docs(deploy): record production access topology

CLAUDE.md

@@ -164,6 +164,20 @@ tests/
 
 `pytest.ini_options` 中 `pythonpath = ["backend", "."]`，测试直接 import `agenteval` 和 `cli`。
 
+## Agent skills
+
+### Issue tracker
+
+工作项跟踪使用 `git.solahqb22.cn/solahqb/AgentEvalTool` 的 Gitea Issues。详见 `docs/agents/issue-tracker.md`。
+
+### Triage labels
+
+Triage 使用五个默认角色标签。详见 `docs/agents/triage-labels.md`。
+
+### Domain docs
+
+仓库采用单一上下文：根目录 `CONTEXT.md` 配合 `docs/adr/`。详见 `docs/agents/domain.md`。
+
 ## 部署关键信息
 
 - **t480**：`sola-t480`（192.168.8.145:8001），Docker Compose **双容器**（`agenteval` + `openclaw-eval`）

backend/agenteval/channels/__init__.py

@@ -1,7 +1,27 @@
 """Message channel adapters."""
 
-from agenteval.channels.base import ChannelHealth, EvalChannel, Reply, SendResult
+from agenteval.channels.base import (
+    ChannelHealth,
+    ChannelTransportError,
+    EvalChannel,
+    ExchangeOutcome,
+    ExchangeStatus,
+    Reply,
+    SendResult,
+    normalize_reply_text,
+)
 from agenteval.channels.factory import ChannelFactory
 from agenteval.channels.tutu import TutuApiChannel
 
-__all__ = ["ChannelFactory", "ChannelHealth", "EvalChannel", "Reply", "SendResult", "TutuApiChannel"]
+__all__ = [
+    "ChannelFactory",
+    "ChannelHealth",
+    "ChannelTransportError",
+    "EvalChannel",
+    "ExchangeOutcome",
+    "ExchangeStatus",
+    "Reply",
+    "SendResult",
+    "TutuApiChannel",
+    "normalize_reply_text",
+]

backend/agenteval/channels/base.py

@@ -1,9 +1,12 @@
 """Abstract base class for message channels."""
 
+import json
+import time
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, Optional
+from enum import Enum
+from typing import Any, Awaitable, Callable, Optional
 
 
 @dataclass
@@ -29,6 +32,143 @@ class Reply:
     raw_message: Optional[dict[str, Any]] = None
 
 
+class ExchangeStatus(str, Enum):
+    """Outcome categories for one complete channel exchange."""
+
+    SUCCESS = "success"
+    SEND_FAILED = "send_failed"
+    REPLY_TIMEOUT = "reply_timeout"
+    POLL_FAILED = "poll_failed"
+
+
+class ChannelTransportError(Exception):
+    """Expected operational transport failure raised by a channel adapter.
+
+    Configuration and programming errors use their original exception types
+    and deliberately cross the exchange interface for diagnosis.
+    """
+
+
+def normalize_reply_text(content: Any) -> str:
+    """Convert a channel reply payload into stable plain text.
+
+    Adapters may return a plain string, a Tutu-style ``msgBody`` object, or a
+    provider-specific mapping.  The exchange interface exposes only text so
+    callers do not need to learn each provider's response shape.
+    """
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, dict):
+        body = content.get("msgBody")
+        if isinstance(body, dict):
+            for key in ("content", "text", "message"):
+                value = body.get(key)
+                if isinstance(value, str) and value:
+                    return value
+        elif isinstance(body, str) and body:
+            return body
+
+        for key in ("content", "text", "message"):
+            value = content.get(key)
+            if isinstance(value, str) and value:
+                return value
+        return json.dumps(content, ensure_ascii=False)
+    return str(content)
+
+
+@dataclass(frozen=True)
+class ExchangeOutcome:
+    """Result of a complete send-and-reply exchange.
+
+    ``diagnostic`` is intentionally opaque to callers.  Adapters may retain
+    provider-specific data there without expanding the public interface.
+    """
+
+    status: ExchangeStatus
+    correlation_id: Optional[str] = None
+    reply_text: Optional[str] = None
+    latency_ms: Optional[int] = None
+    reason: Optional[str] = None
+    diagnostic: Any = None
+
+    def __post_init__(self) -> None:
+        if self.status is ExchangeStatus.SUCCESS:
+            if not self.correlation_id:
+                raise ValueError("a successful exchange requires a correlation_id")
+            if self.reply_text is None:
+                raise ValueError("a successful exchange requires reply_text")
+            if self.latency_ms is None or self.latency_ms < 0:
+                raise ValueError("a successful exchange requires non-negative latency_ms")
+
+    @property
+    def ok(self) -> bool:
+        return self.status is ExchangeStatus.SUCCESS
+
+    @property
+    def expected_failure(self) -> bool:
+        return self.status in {
+            ExchangeStatus.SEND_FAILED,
+            ExchangeStatus.REPLY_TIMEOUT,
+            ExchangeStatus.POLL_FAILED,
+        }
+
+    @classmethod
+    def succeeded(
+        cls,
+        *,
+        correlation_id: str,
+        reply: Any,
+        latency_ms: int,
+        diagnostic: Any = None,
+    ) -> "ExchangeOutcome":
+        return cls(
+            status=ExchangeStatus.SUCCESS,
+            correlation_id=correlation_id,
+            reply_text=normalize_reply_text(reply),
+            latency_ms=latency_ms,
+            diagnostic=diagnostic,
+        )
+
+    @classmethod
+    def send_failed(cls, reason: str, *, diagnostic: Any = None) -> "ExchangeOutcome":
+        return cls(status=ExchangeStatus.SEND_FAILED, reason=reason, diagnostic=diagnostic)
+
+    @classmethod
+    def reply_timeout(
+        cls,
+        *,
+        correlation_id: Optional[str] = None,
+        latency_ms: Optional[int] = None,
+        diagnostic: Any = None,
+    ) -> "ExchangeOutcome":
+        return cls(
+            status=ExchangeStatus.REPLY_TIMEOUT,
+            correlation_id=correlation_id,
+            latency_ms=latency_ms,
+            reason="reply timeout",
+            diagnostic=diagnostic,
+        )
+
+    @classmethod
+    def poll_failed(
+        cls,
+        reason: str,
+        *,
+        correlation_id: Optional[str] = None,
+        latency_ms: Optional[int] = None,
+        diagnostic: Any = None,
+    ) -> "ExchangeOutcome":
+        return cls(
+            status=ExchangeStatus.POLL_FAILED,
+            correlation_id=correlation_id,
+            latency_ms=latency_ms,
+            reason=reason,
+            diagnostic=diagnostic,
+        )
+
+
 class EvalChannel(ABC):
     """Abstract message channel used to interact with an evaluation target.
 
@@ -41,20 +181,79 @@ class EvalChannel(ABC):
         """Verify the channel can reach the target."""
         ...
 
+    async def exchange(
+        self,
+        content: str,
+        *,
+        timeout: float = 30.0,
+        poll_interval: float = 1.0,
+        on_sent: Optional[Callable[[SendResult], Awaitable[None]]] = None,
+        **kwargs: Any,
+    ) -> ExchangeOutcome:
+        """Complete one send-and-reply exchange through this channel.
+
+        ``on_sent`` runs after the adapter confirms a successful send and
+        before reply polling begins.  A hook failure is deliberately allowed
+        to propagate so callers cannot observe a reply without a durable sent
+        record.
+        """
+        started_at = time.monotonic()
+        try:
+            send_result = await self._send(content, **kwargs)
+        except ChannelTransportError as exc:
+            return ExchangeOutcome.send_failed(str(exc))
+        if not send_result.ok:
+            return ExchangeOutcome.send_failed(send_result.error or "channel send failed")
+
+        correlation_id = send_result.question_msg_id
+        if not correlation_id:
+            raise ValueError("a successful send must provide question_msg_id")
+
+        if on_sent is not None:
+            await on_sent(send_result)
+
+        try:
+            reply = await self._poll_reply(
+                correlation_id,
+                timeout=timeout,
+                poll_interval=poll_interval,
+            )
+        except ChannelTransportError as exc:
+            return ExchangeOutcome.poll_failed(
+                str(exc),
+                correlation_id=correlation_id,
+                latency_ms=_elapsed_ms(started_at),
+            )
+
+        latency_ms = _elapsed_ms(started_at)
+        if reply is None:
+            return ExchangeOutcome.reply_timeout(
+                correlation_id=correlation_id,
+                latency_ms=latency_ms,
+            )
+
+        return ExchangeOutcome.succeeded(
+            correlation_id=correlation_id,
+            reply=reply.content,
+            latency_ms=latency_ms,
+            diagnostic={"send": send_result.raw_response, "reply": reply.raw_message},
+        )
+
     @abstractmethod
-    async def send(self, content: str, **kwargs: Any) -> SendResult:
-        """Send a message to the target and return its question message id."""
+    async def _send(self, content: str, **kwargs: Any) -> SendResult:
+        """Transport primitive used by :meth:`exchange`."""
         ...
 
     @abstractmethod
-    async def poll_reply(
+    async def _poll_reply(
         self,
         question_msg_id: str,
         timeout: float = 30.0,
         poll_interval: float = 1.0,
     ) -> Optional[Reply]:
-        """Wait for a reply to a previously sent message.
-
-        Implementations may poll (REST) or await a push event (WebSocket).
-        """
+        """Transport primitive used by :meth:`exchange`."""
         ...
+
+
+def _elapsed_ms(started_at: float) -> int:
+    return max(0, int((time.monotonic() - started_at) * 1000))

backend/agenteval/channels/http.py

@@ -29,7 +29,7 @@ from typing import Any, Optional
 
 import httpx
 
-from agenteval.channels.base import ChannelHealth, EvalChannel, Reply, SendResult
+from agenteval.channels.base import ChannelHealth, ChannelTransportError, EvalChannel, Reply, SendResult
 
 
 def _get_path(data: Any, path: str) -> Any:
@@ -76,7 +76,7 @@ class HttpChannel(EvalChannel):
         except Exception as exc:
             return ChannelHealth(ok=False, message=str(exc))
 
-    async def send(self, content: str, **kwargs: Any) -> SendResult:
+    async def _send(self, content: str, **kwargs: Any) -> SendResult:
         body_str = self.send_body_template.replace("{message}", content)
         try:
             body = json.loads(body_str)
@@ -90,10 +90,10 @@ class HttpChannel(EvalChannel):
             data = resp.json()
             msg_id = str(_get_path(data, self.msg_id_path) or uuid.uuid4())
             return SendResult(ok=True, question_msg_id=msg_id, raw_response=data)
-        except Exception as exc:
+        except (httpx.HTTPError, json.JSONDecodeError) as exc:
             return SendResult(ok=False, error=str(exc))
 
-    async def poll_reply(
+    async def _poll_reply(
         self,
         question_msg_id: str,
         timeout: float = 30.0,
@@ -124,8 +124,8 @@ class HttpChannel(EvalChannel):
                         content=str(reply_text),
                         raw_message=raw,
                     )
-            except Exception:
-                pass
+            except (httpx.HTTPError, json.JSONDecodeError) as exc:
+                raise ChannelTransportError(str(exc)) from exc
 
             await asyncio.sleep(interval)
 

backend/agenteval/channels/openclaw.py

@@ -12,12 +12,13 @@ Configuration keys (in channel_config, all optional — defaults come from setti
 """
 
 import asyncio
+import json
 import uuid
 from typing import Any, Optional
 
 import httpx
 
-from agenteval.channels.base import ChannelHealth, EvalChannel, Reply, SendResult
+from agenteval.channels.base import ChannelHealth, ChannelTransportError, EvalChannel, Reply, SendResult
 from agenteval.config import get_settings
 
 
@@ -50,7 +51,7 @@ class OpenClawChannel(EvalChannel):
         except Exception as exc:
             return ChannelHealth(ok=False, message=str(exc))
 
-    async def send(self, content: str, **kwargs: Any) -> SendResult:
+    async def _send(self, content: str, **kwargs: Any) -> SendResult:
         """Send a chat message to OpenClaw."""
         payload = {
             "model": self.model,
@@ -67,10 +68,10 @@ class OpenClawChannel(EvalChannel):
             # Extract the assistant message ID from the response
             msg_id = data.get("id") or str(uuid.uuid4())
             return SendResult(ok=True, question_msg_id=msg_id, raw_response=data)
-        except Exception as exc:
+        except (httpx.HTTPError, json.JSONDecodeError) as exc:
             return SendResult(ok=False, error=str(exc))
 
-    async def poll_reply(
+    async def _poll_reply(
         self,
         question_msg_id: str,
         timeout: float = 30.0,
@@ -100,8 +101,10 @@ class OpenClawChannel(EvalChannel):
                             content=reply_content,
                             raw_message={"text": reply_content, "_raw": data},
                         )
-            except Exception:
-                pass
+                elif resp.status_code != 404:
+                    raise ChannelTransportError(f"HTTP {resp.status_code}: {resp.text[:500]}")
+            except (httpx.HTTPError, json.JSONDecodeError) as exc:
+                raise ChannelTransportError(str(exc)) from exc
 
             await asyncio.sleep(interval)
 

backend/agenteval/channels/tutu.py

@@ -7,7 +7,7 @@ from typing import Any, Optional
 
 import httpx
 
-from agenteval.channels.base import ChannelHealth, EvalChannel, Reply, SendResult
+from agenteval.channels.base import ChannelHealth, ChannelTransportError, EvalChannel, Reply, SendResult
 
 
 class TutuApiChannel(EvalChannel):
@@ -65,7 +65,7 @@ class TutuApiChannel(EvalChannel):
         except Exception as exc:
             return ChannelHealth(ok=False, message=f"请求异常: {exc}")
 
-    async def send(self, content: str, **kwargs: Any) -> SendResult:
+    async def _send(self, content: str, **kwargs: Any) -> SendResult:
         """Send a text message to the configured chat contact."""
         payload = {
             "chatChannelId": self.chat_channel_id,
@@ -87,10 +87,10 @@ class TutuApiChannel(EvalChannel):
             # The reply references the sent message via metadata.questionMsgId == sent msgId.
             question_msg_id = data.get("msgId")
             return SendResult(ok=True, question_msg_id=question_msg_id, raw_response=data)
-        except Exception as exc:
+        except (httpx.HTTPError, json.JSONDecodeError) as exc:
             return SendResult(ok=False, error=f"发送异常: {exc}")
 
-    async def poll_reply(
+    async def _poll_reply(
         self,
         question_msg_id: str,
         timeout: float = 30.0,
@@ -111,7 +111,9 @@ class TutuApiChannel(EvalChannel):
                     "size": 20,
                 }
                 resp = await client.get(url, headers=self._build_headers(), params=params)
-                if resp.status_code == 200:
+                if resp.status_code != 200:
+                    raise ChannelTransportError(f"轮询失败: HTTP {resp.status_code}: {resp.text[:500]}")
+
                 data = resp.json()
                 records = data.get("data", []) if isinstance(data, dict) else []
                 for msg in records:
@@ -129,8 +131,8 @@ class TutuApiChannel(EvalChannel):
                             msg_time=msg.get("msgTime"),
                             raw_message=msg,
                         )
-            except Exception:
-                pass
+            except (httpx.HTTPError, json.JSONDecodeError) as exc:
+                raise ChannelTransportError(f"轮询异常: {exc}") from exc
 
             await asyncio.sleep(poll_interval)
 

backend/agenteval/evaluation/analysis.py

@@ -342,3 +342,21 @@ def start_campaign_analysis(campaign_id: str, *, triggered_by: str) -> None:
         campaign_id,
         lambda _cancel: execute_campaign_analysis(campaign_id, triggered_by=triggered_by),
     )
+
+
+def enqueue_campaign_analysis(campaign_id: str, *, triggered_by: str) -> None:
+    """Durably queue analysis, then launch the in-process worker."""
+    session = get_session()
+    try:
+        CampaignAnalysisRepository(session).enqueue(campaign_id, triggered_by=triggered_by)
+    finally:
+        session.close()
+    start_campaign_analysis(campaign_id, triggered_by=triggered_by)
+
+
+def resume_queued_campaign_analysis(session: Session) -> int:
+    """Re-launch analysis jobs persisted before a process interruption."""
+    rows = CampaignAnalysisRepository(session).list_queued()
+    for row in rows:
+        start_campaign_analysis(row.campaign_id, triggered_by=row.triggered_by or "manual")
+    return len(rows)

backend/agenteval/evaluation/campaign_lifecycle.py

@@ -0,0 +1,196 @@
+"""Durable Campaign creation lifecycle.
+
+This module owns the create seam: reference validation and Campaign
+persistence happen before the process-local scheduler is launched. The
+database therefore remains authoritative if task startup fails or the process
+stops between commit and launch.
+"""
+
+from dataclasses import dataclass
+from typing import Callable, Optional
+
+from sqlmodel import Session
+
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, ExplorationBudgetConfig, ExplorationSeeds
+from agenteval.storage.db import utc_now
+from agenteval.storage.model_config_repository import ModelConfigRepository
+from agenteval.storage.repository import (
+    CampaignAnalysisRepository,
+    CampaignPeriodComparisonRepository,
+    CampaignRepository,
+    CampaignWriteStatus,
+    RunRepository,
+    ScenarioRepository,
+    TargetRepository,
+)
+
+
+@dataclass(frozen=True)
+class CampaignCreateError(Exception):
+    """Validation failure translated by the HTTP adapter."""
+
+    status_code: int
+    detail: str
+
+
+@dataclass(frozen=True)
+class CampaignLifecycleError(Exception):
+    """Campaign transition failure translated by the HTTP adapter."""
+
+    status_code: int
+    detail: str
+
+
+@dataclass(frozen=True)
+class CampaignRecoverySummary:
+    """Durable work reconciled during process startup."""
+
+    interrupted_runs: int = 0
+    interrupted_analysis: int = 0
+    resumed_campaigns: int = 0
+    resumed_analysis: int = 0
+
+
+def start_campaign(
+    session: Session,
+    campaign_id: str,
+    *,
+    launch: Optional[Callable[[str, Session], object]] = None,
+) -> Optional[Campaign]:
+    """Start a planned Campaign through a conditional lifecycle write."""
+    repo = CampaignRepository(session)
+    campaign = repo.get(campaign_id)
+    if campaign is None or campaign.status in (
+        CampaignStatus.COMPLETED,
+        CampaignStatus.CANCELLED,
+        CampaignStatus.FAILED,
+    ):
+        return None
+    if campaign.status is CampaignStatus.PLANNED:
+        result = repo.start_if_planned(campaign_id, utc_now())
+        campaign = result.campaign
+        if not result.applied or campaign is None:
+            return None
+    if launch is not None and campaign.id:
+        launch(campaign.id, session)
+    return campaign
+
+
+def create_campaign(
+    session: Session,
+    *,
+    name: str,
+    target_id: str,
+    window_seconds: int,
+    time_scale: float,
+    plan: list[CampaignPlanEntry],
+    analysis_model_config_id: Optional[str] = None,
+    exploration_seeds: Optional[ExplorationSeeds] = None,
+    exploration_budget: Optional[ExplorationBudgetConfig] = None,
+    launch: Optional[Callable[[str, Session], object]] = None,
+) -> Campaign:
+    """Validate, commit and then launch one running Campaign.
+
+    ``CampaignRepository.create`` is the transaction seam. The optional
+    ``launch`` callback is invoked only after that commit, which makes this
+    function straightforward to test with a fake scheduler and ensures a
+    launch failure cannot erase the durable Campaign row.
+    """
+    if TargetRepository(session).get(target_id) is None:
+        raise CampaignCreateError(404, "target not found")
+
+    scenario_repo = ScenarioRepository(session)
+    for entry in plan:
+        if scenario_repo.get(entry.scenario_id) is None:
+            raise CampaignCreateError(404, f"scenario not found: {entry.scenario_id}")
+
+    if analysis_model_config_id is not None and ModelConfigRepository(session).get(analysis_model_config_id) is None:
+        raise CampaignCreateError(400, "analysis model config not found")
+
+    if exploration_seeds is not None and not exploration_seeds.personas and not exploration_seeds.goals:
+        exploration_seeds = None
+
+    campaign = Campaign(
+        name=name,
+        target_id=target_id,
+        window_seconds=window_seconds,
+        time_scale=time_scale,
+        plan=plan,
+        status=CampaignStatus.RUNNING,
+        started_at=utc_now(),
+        analysis_model_config_id=analysis_model_config_id,
+        exploration_seeds=exploration_seeds,
+        exploration_budget=exploration_budget,
+    )
+    campaign = CampaignRepository(session).create(campaign)
+
+    # Deliberately after the repository commit. Startup recovery can relaunch
+    # this Campaign if the process dies before the callback runs.
+    if launch is not None and campaign.id:
+        launch(campaign.id, session)
+    return campaign
+
+
+def cancel_campaign(
+    session: Session,
+    campaign_id: str,
+    *,
+    stop: Optional[Callable[[str], object]] = None,
+) -> Campaign:
+    """Cancel, settle exploration, then signal the process-local scheduler."""
+    repo = CampaignRepository(session)
+    result = repo.cancel_if_active(campaign_id, utc_now())
+    campaign = result.campaign
+    if result.status is CampaignWriteStatus.NOT_FOUND:
+        raise CampaignLifecycleError(404, "campaign not found")
+    if result.status is CampaignWriteStatus.CONFLICT or campaign is None:
+        raise CampaignLifecycleError(400, "campaign is not in a cancellable state")
+
+    if stop is not None:
+        stop(campaign_id)
+    return campaign
+
+
+def recover_campaign_runtime(session: Session, *, tick_seconds: float = 1.0) -> CampaignRecoverySummary:
+    """Reconcile all durable Campaign work and relaunch safe tasks.
+
+    The database is inspected and repaired before process-local tasks are
+    launched. Child Run reconciliation remains inside each resumed Campaign
+    loop, so a pending claim is resumed there while an orphaned running claim
+    is marked interrupted without replaying messages.
+    """
+    interrupted_runs = RunRepository(session).mark_orphans_failed()
+    interrupted_analysis = CampaignAnalysisRepository(session).mark_orphans_failed()
+    interrupted_analysis += CampaignPeriodComparisonRepository(session).mark_orphans_failed()
+
+    from agenteval.evaluation.analysis import resume_queued_campaign_analysis
+    from agenteval.evaluation.campaign_runner import resume_running_campaigns
+
+    resumed_campaigns = resume_running_campaigns(session, tick_seconds=tick_seconds)
+    resumed_analysis = resume_queued_campaign_analysis(session)
+    return CampaignRecoverySummary(
+        interrupted_runs=interrupted_runs,
+        interrupted_analysis=interrupted_analysis,
+        resumed_campaigns=resumed_campaigns,
+        resumed_analysis=resumed_analysis,
+    )
+
+
+def complete_campaign(
+    session: Session,
+    campaign_id: str,
+) -> Campaign:
+    """Atomically complete a running Campaign and settle exploration.
+
+    A competing cancellation wins because the status predicate is evaluated in
+    the database. Settlement and the terminal status share one transaction, so
+    a failure leaves the Campaign running for a later tick to retry.
+    """
+    repo = CampaignRepository(session)
+    result = repo.complete_if_running(campaign_id, utc_now())
+    campaign = result.campaign
+    if result.status is CampaignWriteStatus.NOT_FOUND:
+        raise CampaignLifecycleError(404, "campaign not found")
+    if result.status is CampaignWriteStatus.CONFLICT or campaign is None:
+        raise CampaignLifecycleError(409, "campaign is not running")
+    return campaign

backend/agenteval/evaluation/campaign_runner.py

@@ -1,10 +1,11 @@
 """Campaign runner — the side-effect shell around the pure scheduler.
 
 Given a campaign and a clock position (real elapsed seconds), it asks
-``campaign_scheduler.decide_schedule`` what is due, then spawns those child Runs
-by reusing the existing single-run execution path (``EvalEngine.run`` with an
-``existing_run`` that carries ``campaign_id``). Spawned plan-entry indices are
-persisted on the campaign so re-advancing the same clock never double-spawns.
+``campaign_scheduler.decide_schedule`` what is due, then durably claims each
+plan occurrence before reusing the existing single-run execution path
+(``EvalEngine.run`` with an ``existing_run``). Child-Run identities are the
+primary idempotency authority; the legacy campaign summary remains a fallback
+for pre-identity Runs created before this migration.
 
 This module also hosts the durable scheduler loop: a thin async shell that,
 tick by tick, maps real wall-clock elapsed time (since the campaign's persisted
@@ -20,7 +21,9 @@ from typing import Optional
 
 from sqlmodel import Session
 
-from agenteval.evaluation.analysis import resolve_analysis_model, start_campaign_analysis
+from agenteval.evaluation.analysis import enqueue_campaign_analysis, resolve_analysis_model
+from agenteval.evaluation.campaign_lifecycle import complete_campaign
+from agenteval.evaluation.campaign_lifecycle import start_campaign as start_campaign_lifecycle
 from agenteval.evaluation.campaign_scheduler import (
     TickAction,
     clock_offset,
@@ -42,6 +45,7 @@ from agenteval.models import (
 from agenteval.storage.db import get_session, utc_now
 from agenteval.storage.repository import (
     CampaignRepository,
+    CampaignRunClaimStatus,
     RunRepository,
     ScenarioRepository,
     TargetRepository,
@@ -67,10 +71,40 @@ class AdvanceResult:
     finished: bool = False
 
 
-def _spawned_indices(campaign: Campaign) -> set[int]:
-    if campaign.summary is None:
-        return set()
-    return set(campaign.summary.scheduler.spawned_indices)
+@dataclass
+class CampaignRecoveryResult:
+    """Outcome of reconciling child Runs left by a previous process."""
+
+    resumed_run_ids: list[str] = field(default_factory=list)
+    failed_run_ids: list[str] = field(default_factory=list)
+
+
+def _spawned_indices(campaign: Campaign, session: Optional[Session] = None) -> set[int]:
+    """Return plan entries completed by durable identities or legacy progress.
+
+    Once any durable occurrence exists for an entry, identities replace the
+    legacy summary for that entry. This repairs summaries written by older
+    runners after only part of a multi-occurrence entry was claimed.
+    """
+    legacy = set(campaign.summary.scheduler.spawned_indices) if campaign.summary else set()
+    if not campaign.id or session is None:
+        return legacy
+
+    occurrences: dict[int, set[int]] = {}
+    for run in RunRepository(session).list_by_campaign(campaign.id):
+        if run.campaign_plan_index is None or run.campaign_occurrence_index is None:
+            continue
+        occurrences.setdefault(run.campaign_plan_index, set()).add(run.campaign_occurrence_index)
+
+    return {
+        plan_index
+        for plan_index, entry in enumerate(campaign.plan)
+        if (
+            plan_index in occurrences
+            and set(range(entry.count)).issubset(occurrences[plan_index])
+        )
+        or (plan_index not in occurrences and plan_index in legacy)
+    }
 
 
 def current_window_offset(campaign: Campaign) -> float:
@@ -101,29 +135,42 @@ def campaign_progress(campaign: Campaign, runs: list[EvalRun]) -> dict:
     }
 
 
-async def _spawn_child_run(campaign: Campaign, scenario_id: str, session: Session) -> str:
-    """Create a Run owned by the campaign and drive it through the engine.
+async def _spawn_child_run(
+    campaign: Campaign,
+    scenario_id: str,
+    *,
+    plan_index: int,
+    occurrence_index: int,
+    session: Session,
+) -> Optional[str]:
+    """Claim one occurrence and drive its pending Run through the engine.
 
-    The Run row is created on the caller's ``session``, but the engine runs on
-    its own session (obtained internally, closed when the run ends) so one
-    child Run's lifecycle never closes the campaign's session — the same
-    isolation the single-run background task uses.
+    A rejected claim means the Campaign stopped being runnable between the
+    scheduler decision and this write; callers must stop spawning. An existing
+    pending Run is resumed, while an existing running/completed/failed Run is
+    returned without replaying external messages.
     """
-    target = TargetRepository(session).get(campaign.target_id)
     scenario = ScenarioRepository(session).get(scenario_id)
-    if not target or not scenario:
-        raise ValueError(f"campaign target or scenario missing: {campaign.target_id}/{scenario_id}")
+    if not scenario:
+        raise ValueError(f"campaign scenario missing: {scenario_id}")
 
-    run = RunRepository(session).create(
-        EvalRun(
-            target_id=campaign.target_id,
+    claim = RunRepository(session).claim_campaign_run(
+        campaign_id=campaign.id or "",
         scenario_id=scenario_id,
         scenario_version=scenario.version or 1,
-            campaign_id=campaign.id,
-            triggered_by=RunTrigger.CAMPAIGN,
-            status=RunStatus.PENDING,
-        )
+        plan_index=plan_index,
+        occurrence_index=occurrence_index,
     )
+    if claim.status in (CampaignRunClaimStatus.NOT_FOUND, CampaignRunClaimStatus.CONFLICT):
+        return None
+    run = claim.run
+    if run is None:
+        raise RuntimeError(f"Campaign child claim returned no Run: {claim.status.value}")
+
+    if run.status is RunStatus.PENDING:
+        target = TargetRepository(session).get(campaign.target_id)
+        if not target:
+            raise ValueError(f"campaign target missing: {campaign.target_id}")
         engine = EvalEngine(target=target, scenario=scenario, triggered_by=RunTrigger.CAMPAIGN)
         await engine.run(existing_run=run)
     return run.id or ""
@@ -140,9 +187,8 @@ async def advance_campaign(
 
     ``elapsed_seconds`` is real wall-clock time since the window started; it is
     mapped to a window offset via ``time_scale``. A due entry whose spawning
-    raises is skipped and recorded (so one bad entry never wedges the loop),
-    but still marked spawned to avoid infinite retries. Returns ``None`` if the
-    campaign does not exist.
+    raises is recorded but remains due until every occurrence has a durable
+    child-Run identity. Returns ``None`` if the campaign does not exist.
     """
     repo = CampaignRepository(session)
     campaign = repo.get(campaign_id)
@@ -150,7 +196,7 @@ async def advance_campaign(
         return None
 
     offset = clock_offset(elapsed_seconds=elapsed_seconds, time_scale=campaign.time_scale)
-    spawned = _spawned_indices(campaign)
+    spawned = _spawned_indices(campaign, session)
     decision = decide_schedule(
         plan=campaign.plan,
         window_seconds=campaign.window_seconds,
@@ -164,14 +210,26 @@ async def advance_campaign(
         # Cancellation stops further spawning; runs already in flight finish.
         if cancel_event is not None and cancel_event.is_set():
             break
+        claim_rejected = False
         try:
-            for _ in range(due.entry.count):
-                run_id = await _spawn_child_run(campaign, due.entry.scenario_id, session)
+            for occurrence_index in range(due.entry.count):
+                run_id = await _spawn_child_run(
+                    campaign,
+                    due.entry.scenario_id,
+                    plan_index=due.index,
+                    occurrence_index=occurrence_index,
+                    session=session,
+                )
+                if run_id is None:
+                    claim_rejected = True
+                    break
                 result.spawned_run_ids.append(run_id)
         except Exception as exc:  # one entry failing must not wedge the campaign
             errors.append({"entry_index": due.index, "error": str(exc)})
             _logger.warning("活动 %s 计划条目 %d 派生失败（跳过）: %s", campaign_id, due.index, exc)
-        spawned.add(due.index)
+        if claim_rejected:
+            break
+        spawned = _spawned_indices(campaign, session)
         # Persist progress per entry: a failure partway through a multi-entry
         # advance must never lose which entries already spawned, since restart
         # recovery reads this back from the DB. Mutate the existing summary so
@@ -196,22 +254,73 @@ def _auto_start_analysis(campaign: Campaign, session: Session) -> None:
     try:
         if resolve_analysis_model(campaign, session) is None:
             return
-        start_campaign_analysis(campaign.id, triggered_by="auto")
+        enqueue_campaign_analysis(campaign.id, triggered_by="auto")
     except Exception as exc:
         _logger.warning("活动 %s 自动分析触发失败（已跳过）: %s", campaign.id, exc)
 
 
-def _settle_exploration(campaign_id: str, session: Session) -> None:
-    """Expire dangling exploration sessions on completion.
+async def reconcile_campaign_child_runs(
+    campaign_id: str,
+    session: Session,
+    *,
+    cancel_event: Optional[asyncio.Event] = None,
+) -> Optional[CampaignRecoveryResult]:
+    """Resume safe pending claims and fail running Runs without replaying.
 
