
+ok('无凭证 → waiting_external_evidence 且零写入');
+const dryParent = mkdtempSync(join(tmpdir(), 'nightly-dry-'));
+const dryRoot = join(dryParent, 'evidence', 'm3');
+try {
+    const dry = await runNightlyEvidence({
+        evidenceRoot: dryRoot,
+        dryRun: true,
+        clock: ()=>new Date(executeAt)
+    });
+    assert.equal(dry.state, 'dry_run');
+    assert.equal(dry.record?.real_llm, false);
+    parseNightlyRun(JSON.parse(JSON.stringify(dry.record)), 0);
+    assert.equal(existsSync(dryRoot), false, 'dry-run 不得写任何证据文件');
+} finally{
+    rmSync(dryParent, {
+        recursive: true,
+        force: true
+    });
+}
+ok('dry-run mock 全链演练 + real_llm=false + 零落盘');
+console.log(`nightly-evidence tests: ${passed} 组断言全绿(offline,真跑不在 CI)`);
+
+
+//# sourceURL=/Users/bytedance/work/gotry/ts/scripts/nightly-evidence-tests.ts

package/dist/scripts/nightly-evidence.js

@@ -0,0 +1,233 @@
+import { createHash } from 'node:crypto';
+import { appendFileSync, mkdirSync, readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { createMockLlm } from '../src/mock-llm.js';
+import { createOpenAICompatLlm } from '../src/dsh-llm.js';
+import { newState, runTurn } from '../src/loop.js';
+import { solveUnified } from '../src/unified.js';
+import { realtimeSolvePort } from '../src/realtime-pricing.js';
+import { parseNightlyRun } from './product-metrics.js';
+const SHA256 = /^[0-9a-f]{64}$/;
+const DEFAULT_BUDGET_USD = 1;
+function canonical(value) {
+    if (Array.isArray(value)) return `[${value.map(canonical).join(',')}]`;
+    if (typeof value === 'object' && value !== null) {
+        const record = value;
+        return `{${Object.keys(record).sort().map((key)=>`${JSON.stringify(key)}:${canonical(record[key])}`).join(',')}}`;
+    }
+    return JSON.stringify(value);
+}
+function sha256Of(value) {
+    return createHash('sha256').update(canonical(value)).digest('hex');
+}
+function round6(value) {
+    return Number(value.toFixed(6));
+}
+export function loadPromptSet(path) {
+    const raw = JSON.parse(readFileSync(path, 'utf8'));
+    if (raw['schema_version'] !== 'gotry_m3_nightly_prompt_set_v1') {
+        throw new Error(`prompt set schema_version must be gotry_m3_nightly_prompt_set_v1, got: ${String(raw['schema_version'])}`);
+    }
+    const turns = raw['turns'];
+    if (!Array.isArray(turns) || turns.length === 0 || turns.some((t)=>typeof t !== 'string' || t.length === 0)) {
+        throw new Error('prompt set turns must be a non-empty array of non-empty strings');
+    }
+    return {
+        schema_version: 'gotry_m3_nightly_prompt_set_v1',
+        description: raw['description'],
+        turns: turns
+    };
+}
+function assertPriceEntry(label, value) {
+    const entry = typeof value === 'object' && value !== null ? value : null;
+    if (!entry) throw new Error(`${label} must be an object`);
+    for (const field of [
+        'input_cache_miss_usd_per_1m_peak',
+        'input_cache_hit_usd_per_1m_peak',
+        'output_usd_per_1m_peak'
+    ]){
+        const num = entry[field];
+        if (typeof num !== 'number' || !Number.isFinite(num) || num < 0) {
+            throw new Error(`${label}.${field} must be a non-negative finite number`);
+        }
+    }
+    return value;
+}
+export function loadPriceTable(path) {
+    const raw = JSON.parse(readFileSync(path, 'utf8'));
+    if (raw['schema_version'] !== 'gotry_llm_price_table_v1') {
+        throw new Error(`price table schema_version must be gotry_llm_price_table_v1, got: ${String(raw['schema_version'])}`);
+    }
+    const aliases = typeof raw['aliases'] === 'object' && raw['aliases'] !== null ? raw['aliases'] : {};
+    for (const [from, to] of Object.entries(aliases)){
