@danceiny/gotry 0.0.1-rc.13 → 0.0.1-rc.15

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

package/dist/src/booking-saga.js

@@ -0,0 +1,153 @@
+export const BOOKING_SAGA_SCHEMA = 'booking_saga_fsm.v1';
+export const SAGA_STATUSES = [
+    'pending',
+    'confirmed',
+    'compensated'
+];
+export const SAGA_EDGES = [
+    {
+        from: 'none',
+        trigger: 'propose',
+        to: 'pending',
+        event: 'write.pending',
+        guard: 'L2 只登记不执行;idem_key UNIQUE(重复提议 no-op)'
+    },
+    {
+        from: 'pending',
+        trigger: 'confirm',
+        to: 'confirmed',
+        event: 'write.confirmed',
+        guard: 'L3 具名 seam 确认,必携 receipt;已确认不可再确认'
+    },
+    {
+        from: 'pending',
+        trigger: 'compensate',
+        to: 'compensated',
+        event: 'write.compensated',
+        guard: '外部写未发生,取消即终态(无补偿动作)'
+    },
+    {
+        from: 'confirmed',
+        trigger: 'compensate',
+        to: 'compensated',
+        event: 'write.compensated',
+        guard: '已发生副作用的 saga 补偿(退改),receipt 保留(COALESCE)'
+    }
+];
+export const SAGA_REJECTIONS = [
+    {
+        from: 'none',
+        trigger: 'confirm',
+        reason: 'missing-subject'
+    },
+    {
+        from: 'none',
+        trigger: 'compensate',
+        reason: 'missing-subject'
+    },
+    {
+        from: 'pending',
+        trigger: 'propose',
+        reason: 'idem-exists'
+    },
+    {
+        from: 'confirmed',
+        trigger: 'propose',
+        reason: 'idem-exists'
+    },
+    {
+        from: 'confirmed',
+        trigger: 'confirm',
+        reason: 'already-confirmed'
+    },
+    {
+        from: 'compensated',
+        trigger: 'propose',
+        reason: 'idem-exists'
+    },
+    {
+        from: 'compensated',
+        trigger: 'confirm',
+        reason: 'absorbed-compensated'
+    },
+    {
+        from: 'compensated',
+        trigger: 'compensate',
+        reason: 'already-compensated'
+    }
+];
+export function resolveSagaTrigger(origin, trigger) {
+    const edge = SAGA_EDGES.find((e)=>e.from === origin && e.trigger === trigger);
+    if (edge) return {
+        ok: true,
+        from: origin,
+        trigger,
+        to: edge.to,
+        event: edge.event,
+        guard: edge.guard,
+        reason: null
+    };
+    const rejection = SAGA_REJECTIONS.find((e)=>e.from === origin && e.trigger === trigger);
+    return {
+        ok: false,
+        from: origin,
+        trigger,
+        to: null,
+        event: null,
+        guard: null,
+        reason: rejection.reason
+    };
+}
+export const SAGA_TRIGGER_EVENT = {
+    propose: 'write.pending',
+    confirm: 'write.confirmed',
+    compensate: 'write.compensated'
+};
+export function triggerOfEventKind(kind) {
+    if (kind === 'write.pending') return 'propose';
+    if (kind === 'write.confirmed') return 'confirm';
+    if (kind === 'write.compensated') return 'compensate';
+    return null;
+}
+export function sagaTraceViolations(events) {
+    const violations = [];
+    let stage = 'none';
+    for (const ev of events){
+        const trigger = triggerOfEventKind(ev.kind);
+        if (!trigger) continue;
+        const verdict = resolveSagaTrigger(stage, trigger);
+        const at = `seq=${ev.seq ?? '?'} ${ev.kind}`;
+        if (!verdict.ok) {
+            violations.push(`${at}: 非法(${stage} --${trigger}--> 拒绝 ${verdict.reason})`);
+            continue;
+        }
+        stage = verdict.to;
+        if (trigger === 'confirm' && (ev.receipt === undefined || ev.receipt === null || ev.receipt === '')) {
+            violations.push(`${at}: receipt 为空(L3 具名确认必携外部回执)`);
+        }
+    }
+    return violations;
+}
+export function sagaLegalPaths() {
+    return [
+        [
+            'propose'
+        ],
+        [
+            'propose',
+            'compensate'
+        ],
+        [
+            'propose',
+            'confirm'
+        ],
+        [
+            'propose',
+            'confirm',
+            'compensate'
+        ]
+    ];
+}
+
+
+//# sourceURL=/Users/bytedance/work/gotry/ts/src/booking-saga.ts

