feat: P0 production hardening — LLM cache, semantic routing, state persistence

U1: LLM Cache Core (exact + semantic match, InMemory + Redis backends)
U2: Cache integration into LLMGateway with CacheConfig
U3: Semantic Router as Layer 1.5 in CostAwareRouter
U4: UsageStore persistence (Redis Hash + InMemory fallback)
U5: CascadeStateStore persistence (Redis INCR + InMemory TTL)
U6: EvolutionStore interface unification (Protocol + PostgreSQL backend)
U7: Configuration integration + E2E tests

Code review fixes:
- P0: date iteration bug (day>=28), semantic router index never built,
      Redis connection leak (per-call → persistent pool)
- P1: cache degradation recovery, semantic_search degradation,
      double miss counting, asyncio.Lock for PG init, LIMIT on queries,
      __import__ anti-pattern → _utcnow()
- P2: InMemory TTL cleanup, embedding preservation on put(),
      data TTL = max(exact_ttl, semantic_ttl)

.gitignore

@@ -35,6 +35,7 @@ build/pyinstaller-work/
 
 # Frontend build artifacts
 src/agentkit/server/static/
+**/node_modules/
 
 # Env
 .env

agentkit.yaml

@@ -5,23 +5,12 @@ server:
   rate_limit: 60
 llm:
   providers:
-    bailian-coding:
-      api_key: ${DASHSCOPE_API_KEY}
-      base_url: https://coding.dashscope.aliyuncs.com/v1
+    test:
       type: openai
-      models:
-        qwen3.7-plus:
-          alias: default
-        qwen3.6-plus: {}
-        qwen3.5-plus: {}
-        qwen3-max-2026-01-23: {}
-        qwen3-coder-plus:
-          alias: coder
-        qwen3-coder-next: {}
-        kimi-k2.5: {}
-        glm-5: {}
-        glm-4.7: {}
-        MiniMax-M2.5: {}
+      base_url: ''
+      max_tokens: 4096
+      timeout: 120.0
+      api_key: ''
   model_aliases:
     default: bailian-coding/qwen3.7-plus
     coder: bailian-coding/qwen3-coder-plus

src/agentkit/chat/semantic_router.py

@@ -0,0 +1,207 @@
+"""Semantic Router — Embedding-based intent routing as Layer 1.5.
+
+Uses pre-computed skill embeddings for zero-cost semantic matching,
+inserted between Layer 1 (HeuristicClassifier) and Layer 2 (LLM classification)
+in CostAwareRouter.
+
+Design doc: docs/plans/2026-06-14-004-u3-semantic-router.md
+"""
+
+import logging
+from dataclasses import dataclass
+from typing import Any
+
+from agentkit.memory.embedder import Embedder, EmbeddingCache
+from agentkit.utils.vector_math import compute_cosine_similarity
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SemanticRouteResult:
+    """Result of semantic routing."""
+
+    confidence: str  # "high" | "medium" | "low"
+    skill_name: str | None
+    similarity: float
+
+
+class SkillEmbeddingIndex:
+    """Pre-computed embedding index for registered skills.
+
+    Embeddings are computed at skill registration time and cached.
+    Query-time search is O(n) cosine similarity scan, which is fast
+    for <100 skills with 1024-1536 dim vectors.
+    """
+
+    def __init__(self, embedder: Embedder):
+        self._embedder = embedder
+        # skill_name → (embedding, source_text)
+        self._index: dict[str, tuple[list[float], str]] = {}
+
+    async def build(self, skill_registry: Any) -> None:
+        """Build index from all registered skills."""
+        if skill_registry is None:
+            return
+        skills = skill_registry.list_skills()
+        for skill in skills:
+            await self.update_skill(skill.config.name, skill)
+
+    async def update_skill(self, skill_name: str, skill: Any) -> None:
+        """Re-embed a single skill (on registration/update)."""
+        source_text = self._build_source_text(skill)
+        try:
+            embedding = await self._embedder.embed(source_text)
+            self._index[skill_name] = (embedding, source_text)
+        except Exception as e:
+            logger.warning(f"Failed to embed skill '{skill_name}': {e}")
+
+    def remove_skill(self, skill_name: str) -> None:
+        """Remove a skill from the index."""
+        self._index.pop(skill_name, None)
+
+    async def search(self, query_embedding: list[float], top_k: int = 5) -> list[tuple[str, float]]:
+        """Search for skills matching the query embedding.
+
+        Returns:
+            List of (skill_name, similarity) sorted by similarity descending.
+        """
+        if not self._index:
+            return []
+
+        results: list[tuple[str, float]] = []
+        for skill_name, (emb, _) in self._index.items():
+            sim = compute_cosine_similarity(query_embedding, emb)
+            results.append((skill_name, sim))
+
+        results.sort(key=lambda x: x[1], reverse=True)
+        return results[:top_k]
+
+    @staticmethod
+    def _build_source_text(skill: Any) -> str:
+        """Build embedding source text from skill metadata.
+
+        Combines description, intent keywords, and capability tags
+        for rich semantic representation.
+        """
+        config = skill.config if hasattr(skill, "config") else skill
+        parts = []
+
+        # Description
+        description = getattr(config, "description", "") or ""
+        if description:
+            parts.append(description)
+
+        # Intent keywords
+        intent = getattr(config, "intent", None)
+        if intent and hasattr(intent, "keywords") and intent.keywords:
+            parts.append(" ".join(intent.keywords))
+
+        # Capability tags
+        capabilities = getattr(config, "capabilities", None)
+        if capabilities:
+            tags = []
+            for cap in capabilities:
+                if isinstance(cap, str):
+                    tags.append(cap)
+                elif isinstance(cap, dict):
+                    tags.append(cap.get("tag", ""))
+                elif hasattr(cap, "tag"):
+                    tags.append(cap.tag)
+            if tags:
+                parts.append(" ".join(t for t in tags if t))
+
+        # Fallback: use skill name if no other text available
+        if not parts:
+            parts.append(getattr(config, "name", "unknown"))
+
+        return " | ".join(parts)
+
+    @property
+    def size(self) -> int:
+        """Number of skills in the index."""
+        return len(self._index)
+
+
+class SemanticRouter:
+    """Embedding-based semantic routing as Layer 1.5.
+
+    Three confidence zones:
+    - similarity > similarity_high (0.85): HIGH → direct skill match, skip Layer 2
+    - similarity_low (0.6) <= similarity <= similarity_high: MEDIUM → skill hint for Layer 2
+    - similarity < similarity_low (0.6): LOW → no semantic signal, normal routing
+    """
+
+    def __init__(
+        self,
+        embedder: Embedder,
+        similarity_high: float = 0.85,
+        similarity_low: float = 0.6,
+    ):
+        self._embedder = embedder
+        self._similarity_high = similarity_high
+        self._similarity_low = similarity_low
+        self._index = SkillEmbeddingIndex(embedder)
+        self._query_cache = EmbeddingCache(max_size=500, ttl=1800)
+
+    async def build_index(self, skill_registry: Any) -> None:
+        """Build skill embedding index from registry."""
+        await self._index.build(skill_registry)
+        logger.info(f"Semantic router index built: {self._index.size} skills")
+
+    async def update_skill(self, skill_name: str, skill: Any) -> None:
+        """Update a single skill's embedding."""
+        await self._index.update_skill(skill_name, skill)
+
+    def remove_skill(self, skill_name: str) -> None:
+        """Remove a skill from the index."""
+        self._index.remove_skill(skill_name)
+
+    async def route(self, query: str) -> SemanticRouteResult:
+        """Route a query using semantic similarity.
+
+        Args:
+            query: User's input text.
+
+        Returns:
+            SemanticRouteResult with confidence, skill_name, and similarity.
+        """
+        if self._index.size == 0:
+            return SemanticRouteResult(confidence="low", skill_name=None, similarity=0.0)
+
+        try:
+            # Get query embedding (with cache)
+            query_embedding = self._query_cache.get(query)
+            if query_embedding is None:
+                query_embedding = await self._embedder.embed(query)
+                self._query_cache.put(query, query_embedding)
+
+            # Search skill index
+            results = await self._index.search(query_embedding, top_k=1)
+            if not results:
+                return SemanticRouteResult(confidence="low", skill_name=None, similarity=0.0)
+
+            best_skill, best_sim = results[0]
+
+            if best_sim >= self._similarity_high:
+                return SemanticRouteResult(
+                    confidence="high",
+                    skill_name=best_skill,
+                    similarity=best_sim,
+                )
+            elif best_sim >= self._similarity_low:
+                return SemanticRouteResult(
+                    confidence="medium",
+                    skill_name=best_skill,
+                    similarity=best_sim,
+                )
+            else:
+                return SemanticRouteResult(
+                    confidence="low",
+                    skill_name=None,
+                    similarity=best_sim,
+                )
+
+        except Exception as e:
+            logger.warning(f"Semantic routing failed, returning low confidence: {e}")
+            return SemanticRouteResult(confidence="low", skill_name=None, similarity=0.0)

src/agentkit/chat/skill_routing.py

@@ -6,6 +6,7 @@ and prompt assembly into a single module used by both chat routes.
 
 from __future__ import annotations
 
+import enum
 import json
 import logging
 import re
@@ -21,6 +22,19 @@ logger = logging.getLogger(__name__)
 _SKILL_NAME_RE = re.compile(r"^[a-z0-9][a-z0-9_-]{0,63}$")
 
 
+class ExecutionMode(enum.Enum):
+    """How the downstream should execute this routing result.
+
+    This is the single source of truth for execution path selection.
+    The transport layer (portal.py, chat.py) should branch on this
+    field instead of string-matching match_method.
+    """
+
+    DIRECT_CHAT = "direct_chat"    # Zero-cost: direct LLM call, no ReAct loop
+    REACT = "react"                # Default agent ReAct loop with default tools
+    SKILL_REACT = "skill_react"    # Skill-matched ReAct with skill tools + prompt
+
+
 def validate_skill_name(name: str) -> str:
     """Validate and normalize a skill name. Raises ValueError on invalid input."""
     normalized = name.strip().lower()
@@ -49,6 +63,7 @@ class SkillRoutingResult:
     transparency_level: str = "SILENT"
     execution_trace: list[dict] = field(default_factory=list)
     complexity: float = 0.0
+    execution_mode: ExecutionMode = ExecutionMode.DIRECT_CHAT
 
 
 def parse_skill_prefix(content: str) -> tuple[str | None, str]:
@@ -88,6 +103,7 @@ async def resolve_skill_routing(
     default_agent_name: str = "default",
     agent_tool_registry: Any = None,
     session_id: str = "",
+    force_skill: str | None = None,
 ) -> SkillRoutingResult:
     """Resolve skill routing for a user message.
 
@@ -120,6 +136,20 @@ async def resolve_skill_routing(
             result.skill_name = None
             result.skill_config = None
 
+    # Try force_skill match (from semantic router high confidence)
+    if not result.matched and force_skill and skill_registry:
+        try:
+            matched_skill = skill_registry.get(force_skill)
+            result.skill_name = force_skill
+            result.skill_config = matched_skill.config
+            result.skill_tools = matched_skill.tools or []
+            result.matched = True
+            result.match_method = "semantic_force"
+            result.match_confidence = 1.0
+            logger.info(f"Session {session_id}: using force-matched skill '{force_skill}'")
+        except Exception as e:
+            logger.warning(f"Session {session_id}: force skill '{force_skill}' not found: {e}")
+
     # Try IntentRouter if no explicit match
     if not result.matched and skill_registry and intent_router:
         skills = skill_registry.list_skills()
@@ -205,11 +235,14 @@ async def resolve_skill_routing(
 
         result.model = result.skill_config.llm.get("model", default_model) if result.skill_config.llm else default_model
         result.agent_name = result.skill_name
+        result.execution_mode = ExecutionMode.SKILL_REACT
     else:
         result.system_prompt = default_system_prompt
         result.tools = default_tools
         result.model = default_model
         result.agent_name = default_agent_name
+        # No skill matched — if we have tools, use ReAct; otherwise direct chat
+        result.execution_mode = ExecutionMode.REACT if default_tools else ExecutionMode.DIRECT_CHAT
 
     # Append available tools to system prompt so LLM knows what it can call
     if result.tools:
@@ -257,6 +290,14 @@ _CHAT_MODE_RE = re.compile(
     re.IGNORECASE,
 )
 
+# Simple identity/meta questions — zero-cost direct chat, no skill routing needed
+_IDENTITY_RE = re.compile(
+    r"^(你是谁|你叫什么|你是什么|你是哪个|who are you|what are you|what's your name"
+    r"|介绍一下你自己|自我介绍|你叫啥|你叫什么名字|你的名字)"
+    r"\s*[?？!！.。]*$",
+    re.IGNORECASE,
+)
+
 _SENTENCE_SPLIT_RE = re.compile(r'[，。！？；\n,.!?;]')
 
 
@@ -319,8 +360,9 @@ class HeuristicClassifier:
     }
 
     # 中等复杂度暗示词（简单问题但需思考）
+    # 注意：不包含"怎么"，因为"怎么样"是闲聊而非工具需求
     _MEDIUM_COMPLEXITY_HINTS_CN = {
-        "如何", "怎么", "怎样", "为什么", "什么原因", "区别",
+        "如何", "怎样", "为什么", "什么原因", "区别",
         "推荐", "建议", "选择", "哪个",
     }
 
@@ -428,6 +470,7 @@ class CostAwareRouter:
         auction_enabled: bool = False,
         classifier: str = "heuristic",
         merged_llm_classify: bool = True,
+        semantic_router: Any = None,  # SemanticRouter | None
     ):
         self._llm_gateway = llm_gateway
         self._model = model
@@ -435,6 +478,7 @@ class CostAwareRouter:
         self._auction_enabled = auction_enabled
         self._classifier = classifier
         self._merged_llm_classify = merged_llm_classify
+        self._semantic_router = semantic_router
         self._auction_house = AuctionHouse() if auction_enabled else None
         if classifier not in ("heuristic", "llm"):
             raise ValueError(f"Invalid classifier: {classifier!r}, must be 'heuristic' or 'llm'")
@@ -462,6 +506,10 @@ class CostAwareRouter:
         if _CHAT_MODE_RE.match(stripped):
             return "chat_mode", stripped
 
+        # 身份/元问题模式（"你是谁"等）— 零成本直接对话
+        if _IDENTITY_RE.match(stripped):
+            return "identity", stripped
+
         return None, stripped
 
     # -- Layer 1: LLM quick classify (~100 tokens) -------------------------
@@ -577,6 +625,7 @@ class CostAwareRouter:
                         match_method="merged_llm",
                         match_confidence=0.7,
                         complexity=merged_complexity,
+                        execution_mode=ExecutionMode.SKILL_REACT,
                     )
                     # Merge tools
                     agent_tools = agent_tool_registry.list_tools() if agent_tool_registry else default_tools
@@ -590,6 +639,18 @@ class CostAwareRouter:
                     result.model = result.skill_config.llm.get("model", default_model) if result.skill_config.llm else default_model
                     result.agent_name = skill_hint
                     result.system_prompt = build_skill_system_prompt(result.skill_config) or default_system_prompt
+                    # Append available tools to system prompt so LLM knows what it can call
+                    if result.tools:
+                        tools_desc = _build_tools_description(result.tools)
+                        tool_instruction = (
+                            "\n\n## Tool Usage\n"
+                            "You have access to the following tools. When you need to use a tool, "
+                            "respond with a tool call in the format specified by the system.\n"
+                            "Never make up information or guess answers when you can use a tool to find the answer.\n"
+                            "Always prefer using tools over guessing.\n"
+                        )
+                        if result.system_prompt:
+                            result.system_prompt += f"{tool_instruction}\n## Available Tools\n{tools_desc}"
                     logger.info(
                         f"Session {session_id}: merged LLM classify routed to skill '{skill_hint}' "
                         f"(complexity={merged_complexity:.2f})"
@@ -610,6 +671,7 @@ class CostAwareRouter:
                     match_method="merged_llm_low",
                     match_confidence=1.0 - merged_complexity,
                     complexity=merged_complexity,
+                    execution_mode=ExecutionMode.DIRECT_CHAT,
                 )
             elif merged_complexity > 0.7:
                 # High complexity — delegate to Layer 2
@@ -623,6 +685,7 @@ class CostAwareRouter:
                     match_method="merged_llm_high",
                     match_confidence=merged_complexity,
                     complexity=merged_complexity,
+                    execution_mode=ExecutionMode.REACT,
                 )
             else:
                 # Medium complexity, no skill match — default agent
@@ -636,6 +699,7 @@ class CostAwareRouter:
                     match_method="merged_llm_medium",
                     match_confidence=0.5,
                     complexity=merged_complexity,
+                    execution_mode=ExecutionMode.REACT,
                 )
         except (json.JSONDecodeError, TypeError, ValueError) as e:
             logger.warning(f"CostAwareRouter _classify_merged parse failed: {e}, falling back to default")
@@ -649,6 +713,7 @@ class CostAwareRouter:
                 match_method="merged_llm_fallback",
                 match_confidence=0.5,
                 complexity=0.5,
+                execution_mode=ExecutionMode.REACT,
             )
         except Exception as e:
             logger.warning(f"CostAwareRouter _classify_merged failed: {e}, falling back to default")
@@ -662,6 +727,7 @@ class CostAwareRouter:
                 match_method="merged_llm_fallback",
                 match_confidence=0.5,
                 complexity=0.5,
+                execution_mode=ExecutionMode.REACT,
             )
 