+        if (typeof to !== 'string') throw new Error(`price table alias ${from} must map to a model name`);
+    }
+    const prices = {};
+    for (const [name, entry] of Object.entries(typeof raw['prices'] === 'object' && raw['prices'] !== null ? raw['prices'] : {})){
+        prices[name] = assertPriceEntry(`price table entry ${name}`, entry);
+    }
+    return {
+        schema_version: 'gotry_llm_price_table_v1',
+        aliases,
+        prices
+    };
+}
+export function priceRunCost(model, usage, table) {
+    if (usage.responsesMissingUsage > 0) {
+        throw new Error(`usage provider omitted usage on ${usage.responsesMissingUsage} response(s) — cost unprovable, refusing to price (fail-closed)`);
+    }
+    const alias = table.aliases[model];
+    if (alias && !table.prices[alias]) throw new Error(`price table alias ${model} -> ${alias} has no entry; update data/llm-price-table.json via PR`);
+    const entry = table.prices[alias ?? model];
+    if (!entry) throw new Error(`price table has no entry for model ${model}; update data/llm-price-table.json via PR (fail-closed, no guessed price)`);
+    const cost = (usage.inputCacheMissTokens * entry.input_cache_miss_usd_per_1m_peak + usage.inputCacheHitTokens * entry.input_cache_hit_usd_per_1m_peak + usage.outputTokens * entry.output_usd_per_1m_peak) / 1_000_000;
+    return round6(cost);
+}
+export function buildNightlyRunRecord(input) {
+    const { schema_version, executed_at, real_llm, prompt_set_sha256, output_sha256, cost_usd } = input;
+    const core = {
+        schema_version,
+        executed_at,
+        real_llm,
+        prompt_set_sha256,
+        output_sha256,
+        cost_usd
+    };
+    if (!SHA256.test(prompt_set_sha256) || !SHA256.test(output_sha256)) {
+        throw new Error('nightly record hashes must be lowercase SHA-256 digests');
+    }
+    const record = {
+        ...core,
+        run_key: `hmac-sha256:${createHash('sha256').update(canonical(core)).digest('hex')}`
+    };
+    parseNightlyRun(JSON.parse(JSON.stringify(record)), 0);
+    return record;
+}
+function activeModel() {
+    return process.env['LLM_MODEL'] ?? process.env['DEEPSEEK_MODEL'] ?? 'MiniMax-M2';
+}
+function arg(name) {
+    const index = process.argv.indexOf(name);
+    return index >= 0 ? process.argv[index + 1] : undefined;
+}
+export async function runNightlyEvidence(options) {
+    const clock = options.clock ?? (()=>new Date());
+    const llmKey = process.env['LLM_API_KEY'] ?? process.env['DEEPSEEK_API_KEY'];
+    if (!options.dryRun && !llmKey) {
+        return {
+            state: 'waiting_external_evidence',
+            record: null,
+            reason: 'no real LLM credential (LLM_API_KEY/DEEPSEEK_API_KEY); waiting & backoff & no-spend per issue #22 停机纪律'
+        };
+    }
+    const promptSet = loadPromptSet('data/m3-nightly-prompts.json');
+    const priceTable = loadPriceTable('data/llm-price-table.json');
+    const packPath = join('..', 'data', 'flights_2026.json');
+    const solvePort = realtimeSolvePort(solveUnified);
+    const port = options.dryRun ? createMockLlm(packPath) : createOpenAICompatLlm(packPath);
+    const state = newState();
+    const history = [];
+    for (const turn of promptSet.turns){
+        const { reply } = await runTurn(state, turn, port, [
+            ...history
+        ], solvePort);
+        history.push({
+            role: 'user',
+            text: turn
+        }, {
+            role: 'assistant',
+            text: reply
+        });
+    }
+    const promptSetSha = sha256Of(promptSet);
+    const turnsTranscript = promptSet.turns.map((turn, index)=>({
+            turn_index: index,
+            user: turn,
+            reply: history[index * 2 + 1]?.text ?? ''