package/dist/src/dsh-llm.js

@@ -3,7 +3,17 @@ 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(/\/$/, '');
-async function chat(messages, json) {
+function emptyUsage() {
+    return {
+        calls: 0,
+        inputTokens: 0,
+        outputTokens: 0,
+        inputCacheHitTokens: 0,
+        inputCacheMissTokens: 0,
+        responsesMissingUsage: 0
+    };
+}
+async function chat(messages, json, usage) {
     const key = process.env['LLM_API_KEY'] ?? process.env['DEEPSEEK_API_KEY'];
     if (!key) throw new Error('LLM_API_KEY 未设置(兼容 DEEPSEEK_API_KEY 别名)——真 LLM 路径不可用,请回退 mock(ADR-8)');
     const res = await fetch(`${base()}/chat/completions`, {
@@ -25,6 +35,18 @@ async function chat(messages, json) {
     });
     if (!res.ok) throw new Error(`llm ${res.status}: ${(await res.text()).slice(0, 300)}`);
     const data = await res.json();
+    if (usage) {
+        usage.calls += 1;
+        const u = data.usage;
+        if (u && Number.isFinite(u.prompt_tokens) && Number.isFinite(u.completion_tokens)) {
+            usage.inputTokens += u.prompt_tokens ?? 0;
+            usage.outputTokens += u.completion_tokens ?? 0;
+            usage.inputCacheHitTokens += u.prompt_cache_hit_tokens ?? 0;
+            usage.inputCacheMissTokens += u.prompt_cache_miss_tokens ?? u.prompt_tokens ?? 0;
+        } else {
+            usage.responsesMissingUsage += 1;
+        }
+    }
     const raw = data.choices[0]?.message?.content ?? '';
     return raw.replace(/<think>[\s\S]*?<\/think>/g, '').trim() || raw;
 }
@@ -50,6 +72,7 @@ const SKELETON_SYSTEM = `你是行程骨架抽取器。从对话中抽取行程
 scenario 判定:「洱海/大理/千岛湖/太湖+选目的地」→erhai(候选集);「普吉/workation/远程办公+多城链」→workation(五段链);「云南/大理丽江」→yunnan;不确定→generic。只输出 JSON。`;
 export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
     const pack = flightPackPath;
+    const usage = emptyUsage();
     const historyText = (h)=>h.map((t)=>`${t.role === 'user' ? '用户' : '助手'}: ${t.text}`).join('\n');
     const anchorContext = ()=>`时间锚点卡:\n${buildTimeAnchor(clock()).card}`;
     return {
@@ -63,7 +86,7 @@ export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
                     role: 'user',
                     content: `${anchorContext()}\n\n${historyText(history)}`
                 }
-            ], true);
+            ], true, usage);
             const obj = parseJsonBlock(out);
             if (!obj) return {
                 assumptions: []
@@ -91,7 +114,7 @@ export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
                     role: 'user',
                     content: `${context}\n\n${historyText(history)}`
                 }
-            ], true);
+            ], true, usage);
             const skeleton = parseJsonBlock(out);
             if (!skeleton || !Array.isArray(skeleton['segments']) || skeleton['segments'].length === 0) return null;
             if (!pack) return null;
@@ -148,7 +171,7 @@ export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
                     role: 'user',
                     content: historyText(history)
                 }
-            ], true);
+            ], true, usage);
             const obj = parseJsonBlock(out);
             if (!obj) return null;
             const ext = normalizeExtraction(obj, history.map((t)=>t.text).join('\n'));
@@ -164,9 +187,10 @@ export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
                     role: 'user',
                     content: `【${q.key}】${q.text}(为什么问:${q.why})`
                 }
-            ], false);
+            ], false, usage);
             return out.trim() || `【${q.key}】${q.text}`;
-        }
+        },
+        usage
     };
 }