     # -- Layer 2: Capability matching / Auction (optional) -----------------
@@ -746,6 +812,7 @@ class CostAwareRouter:
                             system_prompt=default_system_prompt,
                             tools=default_tools,
                             complexity=complexity,
+                            execution_mode=ExecutionMode.REACT,
                         )
                         if trace is not None:
                             trace.append({
@@ -776,6 +843,7 @@ class CostAwareRouter:
                         system_prompt=default_system_prompt,
                         tools=default_tools,
                         complexity=complexity,
+                        execution_mode=ExecutionMode.REACT,
                     )
                     if trace is not None:
                         trace.append({
@@ -876,7 +944,7 @@ class CostAwareRouter:
                 span.set_attribute("route.target", result.skill_name or "default")
                 return result
 
-            if match_type in ("greeting", "chat_mode"):
+            if match_type in ("greeting", "chat_mode", "identity"):
                 result = SkillRoutingResult(
                     clean_content=clean_content,
                     system_prompt=default_system_prompt,
@@ -887,6 +955,7 @@ class CostAwareRouter:
                     match_method=match_type,
                     match_confidence=1.0,
                     complexity=0.0,
+                    execution_mode=ExecutionMode.DIRECT_CHAT,
                 )
                 trace.append({
                     "layer": 0,
@@ -916,7 +985,7 @@ class CostAwareRouter:
                     "complexity": complexity,
                 })
 
-            # Low complexity → default agent
+            # Low complexity → direct chat
             if complexity < 0.3:
                 result = SkillRoutingResult(
                     clean_content=clean_content,
@@ -928,6 +997,7 @@ class CostAwareRouter:
                     match_method="low_complexity",
                     match_confidence=1.0 - complexity,
                     complexity=complexity,
+                    execution_mode=ExecutionMode.DIRECT_CHAT,
                 )
                 trace.append({
                     "layer": 1,
@@ -941,6 +1011,59 @@ class CostAwareRouter:
                 span.set_attribute("route.target", "default")
                 return result
 
+            # ---- Layer 1.5: Semantic Router (zero LLM cost) ----
+            skill_hint = None
+            if self._semantic_router is not None and complexity >= 0.3:
+                try:
+                    semantic_result = await self._semantic_router.route(clean_content)
+                    if semantic_result.confidence == "high" and semantic_result.skill_name:
+                        # Direct skill match — skip Layer 2
+                        trace.append({
+                            "layer": 1.5,
+                            "method": "semantic_high",
+                            "skill": semantic_result.skill_name,
+                            "similarity": round(semantic_result.similarity, 3),
+                            "cost": "zero",
+                        })
+                        result = await resolve_skill_routing(
+                            content=content,
+                            skill_registry=skill_registry,
+                            intent_router=intent_router,
+                            default_tools=default_tools,
+                            default_system_prompt=default_system_prompt,
+                            default_model=default_model,
+                            default_agent_name=default_agent_name,
+                            agent_tool_registry=agent_tool_registry,
+                            session_id=session_id,
+                            force_skill=semantic_result.skill_name,
+                        )
+                        result.match_method = "semantic_high"
+                        result.match_confidence = semantic_result.similarity
+                        result.complexity = complexity
+                        if result.matched:
+                            result.execution_mode = ExecutionMode.SKILL_REACT
+                        result.execution_trace = trace if transparency != "SILENT" else []
+                        result.transparency_level = transparency
+                        span.set_attribute("route.layer", "semantic_high")
+                        span.set_attribute("route.target", result.skill_name or "default")
+                        return result
+                    elif semantic_result.confidence == "medium" and semantic_result.skill_name:
+                        # Pass skill hint to Layer 1.5 merged classify or Layer 2
+                        skill_hint = semantic_result.skill_name
+                        trace.append({
+                            "layer": 1.5,
+                            "method": "semantic_medium",
+                            "skill_hint": skill_hint,
+                            "similarity": round(semantic_result.similarity, 3),
+                        })
+                except Exception as e:
+                    logger.warning(f"Semantic routing failed, falling through: {e}")
+                    trace.append({
+                        "layer": 1.5,
+                        "method": "semantic_error",
+                        "error": str(e),
+                    })
+
             # Medium complexity → merged LLM classify or IntentRouter
             if complexity <= 0.7:
                 if self._merged_llm_classify and self._llm_gateway is not None:
@@ -994,7 +1117,7 @@ class CostAwareRouter:
                         agent_tool_registry=agent_tool_registry,
                         session_id=session_id,
                     )
-                result.complexity = result.complexity or complexity
+                result.complexity = result.complexity if result.complexity > 0 else complexity
                 trace.append({
                     "layer": 1,
                     "method": result.match_method or "merged_llm",

src/agentkit/core/config_driven.py

@@ -685,7 +685,7 @@ class ConfigDrivenAgent(BaseAgent, EvolutionMixin):
         retrieval_config = self._config.memory.get("retrieval", {}) if self._config.memory else {}
         result = await self._react_engine.execute(
             messages=user_messages,
-            tools=self._tools if self._tools else None,
+            tools=self.get_tools() or None,
             model=self._config.llm.get("model", "default") if self._config.llm else "default",
             agent_name=self.name,
             task_type=task.task_type,
@@ -735,7 +735,7 @@ class ConfigDrivenAgent(BaseAgent, EvolutionMixin):
 
         result = await rewoo_engine.execute(
             messages=user_messages,
-            tools=self._tools if self._tools else None,
+            tools=self.get_tools() or None,
             model=self._config.llm.get("model", "default") if self._config.llm else "default",
             agent_name=self.name,
             task_type=task.task_type,
@@ -781,7 +781,7 @@ class ConfigDrivenAgent(BaseAgent, EvolutionMixin):
 
         result = await plan_exec_engine.execute(
             messages=user_messages,
-            tools=self._tools if self._tools else None,
+            tools=self.get_tools() or None,
             model=self._config.llm.get("model", "default") if self._config.llm else "default",
             agent_name=self.name,
             task_type=task.task_type,
@@ -829,7 +829,7 @@ class ConfigDrivenAgent(BaseAgent, EvolutionMixin):
 
         result = await reflexion_engine.execute(
             messages=user_messages,
-            tools=self._tools if self._tools else None,
+            tools=self.get_tools() or None,
             model=self._config.llm.get("model", "default") if self._config.llm else "default",
             agent_name=self.name,
             task_type=task.task_type,

src/agentkit/core/react.py

@@ -441,7 +441,14 @@ class ReActEngine:
                                         except Exception as e:
                                             tool_result = {"error": f"Tool '{tc.name}' execution failed: {e}"}
                                     else:
-                                        tool_result = await self._execute_tool(tc.name, tc.arguments, tools)
+                                        # Non-dangerous tool: confirmation was for the overall action,
+                                        # re-execute with skip flag to avoid re-triggering confirmation
+                                        clean_args = {k: v for k, v in tc.arguments.items() if not k.startswith("_")}
+                                        clean_args["_skip_dangerous_check"] = True
+                                        try:
+                                            tool_result = await tool.safe_execute(**clean_args) if tool else {"error": f"Tool '{tc.name}' not found"}
+                                        except Exception as e:
+                                            tool_result = {"error": f"Tool '{tc.name}' execution failed: {e}"}
                                 else:
                                     tool_result = {
                                         "output": "",
@@ -905,7 +912,13 @@ class ReActEngine:
                                         finally:
                                             pass  # No shared state mutation needed
                                     else:
-                                        tool_result = await self._execute_tool(tc.name, tc.arguments, tools)
+                                        # Non-dangerous tool: re-execute with skip flag
+                                        clean_args = {k: v for k, v in tc.arguments.items() if not k.startswith("_")}
+                                        clean_args["_skip_dangerous_check"] = True
+                                        try:
+                                            tool_result = await tool.safe_execute(**clean_args) if tool else {"error": f"Tool '{tc.name}' not found"}
+                                        except Exception as e:
+                                            tool_result = {"error": f"Tool '{tc.name}' execution failed: {e}"}
 
                                     yield ReActEvent(
                                         event_type="confirmation_result",
@@ -1261,7 +1274,13 @@ class ReActEngine:
                     except Exception as e:
                         tool_result = {"error": f"Tool '{tc.name}' execution failed: {e}"}
                 else:
-                    tool_result = await self._execute_tool(tc.name, tc.arguments, tools)
+                    # Non-dangerous tool: re-execute with skip flag
+                    clean_args = {k: v for k, v in tc.arguments.items() if not k.startswith("_")}
+                    clean_args["_skip_dangerous_check"] = True
+                    try:
+                        tool_result = await tool.safe_execute(**clean_args) if tool else {"error": f"Tool '{tc.name}' not found"}
+                    except Exception as e:
+                        tool_result = {"error": f"Tool '{tc.name}' execution failed: {e}"}
 
                 events.append(ReActEvent(
                     event_type="confirmation_result",

src/agentkit/evolution/__init__.py

@@ -13,6 +13,7 @@ from agentkit.evolution.strategy_tuner import StrategyTuner
 from agentkit.evolution.ab_tester import ABTester
 from agentkit.evolution.evolution_store import (
     EvolutionStore,
+    EvolutionStoreProtocol,
     InMemoryEvolutionStore,
     PersistentEvolutionStore,
     create_evolution_store,
@@ -30,6 +31,7 @@ __all__ = [
     "StrategyTuner",
     "ABTester",
     "EvolutionStore",
+    "EvolutionStoreProtocol",
     "PersistentEvolutionStore",
     "InMemoryEvolutionStore",
     "create_evolution_store",

src/agentkit/evolution/evolution_store.py

@@ -1,9 +1,11 @@
 """EvolutionStore - 进化日志存储
 
-提供三种后端实现：
-- EvolutionStore: 基于外部注入的异步 SQLAlchemy session（原有实现）
-- PersistentEvolutionStore: 基于 SQLite 的持久化存储
-- InMemoryEvolutionStore: 基于内存字典的轻量存储（用于测试）
+提供统一 Protocol 和四种后端实现：
+- EvolutionStoreProtocol: 统一接口 Protocol（所有后端必须实现）
+- EvolutionStore: 基于外部注入的异步 SQLAlchemy session（原有实现，仅事件操作）
+- PersistentEvolutionStore: 基于 SQLite 的持久化存储（完整 Protocol）
+- InMemoryEvolutionStore: 基于内存字典的轻量存储（完整 Protocol，用于测试）
+- PostgreSQLEvolutionStore: 基于 PostgreSQL 的异步持久化存储（完整 Protocol，见 pg_store.py）
 """
 
 import asyncio
@@ -13,7 +15,7 @@ import os
 import time
 import uuid as _uuid
 from datetime import datetime, timezone
-from typing import Any
+from typing import Any, Protocol, runtime_checkable
 
 from sqlalchemy import create_engine, event as sa_event, select
 from sqlalchemy.exc import OperationalError
@@ -30,6 +32,34 @@ from agentkit.evolution.models import (
 logger = logging.getLogger(__name__)
 
 
+# ── 统一 Protocol ─────────────────────────────────────────
+
+
+@runtime_checkable
+class EvolutionStoreProtocol(Protocol):
+    """进化存储统一接口 Protocol
+
+    所有后端必须实现以下方法。不支持的操作应抛出 NotImplementedError。
+    """
+
+    async def record(self, event: EvolutionEvent) -> str: ...
+    async def rollback(self, event_id: str) -> bool: ...
+    async def list_events(
+        self,
+        agent_name: str | None = ...,
+        change_type: str | None = ...,
+        status: str | None = ...,
+    ) -> list[dict]: ...
+    async def record_skill_version(
+        self, skill_name: str, version: str, content: str, parent_version: str | None = ...
+    ) -> str: ...
+    async def list_skill_versions(self, skill_name: str) -> list[dict]: ...
+    async def record_ab_test_result(
+        self, test_id: str, variant: str, score: float, sample_count: int = ...
+    ) -> str: ...
+    async def get_ab_test_results(self, test_id: str) -> list[dict]: ...
+
+
 class EvolutionStore:
     """进化日志存储
 
@@ -133,6 +163,40 @@ class EvolutionStore:
                 logger.error(f"Failed to list evolution events: {e}")
                 return []
 
+    # ── Protocol 兼容方法（旧版 EvolutionStore 不支持 skill_version / ab_test）──
+
+    async def record_skill_version(
+        self, skill_name: str, version: str, content: str, parent_version: str | None = None
+    ) -> str:
+        """记录技能版本（旧版 SQL 后端不支持，抛出 NotImplementedError）"""
+        raise NotImplementedError(
+            "EvolutionStore (SQL backend) does not support skill_version operations. "
+            "Use PersistentEvolutionStore or PostgreSQLEvolutionStore instead."
+        )
+
+    async def list_skill_versions(self, skill_name: str) -> list[dict]:
+        """列出技能版本历史（旧版 SQL 后端不支持，抛出 NotImplementedError）"""
+        raise NotImplementedError(
+            "EvolutionStore (SQL backend) does not support skill_version operations. "
+            "Use PersistentEvolutionStore or PostgreSQLEvolutionStore instead."
+        )
+
+    async def record_ab_test_result(
+        self, test_id: str, variant: str, score: float, sample_count: int = 0
+    ) -> str:
+        """记录 A/B 测试结果（旧版 SQL 后端不支持，抛出 NotImplementedError）"""
+        raise NotImplementedError(
+            "EvolutionStore (SQL backend) does not support A/B test operations. "
+            "Use PersistentEvolutionStore or PostgreSQLEvolutionStore instead."
+        )
+
+    async def get_ab_test_results(self, test_id: str) -> list[dict]:
+        """获取 A/B 测试结果（旧版 SQL 后端不支持，抛出 NotImplementedError）"""
+        raise NotImplementedError(
+            "EvolutionStore (SQL backend) does not support A/B test operations. "
+            "Use PersistentEvolutionStore or PostgreSQLEvolutionStore instead."
+        )
+
 
 class PersistentEvolutionStore:
     """SQLite 持久化进化存储
@@ -464,19 +528,32 @@ def create_evolution_store(
     db_path: str = "~/.agentkit/evolution.db",
     session_factory: Any = None,
     evolution_model: Any = None,
+    database_url: str | None = None,
 ) -> EvolutionStore | PersistentEvolutionStore | InMemoryEvolutionStore:
     """工厂函数：创建进化存储实例
 
     Args:
-        backend: 存储后端类型 - "memory" | "sqlite" | "sql"
+        backend: 存储后端类型 - "memory" | "sqlite" | "sql" | "postgresql"
         db_path: SQLite 数据库路径（仅 backend="sqlite" 时使用）
         session_factory: 异步 SQLAlchemy session 工厂（仅 backend="sql" 时使用）
         evolution_model: SQLAlchemy ORM 模型类（仅 backend="sql" 时使用）
+        database_url: PostgreSQL 连接字符串（仅 backend="postgresql" 时使用）
 
