v0.3-s1: 规则层异步化 + 工具函数去重 + HTTP 通道

## 核心变更

### 规则层全面异步化（DEBT-1）
- EvalRule.evaluate() 签名改为 async def，全量同步改造（无兼容层）
- LlmScoreRule._call_llm: requests.post → httpx.AsyncClient，彻底消除事件循环阻塞
- engine._save_rule_results: rule.evaluate() → await rule.evaluate()

### 工具函数去重（DEBT-2）
- 新建 agenteval/utils/llm.py，统一三个函数：
  - extract_reply_text (原 5 处重复)
  - extract_content_from_llm_response (原 2 处重复)
  - parse_json_from_llm_text (统一 LLM 输出 JSON 解析)
- engine.py / llm_score.py / runs.py / report.py 全部切换到 utils.llm

### HTTP 通用通道（S1-3）
- 新建 channels/http.py (HttpChannel)
  - 配置化 send_url / reply_url 模板 ({message}, {msg_id} 占位)
  - dot-path 提取 msg_id 和 reply_text
  - 可选 reply_ready_path 就绪标志
  - 长连接 AsyncClient 复用
- ChannelFactory 注册 ChannelType.HTTP → HttpChannel

### 测试
- 新增 tests/unit/test_http_channel_and_rules.py (19 个测试)
- _get_path / health_check / send / poll_reply / 超时 / 就绪标志 / async 规则评估
- 测试总数：24 → 43，全部通过

Co-Authored-By: Claude <noreply@anthropic.com>

backend/agenteval/channels/factory.py

@@ -1,6 +1,7 @@
 """Factory for creating message channels from target configuration."""
 
 from agenteval.channels.base import EvalChannel
+from agenteval.channels.http import HttpChannel
 from agenteval.channels.tutu import TutuApiChannel
 from agenteval.models import ChannelType, EvalTarget
 
@@ -10,6 +11,7 @@ class ChannelFactory:
 
     _mapping: dict[ChannelType, type[EvalChannel]] = {
         ChannelType.TUTU_API: TutuApiChannel,
+        ChannelType.HTTP: HttpChannel,
     }
 
     @classmethod

backend/agenteval/channels/http.py

@@ -0,0 +1,132 @@
+"""Generic HTTP evaluation channel.
+
+Supports any REST API that follows a configurable request/response template.
+Configuration keys (all under channel_config):
+
+Required:
+  send_url       URL to POST the message to (supports {message} template var)
+  reply_url      URL to GET the reply from (supports {msg_id} template var)
+
+Optional:
+  health_url     URL for health check GET (defaults to send_url)
+  headers        dict of extra request headers (e.g. Authorization)
+  send_body      JSON body template for POST; {message} is substituted
+                 Default: {"text": "{message}"}
+  msg_id_path    dot-separated path to extract msg_id from send response
+                 Default: "id"
+  reply_path     dot-separated path to extract reply text from reply response
+                 Default: "text"
+  reply_ready_path  dot-separated path to a boolean indicating reply is ready
+                 Default: (treat any 200 as ready)
+  poll_interval  seconds between polls (default 1.0)
+  timeout        seconds before poll gives up (default 30.0)
+"""
+
+import asyncio
+import json
+import uuid
+from typing import Any, Optional
+
+import httpx
+
+from agenteval.channels.base import ChannelHealth, EvalChannel, Reply, SendResult
+
+
+def _get_path(data: Any, path: str) -> Any:
+    """Traverse a dot-separated key path into a nested dict/list."""
+    parts = path.split(".")
+    current = data
+    for part in parts:
+        if current is None:
+            return None
+        if isinstance(current, dict):
+            current = current.get(part)
+        elif isinstance(current, list) and part.isdigit():
+            idx = int(part)
+            current = current[idx] if idx < len(current) else None
+        else:
+            return None
+    return current
+
+
+class HttpChannel(EvalChannel):
+    """Generic HTTP channel: POST to send, GET/POST to poll for reply."""
+
+    def __init__(self, config: dict[str, Any]):
+        self.send_url: str = config["send_url"]
+        self.reply_url: str = config.get("reply_url", "")
+        self.health_url: str = config.get("health_url", self.send_url)
+        self.headers: dict[str, str] = config.get("headers", {})
+        self.send_body_template: str = config.get("send_body", '{"text": "{message}"}')
+        self.msg_id_path: str = config.get("msg_id_path", "id")
+        self.reply_path: str = config.get("reply_path", "text")
+        self.reply_ready_path: Optional[str] = config.get("reply_ready_path")
+        self._poll_interval: float = float(config.get("poll_interval", 1.0))
+
+        self._client = httpx.AsyncClient(headers=self.headers, timeout=30)
+
+    async def close(self) -> None:
+        await self._client.aclose()
+
+    async def health_check(self) -> ChannelHealth:
+        try:
+            resp = await self._client.get(self.health_url)
+            resp.raise_for_status()
+            return ChannelHealth(ok=True, message=f"HTTP {resp.status_code}")
+        except Exception as exc:
+            return ChannelHealth(ok=False, message=str(exc))
+
+    async def send(self, content: str, **kwargs: Any) -> SendResult:
+        body_str = self.send_body_template.replace("{message}", content)
+        try:
+            body = json.loads(body_str)
+        except json.JSONDecodeError:
+            body = {"text": content}
+
+        url = self.send_url.replace("{message}", content)
+        try:
+            resp = await self._client.post(url, json=body)
+            resp.raise_for_status()
+            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:
+            return SendResult(ok=False, error=str(exc))
+
+    async def poll_reply(
+        self,
+        question_msg_id: str,
+        timeout: float = 30.0,
+        poll_interval: float = 1.0,
+    ) -> Optional[Reply]:
+        interval = poll_interval or self._poll_interval
+        deadline = asyncio.get_event_loop().time() + timeout
+        url = self.reply_url.replace("{msg_id}", question_msg_id)
+
+        while asyncio.get_event_loop().time() < deadline:
+            try:
+                resp = await self._client.get(url)
+                resp.raise_for_status()
+                data = resp.json()
+
+                # Check if reply is ready (optional readiness flag)
+                if self.reply_ready_path:
+                    ready = _get_path(data, self.reply_ready_path)
+                    if not ready:
+                        await asyncio.sleep(interval)
+                        continue
+
+                reply_text = _get_path(data, self.reply_path)
+                if reply_text is not None:
+                    raw = {"text": str(reply_text), "_raw": data}
+                    return Reply(
+                        question_msg_id=question_msg_id,
+                        content=str(reply_text),
+                        raw_message=raw,
+                    )
+            except Exception:
+                pass
+
+            await asyncio.sleep(interval)
+
+        return None

backend/agenteval/config/settings.py

@@ -16,7 +16,6 @@ from typing import Optional
 from pydantic import Field, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
-
 ROOT_DIR = Path(__file__).resolve().parent.parent.parent.parent
 ENV_FILE = ROOT_DIR / ".env"
 

backend/agenteval/evaluation/engine.py

@@ -6,10 +6,9 @@ via an ``asyncio.Event`` cancel token.
 """
 
 import asyncio
-import json
 import uuid
-from dataclasses import dataclass, field
-from typing import Any, Awaitable, Callable, Optional
+from dataclasses import dataclass
+from typing import Any, Callable, Optional
 
 import httpx
 
@@ -19,7 +18,7 @@ from agenteval.evaluation.rules import get_rule
 from agenteval.models import Case, CaseType, EvalResult, EvalRun, EvalTarget, RunStatus, Scenario, Turn
 from agenteval.storage.db import get_session, utc_now
 from agenteval.storage.repository import ResultRepository, RunRepository
-
+from agenteval.utils.llm import extract_content_from_llm_response, extract_reply_text, 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.
@@ -45,48 +44,6 @@ def _build_send_message(content: str) -> dict[str, Any]:
     }
 
