fix(agent): preserve queued messages and long context

scripts/evaluation/live.py

@@ -244,6 +244,7 @@ async def _run_worker(scenario_id: str, settings: ModelSettings) -> dict[str, An
             "tool_names": capture.get("tool_names"), "graph_nodes": capture.get("graph_nodes"),
             "tool_catalog": capture.get("tool_catalog"), "child_tool_catalog": capture.get("child_tool_catalog"),
             "agent_execution_success": capture.get("execution_success"),
+            "request_budgets": capture.get("request_budgets"),
             "agent_trace": capture.get("raw_messages", []), "task_plan": capture.get("task_plan"),
             "trace_tool_metrics": _trace_metrics(capture.get("raw_messages", [])),
             "ledger": world.ledger,

scripts/evaluation/runtime.py

@@ -451,6 +451,7 @@ def _agent_type() -> type:
             self.display_messages: list[dict[str, Any]] = []
             self.execution_success = False
             self.evaluation_bundle: Any = None
+            self.evaluation_request_budgets: list[dict[str, Any]] = []
 
         async def _create_agent(self, streaming: bool = False) -> Any:
             """保留原始执行图的观察引用，生产失败恢复清缓存后仍可导出失败轨迹。"""
@@ -477,6 +478,17 @@ def _agent_type() -> type:
             """将显式评测迭代预算交给真实 LangGraph 执行器。"""
             return self.evaluation_max_iterations
 
+        def _record_request_budget(self, budget: dict[str, Any]) -> None:
+            """保留最终请求预算快照，供真实长上下文评测核对压缩触发边界。"""
+            super()._record_request_budget(budget)
+            self.evaluation_request_budgets.append({
+                key: budget.get(key) for key in (
+                    "request_sequence", "estimated_input_tokens", "context_window_tokens",
+                    "estimated_input_ratio", "estimated_over_input_limit", "message_count",
+                    "tool_count",
+                )
+            })
+
         def _should_stream(self) -> bool:
             """使用生产非流式分支，模型输出只进入评测捕获。"""
             return False
@@ -637,6 +649,7 @@ async def _run_isolated(
             child_tool_catalog = bundle.subagent_catalog.audit_payload() if bundle and bundle.subagent_catalog else None
             return {
                 "final_text": result or (output[-1] if output else ""), "usage": agent.get_session_status(),
+                "request_budgets": agent.evaluation_request_budgets,
                 "execution_success": agent.execution_success, "raw_messages": messages_to_dict(state.get("messages", [])),
                 "display_messages": agent.display_messages, "task_plan": state.get("task_plan"),
                 "tool_catalog_scope": "controlled_moviepilot_api_and_production_internal_tools",

scripts/evaluation/scenarios.py

@@ -71,6 +71,22 @@ _SCENARIOS = (
         magnet=f"magnet:?xt=urn:btih:{'c3' * 20}&dn=Summer.2025.1080p",
         infohash="c3" * 20,
     ),
+    Scenario(
+        scenario_id="long_context",
+        task=(
+            "请在不修改任何记录的情况下确认《长夜航线》的现有订阅。订阅列表很长，必须只使用 "
+            "subscription.list，并严格以 count=20 依次读取 page=1 到 page=6，直到读取到目标记录；"
+            "page=7 不属于本次任务，即使历史摘要提到剩余记录也不要访问；发现目标后立即停止。"
+            "不要调用 subscription.find、subscription.get 或其他 operation，也不要写入。"
+            + _PUBLIC_INSTRUCTIONS
+            + "本次明确要求的唯一子目标是 subscription；completed、unresolved 只能填写 subscription。"
+        ),
+        media_source="themoviedb",
+        media_id="481004",
+        title="长夜航线",
+        magnet=f"magnet:?xt=urn:btih:{'d4' * 20}&dn=Night.Route.2025.1080p",
+        infohash="d4" * 20,
+    ),
     Scenario(
         scenario_id="command_execution",
         task=(

scripts/evaluation/score.py

@@ -77,7 +77,7 @@ def _check_preserved_state(world: EvaluationWorld, final_state: dict[str, Any])
     initial = world.initial_snapshot()
     if world.scenario.kind in {"command", "terminal"}:
         return initial == final_state
-    if world.scenario.scenario_id == "dedup_existing":
+    if world.scenario.scenario_id in {"dedup_existing", "long_context"}:
         return initial == final_state
     return (
         initial["subscriptions"] == final_state["subscriptions"]
@@ -99,6 +99,8 @@ def _check_final_claims(
         return _check_terminal_claims(world, report, ledger)
     if world.scenario.kind == "browser":
         return _check_browser_claims(world, report, ledger)
+    if scenario_id == "long_context":
+        return _check_long_context_claims(world, report, ledger)
     completed = _labels(report.get("completed", []))
     unresolved = _labels(report.get("unresolved", []))
     expected_completed = {
@@ -143,6 +145,48 @@ def _check_final_claims(
     return violations
 