     Returns:
         对应后端的进化存储实例
     """
-    if backend == "sqlite":
+    if backend == "postgresql":
+        from agentkit.evolution.pg_store import PostgreSQLEvolutionStore
+
+        url = database_url or os.environ.get("AGENTKIT_DATABASE_URL")
+        if url:
+            return PostgreSQLEvolutionStore(database_url=url)
+        logger.warning(
+            "PostgreSQL backend requested but no database_url provided, "
+            "falling back to InMemoryEvolutionStore"
+        )
+        return InMemoryEvolutionStore()
+    elif backend == "sqlite":
         return PersistentEvolutionStore(db_path=db_path)
     elif backend == "sql" and session_factory and evolution_model:
         return EvolutionStore(session_factory=session_factory, evolution_model=evolution_model)

src/agentkit/evolution/lifecycle.py

@@ -483,7 +483,15 @@ class EvolutionMixin:
 
                 tool = MemoryTool(memory_store)
                 section = category
-                content = "; ".join(reflections[-1]["reflection"].suggestions[:2])
+                # 汇总所有累积反思的建议（去重，最多取 5 条）
+                all_suggestions: list[str] = []
+                seen: set[str] = set()
+                for r in reflections:
+                    for suggestion in r["reflection"].suggestions:
+                        if suggestion not in seen:
+                            seen.add(suggestion)
+                            all_suggestions.append(suggestion)
+                content = "; ".join(all_suggestions[:5])
                 reason = f"连续{len(reflections)}次低质量反思 (category: {category})"
 
                 update_result = await tool.execute(

src/agentkit/evolution/pg_store.py

@@ -0,0 +1,329 @@
+"""PostgreSQLEvolutionStore - 基于 PostgreSQL 的异步进化存储
+
+使用 async SQLAlchemy + asyncpg 实现完整的 EvolutionStoreProtocol，
+支持进化事件、技能版本、A/B 测试结果的持久化。
+"""
+
+import asyncio
+import logging
+import uuid as _uuid
+from datetime import datetime, timezone
+from typing import Any
+
+from sqlalchemy import select
+from sqlalchemy.dialects.postgresql import JSONB
+from sqlalchemy import Column, DateTime, Float, Integer, String, Text, UniqueConstraint
+from sqlalchemy.orm import declarative_base
+
+from agentkit.core.protocol import EvolutionEvent
+
+logger = logging.getLogger(__name__)
+
+# PG 专用 Base（与 SQLite models 的 Base 隔离，避免表名冲突）
+PGBase = declarative_base()
+
+
+def _utcnow() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+class PGEvolutionEventModel(PGBase):
+    """PostgreSQL 进化事件 ORM 模型"""
+
+    __tablename__ = "evolution_events"
+
+    id = Column(String, primary_key=True, default=lambda: str(_uuid.uuid4()))
+    agent_name = Column(String, index=True)
+    change_type = Column(String, nullable=True)
+    before = Column(JSONB, nullable=True)
+    after = Column(JSONB, nullable=True)
+    metrics = Column(JSONB, nullable=True)
+    status = Column(String, default="active")
+    created_at = Column(DateTime, default=_utcnow)
+
+
+class PGSkillVersionModel(PGBase):
+    """PostgreSQL 技能版本 ORM 模型"""
+
+    __tablename__ = "skill_versions"
+    __table_args__ = (UniqueConstraint("skill_name", "version"),)
+
+    id = Column(String, primary_key=True, default=lambda: str(_uuid.uuid4()))
+    skill_name = Column(String, index=True)
+    version = Column(String)
+    content = Column(Text)
+    parent_version = Column(String, nullable=True)
+    created_at = Column(DateTime, default=_utcnow)
+
+
+class PGABTestResultModel(PGBase):
+    """PostgreSQL A/B 测试结果 ORM 模型"""
+
+    __tablename__ = "ab_test_results"
+
+    id = Column(String, primary_key=True, default=lambda: str(_uuid.uuid4()))
+    test_id = Column(String, index=True)
+    variant = Column(String)
+    score = Column(Float)
+    sample_count = Column(Integer, default=0)
+    created_at = Column(DateTime, default=_utcnow)
+
+
+class PostgreSQLEvolutionStore:
+    """PostgreSQL 异步进化存储
+
+    使用 async SQLAlchemy session 实现完整的 EvolutionStoreProtocol。
+    支持进化事件、技能版本、A/B 测试结果的 CRUD 操作。
+
+    用法：
+        store = PostgreSQLEvolutionStore(
+            database_url="postgresql+asyncpg://user:pass@localhost/dbname"
+        )
+        await store.ensure_tables()
+        event_id = await store.record(event)
+    """
+
+    def __init__(self, database_url: str) -> None:
+        self._database_url = database_url
+        self._engine: Any = None
+        self._session_factory: Any = None
+        self._initialized = False
+        self._init_lock = asyncio.Lock()
+
+    async def _ensure_initialized(self) -> None:
+        """延迟初始化 async engine 和 session factory（带锁防并发）"""
+        if self._initialized:
+            return
+
+        async with self._init_lock:
+            # Double-check after acquiring lock
+            if self._initialized:
+                return
+
+            from sqlalchemy.ext.asyncio import AsyncSession, create_async_engine
+            from sqlalchemy.orm import sessionmaker
+
+            self._engine = create_async_engine(self._database_url, echo=False)
+            self._session_factory = sessionmaker(
+                self._engine, class_=AsyncSession, expire_on_commit=False
+            )
+            self._initialized = True
+
+    async def ensure_tables(self) -> None:
+        """创建所有表（如果不存在）
+
+        安全的启动调用 — 使用 CREATE TABLE IF NOT EXISTS。
+        """
+        await self._ensure_initialized()
+        async with self._engine.begin() as conn:
+            await conn.run_sync(PGBase.metadata.create_all)
+
+    async def close(self) -> None:
+        """关闭 engine，释放所有连接"""
+        if self._engine is not None:
+            await self._engine.dispose()
+            self._engine = None
+            self._session_factory = None
+            self._initialized = False
+
+    async def __aenter__(self) -> "PostgreSQLEvolutionStore":
+        return self
+
+    async def __aexit__(self, exc_type: Any, exc_val: Any, exc_tb: Any) -> None:
+        await self.close()
+
+    # ── 进化事件 ──────────────────────────────────────────
+
+    async def record(self, event: EvolutionEvent) -> str:
+        """记录进化事件"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                event_id = str(_uuid.uuid4())
+                entry = PGEvolutionEventModel(
+                    id=event_id,
+                    agent_name=event.agent_name,
+                    change_type=event.change_type,
+                    before=event.before,
+                    after=event.after,
+                    metrics=event.metrics,
+                    status="active",
+                )
+                db.add(entry)
+                await db.commit()
+                event.event_id = event_id
+                logger.info(f"Evolution event recorded: {event_id} for agent '{event.agent_name}'")
+                return event_id
+            except Exception as e:
+                await db.rollback()
+                logger.error(f"Failed to record evolution event: {e}")
+                raise
+
+    async def rollback(self, event_id: str) -> bool:
+        """回滚进化事件"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                stmt = select(PGEvolutionEventModel).where(
+                    PGEvolutionEventModel.id == event_id
+                )
+                result = await db.execute(stmt)
+                entry = result.scalar_one_or_none()
+
+                if not entry:
+                    logger.error(f"Evolution event {event_id} not found")
+                    return False
+
+                entry.status = "rolled_back"
+                await db.commit()
+                logger.info(f"Evolution event {event_id} rolled back")
+                return True
+            except Exception as e:
+                await db.rollback()
+                logger.error(f"Failed to rollback evolution event {event_id}: {e}")
+                return False
+
+    async def list_events(
+        self,
+        agent_name: str | None = None,
+        change_type: str | None = None,
+        status: str | None = None,
+        limit: int = 100,
+    ) -> list[dict]:
+        """列出进化事件"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                stmt = select(PGEvolutionEventModel)
+                if agent_name:
+                    stmt = stmt.where(PGEvolutionEventModel.agent_name == agent_name)
+                if change_type:
+                    stmt = stmt.where(PGEvolutionEventModel.change_type == change_type)
+                if status:
+                    stmt = stmt.where(PGEvolutionEventModel.status == status)
+                stmt = stmt.order_by(PGEvolutionEventModel.created_at.desc()).limit(limit)
+                result = await db.execute(stmt)
+                entries = result.scalars().all()
+
+                return [
+                    {
+                        "id": e.id,
+                        "agent_name": e.agent_name,
+                        "change_type": e.change_type,
+                        "before": e.before,
+                        "after": e.after,
+                        "metrics": e.metrics,
+                        "status": e.status,
+                        "created_at": e.created_at.isoformat() if e.created_at else None,
+                    }
+                    for e in entries
+                ]
+            except Exception as e:
+                logger.error(f"Failed to list evolution events: {e}")
+                return []
+
+    # ── 技能版本 ──────────────────────────────────────────
+
+    async def record_skill_version(
+        self, skill_name: str, version: str, content: str, parent_version: str | None = None
+    ) -> str:
+        """记录技能版本"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                vid = str(_uuid.uuid4())
+                entry = PGSkillVersionModel(
+                    id=vid,
+                    skill_name=skill_name,
+                    version=version,
+                    content=content,
+                    parent_version=parent_version,
+                )
+                db.add(entry)
+                await db.commit()
+                return vid
+            except Exception as e:
+                await db.rollback()
+                logger.error(f"Failed to record skill version: {e}")
+                raise
+
+    async def list_skill_versions(self, skill_name: str, limit: int = 100) -> list[dict]:
+        """列出技能版本历史"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                stmt = (
+                    select(PGSkillVersionModel)
+                    .where(PGSkillVersionModel.skill_name == skill_name)
+                    .order_by(PGSkillVersionModel.created_at.desc())
+                    .limit(limit)
+                )
+                result = await db.execute(stmt)
+                entries = result.scalars().all()
+                return [
+                    {
+                        "id": e.id,
+                        "skill_name": e.skill_name,
+                        "version": e.version,
+                        "content": e.content,
+                        "parent_version": e.parent_version,
+                        "created_at": e.created_at.isoformat() if e.created_at else None,
+                    }
+                    for e in entries
+                ]
+            except Exception as e:
+                logger.error(f"Failed to list skill versions: {e}")
+                return []
+
+    # ── A/B 测试结果 ──────────────────────────────────────
+
+    async def record_ab_test_result(
+        self, test_id: str, variant: str, score: float, sample_count: int = 0
+    ) -> str:
+        """记录 A/B 测试结果"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                rid = str(_uuid.uuid4())
+                entry = PGABTestResultModel(
+                    id=rid,
+                    test_id=test_id,
+                    variant=variant,
+                    score=score,
+                    sample_count=sample_count,
+                )
+                db.add(entry)
+                await db.commit()
+                return rid
+            except Exception as e:
+                await db.rollback()
+                logger.error(f"Failed to record A/B test result: {e}")
+                raise
+
+    async def get_ab_test_results(self, test_id: str, limit: int = 100) -> list[dict]:
+        """获取 A/B 测试结果"""
+        await self._ensure_initialized()
+        async with self._session_factory() as db:
+            try:
+                stmt = (
+                    select(PGABTestResultModel)
+                    .where(PGABTestResultModel.test_id == test_id)
+                    .order_by(PGABTestResultModel.created_at.desc())
+                    .limit(limit)
+                )
+                result = await db.execute(stmt)
+                entries = result.scalars().all()
+                return [
+                    {
+                        "id": e.id,
+                        "test_id": e.test_id,
+                        "variant": e.variant,
+                        "score": e.score,
+                        "sample_count": e.sample_count,
+                        "created_at": e.created_at.isoformat() if e.created_at else None,
+                    }
+                    for e in entries
+                ]
+            except Exception as e:
+                logger.error(f"Failed to get A/B test results: {e}")
+                return []

src/agentkit/llm/__init__.py

@@ -5,7 +5,8 @@ from agentkit.llm.gateway import LLMGateway
 from agentkit.llm.protocol import LLMProvider, LLMRequest, LLMResponse, TokenUsage, ToolCall
 from agentkit.llm.providers.anthropic import AnthropicProvider
 from agentkit.llm.providers.openai import OpenAICompatibleProvider
-from agentkit.llm.providers.tracker import UsageRecord, UsageSummary, UsageTracker
+from agentkit.llm.providers.tracker import UsageSummary, UsageTracker
+from agentkit.llm.providers.usage_store import UsageRecord
 from agentkit.llm.retry import (
     CircuitBreaker,
     CircuitBreakerConfig,

