@danceiny/gotry 0.0.1-rc.18 → 0.0.1-rc.20

package/ts/src/artifact-gate.ts

@@ -15,6 +15,7 @@
 
 import {
   flightClaimVerdict,
+  hotelClaimVerdict,
   itineraryInvariants,
   latestFactsForRouteDate,
   type BookableFact,
@@ -62,12 +63,24 @@ export interface ExtractedAirportClaim {
   city: string
 }
 
+/** 酒店 claim(D-26):断言「有房/可订/已核验」的行——目的地与档期缺失时 fail-closed */
+export interface ExtractedHotelClaim {
+  line: number
+  text: string
+  destination?: string
+  check_in?: string
+  check_out?: string
+}
+
 export interface ExtractedClaims {
   flights: ExtractedFlightClaim[]
   policies: ExtractedPolicyClaim[]
   airports: ExtractedAirportClaim[]
   /** 含「直飞」断言且可解析航线的行(通用直飞规则:该日无任何在架直飞即违例) */
   direct_lines: Array<{ line: number; text: string; origin?: string; destination?: string; date?: string }>
+  hotels: ExtractedHotelClaim[]
+  /** 渲染原语锚点(fact:<id>):命中行只走确定性回溯,启发式让位 */
+  anchors: Map<number, string>
 }
 
 /** 航班号:2 位承运码 + 3-4 位数字(UO784/EK328/MF1538/CZ8582/9C8781/FD597) */
@@ -94,6 +107,14 @@ function hasDirectAssertion(line: string): boolean {
 
 interface SectionCtx { origin?: string; destination?: string; date?: string }
 
+/** 行内第一个命中词表的城市中文名(酒店 claim 的目的地键;无命中=undefined) */
+function occ1(line: string, map: AirlineAirportMap): string | undefined {
+  for (const city of Object.keys(map.city_alias)) {
+    if (line.includes(city)) return city
+  }
+  return undefined
+}
+
 /** 从文本(小节标题或行)解析航线上下文:箭头紧邻城市对(方向最可靠)优先,否则按出现位置取前两个;日期 M.D 或 MM-DD */
 function routeCtxOf(text: string, map: AirlineAirportMap, defaultYear?: number): SectionCtx {
   const occ: Array<{ city: string; iata: string; pos: number }> = []
@@ -140,12 +161,19 @@ function routeCtxOf(text: string, map: AirlineAirportMap, defaultYear?: number):
  * 行内自带上下文优先。默认年份由 trip window 提供(远期行程的年份不含糊)。
  */
 export function extractClaims(markdown: string, map: AirlineAirportMap, opts?: { trip_year?: number }): ExtractedClaims {
-  const claims: ExtractedClaims = { flights: [], policies: [], airports: [], direct_lines: [] }
+  const claims: ExtractedClaims = { flights: [], policies: [], airports: [], direct_lines: [], hotels: [], anchors: new Map() }
   const lines = markdown.split('\n')
   let section: SectionCtx = {}
+  // 渲染原语锚点(单向生成):带 fact:<id> 的行只走锚点确定性回溯,启发式抽取让位
+  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 = { origin: ctx.origin, destination: ctx.destination, date: ctx.date ?? section.date }
@@ -179,6 +207,19 @@ export function extractClaims(markdown: string, map: AirlineAirportMap, opts?: {
     if (POLICY_WORD.test(line)) {
       claims.policies.push({ line: lineNo, text: line.trim().slice(0, 160), has_as_of: AS_OF_WORD.test(line) })
     }
+    // 酒店 claim(D-26):断言可住性(有房/可订/已核验/✓)的行——目的地与档期缺失时 fail-closed
+    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],
+      })
+    }
     // 航司→机场映射 claim:同一行出现承运码 + 机场码,且该城是双机场映射面
     for (const [city, carriers] of Object.entries(map.carrier_airport)) {
       const airports = [...new Set(Object.values(carriers))]
@@ -207,6 +248,8 @@ export type GateViolationKind =
   | 'legs_inconsistent'               // flight O&D 段与 legs 口径不一致
   | 'budget_floor_inconsistent'       // 预算下限话术低于分项最低合计
   | 'date_order'                      // 住宿/段日期越窗或倒挂
+  | 'unverifiable_hotel_claim'        // 酒店可住断言无 exact-date 事实回溯(D-26)
+  | 'fact_anchor_unknown'             // 渲染锚点 fact:<id> 在注册表不存在(锚点被手改/伪造)
 
 export interface GateViolation {
   kind: GateViolationKind
@@ -242,6 +285,43 @@ export function gateArtifact(
   const violations: GateViolation[] = []
   let traceable = 0
 