-    Never blocks completion: a settlement failure is logged and skipped, the
-    same non-blocking semantics as ``_auto_start_analysis``.
+    A pending claim has not crossed the Engine's send seam and can be resumed.
+    A running Run may already have reached the target, so recovery records an
+    explicit interruption instead of executing it again.
     """
+    campaign = CampaignRepository(session).get(campaign_id)
+    if campaign is None:
+        return None
+    if campaign.status is not CampaignStatus.RUNNING:
+        return CampaignRecoveryResult()
+
+    run_repo = RunRepository(session)
+    result = CampaignRecoveryResult(
+        failed_run_ids=run_repo.mark_campaign_running_interrupted(campaign_id)
+    )
+    for run in run_repo.list_pending_campaign_children(campaign_id):
+        if cancel_event is not None and cancel_event.is_set():
+            break
+        plan_index = run.campaign_plan_index
+        occurrence_index = run.campaign_occurrence_index
+        valid_identity = (
+            plan_index is not None
+            and occurrence_index is not None
+            and 0 <= plan_index < len(campaign.plan)
+            and 0 <= occurrence_index < campaign.plan[plan_index].count
+            and run.scenario_id == campaign.plan[plan_index].scenario_id
+        )
+        if not valid_identity:
+            failed = run_repo.mark_pending_campaign_child_interrupted(run.id or "")
+            if failed and failed.id:
+                result.failed_run_ids.append(failed.id)
+            continue
+
         try:
-        from agenteval.storage.repository import ExplorationSessionRepository
-        ExplorationSessionRepository(session).expire_running_sessions(campaign_id)
+            resumed_id = await _spawn_child_run(
+                campaign,
+                run.scenario_id,
+                plan_index=plan_index,
+                occurrence_index=occurrence_index,
+                session=session,
+            )
+            if resumed_id is None:
+                break
+            result.resumed_run_ids.append(resumed_id)
         except Exception as exc:
-        _logger.warning("活动 %s 探索会话结算失败（已跳过）: %s", campaign_id, exc)
+            fresh = run_repo.get(run.id or "")
+            if fresh is not None and fresh.status is RunStatus.PENDING:
+                failed = run_repo.mark_pending_campaign_child_interrupted(run.id or "")
+                if failed and failed.id:
+                    result.failed_run_ids.append(failed.id)
+            elif fresh is not None and fresh.status is RunStatus.FAILED and fresh.id:
+                result.failed_run_ids.append(fresh.id)
+            _logger.warning("活动 %s 子运行 %s 恢复失败: %s", campaign_id, run.id, exc)
+
+    return result
 
 
 # ── durable scheduler loop ──────────────────────────────────────────────────
@@ -229,6 +338,7 @@ async def run_campaign_loop(
     """
     session = get_session()
     try:
+        await reconcile_campaign_child_runs(campaign_id, session, cancel_event=cancel)
         while not cancel.is_set():
             repo = CampaignRepository(session)
             campaign = repo.get(campaign_id)
@@ -243,7 +353,7 @@ async def run_campaign_loop(
                 window_seconds=campaign.window_seconds,
                 time_scale=campaign.time_scale,
                 plan=campaign.plan,
-                spawned_indices=_spawned_indices(campaign),
+                spawned_indices=_spawned_indices(campaign, session),
             )
             if decision.action is TickAction.STOP:
                 return
@@ -258,13 +368,14 @@ async def run_campaign_loop(
 
             if decision.action is TickAction.COMPLETE:
                 current = repo.get(campaign_id)
-                # Cancel-race guard: only complete if still RUNNING (pure rule).
+                # The lifecycle CAS is the authoritative cancel-race guard.
                 if current and resolve_finalize(current.status) is TickAction.COMPLETE:
-                    current.status = CampaignStatus.COMPLETED
-                    current.completed_at = utc_now()
-                    repo.update(current)
-                    _settle_exploration(campaign_id, session)
-                    _auto_start_analysis(current, session)
+                    try:
+                        completed = complete_campaign(campaign_id=campaign_id, session=session)
+                    except Exception as exc:
+                        _logger.warning("活动 %s 完成迁移失败: %s", campaign_id, exc)
+                    else:
+                        _auto_start_analysis(completed, session)
                         return
 
             try:
@@ -283,15 +394,13 @@ def start_campaign(campaign_id: str, session: Session, *, tick_seconds: float =
     a PLANNED campaign gets a fresh ``started_at``; an already-RUNNING one keeps
     its original window start so recovery resumes at the correct offset.
     """
-    repo = CampaignRepository(session)
-    campaign = repo.get(campaign_id)
+    campaign = CampaignRepository(session).get(campaign_id)
     if not campaign or campaign.status in (CampaignStatus.COMPLETED, CampaignStatus.CANCELLED, CampaignStatus.FAILED):
         return None
 
-    if campaign.status == CampaignStatus.PLANNED:
-        campaign.status = CampaignStatus.RUNNING
-        campaign.started_at = utc_now()
-        repo.update(campaign)
+    started = start_campaign_lifecycle(session, campaign_id)
+    if started is None:
+        return None
 
     return campaign_registry.launch(
         campaign_id,

backend/agenteval/evaluation/engine.py

@@ -11,7 +11,7 @@ from dataclasses import dataclass
 from datetime import datetime
 from typing import Any, Callable, Optional
 
-from agenteval.channels.base import EvalChannel
+from agenteval.channels.base import EvalChannel, ExchangeOutcome, ExchangeStatus, SendResult
 from agenteval.channels.factory import ChannelFactory
 from agenteval.config import get_settings
 from agenteval.evaluation.implicit_rules import derive_implicit_rules
@@ -35,7 +35,7 @@ from agenteval.models import (
 from agenteval.services.model_configs import ModelConfigService, ModelRuntimeConfig
 from agenteval.storage.db import get_session, utc_now
 from agenteval.storage.repository import ResultRepository, RunRepository
-from agenteval.utils.llm import extract_reply_text, parse_json_from_llm_text
+from agenteval.utils.llm import parse_json_from_llm_text
 
 # Progress callbacks may be sync or async; the engine awaits the result if
 # it is a coroutine, otherwise treats it as a plain function.
@@ -61,6 +61,18 @@ def _build_send_message(content: str) -> dict[str, Any]:
     }
 
 
+def _reply_payload(outcome: ExchangeOutcome) -> Optional[dict[str, Any]]:
+    """Recover the adapter payload kept in the opaque exchange diagnostic."""
+    if not outcome.ok:
+        return None
+    diagnostic = outcome.diagnostic
+    if isinstance(diagnostic, dict):
+        reply = diagnostic.get("reply")
+        if isinstance(reply, dict):
+            return reply
+    return {"msgBody": {"content": outcome.reply_text or ""}}
+
+
 # 需要模型资源的规则类型 → 评测岗位（ModelPurpose）；其余规则无需模型
 RULE_PURPOSE = {
     "llm_score": ModelPurpose.JUDGE,
@@ -277,6 +289,26 @@ class EvalEngine:
         self.result_repo.save_turn(turn)
         return turn
 
+    def _complete_turn(self, turn: Turn, outcome: ExchangeOutcome, received_at: datetime) -> Turn:
+        """Attach exchange facts without replacing the already-sent ledger row."""
+        reply = _reply_payload(outcome)
+        if (
+            self.result_repo.update_turn_exchange(
+                turn.id or "",
+                question_msg_id=outcome.correlation_id,
+                reply=reply,
+                received_at=received_at,
+                latency_ms=outcome.latency_ms,
+            )
+            is None
+        ):
+            raise RuntimeError(f"persisted turn disappeared: {turn.id}")
+        turn.question_msg_id = outcome.correlation_id
+        turn.reply = reply
+        turn.received_at = received_at
+        turn.latency_ms = outcome.latency_ms
+        return turn
+
     async def _run_case(
         self,
         run: EvalRun,
@@ -315,8 +347,25 @@ class EvalEngine:
             )
 
             sent_at = utc_now()
-            send_result = await self.channel.send(message)
-            if not send_result.ok:
+            turn: Optional[Turn] = None
+
+            async def record_sent(send_result: SendResult) -> None:
+                nonlocal turn
+                turn = self._persist_turn(
+                    run,
+                    case,
+                    round_index,
+                    message,
+                    sent_at,
+                    question_msg_id=send_result.question_msg_id,
+                )
+
+            outcome = await self.channel.exchange(
+                message,
+                timeout=self.timeout_config.poll_reply,
+                on_sent=record_sent,
+            )
+            if outcome.status is ExchangeStatus.SEND_FAILED:
                 turn = self._persist_turn(run, case, round_index, message, sent_at)
                 await self._save_rule_results(run, case, turn, [], progress_callback)
                 await self._emit(
@@ -325,54 +374,28 @@ class EvalEngine:
                     {
                         "case_id": case.id,
                         "round": round_index,
-                        "error": send_result.error,
+                        "error": outcome.reason,
                     },
                 )
                 return failed, 0, 0
 
-            try:
-                reply = await self.channel.poll_reply(
-                    send_result.question_msg_id or "",
-                    timeout=self.timeout_config.poll_reply,
-                )
-            except Exception as poll_exc:
+            if turn is None:
+                raise RuntimeError("channel exchange succeeded without invoking the sent hook")
+
             received_at = utc_now()
-                turn = self._persist_turn(
-                    run,
-                    case,
-                    round_index,
-                    message,
-                    sent_at,
-                    question_msg_id=send_result.question_msg_id,
-                    received_at=received_at,
-                )
+            turn = self._complete_turn(turn, outcome, received_at)
+            if outcome.status is ExchangeStatus.POLL_FAILED:
                 await self._emit(
                     progress_callback,
                     "turn_error",
                     {
                         "case_id": case.id,
                         "round": round_index,
-                        "error": f"poll_reply 异常: {poll_exc}",
+                        "error": f"poll_reply 异常: {outcome.reason}",
                     },
                 )
                 return failed, 0, 0
 
-            received_at = utc_now()
-            latency_ms = None
-            if sent_at and received_at:
-                latency_ms = int((received_at - sent_at).total_seconds() * 1000)
-
-            turn = self._persist_turn(
-                run,
-                case,
-                round_index,
-                message,
-                sent_at,
-                question_msg_id=send_result.question_msg_id,
-                reply=reply.raw_message if reply else None,
-                received_at=received_at,
-                latency_ms=latency_ms,
-            )
             dialog.append(turn)
 
             await self._emit(
@@ -382,8 +405,8 @@ class EvalEngine:
                     "run_id": run.id,
                     "case_id": case.id,
                     "round": round_index,
-                    "latency_ms": latency_ms,
-                    "reply_text": extract_reply_text(reply.raw_message if reply else None),
+                    "latency_ms": outcome.latency_ms,
+                    "reply_text": outcome.reply_text or "",
                 },
             )
 

backend/agenteval/exploration/lifecycle.py

@@ -6,11 +6,11 @@
 ChannelFactory 既有接缝。
 """
 
-import json
 from typing import Any, Optional
 
 from sqlmodel import Session
 
+from agenteval.channels.base import ExchangeStatus, SendResult
 from agenteval.channels.factory import ChannelFactory
 from agenteval.config import get_settings
 from agenteval.exploration.errors import (
@@ -38,21 +38,6 @@ from agenteval.storage.repository import (
 )
 
 
-def coerce_reply_text(content: Any) -> str:
-    """Flatten a reply payload to text; tutu returns msgBody as a parsed object,
-    and str(dict) would leak a Python repr into the view."""
-    if isinstance(content, str):
-        return content
-    if isinstance(content, dict):
-        for key in ("content", "text", "message"):
-            value = content.get(key)
-            if isinstance(value, str) and value:
-                return value
-    if content is None:
-        return ""
-    return json.dumps(content, ensure_ascii=False)
-
-
 def _check_creation_guardrails(
     campaign,
     triggered_by: ExplorationTrigger,
@@ -94,14 +79,16 @@ def open_session(
     repo = ExplorationSessionRepository(db_session)
     _check_creation_guardrails(campaign, triggered_by, budget, repo)
 
-    return repo.create(ExplorationSession(
+    return repo.create(
+        ExplorationSession(
             campaign_id=campaign.id,
             target_id=campaign.target_id,
             persona=persona,
             goal=goal,
             seed_ref=seed_ref,
             triggered_by=triggered_by,
-    ))
+        )
+    )
 
 
 async def conduct_turn(db_session: Session, *, session_id: str, content: str) -> dict[str, Any]:
@@ -125,38 +112,43 @@ async def conduct_turn(db_session: Session, *, session_id: str, content: str) ->
     if not target:
         raise ExplorationNotFoundError("session target not found")
 
-    channel = ChannelFactory.create(target)
-    sent_at = utc_now()
-    try:
-        send_result = await channel.send(content)
-    except Exception as exc:  # channel adapters raise transport-specific errors
-        raise ExplorationChannelError(f"评测对象通道发送失败: {exc}") from exc
-    if not send_result.ok:
-        raise ExplorationChannelError(f"评测对象通道发送失败: {send_result.error}")
-
     message_repo = ExplorationMessageRepository(db_session)
     round_index = session_obj.turn_count + 1
+    sent_at = utc_now()
+    channel = ChannelFactory.create(target)
+
+    async def record_sent(_send_result: SendResult) -> None:
         message_repo.save_message(
             ExplorationMessage(
-            session_id=session_obj.id, round_index=round_index, role="user", content=content, created_at=sent_at
+                session_id=session_obj.id,
+                round_index=round_index,
+                role="user",
+                content=content,
+                created_at=sent_at,
             )
         )
         session_obj.turn_count = round_index
         repo.update(session_obj)
 
-    try:
-        reply = await channel.poll_reply(
-            send_result.question_msg_id,
+    outcome = await channel.exchange(
+        content,
         timeout=get_settings().poll_reply_timeout,
+        on_sent=record_sent,
     )
-    except Exception as exc:
-        raise ExplorationChannelError(f"等待评测对象回复失败: {exc}") from exc
-    if reply is None:
+    if outcome.status is ExchangeStatus.SEND_FAILED:
+        raise ExplorationChannelError(f"评测对象通道发送失败: {outcome.reason}")
+    if outcome.status is ExchangeStatus.POLL_FAILED:
+        raise ExplorationChannelError(f"等待评测对象回复失败: {outcome.reason}")
+    if outcome.status is ExchangeStatus.REPLY_TIMEOUT:
         raise ExplorationChannelError("等待评测对象回复超时")
 
     received_at = utc_now()
-    latency_ms = int((received_at - sent_at).total_seconds() * 1000)
-    reply_text = coerce_reply_text(reply.content)
+    latency_ms = (
+        outcome.latency_ms
+        if outcome.latency_ms is not None
+        else int((received_at - sent_at).total_seconds() * 1000)
+    )
+    reply_text = outcome.reply_text or ""
     message_repo.save_message(
         ExplorationMessage(
             session_id=session_obj.id,

backend/agenteval/intelligent_eval/lifecycle.py

@@ -10,11 +10,11 @@
 非法转换抛 IntelligentEvalTransitionError，路由层映射为 409。
 """
 
-import json
 from typing import Any
 
 from sqlmodel import Session
 
+from agenteval.channels.base import ExchangeStatus, SendResult
 from agenteval.channels.factory import ChannelFactory
 from agenteval.config import get_settings
 from agenteval.intelligent_eval.models import (
@@ -25,6 +25,7 @@ from agenteval.intelligent_eval.models import (
     IntelligentEvalStatus,
 )
 from agenteval.intelligent_eval.repository import (
+    CompareAndSetStatus,
     IntelligentEvalMessageRepository,
     IntelligentEvalRepository,
     IntelligentEvalSessionRepository,
@@ -36,8 +37,16 @@ from agenteval.storage.repository import TargetRepository
 _TRANSITIONS: dict[IntelligentEvalStatus, set[IntelligentEvalStatus]] = {
     IntelligentEvalStatus.DRAFT: {IntelligentEvalStatus.PLANNING},
     IntelligentEvalStatus.PLANNING: {IntelligentEvalStatus.PENDING_APPROVAL},
-    IntelligentEvalStatus.PENDING_APPROVAL: {IntelligentEvalStatus.EXECUTING, IntelligentEvalStatus.PLANNING, IntelligentEvalStatus.CANCELLED},
-    IntelligentEvalStatus.EXECUTING: {IntelligentEvalStatus.COMPLETED, IntelligentEvalStatus.CANCELLED, IntelligentEvalStatus.FAILED},
+    IntelligentEvalStatus.PENDING_APPROVAL: {
+        IntelligentEvalStatus.EXECUTING,
+        IntelligentEvalStatus.PLANNING,
+        IntelligentEvalStatus.CANCELLED,
+    },
+    IntelligentEvalStatus.EXECUTING: {
+        IntelligentEvalStatus.COMPLETED,
+        IntelligentEvalStatus.CANCELLED,
+        IntelligentEvalStatus.FAILED,
+    },
     IntelligentEvalStatus.COMPLETED: set(),
     IntelligentEvalStatus.CANCELLED: set(),
     IntelligentEvalStatus.FAILED: set(),
@@ -58,20 +67,6 @@ class IntelligentEvalChannelError(Exception):
     pass
 
 
-def coerce_reply_text(content: Any) -> str:
-    """Flatten a reply payload to text (tutu returns msgBody as a parsed object)."""
-    if isinstance(content, str):
-        return content
-    if isinstance(content, dict):
-        for key in ("content", "text", "message"):
-            value = content.get(key)
-            if isinstance(value, str) and value:
-                return value
-    if content is None:
-        return ""
-    return json.dumps(content, ensure_ascii=False)
-
-
 def _get_or_raise(repo: IntelligentEvalRepository, eval_id: str) -> IntelligentEval:
     ev = repo.get(eval_id)
     if ev is None:
@@ -79,13 +74,34 @@ def _get_or_raise(repo: IntelligentEvalRepository, eval_id: str) -> IntelligentE
     return ev
 
 
+def _resolve_write(
+    eval_id: str,
+    result,
+    *,
+    expected: IntelligentEvalStatus,
+    target: IntelligentEvalStatus,
+) -> IntelligentEval:
+    if result.status is CompareAndSetStatus.NOT_FOUND:
+        raise IntelligentEvalNotFoundError(f"intelligent eval {eval_id} not found")
+    if result.status is CompareAndSetStatus.CONFLICT:
+        raise IntelligentEvalTransitionError(
+            f"cannot transition from {expected.value} to {target.value}; state changed concurrently"
+        )
+    if result.evaluation is None:
+        raise RuntimeError(f"lifecycle write returned no evaluation: {eval_id}")
+    return result.evaluation
+
+
 def _transition(repo: IntelligentEvalRepository, ev: IntelligentEval, target: IntelligentEvalStatus) -> IntelligentEval:
     allowed = _TRANSITIONS.get(ev.status, set())
     if target not in allowed:
-        raise IntelligentEvalTransitionError(
-            f"cannot transition from {ev.status.value} to {target.value}"
+        raise IntelligentEvalTransitionError(f"cannot transition from {ev.status.value} to {target.value}")
+    result = repo._compare_and_set_status(
+        ev.id,
+        expected_status=ev.status,
+        new_status=target,
     )
-    return repo.transition_status(ev.id, target)
+    return _resolve_write(ev.id, result, expected=ev.status, target=target)
 
 
 def create_eval(
@@ -100,28 +116,37 @@ def create_eval(
     time_window_hours: int = 24,
 ) -> IntelligentEval:
     """创建智能评估并直接进入 planning 状态（draft → planning 一步完成）。"""
+    if TargetRepository(session).get(target_id) is None:
+        raise IntelligentEvalNotFoundError(f"target {target_id} not found")
+
     repo = IntelligentEvalRepository(session)
-    ev = repo.create(IntelligentEval(
+    ev = repo._create(
+        IntelligentEval(
             name=name,
             target_id=target_id,
-        status=IntelligentEvalStatus.DRAFT,
+            status=IntelligentEvalStatus.PLANNING,
             goal=goal,
             seeds=seeds,
             intent=intent,
             role_description=role_description,
             time_window_hours=time_window_hours,
-    ))
-    return _transition(repo, ev, IntelligentEvalStatus.PLANNING)
+            created_at=utc_now(),
+            updated_at=utc_now(),
+        )
+    )
+    return ev
 
 
 def submit_plan(session: Session, eval_id: str, plan: dict[str, Any]) -> IntelligentEval:
     """OpenClaw 提交粗计划：planning → pending_approval。"""
     repo = IntelligentEvalRepository(session)
-    ev = _get_or_raise(repo, eval_id)
-    ev.plan = plan
-    ev.plan_feedback = None
-    ev = repo.update(ev)
-    return _transition(repo, ev, IntelligentEvalStatus.PENDING_APPROVAL)
+    result = repo._submit_plan_if_planning(eval_id, plan)
+    return _resolve_write(
+        eval_id,
+        result,
+        expected=IntelligentEvalStatus.PLANNING,
+        target=IntelligentEvalStatus.PENDING_APPROVAL,
+    )
 
 
 def approve(session: Session, eval_id: str) -> IntelligentEval:
@@ -134,10 +159,13 @@ def approve(session: Session, eval_id: str) -> IntelligentEval:
 def reject(session: Session, eval_id: str, feedback: str) -> IntelligentEval:
     """用户打回：pending_approval → planning（附反馈）。"""
     repo = IntelligentEvalRepository(session)
-    ev = _get_or_raise(repo, eval_id)
-    ev.plan_feedback = feedback
-    ev = repo.update(ev)
-    return _transition(repo, ev, IntelligentEvalStatus.PLANNING)
+    result = repo._reject_plan_if_pending(eval_id, feedback)
+    return _resolve_write(
+        eval_id,
+        result,
+        expected=IntelligentEvalStatus.PENDING_APPROVAL,
+        target=IntelligentEvalStatus.PLANNING,
+    )
 
 
 def cancel(session: Session, eval_id: str) -> IntelligentEval:
@@ -153,10 +181,13 @@ def submit_report(session: Session, eval_id: str, report: dict[str, Any]) -> Int
     报告结构校验在路由层（pydantic），此处只负责落库与状态迁移。
     """
     repo = IntelligentEvalRepository(session)
-    ev = _get_or_raise(repo, eval_id)
-    ev.report = report
-    ev = repo.update(ev)
-    return _transition(repo, ev, IntelligentEvalStatus.COMPLETED)
+    result = repo._submit_report_if_executing(eval_id, report)
+    return _resolve_write(
+        eval_id,
+        result,
+        expected=IntelligentEvalStatus.EXECUTING,
+        target=IntelligentEvalStatus.COMPLETED,
+    )
 
 
 def get_eval(session: Session, eval_id: str) -> IntelligentEval:
@@ -187,23 +218,28 @@ def open_session(
     eval_repo = IntelligentEvalRepository(session)
     ev = _get_or_raise(eval_repo, eval_id)
     if ev.status != IntelligentEvalStatus.EXECUTING:
-        raise IntelligentEvalTransitionError(
-            f"评估不在执行中（当前 {ev.status.value}），无法创建会话"
-        )
+        raise IntelligentEvalTransitionError(f"评估不在执行中（当前 {ev.status.value}），无法创建会话")
     repo = IntelligentEvalSessionRepository(session)
-    return repo.create(IntelligentEvalSession(
+    status, created = repo._create_if_executing(
+        IntelligentEvalSession(
             eval_id=ev.id,
             target_id=ev.target_id,
             persona=persona,
             goal=goal,
             dimension=dimension,
-    ))
+        )
+    )
+    if status is CompareAndSetStatus.CONFLICT:
+        raise IntelligentEvalTransitionError("评估不在执行中，会话创建被拒绝")
+    if status is CompareAndSetStatus.NOT_FOUND or created is None:
+        raise IntelligentEvalNotFoundError(f"intelligent eval {eval_id} not found")
+    return created
 
 
-async def conduct_turn(session: Session, *, session_id: str, content: str) -> dict[str, Any]:
+async def conduct_turn(session: Session, *, eval_id: str, session_id: str, content: str) -> dict[str, Any]:
     """一轮完整问答：状态检查 → 通道往返 → 双条消息落库 → 轮次自增。"""
     repo = IntelligentEvalSessionRepository(session)
-    obj = _get_session_or_raise(repo, session_id)
+    obj = _get_owned_session_or_raise(repo, eval_id, session_id)
     if obj.status != IntelligentEvalSessionStatus.RUNNING:
         raise IntelligentEvalTransitionError("会话不在进行中，拒收消息")
 
@@ -211,58 +247,81 @@ async def conduct_turn(session: Session, *, session_id: str, content: str) -> di
     if not target:
         raise IntelligentEvalNotFoundError("session target not found")
 
-    channel = ChannelFactory.create(target)
-    sent_at = utc_now()
-    try:
-        send_result = await channel.send(content)
-    except Exception as exc:  # channel adapters raise transport-specific errors
-        raise IntelligentEvalChannelError(f"评测对象通道发送失败: {exc}") from exc
-    if not send_result.ok:
-        raise IntelligentEvalChannelError(f"评测对象通道发送失败: {send_result.error}")
-
     message_repo = IntelligentEvalMessageRepository(session)
-    message_repo.create(IntelligentEvalMessage(
-        session_id=obj.id, role="user", content=content, created_at=sent_at
-    ))
-    repo.increment_turns(obj.id)
+    sent_at = utc_now()
+    channel = ChannelFactory.create(target)
 
-    try:
-        reply = await channel.poll_reply(
-            send_result.question_msg_id,
-            timeout=get_settings().poll_reply_timeout,
+    async def record_sent(_send_result: SendResult) -> None:
+        message = IntelligentEvalMessage(
+            session_id=obj.id,
+            role="user",
+            content=content,
+            created_at=sent_at,
         )
-    except Exception as exc:
-        raise IntelligentEvalChannelError(f"等待评测对象回复失败: {exc}") from exc
-    if reply is None:
+        status = message_repo._create_user_and_increment(message)
+        if status is CompareAndSetStatus.NOT_FOUND:
+            raise IntelligentEvalNotFoundError(f"intelligent eval session {obj.id} not found")
+        if status is CompareAndSetStatus.CONFLICT:
+            raise IntelligentEvalTransitionError("会话不在进行中，拒收消息")
+
+    outcome = await channel.exchange(
+        content,
+        timeout=get_settings().poll_reply_timeout,
+        on_sent=record_sent,
+    )
+    if outcome.status is ExchangeStatus.SEND_FAILED:
+        raise IntelligentEvalChannelError(f"评测对象通道发送失败: {outcome.reason}")
+    if outcome.status is ExchangeStatus.POLL_FAILED:
+        raise IntelligentEvalChannelError(f"等待评测对象回复失败: {outcome.reason}")
+    if outcome.status is ExchangeStatus.REPLY_TIMEOUT:
         raise IntelligentEvalChannelError("等待评测对象回复超时")
 
     received_at = utc_now()
-    latency_ms = int((received_at - sent_at).total_seconds() * 1000)
-    reply_text = coerce_reply_text(reply.content)
-    message_repo.create(IntelligentEvalMessage(
+    latency_ms = (
+        outcome.latency_ms if outcome.latency_ms is not None else int((received_at - sent_at).total_seconds() * 1000)
+    )
+    reply_text = outcome.reply_text or ""
+    message_repo._create(
+        IntelligentEvalMessage(
             session_id=obj.id,
             role="assistant",
             content=reply_text,
             latency_ms=latency_ms,
             created_at=received_at,
-    ))
+        )
+    )
     return {"reply": reply_text, "latency_ms": latency_ms, "turn_count": obj.turn_count + 1}
 
 
-def close_session(session: Session, *, session_id: str, verdict: dict[str, Any]) -> IntelligentEvalSession:
+def close_session(
+    session: Session,
+    *,
+    eval_id: str,
+    session_id: str,
+    verdict: dict[str, Any],
+) -> IntelligentEvalSession:
     """关闭会话并记录结论（verdict）；仅 running 会话可关闭。"""
     repo = IntelligentEvalSessionRepository(session)
-    obj = _get_session_or_raise(repo, session_id)
-    if obj.status != IntelligentEvalSessionStatus.RUNNING:
-        raise IntelligentEvalTransitionError("会话不在进行中，无法关闭")
-    closed = repo.close(obj.id, verdict)
-    if closed is None:
+    obj = _get_owned_session_or_raise(repo, eval_id, session_id)
+    status, closed = repo._close_if_running(obj.id, verdict)
+    if status is CompareAndSetStatus.NOT_FOUND:
         raise IntelligentEvalNotFoundError(f"intelligent eval session {session_id} not found")
+    if status is CompareAndSetStatus.CONFLICT:
+        raise IntelligentEvalTransitionError("会话不在进行中，无法关闭")
+    if closed is None:
+        raise RuntimeError(f"session close returned no session: {session_id}")
     return closed
 
 
-def get_session_by_id(session: Session, session_id: str) -> IntelligentEvalSession:
-    return _get_session_or_raise(IntelligentEvalSessionRepository(session), session_id)
+def _get_owned_session_or_raise(
+    repo: IntelligentEvalSessionRepository,
+    eval_id: str,
+    session_id: str,
+) -> IntelligentEvalSession:
+    obj = _get_session_or_raise(repo, session_id)
+    if obj.eval_id != eval_id:
+        raise IntelligentEvalNotFoundError(f"intelligent eval session {session_id} not found")
+    return obj
 
 
 def list_sessions(session: Session, eval_id: str) -> list[IntelligentEvalSession]:
@@ -271,6 +330,6 @@ def list_sessions(session: Session, eval_id: str) -> list[IntelligentEvalSession
     return IntelligentEvalSessionRepository(session).list_by_eval(eval_id)
 
 
-def list_messages(session: Session, session_id: str) -> list[IntelligentEvalMessage]:
-    _get_session_or_raise(IntelligentEvalSessionRepository(session), session_id)
+def list_messages(session: Session, *, eval_id: str, session_id: str) -> list[IntelligentEvalMessage]:
+    _get_owned_session_or_raise(IntelligentEvalSessionRepository(session), eval_id, session_id)
     return IntelligentEvalMessageRepository(session).list_by_session(session_id)

backend/agenteval/intelligent_eval/read_model.py

@@ -0,0 +1,196 @@
+"""Read projections for intelligent evaluations.
+
+The lifecycle module owns writes and state transitions.  This module is the
+read seam used by HTTP adapters: it turns domain entities into stable,
+purpose-specific projections and keeps session-count/detail assembly out of
+routers.  Message bodies are intentionally not part of the detail projection;
+transcripts remain available through the dedicated messages endpoint.
+"""
+
+from collections.abc import Sequence
+from datetime import datetime
+from typing import Any
+
+from pydantic import BaseModel, Field
+from sqlmodel import Session
+
+from agenteval.intelligent_eval.models import (
+    IntelligentEval,
+    IntelligentEvalMessage,
+    IntelligentEvalSession,
+    IntelligentEvalSessionStatus,
+    IntelligentEvalStatus,
+)
+from agenteval.intelligent_eval.repository import (
+    IntelligentEvalMessageRepository,
+    IntelligentEvalRepository,
+    IntelligentEvalSessionRepository,
+)
+
+
+class IntelligentEvalProjection(BaseModel):
+    """Fields shared by list and detail representations."""
+
+    id: str
+    name: str
+    target_id: str
+    status: IntelligentEvalStatus
+    goal: str
+    seeds: dict[str, Any] = Field(default_factory=dict)
+    intent: str
+    role_description: str
+    plan: dict[str, Any] | None = None
+    plan_feedback: str | None = None
+    time_window_hours: int
+    report: dict[str, Any] | None = None
+    created_at: datetime | None = None
+    updated_at: datetime | None = None
+    started_at: datetime | None = None
+    completed_at: datetime | None = None
+
+
+class IntelligentEvalListItem(IntelligentEvalProjection):
+    """Compact projection used by collection and mutation responses."""
+
+    session_count: int = 0
+    completed_sessions: int = 0
+
+
+class IntelligentEvalSessionSummary(BaseModel):
+    """Session metadata shown in an evaluation detail view.
+
+    This deliberately contains no chat messages.  Callers that need a
+    transcript must use ``GET /{eval_id}/sessions/{session_id}/messages``.
+    """
+
+    id: str
+    eval_id: str
+    target_id: str
+    persona: dict[str, Any] = Field(default_factory=dict)
+    goal: str
+    dimension: str | None = None
+    status: IntelligentEvalSessionStatus
+    verdict: dict[str, Any] | None = None
+    turn_count: int = 0
+    created_at: datetime | None = None
+    closed_at: datetime | None = None
+
+
+class IntelligentEvalDetail(IntelligentEvalListItem):
+    """Full evaluation projection, including session metadata only."""
+
+    sessions: list[IntelligentEvalSessionSummary] = Field(default_factory=list)
+
+
+def _require_id(eval_obj: IntelligentEval) -> str:
+    if eval_obj.id is None:
+        raise ValueError("intelligent evaluation projection requires an id")
+    return eval_obj.id
+
+
+def _base_fields(eval_obj: IntelligentEval) -> dict[str, Any]:
+    """Copy domain data into the stable projection field set."""
+
+    return {
+        "id": _require_id(eval_obj),
+        "name": eval_obj.name,
+        "target_id": eval_obj.target_id,
+        "status": eval_obj.status,
+        "goal": eval_obj.goal,
+        "seeds": eval_obj.seeds,
+        "intent": eval_obj.intent,
+        "role_description": eval_obj.role_description,
+        "plan": eval_obj.plan,
+        "plan_feedback": eval_obj.plan_feedback,
+        "time_window_hours": eval_obj.time_window_hours,
+        "report": eval_obj.report,
+        "created_at": eval_obj.created_at,
+        "updated_at": eval_obj.updated_at,
+        "started_at": eval_obj.started_at,
+        "completed_at": eval_obj.completed_at,
+    }
+
+
+def _session_summary(session_obj: IntelligentEvalSession) -> IntelligentEvalSessionSummary:
+    if session_obj.id is None:
+        raise ValueError("intelligent evaluation session projection requires an id")
+    return IntelligentEvalSessionSummary.model_validate(session_obj.model_dump())
+
+
+def project_list_item(
+    eval_obj: IntelligentEval,
+    sessions: Sequence[IntelligentEvalSession],
+) -> IntelligentEvalListItem:
+    """Build a list projection from an evaluation and its session metadata."""
+
+    return IntelligentEvalListItem(
+        **_base_fields(eval_obj),
+        session_count=len(sessions),
+        completed_sessions=sum(1 for item in sessions if item.status is IntelligentEvalSessionStatus.COMPLETED),
+    )
+
+
+def project_detail(
+    eval_obj: IntelligentEval,
+    sessions: Sequence[IntelligentEvalSession],
+) -> IntelligentEvalDetail:
+    """Build a detail projection without embedding transcript messages."""
+
+    item = project_list_item(eval_obj, sessions)
+    return IntelligentEvalDetail(
+        **item.model_dump(),
+        sessions=[_session_summary(item) for item in sessions],
+    )
+
+
+class IntelligentEvalReadModel:
+    """Read interface for stable intelligent-evaluation projections.
+
+    The current implementation uses the existing session repository.  Query
+    batching can be added behind this seam without changing Router callers.
+    """
+
+    def __init__(self, session: Session):
+        self._evals = IntelligentEvalRepository(session)
+        self._sessions = IntelligentEvalSessionRepository(session)
+        self._messages = IntelligentEvalMessageRepository(session)
+
+    def list_item(self, eval_obj: IntelligentEval) -> IntelligentEvalListItem:
+        return project_list_item(eval_obj, self._sessions.list_by_eval(_require_id(eval_obj)))
+
+    def list_items(self, evals: Sequence[IntelligentEval]) -> list[IntelligentEvalListItem]:
+        """Project a collection after one batched session lookup."""
+
+        sessions_by_eval = self._sessions.list_by_evals([_require_id(item) for item in evals])
+        return [project_list_item(item, sessions_by_eval[item.id]) for item in evals if item.id is not None]
+
+    def detail_by_id(self, eval_id: str) -> IntelligentEvalDetail | None:
+        """Read one evaluation and its summaries with one snapshot query."""
+
+        snapshot = self._evals.get_with_sessions(eval_id)
+        return project_detail(*snapshot) if snapshot is not None else None
+
+    def sessions_by_eval(self, eval_id: str) -> list[IntelligentEvalSession] | None:
+        """Return session summaries, or ``None`` when the parent is unknown."""
+
+        if self._evals.get(eval_id) is None:
+            return None
+        return self._sessions.list_by_eval(eval_id)
+
+    def messages_by_session(self, eval_id: str, session_id: str) -> list[IntelligentEvalMessage] | None:
+        """Read a transcript only when the session belongs to the evaluation."""
+
+        if self._evals.get(eval_id) is None:
+            return None
+        session_obj = self._sessions.get(session_id)
+        if session_obj is None or session_obj.eval_id != eval_id:
+            return None
+        return self._messages.list_by_session(session_id)
+
+    def report_by_eval(self, eval_id: str) -> tuple[str, dict[str, Any]] | None:
+        """Return the report payload with its evaluation name for renderers."""
+
+        eval_obj = self._evals.get(eval_id)
+        if eval_obj is None or eval_obj.report is None:
+            return None
+        return eval_obj.name, eval_obj.report

backend/agenteval/intelligent_eval/repository.py

@@ -1,7 +1,14 @@
 """Repository for intelligent evaluation entities."""
 