src/agentkit/llm/cache.py

@@ -0,0 +1,632 @@
+"""LLM Response Cache — Exact-match + Semantic-match dual cache for LLM responses.
+
+Architecture:
+    - LLMCache Protocol: async interface for cache backends
+    - InMemoryLLMCache: OrderedDict LRU + embedding index
+    - RedisLLMCache: Redis keys + SET index + lazy init
+    - create_llm_cache(): Factory with auto-detection
+
+Design doc: docs/plans/2026-06-14-002-u1-llm-cache-architecture.md
+"""
+
+import json
+import logging
+import time
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from typing import Any, Protocol, runtime_checkable
+
+from agentkit.llm.protocol import LLMResponse, TokenUsage, ToolCall
+from agentkit.utils.vector_math import compute_cosine_similarity
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Data Classes
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class CacheEntry:
+    """A cached LLM response with metadata."""
+
+    response: LLMResponse
+    query_embedding: list[float] = field(default_factory=list)
+    created_at: float = 0.0
+    hit_count: int = 0
+
+
+@dataclass
+class CacheResult:
+    """Result of a cache lookup."""
+
+    hit: bool = False
+    response: LLMResponse | None = None
+    match_type: str = ""  # "exact" | "semantic" | "" (miss)
+
+
+# ---------------------------------------------------------------------------
+# Serialization helpers (for Redis backend)
+# ---------------------------------------------------------------------------
+
+
+def _serialize_response(response: LLMResponse) -> dict:
+    """Serialize LLMResponse to a JSON-compatible dict."""
+    return {
+        "content": response.content,
+        "model": response.model,
+        "usage": {
+            "prompt_tokens": response.usage.prompt_tokens,
+            "completion_tokens": response.usage.completion_tokens,
+        },
+        "tool_calls": [
+            {"id": tc.id, "name": tc.name, "arguments": tc.arguments}
+            for tc in response.tool_calls
+        ],
+        "latency_ms": response.latency_ms,
+    }
+
+
+def _deserialize_response(data: dict) -> LLMResponse:
+    """Deserialize a dict back to LLMResponse."""
+    usage_data = data.get("usage", {})
+    return LLMResponse(
+        content=data["content"],
+        model=data["model"],
+        usage=TokenUsage(
+            prompt_tokens=usage_data.get("prompt_tokens", 0),
+            completion_tokens=usage_data.get("completion_tokens", 0),
+        ),
+        tool_calls=[
+            ToolCall(id=tc["id"], name=tc["name"], arguments=tc["arguments"])
+            for tc in data.get("tool_calls", [])
+        ],
+        latency_ms=data.get("latency_ms", 0.0),
+    )
+
+
+def _serialize_entry(entry: CacheEntry) -> dict:
+    """Serialize CacheEntry to a JSON-compatible dict."""
+    return {
+        "response": _serialize_response(entry.response),
+        "query_embedding": entry.query_embedding,
+        "created_at": entry.created_at,
+        "hit_count": entry.hit_count,
+    }
+
+
+def _deserialize_entry(data: dict) -> CacheEntry:
+    """Deserialize a dict back to CacheEntry."""
+    return CacheEntry(
+        response=_deserialize_response(data["response"]),
+        query_embedding=data.get("query_embedding", []),
+        created_at=data.get("created_at", 0.0),
+        hit_count=data.get("hit_count", 0),
+    )
+
+
+# ---------------------------------------------------------------------------
+# LLMCache Protocol
+# ---------------------------------------------------------------------------
+
+
+@runtime_checkable
+class LLMCache(Protocol):
+    """LLM response cache interface."""
+
+    async def get(self, key: str) -> CacheResult:
+        """Exact-match lookup by cache key."""
+        ...
+
+    async def semantic_search(
+        self, query_embedding: list[float], threshold: float = 0.92
+    ) -> CacheResult:
+        """Semantic similarity search across cached entries."""
+        ...
+
+    async def put(
+        self,
+        key: str,
+        response: LLMResponse,
+        query_embedding: list[float] | None = None,
+    ) -> None:
+        """Store a response in the cache with optional embedding."""
+        ...
+
+    async def invalidate(self, pattern: str | None = None) -> int:
+        """Invalidate cache entries. Returns count of invalidated entries."""
+        ...
+
+    async def stats(self) -> dict[str, int]:
+        """Return cache statistics."""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# InMemoryLLMCache
+# ---------------------------------------------------------------------------
+
+
+class InMemoryLLMCache:
+    """In-memory LLM cache with LRU eviction and semantic search.
+
+    Uses OrderedDict for O(1) LRU access/eviction (follows EmbeddingCache pattern).
+    Maintains a parallel embedding index for semantic similarity search.
+    """
+
+    def __init__(
+        self,
+        max_entries: int = 10000,
+        exact_ttl: int = 3600,
+        semantic_ttl: int = 86400,
+        similarity_threshold: float = 0.92,
+    ):
+        self._max_entries = max_entries
+        self._exact_ttl = exact_ttl
+        self._semantic_ttl = semantic_ttl
+        self._similarity_threshold = similarity_threshold
+
+        self._cache: OrderedDict[str, CacheEntry] = OrderedDict()
+        self._embeddings: dict[str, list[float]] = {}
+
+        self._hits = 0
+        self._misses = 0
+
+    async def get(self, key: str) -> CacheResult:
+        now = time.monotonic()
+        entry = self._cache.get(key)
+
+        if entry is not None:
+            if now - entry.created_at <= self._exact_ttl:
+                # Hit: update LRU position and stats
+                self._cache.move_to_end(key)
+                entry.hit_count += 1
+                self._hits += 1
+                return CacheResult(hit=True, response=entry.response, match_type="exact")
+            # Expired: remove
+            del self._cache[key]
+            self._embeddings.pop(key, None)
+
+        self._misses += 1
+        return CacheResult(hit=False)
+
+    async def semantic_search(
+        self, query_embedding: list[float], threshold: float | None = None
+    ) -> CacheResult:
+        if not self._embeddings:
+            return CacheResult(hit=False)
+
+        effective_threshold = threshold or self._similarity_threshold
+        now = time.monotonic()
+        best_key: str | None = None
+        best_sim: float = 0.0
+
+        for key, emb in self._embeddings.items():
+            entry = self._cache.get(key)
+            if entry is None:
+                continue
+            # Check semantic TTL
+            if now - entry.created_at > self._semantic_ttl:
+                continue
+            sim = compute_cosine_similarity(query_embedding, emb)
+            if sim > best_sim:
+                best_sim = sim
+                best_key = key
+
+        if best_key is not None and best_sim >= effective_threshold:
+            entry = self._cache[best_key]
+            entry.hit_count += 1
+            self._cache.move_to_end(best_key)
+            self._hits += 1
+            return CacheResult(hit=True, response=entry.response, match_type="semantic")
+
+        self._misses += 1
+        return CacheResult(hit=False)
+
+    async def put(
+        self,
+        key: str,
+        response: LLMResponse,
+        query_embedding: list[float] | None = None,
+    ) -> None:
+        now = time.monotonic()
+
+        if key in self._cache:
+            self._cache.move_to_end(key)
+            existing = self._cache[key]
+            # Preserve existing embedding if new one is None
+            effective_embedding = query_embedding if query_embedding is not None else existing.query_embedding
+        else:
+            effective_embedding = query_embedding or []
+
+        self._cache[key] = CacheEntry(
+            response=response,
+            query_embedding=effective_embedding,
+            created_at=now,
+            hit_count=0,
+        )
+
+        if effective_embedding:
+            self._embeddings[key] = effective_embedding
+
+        # Evict LRU entries if over capacity
+        while len(self._cache) > self._max_entries:
+            evicted_key, _ = self._cache.popitem(last=False)
+            self._embeddings.pop(evicted_key, None)
+
+        # Lazy cleanup: remove a few expired entries on each put to prevent memory leak
+        # Check oldest entries first (they are most likely to be expired)
+        if len(self._cache) > 0:
+            expired_keys = []
+            # Iterate from oldest (front of OrderedDict) to find expired entries
+            for k in list(self._cache.keys())[:20]:
+                entry = self._cache.get(k)
+                if entry is not None and now - entry.created_at > self._semantic_ttl:
+                    expired_keys.append(k)
+            for k in expired_keys:
+                self._cache.pop(k, None)
+                self._embeddings.pop(k, None)
+
+    async def invalidate(self, pattern: str | None = None) -> int:
+        if pattern is None:
+            count = len(self._cache)
+            self._cache.clear()
+            self._embeddings.clear()
+            return count
+
+        # Simple prefix matching for pattern
+        keys_to_remove = [
+            k for k in self._cache if k.startswith(pattern.replace("*", ""))
+        ]
+        for key in keys_to_remove:
+            del self._cache[key]
+            self._embeddings.pop(key, None)
+        return len(keys_to_remove)
+
+    async def stats(self) -> dict[str, int]:
+        return {
+            "total_entries": len(self._cache),
+            "total_hits": self._hits,
+            "total_misses": self._misses,
+        }
+
+
+# ---------------------------------------------------------------------------
+# RedisLLMCache
+# ---------------------------------------------------------------------------
+
+
+class RedisLLMCache:
+    """Redis-backed LLM cache with SET index for semantic search.
+
+    Key schema:
+        agentkit:llm_cache:{sha256_hex}     → JSON(CacheEntry) with TTL
+        agentkit:llm_cache_emb:{sha256_hex} → JSON(list[float]) with TTL
+        agentkit:llm_cache_index            → SET of active cache keys
+    """
+
+    KEY_PREFIX = "agentkit:llm_cache:"
+    EMB_PREFIX = "agentkit:llm_cache_emb:"
+    INDEX_KEY = "agentkit:llm_cache_index"
+
+    def __init__(
+        self,
+        redis_url: str = "redis://localhost:6379",
+        max_entries: int = 10000,
+        exact_ttl: int = 3600,
+        semantic_ttl: int = 86400,
+        similarity_threshold: float = 0.92,
+        max_entries_to_scan: int = 500,
+        fallback: InMemoryLLMCache | None = None,
+    ):
+        self._redis_url = redis_url
+        self._max_entries = max_entries
+        self._exact_ttl = exact_ttl
+        self._semantic_ttl = semantic_ttl
+        self._similarity_threshold = similarity_threshold
+        self._max_entries_to_scan = max_entries_to_scan
+        self._redis: Any = None
+        self._fallback: InMemoryLLMCache | None = fallback  # For auto-degradation
+        self._degraded = False  # True if Redis is unreachable
+
+        self._hits = 0
+        self._misses = 0
+
+    async def _get_redis(self):
+        """Lazy Redis initialization (follows RedisSessionStore pattern)."""
+        if self._redis is None:
+            import redis.asyncio as aioredis
+
+            self._redis = aioredis.from_url(
+                self._redis_url, decode_responses=True
+            )
+        return self._redis
+
+    async def aclose(self) -> None:
+        """Close the Redis connection pool."""
+        if self._redis is not None:
+            await self._redis.aclose()
+            self._redis = None
+
+    def _degrade_to_fallback(self) -> None:
+        """Mark Redis as unreachable and switch to in-memory fallback."""
+        if not self._degraded:
+            self._degraded = True
+            self._degrade_count = 0
+            if self._fallback is None:
+                self._fallback = InMemoryLLMCache(
+                    max_entries=self._max_entries,
+                    exact_ttl=self._exact_ttl,
+                    semantic_ttl=self._semantic_ttl,
+                    similarity_threshold=self._similarity_threshold,
+                )
+            logger.warning("Redis cache unreachable, degraded to in-memory fallback")
+
+    def _try_recover(self) -> None:
+        """Attempt to recover from degraded state after enough operations.
+
+        Resets the degraded flag optimistically. The next actual Redis
+        operation will verify connectivity — if it fails, degradation
+        is re-triggered immediately.
+        """
+        if not self._degraded:
+            return
+        self._degrade_count = getattr(self, "_degrade_count", 0) + 1
+        # Try recovery every 100 operations
+        if self._degrade_count >= 100:
+            self._degrade_count = 0
+            self._degraded = False
+            logger.info("Redis cache: attempting recovery from degraded state")
+
+    async def get(self, key: str) -> CacheResult:
+        # If degraded to fallback, use InMemory cache
+        if self._degraded and self._fallback is not None:
+            self._try_recover()
+            if self._degraded:
+                return await self._fallback.get(key)
+            # Recovery attempted — fall through to try Redis
+
+        try:
+            redis = await self._get_redis()
+            data = await redis.get(f"{self.KEY_PREFIX}{key}")
+            if data is not None:
+                entry = _deserialize_entry(json.loads(data))
+                self._hits += 1
+                return CacheResult(
+                    hit=True, response=entry.response, match_type="exact"
+                )
+            self._misses += 1
+            return CacheResult(hit=False)
+        except Exception as e:
+            logger.warning(f"Redis cache get failed, returning miss: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                return await self._fallback.get(key)
+            return CacheResult(hit=False)
+
+    async def semantic_search(
+        self, query_embedding: list[float], threshold: float | None = None
+    ) -> CacheResult:
+        try:
+            redis = await self._get_redis()
+            effective_threshold = threshold or self._similarity_threshold
+
+            # Get all cache keys from index
+            cache_keys = await redis.smembers(self.INDEX_KEY)
+            if not cache_keys:
+                return CacheResult(hit=False)
+
+            # Limit scan to avoid O(n) memory/network transfer for large caches
+            # Sample up to max_entries_to_scan most recent keys
+            cache_keys_list = list(cache_keys)
+            max_scan = min(len(cache_keys_list), self._max_entries_to_scan)
+            if len(cache_keys_list) > max_scan:
+                # Take a random sample to avoid always scanning the same subset
+                import random
+                cache_keys_list = random.sample(cache_keys_list, max_scan)
+
+            # Batch fetch embeddings
+            emb_keys = [f"{self.EMB_PREFIX}{k}" for k in cache_keys_list]
+            emb_values = await redis.mget(emb_keys)
+
+            best_key: str | None = None
+            best_sim: float = 0.0
+            stale_keys: list[str] = []  # Keys whose data has expired
+
+            for cache_key, emb_json in zip(cache_keys_list, emb_values):
+                if emb_json is None:
+                    # Embedding expired but index entry remains — mark for cleanup
+                    stale_keys.append(cache_key)
+                    continue
+                emb = json.loads(emb_json)
+                sim = compute_cosine_similarity(query_embedding, emb)
+                if sim > best_sim:
+                    best_sim = sim
+                    best_key = cache_key
+
+            # Lazy cleanup: remove stale index entries
+            if stale_keys:
+                try:
+                    pipe = redis.pipeline()
+                    for k in stale_keys:
+                        pipe.srem(self.INDEX_KEY, k)
+                    await pipe.execute()
+                except Exception:
+                    pass  # Best-effort cleanup
+
+            if best_key is not None and best_sim >= effective_threshold:
+                data = await redis.get(f"{self.KEY_PREFIX}{best_key}")
+                if data is not None:
+                    entry = _deserialize_entry(json.loads(data))
+                    self._hits += 1
+                    return CacheResult(
+                        hit=True, response=entry.response, match_type="semantic"
+                    )
+                # Data key expired but embedding still exists — mark for cleanup
+                try:
+                    await redis.srem(self.INDEX_KEY, best_key)
+                except Exception:
+                    pass
+
+            self._misses += 1
+            return CacheResult(hit=False)
+        except Exception as e:
+            logger.warning(f"Redis semantic search failed, returning miss: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                return await self._fallback.semantic_search(query_embedding, threshold)
+            self._misses += 1
+            return CacheResult(hit=False)
+
+    async def put(
+        self,
+        key: str,
+        response: LLMResponse,
+        query_embedding: list[float] | None = None,
+    ) -> None:
+        # If degraded to fallback, use InMemory cache
+        if self._degraded and self._fallback is not None:
+            self._try_recover()
+            if self._degraded:
+                await self._fallback.put(key, response, query_embedding)
+                return
+            # Recovery attempted — fall through to try Redis
+
+        try:
+            redis = await self._get_redis()
+            entry = CacheEntry(
+                response=response,
+                query_embedding=query_embedding or [],
+                created_at=time.time(),  # Wall-clock for cross-process comparability in Redis
+                hit_count=0,
+            )
+
+            pipe = redis.pipeline()
+            # Data key TTL must cover both exact and semantic windows
+            # so semantic hits don't return None data
+            data_ttl = max(self._exact_ttl, self._semantic_ttl)
+            pipe.set(
+                f"{self.KEY_PREFIX}{key}",
+                json.dumps(_serialize_entry(entry)),
+                ex=data_ttl,
+            )
+            if query_embedding is not None:
+                pipe.set(
+                    f"{self.EMB_PREFIX}{key}",
+                    json.dumps(query_embedding),
+                    ex=self._semantic_ttl,
+                )
+            pipe.sadd(self.INDEX_KEY, key)
+            await pipe.execute()
+        except Exception as e:
+            logger.warning(f"Redis cache put failed: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                await self._fallback.put(key, response, query_embedding)
+
+    async def invalidate(self, pattern: str | None = None) -> int:
+        try:
+            redis = await self._get_redis()
+
+            if pattern is None:
+                cache_keys = await redis.smembers(self.INDEX_KEY)
+                if not cache_keys:
+                    return 0
+                pipe = redis.pipeline()
+                for key in cache_keys:
+                    pipe.delete(f"{self.KEY_PREFIX}{key}")
+                    pipe.delete(f"{self.EMB_PREFIX}{key}")
+                pipe.delete(self.INDEX_KEY)
+                await pipe.execute()
+                return len(cache_keys)
+
+            # Pattern-based invalidation (prefix match)
+            prefix = pattern.replace("*", "")
+            cache_keys = await redis.smembers(self.INDEX_KEY)
+            keys_to_remove = [k for k in cache_keys if k.startswith(prefix)]
+
+            if not keys_to_remove:
+                return 0
+
+            pipe = redis.pipeline()
+            for key in keys_to_remove:
+                pipe.delete(f"{self.KEY_PREFIX}{key}")
+                pipe.delete(f"{self.EMB_PREFIX}{key}")
+                pipe.srem(self.INDEX_KEY, key)
+            await pipe.execute()
+            return len(keys_to_remove)
+        except Exception as e:
+            logger.warning(f"Redis cache invalidate failed: {e}")
+            return 0
+
+    async def stats(self) -> dict[str, int]:
+        try:
+            redis = await self._get_redis()
+            total_entries = await redis.scard(self.INDEX_KEY)
+            return {
+                "total_entries": total_entries,
+                "total_hits": self._hits,
+                "total_misses": self._misses,
+            }
+        except Exception:
+            return {
+                "total_entries": 0,
+                "total_hits": self._hits,
+                "total_misses": self._misses,
+            }
+
+
+# ---------------------------------------------------------------------------
+# Factory
+# ---------------------------------------------------------------------------
+
+
+def create_llm_cache(
+    backend: str = "auto",
+    redis_url: str = "redis://localhost:6379",
+    max_entries: int = 10000,
+    exact_ttl: int = 3600,
+    semantic_ttl: int = 86400,
+    similarity_threshold: float = 0.92,
+) -> LLMCache:
+    """Create an LLM cache backend.
+
+    Args:
+        backend: "auto" (try Redis, fallback to memory), "redis", "memory".
+        redis_url: Redis connection URL (only used for "redis"/"auto" backend).
+        max_entries: Maximum number of cache entries.
+        exact_ttl: TTL in seconds for exact-match cache entries.
+        semantic_ttl: TTL in seconds for semantic-match embeddings.
+        similarity_threshold: Cosine similarity threshold for semantic match.
+
+    Returns:
+        An LLMCache instance.
+    """
+    if backend in ("auto", "redis"):
+        try:
+            import redis.asyncio as aioredis  # noqa: F401
+
+            return RedisLLMCache(
+                redis_url=redis_url,
+                max_entries=max_entries,
+                exact_ttl=exact_ttl,
+                semantic_ttl=semantic_ttl,
+                similarity_threshold=similarity_threshold,
+            )
+        except ImportError:
+            logger.warning(
+                "redis package not available, falling back to in-memory cache"
+            )
+            return InMemoryLLMCache(
+                max_entries=max_entries,
+                exact_ttl=exact_ttl,
+                semantic_ttl=semantic_ttl,
+                similarity_threshold=similarity_threshold,
+            )
+    return InMemoryLLMCache(
+        max_entries=max_entries,
+        exact_ttl=exact_ttl,
+        semantic_ttl=semantic_ttl,
+        similarity_threshold=similarity_threshold,
+    )

src/agentkit/llm/cache_key.py

@@ -0,0 +1,66 @@
+"""LLM Cache Key Generation — Deterministic SHA-256 cache key from LLM request parameters."""
+
+import hashlib
+import json
+from typing import Any
+
+
+def generate_cache_key(
+    model: str,
+    messages: list[dict[str, str]],
+    temperature: float,
+    tools: list[dict[str, Any]] | None = None,
+    tool_choice: str = "auto",
+    max_tokens: int = 2000,
+) -> str:
+    """Generate a deterministic SHA-256 cache key from LLM request parameters.
+
+    The key captures ALL inputs that deterministically affect LLM output:
+    model, system_prompt (extracted from messages), messages content,
+    temperature, tools, tool_choice, and max_tokens.
+
+    Args:
+        model: Model identifier (e.g. "openai/gpt-4o").
+        messages: Chat messages list (may include system prompt as first message).
+        temperature: Sampling temperature.
+        tools: Optional list of tool definitions.
+        tool_choice: Tool selection mode ("auto", "none", etc.).
+        max_tokens: Maximum response tokens.
+
+    Returns:
+        64-character hex SHA-256 hash string.
+    """
+    system_prompt = _extract_system_prompt(messages)
+    components = [
+        _hash_str(model),
+        _hash_str(system_prompt),
+        _hash_json(messages),
+        _hash_str(f"{temperature:.2f}"),
+        _hash_json(tools),
+        _hash_str(tool_choice),
+        _hash_str(str(max_tokens)),
+    ]
+    combined = "".join(components)
+    return hashlib.sha256(combined.encode()).hexdigest()
+
+
+def _extract_system_prompt(messages: list[dict[str, str]]) -> str:
+    """Extract system prompt from messages list."""
+    for msg in messages:
+        if msg.get("role") == "system":
+            return msg.get("content", "")
+    return ""
+
+
+def _hash_str(s: str) -> str:
+    """SHA-256 hash of a string."""
+    return hashlib.sha256(s.encode()).hexdigest()
+
+
+def _hash_json(obj: Any) -> str:
+    """SHA-256 hash of a JSON-serializable object."""
+    if obj is None:
+        return hashlib.sha256(b"null").hexdigest()
+    return hashlib.sha256(
+        json.dumps(obj, sort_keys=True, ensure_ascii=False).encode()
+    ).hexdigest()

src/agentkit/llm/config.py

@@ -8,6 +8,43 @@ import yaml
 from agentkit.llm.retry import CircuitBreakerConfig, RetryConfig
 
