@danceiny/gotry 0.0.1-rc.17 → 0.0.1-rc.19

package/dist/src/artifact-gate.js

@@ -1,4 +1,4 @@
-import { flightClaimVerdict, itineraryInvariants, latestFactsForRouteDate } from './bookable-facts.js';
+import { flightClaimVerdict, hotelClaimVerdict, itineraryInvariants, latestFactsForRouteDate } from './bookable-facts.js';
 export const ARTIFACT_GATE_SCHEMA = 'gotry_artifact_gate.v1';
 const FLIGHT_NO = /(?<![A-Za-z0-9])([A-Z0-9]{2}\d{3,4})(?![\d])/g;
 const CARRIER_DIRECT = /(?<![A-Za-z0-9])([A-Z0-9]{2})(?!\d)(?:\s|[一-龥]){0,6}直飞/g;
@@ -26,6 +26,12 @@ function hasDirectAssertion(line) {
     const before = line.slice(Math.max(0, idx - 3), idx);
     return !(before.endsWith('无') || before.endsWith('不') || before.endsWith('没有') || before.endsWith('未见'));
 }
+function occ1(line, map) {
+    for (const city of Object.keys(map.city_alias)){
+        if (line.includes(city)) return city;
+    }
+    return undefined;
+}
 function routeCtxOf(text, map, defaultYear) {
     const occ = [];
     for (const [city, iata] of Object.entries(map.city_alias)){
@@ -72,13 +78,21 @@ export function extractClaims(markdown, map, opts) {
         flights: [],
         policies: [],
         airports: [],
-        direct_lines: []
+        direct_lines: [],
+        hotels: [],
+        anchors: new Map()
     };
     const lines = markdown.split('\n');
     let section = {};
+    const ANCHOR = /fact:([0-9a-f]{16})/;
     for(let i = 0; i < lines.length; i++){
         const line = lines[i];
         const lineNo = i + 1;
+        const anchorMatch = line.match(ANCHOR);
+        if (anchorMatch) {
+            claims.anchors.set(lineNo, anchorMatch[1].toLowerCase());
+            continue;
+        }
         if (/^#{1,4}\s/.test(line)) {
             const ctx = routeCtxOf(line, map, opts?.trip_year);
             section = {
@@ -145,6 +159,20 @@ export function extractClaims(markdown, map, opts) {
                 has_as_of: AS_OF_WORD.test(line)
             });
         }
+        const HOTEL_ASSERT = /酒店|住宿|客栈|民宿/;
+        const HOTEL_BOOKABLE = /有房|可订|在架|已核验|已验证|可住|✓|✅/;
+        if (HOTEL_ASSERT.test(line) && HOTEL_BOOKABLE.test(line)) {
+            const stays = [
+                ...line.matchAll(/(?<!\d)(\d{4}-\d{2}-\d{2})(?!\d)/g)
+            ].map((m)=>m[1]);
+            claims.hotels.push({
+                line: lineNo,
+                text: line.trim().slice(0, 160),
+                destination: occ1(line, map),
+                check_in: stays[0],
+                check_out: stays[1]
+            });
+        }
         for (const [city, carriers] of Object.entries(map.carrier_airport)){
             const airports = [
                 ...new Set(Object.values(carriers))
@@ -176,6 +204,66 @@ export function gateArtifact(markdown, facts, map, opts) {
     const claims = extractClaims(markdown, map, opts);
     const violations = [];
     let traceable = 0;
+    for (const [lineNo, factId] of claims.anchors){
+        const f = facts.find((x)=>x.fact_id === factId);
+        if (!f) {
+            violations.push({
+                kind: 'fact_anchor_unknown',
+                line: lineNo,
+                detail: `渲染锚点 fact:${factId} 不在事实注册表——锚点被手改或伪造,产物不可信`
+            });
+            continue;
+        }
+        if (f.kind !== 'policy' && f.bookability === 'unavailable_exact_date') {
+            violations.push({
+                kind: 'not_in_source',
+                line: lineNo,
+                detail: `锚点事实为 exact-date 负事实(${f.fetched_at ?? ''})——负事实对应的可住/可订断言不得出现`
+            });
+            continue;
+        }
+        traceable++;
+    }
+    for (const c of claims.hotels){
+        if (claims.anchors.has(c.line)) continue;
+        if (!c.destination) {
+            violations.push({
+                kind: 'unverifiable_hotel_claim',
+                line: c.line,
+                detail: `酒店可住断言缺目的地上下文,无法回溯 exact-date 事实——fail closed 按未核验处理`
+            });
+            continue;
+        }
+        const r = hotelClaimVerdict(facts, {
+            destination: c.destination,
+            check_in: c.check_in,
+            check_out: c.check_out
+        });
+        if (r.verdict === 'traceable') {
+            traceable++;
+            continue;
+        }
+        if (r.verdict === 'unverified') {
+            violations.push({
+                kind: 'unverifiable_hotel_claim',
+                line: c.line,
+                detail: r.reason
+            });
+        } else {
+            violations.push({
+                kind: 'not_in_source',
+                line: c.line,
+                detail: r.reason
+            });
+        }
+        if (CHECK_MARK.test(c.text)) {
+            violations.push({
+                kind: 'unconditional_check',
+                line: c.line,
+                detail: '对未核验酒店使用无条件 ✓/✅——只有 bookable_exact_date 才允许确定性标记'
+            });
+        }
+    }
     for (const c of claims.flights){
         if (c.carrier_only) {
             const hasCarrier = facts.some((f)=>f.kind === 'flight' && f.bookability === 'bookable_exact_date' && f.nonstop !== false && f.flight_no.startsWith(c.carrier_only) && (!c.origin || f.route.origin === c.origin || f.route.origin_airport === c.origin) && (!c.destination || f.route.destination === c.destination || f.route.dest_airport === c.destination) && (!c.date || f.date === c.date));
@@ -277,7 +365,7 @@ export function gateArtifact(markdown, facts, map, opts) {
         const nos = [
             ...line.matchAll(FLIGHT_NO)
         ].map((m)=>m[1].toUpperCase());
-        const matched = facts.filter((f)=>f.kind !== 'policy' && nos.includes(f.flight_no.toUpperCase()));
+        const matched = facts.filter((f)=>(f.kind === 'flight' || f.kind === 'train') && nos.includes(f.flight_no.toUpperCase()));
         const allProtected = nos.length > 0 && matched.length === nos.length && matched.every((f)=>f.protected_connection === true);
         if (!allProtected) {
             violations.push({