+import json
+from collections.abc import Sequence
+from dataclasses import dataclass
+from enum import Enum
 from typing import Optional
 
+from sqlalchemy import func, literal
+from sqlalchemy import insert as sql_insert
+from sqlalchemy import update as sql_update
 from sqlmodel import Session, select
 
 from agenteval.intelligent_eval.models import (
@@ -16,10 +23,31 @@ from agenteval.storage.db import (
     IntelligentEvalMessageDB,
     IntelligentEvalSessionDB,
     get_session,
+    new_uuid,
     utc_now,
 )
 
 
+class CompareAndSetStatus(str, Enum):
+    """Outcome of a conditional intelligent-evaluation write."""
+
+    APPLIED = "applied"
+    NOT_FOUND = "not_found"
+    CONFLICT = "conflict"
+
+
+@dataclass(frozen=True)
+class CompareAndSetResult:
+    """Typed result for a status update guarded by an expected status."""
+
+    status: CompareAndSetStatus
+    evaluation: Optional[IntelligentEval] = None
+
+    @property
+    def applied(self) -> bool:
+        return self.status is CompareAndSetStatus.APPLIED
+
+
 class IntelligentEvalRepository:
     """CRUD + lifecycle operations for intelligent evaluations."""
 
@@ -77,49 +105,134 @@ class IntelligentEvalRepository:
         db = self.session.get(IntelligentEvalDB, eval_id)
         return self._from_db(db) if db else None
 
-    def create(self, obj: IntelligentEval) -> IntelligentEval:
+    def get_with_sessions(
+        self,
+        eval_id: str,
+    ) -> tuple[IntelligentEval, list[IntelligentEvalSession]] | None:
+        """Load an evaluation and all session summaries in one DB snapshot."""
+
+        statement = (
+            select(IntelligentEvalDB, IntelligentEvalSessionDB)
+            .join(
+                IntelligentEvalSessionDB,
+                IntelligentEvalSessionDB.eval_id == IntelligentEvalDB.id,
+                isouter=True,
+            )
+            .where(IntelligentEvalDB.id == eval_id)
+            .order_by(IntelligentEvalSessionDB.created_at.asc())
+        )
+        rows = self.session.exec(statement).all()
+        if not rows:
+            return None
+        evaluation = self._from_db(rows[0][0])
+        session_repo = IntelligentEvalSessionRepository(self.session)
+        sessions = [session_repo._from_db(session_db) for _, session_db in rows if session_db is not None]
+        return evaluation, sessions
+
+    def _create(self, obj: IntelligentEval) -> IntelligentEval:
         db = self._to_db(obj)
         self.session.add(db)
         self.session.commit()
         self.session.refresh(db)
         return self._from_db(db)
 
-    def update(self, obj: IntelligentEval) -> IntelligentEval:
-        db = self.session.get(IntelligentEvalDB, obj.id)
-        if not db:
-            raise ValueError(f"IntelligentEval {obj.id} not found")
-        updated = self._to_db(obj)
-        updated.id = db.id
-        # Preserve fields that _to_db doesn't set from None
-        self.session.delete(db)
-        self.session.add(updated)
-        self.session.commit()
-        self.session.refresh(updated)
-        return self._from_db(updated)
+    def _compare_and_set_status(
+        self,
+        eval_id: str,
+        *,
+        expected_status: IntelligentEvalStatus,
+        new_status: IntelligentEvalStatus,
+    ) -> CompareAndSetResult:
+        """Atomically transition status only when the expected state still holds."""
+        now = utc_now()
+        values = {
+            "status": new_status.value,
+            "updated_at": now,
+        }
+        if new_status is IntelligentEvalStatus.EXECUTING:
+            values["started_at"] = func.coalesce(IntelligentEvalDB.started_at, now)
+        if new_status in {
+            IntelligentEvalStatus.COMPLETED,
+            IntelligentEvalStatus.CANCELLED,
+            IntelligentEvalStatus.FAILED,
+        }:
+            values["completed_at"] = now
+
+        return self._compare_and_set_fields(
+            eval_id,
+            expected_status=expected_status,
+            values=values,
+        )
+
+    def _compare_and_set_fields(
+        self,
+        eval_id: str,
+        *,
+        expected_status: IntelligentEvalStatus,
+        values: dict[str, object],
+    ) -> CompareAndSetResult:
+        """Apply one conditional SQL update and classify its outcome."""
+        try:
+            result = self.session.exec(
+                sql_update(IntelligentEvalDB)
+                .where(
+                    IntelligentEvalDB.id == eval_id,
+                    IntelligentEvalDB.status == expected_status.value,
+                )
+                .values(**values)
+            )
+            if result.rowcount != 1:
+                self.session.rollback()
+                current = self.session.get(IntelligentEvalDB, eval_id)
+                status = CompareAndSetStatus.NOT_FOUND if current is None else CompareAndSetStatus.CONFLICT
+                return CompareAndSetResult(status=status)
+            self.session.commit()
+        except Exception:
+            self.session.rollback()
+            raise
 
-    def transition_status(self, eval_id: str, new_status: IntelligentEvalStatus) -> Optional[IntelligentEval]:
         db = self.session.get(IntelligentEvalDB, eval_id)
-        if not db:
-            return None
-        db.status = new_status.value
-        db.updated_at = utc_now()
-        if new_status == IntelligentEvalStatus.EXECUTING and db.started_at is None:
-            db.started_at = utc_now()
-        if new_status in (IntelligentEvalStatus.COMPLETED, IntelligentEvalStatus.CANCELLED, IntelligentEvalStatus.FAILED):
-            db.completed_at = utc_now()
-        self.session.add(db)
-        self.session.commit()
-        self.session.refresh(db)
-        return self._from_db(db)
+        if db is None:
+            return CompareAndSetResult(status=CompareAndSetStatus.NOT_FOUND)
+        return CompareAndSetResult(status=CompareAndSetStatus.APPLIED, evaluation=self._from_db(db))
 
-    def delete(self, eval_id: str) -> bool:
-        db = self.session.get(IntelligentEvalDB, eval_id)
-        if not db:
-            return False
-        self.session.delete(db)
-        self.session.commit()
-        return True
+    def _submit_plan_if_planning(self, eval_id: str, plan: dict) -> CompareAndSetResult:
+        now = utc_now()
+        return self._compare_and_set_fields(
+            eval_id,
+            expected_status=IntelligentEvalStatus.PLANNING,
+            values={
+                "status": IntelligentEvalStatus.PENDING_APPROVAL.value,
+                "plan": json.dumps(plan, ensure_ascii=False),
+                "plan_feedback": None,
+                "updated_at": now,
+            },
+        )
 
+    def _reject_plan_if_pending(self, eval_id: str, feedback: str) -> CompareAndSetResult:
+        now = utc_now()
+        return self._compare_and_set_fields(
+            eval_id,
+            expected_status=IntelligentEvalStatus.PENDING_APPROVAL,
+            values={
+                "status": IntelligentEvalStatus.PLANNING.value,
+                "plan_feedback": feedback,
+                "updated_at": now,
+            },
+        )
+
+    def _submit_report_if_executing(self, eval_id: str, report: dict) -> CompareAndSetResult:
+        now = utc_now()
+        return self._compare_and_set_fields(
+            eval_id,
+            expected_status=IntelligentEvalStatus.EXECUTING,
+            values={
+                "status": IntelligentEvalStatus.COMPLETED.value,
+                "report": json.dumps(report, ensure_ascii=False),
+                "updated_at": now,
+                "completed_at": now,
+            },
+        )
 
 class IntelligentEvalSessionRepository:
     """CRUD for intelligent eval sessions."""
@@ -150,11 +263,31 @@ class IntelligentEvalSessionRepository:
         )
         return [self._from_db(r) for r in self.session.exec(statement).all()]
 
+    def list_by_evals(self, eval_ids: Sequence[str]) -> dict[str, list[IntelligentEvalSession]]:
+        """Load sessions for several evaluations with one query.
+
+        The result contains an empty list for every requested evaluation. This
+        keeps read-model assembly deterministic while avoiding one count query
+        per row in the evaluation list.
+        """
+
+        grouped = {eval_id: [] for eval_id in eval_ids}
+        if not grouped:
+            return grouped
+        statement = (
+            select(IntelligentEvalSessionDB)
+            .where(IntelligentEvalSessionDB.eval_id.in_(tuple(grouped)))
+            .order_by(IntelligentEvalSessionDB.created_at.asc())
+        )
+        for row in self.session.exec(statement).all():
+            grouped.setdefault(row.eval_id, []).append(self._from_db(row))
+        return grouped
+
     def get(self, session_id: str) -> Optional[IntelligentEvalSession]:
         db = self.session.get(IntelligentEvalSessionDB, session_id)
         return self._from_db(db) if db else None
 
-    def create(self, obj: IntelligentEvalSession) -> IntelligentEvalSession:
+    def _create(self, obj: IntelligentEvalSession) -> IntelligentEvalSession:
         db = IntelligentEvalSessionDB(
             id=obj.id,
             eval_id=obj.eval_id,
@@ -172,25 +305,92 @@ class IntelligentEvalSessionRepository:
         self.session.refresh(db)
         return self._from_db(db)
 
-    def close(self, session_id: str, verdict: dict, status: IntelligentEvalSessionStatus = IntelligentEvalSessionStatus.COMPLETED) -> Optional[IntelligentEvalSession]:
-        db = self.session.get(IntelligentEvalSessionDB, session_id)
-        if not db:
-            return None
-        db.status = status.value
-        db.set_verdict(verdict)
-        db.closed_at = utc_now()
-        self.session.add(db)
+    def _create_if_executing(
+        self,
+        obj: IntelligentEvalSession,
+    ) -> tuple[CompareAndSetStatus, Optional[IntelligentEvalSession]]:
+        """Insert a session only while its parent evaluation is executing."""
+        session_id = obj.id or new_uuid()
+        now = obj.created_at or utc_now()
+        statement = sql_insert(IntelligentEvalSessionDB).from_select(
+            [
+                "id",
+                "eval_id",
+                "target_id",
+                "persona",
+                "goal",
+                "dimension",
+                "status",
+                "turn_count",
+                "created_at",
+            ],
+            select(
+                literal(session_id),
+                IntelligentEvalDB.id,
+                IntelligentEvalDB.target_id,
+                literal(json.dumps(obj.persona, ensure_ascii=False)),
+                literal(obj.goal),
+                literal(obj.dimension),
+                literal(IntelligentEvalSessionStatus.RUNNING.value),
+                literal(0),
+                literal(now),
+            ).where(
+                IntelligentEvalDB.id == obj.eval_id,
+                IntelligentEvalDB.status == IntelligentEvalStatus.EXECUTING.value,
+            ),
+        )
+        try:
+            result = self.session.exec(statement)
+            if result.rowcount != 1:
+                self.session.rollback()
+                evaluation = self.session.get(IntelligentEvalDB, obj.eval_id)
+                status = CompareAndSetStatus.NOT_FOUND if evaluation is None else CompareAndSetStatus.CONFLICT
+                return status, None
             self.session.commit()
-        self.session.refresh(db)
-        return self._from_db(db)
+        except Exception:
+            self.session.rollback()
+            raise
 
-    def increment_turns(self, session_id: str) -> None:
         db = self.session.get(IntelligentEvalSessionDB, session_id)
-        if db:
-            db.turn_count += 1
-            self.session.add(db)
-            self.session.commit()
+        if db is None:
+            return CompareAndSetStatus.NOT_FOUND, None
+        return CompareAndSetStatus.APPLIED, self._from_db(db)
 
+    def _close_if_running(
+        self,
+        session_id: str,
+        verdict: dict,
+        status: IntelligentEvalSessionStatus = IntelligentEvalSessionStatus.COMPLETED,
+    ) -> tuple[CompareAndSetStatus, Optional[IntelligentEvalSession]]:
+        """Close exactly one running session without overwriting a race winner."""
+        now = utc_now()
+        try:
+            result = self.session.exec(
+                sql_update(IntelligentEvalSessionDB)
+                .where(
+                    IntelligentEvalSessionDB.id == session_id,
+                    IntelligentEvalSessionDB.status == IntelligentEvalSessionStatus.RUNNING.value,
+                )
+                .values(
+                    status=status.value,
+                    verdict=json.dumps(verdict, ensure_ascii=False),
+                    closed_at=now,
+                )
+            )
+            if result.rowcount != 1:
+                self.session.rollback()
+                current = self.session.get(IntelligentEvalSessionDB, session_id)
+                outcome = CompareAndSetStatus.NOT_FOUND if current is None else CompareAndSetStatus.CONFLICT
+                return outcome, None
+            self.session.commit()
+        except Exception:
+            self.session.rollback()
+            raise
+
+        db = self.session.get(IntelligentEvalSessionDB, session_id)
+        if db is None:
+            return CompareAndSetStatus.NOT_FOUND, None
+        return CompareAndSetStatus.APPLIED, self._from_db(db)
 
 class IntelligentEvalMessageRepository:
     """CRUD for intelligent eval session messages."""
@@ -216,7 +416,7 @@ class IntelligentEvalMessageRepository:
         )
         return [self._from_db(r) for r in self.session.exec(statement).all()]
 
-    def create(self, obj: IntelligentEvalMessage) -> IntelligentEvalMessage:
+    def _create(self, obj: IntelligentEvalMessage) -> IntelligentEvalMessage:
         db = IntelligentEvalMessageDB(
             id=obj.id,
             session_id=obj.session_id,
@@ -229,3 +429,31 @@ class IntelligentEvalMessageRepository:
         self.session.commit()
         self.session.refresh(db)
         return self._from_db(db)
+
+    def _create_user_and_increment(self, obj: IntelligentEvalMessage) -> CompareAndSetStatus:
+        """Persist the sent user message and consume one turn atomically."""
+        session_db = self.session.get(IntelligentEvalSessionDB, obj.session_id)
+        if session_db is None:
+            return CompareAndSetStatus.NOT_FOUND
+        if session_db.status != IntelligentEvalSessionStatus.RUNNING.value:
+            return CompareAndSetStatus.CONFLICT
+
+        message_db = IntelligentEvalMessageDB(
+            id=obj.id,
+            session_id=obj.session_id,
+            role="user",
+            content=obj.content,
+            latency_ms=None,
+            created_at=obj.created_at,
+        )
+        try:
+            session_db.turn_count += 1
+            self.session.add(session_db)
+            self.session.add(message_db)
+            self.session.commit()
+            self.session.refresh(message_db)
+        except Exception:
+            self.session.rollback()
+            raise
+        obj.id = message_db.id
+        return CompareAndSetStatus.APPLIED

backend/agenteval/models.py

@@ -222,6 +222,9 @@ class EvalRun(BaseModel):
     scenario_version: int = 1
     # 归属的评估活动（Campaign）；手动/单次运行为空
     campaign_id: Optional[str] = None
+    # Campaign 计划条目的持久化身份；历史 Run 保持为空，不做回填。
+    campaign_plan_index: Optional[int] = Field(default=None, ge=0)
+    campaign_occurrence_index: Optional[int] = Field(default=None, ge=0)
     status: RunStatus = RunStatus.PENDING
     triggered_by: RunTrigger = RunTrigger.MANUAL
     started_at: Optional[datetime] = None

backend/agenteval/storage/db.py

@@ -6,6 +6,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, Optional
 
+import sqlalchemy as sa
 from sqlalchemy.pool import StaticPool
 from sqlmodel import Field, Relationship, Session, SQLModel, create_engine
 
@@ -348,12 +349,23 @@ class EvalRunDB(SQLModel, table=True):
     """Database table for evaluation runs."""
 
     __tablename__ = "eval_runs"
+    __table_args__ = (
+        sa.Index(
+            "uq_eval_runs_campaign_occurrence",
+            "campaign_id",
+            "campaign_plan_index",
+            "campaign_occurrence_index",
+            unique=True,
+        ),
+    )
 
     id: Optional[str] = Field(default_factory=new_uuid, primary_key=True)
     target_id: Optional[str] = Field(default=None, foreign_key="eval_targets.id")
     scenario_id: Optional[str] = Field(default=None, foreign_key="scenarios.id")
     scenario_version: int = Field(default=1)
     campaign_id: Optional[str] = Field(default=None, foreign_key="campaigns.id")
+    campaign_plan_index: Optional[int] = Field(default=None, ge=0)
+    campaign_occurrence_index: Optional[int] = Field(default=None, ge=0)
     status: str = "pending"
     triggered_by: str = Field(default="manual")
     started_at: Optional[datetime] = Field(default_factory=utc_now)

backend/agenteval/storage/repository.py

@@ -1,12 +1,29 @@
 """Repository layer for database access."""
 
+from dataclasses import dataclass
 from datetime import datetime
+from enum import Enum
 from typing import Generic, Optional, TypeVar
 
+from sqlalchemy import insert as sql_insert
+from sqlalchemy import literal
+from sqlalchemy import update as sql_update
+from sqlalchemy.exc import IntegrityError
 from sqlmodel import Session, select
 
 from agenteval.exploration.models import ExplorationMessage, ExplorationSession, ExplorationSessionStatus
-from agenteval.models import Campaign, CampaignStatus, CampaignSummary, Case, EvalResult, EvalRun, EvalTarget, Scenario
+from agenteval.models import (
+    Campaign,
+    CampaignStatus,
+    CampaignSummary,
+    Case,
+    EvalResult,
+    EvalRun,
+    EvalTarget,
+    RunStatus,
+    RunTrigger,
+    Scenario,
+)
 from agenteval.services.model_configs import ModelConfigService
 from agenteval.storage.db import (
     CampaignAnalysisDB,
@@ -20,6 +37,7 @@ from agenteval.storage.db import (
     ScenarioDB,
     TurnDB,
     get_session,
+    new_uuid,
     utc_now,
 )
 from agenteval.storage.model_config_repository import ScenarioModelBindingRepository
@@ -253,12 +271,122 @@ class ScenarioRepository(BaseRepository[Scenario, ScenarioDB]):
         return {sid: name for sid, name in self.session.exec(select(ScenarioDB.id, ScenarioDB.name)).all()}
 
 
+class CampaignRunClaimStatus(str, Enum):
+    """Outcome of a durable Campaign child-Run claim."""
+
+    CLAIMED = "claimed"
+    EXISTING = "existing"
+    NOT_FOUND = "not_found"
+    CONFLICT = "conflict"
+
+
+@dataclass(frozen=True)
+class CampaignRunClaimResult:
+    """Typed result for the idempotent child-Run claim seam."""
+
+    status: CampaignRunClaimStatus
+    run: Optional[EvalRun] = None
+
+    @property
+    def accepted(self) -> bool:
+        return self.status in (CampaignRunClaimStatus.CLAIMED, CampaignRunClaimStatus.EXISTING)
+
+    @property
+    def created(self) -> bool:
+        return self.status is CampaignRunClaimStatus.CLAIMED
+
+
 class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
     """Repository for evaluation runs."""
 
     _table = EvalRunDB
     _order_by = "started_at"
 
+    def claim_campaign_run(
+        self,
+        *,
+        campaign_id: str,
+        scenario_id: str,
+        scenario_version: int,
+        plan_index: int,
+        occurrence_index: int,
+    ) -> "CampaignRunClaimResult":
+        """Claim one durable child Run occurrence while its Campaign is running.
+
+        The conditional ``INSERT ... SELECT`` makes the Campaign status check
+        part of the write.  The composite unique index turns retries or a
+        competing scheduler into an ``EXISTING`` result instead of a second
+        pending Run.  No external messages are sent by this operation.
+        """
+        if plan_index < 0 or occurrence_index < 0:
+            raise ValueError("plan_index and occurrence_index must be non-negative")
+
+        run_id = new_uuid()
+        now = utc_now()
+        statement = sql_insert(EvalRunDB).from_select(
+            [
+                "id",
+                "target_id",
+                "scenario_id",
+                "scenario_version",
+                "campaign_id",
+                "campaign_plan_index",
+                "campaign_occurrence_index",
+                "status",
+                "triggered_by",
+                "started_at",
+            ],
+            select(
+                literal(run_id),
+                CampaignDB.target_id,
+                literal(scenario_id),
+                literal(scenario_version),
+                CampaignDB.id,
+                literal(plan_index),
+                literal(occurrence_index),
+                literal(RunStatus.PENDING.value),
+                literal(RunTrigger.CAMPAIGN.value),
+                literal(now),
+            ).where(
+                CampaignDB.id == campaign_id,
+                CampaignDB.status == CampaignStatus.RUNNING.value,
+                CampaignDB.target_id.is_not(None),
+            ),
+        )
+
+        try:
+            result = self.session.exec(statement)
+            if result.rowcount == 1:
+                self.session.commit()
+                db = self.session.get(EvalRunDB, run_id)
+                return CampaignRunClaimResult(CampaignRunClaimStatus.CLAIMED, self._from_db(db))
+            self.session.rollback()
+        except IntegrityError:
+            self.session.rollback()
+        except Exception:
+            self.session.rollback()
+            raise
+
+        existing = self._get_campaign_identity(campaign_id, plan_index, occurrence_index)
+        campaign = self.session.get(CampaignDB, campaign_id)
+        if campaign is None:
+            return CampaignRunClaimResult(CampaignRunClaimStatus.NOT_FOUND)
+        if campaign.status != CampaignStatus.RUNNING.value:
+            return CampaignRunClaimResult(CampaignRunClaimStatus.CONFLICT)
+        if existing is not None:
+            return CampaignRunClaimResult(CampaignRunClaimStatus.EXISTING, self._from_db(existing))
+        return CampaignRunClaimResult(CampaignRunClaimStatus.CONFLICT)
+
+    def _get_campaign_identity(
+        self, campaign_id: str, plan_index: int, occurrence_index: int
+    ) -> Optional[EvalRunDB]:
+        statement = select(EvalRunDB).where(
+            EvalRunDB.campaign_id == campaign_id,
+            EvalRunDB.campaign_plan_index == plan_index,
+            EvalRunDB.campaign_occurrence_index == occurrence_index,
+        )
+        return self.session.exec(statement).first()
+
     def _to_db(self, run: EvalRun) -> EvalRunDB:
         db = EvalRunDB(
             id=run.id,
@@ -266,6 +394,8 @@ class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
             scenario_id=run.scenario_id,
             scenario_version=run.scenario_version,
             campaign_id=run.campaign_id,
+            campaign_plan_index=run.campaign_plan_index,
+            campaign_occurrence_index=run.campaign_occurrence_index,
             status=run.status.value,
             triggered_by=run.triggered_by.value,
             started_at=run.started_at,
@@ -282,6 +412,8 @@ class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
             scenario_id=db.scenario_id,
             scenario_version=db.scenario_version or 1,
             campaign_id=db.campaign_id,
+            campaign_plan_index=db.campaign_plan_index,
+            campaign_occurrence_index=db.campaign_occurrence_index,
             status=db.status,
             triggered_by=db.triggered_by or "manual",
             started_at=db.started_at,
@@ -294,23 +426,81 @@ class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
         return [self._from_db(r) for r in self.session.exec(statement).all()]
 
     def mark_orphans_failed(self) -> int:
-        """服务启动时清理：把遗留的 running/pending 运行标记为 failed。
+        """Fail process-orphaned Runs while preserving recoverable child claims.
 
-        评测任务是进程内 asyncio 任务，服务重启后不会恢复；不清理则这些
-        运行永远停留在 running（僵尸运行）。
+        Every RUNNING Run may already have sent messages and must not replay.
+        A PENDING Run is preserved only when it has a complete Campaign plan
+        identity and its parent Campaign is still running; the Campaign loop
+        can safely resume those claims before any external send occurred.
         """
         statement = select(EvalRunDB).where(EvalRunDB.status.in_(["running", "pending"]))  # type: ignore[attr-defined]
-        orphans = self.session.exec(statement).all()
-        for db in orphans:
-            db.status = "failed"
+        candidates = self.session.exec(statement).all()
+        failed: list[EvalRunDB] = []
+        for db in candidates:
+            if db.status == RunStatus.PENDING.value and self._is_recoverable_campaign_pending(db):
+                continue
+            self._set_interrupted(db)
+            failed.append(db)
+        if failed:
+            self.session.commit()
+        return len(failed)
+
+    def _is_recoverable_campaign_pending(self, db: EvalRunDB) -> bool:
+        if (
+            db.campaign_id is None
+            or db.campaign_plan_index is None
+            or db.campaign_occurrence_index is None
+        ):
+            return False
+        campaign = self.session.get(CampaignDB, db.campaign_id)
+        return campaign is not None and campaign.status == CampaignStatus.RUNNING.value
+
+    def _set_interrupted(self, db: EvalRunDB) -> None:
+        db.status = RunStatus.FAILED.value
         db.completed_at = db.completed_at or utc_now()
         summary = db.get_summary() or {}
         summary["error"] = {"code": "interrupted", "message": "服务重启导致评测中断"}
         db.set_summary(summary)
         self.session.add(db)
-        if orphans:
+
+    def list_pending_campaign_children(self, campaign_id: str) -> list[EvalRun]:
+        """Return durable pending claims in deterministic plan order."""
+        statement = (
+            select(EvalRunDB)
+            .where(
+                EvalRunDB.campaign_id == campaign_id,
+                EvalRunDB.campaign_plan_index.is_not(None),
+                EvalRunDB.campaign_occurrence_index.is_not(None),
+                EvalRunDB.status == RunStatus.PENDING.value,
+            )
+            .order_by(EvalRunDB.campaign_plan_index, EvalRunDB.campaign_occurrence_index)
+        )
+        return [self._from_db(db) for db in self.session.exec(statement).all()]
+
+    def mark_campaign_running_interrupted(self, campaign_id: str) -> list[str]:
+        """Fail running child Runs from a previous process without replaying."""
+        statement = select(EvalRunDB).where(
+            EvalRunDB.campaign_id == campaign_id,
+            EvalRunDB.campaign_plan_index.is_not(None),
+            EvalRunDB.campaign_occurrence_index.is_not(None),
+            EvalRunDB.status == RunStatus.RUNNING.value,
+        )
+        rows = list(self.session.exec(statement).all())
+        for db in rows:
+            self._set_interrupted(db)
+        if rows:
             self.session.commit()
-        return len(orphans)
+        return [db.id or "" for db in rows]
+
+    def mark_pending_campaign_child_interrupted(self, run_id: str) -> Optional[EvalRun]:
+        """Fail one still-pending child claim that cannot be safely recovered."""
+        db = self.session.get(EvalRunDB, run_id)
+        if db is None or db.status != RunStatus.PENDING.value:
+            return self._from_db(db) if db is not None else None
+        self._set_interrupted(db)
+        self.session.commit()
+        self.session.refresh(db)
+        return self._from_db(db)
 
     def update(self, run: EvalRun) -> Optional[EvalRun]:
         existing = self.session.get(EvalRunDB, run.id)
@@ -320,6 +510,8 @@ class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
         existing.scenario_id = run.scenario_id
         existing.scenario_version = run.scenario_version
         existing.campaign_id = run.campaign_id
+        existing.campaign_plan_index = run.campaign_plan_index
+        existing.campaign_occurrence_index = run.campaign_occurrence_index
         existing.status = run.status.value
         existing.triggered_by = run.triggered_by.value
         existing.completed_at = run.completed_at
@@ -339,6 +531,26 @@ class RunRepository(BaseRepository[EvalRun, EvalRunDB]):
         return [_result_from_db(r) for r in self.session.exec(statement).all()]
 
 
+class CampaignWriteStatus(str, Enum):
+    """Outcome of a conditional Campaign lifecycle write."""
+
+    APPLIED = "applied"
+    NOT_FOUND = "not_found"
+    CONFLICT = "conflict"
+
+
+@dataclass(frozen=True)
+class CampaignWriteResult:
+    """Typed result returned by Campaign compare-and-set operations."""
+
+    status: CampaignWriteStatus
+    campaign: Optional[Campaign] = None
+
+    @property
+    def applied(self) -> bool:
+        return self.status is CampaignWriteStatus.APPLIED
+
+
 class CampaignRepository(BaseRepository[Campaign, CampaignDB]):
     """Repository for evaluation campaigns (评估活动)."""
 