 
+@dataclass
+class CacheConfig:
+    """LLM Cache 配置"""
+
+    enabled: bool = False
+    backend: str = "auto"  # "auto" | "redis" | "memory"
+    redis_url: str = "redis://localhost:6379"
+    exact_ttl: int = 3600
+    semantic_ttl: int = 86400
+    similarity_threshold: float = 0.92
+    max_entries: int = 10000
+    # Embedding config for semantic cache (Chinese-first: bge-m3 via Xinference)
+    embedding_provider: str = "openai"  # "openai" | "xinference" | "local"
+    embedding_model: str = "bge-m3"
+    embedding_base_url: str | None = None
+    embedding_api_key: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict) -> "CacheConfig":
+        if not data:
+            return cls()
+        emb = data.get("embedding", {})
+        return cls(
+            enabled=data.get("enabled", False),
+            backend=data.get("backend", "auto"),
+            redis_url=data.get("redis_url", "redis://localhost:6379"),
+            exact_ttl=data.get("exact_ttl", 3600),
+            semantic_ttl=data.get("semantic_ttl", 86400),
+            similarity_threshold=data.get("similarity_threshold", 0.92),
+            max_entries=data.get("max_entries", 10000),
+            embedding_provider=emb.get("provider", "openai"),
+            embedding_model=emb.get("model", "bge-m3"),
+            embedding_base_url=emb.get("base_url"),
+            embedding_api_key=emb.get("api_key"),
+        )
+
+
 @dataclass
 class ProviderConfig:
     """Provider 配置"""
@@ -32,6 +69,7 @@ class LLMConfig:
     providers: dict[str, ProviderConfig] = field(default_factory=dict)
     model_aliases: dict[str, str] = field(default_factory=dict)
     fallbacks: dict[str, list[str]] = field(default_factory=dict)
+    cache: CacheConfig | None = None
 
     @classmethod
     def from_yaml(cls, path: str) -> "LLMConfig":
@@ -77,8 +115,14 @@ class LLMConfig:
                 retry=retry,
                 circuit_breaker=circuit_breaker,
             )
+        cache = None
+        cache_data = data.get("cache")
+        if cache_data:
+            cache = CacheConfig.from_dict(cache_data)
+
         return cls(
             providers=providers,
             model_aliases=data.get("model_aliases", {}),
             fallbacks=data.get("fallbacks", {}),
+            cache=cache,
         )

src/agentkit/llm/gateway.py

@@ -2,6 +2,7 @@
 
 import logging
 import time
+from typing import Any
 
 from agentkit.core.exceptions import LLMProviderError, ModelNotFoundError
 from agentkit.llm.config import LLMConfig
@@ -14,13 +15,53 @@ logger = logging.getLogger(__name__)
 
 
 class LLMGateway:
-    """LLM 网关 - Provider 注册、模型别名解析、Fallback、Usage 追踪"""
+    """LLM 网关 - Provider 注册、模型别名解析、Fallback、Usage 追踪、Cache"""
 
-    def __init__(self, config: LLMConfig | None = None):
+    def __init__(self, config: LLMConfig | None = None, usage_store: Any = None):
         self._providers: dict[str, LLMProvider] = {}
-        self._usage_tracker = UsageTracker()
+        self._usage_tracker = UsageTracker(store=usage_store) if usage_store else UsageTracker()
         self._config = config or LLMConfig()
 
+        # Cache (opt-in, disabled by default)
+        self._cache: Any = None  # LLMCache | None
+        self._embedder: Any = None  # Embedder | None
+        if self._config.cache and self._config.cache.enabled:
+            from agentkit.llm.cache import create_llm_cache
+            self._cache = create_llm_cache(
+                backend=self._config.cache.backend,
+                redis_url=self._config.cache.redis_url,
+                max_entries=self._config.cache.max_entries,
+                exact_ttl=self._config.cache.exact_ttl,
+                semantic_ttl=self._config.cache.semantic_ttl,
+                similarity_threshold=self._config.cache.similarity_threshold,
+            )
+            self._embedder = self._create_embedder(self._config.cache)
+            logger.info(
+                f"LLM cache enabled (backend={self._config.cache.backend}, "
+                f"embedder={self._config.cache.embedding_provider}/{self._config.cache.embedding_model})"
+            )
+
+    def _create_embedder(self, cache_config) -> Any:
+        """Create embedder for semantic cache based on config."""
+        try:
+            from agentkit.memory.embedder import OpenAIEmbedder
+
+            if cache_config.embedding_provider in ("xinference", "local"):
+                return OpenAIEmbedder(
+                    api_key=cache_config.embedding_api_key or "not-needed",
+                    model=cache_config.embedding_model,
+                    base_url=cache_config.embedding_base_url or "http://localhost:9997/v1",
+                )
+            # Default: OpenAI
+            return OpenAIEmbedder(
+                api_key=cache_config.embedding_api_key,
+                model=cache_config.embedding_model,
+                base_url=cache_config.embedding_base_url,
+            )
+        except Exception as e:
+            logger.warning(f"Failed to create embedder for semantic cache: {e}")
+            return None
+
     def register_provider(self, name: str, provider: LLMProvider) -> None:
         """注册 Provider"""
         self._providers[name] = provider
@@ -66,6 +107,66 @@ class LLMGateway:
                 _span = _span_cm.__enter__()
 
         start = time.monotonic()
+
+        # ── Cache check ──
+        cache_key = None
+        query_embedding = None
+        if self._cache is not None:
+            from agentkit.llm.cache_key import generate_cache_key
+
+            cache_key = generate_cache_key(
+                model=resolved_model,
+                messages=messages,
+                temperature=kwargs.get("temperature", 0.7),
+                tools=tools,
+                tool_choice=tool_choice,
+                max_tokens=kwargs.get("max_tokens", 2000),
+            )
+            result = await self._cache.get(cache_key)
+            if result.hit:
+                latency_ms = (time.monotonic() - start) * 1000
+                self._usage_tracker.record(
+                    agent_name=agent_name,
+                    model=result.response.model,
+                    usage=result.response.usage,
+                    cost=0.0,
+                    latency_ms=latency_ms,
+                )
+                if _span is not None:
+                    _span.set_attribute("gen_ai.cache.hit", True)
+                    _span.set_attribute("gen_ai.cache.match_type", result.match_type)
+                return result.response
+
+            # Semantic match (only for temperature == 0)
+            temperature = kwargs.get("temperature", 0.7)
+            if temperature == 0 and self._embedder is not None:
+                try:
+                    # Embed last N messages for context-aware semantic matching
+                    # (not just last user message — avoids cross-context false hits)
+                    recent_messages = messages[-3:] if len(messages) > 3 else messages
+                    embed_text = " | ".join(
+                        m.get("content", "") for m in recent_messages if m.get("content")
+                    )
+                    if embed_text:
+                        query_embedding = await self._embedder.embed(embed_text)
+                        result = await self._cache.semantic_search(query_embedding)
+                        if result.hit:
+                            latency_ms = (time.monotonic() - start) * 1000
+                            self._usage_tracker.record(
+                                agent_name=agent_name,
+                                model=result.response.model,
+                                usage=result.response.usage,
+                                cost=0.0,
+                                latency_ms=latency_ms,
+                            )
+                            if _span is not None:
+                                _span.set_attribute("gen_ai.cache.hit", True)
+                                _span.set_attribute("gen_ai.cache.match_type", "semantic")
+                            return result.response
+                except Exception as e:
+                    logger.warning(f"Semantic cache search failed: {e}")
+
+        # ── Normal provider call ──
         models_to_try = self._get_models_to_try(resolved_model)
         last_error: LLMProviderError | None = None
 
@@ -95,6 +196,13 @@ class LLMGateway:
 
             latency_ms = (time.monotonic() - start) * 1000
 
+            # ── Cache write ──
+            if self._cache is not None and cache_key is not None:
+                try:
+                    await self._cache.put(cache_key, response, query_embedding)
+                except Exception as e:
+                    logger.warning(f"Cache write failed: {e}")
+
             # 计算成本
             cost = self._calculate_cost(response.model, response.usage)
 
@@ -112,7 +220,9 @@ class LLMGateway:
                 _span.set_attribute("gen_ai.usage.input_tokens", response.usage.prompt_tokens)
                 _span.set_attribute("gen_ai.usage.output_tokens", response.usage.completion_tokens)
                 _span.set_attribute("gen_ai.response.model", response.model)
-                _span.set_attribute("gen_ai.duration_ms", int(latency_ms))
+                _span.set_attribute("gen_ai.duration.ms", int(latency_ms))
+                if self._cache is not None:
+                    _span.set_attribute("gen_ai.cache.hit", False)
             llm_token_histogram().record(
                 response.usage.total_tokens,
                 {"gen_ai.request.model": resolved_model},
@@ -138,6 +248,8 @@ class LLMGateway:
         If the primary model fails before any chunk is yielded, tries fallback
         models. If it fails after chunks have been sent, yields an error chunk
         and terminates (cannot switch mid-stream).
+
+        Note: Streaming responses are NOT cached in this iteration.
         """
         resolved_model = self._resolve_model_alias(model)
 

src/agentkit/llm/providers/__init__.py

@@ -4,7 +4,8 @@ from agentkit.llm.providers.anthropic import AnthropicProvider
 from agentkit.llm.providers.doubao import DoubaoProvider
 from agentkit.llm.providers.gemini import GeminiProvider
 from agentkit.llm.providers.openai import OpenAICompatibleProvider
-from agentkit.llm.providers.tracker import UsageRecord, UsageSummary, UsageTracker
+from agentkit.llm.providers.tracker import UsageSummary, UsageTracker
+from agentkit.llm.providers.usage_store import UsageRecord
 from agentkit.llm.providers.wenxin import WenxinProvider
 from agentkit.llm.providers.yuanbao import YuanbaoProvider
 

src/agentkit/llm/providers/tracker.py

@@ -1,42 +1,20 @@
-"""Usage Tracker - 使用量追踪"""
+"""Usage Tracker - 使用量追踪（委托给 UsageStore）"""
 
-from dataclasses import dataclass, field
-from datetime import datetime, timezone
+from datetime import datetime
 
 from agentkit.llm.protocol import TokenUsage
-
-
-@dataclass
-class UsageRecord:
-    """使用量记录"""
-
-    agent_name: str
-    model: str
-    prompt_tokens: int
-    completion_tokens: int
-    total_tokens: int
-    cost: float
-    latency_ms: float
-    timestamp: datetime = field(default_factory=lambda: datetime.now(timezone.utc))
-
-
-@dataclass
-class UsageSummary:
-    """使用量汇总"""
-
-    total_tokens: int = 0
-    total_cost: float = 0.0
-    by_model: dict[str, dict[str, int | float]] = field(default_factory=dict)
-    records: list[UsageRecord] = field(default_factory=list)
+from agentkit.llm.providers.usage_store import (
+    InMemoryUsageStore,
+    UsageStore,
+    UsageSummary,
+)
 
 
 class UsageTracker:
-    """使用量追踪器"""
+    """使用量追踪器 — 委托给可插拔的 UsageStore"""
 
-    MAX_RECORDS = 10000  # 最大记录数，防止内存无限增长
-
-    def __init__(self) -> None:
-        self._records: list[UsageRecord] = []
+    def __init__(self, store: UsageStore | None = None) -> None:
+        self._store: UsageStore = store or InMemoryUsageStore()
 
     def record(
         self,
@@ -47,19 +25,7 @@ class UsageTracker:
         latency_ms: float,
     ) -> None:
         """记录一次使用"""
-        rec = UsageRecord(
-            agent_name=agent_name,
-            model=model,
-            prompt_tokens=usage.prompt_tokens,
-            completion_tokens=usage.completion_tokens,
-            total_tokens=usage.total_tokens,
-            cost=cost,
-            latency_ms=latency_ms,
-        )
-        self._records.append(rec)
-        # 超过上限时删除最早的记录
-        if len(self._records) > self.MAX_RECORDS:
-            self._records = self._records[-self.MAX_RECORDS:]
+        self._store.record(agent_name, model, usage, cost, latency_ms)
 
     def get_usage(
         self,
@@ -68,32 +34,4 @@ class UsageTracker:
         end_time: datetime | None = None,
     ) -> UsageSummary:
         """查询使用量汇总"""
-        filtered = self._records
-
-        if agent_name is not None:
-            filtered = [r for r in filtered if r.agent_name == agent_name]
-        if start_time is not None:
-            filtered = [r for r in filtered if r.timestamp >= start_time]
-        if end_time is not None:
-            filtered = [r for r in filtered if r.timestamp <= end_time]
-
-        if not filtered:
-            return UsageSummary()
-
-        total_tokens = sum(r.total_tokens for r in filtered)
-        total_cost = sum(r.cost for r in filtered)
-
-        by_model: dict[str, dict[str, int | float]] = {}
-        for r in filtered:
-            if r.model not in by_model:
-                by_model[r.model] = {"total_tokens": 0, "total_cost": 0.0, "count": 0}
-            by_model[r.model]["total_tokens"] += r.total_tokens
-            by_model[r.model]["total_cost"] += r.cost
-            by_model[r.model]["count"] += 1
-
-        return UsageSummary(
-            total_tokens=total_tokens,
-            total_cost=total_cost,
-            by_model=by_model,
-            records=filtered,
-        )
+        return self._store.get_usage(agent_name, start_time, end_time)