package/dist/src/benchmark-agent-conformance.js

@@ -29,10 +29,13 @@ function invalidTerminal() {
         error: 'invalid_terminal_output'
     };
 }
+function stripReasoningBlocks(raw) {
+    return raw.replace(/<think>[\s\S]*?<\/think>\s*/gi, '');
+}
 export function parseBenchmarkTerminal(raw, config) {
     if (!validateTerminalOutputConfig(config)) return invalidTerminal();
     if (Buffer.byteLength(raw, 'utf8') > config.max_bytes) return invalidTerminal();
-    const trimmed = raw.trim();
+    const trimmed = stripReasoningBlocks(raw).trim();
     const opening = `<${config.tag}>`;
     const closing = `</${config.tag}>`;
     if (!trimmed.startsWith(opening) || !trimmed.endsWith(closing)) return invalidTerminal();
@@ -150,8 +153,8 @@ export function createBenchmarkAgentConformance(projection) {
                 }
                 if (typeof event.data.callId !== 'string') return;
                 const args = parseToolArguments(event.data.arguments);
-                const query = args && plainObject(args.query) ? args.query : undefined;
-                if (query?.action === 'call' && typeof query.tool === 'string' && projection.allowedTools.includes(query.tool)) {
+                const flat = args && plainObject(args) ? args : undefined;
+                if (flat?.action === 'call' && typeof flat.tool === 'string' && projection.allowedTools.includes(flat.tool)) {
                     state.validCallIds.add(event.data.callId);
                 }
                 return;
@@ -261,7 +264,7 @@ function systemSection(projection) {
         text: [
             'Benchmark execution contract:',
             `- Translate every task instruction to use a CLI, shell, Python, or agent_env.cli into the native tool ${projection.toolName}; do not merely describe the intended command.`,
-            `- Call it with exactly {"query":{"action":"call","tool":"<one of: ${allowed}>","arguments":{...}}}.`,
+            `- Call it with exactly {"action":"call","tool":"<one of: ${allowed}>","arguments":{...}}.`,
             '- action:"tools" is discovery only and does not satisfy the required environment call.',
             `- After a successful tool result, reply only <${projection.terminal.tag}>{...one JSON object...}</${projection.terminal.tag}> with no prose or code fence.`,
             '- If a terminal-format correction arrives, reuse the existing result and do not call the tool again.'