+        }));
+    const outputSha = sha256Of(turnsTranscript);
+    let record;
+    if (options.dryRun) {
+        record = buildNightlyRunRecord({
+            schema_version: 'gotry_m3_nightly_run_v1',
+            executed_at: clock().toISOString(),
+            real_llm: false,
+            prompt_set_sha256: promptSetSha,
+            output_sha256: outputSha,
+            cost_usd: 0
+        });
+        return {
+            state: 'dry_run',
+            record,
+            reason: 'dry-run exercises the pipeline against mock LLM; no evidence written'
+        };
+    }
+    const usage = port.usage;
+    if (!usage) throw new Error('real LLM port did not expose usage tracker (dsh-llm regression) — refusing to write cost evidence');
+    if (usage.calls !== promptSet.turns.length) {
+        throw new Error(`usage calls ${usage.calls} != expected turns ${promptSet.turns.length} — fail-closed, no evidence written`);
+    }
+    const costUsd = priceRunCost(activeModel(), usage, priceTable);
+    record = buildNightlyRunRecord({
+        schema_version: 'gotry_m3_nightly_run_v1',
+        executed_at: clock().toISOString(),
+        real_llm: true,
+        prompt_set_sha256: promptSetSha,
+        output_sha256: outputSha,
+        cost_usd: costUsd
+    });
+    mkdirSync(options.evidenceRoot, {
+        recursive: true
+    });
+    appendFileSync(join(options.evidenceRoot, 'cohort.jsonl'), `${JSON.stringify(record)}\n`, 'utf8');
+    const budget = Number(process.env['GOTRY_NIGHTLY_BUDGET_USD'] ?? DEFAULT_BUDGET_USD);
+    const overBudget = Number.isFinite(budget) && budget > 0 ? costUsd > budget : false;
+    return {
+        state: 'evidence_written',
+        record,
+        cost_over_budget: overBudget,
+        reason: overBudget ? `cost_usd ${costUsd} exceeds GOTRY_NIGHTLY_BUDGET_USD ${budget} — backoff suggested` : undefined
+    };
+}
+function output(payload, asJson) {
+    if (asJson) console.log(JSON.stringify(payload));
+    else {
+        console.log(`nightly-evidence state: ${payload.state}${payload.error ? ` — ${payload.error}` : ''}`);
+        if (payload.record) console.log(JSON.stringify(payload.record, null, 2));
+        if (payload.reason) console.log(payload.reason);
+    }
+}
+async function main() {
+    if (!process.argv.includes('--no-env-file')) {
+        try {
+            for (const line of readFileSync(join('..', '.env'), 'utf-8').split('\n')){
+                const m = line.match(/^([A-Z_]+)=(.*)$/);
+                if (m && !process.env[m[1]]) process.env[m[1]] = m[2].trim();
+            }
+        } catch  {}
+    }
+    const asJson = (arg('--format') ?? 'markdown') === 'json';
+    const evidenceRoot = arg('--evidence-root') ?? 'gotry-state/evidence/m3';
+    const dryRun = process.argv.includes('--dry-run');
+    try {
+        const result = await runNightlyEvidence({
+            evidenceRoot,
+            dryRun
+        });
+        output(result, asJson);
+        if (result.cost_over_budget) process.exitCode = 3;
+    } catch (error) {
+        output({
+            state: 'error',
+            record: null,
+            error: error.message
+        }, asJson);
+        process.exitCode = 1;
+    }
+}
+if (process.argv[1]?.endsWith('nightly-evidence.ts')) {
+    await main();
+}
+
+
+//# sourceURL=/Users/bytedance/work/gotry/ts/scripts/nightly-evidence.ts

package/dist/scripts/smoke.js

@@ -410,6 +410,89 @@ async function main() {
         }
         console.log('login tool: registered; 登录=外部网站+用户自己的浏览器,gotry 只读票据名(0 值过手)语义钉死');
     }
+    {
+        const { writeFileSync } = await import('node:fs');
+        const { listArtifacts, readArtifact } = await import('../capabilities/artifacts.js');