src/agentkit/llm/providers/usage_store.py

@@ -0,0 +1,373 @@
+"""Usage Store — Persistent usage tracking with Redis Hash backend.
+
+Provides UsageStore Protocol with InMemoryUsageStore and RedisUsageStore
+backends. Replaces the in-memory list in UsageTracker with a pluggable
+store that survives restarts and supports multi-instance deployment.
+
+Key schema (Redis):
+    agentkit:usage:{date}           → Hash: {agent_name:model → JSON(UsageBucket)}
+    agentkit:usage_records:{date}   → List: JSON(UsageRecord) with LTRIM
+"""
+
+import json
+import logging
+from dataclasses import dataclass, field
+from datetime import datetime, timedelta, timezone
+from typing import Any, Protocol, runtime_checkable
+
+from agentkit.llm.protocol import TokenUsage
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class UsageRecord:
+    """使用量记录"""
+
+    agent_name: str
+    model: str
+    prompt_tokens: int
+    completion_tokens: int
+    total_tokens: int
+    cost: float
+    latency_ms: float
+    timestamp: str = ""  # ISO 8601 string for JSON serialization
+
+    def __post_init__(self):
+        if not self.timestamp:
+            self.timestamp = datetime.now(timezone.utc).isoformat()
+
+
+@dataclass
+class UsageBucket:
+    """Aggregated usage for an agent+model pair on a given date."""
+
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    total_tokens: int = 0
+    cost: float = 0.0
+    count: int = 0
+
+
+@dataclass
+class UsageSummary:
+    """使用量汇总"""
+
+    total_tokens: int = 0
+    total_cost: float = 0.0
+    by_model: dict[str, dict[str, int | float]] = field(default_factory=dict)
+    records: list[UsageRecord] = field(default_factory=list)
+
+
+# ---------------------------------------------------------------------------
+# UsageStore Protocol
+# ---------------------------------------------------------------------------
+
+
+@runtime_checkable
+class UsageStore(Protocol):
+    """Persistent usage store interface."""
+
+    def record(
+        self,
+        agent_name: str,
+        model: str,
+        usage: TokenUsage,
+        cost: float,
+        latency_ms: float,
+    ) -> None:
+        """Record a usage event."""
+        ...
+
+    def get_usage(
+        self,
+        agent_name: str | None = None,
+        start_time: datetime | None = None,
+        end_time: datetime | None = None,
+    ) -> UsageSummary:
+        """Query usage summary."""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# InMemoryUsageStore
+# ---------------------------------------------------------------------------
+
+
+class InMemoryUsageStore:
+    """In-memory usage store (drop-in replacement for old UsageTracker)."""
+
+    MAX_RECORDS = 10000
+
+    def __init__(self):
+        self._records: list[UsageRecord] = []
+
+    def record(
+        self,
+        agent_name: str,
+        model: str,
+        usage: TokenUsage,
+        cost: float,
+        latency_ms: float,
+    ) -> None:
+        rec = UsageRecord(
+            agent_name=agent_name,
+            model=model,
+            prompt_tokens=usage.prompt_tokens,
+            completion_tokens=usage.completion_tokens,
+            total_tokens=usage.total_tokens,
+            cost=cost,
+            latency_ms=latency_ms,
+        )
+        self._records.append(rec)
+        if len(self._records) > self.MAX_RECORDS:
+            self._records = self._records[-self.MAX_RECORDS:]
+
+    def get_usage(
+        self,
+        agent_name: str | None = None,
+        start_time: datetime | None = None,
+        end_time: datetime | None = None,
+    ) -> UsageSummary:
+        filtered = self._records
+
+        if agent_name is not None:
+            filtered = [r for r in filtered if r.agent_name == agent_name]
+        if start_time is not None:
+            filtered = [r for r in filtered if datetime.fromisoformat(r.timestamp) >= start_time]
+        if end_time is not None:
+            filtered = [r for r in filtered if datetime.fromisoformat(r.timestamp) <= end_time]
+
+        if not filtered:
+            return UsageSummary()
+
+        total_tokens = sum(r.total_tokens for r in filtered)
+        total_cost = sum(r.cost for r in filtered)
+
+        by_model: dict[str, dict[str, int | float]] = {}
+        for r in filtered:
+            if r.model not in by_model:
+                by_model[r.model] = {"total_tokens": 0, "total_cost": 0.0, "count": 0}
+            by_model[r.model]["total_tokens"] += r.total_tokens
+            by_model[r.model]["total_cost"] += r.cost
+            by_model[r.model]["count"] += 1
+
+        return UsageSummary(
+            total_tokens=total_tokens,
+            total_cost=total_cost,
+            by_model=by_model,
+            records=filtered,
+        )
+
+
+# ---------------------------------------------------------------------------
+# RedisUsageStore
+# ---------------------------------------------------------------------------
+
+
+class RedisUsageStore:
+    """Redis-backed usage store using Hash per date for O(1) writes.
+
+    Key schema:
+        agentkit:usage:{YYYY-MM-DD}         → Hash: {agent:model → JSON(UsageBucket)}
+        agentkit:usage_records:{YYYY-MM-DD}  → List: JSON(UsageRecord) with LTRIM
+    """
+
+    USAGE_PREFIX = "agentkit:usage:"
+    RECORDS_PREFIX = "agentkit:usage_records:"
+    MAX_RECORDS_PER_DAY = 50000
+    TTL_DAYS = 90  # Auto-expire after 90 days
+
+    def __init__(self, redis_url: str = "redis://localhost:6379"):
+        self._redis_url = redis_url
+        self._redis: Any = None
+        self._sync_redis: Any = None
+        self._fallback: InMemoryUsageStore | None = None
+        self._degraded = False
+
+    async def _get_redis(self):
+        if self._redis is None:
+            import redis.asyncio as aioredis
+            self._redis = aioredis.from_url(self._redis_url, decode_responses=True)
+        return self._redis
+
+    def _get_sync_redis(self):
+        """Get or create a persistent sync Redis client (connection pool backed)."""
+        if self._sync_redis is None:
+            import redis as sync_redis
+            self._sync_redis = sync_redis.from_url(
+                self._redis_url, decode_responses=True
+            )
+        return self._sync_redis
+
+    async def aclose(self) -> None:
+        if self._redis is not None:
+            await self._redis.aclose()
+            self._redis = None
+        if self._sync_redis is not None:
+            self._sync_redis.aclose()
+            self._sync_redis = None
+
+    def _degrade_to_fallback(self) -> None:
+        if not self._degraded:
+            self._degraded = True
+            if self._fallback is None:
+                self._fallback = InMemoryUsageStore()
+            logger.warning("Redis usage store unreachable, degraded to in-memory")
+
+    def _today_key(self) -> str:
+        return datetime.now(timezone.utc).strftime("%Y-%m-%d")
+
+    def record(
+        self,
+        agent_name: str,
+        model: str,
+        usage: TokenUsage,
+        cost: float,
+        latency_ms: float,
+    ) -> None:
+        """Record usage — sync wrapper for async Redis.
+
+        Note: This is a sync method because UsageTracker.record() is sync.
+        For Redis, we use a sync Redis client for writes to avoid
+        needing an event loop in the caller.
+        """
+        if self._degraded and self._fallback is not None:
+            self._fallback.record(agent_name, model, usage, cost, latency_ms)
+            return
+
+        try:
+            r = self._get_sync_redis()
+
+            date_key = self._today_key()
+            hash_key = f"{self.USAGE_PREFIX}{date_key}"
+            list_key = f"{self.RECORDS_PREFIX}{date_key}"
+            bucket_field = f"{agent_name}:{model}"
+
+            # Atomic HINCRBYFLOAT for bucket aggregation
+            pipe = r.pipeline()
+            pipe.hincrbyfloat(hash_key, f"{bucket_field}:cost", cost)
+            pipe.hincrby(hash_key, f"{bucket_field}:prompt_tokens", usage.prompt_tokens)
+            pipe.hincrby(hash_key, f"{bucket_field}:completion_tokens", usage.completion_tokens)
+            pipe.hincrby(hash_key, f"{bucket_field}:total_tokens", usage.total_tokens)
+            pipe.hincrby(hash_key, f"{bucket_field}:count", 1)
+
+            # Append record
+            rec = UsageRecord(
+                agent_name=agent_name,
+                model=model,
+                prompt_tokens=usage.prompt_tokens,
+                completion_tokens=usage.completion_tokens,
+                total_tokens=usage.total_tokens,
+                cost=cost,
+                latency_ms=latency_ms,
+            )
+            pipe.rpush(list_key, json.dumps({
+                "agent_name": rec.agent_name,
+                "model": rec.model,
+                "prompt_tokens": rec.prompt_tokens,
+                "completion_tokens": rec.completion_tokens,
+                "total_tokens": rec.total_tokens,
+                "cost": rec.cost,
+                "latency_ms": rec.latency_ms,
+                "timestamp": rec.timestamp,
+            }))
+            pipe.ltrim(list_key, -self.MAX_RECORDS_PER_DAY, -1)
+
+            # Set TTL on first write of the day
+            pipe.expire(hash_key, self.TTL_DAYS * 86400)
+            pipe.expire(list_key, self.TTL_DAYS * 86400)
+
+            pipe.execute()
+        except Exception as e:
+            logger.warning(f"Redis usage record failed: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                self._fallback.record(agent_name, model, usage, cost, latency_ms)
+
+    def get_usage(
+        self,
+        agent_name: str | None = None,
+        start_time: datetime | None = None,
+        end_time: datetime | None = None,
+    ) -> UsageSummary:
+        """Query usage summary from Redis."""
+        if self._degraded and self._fallback is not None:
+            return self._fallback.get_usage(agent_name, start_time, end_time)
+
+        try:
+            r = self._get_sync_redis()
+
+            # Determine date range to scan
+            start = start_time or datetime(2020, 1, 1, tzinfo=timezone.utc)
+            end = end_time or datetime.now(timezone.utc)
+
+            all_records: list[UsageRecord] = []
+            # Scan date keys in range
+            current = start.date()
+            end_date = end.date()
+            while current <= end_date:
+                list_key = f"{self.RECORDS_PREFIX}{current.isoformat()}"
+                raw_records = r.lrange(list_key, 0, -1)
+                for raw in raw_records:
+                    data = json.loads(raw)
+                    rec = UsageRecord(**data)
+                    rec_ts = datetime.fromisoformat(rec.timestamp)
+                    if rec_ts >= start and rec_ts <= end:
+                        if agent_name is None or rec.agent_name == agent_name:
+                            all_records.append(rec)
+                current = current + timedelta(days=1)
+
+            if not all_records:
+                return UsageSummary()
+
+            total_tokens = sum(r.total_tokens for r in all_records)
+            total_cost = sum(r.cost for r in all_records)
+
+            by_model: dict[str, dict[str, int | float]] = {}
+            for r in all_records:
+                if r.model not in by_model:
+                    by_model[r.model] = {"total_tokens": 0, "total_cost": 0.0, "count": 0}
+                by_model[r.model]["total_tokens"] += r.total_tokens
+                by_model[r.model]["total_cost"] += r.cost
+                by_model[r.model]["count"] += 1
+
+            return UsageSummary(
+                total_tokens=total_tokens,
+                total_cost=total_cost,
+                by_model=by_model,
+                records=all_records,
+            )
+        except Exception as e:
+            logger.warning(f"Redis usage query failed: {e}")
+            if self._fallback is not None:
+                return self._fallback.get_usage(agent_name, start_time, end_time)
+            return UsageSummary()
+
+
+# ---------------------------------------------------------------------------
+# Factory
+# ---------------------------------------------------------------------------
+
+
+def create_usage_store(
+    backend: str = "auto",
+    redis_url: str = "redis://localhost:6379",
+) -> UsageStore:
+    """Create a usage store backend.
+
+    Args:
+        backend: "auto" (try Redis, fallback to memory), "redis", "memory".
+        redis_url: Redis connection URL.
+
+    Returns:
+        A UsageStore instance.
+    """
+    if backend in ("auto", "redis"):
+        try:
+            import redis  # noqa: F401
+            return RedisUsageStore(redis_url=redis_url)
+        except ImportError:
+            logger.warning("redis package not available, falling back to in-memory usage store")
+            return InMemoryUsageStore()
+    return InMemoryUsageStore()

src/agentkit/memory/profile.py

@@ -10,7 +10,7 @@ import re
 from dataclasses import dataclass, field
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import Any
+from typing import Any, Callable
 
 
 class MemoryFile:
@@ -26,9 +26,11 @@ class MemoryFile:
 
     """
 
-    def __init__(self, path: Path, char_budget: int | None = None):
+    def __init__(self, path: Path, char_budget: int | None = None,
+                 protected_sections: set[str] | None = None):
         self.path = Path(path)
         self.char_budget = char_budget
+        self._protected_sections = protected_sections or set()
 
     def read(self) -> str:
         """读取整个文件内容，文件不存在返回空字符串."""
@@ -37,11 +39,14 @@ class MemoryFile:
         return self.path.read_text(encoding="utf-8")
 
     def write(self, content: str) -> None:
-        """写入内容，自动创建父目录，超容量时自动裁剪."""
+        """写入内容，自动创建父目录，超容量时自动裁剪.
+
+        在内存中完成裁剪后一次性写入，避免中间不一致状态。
+        """
         self.path.parent.mkdir(parents=True, exist_ok=True)
-        self.path.write_text(content, encoding="utf-8")
         if self.char_budget and len(content) > self.char_budget:
-            self.trim_to_budget()
+            content = self._trim_content(content, self._protected_sections or None)
+        self.path.write_text(content, encoding="utf-8")
 
     def read_section(self, name: str) -> str:
         """读取指定 section 的内容（不含标题行）."""
@@ -104,15 +109,64 @@ class MemoryFile:
             return []
         return re.findall(r"^## (.+)$", content, re.MULTILINE)
 
-    def trim_to_budget(self) -> None:
-        """裁剪内容到容量上限，优先保留前面的 section."""
+    def trim_to_budget(self, protected_sections: set[str] | None = None) -> None:
+        """裁剪内容到容量上限，按 section 边界截断.
+
+        保持原始 section 顺序，仅从后向前移除非保护 section。
+        protected_sections 中的 section 始终保留，不参与裁剪。
+        """
         if not self.char_budget:
             return
         content = self.read()
         if len(content) <= self.char_budget:
             return
-        # 从末尾裁剪，保留前面的 section
-        self.write(content[: self.char_budget])
+        trimmed = self._trim_content(content, protected_sections)
+        self.path.write_text(trimmed, encoding="utf-8")
+
+    def _trim_content(self, content: str, protected_sections: set[str] | None = None) -> str:
+        """在内存中裁剪内容到容量上限，返回裁剪后的字符串（不写文件）.
+
+        保持原始 section 顺序，仅从后向前移除非保护 section。
+        """
+        if not self.char_budget or len(content) <= self.char_budget:
+            return content
+
+        protected = protected_sections or set()
+
+        # 解析所有 section 及其位置
+        sections: list[tuple[str, int, int]] = []  # (name, start, end)
+        for match in re.finditer(r"^## (.+)$", content, re.MULTILINE):
+            name = match.group(1).strip()
+            start = match.start()
+            next_match = re.search(r"^## ", content[match.end():], re.MULTILINE)
+            if next_match:
+                end = match.end() + next_match.start()
+            else:
+                end = len(content)
+            sections.append((name, start, end))
+
+        if not sections:
+            return content[:self.char_budget]
+
+        # 保持原始顺序，标记每个 section 是否受保护
+        ordered: list[tuple[str, str, bool]] = []  # (name, text, is_protected)
+        for name, start, end in sections:
+            ordered.append((name, content[start:end], name in protected))
+
+        # 从后向前移除非保护 section，直到总长度在预算内
+        while ordered:
+            total = len("\n\n".join(s[1] for s in ordered))
+            if total <= self.char_budget:
+                break
+            # 从后向前找第一个非保护 section 移除
+            for i in range(len(ordered) - 1, -1, -1):
+                if not ordered[i][2]:
+                    ordered.pop(i)
+                    break
+            else:
+                break  # 所有剩余 section 都是受保护的
+
+        return "\n\n".join(s[1] for s in ordered).strip()
 
 
 @dataclass
@@ -168,14 +222,21 @@ class MemoryStore:
 
     """
 
-    def __init__(self, base_dir: Path | str | None = None):
+    def __init__(self, base_dir: Path | str | None = None,
+                 on_change: Callable[[str], None] | None = None):
         if base_dir is None:
             base_dir = Path.home() / ".agentkit"
         self.base_dir = Path(base_dir)
         self.base_dir.mkdir(parents=True, exist_ok=True)
+        self._on_change = on_change
+        self._base_prompt: str = ""
 
         # 初始化四个 MemoryFile
-        self._soul = MemoryFile(self.base_dir / "SOUL.md", char_budget=SOUL_BUDGET)
+        self._soul = MemoryFile(
+            self.base_dir / "SOUL.md",
+            char_budget=SOUL_BUDGET,
+            protected_sections={"版本", "更新历史"},
+        )
         self._user = MemoryFile(self.base_dir / "memories" / "USER.md", char_budget=USER_BUDGET)
         self._memory = MemoryFile(self.base_dir / "memories" / "MEMORY.md", char_budget=MEMORY_BUDGET)
         self._daily_dir = self.base_dir / "memories" / "daily"
@@ -277,6 +338,10 @@ class MemoryStore:
 
             [base_prompt]
         """
+        # 保存 base_prompt 供后续刷新使用
+        if base_prompt:
+            self._base_prompt = base_prompt
+
         parts: list[str] = []
 
         if snapshot.soul:
@@ -292,3 +357,23 @@ class MemoryStore:
             parts.append(base_prompt)
 
         return "\n\n".join(parts) if parts else base_prompt
+
+    def refresh_system_prompt(self) -> str:
+        """重新加载所有记忆文件并构建 system prompt.
+
+        在 MemoryTool 写入记忆后调用，确保 agent 的 _system_prompt
+        反映最新的记忆内容。
+        """
+        snapshot = self.load_all()
+        return self.build_system_prompt(snapshot, self._base_prompt)
+
+    def notify_change(self) -> None:
+        """记忆文件变更后通知回调，刷新所有订阅者的 system prompt."""
+        if self._on_change is None:
+            return
+        try:
+            new_prompt = self.refresh_system_prompt()
+            self._on_change(new_prompt)
+        except Exception:
+            import logging
+            logging.getLogger(__name__).warning("Memory notify_change failed", exc_info=True)

src/agentkit/quality/cascade_detector.py

@@ -1,9 +1,14 @@
-"""CascadeDetector - 独立的级联故障检测工具"""
+"""CascadeDetector - 独立的级联故障检测工具（委托给 CascadeStateStore）"""
 
 from __future__ import annotations
 
 from dataclasses import dataclass
 
+from agentkit.quality.cascade_state_store import (
+    CascadeStateStore,
+    InMemoryCascadeStateStore,
+)
+
 
 @dataclass
 class CascadeAlert:
@@ -19,18 +24,19 @@ class CascadeAlert:
 class CascadeDetector:
     """检测多 agent 交互中的级联故障"""
 