 
-def _extract_content_from_api_response(data: dict) -> str:
-    """Extract text content from an LLM API response.
-
-    Handles both OpenAI format (choices[0].message.content as string)
-    and content-block-array format used by Anthropic-compatible APIs
-    (choices[0].message.content as list of {type, text/text} blocks).
-    """
-    try:
-        content = data["choices"][0]["message"]["content"]
-    except (KeyError, IndexError, TypeError):
-        return ""
-    if isinstance(content, str):
-        return content
-    if isinstance(content, list):
-        parts = []
-        for block in content:
-            if not isinstance(block, dict):
-                continue
-            # Anthropic format: {"type": "text", "text": "..."}
-            # Some APIs may use "content" key instead of "text"
-            if block.get("type") == "text":
-                parts.append(block.get("text") or block.get("content") or "")
-            else:
-                # Non-text blocks (tool_use, image, etc.) — skip
-                pass
-        return "\n".join(parts)
-    return str(content)
-
-
-def _extract_reply_text(reply: Any) -> str:
-    if reply is None:
-        return ""
-    if isinstance(reply, str):
-        return reply
-    if isinstance(reply, dict):
-        body = reply.get("msgBody") or reply.get("content", "")
-        if isinstance(body, dict):
-            return body.get("content", "")
-        return str(body)
-    return str(reply)
-
-
 class EvalEngine:
     """Execute evaluation scenarios against targets.
 
@@ -151,22 +108,37 @@ class EvalEngine:
 
             for idx, case in enumerate(self.scenario.cases, start=1):
                 self._check_cancel()
-                await self._emit(progress_callback, "case_start", {
-                    "index": idx, "total": total_cases, "case_id": case.id,
-                })
+                await self._emit(
+                    progress_callback,
+                    "case_start",
+                    {
+                        "index": idx,
+                        "total": total_cases,
+                        "case_id": case.id,
+                    },
+                )
                 async with self._case_semaphore:
                     case_passed, rule_pass, rule_total = await self._run_case(
-                        run, case, progress_callback,
+                        run,
+                        case,
+                        progress_callback,
                     )
                 if case_passed:
                     passed_cases += 1
                 else:
                     failed_cases += 1
-                await self._emit(progress_callback, "case_end", {
-                    "index": idx, "total": total_cases, "case_id": case.id,
-                    "passed": case_passed,
-                    "rule_pass_count": rule_pass, "rule_total": rule_total,
-                })
+                await self._emit(
+                    progress_callback,
+                    "case_end",
+                    {
+                        "index": idx,
+                        "total": total_cases,
+                        "case_id": case.id,
+                        "passed": case_passed,
+                        "rule_pass_count": rule_pass,
+                        "rule_total": rule_total,
+                    },
+                )
 
             results = self.run_repo.get_results(run.id)
             total_rules = len(results)
@@ -183,27 +155,42 @@ class EvalEngine:
             run.status = RunStatus.COMPLETED
             run.completed_at = utc_now()
             run.summary = summary
-            await self._emit(progress_callback, "run_completed", {
-                "status": "completed", "summary": summary,
-            })
+            await self._emit(
+                progress_callback,
+                "run_completed",
+                {
+                    "status": "completed",
+                    "summary": summary,
+                },
+            )
         except CancelledError:
             run.status = RunStatus.FAILED
             run.completed_at = utc_now()
             run.summary = {
                 "error": {"code": "cancelled_by_user", "message": "评测已手动停止"},
             }
-            await self._emit(progress_callback, "run_completed", {
-                "status": "failed", "reason": "cancelled",
-                "error": {"code": "cancelled_by_user", "message": "评测已手动停止"},
-            })
+            await self._emit(
+                progress_callback,
+                "run_completed",
+                {
+                    "status": "failed",
+                    "reason": "cancelled",
+                    "error": {"code": "cancelled_by_user", "message": "评测已手动停止"},
+                },
+            )
         except Exception as exc:
             run.status = RunStatus.FAILED
             run.completed_at = utc_now()
             run.summary = {"error": str(exc)}
             await self._emit(progress_callback, "error", {"error": str(exc)})
-            await self._emit(progress_callback, "run_completed", {
-                "status": "failed", "error": str(exc),
-            })
+            await self._emit(
+                progress_callback,
+                "run_completed",
+                {
+                    "status": "failed",
+                    "error": str(exc),
+                },
+            )
             raise
         finally:
             run = self.run_repo.update(run) or run
@@ -235,9 +222,14 @@ class EvalEngine:
         if case.type == CaseType.DYNAMIC:
             generated = await self._generate_messages(case, progress_callback)
             if not generated:
-                await self._emit(progress_callback, "error", {
-                    "error": "LLM 未能生成测试消息", "case_id": case.id,
-                })
+                await self._emit(
+                    progress_callback,
+                    "error",
+                    {
+                        "error": "LLM 未能生成测试消息",
+                        "case_id": case.id,
+                    },
+                )
                 return False, 0, 0
             case = case.model_copy(update={"messages": generated})
 
@@ -245,24 +237,39 @@ class EvalEngine:
 
         for round_index, message in enumerate(case.messages, start=1):
             self._check_cancel()
-            await self._emit(progress_callback, "turn_start", {
-                "run_id": run.id, "case_id": case.id,
-                "round": round_index, "message": message,
-            })
+            await self._emit(
+                progress_callback,
+                "turn_start",
+                {
+                    "run_id": run.id,
+                    "case_id": case.id,
+                    "round": round_index,
+                    "message": message,
+                },
+            )
 
             sent_at = utc_now()
             send_result = await self.channel.send(message)
             if not send_result.ok:
                 turn = Turn(
                     id=str(uuid.uuid4()),
-                    run_id=run.id, case_id=case.id, round_index=round_index,
-                    sent_message=_build_send_message(message), sent_at=sent_at,
+                    run_id=run.id,
+                    case_id=case.id,
+                    round_index=round_index,
+                    sent_message=_build_send_message(message),
+                    sent_at=sent_at,
                 )
                 self.result_repo.save_turn(turn)
                 await self._save_rule_results(run, case, turn, [], progress_callback)
-                await self._emit(progress_callback, "turn_error", {
-                    "case_id": case.id, "round": round_index, "error": send_result.error,
-                })
+                await self._emit(
+                    progress_callback,
+                    "turn_error",
+                    {
+                        "case_id": case.id,
+                        "round": round_index,
+                        "error": send_result.error,
+                    },
+                )
                 return False, 0, 0
 