@@ -434,6 +646,97 @@ class CampaignRepository(BaseRepository[Campaign, CampaignDB]):
         self.session.refresh(db)
         return self._from_db(db)
 
+    def _compare_and_set_status(
+        self,
+        campaign_id: str,
+        *,
+        expected_statuses: tuple[CampaignStatus, ...],
+        target_status: CampaignStatus,
+        at: datetime,
+        settle_exploration: bool = False,
+    ) -> CampaignWriteResult:
+        """Apply one conditional Campaign transition and optional settlement.
+
+        Final Campaign status and expiration of running exploration sessions
+        share one transaction. A settlement failure therefore rolls the
+        Campaign back to its prior runnable state.
+        """
+
+        values: dict[str, object] = {"status": target_status.value}
+        if target_status is CampaignStatus.RUNNING:
+            values["started_at"] = at
+        if target_status in (CampaignStatus.CANCELLED, CampaignStatus.COMPLETED, CampaignStatus.FAILED):
+            values["completed_at"] = at
+
+        statement = (
+            sql_update(CampaignDB)
+            .where(
+                CampaignDB.id == campaign_id,
+                CampaignDB.status.in_([status.value for status in expected_statuses]),
+            )
+            .values(**values)
+        )
+        try:
+            result = self.session.exec(statement)
+            if result.rowcount != 1:
+                self.session.rollback()
+                db = self.session.get(CampaignDB, campaign_id)
+                status = CampaignWriteStatus.NOT_FOUND if db is None else CampaignWriteStatus.CONFLICT
+                return CampaignWriteResult(status=status, campaign=self._from_db(db) if db else None)
+
+            if settle_exploration:
+                self.session.exec(
+                    sql_update(ExplorationSessionDB)
+                    .where(
+                        ExplorationSessionDB.campaign_id == campaign_id,
+                        ExplorationSessionDB.status == ExplorationSessionStatus.RUNNING.value,
+                    )
+                    .values(status=ExplorationSessionStatus.EXPIRED.value, closed_at=at)
+                )
+            self.session.commit()
+        except Exception:
+            self.session.rollback()
+            raise
+
+        self.session.expire_all()
+        db = self.session.get(CampaignDB, campaign_id)
+        return CampaignWriteResult(
+            status=CampaignWriteStatus.APPLIED,
+            campaign=self._from_db(db) if db else None,
+        )
+
+    def cancel_if_active(self, campaign_id: str, at: datetime) -> CampaignWriteResult:
+        """Atomically cancel and expire running exploration sessions."""
+
+        return self._compare_and_set_status(
+            campaign_id,
+            expected_statuses=(CampaignStatus.PLANNED, CampaignStatus.RUNNING),
+            target_status=CampaignStatus.CANCELLED,
+            at=at,
+            settle_exploration=True,
+        )
+
+    def complete_if_running(self, campaign_id: str, at: datetime) -> CampaignWriteResult:
+        """Atomically complete and expire running exploration sessions."""
+
+        return self._compare_and_set_status(
+            campaign_id,
+            expected_statuses=(CampaignStatus.RUNNING,),
+            target_status=CampaignStatus.COMPLETED,
+            at=at,
+            settle_exploration=True,
+        )
+
+    def start_if_planned(self, campaign_id: str, at: datetime) -> CampaignWriteResult:
+        """Atomically stamp a planned Campaign as running."""
+
+        return self._compare_and_set_status(
+            campaign_id,
+            expected_statuses=(CampaignStatus.PLANNED,),
+            target_status=CampaignStatus.RUNNING,
+            at=at,
+        )
+
     def save_scheduler_state(self, campaign_id: str, summary: CampaignSummary) -> None:
         """窄口径调度进度持久化：只写 summary 列，不覆写并发的状态 / 水位变更。"""
         db = self.session.get(CampaignDB, campaign_id)
@@ -485,6 +788,24 @@ class CampaignAnalysisRepository(AsyncJobRepository[CampaignAnalysisDB]):
 
     _table = CampaignAnalysisDB
 
+    def enqueue(self, campaign_id: str, *, triggered_by: str = "manual") -> CampaignAnalysisDB:
+        """Persist an analysis job before launching its process-local task."""
+        row = self.get_by_campaign(campaign_id)
+        if row is None:
+            row = CampaignAnalysisDB(campaign_id=campaign_id)
+        row.status = "queued"
+        row.result = None
+        row.error = None
+        row.triggered_by = triggered_by
+        row.updated_at = utc_now()
+        self.session.add(row)
+        self.session.commit()
+        self.session.refresh(row)
+        return row
+
+    def list_queued(self) -> list[CampaignAnalysisDB]:
+        return list(self.session.exec(select(CampaignAnalysisDB).where(CampaignAnalysisDB.status == "queued")).all())
+
     def upsert(
         self,
         campaign_id: str,
@@ -584,6 +905,34 @@ class ResultRepository:
         self.session.refresh(db)
         return db
 
+    def update_turn_exchange(
+        self,
+        turn_id: str,
+        *,
+        question_msg_id: Optional[str],
+        reply: Optional[dict],
+        received_at: Optional[datetime],
+        latency_ms: Optional[int],
+    ) -> Optional[TurnDB]:
+        """Attach exchange facts to an already-persisted Turn.
+
+        This is deliberately narrower than updating a whole Turn: the sent
+        message, case identity, ordering and send timestamp remain untouched.
+        A caller can therefore commit the sent fact before polling and safely
+        complete the same ledger row after a reply, timeout or poll failure.
+        """
+        db = self.session.get(TurnDB, turn_id)
+        if db is None:
+            return None
+        db.question_msg_id = question_msg_id
+        db.set_reply(reply)
+        db.received_at = received_at
+        db.latency_ms = latency_ms
+        self.session.add(db)
+        self.session.commit()
+        self.session.refresh(db)
+        return db
+
     def save_result(self, result: EvalResult) -> EvalResult:
         db = _result_to_db(result)
         self.session.add(db)

backend/agenteval/web/app.py

@@ -10,11 +10,6 @@ from fastapi.responses import FileResponse, JSONResponse
 
 from agenteval.config import get_settings
 from agenteval.storage.db import get_session, init_db
-from agenteval.storage.repository import (
-    CampaignAnalysisRepository,
-    CampaignPeriodComparisonRepository,
-    RunRepository,
-)
 from agenteval.version import get_build_info, get_version
 from agenteval.web.deps import require_api_key
 from agenteval.web.routers import (
@@ -41,19 +36,25 @@ async def lifespan(_: FastAPI):
     try:
         session = get_session()
         try:
-            count = RunRepository(session).mark_orphans_failed()
-            if count:
-                logging.getLogger("agenteval").warning("启动清理：%d 个中断的运行已标记为 failed", count)
-            llm_orphans = CampaignAnalysisRepository(session).mark_orphans_failed()
-            llm_orphans += CampaignPeriodComparisonRepository(session).mark_orphans_failed()
-            if llm_orphans:
-                logging.getLogger("agenteval").warning("启动清理：%d 条中断的分析/对比已标记为 failed", llm_orphans)
-            # 据库恢复所有未完成的评估活动，重建其调度循环（不重复派生已到点条目）
-            from agenteval.evaluation.campaign_runner import resume_running_campaigns
+            from agenteval.evaluation.campaign_lifecycle import recover_campaign_runtime
 
-            resumed = resume_running_campaigns(session)
-            if resumed:
-                logging.getLogger("agenteval").warning("启动恢复：%d 个进行中的评估活动已续跑", resumed)
+            recovery = recover_campaign_runtime(session)
+            if recovery.interrupted_runs:
+                logging.getLogger("agenteval").warning(
+                    "启动清理：%d 个中断的运行已标记为 failed", recovery.interrupted_runs
+                )
+            if recovery.interrupted_analysis:
+                logging.getLogger("agenteval").warning(
+                    "启动清理：%d 条中断的分析/对比已标记为 failed", recovery.interrupted_analysis
+                )
+            if recovery.resumed_campaigns:
+                logging.getLogger("agenteval").warning(
+                    "启动恢复：%d 个进行中的评估活动已续跑", recovery.resumed_campaigns
+                )
+            if recovery.resumed_analysis:
+                logging.getLogger("agenteval").warning(
+                    "启动恢复：%d 条排队中的活动分析已续跑", recovery.resumed_analysis
+                )
         finally:
             session.close()
     except Exception as exc:

backend/agenteval/web/routers/campaigns.py

@@ -11,7 +11,10 @@ from fastapi import APIRouter, Body, Depends, HTTPException, Response
 from pydantic import BaseModel, Field
 from sqlmodel import Session
 
-from agenteval.evaluation.analysis import resolve_analysis_model, start_campaign_analysis
+from agenteval.evaluation.analysis import enqueue_campaign_analysis, resolve_analysis_model
+from agenteval.evaluation.campaign_lifecycle import CampaignCreateError, CampaignLifecycleError
+from agenteval.evaluation.campaign_lifecycle import cancel_campaign as cancel_campaign_lifecycle
+from agenteval.evaluation.campaign_lifecycle import create_campaign as create_campaign_lifecycle
 from agenteval.evaluation.campaign_runner import campaign_progress, request_cancel, start_campaign
 from agenteval.evaluation.comparison import (
     ComparisonError,
@@ -27,9 +30,8 @@ from agenteval.evaluation.report import (
 )
 from agenteval.evaluation.report_render import render_campaign_markdown
 from agenteval.exploration.summary import summarize_campaign_exploration
-from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, ExplorationBudgetConfig, ExplorationSeeds
-from agenteval.storage.db import iso_utc, utc_now
-from agenteval.storage.model_config_repository import ModelConfigRepository
+from agenteval.models import CampaignPlanEntry, CampaignStatus, ExplorationBudgetConfig, ExplorationSeeds
+from agenteval.storage.db import iso_utc
 from agenteval.storage.repository import (
     CampaignAnalysisRepository,
     CampaignRepository,
@@ -71,55 +73,30 @@ async def create_campaign(
     request: CreateCampaignRequest,
     session: Session = Depends(get_db),
 ) -> dict:
-    if not TargetRepository(session).get(request.target_id):
-        raise HTTPException(status_code=404, detail="target not found")
-
-    scenario_repo = ScenarioRepository(session)
-    for entry in request.plan:
-        if not scenario_repo.get(entry.scenario_id):
-            raise HTTPException(
-                status_code=404,
-                detail=f"scenario not found: {entry.scenario_id}",
-            )
-
-    if request.analysis_model_config_id is not None:
-        if not ModelConfigRepository(session).get(request.analysis_model_config_id):
-            raise HTTPException(status_code=400, detail="analysis model config not found")
-
-    seeds = request.exploration_seeds
-    if seeds is not None and not seeds.personas and not seeds.goals:
-        seeds = None  # 种子留空 = 该活动不参与探索
-
-    campaign = Campaign(
+    try:
+        campaign = create_campaign_lifecycle(
+            session,
             name=request.name,
             target_id=request.target_id,
             window_seconds=request.window_seconds,
             time_scale=request.time_scale,
             plan=request.plan,
             analysis_model_config_id=request.analysis_model_config_id,
-        exploration_seeds=seeds,
+            exploration_seeds=request.exploration_seeds,
             exploration_budget=request.exploration_budget,
+            launch=start_campaign,
         )
-    repo = CampaignRepository(session)
-    campaign = repo.create(campaign)
-    # Kick off the durable loop; it moves the campaign into RUNNING.
-    start_campaign(campaign.id, session)
-    return (repo.get(campaign.id) or campaign).model_dump()
+    except CampaignCreateError as exc:
+        raise HTTPException(status_code=exc.status_code, detail=exc.detail) from exc
+    return campaign.model_dump()
 
 
 @router.post("/{campaign_id}/cancel")
 async def cancel_campaign(campaign_id: str, session: Session = Depends(get_db)) -> dict:
-    repo = CampaignRepository(session)
-    campaign = repo.get(campaign_id)
-    if not campaign:
-        raise HTTPException(status_code=404, detail="campaign not found")
-    if campaign.status not in (CampaignStatus.PLANNED, CampaignStatus.RUNNING):
-        raise HTTPException(status_code=400, detail="campaign is not in a cancellable state")
-
-    campaign = repo.mark_cancelled(campaign_id, utc_now()) or campaign
-    request_cancel(campaign_id)
-    from agenteval.storage.repository import ExplorationSessionRepository
-    ExplorationSessionRepository(session).expire_running_sessions(campaign_id)
+    try:
+        campaign = cancel_campaign_lifecycle(session, campaign_id, stop=request_cancel)
+    except CampaignLifecycleError as exc:
+        raise HTTPException(status_code=exc.status_code, detail=exc.detail) from exc
     return campaign.model_dump()
 
 
@@ -213,7 +190,7 @@ async def trigger_campaign_analysis(campaign_id: str, session: Session = Depends
             status_code=400,
             detail="未配置分析模型：请在模型配置中心将某个 chat 配置设为「分析默认」，或为该活动指定分析模型",
         )
-    start_campaign_analysis(campaign_id, triggered_by="manual")
+    enqueue_campaign_analysis(campaign_id, triggered_by="manual")
     return {"status": "generating"}
 
 

backend/agenteval/web/routers/intelligent_evals.py

@@ -17,8 +17,8 @@ from agenteval.intelligent_eval.lifecycle import (
     IntelligentEvalNotFoundError,
     IntelligentEvalTransitionError,
 )
+from agenteval.intelligent_eval.read_model import IntelligentEvalReadModel
 from agenteval.intelligent_eval.report import ReportModel, render_report_markdown
-from agenteval.intelligent_eval.repository import IntelligentEvalSessionRepository
 from agenteval.web.deps import get_db
 
 router = APIRouter()
@@ -69,15 +69,15 @@ def _translate(exc: Exception) -> HTTPException:
 
 
 def _eval_response(ev, session: Session) -> dict:
-    data = ev.model_dump(mode="json")
-    sessions = IntelligentEvalSessionRepository(session).list_by_eval(ev.id)
-    data["session_count"] = len(sessions)
-    data["completed_sessions"] = sum(1 for s in sessions if s.status.value == "completed")
-    return data
+    """Serialize one stable intelligent-evaluation read projection."""
+
+    projection = IntelligentEvalReadModel(session).list_item(ev)
+    return projection.model_dump(mode="json")
 
 
 @router.post("")
 async def create_eval(request: CreateEvalRequest, session: Session = Depends(get_db)) -> dict:
+    try:
         ev = lifecycle.create_eval(
             session,
             name=request.name,
@@ -88,22 +88,24 @@ async def create_eval(request: CreateEvalRequest, session: Session = Depends(get
             role_description=request.role_description,
             time_window_hours=request.time_window_hours,
         )
+    except IntelligentEvalNotFoundError as exc:
+        raise HTTPException(status_code=404, detail=str(exc)) from exc
     return _eval_response(ev, session)
 
 
 @router.get("")
 async def list_evals(session: Session = Depends(get_db)) -> dict:
     evals = lifecycle.list_evals(session)
-    return {"intelligent_evals": [_eval_response(ev, session) for ev in evals]}
+    reader = IntelligentEvalReadModel(session)
+    return {"intelligent_evals": [item.model_dump(mode="json") for item in reader.list_items(evals)]}
 
 
 @router.get("/{eval_id}")
 async def get_eval(eval_id: str, session: Session = Depends(get_db)) -> dict:
-    try:
-        ev = lifecycle.get_eval(session, eval_id)
-    except IntelligentEvalNotFoundError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
-    return _eval_response(ev, session)
+    projection = IntelligentEvalReadModel(session).detail_by_id(eval_id)
+    if projection is None:
+        raise HTTPException(status_code=404, detail=f"intelligent eval {eval_id} not found")
+    return projection.model_dump(mode="json")
 
 
 @router.put("/{eval_id}/plan")
@@ -153,31 +155,24 @@ async def submit_report(eval_id: str, request: SubmitReportRequest, session: Ses
 
 @router.get("/{eval_id}/report")
 async def get_report(eval_id: str, session: Session = Depends(get_db)) -> dict:
-    try:
-        ev = lifecycle.get_eval(session, eval_id)
-    except IntelligentEvalNotFoundError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
-    if ev.report is None:
+    report = IntelligentEvalReadModel(session).report_by_eval(eval_id)
+    if report is None:
         raise HTTPException(status_code=404, detail="report not submitted yet")
-    return ev.report
+    return report[1]
 
 
 @router.get("/{eval_id}/report/markdown", response_class=PlainTextResponse)
 async def get_report_markdown(eval_id: str, session: Session = Depends(get_db)) -> PlainTextResponse:
-    try:
-        ev = lifecycle.get_eval(session, eval_id)
-    except IntelligentEvalNotFoundError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
-    if ev.report is None:
+    report = IntelligentEvalReadModel(session).report_by_eval(eval_id)
+    if report is None:
         raise HTTPException(status_code=404, detail="report not submitted yet")
-    markdown = render_report_markdown(ev.report, name=ev.name, eval_id=ev.id)
+    name, payload = report
+    markdown = render_report_markdown(payload, name=name, eval_id=eval_id)
     return PlainTextResponse(markdown, media_type="text/markdown; charset=utf-8")
 
 
 @router.post("/{eval_id}/sessions")
-async def create_session(
-    eval_id: str, request: CreateSessionRequest, session: Session = Depends(get_db)
-) -> dict:
+async def create_session(eval_id: str, request: CreateSessionRequest, session: Session = Depends(get_db)) -> dict:
     try:
         obj = lifecycle.open_session(
             session,
@@ -193,27 +188,23 @@ async def create_session(
 
 @router.get("/{eval_id}/sessions")
 async def list_sessions(eval_id: str, session: Session = Depends(get_db)) -> dict:
-    try:
-        sessions = lifecycle.list_sessions(session, eval_id)
-    except IntelligentEvalNotFoundError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    sessions = IntelligentEvalReadModel(session).sessions_by_eval(eval_id)
+    if sessions is None:
+        raise HTTPException(status_code=404, detail=f"intelligent eval {eval_id} not found")
     return {"sessions": [s.model_dump(mode="json") for s in sessions]}
 
 
-def _get_owned_session(eval_id: str, session_id: str, session: Session):
-    obj = lifecycle.get_session_by_id(session, session_id)
-    if obj.eval_id != eval_id:
-        raise IntelligentEvalNotFoundError(f"intelligent eval session {session_id} not found")
-    return obj
-
-
 @router.post("/{eval_id}/sessions/{session_id}/messages")
 async def send_message(
     eval_id: str, session_id: str, request: SendMessageRequest, session: Session = Depends(get_db)
 ) -> dict:
     try:
-        _get_owned_session(eval_id, session_id, session)
-        return await lifecycle.conduct_turn(session, session_id=session_id, content=request.content)
+        return await lifecycle.conduct_turn(
+            session,
+            eval_id=eval_id,
+            session_id=session_id,
+            content=request.content,
+        )
     except (
         IntelligentEvalNotFoundError,
         IntelligentEvalTransitionError,
@@ -227,8 +218,12 @@ async def close_session(
     eval_id: str, session_id: str, request: CloseSessionRequest, session: Session = Depends(get_db)
 ) -> dict:
     try:
-        _get_owned_session(eval_id, session_id, session)
-        obj = lifecycle.close_session(session, session_id=session_id, verdict=request.verdict)
+        obj = lifecycle.close_session(
+            session,
+            eval_id=eval_id,
+            session_id=session_id,
+            verdict=request.verdict,
+        )
     except (IntelligentEvalNotFoundError, IntelligentEvalTransitionError) as exc:
         raise _translate(exc) from exc
     return obj.model_dump(mode="json")
@@ -236,9 +231,7 @@ async def close_session(
 
 @router.get("/{eval_id}/sessions/{session_id}/messages")
 async def list_messages(eval_id: str, session_id: str, session: Session = Depends(get_db)) -> dict:
-    try:
-        _get_owned_session(eval_id, session_id, session)
-        messages = lifecycle.list_messages(session, session_id)
-    except IntelligentEvalNotFoundError as exc:
-        raise HTTPException(status_code=404, detail=str(exc)) from exc
+    messages = IntelligentEvalReadModel(session).messages_by_session(eval_id, session_id)
+    if messages is None:
+        raise HTTPException(status_code=404, detail=f"intelligent eval session {session_id} not found")
     return {"messages": [m.model_dump(mode="json") for m in messages]}

deploy/volcengine-102/Dockerfile

@@ -44,4 +44,7 @@ RUN mkdir -p ./data/scenarios ./data/reports
 EXPOSE 8000
 HEALTHCHECK --interval=30s --timeout=5s --retries=3 \
     CMD curl -f http://localhost:8000/api/health || exit 1
-CMD ["sh", "-c", "python -c 'from agenteval.storage.db import init_db; init_db()' && alembic stamp head && agenteval server start --host 0.0.0.0 --port 8000"]
+COPY scripts/production-entrypoint.sh ./scripts/production-entrypoint.sh
+RUN chmod +x ./scripts/production-entrypoint.sh
+
+CMD ["./scripts/production-entrypoint.sh"]

deploy/volcengine-102/docker-compose.yml

@@ -11,7 +11,8 @@ services:
     ports:
       - "8002:8000"
     env_file:
-      - .env
+      - path: .env
+        required: false
     volumes:
       - agenteval_data:/app/data
     restart: unless-stopped

docs/adr/0003-campaign-phased-static-then-adaptive.md

@@ -41,3 +41,16 @@
 - **产出接入**：报告新增"探索发现"维度 + 喂 v0.7 分析 + Markdown 导出；周期对比暂不扩展，待口径稳定。
 
 词汇表更新见 `CONTEXT.md`「探索式评测」章节。
+
+## v1 持久化与恢复修订（2026-08-06）
+
+为使静态活动真正可跨进程恢复，v1 的耐久边界进一步明确：
+
+- **数据库是唯一权威**：`TaskRegistry` 只保存进程内任务句柄，不决定活动或 Run 的持久状态。
+- **子 Run 身份持久化**：新建 child Run 保存 `campaign_id`、计划条目索引和 occurrence 索引；三者由 Campaign 范围内唯一索引约束。历史 Run 保持空值，不回填。
+- **Claim 先于执行**：调度器对每个 occurrence 先执行条件 claim，再进入 `EvalEngine`。重复 tick、并发调度或重启不会创建第二个 Run；取消后的 Campaign 不允许新 claim，已启动 Run 可继续完成。
+- **恢复边界**：持久化为 `pending` 且仍属于 running Campaign 的子 Run 可以恢复；进程中断遗留的 `running` 子 Run 统一标记为 `failed/interrupted`，禁止重放可能已经发送的外部消息。
+- **生命周期 CAS**：创建、启动、取消和完成均通过生命周期模块与条件更新完成。取消竞态优先于完成；Campaign 终态与 running 探索会话结算在同一事务提交，任一步失败均保持活动与探索会话为 `running`，由后续调度 tick 重试。
+- **分析任务耐久化**：活动分析先写入 `queued` 再启动进程内 worker；重启恢复 queued 任务，遗留 `generating` 任务标记为中断失败。分析失败或重复执行不改变 Campaign 完成状态。
+
+启动恢复顺序固定为：清理中断 Run 与 LLM 任务 → 重建 running Campaign 调度循环（其中包含 child Run reconciliation）→ 重启 queued 分析任务。该顺序保证恢复动作只依据已提交的数据库事实，不依赖上一次进程的内存状态。

docs/agents/domain.md

@@ -0,0 +1,14 @@
+# Domain Docs
+
+本仓库采用单一上下文领域文档布局。
+
+## 探索前必读
+
+- 根目录 `CONTEXT.md`：领域词汇与边界。
+- `docs/adr/`：与当前修改相关的架构决策。
+
+文件缺失时直接继续，不要预先创建空文档。只有在术语或决策确实需要澄清时，才通过领域建模流程补充。
+
+## 使用规则
+
+Issue、规格、测试和代码应使用 `CONTEXT.md` 定义的领域术语，避免引入同义但含义不一致的新词。若修改与现有 ADR 冲突，必须显式指出冲突及重新决策的理由，不能静默覆盖。

docs/agents/issue-tracker.md

@@ -0,0 +1,18 @@
+# Issue Tracker
+
+本仓库使用自托管 Gitea Issues：
+
+- 实例：`https://git.solahqb22.cn`
+- 仓库：`solahqb/AgentEvalTool`
+- Web：`https://git.solahqb22.cn/solahqb/AgentEvalTool/issues`
+- API：`/api/v1/repos/solahqb/AgentEvalTool/issues`
+
+## 工作流
+
+读取、创建和更新 Issue 时优先使用 Gitea REST API。认证 token 必须通过临时文件或进程环境注入，使用请求头传递；禁止写入仓库、remote URL、命令参数、日志或 Issue 正文。
+
+创建前按标题搜索现有 Issue，避免重复。Issue 正文使用 Markdown；阶段进展写入评论；完成后更新状态为 closed。需要表达依赖时使用 Gitea Issue dependency API，而不是仅在正文中描述。
+
+若本机已配置 `tea`，可使用非交互 flags；否则使用 `curl` 调用当前实例 Swagger 定义的 REST 接口。不要用 GitHub `gh` 命令或假设 GitHub 字段与 Gitea 完全一致。
+
+Pull Request 默认不作为 triage 输入来源。

docs/agents/triage-labels.md

@@ -0,0 +1,11 @@
+# Triage Labels
+
+| 技能角色 | Gitea 标签 | 含义 |
+| --- | --- | --- |
+| `needs-triage` | `needs-triage` | 等待维护者评估 |
+| `needs-info` | `needs-info` | 等待报告者补充信息 |
+| `ready-for-agent` | `ready-for-agent` | 规格完整，可由智能体执行 |
+| `ready-for-human` | `ready-for-human` | 需要人工处理 |
+| `wontfix` | `wontfix` | 决定不处理 |
+
+技能提到 triage 角色时，使用表中对应的 Gitea 标签。若仓库标签命名调整，应同步更新此表。

docs/deployment/README.md

@@ -23,6 +23,7 @@ Docker 部署是推荐的生产环境部署方式，提供以下优势：
 
 **部署文档**：
 - [t480 服务器部署指南](t480-v1.0.md) - 针对 t480 测试服务器的具体部署步骤
+- [volcengine-102 正式线部署指南](volcengine-102-v1.0.md) - 版本化发布、备份、迁移、健康检查与回滚
 
 ### 2.2 裸机部署
 

docs/deployment/volcengine-102-v1.0.md

@@ -0,0 +1,100 @@
+# volcengine-102 正式线部署指南
+
+**状态**：已部署并通过正式域名验收（最后验证：2026-08-09）。
+**Compose**：`deploy/volcengine-102/docker-compose.yml`
+**入口脚本**：`scripts/deploy-volcengine-102.sh`
+
+## 1. 正式入口与流量路径
+
+- 正式地址：<https://agenteval.solahqb22.cn/>
+- DNS：`agenteval.solahqb22.cn` A 记录指向 `47.111.21.147`（`sola-aliyun-147`）。
+- HTTPS：147 Nginx 终止 TLS，并反代到 `101.96.206.102:8002`（`sola-volcengine-102`）。
+- WebSocket：Nginx 为 `/ws` 和 `/openclaw` 配置 HTTP/1.1 Upgrade。
+- 网络边界：102 的 `8002` 仅允许 `47.111.21.147/32` 访问，公网直连超时属于预期行为。
+- 证书：Let's Encrypt，实测有效期为 2026-07-20 至 2026-10-18；147 的 `certbot.timer` 已启用并处于 active 状态。
+
+```mermaid
+flowchart LR
+    Browser[浏览器] -->|HTTPS 443| DNS[agenteval.solahqb22.cn]
+    DNS --> Nginx[aliyun-147 Nginx]
+    Nginx -->|HTTP 8002，仅白名单| App[volcengine-102 AgentEval]
+    App --> OpenClaw[openclaw-eval]
+```
+
+## 2. 部署边界
+
+正式线使用两个容器：`agenteval`（端口 `8002`）和 `openclaw-eval`（端口 `28789`）。SQLite 数据保存在 Docker named volume 中；生产 `.env` 不从代码仓库同步，必须预先放在远端：
+
+```text
+/opt/sola/AgentEvalTool/deploy/volcengine-102/.env
+```
+
+环境变量模板见 [`deploy/volcengine-102/.env.example`](../../deploy/volcengine-102/.env.example)。API Key、OpenClaw token 和 webhook secret 不得写入 Git；正式域名和公开网络拓扑应记录在本指南中。
+
+## 3. 首次准备
+
+在正式主机完成 Docker/Compose、SSH 和镜像仓库访问配置，然后建立远端目录并准备 `.env`。本地只需要配置 SSH 别名或地址：
+
+```bash
+export AGENTEVAL_PROD_HOST=sola-volcengine-102
+export AGENTEVAL_PROD_DIR=/opt/sola/AgentEvalTool
+export AGENTEVAL_PROD_BACKUP_DIR=/var/backups/agenteval
+```
+
+首次正式发布前必须确认：
+
+- 生产 `.env` 已填写非空 `AGENTEVAL_API_KEY`、`AGENTEVAL_SECRET_KEY` 和 OpenClaw token。
+- 生产域名已加入 `AGENTEVAL_ALLOWED_ORIGINS`。
+- 正式主机可拉取 `registry.solahqb22.cn/sola/openclaw:latest`。
+- Docker named volume 和备份目录具备写权限。
+
+## 4. 发布流程
+
+发布必须基于已提交的 Git 版本；脚本会拒绝存在 tracked 未提交修改的工作区。推荐使用版本号加 commit 作为不可变镜像标签：
+
+```bash
+scripts/deploy-volcengine-102.sh --dry-run --tag 0.8.0-6248568
+scripts/deploy-volcengine-102.sh --tag 0.8.0-6248568
+```
+
+脚本依次执行：SSH 检查 → 版本读取 → 远端 `.env` 检查 → 通过 `git archive HEAD` 生成并同步已提交源代码 → 备份数据 volume → 远端构建带版本元数据的镜像 → 启动双容器 → 等待健康接口 → 校验版本/commit → 鉴权 API 冒烟。
+
+后端启动入口 [`scripts/production-entrypoint.sh`](../../scripts/production-entrypoint.sh) 的规则是：
+
+- 全新数据库：创建初始 SQLModel schema，记录当前 Alembic 基线，再执行后续迁移。
+- 旧版无 `alembic_version` 的数据库：保留旧数据，标记为 pre-Alembic base，再依次执行全部迁移。
+- 已有 Alembic 版本的数据库：只执行 `alembic upgrade head`。
+
+## 5. 回滚
+
+回滚只切换应用镜像，不自动降级数据库：
+
+```bash
+scripts/deploy-volcengine-102.sh --rollback 0.8.0-previous
+```
+
+每次发布和回滚前都会在远端备份 named volume 到 `AGENTEVAL_PROD_BACKUP_DIR`。如果某个版本包含不可逆数据迁移，应先停止服务、恢复匹配的数据备份，再启动旧镜像；不得只回滚代码而忽略数据库版本。
+
+## 6. 发布后检查
+
+```bash
+ssh "$AGENTEVAL_PROD_HOST" \
+  'cd /opt/sola/AgentEvalTool && docker compose -f deploy/volcengine-102/docker-compose.yml ps'
+```
+
+至少通过正式域名验证：`/api/health`、`/api/targets`、`/api/scenarios`、`/api/runs`、`/api/campaigns`、`/api/intelligent-evals`、`/api/model-configs` 和 `/openclaw/`。`/ws` 应返回 `101 Switching Protocols`。发现异常时保留部署输出、容器日志和备份文件，再执行回滚判断。
+
+## 7. 线路关系
+
+```mermaid
+flowchart LR
+    Commit[已提交 Git 版本] --> Dev[scripts/deploy-t480.sh]
+    Dev --> DevCheck[t480 健康检查与业务冒烟]
+    DevCheck --> Tag[Gitea Tag / 发布审批]
+    Tag --> Prod[scripts/deploy-volcengine-102.sh]
+    Prod --> Backup[备份 named volume]
+    Backup --> Migrate[Alembic upgrade head]
+    Migrate --> Smoke[正式线 API / OpenClaw 冒烟]
+    Smoke --> Live[正式服务]
+    Smoke --> Rollback[旧镜像 + 数据备份恢复]
+```