-    def __init__(self, max_interactions: int = 10, max_depth: int = 3):
+    def __init__(
+        self,
+        max_interactions: int = 10,
+        max_depth: int = 3,
+        store: CascadeStateStore | None = None,
+    ):
         self._max_interactions = max_interactions
         self._max_depth = max_depth
-        self._interaction_counts: dict[str, int] = {}
-        self._loop_depths: dict[str, int] = {}
+        self._store: CascadeStateStore = store or InMemoryCascadeStateStore()
 
     def check_interaction(self, session_id: str) -> CascadeAlert | None:
         """递增并检查交互计数"""
-        self._interaction_counts[session_id] = (
-            self._interaction_counts.get(session_id, 0) + 1
-        )
-        count = self._interaction_counts[session_id]
+        count = self._store.increment_interaction(session_id)
         if count > self._max_interactions:
             return CascadeAlert(
                 session_id=session_id,
@@ -46,7 +52,7 @@ class CascadeDetector:
 
     def check_depth(self, session_id: str, depth: int) -> CascadeAlert | None:
         """检查循环深度"""
-        self._loop_depths[session_id] = depth
+        self._store.set_depth(session_id, depth)
         if depth > self._max_depth:
             return CascadeAlert(
                 session_id=session_id,
@@ -62,12 +68,11 @@ class CascadeDetector:
 
     def reset(self, session_id: str) -> None:
         """重置某个 session 的计数器"""
-        self._interaction_counts.pop(session_id, None)
-        self._loop_depths.pop(session_id, None)
+        self._store.reset(session_id)
 
     def get_stats(self, session_id: str) -> dict[str, int]:
         """获取某个 session 的当前统计"""
         return {
-            "interactions": self._interaction_counts.get(session_id, 0),
-            "depth": self._loop_depths.get(session_id, 0),
+            "interactions": self._store.get_interaction(session_id),
+            "depth": self._store.get_depth(session_id),
         }

src/agentkit/quality/cascade_state_store.py

@@ -0,0 +1,245 @@
+"""Cascade State Store — Persistent cascade detection state with Redis INCR backend.
+
+Provides CascadeStateStore Protocol with InMemoryCascadeStateStore and
+RedisCascadeStateStore backends. Enables CascadeDetector state to survive
+restarts and work across multiple instances.
+
+Key schema (Redis):
+    agentkit:cascade:interactions:{session_id}  → INCR counter with TTL
+    agentkit:cascade:depths:{session_id}        → SET counter with TTL
+"""
+
+import logging
+import time
+from typing import Any, Protocol, runtime_checkable
+
+logger = logging.getLogger(__name__)
+
+
+@runtime_checkable
+class CascadeStateStore(Protocol):
+    """Persistent cascade detection state interface."""
+
+    def increment_interaction(self, session_id: str) -> int:
+        """Atomically increment interaction count. Returns new count."""
+        ...
+
+    def get_interaction(self, session_id: str) -> int:
+        """Get current interaction count."""
+        ...
+
+    def set_depth(self, session_id: str, depth: int) -> None:
+        """Set loop depth for a session."""
+        ...
+
+    def get_depth(self, session_id: str) -> int:
+        """Get current loop depth."""
+        ...
+
+    def reset(self, session_id: str) -> None:
+        """Reset all counters for a session."""
+        ...
+
+
+# ---------------------------------------------------------------------------
+# InMemoryCascadeStateStore
+# ---------------------------------------------------------------------------
+
+
+class InMemoryCascadeStateStore:
+    """In-memory cascade state store (default, process-local).
+
+    Supports optional session TTL to prevent unbounded memory growth.
+    Expired entries are lazily cleaned up on access.
+    """
+
+    DEFAULT_SESSION_TTL = 86400  # 24 hours
+
+    def __init__(self, session_ttl: int = 86400):
+        self._session_ttl = session_ttl
+        self._interaction_counts: dict[str, int] = {}
+        self._loop_depths: dict[str, int] = {}
+        self._timestamps: dict[str, float] = {}
+
+    def _is_expired(self, session_id: str) -> bool:
+        ts = self._timestamps.get(session_id)
+        if ts is None:
+            return False
+        return (time.monotonic() - ts) > self._session_ttl
+
+    def _cleanup_expired(self) -> None:
+        """Lazy cleanup: remove expired sessions."""
+        expired = [sid for sid in self._timestamps if self._is_expired(sid)]
+        for sid in expired:
+            self._interaction_counts.pop(sid, None)
+            self._loop_depths.pop(sid, None)
+            self._timestamps.pop(sid, None)
+
+    def _touch(self, session_id: str) -> None:
+        self._timestamps[session_id] = time.monotonic()
+
+    def increment_interaction(self, session_id: str) -> int:
+        self._cleanup_expired()
+        self._interaction_counts[session_id] = self._interaction_counts.get(session_id, 0) + 1
+        self._touch(session_id)
+        return self._interaction_counts[session_id]
+
+    def get_interaction(self, session_id: str) -> int:
+        if self._is_expired(session_id):
+            self.reset(session_id)
+            return 0
+        return self._interaction_counts.get(session_id, 0)
+
+    def set_depth(self, session_id: str, depth: int) -> None:
+        self._touch(session_id)
+        self._loop_depths[session_id] = depth
+
+    def get_depth(self, session_id: str) -> int:
+        if self._is_expired(session_id):
+            self.reset(session_id)
+            return 0
+        return self._loop_depths.get(session_id, 0)
+
+    def reset(self, session_id: str) -> None:
+        self._interaction_counts.pop(session_id, None)
+        self._loop_depths.pop(session_id, None)
+        self._timestamps.pop(session_id, None)
+
+
+# ---------------------------------------------------------------------------
+# RedisCascadeStateStore
+# ---------------------------------------------------------------------------
+
+
+class RedisCascadeStateStore:
+    """Redis-backed cascade state store using INCR for atomic increments.
+
+    Key schema:
+        agentkit:cascade:interactions:{session_id}  → INCR counter with TTL
+        agentkit:cascade:depths:{session_id}         → SET counter with TTL
+    """
+
+    INTER_PREFIX = "agentkit:cascade:interactions:"
+    DEPTH_PREFIX = "agentkit:cascade:depths:"
+    SESSION_TTL = 86400  # 24 hours — sessions rarely last longer
+
+    def __init__(self, redis_url: str = "redis://localhost:6379", session_ttl: int = 86400):
+        self._redis_url = redis_url
+        self._session_ttl = session_ttl
+        self._sync_redis: Any = None
+        self._fallback: InMemoryCascadeStateStore | None = None
+        self._degraded = False
+
+    def _get_sync_redis(self):
+        """Get or create a persistent sync Redis client (connection pool backed)."""
+        if self._sync_redis is None:
+            import redis as sync_redis
+            self._sync_redis = sync_redis.from_url(
+                self._redis_url, decode_responses=True
+            )
+        return self._sync_redis
+
+    def _degrade_to_fallback(self) -> None:
+        if not self._degraded:
+            self._degraded = True
+            if self._fallback is None:
+                self._fallback = InMemoryCascadeStateStore()
+            logger.warning("Redis cascade store unreachable, degraded to in-memory")
+
+    def increment_interaction(self, session_id: str) -> int:
+        if self._degraded and self._fallback is not None:
+            return self._fallback.increment_interaction(session_id)
+        try:
+            r = self._get_sync_redis()
+            key = f"{self.INTER_PREFIX}{session_id}"
+            pipe = r.pipeline()
+            pipe.incr(key)
+            pipe.expire(key, self._session_ttl)
+            results = pipe.execute()
+            return results[0]
+        except Exception as e:
+            logger.warning(f"Redis cascade increment failed: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                return self._fallback.increment_interaction(session_id)
+            return 0
+
+    def get_interaction(self, session_id: str) -> int:
+        if self._degraded and self._fallback is not None:
+            return self._fallback.get_interaction(session_id)
+        try:
+            r = self._get_sync_redis()
+            val = r.get(f"{self.INTER_PREFIX}{session_id}")
+            return int(val) if val is not None else 0
+        except Exception as e:
+            logger.warning(f"Redis cascade get failed: {e}")
+            if self._fallback is not None:
+                return self._fallback.get_interaction(session_id)
+            return 0
+
+    def set_depth(self, session_id: str, depth: int) -> None:
+        if self._degraded and self._fallback is not None:
+            self._fallback.set_depth(session_id, depth)
+            return
+        try:
+            r = self._get_sync_redis()
+            key = f"{self.DEPTH_PREFIX}{session_id}"
+            pipe = r.pipeline()
+            pipe.set(key, depth)
+            pipe.expire(key, self._session_ttl)
+            pipe.execute()
+        except Exception as e:
+            logger.warning(f"Redis cascade set_depth failed: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                self._fallback.set_depth(session_id, depth)
+
+    def get_depth(self, session_id: str) -> int:
+        if self._degraded and self._fallback is not None:
+            return self._fallback.get_depth(session_id)
+        try:
+            r = self._get_sync_redis()
+            val = r.get(f"{self.DEPTH_PREFIX}{session_id}")
+            return int(val) if val is not None else 0
+        except Exception as e:
+            logger.warning(f"Redis cascade get_depth failed: {e}")
+            if self._fallback is not None:
+                return self._fallback.get_depth(session_id)
+            return 0
+
+    def reset(self, session_id: str) -> None:
+        if self._degraded and self._fallback is not None:
+            self._fallback.reset(session_id)
+            return
+        try:
+            r = self._get_sync_redis()
+            pipe = r.pipeline()
+            pipe.delete(f"{self.INTER_PREFIX}{session_id}")
+            pipe.delete(f"{self.DEPTH_PREFIX}{session_id}")
+            pipe.execute()
+        except Exception as e:
+            logger.warning(f"Redis cascade reset failed: {e}")
+            self._degrade_to_fallback()
+            if self._fallback is not None:
+                self._fallback.reset(session_id)
+
+
+# ---------------------------------------------------------------------------
+# Factory
+# ---------------------------------------------------------------------------
+
+
+def create_cascade_state_store(
+    backend: str = "auto",
+    redis_url: str = "redis://localhost:6379",
+    session_ttl: int = 86400,
+) -> CascadeStateStore:
+    """Create a cascade state store backend."""
+    if backend in ("auto", "redis"):
+        try:
+            import redis  # noqa: F401
+            return RedisCascadeStateStore(redis_url=redis_url, session_ttl=session_ttl)
+        except ImportError:
+            logger.warning("redis package not available, falling back to in-memory cascade store")
+            return InMemoryCascadeStateStore()
+    return InMemoryCascadeStateStore()

src/agentkit/server/app.py

@@ -40,7 +40,19 @@ _ALLOWED_ENV_EXACT = {'DATABASE_URL', 'REDIS_URL'}
 
 def _build_llm_gateway(config: ServerConfig) -> LLMGateway:
     """Build LLMGateway from ServerConfig, registering all providers."""
-    gateway = LLMGateway(config=config.llm_config)
+    # Initialize UsageStore if configured
+    usage_store = None
+    if config.usage_store:
+        try:
+            from agentkit.llm.providers.usage_store import create_usage_store
+            usage_store = create_usage_store(
+                backend=config.usage_store.get("backend", "memory"),
+                redis_url=config.usage_store.get("redis_url", "redis://localhost:6379"),
+            )
+        except Exception as e:
+            logger.warning(f"Failed to initialize usage store: {e}, using in-memory")
+
+    gateway = LLMGateway(config=config.llm_config, usage_store=usage_store)
 
     for name, pconf in config.llm_config.providers.items():
         if not pconf.api_key:
@@ -111,6 +123,15 @@ async def lifespan(app: FastAPI):
 
     # Start MCP servers if configured
     mcp_manager = getattr(app.state, "mcp_manager", None)
+
+    # Build semantic router index after skill registry is populated
+    semantic_router = getattr(getattr(app.state, "cost_aware_router", None), "_semantic_router", None)
+    if semantic_router is not None:
+        try:
+            await semantic_router.build_index(app.state.skill_registry)
+            logger.info(f"Semantic router index built with {len(app.state.skill_registry.list_skills())} skills")
+        except Exception as e:
+            logger.warning(f"Failed to build semantic router index: {e}")
     if mcp_manager is not None:
         await mcp_manager.start_all()
 
@@ -142,6 +163,23 @@ async def lifespan(app: FastAPI):
         )
         effective_system_prompt = memory_store.build_system_prompt(memory_snapshot, base_prompt)
 
+        # Register on_change callback to refresh all agents' system prompts
+        # when MemoryTool writes to memory files
+        def _on_memory_change(new_prompt: str) -> None:
+            pool = app.state.agent_pool
+            updated = 0
+            for agent_name in pool.list_agents():
+                try:
+                    agent = pool.get_agent(agent_name)
+                    if agent is not None:
+                        agent._system_prompt = new_prompt
+                        updated += 1
+                except Exception:
+                    logger.warning(f"Failed to update system prompt for agent '{agent_name}'", exc_info=True)
+            logger.info(f"Memory changed: refreshed system prompt for {updated}/{len(pool.list_agents())} agents")
+
+        memory_store._on_change = _on_memory_change
+
         # Store memory_store on app.state for chat routes to use
         app.state.memory_store = memory_store
 
@@ -219,6 +257,34 @@ async def lifespan(app: FastAPI):
             from agentkit.memory.profile import MemoryStore
             memory_store = MemoryStore()
             memory_store.ensure_defaults()
+            # Initialize _base_prompt so refresh_system_prompt works correctly
+            snapshot = memory_store.load_all()
+            base_prompt = (
+                "你是一个有帮助的AI助手。请记住我们对话的上下文，并在后续对话中引用之前的内容。回答要清晰简洁，请使用中文回复。\n\n"
+                "重要提示：当你不确定事实信息、时事新闻或任何你不确信的话题时，"
+                "你必须先使用搜索工具查找准确和最新的信息，然后再回答。"
+                "中文内容优先使用 baidu_search 工具，英文/国际内容使用 web_search。"
+                "在能够搜索到真相的情况下，绝不猜测或编造答案。"
+                "始终优先搜索而不是给出可能不正确的信息。\n\n"
+                "技能安装：当需要安装技能时，使用 skill_install 工具，不要用 shell 执行 npm install。"
+                "skill_install 的 source 参数格式为 owner/repo@skill，例如 vercel-labs/skills@find-skills。"
+                "如果不知道完整 source，先用 shell 执行 `npx skills search <name>` 搜索。"
+            )
+            memory_store.build_system_prompt(snapshot, base_prompt)
+            # Register on_change callback for existing agents
+            def _on_memory_change(new_prompt: str) -> None:
+                pool = app.state.agent_pool
+                updated = 0
+                for agent_name in pool.list_agents():
+                    try:
+                        agent = pool.get_agent(agent_name)
+                        if agent is not None:
+                            agent._system_prompt = new_prompt
+                            updated += 1
+                    except Exception:
+                        logger.warning(f"Failed to update system prompt for agent '{agent_name}'", exc_info=True)
+                logger.info(f"Memory changed: refreshed system prompt for {updated}/{len(pool.list_agents())} agents")
+            memory_store._on_change = _on_memory_change
             app.state.memory_store = memory_store
 
     yield
@@ -502,12 +568,28 @@ def create_app(
     auction_enabled = False
     if server_config and hasattr(server_config, "marketplace") and server_config.marketplace:
         auction_enabled = server_config.marketplace.get("auction_enabled", False)
+
+    # Initialize semantic router if configured
+    semantic_router = None
+    router_conf = server_config.router if server_config and server_config.router else {}
+    if router_conf.get("semantic", {}).get("enabled"):
+        try:
+            from agentkit.chat.semantic_router import SemanticRouter
+            semantic_router = SemanticRouter(
+                embedder=app.state.llm_gateway._embedder,
+                similarity_high=router_conf["semantic"].get("similarity_high", 0.85),
+                similarity_low=router_conf["semantic"].get("similarity_low", 0.6),
+            )
+        except Exception as e:
+            logger.warning(f"Failed to initialize semantic router: {e}")
+
     cost_aware_router = CostAwareRouter(
         llm_gateway=app.state.llm_gateway,
         org_context=org_context,
         auction_enabled=auction_enabled,
-        classifier=server_config.router.get("classifier", "heuristic") if server_config and server_config.router else "heuristic",
-        merged_llm_classify=server_config.router.get("merged_llm_classify", True) if server_config and server_config.router else True,
+        classifier=router_conf.get("classifier", "heuristic"),
+        merged_llm_classify=router_conf.get("merged_llm_classify", True),
+        semantic_router=semantic_router,
     )
     app.state.cost_aware_router = cost_aware_router
     # Initialize task store from config
@@ -555,14 +637,30 @@ def create_app(
             app.state.evolution_store = create_evolution_store(
                 backend=evo_conf.get("backend", "memory"),
                 db_path=evo_conf.get("db_path", "~/.agentkit/evolution.db"),
+                database_url=evo_conf.get("database_url"),
             )
         except Exception as e:
-            import logging
-            logging.getLogger(__name__).warning(f"Failed to initialize evolution store: {e}")
+            logger.warning(f"Failed to initialize evolution store: {e}")
             app.state.evolution_store = None
     else:
         app.state.evolution_store = None
 
+    # Initialize cascade state store if configured
+    if server_config and hasattr(server_config, 'cascade_store') and server_config.cascade_store:
+        try:
+            from agentkit.quality.cascade_state_store import create_cascade_state_store
+            cs_conf = server_config.cascade_store
+            app.state.cascade_state_store = create_cascade_state_store(
+                backend=cs_conf.get("backend", "memory"),
+                redis_url=cs_conf.get("redis_url", "redis://localhost:6379"),
+                session_ttl=cs_conf.get("session_ttl", 86400),
+            )
+        except Exception as e:
+            logger.warning(f"Failed to initialize cascade state store: {e}")
+            app.state.cascade_state_store = None
+    else:
+        app.state.cascade_state_store = None
+
     # Initialize memory components if configured
     if server_config and hasattr(server_config, 'memory') and server_config.memory:
         try:

src/agentkit/server/config.py

@@ -111,6 +111,9 @@ class ServerConfig:
         marketplace: dict[str, Any] | None = None,
         alignment: dict[str, Any] | None = None,
         router: dict[str, Any] | None = None,
+        usage_store: dict[str, Any] | None = None,
+        cascade_store: dict[str, Any] | None = None,
+        evolution: dict[str, Any] | None = None,
         on_change: Callable[["ServerConfig"], None] | None = None,
     ):
         self.host = host
@@ -134,6 +137,9 @@ class ServerConfig:
         self.marketplace = marketplace or {}
         self.alignment = alignment or {}
         self.router = router or {}
+        self.usage_store = usage_store or {}
+        self.cascade_store = cascade_store or {}
+        self.evolution = evolution or {}
         self.on_change = on_change
 