+  // 渲染锚点优先(issue #118 单向生成):带 fact:<id> 的行确定性回溯注册表——
+  // 锚点在=按事实 bookability 判;锚点不存在=手改/伪造,直接违例。启发式对锚点行让位。
+  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 as { fetched_at?: string }).fetched_at ?? ''})——负事实对应的可住/可订断言不得出现` })
+      continue
+    }
+    traceable++
+  }
+
+  // 酒店 claim(D-26):目的地+档期回溯酒店事实;无事实 fail-closed
+  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) {
       // 承运级直飞断言:注册表该 route+date 须存在该承运的 bookable 直飞
@@ -334,7 +414,7 @@ export function gateArtifact(
     if (!line.includes('联程')) return
     const nos = [...line.matchAll(FLIGHT_NO)].map(m => m[1]!.toUpperCase())
     const matched = facts.filter((f): f is Extract<BookableFact, { kind: 'flight' | 'train' }> =>
-      f.kind !== 'policy' && nos.includes(f.flight_no.toUpperCase()))
+      (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)

package/ts/src/benchmark-agent-conformance.ts

@@ -22,6 +22,8 @@ export const BENCHMARK_CONFORMANCE_STATE_UNAVAILABLE = 'BENCHMARK_CONFORMANCE_ST
 
 const IDENTIFIER = /^[A-Za-z][A-Za-z0-9_.-]*$/
 const MAX_TERMINAL_BYTES = 1024 * 1024
+const BENCHMARK_TOOL_RESULT_SCHEMA_VERSION = 'gotry_benchmark_tool_result_v1'
+const BENCHMARK_DOMAIN_RECOVERIES = new Set(['none', 'retry_same', 'revise_arguments', 'choose_alternative'])
 
 export interface TerminalOutputConfig {
   tag: string
@@ -46,6 +48,8 @@ interface TurnState {
   retryCount: number
   retryMode: RetryMode
   successfulResultStep?: number
+  domainOutcomeStep?: number
+  lastBridgeResponseStep?: number
   bridgeCallFailed: boolean
   bridgeFailureCode?: string
   retryRedispatchAttempted: boolean
@@ -89,12 +93,19 @@ function invalidTerminal(): TerminalOutputValue {
   return { ok: false, error: 'invalid_terminal_output' }
 }
 
+/** 剥离推理模型的思考标签对(如 MiniMax/DeepSeek 的 <think>…</think>):当代模型
+ *  即便被明令「只回终态包」也会以思考块前缀输出——不剥则严格终态门对推理模型
+ *  结构性失效(Round 9 治理面实测)。只剥配对的推理标签本体,其余前后缀仍 fail-closed。 */
+function stripReasoningBlocks(raw: string): string {
+  return raw.replace(/<think>[\s\S]*?<\/think>\s*/gi, '')
+}
+
 /** Parse exactly one configured tag pair containing one JSON object. */
 export function parseBenchmarkTerminal(raw: string, config: TerminalOutputConfig): TerminalOutputValue {
   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()
@@ -143,7 +154,9 @@ function assistantText(data: Record<string, unknown>): { text: string; interrupt
   return { text, interrupted: data.interrupted === true }
 }
 