package/dist/src/benchmark-environment-bridge.js

@@ -211,18 +211,74 @@ function serializedArguments(value) {
         };
     }
 }
+export const BRIDGE_ERROR_CONTRACT = Object.freeze({
+    invalid_action: {
+        recoverable: true,
+        remedy: 'action 只允许 tools 或 call'
+    },
+    disallowed_tool: {
+        recoverable: true,
+        remedy: '先 action=tools 列出 allowed_tools 清单,再从清单内选工具'
+    },
+    invalid_arguments: {
+        recoverable: true,
+        remedy: 'arguments 必须是可序列化 JSON 对象(≤64KB、深度≤12);收窄后重试'
+    },
+    timed_out: {
+        recoverable: true,
+        remedy: '上游超时;可原样重试一次或换其他工具'
+    },
+    output_truncated: {
+        recoverable: true,
+        remedy: '输出超出上限;让被调工具收窄查询范围后重试'
+    },
+    invalid_json: {
+        recoverable: true,
+        remedy: '上游输出不是合法 JSON;重试一次或换工具'
+    },
+    invalid_output: {
+        recoverable: true,
+        remedy: '上游输出结构不符(如原始字符串);重试或换工具'
+    },
+    runner_failed: {
+        recoverable: true,
+        remedy: '上游进程非零退出;可重试一次,持续失败换工具'
+    },
+    spawn_failed: {
+        recoverable: false,
+        remedy: '可执行文件不可用——环境问题,调用方不可恢复'
+    },
+    forbidden_output: {
+        recoverable: false,
+        remedy: '输出含未授权键(策略边界)——不可恢复,换工具或放弃'
+    }
+});
 export function registerBenchmarkEnvironmentBridge(path, register, subprocess) {
     const bridge = loadBenchmarkEnvironmentConfig(path);
     if (!bridge) throw new Error('benchmark environment bridge configuration unavailable');
     if (!subprocess) throw new Error('benchmark environment bridge subprocess unavailable');
     register(defineTool({
         name: 'gotry_benchmark_environment',
-        description: 'When a prompt asks to run agent_env.cli, use action=call here with the mapped tool; arbitrary shell is not exposed.',
+        description: 'When a prompt asks to run agent_env.cli, use action=call here with the mapped tool; arbitrary shell is not exposed. ' + 'Recoverable domain-error contract: every failure returns { ok:false, error:<code>, ... } from a closed vocabulary; ' + 'call action=errors to fetch the per-code recoverable flag and remedy before retrying.',
         parameters: {
-            query: {
-                type: 'json',
+            action: {
+                type: 'string',
+                enum: [
+                    'tools',
+                    'call',
+                    'errors'
+                ],
                 required: true,
-                description: '{ action: "tools" | "call", tool?: string, arguments?: object }'
+                description: 'tools=列出可调工具;call=执行被映射工具;errors=拉取可恢复错误契约表'
+            },
+            tool: {
+                type: 'string',
+                description: '被调工具名(必须在 allowed_tools 清单内;action=call 时必填)'
+            },
+            arguments: {
+                type: 'object',
+                additionalProperties: true,
+                description: '传给被调工具的参数对象(action=call 时可选;≤64KB、深度≤12)'
             }
         },
         output: {
@@ -237,15 +293,20 @@ export function registerBenchmarkEnvironmentBridge(path, register, subprocess) {
                 ]
         },
         async execute (args) {
-            const query = plainObject(args?.query) ? args.query : {};
-            if (query.action === 'tools') return jsonObject({
+            const action = args?.action;
+            if (action === 'errors') return jsonObject({
+                ok: true,
+                errors: BRIDGE_ERROR_CONTRACT
+            });
+            if (action === 'tools') return jsonObject({
                 ok: true,
                 tools: bridge.allowed_tools
             });
-            if (query.action !== 'call') return jsonObject({
+            if (action !== 'call') return jsonObject({
                 ok: false,
                 error: 'invalid_action'
             });
+            const query = args;
             if (typeof query.tool !== 'string' || !bridge.allowed_tools.includes(query.tool)) {
                 return jsonObject({
                     ok: false,

