{
  "title": "unified@3 PPE 消息映射与三轮任务对照",
  "date": "2026-10-05",
  "status": "planned_live_trials_completed",
  "environment": {
    "model": "qwen3.8-flash"
  },
  "conclusion": "保留底层上下文能力和实验协议；当前证据不足以将LLM自主组合管理设为默认。先以普通历史/宿主维护为基线，把多份记忆的权威性与失效规则收敛，再评估是否继续增加自治程度。",
  "phase1": {
    "cases": [
      {
        "arm": "A",
        "phase": "cold",
        "rep": 1,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4236
      },
      {
        "arm": "B",
        "phase": "cold",
        "rep": 1,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4224
      },
      {
        "arm": "C",
        "phase": "cold",
        "rep": 1,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4284
      },
      {
        "arm": "V2",
        "phase": "cold",
        "rep": 1,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4225
      },
      {
        "arm": "A",
        "phase": "cold",
        "rep": 2,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4254
      },
      {
        "arm": "B",
        "phase": "cold",
        "rep": 2,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4216
      },
      {
        "arm": "C",
        "phase": "cold",
        "rep": 2,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4245
      },
      {
        "arm": "V2",
        "phase": "cold",
        "rep": 2,
        "pass": false,
        "reason": "请求含预算题，回复声称无任务",
        "tokens": 4580
      },
      {
        "arm": "A",
        "phase": "cold",
        "rep": 3,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4251
      },
      {
        "arm": "B",
        "phase": "cold",
        "rep": 3,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4226
      },
      {
        "arm": "C",
        "phase": "cold",
        "rep": 3,
        "pass": true,
        "reason": "9500、余额300、未超预算，0工具",
        "tokens": 4245
      },
      {
        "arm": "A",
        "phase": "tool",
        "rep": 1,
        "pass": true,
        "reason": "一次list成功，完整答复main/9500/9800/300",
        "tokens": 8453
      },
      {
        "arm": "B",
        "phase": "tool",
        "rep": 1,
        "pass": false,
        "reason": "只调用一次list且成功，但最终漏掉明确要求的余额300，未通过完整答复门槛",
        "tokens": 8675
      },
      {
        "arm": "C",
        "phase": "tool",
        "rep": 1,
        "pass": true,
        "reason": "一次list成功，完整答复main/9500/9800/300",
        "tokens": 8433
      },
      {
        "arm": "A",
        "phase": "tool",
        "rep": 2,
        "pass": true,
        "reason": "一次list成功，完整答复main/9500/9800/300",
        "tokens": 8612
      },
      {
        "arm": "C",
        "phase": "tool",
        "rep": 2,
        "pass": true,
        "reason": "一次list成功，完整答复main/9500/9800/300",
        "tokens": 8437
      }
    ],
    "totalTokens": 89596,
    "isolation": {
      "systemIdentical": true,
      "toolsIdentical": true,
      "stateContentIdentical": true