             try:
@@ -274,16 +281,24 @@ class EvalEngine:
                 received_at = utc_now()
                 turn = Turn(
                     id=str(uuid.uuid4()),
-                    run_id=run.id, case_id=case.id, round_index=round_index,
-                    sent_message=_build_send_message(message), sent_at=sent_at,
+                    run_id=run.id,
+                    case_id=case.id,
+                    round_index=round_index,
+                    sent_message=_build_send_message(message),
+                    sent_at=sent_at,
                     question_msg_id=send_result.question_msg_id,
                     received_at=received_at,
                 )
                 self.result_repo.save_turn(turn)
-                await self._emit(progress_callback, "turn_error", {
-                    "case_id": case.id, "round": round_index,
-                    "error": f"poll_reply 异常: {poll_exc}",
-                })
+                await self._emit(
+                    progress_callback,
+                    "turn_error",
+                    {
+                        "case_id": case.id,
+                        "round": round_index,
+                        "error": f"poll_reply 异常: {poll_exc}",
+                    },
+                )
                 return False, 0, 0
 
             received_at = utc_now()
@@ -293,20 +308,30 @@ class EvalEngine:
 
             turn = Turn(
                 id=str(uuid.uuid4()),
-                run_id=run.id, case_id=case.id, round_index=round_index,
-                sent_message=_build_send_message(message), sent_at=sent_at,
+                run_id=run.id,
+                case_id=case.id,
+                round_index=round_index,
+                sent_message=_build_send_message(message),
+                sent_at=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,
+                received_at=received_at,
+                latency_ms=latency_ms,
             )
             self.result_repo.save_turn(turn)
             dialog.append(turn)
 
-            await self._emit(progress_callback, "turn_end", {
-                "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),
-            })
+            await self._emit(
+                progress_callback,
+                "turn_end",
+                {
+                    "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),
+                },
+            )
 
         if not dialog:
             return False, 0, 0
@@ -329,9 +354,14 @@ class EvalEngine:
         # If no explicit rules, derive implicit rules from expectations.
         if not rules_config:
             if case.expectations.response_time_max_ms:
-                rules_config.append(EvalRuleConfig(type="response_time", params={
-                    "max_ms": case.expectations.response_time_max_ms,
-                }))
+                rules_config.append(
+                    EvalRuleConfig(
+                        type="response_time",
+                        params={
+                            "max_ms": case.expectations.response_time_max_ms,
+                        },
+                    )
+                )
             if case.expectations.keywords_include or case.expectations.keywords_exclude:
                 rules_config.append(
                     EvalRuleConfig(
@@ -348,12 +378,16 @@ class EvalEngine:
         total_count = 0
         for rule_config in rules_config:
             rule = get_rule(rule_config.type, rule_config.params)
-            result = rule.evaluate(case, dialog)
+            result = await rule.evaluate(case, dialog)
             eval_result = EvalResult(
                 id=str(uuid.uuid4()),
-                run_id=run.id, case_id=case.id, turn_id=turn.id or "",
+                run_id=run.id,
+                case_id=case.id,
+                turn_id=turn.id or "",
                 rule_type=rule_config.type,
-                passed=result.passed, score=result.score, reason=result.reason,
+                passed=result.passed,
+                score=result.score,
+                reason=result.reason,
             )
             self.result_repo.save_result(eval_result)
             total_count += 1
@@ -361,23 +395,36 @@ class EvalEngine:
                 passed_count += 1
             else:
                 all_passed = False
-            await self._emit(progress_callback, "rule_result", {
-                "run_id": run.id, "case_id": case.id,
-                "rule_type": rule_config.type,
-                "passed": result.passed, "score": result.score, "reason": result.reason,
-            })
+            await self._emit(
+                progress_callback,
+                "rule_result",
+                {
+                    "run_id": run.id,
+                    "case_id": case.id,
+                    "rule_type": rule_config.type,
+                    "passed": result.passed,
+                    "score": result.score,
+                    "reason": result.reason,
+                },
+            )
 
         return all_passed, passed_count, total_count
 
     async def _generate_messages(
-        self, case: Case, progress_callback: Optional[ProgressCallback],
+        self,
+        case: Case,
+        progress_callback: Optional[ProgressCallback],
     ) -> list[str]:
         """Use LLM to generate test messages for dynamic cases."""
         llm_config = self.scenario.llm_config
         if not llm_config:
-            await self._emit(progress_callback, "error", {
-                "error": "动态用例需要配置 llm_config",
-            })
+            await self._emit(
+                progress_callback,
+                "error",
+                {
+                    "error": "动态用例需要配置 llm_config",
+                },
+            )
             return []
 
         api_url = llm_config.get("api_url")
@@ -415,25 +462,28 @@ class EvalEngine:
             async with httpx.AsyncClient(timeout=self.timeout_config.llm_generate) as client:
                 resp = await client.post(api_url, headers=headers, json=payload)
                 resp.raise_for_status()
-                data = resp.json()
-            content = _extract_content_from_api_response(data)
+            content = extract_content_from_llm_response(resp.json())
             if not content:
-                await self._emit(progress_callback, "error", {
-                    "error": "LLM 返回内容为空或无法解析",
-                })
+                await self._emit(
+                    progress_callback,
+                    "error",
+                    {
+                        "error": "LLM 返回内容为空或无法解析",
+                    },
+                )
                 return []
 
             try:
-                parsed = json.loads(content)
-            except json.JSONDecodeError:
-                start = content.find("[")
-                end = content.rfind("]")
-                if start == -1 or end == -1:
-                    await self._emit(progress_callback, "error", {
-                        "error": f"LLM 返回无法解析为数组: {content[:200]}",
-                    })
-                    return []
-                parsed = json.loads(content[start:end + 1])
+                parsed = parse_json_from_llm_text(content)
+            except (ValueError, Exception) as parse_exc:
+                await self._emit(
+                    progress_callback,
+                    "error",
+                    {
+                        "error": f"LLM 返回无法解析为数组: {parse_exc}",
+                    },
+                )
+                return []
 
             if not isinstance(parsed, list):
                 await self._emit(progress_callback, "error", {"error": "LLM 返回的不是数组"})
@@ -444,9 +494,14 @@ class EvalEngine:
                 await self._emit(progress_callback, "error", {"error": "LLM 返回的消息为空"})
                 return []
 
-            await self._emit(progress_callback, "messages_generated", {
-                "case_id": case.id, "messages": messages,
-            })
+            await self._emit(
+                progress_callback,
+                "messages_generated",
+                {
+                    "case_id": case.id,
+                    "messages": messages,
+                },
+            )
             return messages
 
         except Exception as exc:
@@ -460,7 +515,10 @@ class EvalEngine:
             raise CancelledError("run cancelled")
 
     async def _emit(
-        self, callback: Optional[ProgressCallback], event: str, data: dict[str, Any],
+        self,
+        callback: Optional[ProgressCallback],
+        event: str,
+        data: dict[str, Any],
     ) -> None:
         if not callback:
             return

backend/agenteval/evaluation/report.py

@@ -9,6 +9,7 @@ from jinja2 import Template
 
 from agenteval.storage.db import DATA_DIR
 from agenteval.storage.repository import RunRepository, ScenarioRepository, TargetRepository
+from agenteval.utils.llm import extract_reply_text
 