frontend/web/package.json

@@ -6,7 +6,8 @@
   "scripts": {
     "dev": "vite",
     "build": "tsc && vite build",
-    "preview": "vite preview"
+    "preview": "vite preview",
+    "test": "vitest run"
   },
   "dependencies": {
     "@ant-design/charts": "^2.6.7",
@@ -22,10 +23,14 @@
     "zustand": "^5.0.14"
   },
   "devDependencies": {
+    "@testing-library/jest-dom": "^7.0.0",
+    "@testing-library/react": "^16.3.2",
     "@types/react": "^18.3.3",
     "@types/react-dom": "^18.3.0",
     "@vitejs/plugin-react": "^4.3.1",
+    "jsdom": "^29.1.1",
     "typescript": "^5.5.3",
-    "vite": "^5.3.3"
+    "vite": "^5.3.3",
+    "vitest": "^4.1.10"
   }
 }

frontend/web/src/api.ts

@@ -702,6 +702,7 @@ export interface IntelligentEval {
   completed_at: string | null
   session_count: number
   completed_sessions: number
+  sessions?: IntelligentEvalSession[]
 }
 
 export interface CreateIntelligentEvalPayload {

frontend/web/src/components/intelligent_eval/EvalDetail.tsx

@@ -1,12 +1,11 @@
-import { useCallback, useEffect, useState } from 'react'
+import { useState } from 'react'
 import {
   Alert, Button, Card, Col, Descriptions, Empty, Input, Modal, Popconfirm, Progress, Row, Space, Spin, Tag, message,
 } from 'antd'
 import { FileTextOutlined, StopOutlined } from '@ant-design/icons'
-import { intelligentEvalsApi, type IntelligentEval, type IntelligentEvalSession } from '../../api'
+import { intelligentEvalsApi, type IntelligentEval } from '../../api'
 import { colors } from '../../tokens'
 import { formatDateTime, shortDateTime } from '../../utils/date'
-import { usePolling } from '../../hooks/usePolling'
 import { EVAL_STATUS, SESSION_STATUS } from './status'
 
 const sectionCard: React.CSSProperties = { marginBottom: 16 }
@@ -71,22 +70,9 @@ export default function EvalDetail({ ev, targetName, onOpenReport, onChanged }:
   const [busy, setBusy] = useState(false)
   const [rejectOpen, setRejectOpen] = useState(false)
   const [feedback, setFeedback] = useState('')
-  const [sessions, setSessions] = useState<IntelligentEvalSession[] | null>(null)
   const meta = EVAL_STATUS[ev.status] ?? { label: ev.status, color: 'default' }
   const showSessions = ev.status === 'executing' || ev.status === 'completed'
-
-  const loadSessions = useCallback(() => {
-    intelligentEvalsApi.listSessions(ev.id)
-      .then((res) => setSessions(res.data.sessions))
-      .catch(() => undefined)
-  }, [ev.id])
-
-  useEffect(() => {
-    if (!showSessions) { setSessions(null); return }
-    loadSessions()
-  }, [showSessions, loadSessions])
-
-  usePolling(loadSessions, 5000, showSessions && ev.status === 'executing')
+  const sessions = showSessions ? ev.sessions ?? [] : []
 
   const runAction = async (fn: () => Promise<unknown>, okMsg: string) => {
     setBusy(true)
@@ -211,13 +197,12 @@ export default function EvalDetail({ ev, targetName, onOpenReport, onChanged }:
                 />
               )}
             >
-              {sessions === null && <Spin size="small" />}
-              {sessions !== null && sessions.length === 0 && (
+              {sessions.length === 0 && (
                 <div style={{ fontSize: 13, color: colors.textSecondary }}>
                   等待 OpenClaw 按粗计划的时间分布创建会话…
                 </div>
               )}
-              {sessions?.map((s) => {
+              {sessions.map((s) => {
                 const sMeta = SESSION_STATUS[s.status]
                 return (
                   <div

frontend/web/src/components/intelligent_eval/EvalReport.tsx

@@ -107,29 +107,11 @@ interface EvalReportProps {
 }
 
 export default function EvalReport({ ev, onBack }: EvalReportProps) {
-  const [report, setReport] = useState<IntelligentEvalReport | null>(ev.report)
-  const [reportLoading, setReportLoading] = useState(!ev.report)
-  const [sessions, setSessions] = useState<IntelligentEvalSession[] | null>(null)
-  const [sessionsLoading, setSessionsLoading] = useState(true)
   const [exporting, setExporting] = useState(false)
-
-  useEffect(() => {
-    let cancelled = false
-    setReport(ev.report)
-    setReportLoading(!ev.report)
-    if (!ev.report) {
-      intelligentEvalsApi.getReport(ev.id)
-        .then((res) => { if (!cancelled) setReport(res.data) })
-        .catch(() => undefined)
-        .finally(() => { if (!cancelled) setReportLoading(false) })
-    }
-    setSessionsLoading(true)
-    intelligentEvalsApi.listSessions(ev.id)
-      .then((res) => { if (!cancelled) setSessions(res.data.sessions) })
-      .catch(() => undefined)
-      .finally(() => { if (!cancelled) setSessionsLoading(false) })
-    return () => { cancelled = true }
-  }, [ev.id, ev.report])
+  const report: IntelligentEvalReport | null = ev.report
+  const reportLoading = false
+  const sessions: IntelligentEvalSession[] = ev.sessions ?? []
+  const sessionsLoading = false
 
   const exportMarkdown = async () => {
     setExporting(true)

frontend/web/src/pages/IntelligentEvals.tsx

@@ -9,7 +9,6 @@ import PageWrapper from '../components/PageWrapper'
 import EvalDetail from '../components/intelligent_eval/EvalDetail'
 import EvalReport from '../components/intelligent_eval/EvalReport'
 import { EVAL_STATUS } from '../components/intelligent_eval/status'
-import { usePolling } from '../hooks/usePolling'
 import { useResource } from '../hooks/useResource'
 import {
   intelligentEvalsApi, targetsApi,
@@ -17,6 +16,7 @@ import {
 } from '../api'
 import { colors } from '../tokens'
 import { formatDateTime } from '../utils/date'
+import { useIntelligentEvalRead } from '../read/useIntelligentEvalRead'
 
 type DrawerView = 'detail' | 'report'
 
@@ -33,29 +33,19 @@ interface CreateFormValues {
 export default function IntelligentEvalsPage() {
   const [drawerView, setDrawerView] = useState<DrawerView>('detail')
   const [selectedId, setSelectedId] = useState<string | null>(null)
-  const [detailTick, setDetailTick] = useState(0)
   const [createOpen, setCreateOpen] = useState(false)
   const [submitting, setSubmitting] = useState(false)
   const [form] = Form.useForm<CreateFormValues>()
 
-  const { data: evals, loading, reload } = useResource(
-    () => intelligentEvalsApi.list().then((r) => r.data.intelligent_evals),
-    { tabPath: '/intelligent-evals' },
-  )
+  const { list, detail, reloadList, reloadDetail } = useIntelligentEvalRead(selectedId)
+  const evals = list.value
+  const loading = list.phase === 'loading'
   const { data: targets } = useResource(
     () => targetsApi.list().then((r) => r.data),
     { tabPath: '/intelligent-evals' },
   )
 
-  const { data: selected, reload: reloadSelected } = useResource(
-    () => (selectedId ? intelligentEvalsApi.get(selectedId).then((r) => r.data) : Promise.resolve(null)),
-    { deps: [selectedId, detailTick] },
-  )
-
-  const pollActive = selected != null
-    && (selected.status === 'planning' || selected.status === 'executing' || selected.status === 'pending_approval')
-
-  usePolling(() => void reloadSelected(true), 5000, pollActive)
+  const selected = selectedId != null && detail.value?.id === selectedId ? detail.value : null
 
   const targetName = (id: string) =>
     targets?.find((t) => t.id === id)?.name ?? id.slice(0, 8)
@@ -73,7 +63,7 @@ export default function IntelligentEvalsPage() {
   const closeDrawer = () => {
     setSelectedId(null)
     setDrawerView('detail')
-    void reload()
+    void reloadList()
   }
 
   const submitCreate = async () => {
@@ -108,7 +98,7 @@ export default function IntelligentEvalsPage() {
       message.success('已创建，OpenClaw 开始规划')
       setCreateOpen(false)
       form.resetFields()
-      void reload()
+      void reloadList()
     } finally {
       setSubmitting(false)
     }
@@ -163,7 +153,7 @@ export default function IntelligentEvalsPage() {
       fullHeight
       extra={
         <Space>
-          <Button icon={<ReloadOutlined />} onClick={() => reload()} />
+          <Button icon={<ReloadOutlined />} onClick={() => void reloadList()} />
           <Button type="primary" icon={<PlusOutlined />} onClick={() => setCreateOpen(true)}>
             新建智能评估
           </Button>
@@ -198,7 +188,7 @@ export default function IntelligentEvalsPage() {
             ev={selected}
             targetName={targetName(selected.target_id)}
             onOpenReport={() => setDrawerView('report')}
-            onChanged={() => setDetailTick((t) => t + 1)}
+            onChanged={() => void reloadDetail()}
           />
         )}
       </Drawer>

frontend/web/src/read/intelligentEval.test.ts

@@ -0,0 +1,74 @@
+import { describe, expect, it } from 'vitest'
+import {
+  initialIntelligentEvalReadState,
+  intelligentEvalReadReducer,
+  type IntelligentEvalReadState,
+} from './intelligentEval'
+
+const evaluation = { id: 'eval-1', name: '评估', status: 'planning' } as never
+
+describe('intelligent evaluation read state', () => {
+  it('represents initial loading and a complete list snapshot', () => {
+    const loading = intelligentEvalReadReducer(initialIntelligentEvalReadState, { type: 'list_requested' })
+    expect(loading.list.phase).toBe('loading')
+
+    const ready = intelligentEvalReadReducer(loading, { type: 'list_succeeded', value: [evaluation] })
+    expect(ready.list).toEqual({ phase: 'ready', value: [evaluation], error: null })
+  })
+
+  it('keeps the last list snapshot during silent refresh and failure', () => {
+    const ready: IntelligentEvalReadState = {
+      ...initialIntelligentEvalReadState,
+      list: { phase: 'ready', value: [evaluation], error: null },
+    }
+    const refreshing = intelligentEvalReadReducer(ready, { type: 'list_requested', silent: true })
+    expect(refreshing.list.phase).toBe('refreshing')
+
+    const afterFailure = intelligentEvalReadReducer(refreshing, { type: 'list_failed', error: '网络错误' })
+    expect(afterFailure.list).toEqual({ phase: 'ready', value: [evaluation], error: null })
+  })
+
+  it('surfaces an initial detail failure without an existing snapshot', () => {
+    const loading = intelligentEvalReadReducer(initialIntelligentEvalReadState, {
+      type: 'detail_requested',
+      id: 'eval-1',
+      requestId: 1,
+    })
+    const failed = intelligentEvalReadReducer(loading, {
+      type: 'detail_failed',
+      id: 'eval-1',
+      requestId: 1,
+      error: '不存在',
+    })
+    expect(failed.detail).toEqual({
+      phase: 'error',
+      value: null,
+      error: '不存在',
+      selectedId: 'eval-1',
+      requestId: 1,
+    })
+  })
+
+  it('ignores a stale detail response after the selection changes', () => {
+    const first = intelligentEvalReadReducer(initialIntelligentEvalReadState, {
+      type: 'detail_requested',
+      id: 'eval-1',
+      requestId: 1,
+    })
+    const second = intelligentEvalReadReducer(first, {
+      type: 'detail_requested',
+      id: 'eval-2',
+      requestId: 2,
+    })
+    const stale = intelligentEvalReadReducer(second, {
+      type: 'detail_succeeded',
+      id: 'eval-1',
+      requestId: 1,
+      value: evaluation,
+    })
+
+    expect(stale).toBe(second)
+    expect(stale.detail.value).toBeNull()
+    expect(stale.detail.selectedId).toBe('eval-2')
+  })
+})

frontend/web/src/read/intelligentEval.ts

@@ -0,0 +1,99 @@
+import { intelligentEvalsApi, type IntelligentEval } from '../api'
+
+export type ReadPhase = 'idle' | 'loading' | 'refreshing' | 'ready' | 'error'
+
+export interface ReadSlot<T> {
+  phase: ReadPhase
+  value: T
+  error: string | null
+}
+
+export interface IntelligentEvalReadState {
+  list: ReadSlot<IntelligentEval[]>
+  detail: ReadSlot<IntelligentEval | null> & {
+    selectedId: string | null
+    requestId: number
+  }
+}
+
+export type IntelligentEvalReadAction =
+  | { type: 'list_requested'; silent?: boolean }
+  | { type: 'list_succeeded'; value: IntelligentEval[] }
+  | { type: 'list_failed'; error: string }
+  | { type: 'detail_cleared'; requestId: number }
+  | { type: 'detail_requested'; id: string; requestId: number; silent?: boolean }
+  | { type: 'detail_succeeded'; id: string; requestId: number; value: IntelligentEval }
+  | { type: 'detail_failed'; id: string; requestId: number; error: string }
+
+export interface IntelligentEvalReadAdapter {
+  list: () => Promise<IntelligentEval[]>
+  get: (id: string) => Promise<IntelligentEval>
+}
+
+export const intelligentEvalReadAdapter: IntelligentEvalReadAdapter = {
+  list: () => intelligentEvalsApi.list().then((response) => response.data.intelligent_evals),
+  get: (id) => intelligentEvalsApi.get(id).then((response) => response.data),
+}
+
+export const initialIntelligentEvalReadState: IntelligentEvalReadState = {
+  list: { phase: 'idle', value: [], error: null },
+  detail: { phase: 'idle', value: null, error: null, selectedId: null, requestId: 0 },
+}
+
+function requestPhase<T>(slot: ReadSlot<T>, silent: boolean | undefined): ReadSlot<T> {
+  return { ...slot, phase: silent && slot.phase === 'ready' ? 'refreshing' : 'loading', error: null }
+}
+
+export function intelligentEvalReadReducer(
+  state: IntelligentEvalReadState,
+  action: IntelligentEvalReadAction,
+): IntelligentEvalReadState {
+  switch (action.type) {
+    case 'list_requested':
+      return { ...state, list: requestPhase(state.list, action.silent) }
+    case 'list_succeeded':
+      return { ...state, list: { phase: 'ready', value: action.value, error: null } }
+    case 'list_failed':
+      return {
+        ...state,
+        list: state.list.value.length > 0
+          ? { ...state.list, phase: 'ready', error: null }
+          : { ...state.list, phase: 'error', error: action.error },
+      }
+    case 'detail_cleared':
+      return {
+        ...state,
+        detail: { phase: 'idle', value: null, error: null, selectedId: null, requestId: action.requestId },
+      }
+    case 'detail_requested': {
+      const sameSelection = state.detail.selectedId === action.id
+      const current = sameSelection
+        ? state.detail
+        : { ...state.detail, value: null, selectedId: action.id }
+      return {
+        ...state,
+        detail: { ...requestPhase(current, action.silent), selectedId: action.id, requestId: action.requestId },
+      }
+    }
+    case 'detail_succeeded':
+      if (state.detail.selectedId !== action.id || state.detail.requestId !== action.requestId) return state
+      return {
+        ...state,
+        detail: {
+          phase: 'ready',
+          value: action.value,
+          error: null,
+          selectedId: action.id,
+          requestId: action.requestId,
+        },
+      }
+    case 'detail_failed':
+      if (state.detail.selectedId !== action.id || state.detail.requestId !== action.requestId) return state
+      return {
+        ...state,
+        detail: state.detail.value
+          ? { ...state.detail, phase: 'ready', error: null }
+          : { ...state.detail, phase: 'error', error: action.error },
+      }
+  }
+}

frontend/web/src/read/useIntelligentEvalRead.test.tsx

@@ -0,0 +1,113 @@
+import { act, renderHook } from '@testing-library/react'
+import { afterEach, describe, expect, it, vi } from 'vitest'
+import { useIntelligentEvalRead } from './useIntelligentEvalRead'
+import type { IntelligentEvalReadAdapter } from './intelligentEval'
+import type { IntelligentEval } from '../api'
+
+const evaluation = { id: 'eval-1', name: '评估', status: 'executing' } as unknown as IntelligentEval
+
+function deferred<T>() {
+  let resolve!: (value: T) => void
+  let reject!: (error: Error) => void
+  const promise = new Promise<T>((resolvePromise, rejectPromise) => {
+    resolve = resolvePromise
+    reject = rejectPromise
+  })
+  return { promise, resolve, reject }
+}
+
+afterEach(() => {
+  vi.useRealTimers()
+})
+
+async function settle() {
+  await act(async () => { await Promise.resolve() })
+}
+
+describe('useIntelligentEvalRead', () => {
+  it('refreshes active evaluations every five seconds and stops at terminal state', async () => {
+    vi.useFakeTimers()
+    const adapter: IntelligentEvalReadAdapter = {
+      list: vi.fn().mockResolvedValue([evaluation]),
+      get: vi.fn().mockResolvedValue(evaluation),
+    }
+    renderHook(() => useIntelligentEvalRead(null, adapter))
+    await settle()
+    expect(adapter.list).toHaveBeenCalledTimes(1)
+
+    await act(async () => { await vi.advanceTimersByTimeAsync(5000) })
+    await settle()
+    expect(adapter.list).toHaveBeenCalledTimes(2)
+
+    const terminal = { ...evaluation, status: 'completed' } as IntelligentEval
+    vi.mocked(adapter.list).mockResolvedValue([terminal])
+    await act(async () => { await vi.advanceTimersByTimeAsync(5000) })
+    await settle()
+    await act(async () => { await vi.advanceTimersByTimeAsync(10000) })
+    expect(adapter.list).toHaveBeenCalledTimes(3)
+  })
+
+  it('retains a complete list when a silent refresh fails', async () => {
+    vi.useFakeTimers()
+    const adapter: IntelligentEvalReadAdapter = {
+      list: vi.fn()
+        .mockResolvedValueOnce([evaluation])
+        .mockRejectedValueOnce(new Error('网络错误')),
+      get: vi.fn().mockResolvedValue(evaluation),
+    }
+    const { result } = renderHook(() => useIntelligentEvalRead(null, adapter))
+    await settle()
+    await act(async () => { await vi.advanceTimersByTimeAsync(5000) })
+    await settle()
+
+    expect(result.current.list.phase).toBe('ready')
+    expect(result.current.list.value).toEqual([evaluation])
+    expect(result.current.list.error).toBeNull()
+  })
+
+  it('clears the old detail and ignores a response for the previous selection', async () => {
+    const first = deferred<IntelligentEval>()
+    const second = deferred<IntelligentEval>()
+    const secondEvaluation = { ...evaluation, id: 'eval-2', name: '评估二' } as IntelligentEval
+    const adapter: IntelligentEvalReadAdapter = {
+      list: vi.fn().mockResolvedValue([]),
+      get: vi.fn((id: string) => (id === 'eval-1' ? first.promise : second.promise)),
+    }
+    const { result, rerender } = renderHook(
+      ({ selectedId }) => useIntelligentEvalRead(selectedId, adapter),
+      { initialProps: { selectedId: 'eval-1' as string | null } },
+    )
+
+    rerender({ selectedId: 'eval-2' })
+    expect(result.current.detail.value).toBeNull()
+    expect(result.current.detail.selectedId).toBe('eval-2')
+
+    await act(async () => { second.resolve(secondEvaluation) })
+    expect(result.current.detail.value?.id).toBe('eval-2')
+
+    await act(async () => { first.resolve(evaluation) })
+    expect(result.current.detail.value?.id).toBe('eval-2')
+  })
+
+  it('surfaces failure for a newly selected detail instead of retaining the old snapshot', async () => {
+    const adapter: IntelligentEvalReadAdapter = {
+      list: vi.fn().mockResolvedValue([]),
+      get: vi.fn()
+        .mockResolvedValueOnce(evaluation)
+        .mockRejectedValueOnce(new Error('不存在')),
+    }
+    const { result, rerender } = renderHook(
+      ({ selectedId }) => useIntelligentEvalRead(selectedId, adapter),
+      { initialProps: { selectedId: 'eval-1' as string | null } },
+    )
+    await settle()
+    expect(result.current.detail.value?.id).toBe('eval-1')
+
+    rerender({ selectedId: 'eval-2' })
+    await settle()
+
+    expect(result.current.detail.phase).toBe('error')
+    expect(result.current.detail.value).toBeNull()
+    expect(result.current.detail.error).toBe('不存在')
+  })
+})

frontend/web/src/read/useIntelligentEvalRead.ts

@@ -0,0 +1,80 @@
+import { useCallback, useEffect, useLayoutEffect, useReducer, useRef } from 'react'
+import {
+  intelligentEvalReadAdapter,
+  intelligentEvalReadReducer,
+  initialIntelligentEvalReadState,
+  type IntelligentEvalReadAdapter,
+  type IntelligentEvalReadState,
+} from './intelligentEval'
+import { usePolling } from '../hooks/usePolling'
+
+const ACTIVE_STATUSES = new Set(['planning', 'pending_approval', 'executing'])
+
+function errorMessage(error: unknown): string {
+  return error instanceof Error ? error.message : '读取智能评估失败'
+}
+
+function isActive(status: string | undefined): boolean {
+  return status != null && ACTIVE_STATUSES.has(status)
+}
+
+export function useIntelligentEvalRead(
+  selectedId: string | null,
+  adapter: IntelligentEvalReadAdapter = intelligentEvalReadAdapter,
+): IntelligentEvalReadState & { reloadList: () => Promise<void>; reloadDetail: () => Promise<void> } {
+  const [state, dispatch] = useReducer(intelligentEvalReadReducer, initialIntelligentEvalReadState)
+  const detailRequestId = useRef(0)
+
+  const loadList = useCallback(async (silent = false) => {
+    dispatch({ type: 'list_requested', silent })
+    try {
+      const value = await adapter.list()
+      dispatch({ type: 'list_succeeded', value })
+    } catch (error) {
+      dispatch({ type: 'list_failed', error: errorMessage(error) })
+    }
+  }, [adapter])
+
+  const loadDetail = useCallback(async (id: string, silent = false) => {
+    const requestId = ++detailRequestId.current
+    dispatch({ type: 'detail_requested', id, requestId, silent })
+    try {
+      const value = await adapter.get(id)
+      dispatch({ type: 'detail_succeeded', id, requestId, value })
+    } catch (error) {
+      dispatch({ type: 'detail_failed', id, requestId, error: errorMessage(error) })
+    }
+  }, [adapter])
+
+  useEffect(() => {
+    void loadList()
+  }, [loadList])
+
+  useLayoutEffect(() => {
+    if (selectedId == null) {
+      dispatch({ type: 'detail_cleared', requestId: ++detailRequestId.current })
+      return
+    }
+    void loadDetail(selectedId)
+  }, [loadDetail, selectedId])
+
+  const listActive = state.list.value.some((item) => isActive(item.status))
+  const detailActive = isActive(state.detail.value?.status)
+
+  usePolling(() => { void loadList(true) }, 5000, listActive)
+  usePolling(
+    () => {
+      if (selectedId != null) void loadDetail(selectedId, true)
+    },
+    5000,
+    selectedId != null && detailActive,
+  )
+
+  const reloadList = useCallback(() => loadList(true), [loadList])
+  const reloadDetail = useCallback(
+    () => (selectedId == null ? Promise.resolve() : loadDetail(selectedId, true)),
+    [loadDetail, selectedId],
+  )
+
+  return { ...state, reloadList, reloadDetail }
+}

frontend/web/src/test/setup.ts

@@ -0,0 +1 @@
+import '@testing-library/jest-dom/vitest'

frontend/web/src/test/smoke.test.tsx

@@ -0,0 +1,13 @@
+import { render, screen } from '@testing-library/react'
+import { describe, expect, it } from 'vitest'
+
+function SmokeView() {
+  return <p>测试环境已就绪</p>
+}
+
+describe('frontend test foundation', () => {
+  it('renders a React component in jsdom', () => {
+    render(<SmokeView />)
+    expect(screen.getByText('测试环境已就绪')).toBeInTheDocument()
+  })
+})

frontend/web/vite.config.ts

@@ -1,4 +1,4 @@
-import { defineConfig } from 'vite'
+import { defineConfig } from 'vitest/config'
 import react from '@vitejs/plugin-react'
 
 export default defineConfig({
@@ -54,4 +54,8 @@ export default defineConfig({
       },
     },
   },
+  test: {
+    environment: 'jsdom',
+    setupFiles: './src/test/setup.ts',
+  },
 })

migrations/versions/c2f4a6b8d0e1_add_campaign_run_identity.py

@@ -0,0 +1,54 @@
+"""add durable Campaign child Run identity
+
+Revision ID: c2f4a6b8d0e1
+Revises: 99dbae2a20da
+Create Date: 2026-08-06
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+import sqlmodel  # noqa: F401
+from alembic import op
+
+
+revision: str = "c2f4a6b8d0e1"
+down_revision: Union[str, Sequence[str], None] = "99dbae2a20da"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    """Add nullable identity columns without backfilling historical Runs."""
+    # Adding nullable columns directly avoids rebuilding ``eval_runs``.  That
+    # table participates in a foreign-key cycle with ``campaigns`` and
+    # ``eval_targets``; SQLite batch recreation can otherwise fail while
+    # topologically sorting the copied table.  A unique index has the same
+    # database-level idempotency semantics as a composite unique constraint.
+    inspector = sa.inspect(op.get_bind())
+    columns = {column["name"] for column in inspector.get_columns("eval_runs")}
+    if "campaign_plan_index" not in columns:
+        op.add_column("eval_runs", sa.Column("campaign_plan_index", sa.Integer(), nullable=True))
+    if "campaign_occurrence_index" not in columns:
+        op.add_column("eval_runs", sa.Column("campaign_occurrence_index", sa.Integer(), nullable=True))
+    index_names = {index["name"] for index in inspector.get_indexes("eval_runs")}
+    if "uq_eval_runs_campaign_occurrence" not in index_names:
+        op.create_index(
+            "uq_eval_runs_campaign_occurrence",
+            "eval_runs",
+            ["campaign_id", "campaign_plan_index", "campaign_occurrence_index"],
+            unique=True,
+        )
+
+
+def downgrade() -> None:
+    """Remove the child identity fields and their uniqueness constraint."""
+    inspector = sa.inspect(op.get_bind())
+    index_names = {index["name"] for index in inspector.get_indexes("eval_runs")}
+    if "uq_eval_runs_campaign_occurrence" in index_names:
+        op.drop_index("uq_eval_runs_campaign_occurrence", table_name="eval_runs")
+    columns = {column["name"] for column in inspector.get_columns("eval_runs")}
+    if "campaign_occurrence_index" in columns:
+        op.drop_column("eval_runs", "campaign_occurrence_index")
+    if "campaign_plan_index" in columns:
+        op.drop_column("eval_runs", "campaign_plan_index")

scripts/ci-check.sh

@@ -23,6 +23,10 @@ else
 fi
 
 echo "==> 3/4 pytest"
+# Tests use mocked HTTP boundaries and should not inherit a developer's local
+# SOCKS/HTTP proxy. Otherwise httpx may require an optional socksio package
+# merely while importing the reverse-proxy router.
+env -u ALL_PROXY -u all_proxy -u HTTP_PROXY -u http_proxy -u HTTPS_PROXY -u https_proxy \
     python -m pytest -q
 
 if [[ $FAST -eq 1 ]]; then

scripts/deploy-t480.sh

@@ -86,6 +86,7 @@ run rsync -az --delete \
     --exclude='.pytest_cache' \
     --exclude='.ruff_cache' \
     --exclude='AGENTS.md' \
+    --exclude='.scratch' \
     --exclude='data' \
     --exclude='.env' \
     --exclude='config/config.json' \