package/dist/src/bookable-facts.js

@@ -52,7 +52,7 @@ export function factsFromFlyai(q, r, fetchedAt, alias) {
     const queryId = makeQueryId(source, q.kind, q.origin, q.destination, q.date);
     const origin = iataOf(alias, q.origin);
     const destination = iataOf(alias, q.destination);
-    if (r.verdict === 'error') return [];
+    if (r.verdict === 'error' || r.verdict === 'needs-setup') return [];
     if (r.verdict === 'miss') return [
         negativeFact(queryId, q.kind, q.origin, q.destination, q.date, source, fetchedAt, alias)
     ];
@@ -159,7 +159,7 @@ export function dedupeFacts(facts) {
     ].sort((a, b)=>a.fetched_at.localeCompare(b.fetched_at) || a.fact_id.localeCompare(b.fact_id));
 }
 export function latestFactsForRouteDate(facts, origin, destination, date) {
-    const hits = facts.filter((f)=>f.kind !== 'policy' && f.route.origin === origin && f.route.destination === destination && f.date === date);
+    const hits = facts.filter((f)=>(f.kind === 'flight' || f.kind === 'train') && f.route.origin === origin && f.route.destination === destination && f.date === date);
     const latestQueryTs = new Map();
     for (const f of hits){
         const prev = latestQueryTs.get(f.query_id);
@@ -173,7 +173,7 @@ export function latestFactsForRouteDate(facts, origin, destination, date) {
 }
 export function flightClaimVerdict(facts, claim) {
     const no = claim.flight_no.toUpperCase().replace(/\s+/g, '');
-    const sameNo = facts.filter((f)=>f.kind !== 'policy' && f.flight_no.toUpperCase() === no && f.bookability === 'bookable_exact_date');
+    const sameNo = facts.filter((f)=>(f.kind === 'flight' || f.kind === 'train') && f.flight_no.toUpperCase() === no && f.bookability === 'bookable_exact_date');
     const routeMatched = sameNo.filter((f)=>(!claim.origin || f.route.origin === claim.origin || f.route.origin_airport === claim.origin) && (!claim.destination || f.route.destination === claim.destination || f.route.dest_airport === claim.destination) && (!claim.date || f.date === claim.date));
     if (routeMatched.length > 0) return {
         verdict: 'traceable',
@@ -310,6 +310,10 @@ export function defaultReviewBy(tripDate) {
     return d.toISOString().slice(0, 10);
 }
 export function renderFlightFact(f) {
+    return renderFlightFactInner(f) + ` <!-- fact:${f.fact_id} -->`;
+}
+function renderFlightFactInner(f) {
+    const anchor = ` <!-- fact:${f.fact_id} -->`;
     const route = `${f.route.origin}→${f.route.destination}`;
     const evidence = `[${f.source}@${f.fetched_at} #${f.query_id}]`;
     if (f.bookability === 'unavailable_exact_date') {
@@ -344,6 +348,61 @@ export function renderNightsLine(it) {
     const flightLegs = it.od_segments.filter((s)=>s.mode === 'flight').reduce((acc, s)=>acc + s.legs, 0);
     return `共 ${daysBetween(it.trip_start, it.trip_end) + 1} 天 ${totalNights} 晚 = ${hotelNights} 个酒店夜 + ${it.onboard_nights} 个机上夜;航班按 O&D ${flightOd} 段(共 ${flightLegs} 个 flight legs)`;
 }
+export function factsFromHotel(input) {
+    if (!input.checkIn || !input.checkOut) return [];
+    if (input.verdict !== 'hit' && input.verdict !== 'miss') return [];
+    const queryId = `hotel:${input.source}:${input.destination}:${input.checkIn}_${input.checkOut}`;
+    return [
+        {
+            schema: BOOKABLE_FACT_SCHEMA,
+            kind: 'hotel',
+            fact_id: makeFactId([
+                input.destination,
+                input.checkIn,
+                input.checkOut,
+                queryId
+            ]),
+            destination: input.destination,
+            check_in: input.checkIn,
+            check_out: input.checkOut,
+            verdict: input.verdict,
+            bookability: input.verdict === 'hit' ? 'bookable_exact_date' : 'unavailable_exact_date',
+            options_masked: input.options,
+            source: input.source,
+            query_id: queryId,
+            fetched_at: input.fetchedAt,
+            as_of: input.fetchedAt.slice(0, 10),
+            tier: 'live_inventory'
+        }
+    ];
+}
+export function hotelClaimVerdict(facts, claim) {
+    const scoped = facts.filter((f)=>f.kind === 'hotel' && f.destination === claim.destination && (!claim.check_in || f.check_in === claim.check_in) && (!claim.check_out || f.check_out === claim.check_out));
+    if (scoped.length === 0) {
+        return {
+            verdict: 'unverified',
+            reason: `酒店 claim(${claim.destination}${claim.check_in ? ` ${claim.check_in}→${claim.check_out ?? '?'}` : ''})未经 exact-date 源核验——先检索再写结论`
+        };
+    }
+    const hit = scoped.find((f)=>f.bookability === 'bookable_exact_date');
+    if (hit) return {
+        verdict: 'traceable',
+        fact: hit,
+        reason: 'exact-date 命中'
+    };
+    const latest = scoped[scoped.length - 1];
+    return {
+        verdict: 'not_in_source',
+        fact: latest,
+        reason: `该档期 exact-date 源正常返回 0 条(${latest.fetched_at})——不得写有房/可订`
+    };
+}
+export function renderHotelFact(f) {
+    if (f.verdict !== 'hit') {
+        return `- ${f.destination} ${f.check_in}→${f.check_out}:exact-date 源 0 条在架(不可订) <!-- fact:${f.fact_id} -->`;
+    }
+    return `- ${f.destination} ${f.check_in}→${f.check_out}:住宿已核验(在架 ${f.options_masked} 家;价格上游打码,真实价以落地页为准) <!-- fact:${f.fact_id} -->`;
+}
 
 
 //# sourceURL=ts/src/bookable-facts.ts