 HTML_TEMPLATE = """<!DOCTYPE html>
 <html lang="zh-CN">
@@ -88,16 +89,7 @@ HTML_TEMPLATE = """<!DOCTYPE html>
 
 
 def _extract_text(data: Any) -> str:
-    if data is None:
-        return ""
-    if isinstance(data, str):
-        return data
-    if isinstance(data, dict):
-        body = data.get("msgBody") or data.get("content", "")
-        if isinstance(body, dict):
-            return body.get("content", "")
-        return str(body)
-    return str(data)
+    return extract_reply_text(data)
 
 
 def generate_report(run_id: str, session=None) -> dict[str, Any]:

backend/agenteval/evaluation/rules/base.py

@@ -17,7 +17,7 @@ class RuleResult:
 
 
 class EvalRule(ABC):
-    """Abstract evaluation rule."""
+    """Abstract evaluation rule. All implementations must be async."""
 
     name: str = ""
 
@@ -25,7 +25,7 @@ class EvalRule(ABC):
         self.params = params
 
     @abstractmethod
-    def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
+    async def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
         """Evaluate the dialog against the case expectation."""
         ...
 

backend/agenteval/evaluation/rules/keyword.py

@@ -2,21 +2,7 @@
 
 from agenteval.evaluation.rules.base import EvalRule, RuleResult, register_rule
 from agenteval.models import Case, Turn
-
-
-def _extract_text(reply) -> str:
-    """Extract plain text from a reply object for matching."""
-    if reply is None:
-        return ""
-    if isinstance(reply, str):
-        return reply
-    if isinstance(reply, dict):
-        # Tutu-api message structure: msgBody.content
-        body = reply.get("msgBody") or reply.get("content", "")
-        if isinstance(body, dict):
-            return body.get("content", "")
-        return str(body)
-    return str(reply)
+from agenteval.utils.llm import extract_reply_text
 
 
 @register_rule
@@ -25,13 +11,12 @@ class KeywordMatchRule(EvalRule):
 
     name = "keyword_match"
 
-    def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
+    async def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
         if not dialog:
             return RuleResult(passed=False, reason="无回复记录")
 
         last_turn = dialog[-1]
-        reply = last_turn.reply
-        text = _extract_text(reply).lower()
+        text = extract_reply_text(last_turn.reply).lower()
 
         params = self.params
         include = [k.lower() for k in params.get("keywords", [])]

backend/agenteval/evaluation/rules/llm_score.py

@@ -1,49 +1,12 @@
 """LLM-based scoring evaluation rule."""
 
 import json
-from typing import Any
 
-import requests
+import httpx
 
 from agenteval.evaluation.rules.base import EvalRule, RuleResult, register_rule
 from agenteval.models import Case, Turn
-
-
-def _extract_text(reply: Any) -> str:
-    if reply is None:
-        return ""
-    if isinstance(reply, str):
-        return reply
-    if isinstance(reply, dict):
-        body = reply.get("msgBody") or reply.get("content", "")
-        if isinstance(body, dict):
-            return body.get("content", "")
-        return str(body)
-    return str(reply)
-
-
-def _extract_content_from_api_response(data: dict) -> str:
-    """Extract text content from an LLM API response.
-
-    Handles both OpenAI format (choices[0].message.content as string)
-    and content-block-array format used by Anthropic-compatible APIs
-    (choices[0].message.content as list of {type, text/text} blocks).
-    """
-    try:
-        content = data["choices"][0]["message"]["content"]
-    except (KeyError, IndexError, TypeError):
-        return ""
-    if isinstance(content, str):
-        return content
-    if isinstance(content, list):
-        parts = []
-        for block in content:
-            if not isinstance(block, dict):
-                continue
-            if block.get("type") == "text":
-                parts.append(block.get("text") or block.get("content") or "")
-        return "\n".join(parts)
-    return str(content)
+from agenteval.utils.llm import extract_content_from_llm_response, extract_reply_text, parse_json_from_llm_text
 
 
 @register_rule
@@ -52,23 +15,23 @@ class LlmScoreRule(EvalRule):
 
     name = "llm_score"
 
-    def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
+    async def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
         if not dialog:
             return RuleResult(passed=False, reason="无回复记录")
 
         last_turn = dialog[-1]
-        reply_text = _extract_text(last_turn.reply)
+        reply_text = extract_reply_text(last_turn.reply)
+
         question_text = ""
         if len(dialog) >= 2:
-            question_text = _extract_text(dialog[-2].reply) or ""
+            question_text = extract_reply_text(dialog[-2].reply) or ""
         if not question_text and last_turn.sent_message:
             body = last_turn.sent_message.get("msgBody", "")
             if isinstance(body, dict):
                 question_text = body.get("content", "")
             else:
                 try:
-                    parsed = json.loads(body)
-                    question_text = parsed.get("content", "")
+                    question_text = json.loads(body).get("content", "")
                 except Exception:
                     question_text = str(body)
 
@@ -81,7 +44,7 @@ class LlmScoreRule(EvalRule):
         if not api_url:
             return RuleResult(passed=False, reason="LLM 评分规则未配置 api_url")
 
-        score, reason = self._call_llm(api_url, api_key, model, question_text, reply_text, criteria)
+        score, reason = await self._call_llm(api_url, api_key, model, question_text, reply_text, criteria)
         if score is None:
             return RuleResult(passed=False, reason=f"LLM 评分失败: {reason}")
 
@@ -92,7 +55,7 @@ class LlmScoreRule(EvalRule):
             reason=f"LLM 评分 {score}/10，{'通过' if passed else '未通过'} (阈值 {min_score})",
         )
 
-    def _call_llm(
+    async def _call_llm(
         self,
         api_url: str,
         api_key: str | None,
@@ -105,7 +68,7 @@ class LlmScoreRule(EvalRule):
         system_prompt = (
             "你是一位严格的智能客服质量评估专家。请根据用户问题和智能体回复，"
             f"按照以下标准打分（0-10分，10分最高）：{criteria}\n"
-            "只输出一个 JSON 对象：{\"score\": number, \"reason\": \"简短说明\"}"
+            '只输出一个 JSON 对象：{"score": number, "reason": "简短说明"}'
         )
         user_prompt = f"用户问题：{question}\n智能体回复：{reply}"
 
@@ -123,24 +86,14 @@ class LlmScoreRule(EvalRule):
         }
 
         try:
-            resp = requests.post(api_url, headers=headers, json=payload, timeout=60)
+            async with httpx.AsyncClient(timeout=60) as client:
+                resp = await client.post(api_url, headers=headers, json=payload)
             resp.raise_for_status()
-            data = resp.json()
-            content = _extract_content_from_api_response(data)
+            content = extract_content_from_llm_response(resp.json())
             if not content:
                 return None, "LLM 返回内容为空"
 
