@danceiny/gotry 0.0.1-rc.20 → 0.0.1-rc.21

package/ts/src/benchmark-environment-bridge.ts

@@ -5,7 +5,6 @@ import {
   assertSupportedJsonSchema,
   validateJsonSchemaValue,
   ToolArgsError,
-  type JsonSchemaNode,
   type ObjectJsonSchema,
   type ToolDefinition,
 } from '@deepseek-ai/dsh-tools'
@@ -15,7 +14,9 @@ import {
   type TerminalOutputConfig,
 } from './benchmark-agent-conformance.ts'
 
-const SCHEMA_VERSION = 'gotry_benchmark_environment_bridge_v3'
+// v4(#215 Round 12):terminal_output 携带 closed body schema(结构关键字白名单,
+// 禁数据值注解面);v3 及更早配置缺该面,按契约不兼容 fail-closed。
+const SCHEMA_VERSION = 'gotry_benchmark_environment_bridge_v4'
 export const BENCHMARK_TOOL_RESULT_SCHEMA_VERSION = 'gotry_benchmark_tool_result_v1'
 export const BENCHMARK_DOMAIN_RECOVERIES = ['none', 'retry_same', 'revise_arguments', 'choose_alternative'] as const
 type BenchmarkDomainRecovery = typeof BENCHMARK_DOMAIN_RECOVERIES[number]
@@ -55,38 +56,6 @@ export interface BenchmarkToolDescriptor {
   domain_outcomes: Array<{ status: 'miss' | 'error'; code: string; recovery: BenchmarkDomainRecovery }>
 }
 
-function descriptorSchema(tools: readonly BenchmarkToolDescriptor[]): JsonSchemaNode {
-  return {
-    oneOf: [
-      {
-        type: 'object',
-        description: 'List the frozen benchmark tool descriptors.',
-        properties: { action: { type: 'string', const: 'tools' } },
-        required: ['action'],
-        additionalProperties: false,
-      },
-      {
-        type: 'object',
-        description: 'List the frozen bridge error contract.',
-        properties: { action: { type: 'string', const: 'errors' } },
-        required: ['action'],
-        additionalProperties: false,
-      },
-      ...tools.map<JsonSchemaNode>(tool => ({
-        type: 'object',
-        description: tool.description,
-        properties: {
-          action: { type: 'string', const: 'call' },
-          tool: { type: 'string', const: tool.name },
-          arguments: tool.input_schema,
-        },
-        required: ['action', 'tool', 'arguments'],
-        additionalProperties: false,
-      })),
-    ],
-  }
-}
-
 function plainObject(value: unknown): value is Record<string, unknown> {
   return typeof value === 'object' && value !== null && !Array.isArray(value)
 }
@@ -426,7 +395,15 @@ export function registerBenchmarkEnvironmentBridge(
   if (!subprocess) throw new Error('benchmark environment bridge subprocess unavailable')
 
   const parameters: Record<string, unknown> = deepFreezeJson({
-    oneOf: descriptorSchema(bridge.tools).oneOf,
+    type: 'object',
+    description: 'Flat benchmark bridge wire: choose an action, then provide the mapped tool and its arguments.',
+    properties: {
+      action: { type: 'string', enum: ['tools', 'call', 'errors'], description: 'tools=列出可调工具;call=执行工具;errors=拉取错误契约' },
+      tool: { type: 'string', enum: bridge.tools.map(item => item.name), description: 'action=call 时选择的冻结工具名' },
+      arguments: { type: 'object', additionalProperties: true, description: 'action=call 时传给工具的参数对象' },
+    },
+    required: ['action'],
+    additionalProperties: false,
   })
   assertSupportedJsonSchema(parameters)
 
@@ -444,11 +421,21 @@ export function registerBenchmarkEnvironmentBridge(
       const query = args as Record<string, unknown>
       const tool = query.tool
       const descriptor = bridge.tools.find(item => item.name === tool)
-      if (query.action === 'tools') return jsonObject({ ok: true, tools: bridge.tools })
-      if (query.action === 'errors') return jsonObject({ ok: true, errors: BRIDGE_ERROR_CONTRACT })
+      const keys = Object.keys(query)
+      if (query.action === 'tools') {
+        if (keys.length !== 1) throw new ToolArgsError(['arguments: tools action accepts no tool or arguments'])
+        return jsonObject({ ok: true, tools: bridge.tools })
+      }
+      if (query.action === 'errors') {
+        if (keys.length !== 1) throw new ToolArgsError(['arguments: errors action accepts no tool or arguments'])
+        return jsonObject({ ok: true, errors: BRIDGE_ERROR_CONTRACT })
+      }
       if (query.action !== 'call') {
         return jsonObject({ ok: false, error: 'invalid_action' })
       }
+      if (keys.length !== 3 || !Object.hasOwn(query, 'tool') || !Object.hasOwn(query, 'arguments')) {
+        throw new ToolArgsError(['arguments: call action requires exactly tool and arguments'])
+      }
       if (typeof tool !== 'string' || !descriptor) {
         return jsonObject({ ok: false, error: 'disallowed_tool' })
       }
@@ -457,9 +444,7 @@ export function registerBenchmarkEnvironmentBridge(
       }
       const callArguments = query.arguments
       const argumentViolations = validateJsonSchemaValue(descriptor.input_schema, callArguments, 'arguments')
-      if (argumentViolations.length > 0) {
-        return jsonObject({ ok: false, error: 'invalid_arguments' })
-      }
+      if (argumentViolations.length > 0) throw new ToolArgsError(argumentViolations)
       const serialized = serializedArguments(callArguments)
       if (serialized.reason) return jsonObject({ ok: false, error: 'invalid_arguments', reason: serialized.reason })
       const controller = new AbortController()