scripts/deploy-volcengine-102.sh

@@ -0,0 +1,207 @@
+#!/usr/bin/env bash
+# Versioned production deployment for the volcengine-102 environment.
+#
+# Required:
+#   AGENTEVAL_PROD_HOST=<ssh host alias>
+#
+# Optional:
+#   AGENTEVAL_PROD_DIR=/opt/agenteval
+#   AGENTEVAL_PROD_PORT=8002
+#   AGENTEVAL_PROD_BACKUP_DIR=/var/backups/agenteval
+#
+# Usage:
+#   scripts/deploy-volcengine-102.sh --tag 0.8.0-6248568
+#   scripts/deploy-volcengine-102.sh --dry-run --tag 0.8.0-6248568
+#   scripts/deploy-volcengine-102.sh --rollback 0.8.0-6248568
+
+set -euo pipefail
+
+HOST="${AGENTEVAL_PROD_HOST:-}"
+REMOTE_DIR="${AGENTEVAL_PROD_DIR:-/opt/agenteval}"
+APP_PORT="${AGENTEVAL_PROD_PORT:-8002}"
+BACKUP_DIR="${AGENTEVAL_PROD_BACKUP_DIR:-/var/backups/agenteval}"
+COMPOSE_FILE="deploy/volcengine-102/docker-compose.yml"
+
+DRY_RUN=0
+IMAGE_TAG=""
+ROLLBACK_TAG=""
+
+usage() {
+    sed -n '2,18p' "$0"
+}
+
+log()  { printf '\033[1;36m>> %s\033[0m\n' "$*"; }
+warn() { printf '\033[1;33m!! %s\033[0m\n' "$*" >&2; }
+die()  { printf '\033[1;31m!! %s\033[0m\n' "$*" >&2; exit 1; }
+run() {
+    if [[ "$DRY_RUN" == "1" ]]; then
+        printf '\033[2m[dry-run]'
+        printf ' %q' "$@"
+        printf '\033[0m\n'
+    else
+        "$@"
+    fi
+}
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --tag)
+            [[ $# -ge 2 ]] || die "--tag requires a value"
+            IMAGE_TAG="$2"
+            shift 2
+            ;;
+        --rollback)
+            [[ $# -ge 2 ]] || die "--rollback requires an image tag"
+            ROLLBACK_TAG="$2"
+            shift 2
+            ;;
+        --dry-run)
+            DRY_RUN=1
+            shift
+            ;;
+        -h|--help)
+            usage
+            exit 0
+            ;;
+        *)
+            die "unknown argument: $1"
+            ;;
+    esac
+done
+
+[[ -n "$HOST" ]] || die "AGENTEVAL_PROD_HOST is required"
+[[ -z "$IMAGE_TAG" || -z "$ROLLBACK_TAG" ]] || die "--tag and --rollback cannot be used together"
+
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$ROOT"
+
+RELEASE_DIR=""
+cleanup_release_source() {
+    if [[ -n "$RELEASE_DIR" && -d "$RELEASE_DIR" ]]; then
+        rm -rf "$RELEASE_DIR"
+    fi
+}
+trap cleanup_release_source EXIT
+
+validate_tag() {
+    [[ "$1" =~ ^[A-Za-z0-9][A-Za-z0-9._-]*$ ]] || die "invalid image tag: $1"
+}
+
+backup_remote_data() {
+    local release_tag="$1"
+    local timestamp backup_file
+    timestamp=$(date -u +"%Y%m%dT%H%M%SZ")
+    backup_file="agenteval-${release_tag}-${timestamp}.tgz"
+
+    log "backup production data volume → $BACKUP_DIR/$backup_file"
+    run ssh "$HOST" "set -eu; \
+        mkdir -p '$BACKUP_DIR'; \
+        if docker inspect agenteval >/dev/null 2>&1; then \
+            DATA_VOLUME=\$(docker inspect agenteval --format '{{range .Mounts}}{{if eq .Destination \"/app/data\"}}{{.Name}}{{end}}{{end}}'); \
+            [ -n \"\$DATA_VOLUME\" ] || { echo 'agenteval data volume not found' >&2; exit 1; }; \
+            docker run --rm -v \"\$DATA_VOLUME:/source:ro\" -v '$BACKUP_DIR:/backup' alpine:3.20 \
+                tar -czf '/backup/$backup_file' -C /source .; \
+        else \
+            echo 'first deployment: no existing agenteval container, backup skipped'; \
+        fi"
+}
+
+wait_for_health() {
+    local body=""
+    log "waiting for production /api/health" >&2
+    for _attempt in $(seq 1 30); do
+        body=$(ssh "$HOST" "curl -sf http://localhost:$APP_PORT/api/health 2>/dev/null" || true)
+        [[ -n "$body" ]] && break
+        sleep 2
+    done
+    [[ -n "$body" ]] || die "production health check did not respond after 60s"
+    printf '%s' "$body"
+}
+
+log "pre-flight: ssh connectivity"
+run ssh -o ConnectTimeout=10 "$HOST" "echo ok >/dev/null" || die "cannot reach $HOST via ssh"
+
+if [[ -n "$ROLLBACK_TAG" ]]; then
+    validate_tag "$ROLLBACK_TAG"
+    backup_remote_data "pre-rollback-$ROLLBACK_TAG"
+    log "rollback production image → agenteval:$ROLLBACK_TAG"
+    run ssh "$HOST" "set -eu; cd '$REMOTE_DIR'; \
+        docker image inspect 'agenteval:$ROLLBACK_TAG' >/dev/null; \
+        IMAGE_TAG='$ROLLBACK_TAG' docker compose -f '$COMPOSE_FILE' up -d --no-build agenteval"
+    if [[ "$DRY_RUN" == "1" ]]; then
+        log "rollback dry-run complete"
+        exit 0
+    fi
+    HEALTH_BODY=$(wait_for_health)
+    log "/api/health: $HEALTH_BODY"
+    warn "database migrations are not downgraded; restore a matching backup only when the release requires it"
+    log "rollback OK: image=agenteval:$ROLLBACK_TAG"
+    exit 0
+fi
+
+log "pre-flight: sync version metadata"
+run python3 scripts/sync_version.py
+
+if [[ "$DRY_RUN" != "1" ]] && { ! git diff --quiet HEAD -- || ! git diff --cached --quiet; }; then
+    die "tracked changes exist; commit them before a production deployment"
+fi
+
+VERSION=$(python3 -c "import re; print(re.search(r'^version\s*=\s*\"([^\"]+)\"', open('pyproject.toml').read(), re.M).group(1))")
+COMMIT=$(git rev-parse --short HEAD)
+BUILD_TIME=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
+IMAGE_TAG="${IMAGE_TAG:-${VERSION}-${COMMIT}}"
+validate_tag "$IMAGE_TAG"
+
+log "release: version=$VERSION image=agenteval:$IMAGE_TAG commit=$COMMIT"
+
+log "pre-flight: remote production configuration"
+run ssh "$HOST" "set -eu; mkdir -p '$REMOTE_DIR'; \
+    test -f '$REMOTE_DIR/deploy/volcengine-102/.env' \
+        || { echo 'missing production deploy/volcengine-102/.env' >&2; exit 1; }"
+
+log "prepare committed release source from HEAD"
+RELEASE_DIR=$(mktemp -d "${TMPDIR:-/tmp}/agenteval-release.XXXXXX")
+git archive --format=tar HEAD | tar -xf - -C "$RELEASE_DIR"
+
+log "sync committed release source → $HOST:$REMOTE_DIR"
+run rsync -az --delete \
+    --exclude='.env' \
+    --exclude='data' \
+    --exclude='config/config.json' \
+    --exclude='*.db*' \
+    "$RELEASE_DIR/" "$HOST:$REMOTE_DIR/"
+
+backup_remote_data "$IMAGE_TAG"
+
+log "build immutable production image agenteval:$IMAGE_TAG"
+run ssh "$HOST" "set -eu; cd '$REMOTE_DIR'; \
+    IMAGE_TAG='$IMAGE_TAG' BUILD_COMMIT='$COMMIT' BUILD_TIME='$BUILD_TIME' \
+    docker compose -f '$COMPOSE_FILE' build agenteval"
+
+log "start production services"
+run ssh "$HOST" "set -eu; cd '$REMOTE_DIR'; \
+    IMAGE_TAG='$IMAGE_TAG' BUILD_COMMIT='$COMMIT' BUILD_TIME='$BUILD_TIME' \
+    docker compose -f '$COMPOSE_FILE' up -d --remove-orphans"
+
+if [[ "$DRY_RUN" == "1" ]]; then
+    log "production deployment dry-run complete"
+    exit 0
+fi
+
+HEALTH_BODY=$(wait_for_health)
+REMOTE_VERSION=$(printf '%s' "$HEALTH_BODY" | python3 -c "import json, sys; print(json.load(sys.stdin).get('version', ''))")
+REMOTE_COMMIT=$(printf '%s' "$HEALTH_BODY" | python3 -c "import json, sys; print(json.load(sys.stdin).get('commit', ''))")
+[[ "$REMOTE_VERSION" == "$VERSION" ]] || die "version mismatch: local=$VERSION remote=$REMOTE_VERSION"
+[[ "$REMOTE_COMMIT" == "$COMMIT" ]] || die "commit mismatch: local=$COMMIT remote=$REMOTE_COMMIT"
+
+log "authenticated API smoke checks"
+run ssh "$HOST" "set -eu; cd '$REMOTE_DIR'; \
+    API_KEY_VALUE=\$(sed -n 's/^AGENTEVAL_API_KEY=//p' deploy/volcengine-102/.env | head -1); \
+    for API_PATH in /api/targets /api/scenarios /api/runs /api/campaigns /api/intelligent-evals /api/model-configs; do \
+        HTTP_CODE=\$(curl -sS -o /dev/null -w '%{http_code}' -H \"X-API-Key: \$API_KEY_VALUE\" \"http://localhost:$APP_PORT\$API_PATH\"); \
+        [ \"\$HTTP_CODE\" = '200' ] || { echo \"\$API_PATH returned \$HTTP_CODE\" >&2; exit 1; }; \
+        echo \"\$API_PATH 200\"; \
+    done"
+
+log "/api/health: $HEALTH_BODY"
+log "production deploy OK: version=$VERSION image=agenteval:$IMAGE_TAG commit=$COMMIT"

scripts/production-entrypoint.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env sh
+# Bootstrap legacy SQLite databases once, then apply real migrations on every
+# subsequent start. This keeps the old pre-Alembic production database usable
+# without stamping away future migrations.
+
+set -eu
+
+DB_PATH="${AGENTEVAL_DB_PATH:-data/agenteval.db}"
+
+if [ ! -f "$DB_PATH" ]; then
+    echo "production db not found; creating the initial SQLModel schema"
+    python -c "from agenteval.storage.db import init_db; init_db()"
+    alembic stamp head
+else
+    HAS_ALEMBIC_VERSION=$(python -c "import sqlite3, sys; connection = sqlite3.connect(sys.argv[1]); row = connection.execute(\"SELECT 1 FROM sqlite_master WHERE type = 'table' AND name = 'alembic_version'\").fetchone(); connection.close(); print('yes' if row else 'no')" "$DB_PATH")
+    if [ "$HAS_ALEMBIC_VERSION" = "no" ]; then
+        echo "legacy production db detected; starting migrations from the pre-Alembic baseline"
+        alembic stamp base
+    fi
+fi
+
+alembic upgrade head
+exec agenteval server start --host 0.0.0.0 --port 8000

tests/integration/test_campaign_analysis_api.py

@@ -151,7 +151,7 @@ async def test_post_then_get_completed_analysis(client, seeded_db, monkeypatch):
     # 假后台任务：同步写入 completed 行（真任务的单测覆盖在 test_campaign_analysis.py）
     monkeypatch.setattr(
         campaigns_module,
-        "start_campaign_analysis",
+        "enqueue_campaign_analysis",
         lambda cid, *, triggered_by: _complete_analysis_row(seeded_db, cid),
     )
 
@@ -172,7 +172,7 @@ async def test_rerun_upserts_without_new_row(client, seeded_db, monkeypatch):
     campaign_id = await _create_campaign(client, seeded_db)
     monkeypatch.setattr(
         campaigns_module,
-        "start_campaign_analysis",
+        "enqueue_campaign_analysis",
         lambda cid, *, triggered_by: _complete_analysis_row(seeded_db, cid),
     )
 

tests/integration/test_campaign_analysis_auto_trigger.py

@@ -63,7 +63,7 @@ def analysis_spy(monkeypatch):
     """Spy the analysis seam: resolvable model, recorded enqueue calls."""
     calls: list[tuple[str, str]] = []
     monkeypatch.setattr(
-        campaign_runner, "start_campaign_analysis",
+        campaign_runner, "enqueue_campaign_analysis",
         lambda cid, *, triggered_by: calls.append((cid, triggered_by)),
     )
     monkeypatch.setattr(campaign_runner, "resolve_analysis_model", lambda campaign, session: object())

tests/integration/test_campaign_runner.py

@@ -8,10 +8,11 @@ No real timer is used — the clock is injected, mirroring how the durable loop
 """
 
 import pytest
-from agenteval.evaluation.campaign_runner import advance_campaign
+from agenteval.evaluation.campaign_runner import advance_campaign, reconcile_campaign_child_runs
 from agenteval.models import (
     Campaign,
     CampaignPlanEntry,
+    CampaignSummary,
     Case,
     CaseType,
     ChannelType,
@@ -20,6 +21,7 @@ from agenteval.models import (
     RunStatus,
     RunTrigger,
     Scenario,
+    SchedulerState,
     TargetStatus,
 )
 from agenteval.storage.repository import CampaignRepository, RunRepository, ScenarioRepository, TargetRepository
@@ -44,6 +46,7 @@ def seeded_db(db_session, monkeypatch):
     monkeypatch.setattr(engine_module, "get_session", _test_get_session)
 
     channel = MockChannel(reply_delay=0.0)
+    db_session.info["mock_channel"] = channel
     monkeypatch.setattr(factory_module.ChannelFactory, "create", lambda target: channel)
 
     target = EvalTarget(
@@ -68,6 +71,7 @@ def _make_campaign(session) -> Campaign:
     return CampaignRepository(session).create(Campaign(
         name="compressed",
         target_id="t-1",
+        status="running",
         window_seconds=7200,
         time_scale=3600.0,  # 1 real second == 3600 window seconds
         plan=[
@@ -91,6 +95,9 @@ async def test_advance_spawns_due_runs_matching_plan(seeded_db):
 
     runs = RunRepository(seeded_db).list_all()
     assert len(runs) == 3
+    assert {
+        (run.campaign_plan_index, run.campaign_occurrence_index) for run in runs
+    } == {(0, 0), (0, 1), (1, 0)}
     for run in runs:
         assert run.campaign_id == campaign.id
         assert run.scenario_id == "s-1"
@@ -118,3 +125,140 @@ async def test_advance_reports_finished_at_window_end(seeded_db):
 
 async def test_advance_missing_campaign_returns_none(seeded_db):
     assert await advance_campaign(campaign_id="nope", elapsed_seconds=0.0, session=seeded_db) is None
+
+
+async def test_restart_after_claim_does_not_create_second_run(seeded_db):
+    campaign = CampaignRepository(seeded_db).create(
+        Campaign(
+            name="claimed-before-crash",
+            target_id="t-1",
+            status="running",
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+    )
+    claim = RunRepository(seeded_db).claim_campaign_run(
+        campaign_id=campaign.id,
+        scenario_id="s-1",
+        scenario_version=1,
+        plan_index=0,
+        occurrence_index=0,
+    )
+    assert claim.run is not None
+
+    result = await advance_campaign(campaign_id=campaign.id, elapsed_seconds=0.0, session=seeded_db)
+
+    runs = RunRepository(seeded_db).list_by_campaign(campaign.id)
+    assert result.spawned_run_ids == []
+    assert len(runs) == 1
+    assert runs[0].id == claim.run.id
+    assert runs[0].status is RunStatus.PENDING
+
+
+async def test_partial_claim_does_not_hide_remaining_occurrences(seeded_db):
+    campaign = CampaignRepository(seeded_db).create(
+        Campaign(
+            name="partial-claim",
+            target_id="t-1",
+            status="running",
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=2)],
+        )
+    )
+    first_claim = RunRepository(seeded_db).claim_campaign_run(
+        campaign_id=campaign.id,
+        scenario_id="s-1",
+        scenario_version=1,
+        plan_index=0,
+        occurrence_index=0,
+    )
+    assert first_claim.run is not None
+    campaign.summary = CampaignSummary(scheduler=SchedulerState(spawned_indices=[0]))
+    CampaignRepository(seeded_db).update(campaign)
+
+    await advance_campaign(campaign_id=campaign.id, elapsed_seconds=0.0, session=seeded_db)
+
+    runs = RunRepository(seeded_db).list_by_campaign(campaign.id)
+    assert len(runs) == 2
+    assert {run.campaign_occurrence_index for run in runs} == {0, 1}
+    assert all(run.status is RunStatus.COMPLETED for run in runs)
+
+
+async def test_partial_spawn_failure_is_retried_until_all_occurrences_are_claimed(seeded_db, monkeypatch):
+    from agenteval.evaluation import campaign_runner
+
+    campaign = CampaignRepository(seeded_db).create(
+        Campaign(
+            name="partial-failure",
+            target_id="t-1",
+            status="running",
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=2)],
+        )
+    )
+    real_spawn = campaign_runner._spawn_child_run
+    failed_once = False
+
+    async def flaky_spawn(*args, occurrence_index, **kwargs):
+        nonlocal failed_once
+        if occurrence_index == 1 and not failed_once:
+            failed_once = True
+            raise RuntimeError("transient spawn failure")
+        return await real_spawn(*args, occurrence_index=occurrence_index, **kwargs)
+
+    monkeypatch.setattr(campaign_runner, "_spawn_child_run", flaky_spawn)
+
+    await advance_campaign(campaign_id=campaign.id, elapsed_seconds=0.0, session=seeded_db)
+    after_failure = CampaignRepository(seeded_db).get(campaign.id)
+    assert after_failure.summary.scheduler.spawned_indices == []
+
+    await advance_campaign(campaign_id=campaign.id, elapsed_seconds=0.0, session=seeded_db)
+    runs = RunRepository(seeded_db).list_by_campaign(campaign.id)
+    recovered = CampaignRepository(seeded_db).get(campaign.id)
+    assert {run.campaign_occurrence_index for run in runs} == {0, 1}
+    assert recovered.summary.scheduler.spawned_indices == [0]
+
+
+async def test_recovery_resumes_pending_claim_without_replacing_identity(seeded_db):
+    campaign = _make_campaign(seeded_db)
+    claim = RunRepository(seeded_db).claim_campaign_run(
+        campaign_id=campaign.id,
+        scenario_id="s-1",
+        scenario_version=1,
+        plan_index=0,
+        occurrence_index=0,
+    )
+    assert claim.run is not None
+
+    result = await reconcile_campaign_child_runs(campaign.id, seeded_db)
+
+    recovered = RunRepository(seeded_db).get(claim.run.id)
+    assert result is not None
+    assert result.resumed_run_ids == [claim.run.id]
+    assert recovered.status is RunStatus.COMPLETED
+    assert len(RunRepository(seeded_db).list_by_campaign(campaign.id)) == 1
+
+
+async def test_recovery_fails_running_claim_without_replaying_messages(seeded_db):
+    campaign = _make_campaign(seeded_db)
+    repo = RunRepository(seeded_db)
+    claim = repo.claim_campaign_run(
+        campaign_id=campaign.id,
+        scenario_id="s-1",
+        scenario_version=1,
+        plan_index=0,
+        occurrence_index=0,
+    )
+    assert claim.run is not None
+    claim.run.status = RunStatus.RUNNING
+    repo.update(claim.run)
+
+    result = await reconcile_campaign_child_runs(campaign.id, seeded_db)
+
+    interrupted = repo.get(claim.run.id)
+    assert result is not None
+    assert result.resumed_run_ids == []
+    assert result.failed_run_ids == [claim.run.id]
+    assert interrupted.status is RunStatus.FAILED
+    assert interrupted.summary.error.code == "interrupted"
+    assert seeded_db.info["mock_channel"].send_calls == 0

tests/integration/test_campaign_scheduler_loop.py

@@ -98,6 +98,26 @@ async def test_loop_runs_to_completion(seeded_db):
     assert all(r.status == RunStatus.COMPLETED for r in runs)
 
 
+async def test_loop_retries_completion_after_settlement_failure(seeded_db, monkeypatch):
+    campaign = _make_campaign(seeded_db)
+    real_complete = campaign_runner.complete_campaign
+    attempts = 0
+
+    def flaky_complete(*args, **kwargs):
+        nonlocal attempts
+        attempts += 1
+        if attempts == 1:
+            raise RuntimeError("settlement failed")
+        return real_complete(*args, **kwargs)
+
+    monkeypatch.setattr(campaign_runner, "complete_campaign", flaky_complete)
+    start_campaign(campaign.id, seeded_db, tick_seconds=TICK)
+    await _await_task(campaign.id)
+
+    assert attempts == 2
+    assert CampaignRepository(seeded_db).get(campaign.id).status is CampaignStatus.COMPLETED
+
+
 async def test_restart_recovery_does_not_respawn(seeded_db):
     # Simulate a campaign that was already RUNNING before a restart, with its
     # window start well in the past and entry 0 already recorded as spawned.

tests/integration/test_campaigns_api.py

@@ -106,7 +106,7 @@ async def test_create_campaign_then_get(client, seeded_db):
     assert resp.status_code == 200
     body = resp.json()
     campaign_id = body["id"]
-    assert body["status"] == "planned"
+    assert body["status"] == "running"
     assert body["name"] == "24h-cycle"
     assert body["target_id"] == "t-1"
     assert body["window_seconds"] == 86400
@@ -115,7 +115,7 @@ async def test_create_campaign_then_get(client, seeded_db):
 
     got = (await client.get(f"/api/campaigns/{campaign_id}")).json()
     assert got["id"] == campaign_id
-    assert got["status"] == "planned"
+    assert got["status"] == "running"
     assert got["plan"][0]["scenario_id"] == "s-1"
     assert got["plan"][0]["offset_seconds"] == 0
     assert got["plan"][0]["count"] == 2
@@ -245,7 +245,7 @@ async def test_detail_includes_progress_fields(client, seeded_db):
     progress = got["progress"]
     assert progress["spawned_runs"] == 0
     assert progress["completed_runs"] == 0
-    assert progress["current_offset_seconds"] == 0.0
+    assert progress["current_offset_seconds"] >= 0.0
 
 
 # ── campaign report endpoint (ticket 04) ─────────────────────────────────────

tests/integration/test_exploration_api.py

@@ -182,7 +182,7 @@ async def test_dict_reply_content_is_flattened_to_text(seeded_db, monkeypatch, c
     from tests.unit.mock_channel import MockChannel
 
     class _DictReplyChannel(MockChannel):
-        async def poll_reply(self, question_msg_id, timeout=30.0, poll_interval=1.0):
+        async def _poll_reply(self, question_msg_id, timeout=30.0, poll_interval=1.0):
             return Reply(
                 question_msg_id=question_msg_id,
                 content={"content": "您好，我是客服"},

tests/integration/test_intelligent_evals_api.py

@@ -101,17 +101,34 @@ class TestCreateEval:
         assert "id" in data
 
     async def test_create_requires_name(self, client, seeded_db):
-        resp = await client.post("/api/intelligent-evals", json={
-            "name": "", "target_id": "t-1", "goal": "test",
-        })
+        resp = await client.post(
+            "/api/intelligent-evals",
+            json={
+                "name": "",
+                "target_id": "t-1",
+                "goal": "test",
+            },
+        )
         assert resp.status_code == 422
 
     async def test_create_requires_goal(self, client, seeded_db):
-        resp = await client.post("/api/intelligent-evals", json={
-            "name": "test", "target_id": "t-1", "goal": "",
-        })
+        resp = await client.post(
+            "/api/intelligent-evals",
+            json={
+                "name": "test",
+                "target_id": "t-1",
+                "goal": "",
+            },
+        )
         assert resp.status_code == 422
 
+    async def test_create_requires_existing_target(self, client, seeded_db):
+        resp = await client.post(
+            "/api/intelligent-evals",
+            json={"name": "test", "target_id": "missing", "goal": "evaluate"},
+        )
+        assert resp.status_code == 404
+
 
 class TestPlanApproval:
     async def test_submit_plan_transitions_to_pending_approval(self, client, seeded_db):
@@ -267,6 +284,8 @@ class TestCreateSession:
         data = resp.json()
         assert data["session_count"] == 1
         assert data["completed_sessions"] == 0
+        assert len(data["sessions"]) == 1
+        assert "messages" not in data["sessions"][0]
 
 
 class TestConductTurn:
@@ -424,6 +443,23 @@ class TestListSessions:
         resp = await client.get(f"/api/intelligent-evals/{other['id']}/sessions/{session['id']}/messages")
         assert resp.status_code == 404
 
+    async def test_session_mutation_via_wrong_eval_returns_404(self, client, seeded_db):
+        eval_id = await _create_executing_eval(client)
+        session = await _create_session(client, eval_id)
+        other = await _create_executing_eval(client)
+
+        message = await client.post(
+            f"/api/intelligent-evals/{other}/sessions/{session['id']}/messages",
+            json={"content": "越权消息"},
+        )
+        assert message.status_code == 404
+
+        closed = await client.post(
+            f"/api/intelligent-evals/{other}/sessions/{session['id']}/close",
+            json={"verdict": {"passed": False}},
+        )
+        assert closed.status_code == 404
+
 
 def _report_payload() -> dict:
     return {
@@ -453,9 +489,7 @@ def _report_payload() -> dict:
 class TestReport:
     async def test_submit_report_completes_eval(self, client, seeded_db):
         eval_id = await _create_executing_eval(client)
-        resp = await client.put(
-            f"/api/intelligent-evals/{eval_id}/report", json={"report": _report_payload()}
-        )
+        resp = await client.put(f"/api/intelligent-evals/{eval_id}/report", json={"report": _report_payload()})
         assert resp.status_code == 200, resp.text
         data = resp.json()
         assert data["status"] == "completed"
@@ -482,9 +516,7 @@ class TestReport:
 
     async def test_submit_report_requires_executing(self, client, seeded_db):
         ev = await _create_eval(client)  # planning state
-        resp = await client.put(
-            f"/api/intelligent-evals/{ev['id']}/report", json={"report": _report_payload()}
-        )
+        resp = await client.put(f"/api/intelligent-evals/{ev['id']}/report", json={"report": _report_payload()})
         assert resp.status_code == 409
 
     async def test_submit_report_rejects_invalid_structure(self, client, seeded_db):

tests/unit/mock_channel.py

@@ -37,7 +37,7 @@ class MockChannel(EvalChannel):
     async def health_check(self) -> ChannelHealth:
         return ChannelHealth(ok=True, message="mock")
 
-    async def send(self, content: str, **kwargs: Any) -> SendResult:
+    async def _send(self, content: str, **kwargs: Any) -> SendResult:
         self.send_calls += 1
         self.sent.append(content)
         if self.raise_on_send:
@@ -47,7 +47,7 @@ class MockChannel(EvalChannel):
         self._msg_counter += 1
         return SendResult(ok=True, question_msg_id=f"q-{self._msg_counter}")
 
-    async def poll_reply(
+    async def _poll_reply(
         self,
         question_msg_id: str,
         timeout: float = 30.0,

tests/unit/test_campaign_cancel_lifecycle.py

@@ -0,0 +1,77 @@
+"""Tests for atomic Campaign cancellation and settlement ordering."""
+
+import pytest
+from agenteval.evaluation.campaign_lifecycle import CampaignLifecycleError, cancel_campaign
+from agenteval.exploration.models import ExplorationSession, ExplorationSessionStatus
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus
+from agenteval.storage.repository import CampaignRepository, ExplorationSessionRepository
+from sqlalchemy import event
+
+
+def _campaign(session, status=CampaignStatus.RUNNING):
+    return CampaignRepository(session).create(
+        Campaign(
+            id="campaign-1",
+            name="campaign",
+            target_id="t-1",
+            status=status,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+    )
+
+
+def test_cancel_is_atomic_and_stop_happens_after_durable_state(db_session):
+    _campaign(db_session)
+    observed = []
+
+    def stop(campaign_id):
+        observed.append(CampaignRepository(db_session).get(campaign_id).status)
+
+    cancelled = cancel_campaign(db_session, "campaign-1", stop=stop)
+
+    assert cancelled.status is CampaignStatus.CANCELLED
+    assert cancelled.completed_at is not None
+    assert observed == [CampaignStatus.CANCELLED]
+
+
+def test_cancel_distinguishes_missing_and_terminal_campaign(db_session):
+    with pytest.raises(CampaignLifecycleError) as missing:
+        cancel_campaign(db_session, "missing")
+    assert missing.value.status_code == 404
+
+    _campaign(db_session, CampaignStatus.COMPLETED)
+    with pytest.raises(CampaignLifecycleError) as terminal:
+        cancel_campaign(db_session, "campaign-1")
+    assert terminal.value.status_code == 400
+
+
+def test_cancel_is_idempotency_guarded_by_status(db_session):
+    _campaign(db_session)
+    cancel_campaign(db_session, "campaign-1")
+    with pytest.raises(CampaignLifecycleError) as again:
+        cancel_campaign(db_session, "campaign-1")
+    assert again.value.status_code == 400
+
+
+def test_cancel_settlement_failure_rolls_back_campaign_and_session(db_session):
+    _campaign(db_session)
+    session_obj = ExplorationSessionRepository(db_session).create(
+        ExplorationSession(campaign_id="campaign-1", target_id="t-1")
+    )
+    engine = db_session.get_bind()
+
+    def fail_exploration_update(_connection, _cursor, statement, _parameters, _context, _executemany):
+        if statement.lstrip().upper().startswith("UPDATE EXPLORATION_SESSIONS"):
+            raise RuntimeError("settlement failed")
+
+    event.listen(engine, "before_cursor_execute", fail_exploration_update)
+    try:
+        with pytest.raises(RuntimeError, match="settlement failed"):
+            cancel_campaign(db_session, "campaign-1")
+    finally:
+        event.remove(engine, "before_cursor_execute", fail_exploration_update)
+        db_session.rollback()
+
+    assert CampaignRepository(db_session).get("campaign-1").status is CampaignStatus.RUNNING
+    assert ExplorationSessionRepository(db_session).get(session_obj.id).status is ExplorationSessionStatus.RUNNING

tests/unit/test_campaign_complete_lifecycle.py

@@ -0,0 +1,72 @@
+"""Tests for idempotent Campaign completion and settlement ordering."""
+
+import pytest
+from agenteval.evaluation.campaign_lifecycle import CampaignLifecycleError, complete_campaign
+from agenteval.exploration.models import ExplorationSession, ExplorationSessionStatus
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus
+from agenteval.storage.repository import CampaignRepository, ExplorationSessionRepository
+from sqlalchemy import event
+
+
+def _campaign(session, status=CampaignStatus.RUNNING):
+    return CampaignRepository(session).create(
+        Campaign(
+            id="campaign-1",
+            name="campaign",
+            target_id="t-1",
+            status=status,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+    )
+
+
+def test_complete_atomically_settles_running_exploration(db_session):
+    _campaign(db_session)
+    session_obj = ExplorationSessionRepository(db_session).create(
+        ExplorationSession(campaign_id="campaign-1", target_id="t-1")
+    )
+
+    completed = complete_campaign(db_session, "campaign-1")
+
+    assert completed.status is CampaignStatus.COMPLETED
+    assert completed.completed_at is not None
+    assert ExplorationSessionRepository(db_session).get(session_obj.id).status is ExplorationSessionStatus.EXPIRED
+
+
+def test_complete_is_idempotency_guarded(db_session):
+    _campaign(db_session)
+    complete_campaign(db_session, "campaign-1")
+    with pytest.raises(CampaignLifecycleError) as again:
+        complete_campaign(db_session, "campaign-1")
+    assert again.value.status_code == 409
+
+
+def test_cancel_wins_completion_race(db_session):
+    _campaign(db_session, CampaignStatus.CANCELLED)
+    with pytest.raises(CampaignLifecycleError) as conflict:
+        complete_campaign(db_session, "campaign-1")
+    assert conflict.value.status_code == 409
+
+
+def test_settlement_failure_keeps_campaign_running_and_session_open(db_session):
+    _campaign(db_session)
+    session_obj = ExplorationSessionRepository(db_session).create(
+        ExplorationSession(campaign_id="campaign-1", target_id="t-1")
+    )
+    engine = db_session.get_bind()
+
+    def fail_exploration_update(_connection, _cursor, statement, _parameters, _context, _executemany):
+        if statement.lstrip().upper().startswith("UPDATE EXPLORATION_SESSIONS"):
+            raise RuntimeError("settlement failed")
+
+    event.listen(engine, "before_cursor_execute", fail_exploration_update)
+    try:
+        with pytest.raises(RuntimeError, match="settlement failed"):
+            complete_campaign(db_session, "campaign-1")
+    finally:
+        event.remove(engine, "before_cursor_execute", fail_exploration_update)
+        db_session.rollback()
+
+    assert CampaignRepository(db_session).get("campaign-1").status is CampaignStatus.RUNNING
+    assert ExplorationSessionRepository(db_session).get(session_obj.id).status is ExplorationSessionStatus.RUNNING

tests/unit/test_campaign_lifecycle.py

@@ -0,0 +1,70 @@
+"""Tests for the durable Campaign creation lifecycle seam."""
+
+import pytest
+from agenteval.evaluation.campaign_lifecycle import CampaignCreateError, create_campaign
+from agenteval.models import CampaignPlanEntry, CampaignStatus, Case, Scenario
+from agenteval.storage.repository import CampaignRepository, ScenarioRepository, TargetRepository
+
+from tests.unit.test_repository import _make_target
+
+
+def test_create_campaign_commits_running_before_launch(db_session):
+    TargetRepository(db_session).create(_make_target())
+    ScenarioRepository(db_session).create(Scenario(id="s-1", name="scenario", cases=[Case(id="c-1", messages=["hi"])]))
+    launched: list[str] = []
+
+    def launch(campaign_id, session):
+        launched.append(campaign_id)
+        assert CampaignRepository(session).get(campaign_id).status is CampaignStatus.RUNNING
+
+    campaign = create_campaign(
+        db_session,
+        name="campaign",
+        target_id="t-1",
+        window_seconds=60,
+        time_scale=1,
+        plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        launch=launch,
+    )
+
+    assert campaign.status is CampaignStatus.RUNNING
+    assert campaign.started_at is not None
+    assert launched == [campaign.id]
+
+
+def test_create_campaign_launch_failure_leaves_durable_row(db_session):
+    TargetRepository(db_session).create(_make_target())
+    ScenarioRepository(db_session).create(Scenario(id="s-1", name="scenario", cases=[Case(id="c-1", messages=["hi"])]))
+
+    def launch(_campaign_id, _session):
+        raise RuntimeError("scheduler unavailable")
+
+    with pytest.raises(RuntimeError):
+        create_campaign(
+            db_session,
+            name="campaign",
+            target_id="t-1",
+            window_seconds=60,
+            time_scale=1,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+            launch=launch,
+        )
+
+    rows = CampaignRepository(db_session).list_all()
+    assert len(rows) == 1
+    assert rows[0].status is CampaignStatus.RUNNING
+
+
+def test_create_campaign_rejects_invalid_references_without_writing(db_session):
+    with pytest.raises(CampaignCreateError) as exc:
+        create_campaign(
+            db_session,
+            name="campaign",
+            target_id="missing",
+            window_seconds=60,
+            time_scale=1,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+
+    assert exc.value.status_code == 404
+    assert CampaignRepository(db_session).list_all() == []

tests/unit/test_campaign_recovery.py

@@ -0,0 +1,46 @@
+"""Tests for the unified durable Campaign startup recovery seam."""
+
+from agenteval.evaluation import campaign_lifecycle
+
+
+def test_recovery_coordinator_orders_cleanup_before_relaunch(db_session, monkeypatch):
+    calls: list[str] = []
+
+    class FakeRuns:
+        def __init__(self, _session):
+            pass
+
+        def mark_orphans_failed(self):
+            calls.append("runs")
+            return 2
+
+    class FakeAnalysis:
+        def __init__(self, _session):
+            pass
+
+        def mark_orphans_failed(self):
+            calls.append("analysis")
+            return 1
+
+    class FakeComparison(FakeAnalysis):
+        def mark_orphans_failed(self):
+            calls.append("comparison")
+            return 1
+
+    monkeypatch.setattr(campaign_lifecycle, "RunRepository", FakeRuns)
+    monkeypatch.setattr(campaign_lifecycle, "CampaignAnalysisRepository", FakeAnalysis)
+    monkeypatch.setattr(campaign_lifecycle, "CampaignPeriodComparisonRepository", FakeComparison)
+
+    import agenteval.evaluation.analysis as analysis
+    import agenteval.evaluation.campaign_runner as runner
+
+    monkeypatch.setattr(analysis, "resume_queued_campaign_analysis", lambda _session: calls.append("queued") or 3)
+    monkeypatch.setattr(runner, "resume_running_campaigns", lambda _session, **_: calls.append("campaigns") or 4)
+
+    summary = campaign_lifecycle.recover_campaign_runtime(db_session)
+
+    assert calls == ["runs", "analysis", "comparison", "campaigns", "queued"]
+    assert summary.interrupted_runs == 2
+    assert summary.interrupted_analysis == 2
+    assert summary.resumed_campaigns == 4
+    assert summary.resumed_analysis == 3