-            # Try to parse JSON from the content
-            try:
-                parsed = json.loads(content)
-            except json.JSONDecodeError:
-                # Fallback: extract JSON substring
-                start = content.find("{")
-                end = content.rfind("}")
-                if start == -1 or end == -1:
-                    return None, "LLM 返回格式无法解析"
-                parsed = json.loads(content[start : end + 1])
-
+            parsed = parse_json_from_llm_text(content)
             score = float(parsed["score"])
             reason = parsed.get("reason", "")
             return max(0.0, min(10.0, score)), reason

backend/agenteval/evaluation/rules/response_time.py

@@ -10,7 +10,7 @@ class ResponseTimeRule(EvalRule):
 
     name = "response_time"
 
-    def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
+    async def evaluate(self, case: Case, dialog: list[Turn]) -> RuleResult:
         if not dialog:
             return RuleResult(passed=False, reason="无回复记录")
 

backend/agenteval/storage/db.py

@@ -3,7 +3,6 @@
 import json
 import uuid
 from datetime import datetime, timezone
-import os
 from pathlib import Path
 from typing import Any, Optional
 

backend/agenteval/storage/file_repository.py

@@ -6,9 +6,9 @@ from typing import Optional
 from sqlmodel import Session, select
 
 from agenteval.storage.db import (
+    FILES_DIR,
     FileCategoryDB,
     FileRecordDB,
-    FILES_DIR,
     get_session,
     new_uuid,
     utc_now,
@@ -183,4 +183,4 @@ class FileRecordRepository:
             return
         file_path = FILES_DIR / record.storage_name
         if file_path.exists():
-            os.remove(file_path)
+            os.remove(file_path)

backend/agenteval/utils/llm.py

@@ -0,0 +1,72 @@
+"""Shared utilities for LLM API interaction and message extraction.
+
+Consolidates the duplicated _extract_text / _extract_reply_text pattern
+(previously repeated in 5 places) and _extract_content_from_api_response
+(previously duplicated in 2 places).
+"""
+
+from typing import Any
+
+
+def extract_reply_text(reply: Any) -> str:
+    """Extract plain text from a tutu-api reply object.
+
+    Handles: None | str | dict with msgBody.content or content key.
+    """
+    if reply is None:
+        return ""
+    if isinstance(reply, str):
+        return reply
+    if isinstance(reply, dict):
+        body = reply.get("msgBody") or reply.get("content", "")
+        if isinstance(body, dict):
+            return body.get("content", "")
+        return str(body)
+    return str(reply)
+
+
+def extract_content_from_llm_response(data: dict) -> str:
+    """Extract text content from an LLM API response dict.
+
+    Handles:
+    - OpenAI format: choices[0].message.content as a plain string
+    - Anthropic-compatible format: choices[0].message.content as a list of
+      content blocks {type: "text", text: "..."}  (non-text blocks are skipped)
+    """
+    try:
+        content = data["choices"][0]["message"]["content"]
+    except (KeyError, IndexError, TypeError):
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = []
+        for block in content:
+            if not isinstance(block, dict):
+                continue
+            if block.get("type") == "text":
+                parts.append(block.get("text") or block.get("content") or "")
+        return "\n".join(parts)
+    return str(content)
+
+
+def parse_json_from_llm_text(content: str) -> Any:
+    """Parse JSON from LLM output, falling back to bracket-delimited substring.
+
+    Returns parsed JSON object, or raises json.JSONDecodeError if unparseable.
+    """
+    import json
+
+    try:
+        return json.loads(content)
+    except json.JSONDecodeError:
+        pass
+
+    # Fallback: find the outermost JSON object or array
+    for open_char, close_char in [("{", "}"), ("[", "]")]:
+        start = content.find(open_char)
+        end = content.rfind(close_char)
+        if start != -1 and end != -1 and end > start:
+            return json.loads(content[start : end + 1])
+
+    raise ValueError(f"No JSON found in LLM output: {content[:200]}")

backend/agenteval/web/app.py

@@ -70,6 +70,7 @@ _default_dist = Path(__file__).resolve().parent.parent.parent.parent / "frontend
 WEB_DIST = Path(settings.frontend_dist_path) if settings.frontend_dist_path else _default_dist
 if WEB_DIST.exists():
     from fastapi.staticfiles import StaticFiles
+
     app.mount("/assets", StaticFiles(directory=WEB_DIST / "assets"), name="assets")
 
 

backend/agenteval/web/deps.py

@@ -7,7 +7,6 @@ longer re-declare these helpers themselves.
 from typing import Optional
 
 from fastapi import Depends, Header, HTTPException, status
-from sqlmodel import Session
 
 from agenteval.config import get_settings
 from agenteval.storage.db import get_session

backend/agenteval/web/routers/files.py

@@ -10,7 +10,7 @@ from pydantic import BaseModel
 from sqlmodel import Session
 
 from agenteval.config import get_settings
-from agenteval.storage.db import FILES_DIR, FileRecordDB
+from agenteval.storage.db import FILES_DIR
 from agenteval.storage.file_repository import FileCategoryRepository, FileRecordRepository
 from agenteval.web.deps import get_db
 
@@ -200,4 +200,4 @@ def download_file(file_id: str, session: Session = Depends(get_db)):
 def delete_file(file_id: str, session: Session = Depends(get_db)) -> dict:
     if not FileRecordRepository(session).delete(file_id):
         raise HTTPException(status_code=404, detail="文件不存在")
-    return {"ok": True}
+    return {"ok": True}

backend/agenteval/web/routers/runs.py

@@ -11,6 +11,7 @@ from agenteval.evaluation.engine import EvalEngine
 from agenteval.models import EvalRun, RunStatus
 from agenteval.storage.db import get_session
 from agenteval.storage.repository import RunRepository, ScenarioRepository, TargetRepository
+from agenteval.utils.llm import extract_reply_text
 from agenteval.web.deps import get_db
 from agenteval.web.websocket import ws_manager
 