package/ts/src/booking-surface/dsh-planner.ts

@@ -407,9 +407,11 @@ function recoverFinalResponseDecision(response: string, task: BookingCopilotTask
     console.error('[booking-copilot] finalResponse recovery rejected (invalid action):', JSON.stringify({ kind: action.kind, errors: validation.errors.slice(0, 6) }).slice(0, 600))
     return null
   }
-  repairPlannerFactRefs(action)
+  const repairedRefs = repairPlannerFactRefs(action)
+  const repairedValidation = validateBookingReadAction(action as unknown as BookingReadAction)
+  if (!repairedValidation.ok) return null
   try {
-    assertPlannerSafeRefs(action)
+    assertPlannerSafeRefs(action, repairedRefs)
   } catch (error) {
     console.error('[booking-copilot] finalResponse recovery rejected (unsafe ref):', JSON.stringify({ actionId: action.actionId, factRefs: action.factRefs }).slice(0, 600))
     return null
@@ -426,17 +428,22 @@ function recoverFinalResponseDecision(response: string, task: BookingCopilotTask
 const PLANNER_SAFE_REF_PATTERN = /^[A-Za-z0-9][A-Za-z0-9:._-]*$/
 
 // Models cite prompt facts in URI-ish syntax (`fact://turn_X/request`,
-// `turn_X#request`); characters outside the runtime ref charset are mapped
-// deterministically to '.' so repair succeeds without burning the retry
-// budget. Anything the sanitizer cannot make unique enough still fails the
-// safe-ref gate into the retry path.
-function repairPlannerFactRefs(action: Record<string, unknown>): void {
+// `turn_X#request`). Preserve already-safe refs exactly; unsafe refs enter the
+// reserved modelref namespace with the full SHA-256 of the raw UTF-8 value.
+function repairPlannerFactRefs(action: Record<string, unknown>): Set<string> {
   const factRefs = action.factRefs
-  if (!Array.isArray(factRefs)) return
-  action.factRefs = factRefs.map((ref) => (typeof ref === 'string' ? ref.replace(/#/g, ':').replace(/[^A-Za-z0-9:._-]/g, '.') : ref))
+  const repairedRefs = new Set<string>()
+  if (!Array.isArray(factRefs)) return repairedRefs
+  action.factRefs = factRefs.map((ref) => {
+    if (typeof ref !== 'string' || PLANNER_SAFE_REF_PATTERN.test(ref)) return ref
+    const alias = `modelref:${createHash('sha256').update(ref, 'utf8').digest('hex')}`
+    repairedRefs.add(alias)
+    return alias
+  })
+  return repairedRefs
 }
 
-function assertPlannerSafeRefs(action: Record<string, unknown>): void {
+function assertPlannerSafeRefs(action: Record<string, unknown>, repairedRefs: Set<string>): void {
   const actionId = action.actionId
   if (typeof actionId !== 'string' || !PLANNER_SAFE_REF_PATTERN.test(actionId)) {
     throw new Error(`planner_invalid_action:unsafe_action_id:${String(actionId).slice(0, 60)}`)
@@ -444,7 +451,7 @@ function assertPlannerSafeRefs(action: Record<string, unknown>): void {
   const factRefs = action.factRefs
   if (Array.isArray(factRefs)) {
     for (const ref of factRefs) {
-      if (typeof ref !== 'string' || !PLANNER_SAFE_REF_PATTERN.test(ref)) {
+      if (typeof ref !== 'string' || !PLANNER_SAFE_REF_PATTERN.test(ref) || (ref.startsWith('modelref:') && !repairedRefs.has(ref)) || ref.length > 512) {
         throw new Error(`planner_invalid_action:unsafe_fact_ref:${String(ref).slice(0, 60)}`)
       }
     }
@@ -502,8 +509,13 @@ function parseToolDecision(event: unknown, task: BookingCopilotTaskState): Booki
   // boundary, past the retry budget. Repair the common fragment syntax first,
   // then enforce the same charset here so remaining violations retry as
   // parse-class failures instead of failing the turn as PLANNER_FAILED.
-  repairPlannerFactRefs(decision.action)
-  assertPlannerSafeRefs(decision.action)
+  const repairedRefs = repairPlannerFactRefs(decision.action)
+  const repairedValidation = validateBookingReadAction(decision.action)
+  if (!repairedValidation.ok) {
+    console.error(`[booking-copilot] repaired action rejected:`, JSON.stringify({ errors: repairedValidation.errors.slice(0, 8) }).slice(0, 800))
+    throw new Error(`planner_invalid_action:${repairedValidation.errors.join('; ')}`)
+  }
+  assertPlannerSafeRefs(decision.action, repairedRefs)
   const action = decision.action as unknown as BookingReadAction
   const capability = TOOL_TO_CAPABILITY.get(name as DshEmbeddedBookingToolName)
   if (!capability || !actionsForEmbeddedCapability(capability).includes(action.kind)) {