tests/unit/test_campaign_run_claim.py

@@ -0,0 +1,115 @@
+"""Tests for the durable, idempotent Campaign child-Run claim seam."""
+
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, RunStatus, RunTrigger
+from agenteval.storage.repository import (
+    CampaignRepository,
+    CampaignRunClaimStatus,
+    RunRepository,
+    TargetRepository,
+)
+
+from tests.unit.test_repository import _make_target
+
+
+def _running_campaign(session, *, status=CampaignStatus.RUNNING):
+    TargetRepository(session).create(_make_target())
+    return CampaignRepository(session).create(
+        Campaign(
+            id="campaign-1",
+            name="campaign",
+            target_id="t-1",
+            status=status,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="scenario-1", offset_seconds=0, count=2)],
+        )
+    )
+
+
+def _claim(repo: RunRepository, *, occurrence_index=0):
+    return repo.claim_campaign_run(
+        campaign_id="campaign-1",
+        scenario_id="scenario-1",
+        scenario_version=3,
+        plan_index=0,
+        occurrence_index=occurrence_index,
+    )
+
+
+def test_claim_creates_pending_campaign_run_with_durable_identity(db_session):
+    _running_campaign(db_session)
+
+    result = _claim(RunRepository(db_session))
+
+    assert result.status is CampaignRunClaimStatus.CLAIMED
+    assert result.accepted is True
+    assert result.created is True
+    assert result.run is not None
+    assert result.run.status is RunStatus.PENDING
+    assert result.run.triggered_by is RunTrigger.CAMPAIGN
+    assert result.run.target_id == "t-1"
+    assert result.run.campaign_id == "campaign-1"
+    assert result.run.campaign_plan_index == 0
+    assert result.run.campaign_occurrence_index == 0
+
+
+def test_repeated_claim_returns_same_run_without_duplicate(db_session):
+    _running_campaign(db_session)
+    repo = RunRepository(db_session)
+
+    first = _claim(repo)
+    second = _claim(repo)
+
+    assert first.run is not None and second.run is not None
+    assert second.status is CampaignRunClaimStatus.EXISTING
+    assert second.created is False
+    assert second.run.id == first.run.id
+    assert len(repo.list_by_campaign("campaign-1")) == 1
+
+
+def test_claim_rejects_missing_or_non_running_campaign(db_session):
+    repo = RunRepository(db_session)
+
+    missing = _claim(repo)
+    assert missing.status is CampaignRunClaimStatus.NOT_FOUND
+    assert missing.accepted is False
+
+    _running_campaign(db_session, status=CampaignStatus.CANCELLED)
+    cancelled = _claim(repo)
+    assert cancelled.status is CampaignRunClaimStatus.CONFLICT
+    assert cancelled.run is None
+
+
+def test_each_occurrence_has_independent_claim_identity(db_session):
+    _running_campaign(db_session)
+    repo = RunRepository(db_session)
+
+    first = _claim(repo, occurrence_index=0)
+    second = _claim(repo, occurrence_index=1)
+
+    assert first.created is True
+    assert second.created is True
+    assert first.run is not None and second.run is not None
+    assert first.run.id != second.run.id
+    assert len(repo.list_by_campaign("campaign-1")) == 2
+
+
+def test_competing_sessions_converge_on_one_claim(tmp_path):
+    from agenteval.storage.db import EvalRunDB  # noqa: F401 - register metadata
+    from sqlmodel import Session, SQLModel, create_engine
+
+    engine = create_engine(
+        f"sqlite:///{tmp_path / 'competing-claims.db'}",
+        connect_args={"check_same_thread": False},
+    )
+    SQLModel.metadata.create_all(engine)
+    with Session(engine) as first_session, Session(engine) as second_session:
+        _running_campaign(first_session)
+
+        first = _claim(RunRepository(first_session))
+        second = _claim(RunRepository(second_session))
+
+        assert first.status is CampaignRunClaimStatus.CLAIMED
+        assert second.status is CampaignRunClaimStatus.EXISTING
+        assert first.run is not None and second.run is not None
+        assert second.run.id == first.run.id
+        assert len(RunRepository(second_session).list_by_campaign("campaign-1")) == 1

tests/unit/test_campaign_run_identity_migration.py

@@ -0,0 +1,85 @@
+"""Verify durable Campaign child-Run identity migration semantics."""
+
+import importlib
+
+import sqlalchemy as sa
+from alembic.migration import MigrationContext
+from alembic.operations import Operations
+
+
+def _legacy_eval_runs(connection) -> None:
+    connection.execute(sa.text("CREATE TABLE campaigns (id VARCHAR PRIMARY KEY)"))
+    connection.execute(
+        sa.text(
+            "CREATE TABLE eval_runs ("
+            "id VARCHAR PRIMARY KEY, campaign_id VARCHAR NULL, "
+            "status VARCHAR NOT NULL)"
+        )
+    )
+    connection.execute(sa.text("INSERT INTO eval_runs (id, campaign_id, status) VALUES ('old', NULL, 'completed')"))
+
+
+def test_campaign_run_identity_migration_is_nullable_unique_and_reversible(tmp_path):
+    engine = sa.create_engine(f"sqlite:///{tmp_path / 'campaign-run-identity.db'}")
+    migration = importlib.import_module(
+        "migrations.versions.c2f4a6b8d0e1_add_campaign_run_identity"
+    )
+
+    with engine.begin() as connection:
+        _legacy_eval_runs(connection)
+        operations = Operations(MigrationContext.configure(connection))
+        migration.op = operations
+        migration.upgrade()
+
+        inspector = sa.inspect(connection)
+        columns = {column["name"]: column for column in inspector.get_columns("eval_runs")}
+        assert columns["campaign_plan_index"]["nullable"] is True
+        assert columns["campaign_occurrence_index"]["nullable"] is True
+        assert connection.execute(sa.text("SELECT campaign_plan_index, campaign_occurrence_index FROM eval_runs WHERE id='old'")) .one() == (None, None)
+        indexes = inspector.get_indexes("eval_runs")
+        assert any(
+            index["name"] == "uq_eval_runs_campaign_occurrence"
+            and bool(index["unique"])
+            and index["column_names"] == ["campaign_id", "campaign_plan_index", "campaign_occurrence_index"]
+            for index in indexes
+        )
+
+        connection.execute(
+            sa.text(
+                "INSERT INTO eval_runs (id, campaign_id, campaign_plan_index, campaign_occurrence_index, status) "
+                "VALUES ('child-1', 'campaign-1', 0, 0, 'pending')"
+            )
+        )
+        try:
+            connection.execute(
+                sa.text(
+                    "INSERT INTO eval_runs (id, campaign_id, campaign_plan_index, campaign_occurrence_index, status) "
+                    "VALUES ('child-duplicate', 'campaign-1', 0, 0, 'pending')"
+                )
+            )
+        except sa.exc.IntegrityError:
+            pass
+        else:  # pragma: no cover - assertion guard
+            raise AssertionError("Campaign child identity must be unique")
+
+        migration.downgrade()
+        assert "campaign_plan_index" not in {column["name"] for column in sa.inspect(connection).get_columns("eval_runs")}
+        assert "campaign_occurrence_index" not in {column["name"] for column in sa.inspect(connection).get_columns("eval_runs")}
+
+
+def test_fresh_schema_exposes_same_identity_columns_and_unique_index(tmp_path):
+    """SQLModel's fresh-database path matches the Alembic contract."""
+    from agenteval.storage.db import EvalRunDB  # noqa: F401 - register table
+    from sqlmodel import SQLModel, create_engine
+
+    engine = create_engine(f"sqlite:///{tmp_path / 'fresh.db'}")
+    SQLModel.metadata.create_all(engine)
+    inspector = sa.inspect(engine)
+    columns = {column["name"] for column in inspector.get_columns("eval_runs")}
+    assert {"campaign_plan_index", "campaign_occurrence_index"} <= columns
+    assert any(
+        index["name"] == "uq_eval_runs_campaign_occurrence"
+        and bool(index["unique"])
+        and index["column_names"] == ["campaign_id", "campaign_plan_index", "campaign_occurrence_index"]
+        for index in inspector.get_indexes("eval_runs")
+    )

tests/unit/test_campaign_start_lifecycle.py

@@ -0,0 +1,42 @@
+"""Tests for the conditional Campaign start lifecycle seam."""
+
+from agenteval.evaluation.campaign_lifecycle import start_campaign
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus
+from agenteval.storage.repository import CampaignRepository
+
+
+def test_start_planned_campaign_is_atomic_and_launches_after_commit(db_session):
+    CampaignRepository(db_session).create(
+        Campaign(
+            id="campaign-1",
+            name="campaign",
+            target_id="t-1",
+            status=CampaignStatus.PLANNED,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+    )
+    observed = []
+
+    def launch(campaign_id, session):
+        observed.append(CampaignRepository(session).get(campaign_id).status)
+
+    started = start_campaign(db_session, "campaign-1", launch=launch)
+
+    assert started.status is CampaignStatus.RUNNING
+    assert started.started_at is not None
+    assert observed == [CampaignStatus.RUNNING]
+
+
+def test_start_terminal_campaign_is_noop(db_session):
+    CampaignRepository(db_session).create(
+        Campaign(
+            id="campaign-1",
+            name="campaign",
+            target_id="t-1",
+            status=CampaignStatus.COMPLETED,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=1)],
+        )
+    )
+    assert start_campaign(db_session, "campaign-1") is None

tests/unit/test_channel_contract.py

@@ -0,0 +1,147 @@
+"""Contract tests for the complete channel exchange outcome."""
+
+import pytest
+from agenteval.channels.base import (
+    ChannelHealth,
+    ChannelTransportError,
+    EvalChannel,
+    ExchangeOutcome,
+    ExchangeStatus,
+    Reply,
+    SendResult,
+    normalize_reply_text,
+)
+
+
+class ContractChannel(EvalChannel):
+    def __init__(self, *, send_result=None, reply=None, poll_error=None):
+        self.send_result = send_result or SendResult(ok=True, question_msg_id="question-7")
+        self.reply = reply
+        self.poll_error = poll_error
+        self.events: list[str] = []
+
+    async def health_check(self):
+        return ChannelHealth(ok=True)
+
+    async def _send(self, content, **kwargs):
+        self.events.append("send")
+        return self.send_result
+
+    async def _poll_reply(self, question_msg_id, timeout=30.0, poll_interval=1.0):
+        self.events.append("poll")
+        if self.poll_error:
+            raise self.poll_error
+        return self.reply
+
+
+def test_success_normalizes_reply_and_preserves_exchange_metadata():
+    diagnostic = {"provider": "tutu", "request_id": "req-7"}
+
+    outcome = ExchangeOutcome.succeeded(
+        correlation_id="question-7",
+        reply={"msgBody": {"content": "你好"}},
+        latency_ms=128,
+        diagnostic=diagnostic,
+    )
+
+    assert outcome.status is ExchangeStatus.SUCCESS
+    assert outcome.ok is True
+    assert outcome.expected_failure is False
+    assert outcome.correlation_id == "question-7"
+    assert outcome.reply_text == "你好"
+    assert outcome.latency_ms == 128
+    assert outcome.diagnostic is diagnostic
+
+
+@pytest.mark.parametrize(
+    ("outcome", "status"),
+    [
+        (ExchangeOutcome.send_failed("connection refused"), ExchangeStatus.SEND_FAILED),
+        (ExchangeOutcome.reply_timeout(correlation_id="question-7", latency_ms=30_000), ExchangeStatus.REPLY_TIMEOUT),
+        (
+            ExchangeOutcome.poll_failed("upstream returned 502", correlation_id="question-7"),
+            ExchangeStatus.POLL_FAILED,
+        ),
+    ],
+)
+def test_expected_transport_failures_have_typed_status(outcome, status):
+    assert outcome.status is status
+    assert outcome.ok is False
+    assert outcome.expected_failure is True
+
+
+def test_normalize_reply_text_handles_provider_shapes():
+    assert normalize_reply_text("plain text") == "plain text"
+    assert normalize_reply_text({"content": "content field"}) == "content field"
+    assert normalize_reply_text({"msgBody": {"text": "nested text"}}) == "nested text"
+    assert normalize_reply_text(None) == ""
+
+
+def test_success_requires_correlation_text_and_latency():
+    with pytest.raises(ValueError, match="correlation_id"):
+        ExchangeOutcome(status=ExchangeStatus.SUCCESS, reply_text="ok", latency_ms=1)
+
+    with pytest.raises(ValueError, match="reply_text"):
+        ExchangeOutcome(status=ExchangeStatus.SUCCESS, correlation_id="question-7", latency_ms=1)
+
+    with pytest.raises(ValueError, match="latency_ms"):
+        ExchangeOutcome(status=ExchangeStatus.SUCCESS, correlation_id="question-7", reply_text="ok", latency_ms=-1)
+
+
+async def test_exchange_runs_send_hook_before_polling_and_returns_success():
+    channel = ContractChannel(reply=Reply(question_msg_id="question-7", content={"content": "答复"}))
+
+    async def on_sent(send_result):
+        channel.events.append(f"hook:{send_result.question_msg_id}")
+
+    outcome = await channel.exchange("问题", on_sent=on_sent)
+
+    assert outcome.ok is True
+    assert outcome.reply_text == "答复"
+    assert outcome.correlation_id == "question-7"
+    assert channel.events == ["send", "hook:question-7", "poll"]
+
+
+async def test_exchange_send_failure_skips_hook_and_polling():
+    channel = ContractChannel(send_result=SendResult(ok=False, error="offline"))
+
+    async def on_sent(_send_result):
+        raise AssertionError("send hook must not run after a failed send")
+
+    outcome = await channel.exchange("问题", on_sent=on_sent)
+
+    assert outcome.status is ExchangeStatus.SEND_FAILED
+    assert outcome.reason == "offline"
+    assert channel.events == ["send"]
+
+
+async def test_exchange_hook_failure_skips_polling():
+    channel = ContractChannel(reply=Reply(question_msg_id="question-7", content="答复"))
+
+    async def on_sent(_send_result):
+        channel.events.append("hook")
+        raise RuntimeError("ledger unavailable")
+
+    with pytest.raises(RuntimeError, match="ledger unavailable"):
+        await channel.exchange("问题", on_sent=on_sent)
+
+    assert channel.events == ["send", "hook"]
+
+
+async def test_exchange_distinguishes_timeout_and_poll_failure():
+    timeout_channel = ContractChannel()
+    timeout = await timeout_channel.exchange("问题")
+    assert timeout.status is ExchangeStatus.REPLY_TIMEOUT
+    assert timeout.correlation_id == "question-7"
+
+    failed_channel = ContractChannel(poll_error=ChannelTransportError("upstream unavailable"))
+    failed = await failed_channel.exchange("问题")
+    assert failed.status is ExchangeStatus.POLL_FAILED
+    assert failed.reason == "upstream unavailable"
+
+
+async def test_exchange_does_not_mask_programming_or_configuration_errors():
+    channel = ContractChannel(poll_error=ValueError("invalid adapter configuration"))
+
+    with pytest.raises(ValueError, match="invalid adapter configuration"):
+        await channel.exchange("问题")

tests/unit/test_engine.py

@@ -7,7 +7,7 @@ timeouts, and concurrent case execution via the semaphore.
 import asyncio
 from typing import Any
 
-from agenteval.channels.base import SendResult
+from agenteval.channels.base import ChannelTransportError, SendResult
 from agenteval.evaluation.engine import EvalEngine, TimeoutConfig
 from agenteval.models import (
     Case,
@@ -21,7 +21,9 @@ from agenteval.models import (
     Scenario,
     TargetStatus,
 )
+from agenteval.storage.db import TurnDB
 from agenteval.storage.repository import RunRepository
+from sqlmodel import select
 
 from tests.unit.mock_channel import MockChannel
 
@@ -71,9 +73,11 @@ def _build_engine(
 
 # ── basic happy path ─────────────────────────────────────────────────────
 
+
 async def test_run_single_case_completes(db_session):
     scenario = Scenario(
-        id="s-1", name="single",
+        id="s-1",
+        name="single",
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hello"])],
     )
     channel = MockChannel()
@@ -90,7 +94,8 @@ async def test_run_single_case_completes(db_session):
 
 async def test_run_multi_turn_collects_all_turns(db_session):
     scenario = Scenario(
-        id="s-1", name="multi",
+        id="s-1",
+        name="multi",
         cases=[Case(id="c1", type=CaseType.MULTI_TURN, messages=["a", "b", "c"])],
     )
     channel = MockChannel()
@@ -107,7 +112,8 @@ async def test_run_multi_turn_collects_all_turns(db_session):
 
 async def test_run_multiple_cases(db_session):
     scenario = Scenario(
-        id="s-1", name="multi-case",
+        id="s-1",
+        name="multi-case",
         cases=[
             Case(id="c1", type=CaseType.SINGLE, messages=["one"]),
             Case(id="c2", type=CaseType.SINGLE, messages=["two"]),
@@ -126,9 +132,11 @@ async def test_run_multiple_cases(db_session):
 
 # ── progress callback ────────────────────────────────────────────────────
 
+
 async def test_progress_callback_receives_events(db_session):
     scenario = Scenario(
-        id="s-1", name="single",
+        id="s-1",
+        name="single",
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
     )
     events: list[tuple[str, dict]] = []
@@ -151,7 +159,8 @@ async def test_progress_callback_receives_events(db_session):
 async def test_sync_progress_callback_also_works(db_session):
     """Engine must accept sync callbacks (CLI uses them)."""
     scenario = Scenario(
-        id="s-1", name="single",
+        id="s-1",
+        name="single",
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
     )
     seen: list[str] = []
@@ -168,9 +177,11 @@ async def test_sync_progress_callback_also_works(db_session):
 
 # ── cancellation ─────────────────────────────────────────────────────────
 
+
 async def test_cancel_before_run_marks_failed(db_session):
     scenario = Scenario(
-        id="s-1", name="single",
+        id="s-1",
+        name="single",
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
     )
     cancel_token = asyncio.Event()
@@ -189,7 +200,8 @@ async def test_cancel_before_run_marks_failed(db_session):
 async def test_cancel_mid_run_stops_after_current_case(db_session):
     """Cancelling between cases should stop further cases from running."""
     scenario = Scenario(
-        id="s-1", name="multi",
+        id="s-1",
+        name="multi",
         cases=[
             Case(id="c1", type=CaseType.SINGLE, messages=["a"]),
             Case(id="c2", type=CaseType.SINGLE, messages=["b"]),
@@ -220,16 +232,19 @@ async def test_cancel_mid_run_stops_after_current_case(db_session):
 
 # ── timeouts ─────────────────────────────────────────────────────────────
 
+
 async def test_poll_timeout_records_missing_reply(db_session):
     """When the channel returns no reply within the timeout, the turn is saved
     with reply=None but the engine keeps going (doesn't crash)."""
     scenario = Scenario(
-        id="s-1", name="single",
+        id="s-1",
+        name="single",
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
     )
     channel = MockChannel(missing_reply=True)
     engine = _build_engine(
-        scenario, channel,
+        scenario,
+        channel,
         timeout_config=TimeoutConfig(poll_reply=0.1),
         session=db_session,
     )
@@ -241,36 +256,79 @@ async def test_poll_timeout_records_missing_reply(db_session):
     turns = RunRepository(db_session).get_turns(run.id)
     assert len(turns) == 1
     assert turns[0].reply is None
+    assert turns[0].question_msg_id == "q-1"
+
+
+async def test_sent_turn_is_durable_before_reply_polling(db_session):
+    scenario = Scenario(
+        id="s-1",
+        name="durable-send",
+        cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
+    )
+
+    class LedgerInspectingChannel(MockChannel):
+        persisted_before_poll = False
+
+        async def _poll_reply(self, question_msg_id, timeout=30.0, poll_interval=1.0):
+            turns = list(db_session.exec(select(TurnDB)).all())
+            self.persisted_before_poll = len(turns) == 1 and turns[0].question_msg_id == question_msg_id
+            return await super()._poll_reply(question_msg_id, timeout, poll_interval)
+
+    channel = LedgerInspectingChannel()
+    engine = _build_engine(scenario, channel, session=db_session)
+
+    await engine.run()
+
+    assert channel.persisted_before_poll is True
+
+
+async def test_poll_failure_keeps_sent_turn_and_fails_case(db_session):
+    scenario = Scenario(
+        id="s-1",
+        name="poll-failure",
+        cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
+    )
+    channel = MockChannel(raise_on_poll=ChannelTransportError("upstream unavailable"))
+    engine = _build_engine(scenario, channel, session=db_session)
+
+    run = await engine.run()
+
+    assert run.status == RunStatus.COMPLETED
+    assert run.summary.failed_cases == 1
+    turns = RunRepository(db_session).get_turns(run.id)
+    assert len(turns) == 1
+    assert turns[0].question_msg_id == "q-1"
+    assert turns[0].reply is None
 
 
 # ── concurrency ──────────────────────────────────────────────────────────
 
+
 async def test_concurrent_cases_respect_semaphore(db_session):
     """With max_concurrent_cases=2, at most 2 cases should be in-flight."""
     scenario = Scenario(
-        id="s-1", name="concurrent",
-        cases=[
-            Case(id=f"c{i}", type=CaseType.SINGLE, messages=[f"m{i}"])
-            for i in range(5)
-        ],
+        id="s-1",
+        name="concurrent",
+        cases=[Case(id=f"c{i}", type=CaseType.SINGLE, messages=[f"m{i}"]) for i in range(5)],
     )
     in_flight = {"n": 0, "max": 0}
     lock = asyncio.Lock()
 
     class TrackedChannel(MockChannel):
-        async def send(self, content: str, **kwargs):
+        async def _send(self, content: str, **kwargs):
             async with lock:
                 in_flight["n"] += 1
                 in_flight["max"] = max(in_flight["max"], in_flight["n"])
             await asyncio.sleep(0.05)
-            result = await super().send(content, **kwargs)
+            result = await super()._send(content, **kwargs)
             async with lock:
                 in_flight["n"] -= 1
             return result
 
     channel = TrackedChannel()
     engine = _build_engine(
-        scenario, channel,
+        scenario,
+        channel,
         max_concurrent_cases=2,
         session=db_session,
     )
@@ -284,11 +342,13 @@ async def test_concurrent_cases_respect_semaphore(db_session):
 
 # ── send failure ─────────────────────────────────────────────────────────
 
+
 async def test_send_failure_aborts_case(db_session):
     """When channel.send fails, the case is marked failed but the engine
     continues with the next case."""
     scenario = Scenario(
-        id="s-1", name="mixed",
+        id="s-1",
+        name="mixed",
         cases=[
             Case(id="c1", type=CaseType.SINGLE, messages=["a"]),
             Case(id="c2", type=CaseType.SINGLE, messages=["b"]),
@@ -297,11 +357,11 @@ async def test_send_failure_aborts_case(db_session):
     call_count = {"n": 0}
 
     class FlakeyChannel(MockChannel):
-        async def send(self, content, **kwargs):
+        async def _send(self, content, **kwargs):
             call_count["n"] += 1
             if call_count["n"] == 1:
                 return SendResult(ok=False, error="boom")
-            return await super().send(content, **kwargs)
+            return await super()._send(content, **kwargs)
 
     channel = FlakeyChannel()
     engine = _build_engine(scenario, channel, session=db_session)
@@ -315,12 +375,14 @@ async def test_send_failure_aborts_case(db_session):
 
 # ── dynamic case generation failure ──────────────────────────────────────
 
+
 async def test_dynamic_generation_failure_records_case_error(db_session):
     """A dynamic case whose message generation fails must persist the reason
     into run.summary.case_errors — not just emit it transiently. Otherwise a
     run shows 0 rules / failed with no discoverable cause."""
     scenario = Scenario(
-        id="s-1", name="dynamic",
+        id="s-1",
+        name="dynamic",
         cases=[Case(id="dyn-1", type=CaseType.DYNAMIC, prompt="生成问题", turns=3)],
         llm_config=None,  # 缺 llm_config → 生成消息立即失败
     )
@@ -354,7 +416,9 @@ def _case_with(
     rule_pass_threshold: float = 0.6,
 ) -> Case:
     return Case(
-        id="c1", type=CaseType.SINGLE, messages=["hi"],
+        id="c1",
+        type=CaseType.SINGLE,
+        messages=["hi"],
         eval_rules=rules or [],
         expectations=expectations or Expectation(),
         rule_logic=rule_logic,
@@ -364,10 +428,16 @@ def _case_with(
 
 async def test_expectation_fails_case_even_when_rules_pass(db_session):
     """期望不满足 → 用例不通过，即使显式规则全部通过。"""
-    scenario = Scenario(id="s1", name="s", cases=[_case_with(
+    scenario = Scenario(
+        id="s1",
+        name="s",
+        cases=[
+            _case_with(
                 rules=[EvalRuleConfig(type="keyword_match", params={"keywords": ["echo"]})],
                 expectations=Expectation(keywords_include=["__NOPE__"]),
-    )])
+            )
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -379,10 +449,16 @@ async def test_expectation_fails_case_even_when_rules_pass(db_session):
 
 async def test_expectation_and_rules_both_pass(db_session):
     """期望与规则都满足 → 通过，且期望派生判定同构落库、reason 可辨识来源。"""
-    scenario = Scenario(id="s1", name="s", cases=[_case_with(
+    scenario = Scenario(
+        id="s1",
+        name="s",
+        cases=[
+            _case_with(
                 rules=[EvalRuleConfig(type="keyword_match", params={"keywords": ["echo"]})],
                 expectations=Expectation(keywords_include=["echo"], response_time_max_ms=99999),
-    )])
+            )
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -399,11 +475,17 @@ async def test_expectation_and_rules_both_pass(db_session):
 
 async def test_implicit_expectation_not_in_any_combination(db_session):
     """rule_logic=ANY 只组合显式规则：期望通过不能救活全败的显式规则组。"""
-    scenario = Scenario(id="s1", name="s", cases=[_case_with(
+    scenario = Scenario(
+        id="s1",
+        name="s",
+        cases=[
+            _case_with(
                 rules=[EvalRuleConfig(type="keyword_match", params={"keywords": ["__NOPE__"]})],
                 expectations=Expectation(keywords_include=["echo"]),  # 通过
                 rule_logic=RuleLogic.ANY,
-    )])
+            )
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -414,12 +496,18 @@ async def test_implicit_expectation_not_in_any_combination(db_session):
 
 async def test_implicit_expectation_is_hard_constraint_over_weighted(db_session):
     """rule_logic=WEIGHTED 达标但期望不满足 → 仍不通过（期望是硬约束）。"""
-    scenario = Scenario(id="s1", name="s", cases=[_case_with(
+    scenario = Scenario(
+        id="s1",
+        name="s",
+        cases=[
+            _case_with(
                 rules=[EvalRuleConfig(type="keyword_match", params={"keywords": ["echo"]}, weight=1.0)],
                 expectations=Expectation(keywords_include=["__NOPE__"]),
                 rule_logic=RuleLogic.WEIGHTED,
                 rule_pass_threshold=0.5,  # 显式加权得分 1.0 ≥ 0.5
-    )])
+            )
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -430,12 +518,16 @@ async def test_implicit_expectation_is_hard_constraint_over_weighted(db_session)
 
 async def test_pure_expectation_case_behavior_unchanged(db_session):
     """纯期望用例（无显式规则）：满足通过、不满足失败，与升级前一致。"""
-    scenario = Scenario(id="s1", name="s", cases=[
-        Case(id="ok", type=CaseType.SINGLE, messages=["hi"],
-             expectations=Expectation(keywords_include=["echo"])),
-        Case(id="bad", type=CaseType.SINGLE, messages=["hi"],
-             expectations=Expectation(keywords_include=["__NOPE__"])),
-    ])
+    scenario = Scenario(
+        id="s1",
+        name="s",
+        cases=[
+            Case(id="ok", type=CaseType.SINGLE, messages=["hi"], expectations=Expectation(keywords_include=["echo"])),
+            Case(
+                id="bad", type=CaseType.SINGLE, messages=["hi"], expectations=Expectation(keywords_include=["__NOPE__"])
+            ),
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -447,10 +539,13 @@ async def test_pure_expectation_case_behavior_unchanged(db_session):
 
 # ── scenario_version snapshot (ticket 04) ────────────────────────────────
 
+
 async def test_engine_run_snapshots_scenario_version(db_session):
     """引擎直启（CLI 路径）创建的运行快照场景当前版本。"""
     scenario = Scenario(
-        id="s-1", name="versioned", version=3,
+        id="s-1",
+        name="versioned",
+        version=3,
         cases=[Case(id="c1", type=CaseType.SINGLE, messages=["hi"])],
     )
     channel = MockChannel()
@@ -463,17 +558,22 @@ async def test_engine_run_snapshots_scenario_version(db_session):
     assert persisted.scenario_version == 3
 
 
-
 # ── 权威判定写入 summary.case_outcomes（判定语义收敛） ────────────────────
 
+
 async def test_summary_contains_case_outcomes_and_case_level_pass_rate(db_session):
-    scenario = Scenario(id="s-1", name="outcomes", cases=[
+    scenario = Scenario(
+        id="s-1",
+        name="outcomes",
+        cases=[
             # 连通用例（无规则无期望）
             Case(id="conn", type=CaseType.SINGLE, messages=["ping"]),
             # 判定失败用例（期望不满足）
-        Case(id="bad", type=CaseType.SINGLE, messages=["hi"],
-             expectations=Expectation(keywords_include=["__NOPE__"])),
-    ])
+            Case(
+                id="bad", type=CaseType.SINGLE, messages=["hi"], expectations=Expectation(keywords_include=["__NOPE__"])
+            ),
+        ],
+    )
     channel = MockChannel()
     engine = _build_engine(scenario, channel, session=db_session)
 
@@ -488,12 +588,18 @@ async def test_summary_contains_case_outcomes_and_case_level_pass_rate(db_sessio
 
 async def test_connectivity_case_without_reply_fails(db_session):
     """连通用例没收到回复=故障=不通过（此前无条件判通过的 bug）。"""
-    scenario = Scenario(id="s-1", name="conn-fail", cases=[
+    scenario = Scenario(
+        id="s-1",
+        name="conn-fail",
+        cases=[
             Case(id="conn", type=CaseType.SINGLE, messages=["ping"]),
-    ])
+        ],
+    )
     channel = MockChannel(missing_reply=True)
     engine = _build_engine(
-        scenario, channel, session=db_session,
+        scenario,
+        channel,
+        session=db_session,
         timeout_config=TimeoutConfig(poll_reply=0.2),
     )
 

tests/unit/test_http_channel_and_rules.py

@@ -2,8 +2,12 @@
 
 from unittest.mock import AsyncMock, MagicMock, patch
 
+import httpx
+import pytest
+from agenteval.channels.base import ExchangeStatus, SendResult
 from agenteval.channels.http import HttpChannel, _get_path
 from agenteval.channels.openclaw import OpenClawChannel
+from agenteval.channels.tutu import TutuApiChannel
 from agenteval.evaluation.rules.keyword import KeywordMatchRule
 from agenteval.evaluation.rules.response_time import ResponseTimeRule
 from agenteval.models import Case, CaseType, Expectation, Turn
@@ -67,19 +71,65 @@ async def test_http_send_extracts_msg_id():
     mock_resp.raise_for_status = MagicMock()
     mock_resp.json = MagicMock(return_value={"data": {"id": "msg-42"}})
     with patch.object(ch._client, "post", new=AsyncMock(return_value=mock_resp)):
-        result = await ch.send("hello")
+        result = await ch._send("hello")
     assert result.ok is True
     assert result.question_msg_id == "msg-42"
 
 
 async def test_http_send_failure():
     ch = _make_channel()
-    with patch.object(ch._client, "post", new=AsyncMock(side_effect=Exception("timeout"))):
-        result = await ch.send("hi")
+    with patch.object(ch._client, "post", new=AsyncMock(side_effect=httpx.ConnectError("timeout"))):
+        result = await ch._send("hi")
     assert result.ok is False
     assert "timeout" in result.error
 