@@ -42,7 +43,9 @@ async def _run_evaluation(run_id: str, target_id: str, scenario_id: str) -> None
             return
 
         engine = EvalEngine(
-            target=target, scenario=scenario, session=session,
+            target=target,
+            scenario=scenario,
+            session=session,
             cancel_token=cancel_token,
         )
         await engine.run(
@@ -135,7 +138,7 @@ async def get_run_logs(run_id: str, session: Session = Depends(get_db)) -> dict:
             "round_index": t.round_index,
             "latency_ms": t.latency_ms,
             "sent_text": t.get_sent_message().get("msgBody", {}).get("content", ""),
-            "reply_text": _extract_reply_text(t.get_reply()),
+            "reply_text": extract_reply_text(t.get_reply()),
             "sent_at": t.sent_at.isoformat() if t.sent_at else None,
             "received_at": t.received_at.isoformat() if t.received_at else None,
         }
@@ -169,22 +172,7 @@ async def get_run_logs(run_id: str, session: Session = Depends(get_db)) -> dict:
                     "response_time_max_ms": case.expectations.response_time_max_ms,
                     "coherence_min_score": case.expectations.coherence_min_score,
                 },
-                "eval_rules": [
-                    {"type": r.type, "params": dict(r.params)} for r in case.eval_rules
-                ],
+                "eval_rules": [{"type": r.type, "params": dict(r.params)} for r in case.eval_rules],
             }
 
     return {"turns": turns_data, "results": results_data, "scenario_snapshot": scenario_snapshot}
-
-
-def _extract_reply_text(reply) -> str:
-    if reply is None:
-        return ""
-    if isinstance(reply, str):
-        return reply
-    if isinstance(reply, dict):
-        body = reply.get("msgBody") or reply.get("content", "")
-        if isinstance(body, dict):
-            return body.get("content", "")
-        return str(body)
-    return str(reply)

backend/cli/run.py

@@ -20,7 +20,9 @@ def _progress(event: str, data: dict[str, Any]) -> None:
         console.print(f"  第 {data['round']} 轮回复，耗时 {data['latency_ms']}ms")
     elif event == "rule_result":
         style = "green" if data["passed"] else "red"
-        console.print(f"    [{data['rule_type']}] {'通过' if data['passed'] else '失败'} - {data['reason']}", style=style)
+        console.print(
+            f"    [{data['rule_type']}] {'通过' if data['passed'] else '失败'} - {data['reason']}", style=style
+        )
     elif event == "error":
         console.print(f"执行异常: {data['error']}", style="red")
 