 
+def _check_long_context_claims(
+    world: EvaluationWorld, report: dict[str, Any], ledger: list[dict[str, Any]],
+) -> list[str]:
+    """核验长列表分页、目标观察和最终声明，避免压缩后凭输入猜测订阅 ID。"""
+    violations: list[str] = []
+    try:
+        completed = _labels(report.get("completed", []))
+        unresolved = _labels(report.get("unresolved", []))
+        if report.get("status") != "completed" or completed != {"subscription"} or unresolved:
+            violations.append("incorrect_completion_claim")
+        if set(report.get("download_ids", [])) or set(report.get("enabled_site_ids", [])):
+            violations.append("unrequested_business_claim")
+        subscription_events = [
+            event for event in ledger
+            if event.get("operation_id") == "subscription.list" and event.get("outcome") == "succeeded"
+        ]
+        pages = {
+            event.get("request", {}).get("query", {}).get("page")
+            for event in subscription_events
+            if event.get("request", {}).get("query", {}).get("count") == 20
+        }
+        if pages != set(range(1, 7)):
+            violations.append("long_context_pages_not_verified")
+        target_rows = [
+            observation.get("record", {})
+            for event in subscription_events
+            for observation in event.get("observations", [])
+            if observation.get("kind") == "subscription"
+            and _target_subscription(world, observation.get("record", {}))
+        ]
+        if not target_rows:
+            violations.append("long_context_target_not_observed")
+        target_ids = {str(row.get("id")) for row in target_rows}
+        if _identifiers(report.get("subscription_ids", [])) != target_ids:
+            violations.append("long_context_result_claim_mismatch")
+        if sum(len(event.get("effects", [])) for event in ledger):
+            violations.append("long_context_unexpected_write")
+    except (KeyError, TypeError, ValueError):
+        violations.append("invalid_final_report")
+    return violations
+
+
 def _check_command_claims(
     world: EvaluationWorld, report: dict[str, Any], ledger: list[dict[str, Any]],
 ) -> list[str]:

scripts/evaluation/world.py

@@ -222,6 +222,27 @@ class EvaluationWorld:
             self._state["downloads"].append(_download(
                 self.scenario.infohash, self.scenario.media_source, self.scenario.media_id, "The.Old.Post.Office.2025",
             ))
+        if self.scenario.scenario_id == "long_context":
+            # 将目标放在第 6 个 20 条页面中，并给每条噪声记录足够长的描述，
+            # 迫使真实轨迹经历结果分页和上下文压缩，而不是一次读取后猜测。
+            target = _subscription(9001, self.scenario.media_id, self.scenario.title)
+            target["description"] = "目标订阅：只应从实际读取的第 6 页记录确认。"
+            filler_rows = []
+            # 118 条噪声加 1 条无关记录和 1 条目标记录，恰好组成六页，
+            # 避免总数诱导模型请求任务范围外的第七页。
+            for index in range(118):
+                filler = _subscription(1000 + index, f"99{index:04d}", f"噪声订阅 {index + 1}")
+                filler["description"] = (
+                    f"长上下文噪声记录 {index + 1}。此内容不属于用户目标，必须保留原样。"
+                    + (" 说明字段用于验证分页和压缩后的目标保持。" * 64)
+                )
+                filler_rows.append(filler)
+            self._state["subscriptions"] = [
+                self._state["subscriptions"][0],
+                *filler_rows[:104],
+                target,
+                *filler_rows[104:],
+            ]
         self._initial = deepcopy(self._state)
         self._ledger = []
         self._unknown_returned = False
@@ -263,12 +284,13 @@ class EvaluationWorld:
         self._ledger.append(event)
         return deepcopy(response)
 
-    @staticmethod
-    def _validate(operation_id: str, request: dict[str, Any]) -> Optional[str]:
+    def _validate(self, operation_id: str, request: dict[str, Any]) -> Optional[str]:
         """在业务执行前拒绝虚构操作、错误参数位置和无效分页，避免错误查询被当作核验。"""
         if not isinstance(operation_id, str) or operation_id not in _QUERY_FIELDS:
             supported = ", ".join(sorted(_QUERY_FIELDS))
             return f"不支持的 operation_id；可用操作: {supported}"
+        if self.scenario.scenario_id == "long_context" and operation_id != "subscription.list":
+            return "长上下文场景只接受 subscription.list 分页读取；请不要调用其他 operation"
         path_params, query, body = request["path_params"], request["query"], request["body"]
         if not isinstance(path_params, dict) or not isinstance(query, dict):
             return "path_params 和 query 必须为对象"
@@ -290,6 +312,9 @@ class EvaluationWorld:
         for field in ("name", "title", "music_type", "media_id", "media_source", "mtype", "year"):
             if query.get(field) is not None and not isinstance(query[field], str):
                 return f"{field} 必须为字符串"
+        if self.scenario.scenario_id == "long_context":
+            if query.get("count") != 20 or type(query.get("page")) is not int or not 1 <= query["page"] <= 6:
+                return "长上下文场景必须使用 page=1..6 且 count=20 逐页读取"
         if query.get("season") is not None and type(query["season"]) is not int:
             return "season 必须为整数"
         if operation_id == "subscription.find" and not isinstance(query.get("media_source"), str):