-function bridgeResultStatus(data: Record<string, unknown>, callId: string): { ok: true } | { ok: false; code: string } | undefined {
+type BridgeResultStatus = { kind: 'result' } | { kind: 'domain' } | { kind: 'failure'; code: string }
+
+function bridgeResultStatus(data: Record<string, unknown>, callId: string): BridgeResultStatus | undefined {
   if (!plainObject(data.message) || !plainObject(data.message.source)) return undefined
   if (data.message.source.callId !== callId || !Array.isArray(data.message.content)) return undefined
   const block = data.message.content.find(candidate => plainObject(candidate)
@@ -157,8 +170,28 @@ function bridgeResultStatus(data: Record<string, unknown>, callId: string): { ok
   try {
     const parsed: unknown = JSON.parse(text)
     if (!plainObject(parsed)) return undefined
-    if (parsed.ok === true) return { ok: true }
+    const keys = Object.keys(parsed).sort().join(',')
+    if (parsed.ok === true && keys === 'ok,result' && (plainObject(parsed.result) || Array.isArray(parsed.result))) return { kind: 'result' }
+    if (parsed.ok === true && keys === 'ok,outcome' && plainObject(parsed.outcome)) {
+      const outcomeKeys = Object.keys(parsed.outcome).sort().join(',')
+      if (outcomeKeys === 'code,recovery,schema_version,status'
+        && parsed.outcome.schema_version === BENCHMARK_TOOL_RESULT_SCHEMA_VERSION
+        && (parsed.outcome.status === 'miss' || parsed.outcome.status === 'error')
+        && typeof parsed.outcome.code === 'string'
+        && IDENTIFIER.test(parsed.outcome.code)
+        && typeof parsed.outcome.recovery === 'string'
+        && BENCHMARK_DOMAIN_RECOVERIES.has(parsed.outcome.recovery)) return { kind: 'domain' }
+      return undefined
+    }
     if (parsed.ok === false && typeof parsed.error === 'string') {
+      const exactFailure = parsed.error === 'runner_failed'
+        ? (keys === 'error,ok' || (keys === 'error,exit_code,ok,signal'
+          && (parsed.exit_code === null || (typeof parsed.exit_code === 'number' && Number.isInteger(parsed.exit_code)))
+          && (parsed.signal === null || typeof parsed.signal === 'string')))
+        : parsed.error === 'invalid_arguments'
+          ? keys === 'error,ok,reason' && typeof parsed.reason === 'string'
+          : keys === 'error,ok'
+      if (!exactFailure) return undefined
       const code = parsed.error === 'timed_out'
         ? BENCHMARK_BRIDGE_TIMED_OUT
         : parsed.error === 'runner_failed'
@@ -168,7 +201,7 @@ function bridgeResultStatus(data: Record<string, unknown>, callId: string): { ok
             : parsed.error === 'output_truncated'
               ? BENCHMARK_BRIDGE_OUTPUT_TRUNCATED
             : BENCHMARK_BRIDGE_CALL_FAILED
-      return { ok: false, code }
+      return { kind: 'failure', code }
     }
     return undefined
   } catch {
@@ -227,10 +260,11 @@ export function createBenchmarkAgentConformance(projection: BenchmarkBridgeProje
         }
         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)) {
+        // Round 8(issue #100/#102):bridge 工具 typed 泛化后,模型出参为平铺形态
+        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
@@ -241,10 +275,15 @@ export function createBenchmarkAgentConformance(projection: BenchmarkBridgeProje
         const callId = event.data.message.source.callId
         if (typeof callId !== 'string' || !state.validCallIds.has(callId)) return
         const resultStatus = bridgeResultStatus(event.data, callId)
-        if (resultStatus?.ok === true) {
+        if (typeof event.data.step === 'number') state.lastBridgeResponseStep = event.data.step
+        if (resultStatus?.kind === 'result') {
           state.successfulResultStep = typeof event.data.step === 'number'
             ? event.data.step
             : state.successfulResultStep
+        } else if (resultStatus?.kind === 'domain') {
+          state.domainOutcomeStep = typeof event.data.step === 'number'
+            ? event.data.step
+            : state.domainOutcomeStep
         } else {
           state.bridgeCallFailed = true
           state.bridgeFailureCode = resultStatus?.code ?? BENCHMARK_BRIDGE_CALL_FAILED
@@ -266,10 +305,10 @@ export function createBenchmarkAgentConformance(projection: BenchmarkBridgeProje
       if (state.retryRedispatchAttempted) {
         return { kind: 'reject', code: BENCHMARK_BRIDGE_RETRY_CALL_NOT_ALLOWED }
       }
-      if (state.successfulResultStep === undefined) {
-        if (state.bridgeCallFailed) {
-          return { kind: 'reject', code: state.bridgeFailureCode ?? BENCHMARK_BRIDGE_CALL_FAILED }
-        }
+      if (state.bridgeCallFailed && state.successfulResultStep === undefined) {
+        return { kind: 'reject', code: state.bridgeFailureCode ?? BENCHMARK_BRIDGE_CALL_FAILED }
+      }
+      if (state.successfulResultStep === undefined && state.domainOutcomeStep === undefined) {
         if (state.retryCount >= MAX_CONFORMANCE_RETRIES) {
           return { kind: 'reject', code: BENCHMARK_BRIDGE_CALL_REQUIRED }
         }
@@ -281,7 +320,7 @@ export function createBenchmarkAgentConformance(projection: BenchmarkBridgeProje
       const terminal = state.lastAssistant
       const terminalIsValid = terminal !== undefined
         && !terminal.interrupted
-        && terminal.step > state.successfulResultStep
+        && terminal.step > state.lastBridgeResponseStep!
         && parseBenchmarkTerminal(terminal.text, projection.terminal).ok
       if (terminalIsValid) return { kind: 'accept' }
       if (state.retryCount >= MAX_CONFORMANCE_RETRIES) {
@@ -338,7 +377,7 @@ function systemSection(projection: BenchmarkBridgeProjection): { name: string; t
     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.',