docs/plan-v0.3.md

@@ -0,0 +1,195 @@
+# AgentEvalTool v0.3 「拓」迭代规划
+
+**版本**: v0.3.0（规划）
+**制定日期**: 2026-07-16
+**基线**: v0.2.0-dev（已部署 t480，含文件管理 + 布局统一 + 6 bug 修复）
+**代号**: 「拓」(Extensibility milestone)
+**状态**: 待评审
+
+---
+
+## 一、版本定位
+
+v0.3 是「先稳后拓」战略的第二步。v0.2 已让核心闭环具备生产可用性，v0.3 的使命是：
+
+> 把「单通道 + 3 规则」的稳定闭环，扩展成「多通道 + 可组合规则」的评估平台，
+> 并打通 OpenClaw 双向集成，让评估能力可被 AI 助手自主编排。
+
+```
+v0.1 (07-09) MVP 闭环
+v1.1 (07-10) Ant Design 前端重构
+v0.2 (07-14) 「稳」async + 安全 + 测试 + 部署
+  └─ (07-16) 文件管理 + 布局统一 + bug 修复
+v0.3 (规划)  「拓」← 本文档
+  └─ 多通道 · 规则扩展 · OpenClaw 深度集成 · 报告升级
+```
+
+---
+
+## 二、现状评估
+
+### 2.1 已具备的能力
+
+| 层面 | 现状 |
+|---|---|
+| 引擎 | async EvalEngine，真取消，可配置超时，并发信号量 |
+| 通道 | 仅 `tutu-api`（工厂已预留 register 扩展点） |
+| 规则 | 3 种：keyword_match / response_time / llm_score |
+| 报告 | JSON + HTML |
+| 存储 | SQLite + Alembic + 级联删除 + 文件管理 |
+| 安全 | X-API-Key + .env + CORS 收紧 |
+| 测试 | 24 个后端测试，57% 覆盖率 |
+| 部署 | 一键脚本 + 镜像版本 tag + /api/health 校验 |
+
+### 2.2 技术债务盘点（v0.3 前置或并行处理）
+
+| 编号 | 债务 | 严重度 | 影响 |
+|---|---|---|---|
+| **DEBT-1** | **规则评估层同步阻塞**：`EvalRule.evaluate()` 是 `def` 而非 `async def`；`llm_score` 用同步 `requests.post(timeout=60)` | 🔴 架构级 | 阻塞事件循环，与 v0.2 async 化目标矛盾；**是 P0 规则扩展的前置阻塞项** |
+| DEBT-2 | 后端 `_extract_text`/`_extract_reply_text` 重复 5 处，`_extract_content_from_api_response` 重复 2 处 | 🟡 | 一处修复遗漏另一处即产生 bug |
+| DEBT-3 | 前端 `PageWrapper` 已成死代码，6 页面重复布局 JSX；`statusColor` 多页重复 | 🟡 | 改布局需改 6 处 |
+| DEBT-4 | 测试盲区：rules 覆盖率 15-30%，文件管理零测试，`_generate_messages` 无测试 | 🟡 | 规则/文件功能回归无保护 |
+| DEBT-5 | WebSocket 断线无自动重连；前端 bundle ~2.8MB | 🟢 | 体验问题 |
+
+> **关键判断**：DEBT-1 必须在 P0 规则扩展之前解决。semantic_similarity（调 embedding API）、safety（调 moderation API）本质都是网络 IO 规则，若继续在同步 `evaluate()` 上叠加，每条规则都会阻塞整个事件循环，v0.2 的异步化收益被规则层抵消。
+
+---
+
+## 三、v0.3 范围
+
+### 3.1 做什么（In Scope）
+
+#### P0 — 核心能力（必须交付）
+
+**P0-0 规则评估层异步化**（前置技术改造）
+- `EvalRule.evaluate()` → `async def evaluate()`
+- `engine._save_rule_results` 相应 `await rule.evaluate(...)`
+- `llm_score._call_llm`：`requests` → `httpx.AsyncClient`
+- 涉及文件：`evaluation/rules/base.py`、`keyword.py`、`response_time.py`、`llm_score.py`、`evaluation/engine.py`、`tests/unit/test_engine.py`
+- 验收：4 个规则全部 async；`pytest` 全绿；并发评测时 LLM 评分不阻塞 WebSocket 推送
+
+**P0-1 多通道插件化**
+- 新增 `HttpChannel`（通用 HTTP 请求/响应通道，配置化 request 模板 + response JSONPath 提取）
+- 新增 `OpenClawChannel`（直接对接 OpenClaw WS，复用现有 proxy 认证逻辑）
+- `ChannelFactory` 补齐 `HTTP` / `OPENCLAW` 映射
+- （可选）entry_points 插件发现 —— 见决策点 D3
+- 涉及文件：`channels/http.py`（新）、`channels/openclaw.py`（新）、`channels/factory.py`、`models.py`（ChannelType 已有枚举）
+- 验收：可创建 HTTP 通道类型的评测对象并跑通一次评测；连通性测试可用
+
+**P0-2 规则扩展 + 组合逻辑**
+- 新增规则：
+  - `semantic_similarity`（回复与参考答案的语义相似度，embedding + 余弦）
+  - `json_schema`（回复 JSON 结构校验）
+  - `safety`（敏感内容检测，可接 moderation API 或关键词黑名单降级）
+- 组合逻辑：Case 支持 `rule_logic: "all" | "any" | "weighted"`，weighted 支持每规则权重 + 阈值
+- 涉及文件：`rules/semantic.py`（新）、`rules/json_schema.py`（新）、`rules/safety.py`（新）、`rules/__init__.py`、`models.py`（Case 加 rule_logic 字段）、`engine.py`（组合判定逻辑）、Alembic 迁移
+- 验收：3 个新规则注册可用；weighted 组合的 pass_rate 计算正确；前端场景编辑器模板包含新规则示例
+
+#### P1 — 重要能力
+
+**P1-1 OpenClaw 深度集成（双向）**
+- OpenClaw skill 可触发评测（现有 CLI 调用链打通 + 结构化返回）
+- 评测完成 webhook 通知（`POST` 到配置的回调地址，附报告摘要）
+- 涉及文件：`plugins/openclaw/agenteval_skill.py`、`web/routers/runs.py`（webhook hook）、`config/settings.py`（webhook 配置）
+- 验收：OpenClaw 对话触发评测后能收到完成通知
+
+**P1-2 报告能力升级**
+- 对比报告：选两次 run，side-by-side 展示规则/用例差异
+- Markdown 导出（补充现有 JSON/HTML）
+- 报告模板外置（Jinja2 模板从代码抽到 `templates/` 目录）
+- 涉及文件：`web/routers/reports.py`、`evaluation/report.py`、`templates/`（新）、前端 `pages/Reports.tsx`
+- 验收：对比视图可用；Markdown 导出格式正确
+
+**P1-3 测试补全**（还 DEBT-4）
+- rules 单测：每个规则（含 3 个新规则）覆盖 pass/fail/边界
+- 文件管理测试：上传/下载/分类级联删除
+- 前端引入 vitest：`sessionReducer` 单测（v0.2 已抽纯函数，就等测试）
+- 验收：后端覆盖率 → 70%+；前端有首批 reducer 测试
+
+#### P2 — 体验优化
+
+**P2-1 场景能力增强**
+- 内置场景模板库（单轮问答 / 多轮对话 / 压力测试 / 动态生成）
+- 参数化场景（变量占位 + 批量实例化）
+
+**P2-2 前端体验 + 债务清理**（还 DEBT-3 / DEBT-5）
+- WebSocket 自动重连（指数退避）
+- `PageWrapper` 复用改造 或 删除；`statusColor` 抽到 tokens
+- bundle 优化：路由级懒加载
+
+### 3.2 不做什么（Out of Scope）
+
+明确排除，维持「个人/小团队」定位，避免过度设计：
+
+- ❌ 多租户 / RBAC 权限系统
+- ❌ PostgreSQL / 分布式数据库（SQLite 足够）
+- ❌ Celery / RQ 任务队列（async + asyncio.Task 足够）
+- ❌ 移动端适配
+- ❌ entry_points 全动态插件市场（见决策点 D3，倾向轻量注册）
+
+---
+
+## 四、迭代节奏
+
+按「先还债、再扩展、后体验」组织为 4 个 Sprint。工作量按人天（PD）估算，基于项目快速迭代节奏。
+
+| Sprint | 主题 | 工作项 | 估算 |
+|---|---|---|---|
+| **S1** | 还债 + 通道 | P0-0 规则异步化 + P0-1 多通道 + DEBT-2 后端工具函数合并 | 4-5 PD |
+| **S2** | 规则扩展 | P0-2 新规则 + 组合逻辑 + P1-3 规则测试同步 | 5-6 PD |
+| **S3** | 集成 + 报告 | P1-1 OpenClaw 双向 + P1-2 报告升级 | 4-5 PD |
+| **S4** | 体验 + 收尾 | P2-1 场景模板 + P2-2 前端债务清理 + 文件管理测试 | 3-4 PD |
+
+**合计约 16-20 PD（≈ 3-4 周）**，与原 release notes 的 4-5 周估算基本吻合（本计划把还债前置，S1 会略慢）。
+
+**里程碑验收点**：
+- M1（S1 末）：HTTP 通道跑通一次评测，规则层全 async，pytest 全绿
+- M2（S2 末）：5 种规则 + weighted 组合可用，后端覆盖率 65%+
+- M3（S3 末）：OpenClaw 对话触发评测 + 收到 webhook，对比报告可用
+- M4（S4 末）：场景模板库 + WebSocket 重连，v0.3 部署 t480
+
+---
+
+## 五、需要拍板的决策点
+
+以下决策会显著影响实现，建议在 S1 启动前确认：
+
+| 编号 | 决策 | 选项 | 倾向 |
+|---|---|---|---|
+| **D1** | semantic_similarity 的 embedding 来源 | A) 外部 API（豆包/OpenAI embedding，成本+网络依赖） B) 本地 sentence-transformers（部署重，+模型体积） | **A**：与现有 llm_score 一致的外部 API 模式，部署轻 |
+| **D2** | safety 规则实现深度 | A) 接 moderation API B) 关键词黑名单 C) 两者，API 不可用时降级黑名单 | **C**：默认黑名单保证可用，API 可选增强 |
+| **D3** | 通道/规则插件机制 | A) entry_points 动态发现 B) 现有工厂 register + 装饰器注册 | **B**：小团队无需插件市场，现有机制已够，避免过度设计 |
+| **D4** | 规则异步化是否 breaking | 改 `evaluate` 签名会影响所有规则 | 全量改造（4 规则少，一次到位），无需兼容层 |
+| **D5** | v0.3 是否顺带升级前端数据层（TanStack Query） | A) 引入 B) 维持 useState + 手动 load | **B**：当前页面数据量小，S4 仅做重连 + 懒加载，Query 留 v0.4 |
+
+---
+
+## 六、风险
+
+| 风险 | 影响 | 缓解 |
+|---|---|---|
+| 规则异步化牵连 engine + 全部规则 + 测试 | S1 工期 | 规则数量少（4 个），且有 24 个测试兜底，一次改到位 |
+| semantic/safety 依赖外部 API 稳定性 | 评测可靠性 | 规则内 try/except，API 失败降级为「规则跳过 + 明确 reason」，不中止 run |
+| OpenClaw webhook 依赖外部可达性 | P1-1 集成 | webhook 失败不影响评测结果，仅记录日志 |
+| 组合逻辑 weighted 的 pass_rate 语义变化 | 报告一致性 | 明确定义：weighted 下 case 通过 = 加权分 ≥ 阈值；文档 + 测试锁定 |
+
+---
+
+## 七、验收标准（v0.3 Definition of Done）
+
+1. 可创建 tutu-api / HTTP / OpenClaw 三类通道的评测对象并各跑通一次评测
+2. 5 种规则（3 旧 + semantic/json_schema/safety 选交付）可用，支持 all/any/weighted 组合
+3. 规则评估层全 async，并发评测时 LLM/embedding 调用不阻塞 WebSocket 推送
+4. OpenClaw 对话可触发评测并收到完成通知
+5. 报告支持 JSON/HTML/Markdown 导出 + 两次 run 对比
+6. 后端测试覆盖率 ≥ 70%，前端有首批 reducer 测试
+7. 后端工具函数去重（DEBT-2），前端布局债务清理（DEBT-3）
+8. v0.3 部署 t480，`/api/health` 显示 0.3.0 版本
+
+---
+
+## 八、下一步
+
+1. 评审本计划，确认第五节 5 个决策点
+2. 确认后按 Sprint 拆分为可执行 issue（每项含涉及文件 + 验收）
+3. 从 S1（规则异步化 + 多通道）启动