    },
    "selectedArm": "C",
    "selectionNote": "A和C均5项通过；C=29644，A=29806，差162 tokens，仅按预注册规则选C，不是稳定效率优势。",
    "interpretation": "A保留trailing_assistant也全部通过，不能把旧v2错误简单归因于末尾assistant。v3所有组包括A都改成位置中性提示；v2是哨兵而非严格单变量对照。"
  },
  "phase2": {
    "cases": [
      {
        "arm": "C",
        "rep": 1,
        "pass": false,
        "businessTurns": [
          "pass",
          "fail",
          "diagnostic_only_core_pass"
        ],
        "stateConsistency": false,
        "reason": "第2轮仅回显空Agent Inputs；第3轮补答正确但公共记忆仍保留旧预算9800为长期约束，与窗口记忆9000冲突。",
        "notes": [
          "工具成功3次，全在main：task_memory_set一次、window_memory_set两次。未切窗。",
          "阶段记忆加入2024-ROUND-7429和价格含税/最终价假设，非原始证据。"
        ],
        "tokens": 39757,
        "loopSteps": 6,
        "toolCalls": 3
      },
      {
        "arm": "Host",
        "rep": 1,
        "pass": true,
        "passScope": "预注册核心事实/约束",
        "businessTurns": [
          "pass",
          "pass",
          "pass"
        ],
        "reason": "9000当前、9800历史、9500总价、超500、排期3要素均正确；零工具。",
        "qualityNotes": [
          "把两个项目都标为供应商A；题面‘供应商A项目4200元、B项目5300元’本身有归属歧义，不能单独作为协议失败证据。",
          "冗长、添加未要求的未知项与建议；核心通过不等于无质量问题。"
        ],
        "tokens": 8299,
        "loopSteps": 3,
        "toolCalls": 0
      },
      {
        "arm": "Host",
        "rep": 2,
        "pass": true,
        "passScope": "预注册核心事实/约束",
        "businessTurns": [
          "pass",
          "pass",
          "pass"
        ],
        "reason": "核心事实/约束全部保留，零工具。",
        "qualityNotes": [
          "再次把两个项目都标为供应商A，存在题面歧义。",
          "预算表中出现2026-…占位式文本；对未知项和建议扩写过多。"
        ],
        "tokens": 9100,
        "loopSteps": 3,
        "toolCalls": 0
      },
      {
        "arm": "C",
        "rep": 2,
        "pass": true,
        "passScope": "预注册核心事实/约束；不代表严格成本上限通过",
        "businessTurns": [
          "pass",
          "pass",
          "pass"
        ],
        "stateConsistency": true,
        "reason": "三轮业务结果正确；公共记忆改为9000当前/9800历史。创建排期和新报价复核窗口，未恢复已有报价窗口。",
        "notes": [
          "8次工具均成功：task_memory_set×3、window_open×2、window_memory_set×2、window_search×1。无折叠/压缩。",
          "最终余额写0但明确超出500；按预注册超支判断通过，数学有符号余额应为-500。",
          "写记忆时提前称已汇报；最终报告到达前该描述不准确。",
          "排期窗口保留带旧预算的上一阶段引用，未作为当前公共约束；没有回到排期窗口验证其后续解释。"
        ],
        "budget": {
          "configuredMaxRunTokens": 60000,
          "actualTokens": 64104,
          "lastInferenceTokens": 6656,
          "tokensBeforeLastInference": 57448,
          "strictCapMet": false,
          "runtimeFailed": false,
          "explanation": "Runtime允许已获准的最后推理超出剩余额度并保留正常结束答复；没有抬高配置或继续发第4轮。"
        },
        "tokens": 64104,
        "loopSteps": 11,
        "toolCalls": 8
      }
    ],
    "corePasses": {
      "C": "1/2",
      "Host": "2/2"
    },
    "tokens": 121260,
    "softBudget": 260000,
    "costBasis": "每个根Run的最终累计input+output，包含C重复1已失败后的诊断第3轮；不叠加每轮累计读数。不是缓存折扣后的计费金额。",