+        const { ensureLedger } = await import('../src/state-ledger.js');
+        const cwdDir = mkdtempSync(join(tmpdir(), 'gotry-artifacts-cwd-'));
+        const twelve = Array.from({
+            length: 12
+        }, (_, i)=>`第 ${i + 1} 行`).join('\n');
+        writeFileSync(join(cwdDir, 'trip-2027-probe.md'), `# 行程·产物探针\n${twelve}\n`);
+        const ledger = ensureLedger(smokeRoot);
+        ledger.createWorkflowRun({
+            id: 'art-probe-1',
+            goal: '行程·产物探针(账本)',
+            ticket: {
+                objective: 'probe'
+            },
+            state: {}
+        });
+        ledger.settleWorkflowRun('art-probe-1', '# 交付·账本权威\nD1 大理\nD2 洱海');
+        const listTool = byName('gotry_artifacts_list');
+        const ledList = await listTool.execute({
+            query: {}
+        }, null);
+        const run = ledList.artifacts?.find((a)=>a.id === 'art-probe-1');
+        if (!ledList.ok || !run || run.source !== 'async-run' || run.status !== 'settled') {
+            throw new Error(`FAIL: artifacts list 应发现账本已交付工单,实际:${JSON.stringify(ledList).slice(0, 200)}`);
+        }
+        const direct = await listArtifacts({
+            stateRoot: smokeRoot,
+            cwd: cwdDir
+        });
+        if (!direct.artifacts.some((a)=>a.source === 'cwd-file' && a.id === 'trip-2027-probe.md')) {
+            throw new Error(`FAIL: artifacts list 应发现 cwd 顶层 md,实际:${JSON.stringify(direct.artifacts.map((a)=>a.id))}`);
+        }
+        const readTool = byName('gotry_artifacts_read');
+        const r1 = await readTool.execute({
+            query: {
+                path: 'art-probe-1'
+            }
+        }, null);
+        if (!r1.ok || r1.lines?.[0]?.number !== 1 || !r1.lines?.[0]?.text.includes('交付·账本权威') || r1.lang !== 'markdown') {
+            throw new Error(`FAIL: 裸工单 id 应从账本读出行号视图,实际:${JSON.stringify(r1).slice(0, 200)}`);
+        }
+        const view = readTool.presentResult?.({
+            query: {
+                path: 'art-probe-1'
+            }
+        }, r1);
+        if (view?.card !== 'read' || view.path !== r1.path || view.offset !== 1 || view.lines?.length !== r1.lines?.length || view.totalLines !== r1.totalLines || view.lang !== 'markdown') {
+            throw new Error(`FAIL: read 卡字段不齐,实际:${JSON.stringify(view).slice(0, 200)}`);
+        }
+        const r2 = await readArtifact({
+            stateRoot: smokeRoot,
+            cwd: cwdDir,
+            path: 'trip-2027-probe.md',
+            offset: 10,
+            limit: 5
+        });
+        if (!r2.ok || r2.lines[0].number !== 10 || r2.windowed !== false) {
+            throw new Error(`FAIL: 行窗口 12 行文件 offset=10 应从行号 10 起,实际:${JSON.stringify(r2.ok ? {
+                o: r2.offset,
+                first: r2.lines[0]?.number
+            } : r2)}`);
+        }
+        if (r2.lines.length !== 5) throw new Error(`FAIL: 14 行文件 offset=10/limit=5 窗口应恰 5 行(10-14),实际 ${r2.lines.length}`);
+        const badPath = await readArtifact({
+            stateRoot: smokeRoot,
+            cwd: cwdDir,
+            path: '../../../../../etc/passwd'
+        });
+        if (badPath.ok) throw new Error('FAIL: 越界路径必须被拒');
+        const badExt = await readArtifact({
+            stateRoot: smokeRoot,
+            cwd: cwdDir,
+            path: 'gotry-state/gotry-state.db'
+        });
+        if (badExt.ok) throw new Error('FAIL: 白名单外扩展名(.db)必须被拒');
+        rmSync(cwdDir, {
+            recursive: true,
+            force: true
+        });
+        console.log(`artifacts: ledger run + cwd md discovered; read window(${r2.ok ? r2.lines.length : '?'} lines @10) + read card; path/ext guardrails hold`);
+    }
     rmSync(smokeRoot, {
         recursive: true,
         force: true