 
+async def test_http_exchange_classifies_poll_transport_failure():
+    ch = _make_channel()
+    sent = MagicMock()
+    sent.raise_for_status = MagicMock()
+    sent.json = MagicMock(return_value={"id": "msg-1"})
+    with (
+        patch.object(ch._client, "post", new=AsyncMock(return_value=sent)),
+        patch.object(ch._client, "get", new=AsyncMock(side_effect=httpx.ConnectError("offline"))),
+    ):
+        outcome = await ch.exchange("hi")
+
+    assert outcome.status is ExchangeStatus.POLL_FAILED
+    assert "offline" in outcome.reason
+
+
+async def test_http_exchange_does_not_mask_adapter_programming_errors():
+    ch = _make_channel()
+    with patch.object(ch._client, "post", new=AsyncMock(side_effect=RuntimeError("adapter bug"))):
+        with pytest.raises(RuntimeError, match="adapter bug"):
+            await ch.exchange("hi")
+
+
+async def test_tutu_exchange_classifies_non_200_poll_response_as_failure():
+    channel = TutuApiChannel(
+        {
+            "base_url": "http://mock",
+            "token": "token",
+            "tenant": "tenant",
+            "chat_channel_id": "channel",
+            "chat_contact_id": "contact",
+        }
+    )
+    response = MagicMock(status_code=502, text="bad gateway")
+    client = MagicMock()
+    client.get = AsyncMock(return_value=response)
+
+    with (
+        patch.object(channel, "_send", new=AsyncMock(return_value=SendResult(ok=True, question_msg_id="msg-1"))),
+        patch.object(channel, "_get_client", new=AsyncMock(return_value=client)),
+    ):
+        outcome = await channel.exchange("hi", timeout=0.1, poll_interval=0.01)
+
+    assert outcome.status is ExchangeStatus.POLL_FAILED
+    assert "HTTP 502" in outcome.reason
+
+
 async def test_http_poll_reply_found():
     ch = _make_channel(reply_path="answer")
     call_count = {"n": 0}
@@ -92,7 +142,7 @@ async def test_http_poll_reply_found():
         return mock_resp
 
     with patch.object(ch._client, "get", new=mock_get):
-        reply = await ch.poll_reply("msg-1", timeout=5.0)
+        reply = await ch._poll_reply("msg-1", timeout=5.0)
     assert reply is not None
     assert reply.content == "Hello world"
     assert reply.question_msg_id == "msg-1"
@@ -108,7 +158,7 @@ async def test_http_poll_reply_timeout():
         return mock_resp
 
     with patch.object(ch._client, "get", new=mock_get):
-        reply = await ch.poll_reply("msg-1", timeout=0.15, poll_interval=0.05)
+        reply = await ch._poll_reply("msg-1", timeout=0.15, poll_interval=0.05)
     assert reply is None
 
 
@@ -128,7 +178,7 @@ async def test_http_poll_reply_readiness_flag():
         return mock_resp
 
     with patch.object(ch._client, "get", new=mock_get):
-        reply = await ch.poll_reply("msg-1", timeout=5.0, poll_interval=0.05)
+        reply = await ch._poll_reply("msg-1", timeout=5.0, poll_interval=0.05)
     assert reply is not None
     assert reply.content == "final answer"
 
@@ -243,15 +293,15 @@ async def test_openclaw_send_ok():
     mock_resp.raise_for_status = MagicMock()
     mock_resp.json = MagicMock(return_value={"id": "chat-msg-1", "choices": [{"message": {"content": "reply"}}]})
     with patch.object(ch._client, "post", new=AsyncMock(return_value=mock_resp)):
-        result = await ch.send("hello")
+        result = await ch._send("hello")
     assert result.ok is True
     assert result.question_msg_id == "chat-msg-1"
 
 
 async def test_openclaw_send_failure():
     ch = _make_openclaw_channel()
-    with patch.object(ch._client, "post", new=AsyncMock(side_effect=Exception("timeout"))):
-        result = await ch.send("hi")
+    with patch.object(ch._client, "post", new=AsyncMock(side_effect=httpx.ConnectError("timeout"))):
+        result = await ch._send("hi")
     assert result.ok is False
 
 
@@ -261,7 +311,7 @@ async def test_openclaw_poll_reply_found():
     mock_resp.status_code = 200
     mock_resp.json = MagicMock(return_value={"choices": [{"message": {"content": "assistant reply"}}]})
     with patch.object(ch._client, "get", new=AsyncMock(return_value=mock_resp)):
-        reply = await ch.poll_reply("chat-msg-1", timeout=5.0)
+        reply = await ch._poll_reply("chat-msg-1", timeout=5.0)
     assert reply is not None
     assert reply.content == "assistant reply"
 
@@ -272,7 +322,7 @@ async def test_openclaw_poll_reply_timeout():
     mock_resp.status_code = 404
     mock_resp.json = MagicMock(return_value={})
     with patch.object(ch._client, "get", new=AsyncMock(return_value=mock_resp)):
-        reply = await ch.poll_reply("chat-msg-1", timeout=0.15, poll_interval=0.05)
+        reply = await ch._poll_reply("chat-msg-1", timeout=0.15, poll_interval=0.05)
     assert reply is None
 
 

tests/unit/test_intelligent_eval_model.py

@@ -1,160 +1,167 @@
-"""Smoke tests for intelligent eval data model (ticket 01)."""
+"""Lifecycle contract tests for intelligent evaluations."""
 
 import pytest
-from agenteval.intelligent_eval.models import (
-    IntelligentEval,
-    IntelligentEvalMessage,
-    IntelligentEvalSession,
-    IntelligentEvalSessionStatus,
-    IntelligentEvalStatus,
+from agenteval.channels.base import ExchangeOutcome, SendResult
+from agenteval.intelligent_eval import lifecycle
+from agenteval.intelligent_eval.lifecycle import (
+    IntelligentEvalChannelError,
+    IntelligentEvalNotFoundError,
+    IntelligentEvalTransitionError,
 )
-from agenteval.intelligent_eval.repository import (
-    IntelligentEvalMessageRepository,
-    IntelligentEvalRepository,
-    IntelligentEvalSessionRepository,
+from agenteval.intelligent_eval.models import IntelligentEvalSessionStatus, IntelligentEvalStatus
+from agenteval.models import ChannelType, EvalTarget, PlatformType, TargetStatus
+from agenteval.storage.repository import TargetRepository
+
+
+@pytest.fixture()
+def eval_session(db_session):
+    TargetRepository(db_session).create(
+        EvalTarget(
+            id="target-1",
+            name="测试对象",
+            platform=PlatformType.AI_DIGITAL_EMPLOYEE,
+            channel_type=ChannelType.TUTU_API,
+            channel_config={"base_url": "http://mock", "token": "token"},
+            status=TargetStatus.ACTIVE,
         )
-from agenteval.storage.db import EvalTargetDB
-from sqlalchemy.pool import StaticPool
-from sqlmodel import Session, SQLModel, create_engine
+    )
+    return db_session
 
 
-@pytest.fixture
-def db_session():
-    engine = create_engine("sqlite://", connect_args={"check_same_thread": False}, poolclass=StaticPool)
-    SQLModel.metadata.create_all(engine)
-    session = Session(engine)
-    target = EvalTargetDB(id="t1", name="test-target")
-    session.add(target)
-    session.commit()
-    yield session
-    session.close()
+def _create(session, *, name: str = "智能评估"):
+    return lifecycle.create_eval(
+        session,
+        name=name,
+        target_id="target-1",
+        goal="验证退货流程",
+        seeds={"personas": ["老客户"]},
+        intent="流程覆盖",
+        role_description="模拟用户",
+    )
 
 
-class TestIntelligentEvalRepository:
-    def test_create_and_get(self, db_session):
-        repo = IntelligentEvalRepository(db_session)
-        ev = repo.create(IntelligentEval(
-            name="test-eval",
-            target_id="t1",
-            goal="evaluate customer service",
-            seeds={"personas": [], "goals": []},
-            intent="test intent",
-            role_description="impatient customer",
-            time_window_hours=24,
-        ))
-        assert ev.id is not None
-        assert ev.status == IntelligentEvalStatus.DRAFT
-        assert ev.name == "test-eval"
-        assert ev.time_window_hours == 24
-
-        fetched = repo.get(ev.id)
-        assert fetched is not None
-        assert fetched.goal == "evaluate customer service"
-        assert fetched.seeds == {"personas": [], "goals": []}
-
-    def test_list_all(self, db_session):
-        repo = IntelligentEvalRepository(db_session)
-        repo.create(IntelligentEval(name="eval-1", target_id="t1"))
-        repo.create(IntelligentEval(name="eval-2", target_id="t1"))
-        all_evals = repo.list_all()
-        assert len(all_evals) == 2
-
-    def test_status_transitions(self, db_session):
-        repo = IntelligentEvalRepository(db_session)
-        ev = repo.create(IntelligentEval(name="eval", target_id="t1"))
-        assert ev.status == IntelligentEvalStatus.DRAFT
-
-        ev = repo.transition_status(ev.id, IntelligentEvalStatus.PLANNING)
-        assert ev.status == IntelligentEvalStatus.PLANNING
-
-        ev = repo.transition_status(ev.id, IntelligentEvalStatus.PENDING_APPROVAL)
-        assert ev.status == IntelligentEvalStatus.PENDING_APPROVAL
-
-        ev = repo.transition_status(ev.id, IntelligentEvalStatus.EXECUTING)
-        assert ev.status == IntelligentEvalStatus.EXECUTING
-        assert ev.started_at is not None
-
-        ev = repo.transition_status(ev.id, IntelligentEvalStatus.COMPLETED)
-        assert ev.status == IntelligentEvalStatus.COMPLETED
-        assert ev.completed_at is not None
-
-    def test_plan_and_report_json(self, db_session):
-        repo = IntelligentEvalRepository(db_session)
-        ev = repo.create(IntelligentEval(name="eval", target_id="t1"))
-
-        plan = {"dimensions": ["退货"], "virtual_users": [], "estimated_sessions": 3}
-        ev.plan = plan
-        ev.status = IntelligentEvalStatus.PENDING_APPROVAL
-        ev = repo.update(ev)
-        assert ev.plan == plan
-
-        report = {"summary": "good", "findings": []}
-        ev.report = report
-        ev = repo.update(ev)
-        assert ev.report == report
+def _start(session, *, name: str = "智能评估"):
+    evaluation = _create(session, name=name)
+    lifecycle.submit_plan(session, evaluation.id, {"dimensions": ["退货"]})
+    return lifecycle.approve(session, evaluation.id)
 
 
-class TestIntelligentEvalSessionRepository:
-    def test_create_and_list(self, db_session):
-        eval_repo = IntelligentEvalRepository(db_session)
-        ev = eval_repo.create(IntelligentEval(name="eval", target_id="t1"))
+def test_lifecycle_owns_complete_evaluation_state_machine(eval_session) -> None:
+    evaluation = _create(eval_session)
+    assert evaluation.status is IntelligentEvalStatus.PLANNING
 
-        sess_repo = IntelligentEvalSessionRepository(db_session)
-        sess = sess_repo.create(IntelligentEvalSession(
-            eval_id=ev.id,
-            target_id="t1",
-            persona={"name": "user1", "patience": "low"},
-            goal="complete return",
-            dimension="退货流程",
-        ))
-        assert sess.id is not None
-        assert sess.status == IntelligentEvalSessionStatus.RUNNING
-        assert sess.persona == {"name": "user1", "patience": "low"}
+    pending = lifecycle.submit_plan(eval_session, evaluation.id, {"dimensions": ["退货"]})
+    assert pending.status is IntelligentEvalStatus.PENDING_APPROVAL
+    assert pending.plan == {"dimensions": ["退货"]}
 
-        sessions = sess_repo.list_by_eval(ev.id)
-        assert len(sessions) == 1
+    executing = lifecycle.approve(eval_session, evaluation.id)
+    assert executing.status is IntelligentEvalStatus.EXECUTING
+    assert executing.started_at is not None
 
-    def test_close_session(self, db_session):
-        eval_repo = IntelligentEvalRepository(db_session)
-        ev = eval_repo.create(IntelligentEval(name="eval", target_id="t1"))
-
-        sess_repo = IntelligentEvalSessionRepository(db_session)
-        sess = sess_repo.create(IntelligentEvalSession(eval_id=ev.id, target_id="t1"))
-
-        verdict = {"goal_achieved": True, "issues": []}
-        closed = sess_repo.close(sess.id, verdict)
-        assert closed.status == IntelligentEvalSessionStatus.COMPLETED
-        assert closed.verdict == verdict
-        assert closed.closed_at is not None
-
-    def test_increment_turns(self, db_session):
-        eval_repo = IntelligentEvalRepository(db_session)
-        ev = eval_repo.create(IntelligentEval(name="eval", target_id="t1"))
-
-        sess_repo = IntelligentEvalSessionRepository(db_session)
-        sess = sess_repo.create(IntelligentEvalSession(eval_id=ev.id, target_id="t1"))
-        assert sess.turn_count == 0
-
-        sess_repo.increment_turns(sess.id)
-        sess_repo.increment_turns(sess.id)
-        updated = sess_repo.get(sess.id)
-        assert updated.turn_count == 2
+    completed = lifecycle.submit_report(eval_session, evaluation.id, {"summary": "done"})
+    assert completed.status is IntelligentEvalStatus.COMPLETED
+    assert completed.report == {"summary": "done"}
+    assert completed.completed_at is not None
 
 
-class TestIntelligentEvalMessageRepository:
-    def test_create_and_list(self, db_session):
-        eval_repo = IntelligentEvalRepository(db_session)
-        ev = eval_repo.create(IntelligentEval(name="eval", target_id="t1"))
+def test_lifecycle_rejects_invalid_or_repeated_transitions(eval_session) -> None:
+    evaluation = _create(eval_session)
 
-        sess_repo = IntelligentEvalSessionRepository(db_session)
-        sess = sess_repo.create(IntelligentEvalSession(eval_id=ev.id, target_id="t1"))
+    with pytest.raises(IntelligentEvalTransitionError):
+        lifecycle.approve(eval_session, evaluation.id)
 
-        msg_repo = IntelligentEvalMessageRepository(db_session)
-        msg_repo.create(IntelligentEvalMessage(session_id=sess.id, role="user", content="hello"))
-        msg_repo.create(IntelligentEvalMessage(session_id=sess.id, role="assistant", content="hi", latency_ms=120))
+    lifecycle.submit_plan(eval_session, evaluation.id, {"dimensions": ["退货"]})
+    with pytest.raises(IntelligentEvalTransitionError):
+        lifecycle.submit_plan(eval_session, evaluation.id, {"dimensions": ["重复"]})
 
-        messages = msg_repo.list_by_session(sess.id)
-        assert len(messages) == 2
-        assert messages[0].role == "user"
-        assert messages[1].role == "assistant"
-        assert messages[1].latency_ms == 120
+
+def test_plan_transaction_failure_leaves_public_snapshot_unchanged(eval_session, monkeypatch) -> None:
+    evaluation = _create(eval_session)
+    monkeypatch.setattr(eval_session, "commit", lambda: (_ for _ in ()).throw(RuntimeError("commit failed")))
+
+    with pytest.raises(RuntimeError, match="commit failed"):
+        lifecycle.submit_plan(eval_session, evaluation.id, {"dimensions": ["退货"]})
+
+    monkeypatch.undo()
+    unchanged = lifecycle.get_eval(eval_session, evaluation.id)
+    assert unchanged.status is IntelligentEvalStatus.PLANNING
+    assert unchanged.plan is None
+
+
+def test_session_ownership_and_close_invariants_are_lifecycle_rules(eval_session) -> None:
+    first = _start(eval_session, name="first")
+    second = _start(eval_session, name="second")
+    session_obj = lifecycle.open_session(
+        eval_session,
+        eval_id=first.id,
+        persona={"name": "老客户"},
+        goal="完成退货",
+        dimension="退货",
+    )
+
+    with pytest.raises(IntelligentEvalNotFoundError):
+        lifecycle.close_session(
+            eval_session,
+            eval_id=second.id,
+            session_id=session_obj.id,
+            verdict={"goal_achieved": False},
+        )
+
+    closed = lifecycle.close_session(
+        eval_session,
+        eval_id=first.id,
+        session_id=session_obj.id,
+        verdict={"goal_achieved": True},
+    )
+    assert closed.status is IntelligentEvalSessionStatus.COMPLETED
+    assert closed.verdict == {"goal_achieved": True}
+
+
+class _ReplyChannel:
+    async def exchange(self, _content, *, on_sent, **_kwargs):
+        await on_sent(SendResult(ok=True, question_msg_id="message-1"))
+        return ExchangeOutcome.succeeded(correlation_id="message-1", reply="答复", latency_ms=12)
+
+
+class _TimeoutChannel:
+    async def exchange(self, _content, *, on_sent, **_kwargs):
+        await on_sent(SendResult(ok=True, question_msg_id="message-1"))
+        return ExchangeOutcome.reply_timeout(correlation_id="message-1", latency_ms=30_000)
+
+
+async def test_turn_ledger_is_observable_only_through_lifecycle(eval_session, monkeypatch) -> None:
+    evaluation = _start(eval_session)
+    session_obj = lifecycle.open_session(eval_session, eval_id=evaluation.id, persona={}, goal="完成退货")
+    monkeypatch.setattr(lifecycle.ChannelFactory, "create", lambda _target: _ReplyChannel())
+
+    result = await lifecycle.conduct_turn(
+        eval_session,
+        eval_id=evaluation.id,
+        session_id=session_obj.id,
+        content="你好",
+    )
+
+    messages = lifecycle.list_messages(eval_session, eval_id=evaluation.id, session_id=session_obj.id)
+    sessions = lifecycle.list_sessions(eval_session, evaluation.id)
+    assert result == {"reply": "答复", "latency_ms": 12, "turn_count": 1}
+    assert [(item.role, item.content) for item in messages] == [("user", "你好"), ("assistant", "答复")]
+    assert sessions[0].turn_count == 1
+
+
+async def test_timeout_preserves_sent_message_and_turn_count(eval_session, monkeypatch) -> None:
+    evaluation = _start(eval_session)
+    session_obj = lifecycle.open_session(eval_session, eval_id=evaluation.id, persona={}, goal="完成退货")
+    monkeypatch.setattr(lifecycle.ChannelFactory, "create", lambda _target: _TimeoutChannel())
+
+    with pytest.raises(IntelligentEvalChannelError, match="超时"):
+        await lifecycle.conduct_turn(
+            eval_session,
+            eval_id=evaluation.id,
+            session_id=session_obj.id,
+            content="你好",
+        )
+
+    messages = lifecycle.list_messages(eval_session, eval_id=evaluation.id, session_id=session_obj.id)
+    sessions = lifecycle.list_sessions(eval_session, evaluation.id)
+    assert [(item.role, item.content) for item in messages] == [("user", "你好")]
+    assert sessions[0].turn_count == 1

tests/unit/test_intelligent_eval_read_model.py

@@ -0,0 +1,125 @@
+"""Pure contract tests for intelligent-evaluation read projections."""
+
+from datetime import datetime, timezone
+
+from agenteval.intelligent_eval.models import (
+    IntelligentEval,
+    IntelligentEvalSession,
+    IntelligentEvalSessionStatus,
+    IntelligentEvalStatus,
+)
+from agenteval.intelligent_eval.read_model import IntelligentEvalReadModel, project_detail, project_list_item
+from agenteval.intelligent_eval.repository import IntelligentEvalRepository, IntelligentEvalSessionRepository
+from sqlalchemy import event
+from sqlmodel import Session
+
+
+def _evaluation(eval_id: str = "eval-1") -> IntelligentEval:
+    now = datetime(2026, 1, 1, tzinfo=timezone.utc)
+    return IntelligentEval(
+        id=eval_id,
+        name="客服评估",
+        target_id="target-1",
+        status=IntelligentEvalStatus.EXECUTING,
+        goal="验证退货流程",
+        seeds={"personas": ["老客户"]},
+        intent="流程覆盖",
+        role_description="模拟用户",
+        plan={"dimensions": ["退货"]},
+        time_window_hours=24,
+        created_at=now,
+        updated_at=now,
+    )
+
+
+def _session(
+    session_id: str,
+    status: IntelligentEvalSessionStatus,
+    eval_id: str = "eval-1",
+) -> IntelligentEvalSession:
+    return IntelligentEvalSession(
+        id=session_id,
+        eval_id=eval_id,
+        target_id="target-1",
+        persona={"name": session_id},
+        goal="完成退货",
+        dimension="退货",
+        status=status,
+        turn_count=2,
+    )
+
+
+def test_list_projection_contains_stable_fields_and_counts() -> None:
+    projection = project_list_item(
+        _evaluation(),
+        [_session("s-1", IntelligentEvalSessionStatus.RUNNING), _session("s-2", IntelligentEvalSessionStatus.COMPLETED)],
+    )
+
+    assert projection.id == "eval-1"
+    assert projection.session_count == 2
+    assert projection.completed_sessions == 1
+    assert projection.plan == {"dimensions": ["退货"]}
+
+
+def test_detail_projection_contains_session_metadata_without_messages() -> None:
+    projection = project_detail(_evaluation(), [_session("s-1", IntelligentEvalSessionStatus.COMPLETED)])
+    payload = projection.model_dump(mode="json")
+
+    assert payload["sessions"][0]["id"] == "s-1"
+    assert payload["sessions"][0]["turn_count"] == 2
+    assert "messages" not in payload
+    assert "content" not in payload["sessions"][0]
+
+
+def _count_queries(session: Session, action) -> int:
+    count = 0
+
+    def before_cursor_execute(*_args) -> None:
+        nonlocal count
+        count += 1
+
+    engine = session.get_bind()
+    event.listen(engine, "before_cursor_execute", before_cursor_execute)
+    try:
+        action()
+    finally:
+        event.remove(engine, "before_cursor_execute", before_cursor_execute)
+    return count
+
+
+def test_empty_list_projection_does_not_query_sessions(db_session: Session) -> None:
+    reader = IntelligentEvalReadModel(db_session)
+
+    query_count = _count_queries(db_session, lambda: reader.list_items([]))
+
+    assert query_count == 0
+
+
+def test_list_projection_loads_all_sessions_with_one_query(db_session: Session) -> None:
+    sessions = IntelligentEvalSessionRepository(db_session)
+    sessions._create(_session("s-1", IntelligentEvalSessionStatus.RUNNING, "eval-1"))
+    sessions._create(_session("s-2", IntelligentEvalSessionStatus.COMPLETED, "eval-2"))
+    reader = IntelligentEvalReadModel(db_session)
+    evaluations = [_evaluation("eval-1"), _evaluation("eval-2")]
+    result: list = []
+
+    query_count = _count_queries(db_session, lambda: result.extend(reader.list_items(evaluations)))
+
+    assert query_count == 1
+    assert [(item.id, item.session_count) for item in result] == [("eval-1", 1), ("eval-2", 1)]
+
+
+def test_detail_projection_uses_one_snapshot_query(db_session: Session) -> None:
+    evaluations = IntelligentEvalRepository(db_session)
+    sessions = IntelligentEvalSessionRepository(db_session)
+    evaluations._create(_evaluation())
+    sessions._create(_session("s-1", IntelligentEvalSessionStatus.COMPLETED))
+    reader = IntelligentEvalReadModel(db_session)
+    result: list = []
+
+    query_count = _count_queries(db_session, lambda: result.append(reader.detail_by_id("eval-1")))
+
+    assert query_count == 1
+    assert result[0] is not None
+    assert result[0].id == "eval-1"
+    assert [item.id for item in result[0].sessions] == ["s-1"]

tests/unit/test_llm_task_lifecycle.py

@@ -98,6 +98,33 @@ def test_mark_orphans_failed_flips_generating_analysis(db_session):
     assert repo.get_by_campaign("c-done").status == "completed"
 
 
+def test_enqueue_analysis_persists_before_launch(db_session, monkeypatch):
+    from agenteval.evaluation import analysis
+
+    launched = []
+    monkeypatch.setattr(analysis, "get_session", lambda: db_session)
+    monkeypatch.setattr(analysis, "start_campaign_analysis", lambda cid, *, triggered_by: launched.append((cid, triggered_by)))
+
+    analysis.enqueue_campaign_analysis("c-queued", triggered_by="auto")
+
+    assert launched == [("c-queued", "auto")]
+    row = CampaignAnalysisRepository(db_session).get_by_campaign("c-queued")
+    assert row.status == "queued"
+    assert row.triggered_by == "auto"
+
+
+def test_resume_queued_analysis_relaunches_persisted_jobs(db_session, monkeypatch):
+    from agenteval.evaluation import analysis
+
+    repo = CampaignAnalysisRepository(db_session)
+    repo.enqueue("c-queued", triggered_by="auto")
+    launched = []
+    monkeypatch.setattr(analysis, "start_campaign_analysis", lambda cid, *, triggered_by: launched.append((cid, triggered_by)))
+
+    assert analysis.resume_queued_campaign_analysis(db_session) == 1
+    assert launched == [("c-queued", "auto")]
+
+
 def test_mark_orphans_failed_flips_generating_comparison(db_session):
     repo = CampaignPeriodComparisonRepository(db_session)
     repo.upsert("c-gen", status="generating", baseline_campaign_id="b-1", triggered_by="auto")

tests/unit/test_orphan_runs.py

@@ -1,7 +1,7 @@
 """Startup cleanup of orphan runs (interrupted by server restart)."""
 
-from agenteval.models import EvalRun, RunStatus
-from agenteval.storage.repository import RunRepository
+from agenteval.models import Campaign, CampaignPlanEntry, CampaignStatus, EvalRun, RunStatus
+from agenteval.storage.repository import CampaignRepository, RunRepository
 
 
 def _make_run(session, status: RunStatus) -> str:
@@ -38,6 +38,46 @@ def test_mark_orphans_failed_noop_when_clean(db_session):
     assert repo.mark_orphans_failed() == 0
 
 
+def test_mark_orphans_preserves_only_recoverable_campaign_pending(db_session):
+    campaign = CampaignRepository(db_session).create(
+        Campaign(
+            id="campaign-1",
+            name="running",
+            target_id="t-1",
+            status=CampaignStatus.RUNNING,
+            window_seconds=60,
+            plan=[CampaignPlanEntry(scenario_id="s-1", offset_seconds=0, count=2)],
+        )
+    )
+    repo = RunRepository(db_session)
+    pending = repo.create(
+        EvalRun(
+            target_id="t-1",
+            scenario_id="s-1",
+            campaign_id=campaign.id,
+            campaign_plan_index=0,
+            campaign_occurrence_index=0,
+            status=RunStatus.PENDING,
+        )
+    )
+    running = repo.create(
+        EvalRun(
+            target_id="t-1",
+            scenario_id="s-1",
+            campaign_id=campaign.id,
+            campaign_plan_index=0,
+            campaign_occurrence_index=1,
+            status=RunStatus.RUNNING,
+        )
+    )
+
+    count = repo.mark_orphans_failed()
+
+    assert count == 1
+    assert repo.get(pending.id).status is RunStatus.PENDING
+    assert repo.get(running.id).status is RunStatus.FAILED
+
+
 def test_update_preserves_scenario_version_and_triggered_by(db_session):
     """update() 不得丢字段：scenario_version / triggered_by 必须回写（漂移回归）。"""
     from agenteval.models import RunTrigger

tests/unit/test_production_entrypoint.py

@@ -0,0 +1,41 @@
+import os
+import sqlite3
+import subprocess
+import sys
+from pathlib import Path
+
+
+def _write_executable(path: Path, content: str) -> None:
+    path.write_text(content)
+    path.chmod(0o755)
+
+
+def test_legacy_database_starts_migrations_from_pre_alembic_baseline(tmp_path: Path) -> None:
+    database_path = tmp_path / "legacy.db"
+    connection = sqlite3.connect(database_path)
+    connection.execute("CREATE TABLE scenarios (id TEXT PRIMARY KEY, name TEXT NOT NULL)")
+    connection.close()
+
+    command_log = tmp_path / "commands.log"
+    bin_dir = tmp_path / "bin"
+    bin_dir.mkdir()
+    recorder = '#!/bin/sh\nprintf "%s\\n" "$*" >> "$COMMAND_LOG"\n'
+    _write_executable(bin_dir / "alembic", recorder)
+    _write_executable(bin_dir / "agenteval", recorder)
+    _write_executable(bin_dir / "python", f'#!/bin/sh\nexec "{sys.executable}" "$@"\n')
+
+    environment = {
+        **os.environ,
+        "AGENTEVAL_DB_PATH": str(database_path),
+        "COMMAND_LOG": str(command_log),
+        "PATH": f"{bin_dir}:{os.environ['PATH']}",
+    }
+    script = Path(__file__).resolve().parents[2] / "scripts" / "production-entrypoint.sh"
+
+    subprocess.run(["sh", str(script)], check=True, env=environment)
+
+    assert command_log.read_text().splitlines() == [
+        "stamp base",
+        "upgrade head",
+        "server start --host 0.0.0.0 --port 8000",
+    ]

tests/unit/test_repository.py

@@ -19,9 +19,11 @@ from agenteval.models import (
     RunStatus,
     RunSummary,
     TargetStatus,
+    Turn,
 )
 from agenteval.storage.repository import (
     CampaignRepository,
+    ResultRepository,
     RunRepository,
     TargetRepository,
 )
@@ -90,6 +92,28 @@ def test_run_summary_json_round_trip(db_session):
     assert fetched.summary.pass_rate == 0.5
 
 
+def test_campaign_child_identity_round_trip(db_session):
+    """Campaign plan identity survives the repository's model conversion seam."""
+    TargetRepository(db_session).create(_make_target())
+    repo = RunRepository(db_session)
+    repo.create(
+        EvalRun(
+            id="child-1",
+            target_id="t-1",
+            scenario_id="s-1",
+            campaign_id="campaign-1",
+            campaign_plan_index=2,
+            campaign_occurrence_index=1,
+        )
+    )
+
+    fetched = repo.get("child-1")
+    assert fetched is not None
+    assert fetched.campaign_id == "campaign-1"
+    assert fetched.campaign_plan_index == 2
+    assert fetched.campaign_occurrence_index == 1
+
+
 def test_campaign_crud_and_plan_round_trip(db_session):
     TargetRepository(db_session).create(_make_target())
     repo = CampaignRepository(db_session)
@@ -111,3 +135,48 @@ def test_campaign_crud_and_plan_round_trip(db_session):
     # Campaign inherits the shared delete() from the base repository.
     assert repo.delete("cp-1") is True
     assert repo.get("cp-1") is None
+
+
+def test_update_turn_exchange_preserves_sent_fact(db_session):
+    TargetRepository(db_session).create(_make_target())
+    RunRepository(db_session).create(EvalRun(id="r-1", target_id="t-1", scenario_id="s-1", status=RunStatus.RUNNING))
+    repo = ResultRepository(db_session)
+    repo.save_turn(
+        Turn(
+            id="turn-1",
+            run_id="r-1",
+            case_id="case-1",
+            round_index=1,
+            sent_message={"msgType": "text", "msgBody": {"content": "hello"}},
+            sent_at=datetime(2026, 8, 6, 1, 0, tzinfo=timezone.utc),
+        )
+    )
+
+    updated = repo.update_turn_exchange(
+        "turn-1",
+        question_msg_id="question-1",
+        reply={"msgBody": {"content": "world"}},
+        received_at=datetime(2026, 8, 6, 1, 0, 1, tzinfo=timezone.utc),
+        latency_ms=1000,
+    )
+
+    assert updated is not None
+    assert updated.get_sent_message() == {"msgType": "text", "msgBody": {"content": "hello"}}
+    assert updated.case_id == "case-1"
+    assert updated.round_index == 1
+    assert updated.question_msg_id == "question-1"
+    assert updated.get_reply() == {"msgBody": {"content": "world"}}
+    assert updated.latency_ms == 1000
+
+
+def test_update_turn_exchange_returns_none_for_unknown_turn(db_session):
+    assert (
+        ResultRepository(db_session).update_turn_exchange(
+            "missing",
+            question_msg_id="question-1",
+            reply=None,
+            received_at=None,
+            latency_ms=None,
+        )
+        is None
+    )