    "comparison": {
      "C_total": 103861,
      "Host_total": 17399,
      "ratio": 5.969366055520433,
      "successfulRepeat2Ratio": 7.044395604395604
    },
    "scope": "比较整套协议与最小普通历史基线，system协议提示、暴露工具、上下文投影均不同，不能把成本/质量差归因于某一个映射因素。"
  },
  "totalTokens": 210856,
  "evidenceHighlights": [
    {
      "finding": "输入已送达但回复只回显元信息",
      "step": 4,
      "observation": "实际providerRequest包含完整SCHEDULE-2086用户排期消息；最终assistant只输出空Agent Inputs块。",
      "causality": "说明此次偏离发生在输入包含任务的条件下，尚不能断言由某一角色/位置单独导致。"
    },
    {
      "finding": "公共记忆与窗口记忆冲突",
      "step": 6,
      "observation": "header公共任务约束仍写预算9800长期有效；同一header窗口memory写9000当前、9800历史。最终答案虽正确，派生状态不一致。"
    },
    {
      "finding": "自主切窗成功但路径不稳定",
      "observation": "第二次先main→schedule-7429，再搜索报价并创建quote-review-7429；未返回main，未恢复已完成窗口。全部上下文工具业务返回success=true。"
    },
    {
      "finding": "源事实表述质量",
      "observation": "两组均有额外展开；C首次记忆加入2024编号及含税假设。Host把两项都归供应商A，但原题供应商/项目归属有歧义，该项只作措辞质量记录，不作为严格失败判据。"
    }
  ],
  "coverage": {
    "exercised": [
      "冷启动预算题",
      "一次window_list后的工具续接",
      "真实同一会话三轮输入",
      "task_memory_set",
      "window_memory_set",
      "window_open创建/切换新窗口",
      "window_search",
      "变更约束覆盖旧预算"
    ],
    "notExercised": [
      "恢复已有/已完成窗口",
      "窗口complete",
      "消息fold_create/fold_read",
      "前缀压缩与维护子Agent",
      "长历史容量阈值触发",
      "跨模型泛化",
      "监控Agent/经验路由方案"
    ]
  },
  "limitations": [
    "每映射仅3个冷启动+最多2个续接；每多轮组只有2次同题重复，不构成统计显著性或普遍可靠性结论。",
    "phase1 A与C均通过，不支持‘末尾assistant必然有问题’。v2与v3提示词也变化，v2只作为漂移哨兵。",
    "prefix@2基线保留普通历史；本轮没有自动压缩，也没有实现阶段意图自动路由，不能宣称宿主自动窗口方案已经验证。",
    "没有调用三个复杂能力的完整组合；未调用折叠/压缩符合短任务提示，不能把未触发本身当作失败。",
    "当前Run全部用户原文跨窗口可见，最终回忆正确不能证明仅靠摘要或记忆即可恢复。",
    "工具接口成功与LLM正确选择时机、维护语义一致性是不同维度。"
  ],
  "deviations": [
    "接口run_send_input复用Run，60000自身token/12步/10工具上限约束整个会话，不是每轮；两组保持相同冻结配置，未放宽。",
    "C重复1第2轮仅回显Agent Inputs空块。读取结果与发送第3轮在同一次编排调用中完成，模型审阅前第3轮已发出；该会话已判失败，第3轮只作诊断，不用于挽回通过。",
    "第二组初始排队超过10分钟。源码显示waiting_for_input仍占Agent调度并发名额；在已保存所有第一组证据后取消第一组开放会话释放名额。清理取消不计模型失败。",
    "取消第一组开放会话后，第二组立即启动，支持并发名额被waiting_for_input占用的诊断；无需重发或修改PPE。",
    "C重复2最后一次推理前累计57448<60000，最后调用6656使总额64104；源码明确允许正常结束答复超出剩余额度。业务核心通过，但不能称严格60000成本上限通过。"
  ],
  "nextDirection": [
    "不继续用无边界提示词微调和重复重跑追求单次成功；保持此轮结果不变。",
    "先定义当前约束的唯一权威来源、版本与失效关系；历史摘要不能继续宣称被新输入取代的约束长期有效。",
    "若后续继续实现，优先减少主Agent需要决定的上下文动作，将机械折叠/容量维护交给宿主，复杂窗口作为有明确收益的可选能力。"
  ],
  "cleanup": {
    "description": "取证后取消4个已完成答题但仍waiting_for_input的测试会话，以释放Agent并发名额；不删除会话历史，不改变判分。"
  },
  "sourceReferences": [
    "AgentCoreService/src/AgentCore.Runtime/Loop/AgentLoopDecisionEngine.cs:97",
    "AgentCoreService/src/AgentCore.Infrastructure/Events/EventTaskDispatcher.cs:64"
  ],
  "artifacts": {
    "preRegisteredPlan": "experiment-plan.json",
    "coldRequestHashes": "cold-request-hashes.json"
  },
  "publicationNote": "公开版仅保留实验汇总、结果统计与分析；原始请求、消息、工具轨迹及推理快照不随本仓库公开，内部运行标识已移除。"
}