tests/unit/test_http_channel_and_rules.py

@@ -0,0 +1,213 @@
+"""Unit tests for HttpChannel and async rule evaluation."""
+
+import asyncio
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from agenteval.channels.http import HttpChannel, _get_path
+from agenteval.channels.base import SendResult
+from agenteval.evaluation.rules.keyword import KeywordMatchRule
+from agenteval.evaluation.rules.response_time import ResponseTimeRule
+from agenteval.models import Case, CaseType, Expectation, Turn
+
+
+# ── _get_path helper ─────────────────────────────────────────────────────
+
+def test_get_path_simple():
+    assert _get_path({"id": "abc"}, "id") == "abc"
+
+
+def test_get_path_nested():
+    assert _get_path({"a": {"b": {"c": 42}}}, "a.b.c") == 42
+
+
+def test_get_path_missing():
+    assert _get_path({"a": 1}, "a.b") is None
+
+
+def test_get_path_none_data():
+    assert _get_path(None, "x") is None
+
+
+def test_get_path_list_index():
+    assert _get_path({"items": ["x", "y"]}, "items.1") == "y"
+
+
+# ── HttpChannel ──────────────────────────────────────────────────────────
+
+def _make_channel(**extra) -> HttpChannel:
+    config = {
+        "send_url": "http://mock/send",
+        "reply_url": "http://mock/reply/{msg_id}",
+        "health_url": "http://mock/health",
+        **extra,
+    }
+    return HttpChannel(config)
+
+
+async def test_http_health_check_ok():
+    ch = _make_channel()
+    mock_resp = MagicMock()
+    mock_resp.status_code = 200
+    mock_resp.raise_for_status = MagicMock()
+    with patch.object(ch._client, "get", new=AsyncMock(return_value=mock_resp)):
+        result = await ch.health_check()
+    assert result.ok is True
+    assert "200" in result.message
+
+
+async def test_http_health_check_fail():
+    ch = _make_channel()
+    with patch.object(ch._client, "get", new=AsyncMock(side_effect=Exception("conn refused"))):
+        result = await ch.health_check()
+    assert result.ok is False
+    assert "conn refused" in result.message
+
+
+async def test_http_send_extracts_msg_id():
+    ch = _make_channel(msg_id_path="data.id")
+    mock_resp = MagicMock()
+    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")
+    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")
+    assert result.ok is False
+    assert "timeout" in result.error
+
+
+async def test_http_poll_reply_found():
+    ch = _make_channel(reply_path="answer")
+    call_count = {"n": 0}
+
+    async def mock_get(url, **kwargs):
+        call_count["n"] += 1
+        mock_resp = MagicMock()
+        mock_resp.raise_for_status = MagicMock()
+        mock_resp.json = MagicMock(return_value={"answer": "Hello world"})
+        return mock_resp
+
+    with patch.object(ch._client, "get", new=mock_get):
+        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"
+
+
+async def test_http_poll_reply_timeout():
+    ch = _make_channel(reply_path="missing_field", poll_interval=0.05)
+
+    async def mock_get(url, **kwargs):
+        mock_resp = MagicMock()
+        mock_resp.raise_for_status = MagicMock()
+        mock_resp.json = MagicMock(return_value={})  # field not present
+        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)
+    assert reply is None
+
+
+async def test_http_poll_reply_readiness_flag():
+    ch = _make_channel(reply_path="text", reply_ready_path="ready")
+    responses = [
+        {"ready": False, "text": "not ready"},
+        {"ready": True, "text": "final answer"},
+    ]
+    call_idx = {"n": 0}
+
+    async def mock_get(url, **kwargs):
+        mock_resp = MagicMock()
+        mock_resp.raise_for_status = MagicMock()
+        mock_resp.json = MagicMock(return_value=responses[min(call_idx["n"], 1)])
+        call_idx["n"] += 1
+        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)
+    assert reply is not None
+    assert reply.content == "final answer"
+
+
+# ── Async rule evaluation ────────────────────────────────────────────────
+
+def _make_turn(reply_text: str, latency_ms: int = 100) -> Turn:
+    return Turn(
+        id="t1", run_id="r1", case_id="c1", round_index=1,
+        reply={"msgBody": {"content": reply_text}},
+        latency_ms=latency_ms,
+    )
+
+
+def _make_case(*, keywords: list[str] | None = None, max_ms: int | None = None) -> Case:
+    return Case(
+        id="c1", type=CaseType.SINGLE, messages=["hi"],
+        expectations=Expectation(
+            keywords_include=keywords or [],
+            response_time_max_ms=max_ms,
+        ),
+    )
+
+
+async def test_keyword_rule_is_async_and_passes():
+    rule = KeywordMatchRule({"keywords": ["hello", "world"]})
+    turn = _make_turn("hello world here")
+    result = await rule.evaluate(_make_case(), [turn])
+    assert result.passed is True
+    assert result.score == 1.0
+
+
+async def test_keyword_rule_fails_missing_keyword():
+    rule = KeywordMatchRule({"keywords": ["missing"]})
+    turn = _make_turn("some other text")
+    result = await rule.evaluate(_make_case(), [turn])
+    assert result.passed is False
+    assert "missing" in result.reason
+
+
+async def test_keyword_rule_fails_excluded_keyword():
+    rule = KeywordMatchRule({"exclude_keywords": ["banned"]})
+    turn = _make_turn("this is banned content")
+    result = await rule.evaluate(_make_case(), [turn])
+    assert result.passed is False
+
+
+async def test_response_time_rule_is_async_and_passes():
+    rule = ResponseTimeRule({"max_ms": 500})
+    turn = _make_turn("ok", latency_ms=200)
+    result = await rule.evaluate(_make_case(), [turn])
+    assert result.passed is True
+
+
+async def test_response_time_rule_fails_over_threshold():
+    rule = ResponseTimeRule({"max_ms": 100})
+    turn = _make_turn("slow response", latency_ms=5000)
+    result = await rule.evaluate(_make_case(), [turn])
+    assert result.passed is False
+    assert "5000ms" in result.reason
+
+
+async def test_response_time_uses_expectation_fallback():
+    rule = ResponseTimeRule({})  # no max_ms in params
+    case = _make_case(max_ms=200)
+    turn = _make_turn("ok", latency_ms=100)
+    result = await rule.evaluate(case, [turn])
+    assert result.passed is True
+
+
+async def test_rules_empty_dialog_fail():
+    keyword_rule = KeywordMatchRule({"keywords": ["x"]})
+    rt_rule = ResponseTimeRule({"max_ms": 1000})
+    case = _make_case()
+    for rule in [keyword_rule, rt_rule]:
+        result = await rule.evaluate(case, [])
+        assert result.passed is False
+        assert "无回复" in result.reason