         # Config watching state
@@ -201,6 +207,15 @@ class ServerConfig:
         # Router config
         router_data = data.get("router", {})
 
+        # Usage store config
+        usage_store_data = data.get("usage_store", {})
+
+        # Cascade store config
+        cascade_store_data = data.get("cascade_store", {})
+
+        # Evolution store config
+        evolution_data = data.get("evolution", {})
+
         return cls(
             host=server.get("host", "0.0.0.0"),
             port=server.get("port", 8001),
@@ -223,11 +238,16 @@ class ServerConfig:
             marketplace=marketplace_data,
             alignment=alignment_data,
             router=router_data,
+            usage_store=usage_store_data,
+            cascade_store=cascade_store_data,
+            evolution=evolution_data,
         )
 
     @staticmethod
     def _build_llm_config(data: dict) -> LLMConfig:
         """Build LLMConfig from the llm section of agentkit.yaml."""
+        from agentkit.llm.config import CacheConfig
+
         providers = {}
         model_aliases = {}
 
@@ -254,10 +274,17 @@ class ServerConfig:
                 keepalive_expiry=pconf.get("keepalive_expiry", 30.0),
             )
 
+        # Build CacheConfig if cache section is present
+        cache_config = None
+        cache_data = data.get("cache")
+        if cache_data and isinstance(cache_data, dict):
+            cache_config = CacheConfig.from_dict(cache_data)
+
         return LLMConfig(
             providers=providers,
             model_aliases=model_aliases,
             fallbacks=data.get("fallbacks", {}),
+            cache=cache_config,
         )
 
     @staticmethod

tests/integration/test_p0_hardening.py

@@ -0,0 +1,422 @@
+"""P0 Production Hardening — End-to-End Integration Tests
+
+Verifies the full configuration wiring and feature integration:
+- Config from YAML → all features configured correctly
+- Cache + usage tracking: cached requests show 0 cost
+- UsageStore persistence via config
+- CascadeStateStore persistence via config
+- EvolutionStore via config
+- Semantic router via config
+- Graceful degradation when backends unavailable
+"""
+
+import os
+import tempfile
+
+import pytest
+
+from agentkit.core.protocol import EvolutionEvent
+from agentkit.llm.config import CacheConfig, LLMConfig, ProviderConfig
+from agentkit.llm.gateway import LLMGateway
+from agentkit.llm.protocol import TokenUsage
+from agentkit.llm.providers.usage_store import (
+    InMemoryUsageStore,
+    create_usage_store,
+)
+from agentkit.quality.cascade_detector import CascadeDetector
+from agentkit.quality.cascade_state_store import (
+    InMemoryCascadeStateStore,
+    create_cascade_state_store,
+)
+from agentkit.evolution.evolution_store import (
+    EvolutionStoreProtocol,
+    InMemoryEvolutionStore,
+    PersistentEvolutionStore,
+    create_evolution_store,
+)
+from agentkit.server.config import ServerConfig
+
+
+# ── Config parsing tests ──────────────────────────────────
+
+
+class TestConfigParsing:
+    """Verify ServerConfig correctly parses all new config sections."""
+
+    def test_usage_store_config_parsed(self):
+        data = {
+            "usage_store": {
+                "backend": "redis",
+                "redis_url": "redis://custom:6380",
+            }
+        }
+        config = ServerConfig.from_dict(data)
+        assert config.usage_store["backend"] == "redis"
+        assert config.usage_store["redis_url"] == "redis://custom:6380"
+
+    def test_cascade_store_config_parsed(self):
+        data = {
+            "cascade_store": {
+                "backend": "redis",
+                "redis_url": "redis://custom:6380",
+                "session_ttl": 3600,
+            }
+        }
+        config = ServerConfig.from_dict(data)
+        assert config.cascade_store["backend"] == "redis"
+        assert config.cascade_store["session_ttl"] == 3600
+
+    def test_evolution_config_parsed(self):
+        data = {
+            "evolution": {
+                "backend": "sqlite",
+                "db_path": "/tmp/test.db",
+            }
+        }
+        config = ServerConfig.from_dict(data)
+        assert config.evolution["backend"] == "sqlite"
+        assert config.evolution["db_path"] == "/tmp/test.db"
+
+    def test_llm_cache_config_parsed(self):
+        data = {
+            "llm": {
+                "providers": {},
+                "cache": {
+                    "enabled": True,
+                    "backend": "memory",
+                    "exact_ttl": 7200,
+                },
+            }
+        }
+        config = ServerConfig.from_dict(data)
+        assert config.llm_config.cache is not None
+        assert config.llm_config.cache.enabled is True
+        assert config.llm_config.cache.backend == "memory"
+        assert config.llm_config.cache.exact_ttl == 7200
+
+    def test_router_semantic_config_parsed(self):
+        data = {
+            "router": {
+                "classifier": "heuristic",
+                "semantic": {
+                    "enabled": True,
+                    "similarity_high": 0.9,
+                    "similarity_low": 0.5,
+                },
+            }
+        }
+        config = ServerConfig.from_dict(data)
+        assert config.router["semantic"]["enabled"] is True
+        assert config.router["semantic"]["similarity_high"] == 0.9
+
+    def test_empty_config_defaults(self):
+        config = ServerConfig.from_dict({})
+        assert config.usage_store == {}
+        assert config.cascade_store == {}
+        assert config.evolution == {}
+        assert config.llm_config.cache is None
+
+    def test_config_from_yaml_roundtrip(self, tmp_path):
+        """Config can be loaded from a YAML file with all new sections."""
+        yaml_content = """
+server:
+  host: 0.0.0.0
+  port: 8001
+llm:
+  providers: {}
+  cache:
+    enabled: true
+    backend: memory
+router:
+  classifier: heuristic
+  semantic:
+    enabled: false
+usage_store:
+  backend: memory
+cascade_store:
+  backend: memory
+evolution:
+  backend: memory
+"""
+        yaml_path = str(tmp_path / "test_config.yaml")
+        with open(yaml_path, "w") as f:
+            f.write(yaml_content)
+
+        config = ServerConfig.from_yaml(yaml_path)
+        assert config.llm_config.cache is not None
+        assert config.llm_config.cache.enabled is True
+        assert config.usage_store["backend"] == "memory"
+        assert config.cascade_store["backend"] == "memory"
+        assert config.evolution["backend"] == "memory"
+
+
+# ── UsageStore integration tests ───────────────────────────
+
+
+class TestUsageStoreIntegration:
+    """Verify UsageStore works with LLMGateway."""
+
+    async def test_gateway_with_usage_store(self):
+        """LLMGateway uses injected UsageStore for tracking."""
+        store = InMemoryUsageStore()
+        gateway = LLMGateway(usage_store=store)
+
+        # Record usage directly through the tracker
+        gateway._usage_tracker.record(
+            agent_name="test_agent",
+            model="gpt-4",
+            usage=TokenUsage(prompt_tokens=100, completion_tokens=50),
+            cost=0.01,
+            latency_ms=100.0,
+        )
+
+        usage = gateway.get_usage()
+        assert usage.total_tokens > 0
+        assert usage.total_cost > 0
+
+    async def test_gateway_without_usage_store(self):
+        """LLMGateway works without explicit UsageStore (uses InMemory)."""
+        gateway = LLMGateway()
+        gateway._usage_tracker.record(
+            agent_name="test_agent",
+            model="gpt-4",
+            usage=TokenUsage(prompt_tokens=100, completion_tokens=50),
+            cost=0.01,
+            latency_ms=100.0,
+        )
+
+        usage = gateway.get_usage()
+        assert usage.total_tokens > 0
+
+    def test_create_usage_store_memory(self):
+        store = create_usage_store(backend="memory")
+        assert isinstance(store, InMemoryUsageStore)
+
+    def test_create_usage_store_redis_lazy(self):
+        """Redis backend creates RedisUsageStore (lazy connection, degrades on first op)."""
+        from agentkit.llm.providers.usage_store import RedisUsageStore
+
+        store = create_usage_store(
+            backend="redis",
+            redis_url="redis://nonexistent:6379",
+        )
+        # RedisUsageStore is created (lazy connection), degrades on first operation
+        assert isinstance(store, RedisUsageStore)
+
+
+# ── CascadeStateStore integration tests ────────────────────
+
+
+class TestCascadeStateStoreIntegration:
+    """Verify CascadeStateStore works with CascadeDetector."""
+
+    async def test_cascade_detector_with_store(self):
+        """CascadeDetector uses injected CascadeStateStore."""
+        store = InMemoryCascadeStateStore()
+        detector = CascadeDetector(store=store)
+
+        # Check interaction — should not trigger cascade
+        result = detector.check_interaction(session_id="test-session")
+        assert result is None  # No cascade alert
+
+    async def test_cascade_detector_without_store(self):
+        """CascadeDetector works without explicit store (uses InMemory)."""
+        detector = CascadeDetector()
+        result = detector.check_interaction(session_id="test-session")
+        assert result is None
+
+    def test_create_cascade_state_store_memory(self):
+        store = create_cascade_state_store(backend="memory")
+        assert isinstance(store, InMemoryCascadeStateStore)
+
+
+# ── EvolutionStore integration tests ───────────────────────
+
+
+class TestEvolutionStoreIntegration:
+    """Verify EvolutionStore creation from config."""
+
+    async def test_create_evolution_store_from_config(self, tmp_path):
+        """EvolutionStore created from config dict works correctly."""
+        db_path = str(tmp_path / "evo_test.db")
+        store = create_evolution_store(backend="sqlite", db_path=db_path)
+        assert isinstance(store, PersistentEvolutionStore)
+
+        event = EvolutionEvent(
+            agent_name="test_agent",
+            change_type="prompt",
+            before={"old": 1},
+            after={"new": 2},
+        )
+        event_id = await store.record(event)
+        assert event_id is not None
+
+        events = await store.list_events()
+        assert len(events) == 1
+        assert events[0]["agent_name"] == "test_agent"
+
+    async def test_create_evolution_store_memory(self):
+        store = create_evolution_store(backend="memory")
+        assert isinstance(store, InMemoryEvolutionStore)
+
+        event = EvolutionEvent(
+            agent_name="test_agent",
+            change_type="prompt",
+            before={},
+            after={},
+        )
+        event_id = await store.record(event)
+        assert event_id is not None
+
+    async def test_evolution_store_protocol_compliance(self):
+        """All created stores satisfy EvolutionStoreProtocol."""
+        memory_store = create_evolution_store(backend="memory")
+        assert isinstance(memory_store, EvolutionStoreProtocol)
+
+
+# ── Cache integration tests ────────────────────────────────
+
+
+class TestCacheIntegration:
+    """Verify LLMCache integration with LLMGateway via config."""
+
+    def test_gateway_with_cache_config(self):
+        """LLMGateway initializes cache when CacheConfig is provided."""
+        config = LLMConfig(
+            providers={},
+            cache=CacheConfig(enabled=True, backend="memory"),
+        )
+        gateway = LLMGateway(config=config)
+        assert gateway._cache is not None
+
+    def test_gateway_without_cache_config(self):
+        """LLMGateway works without cache (default)."""
+        config = LLMConfig(providers={})
+        gateway = LLMGateway(config=config)
+        assert gateway._cache is None
+
+    def test_gateway_cache_disabled(self):
+        """LLMGateway does not initialize cache when disabled."""
+        config = LLMConfig(
+            providers={},
+            cache=CacheConfig(enabled=False),
+        )
+        gateway = LLMGateway(config=config)
+        assert gateway._cache is None
+
+
+# ── Graceful degradation tests ─────────────────────────────
+
+
+class TestGracefulDegradation:
+    """Verify all features degrade gracefully when backends unavailable."""
+
+    def test_usage_store_auto_creates_redis(self):
+        """auto backend creates RedisUsageStore (lazy connection)."""
+        from agentkit.llm.providers.usage_store import RedisUsageStore
+
+        store = create_usage_store(
+            backend="auto",
+            redis_url="redis://nonexistent:6379",
+        )
+        # Redis is available as package, so RedisUsageStore is created
+        assert isinstance(store, RedisUsageStore)
+
+    def test_cascade_store_redis_lazy(self):
+        """CascadeStateStore Redis backend creates instance (lazy connection)."""
+        from agentkit.quality.cascade_state_store import RedisCascadeStateStore
+
+        store = create_cascade_state_store(
+            backend="redis",
+            redis_url="redis://nonexistent:6379",
+        )
+        assert isinstance(store, RedisCascadeStateStore)
+
+    def test_evolution_store_postgresql_unavailable(self):
+        """EvolutionStore falls back to InMemory when PG unavailable."""
+        store = create_evolution_store(
+            backend="postgresql",
+            database_url=None,
+        )
+        assert isinstance(store, InMemoryEvolutionStore)
+
+    def test_cache_auto_creates_redis(self):
+        """LLMCache auto backend creates RedisLLMCache (lazy connection)."""
+        from agentkit.llm.cache import create_llm_cache, RedisLLMCache
+
+        cache = create_llm_cache(
+            backend="auto",
+            redis_url="redis://nonexistent:6379",
+        )
+        # Redis package is available, so RedisLLMCache is created (lazy connection)
+        assert isinstance(cache, RedisLLMCache)
+
+
+# ── Full flow test (in-memory) ─────────────────────────────
+
+
+class TestFullFlowInMemory:
+    """End-to-end flow test using in-memory backends (no external deps)."""
+
+    async def test_config_to_components(self):
+        """ServerConfig → all components initialized correctly."""
+        data = {
+            "llm": {
+                "providers": {},
+                "cache": {
+                    "enabled": True,
+                    "backend": "memory",
+                },
+            },
+            "router": {
+                "classifier": "heuristic",
+            },
+            "usage_store": {"backend": "memory"},
+            "cascade_store": {"backend": "memory"},
+            "evolution": {"backend": "memory"},
+        }
+        config = ServerConfig.from_dict(data)
+
+        # Verify config parsed correctly
+        assert config.llm_config.cache is not None
+        assert config.llm_config.cache.enabled is True
+        assert config.usage_store["backend"] == "memory"
+        assert config.cascade_store["backend"] == "memory"
+        assert config.evolution["backend"] == "memory"
+
+        # Create components from config
+        usage_store = create_usage_store(
+            backend=config.usage_store.get("backend", "memory"),
+            redis_url=config.usage_store.get("redis_url", "redis://localhost:6379"),
+        )
+        gateway = LLMGateway(config=config.llm_config, usage_store=usage_store)
+        assert gateway._cache is not None
+
+        cascade_store = create_cascade_state_store(
+            backend=config.cascade_store.get("backend", "memory"),
+        )
+        detector = CascadeDetector(store=cascade_store)
+
+        evo_store = create_evolution_store(
+            backend=config.evolution.get("backend", "memory"),
+        )
+
+        # Exercise the components
+        gateway._usage_tracker.record(
+            agent_name="test",
+            model="gpt-4",
+            usage=TokenUsage(prompt_tokens=100, completion_tokens=50),
+            cost=0.01,
+            latency_ms=100.0,
+        )
+        usage = gateway.get_usage()
+        assert usage.total_tokens == 150
+
+        result = detector.check_interaction("s1")
+        assert result is None  # No cascade alert
+
+        event = EvolutionEvent(
+            agent_name="test", change_type="prompt", before={}, after={}
+        )
+        event_id = await evo_store.record(event)
+        assert event_id is not None

tests/unit/test_gateway_cache.py

@@ -0,0 +1,162 @@
+"""Integration tests for LLM Cache integration into LLMGateway (U2)."""
+
+import pytest
+
+from agentkit.llm.cache import InMemoryLLMCache
+from agentkit.llm.config import CacheConfig, LLMConfig
+from agentkit.llm.gateway import LLMGateway
+from agentkit.llm.protocol import LLMProvider, LLMRequest, LLMResponse, TokenUsage
+
+
+class MockProvider(LLMProvider):
+    """Mock LLM provider that tracks call count."""
+
+    def __init__(self, response_content: str = "Mock response"):
+        self.call_count = 0
+        self._response_content = response_content
+
+    async def chat(self, request: LLMRequest) -> LLMResponse:
+        self.call_count += 1
+        return LLMResponse(
+            content=self._response_content,
+            model=request.model,
+            usage=TokenUsage(prompt_tokens=10, completion_tokens=20),
+        )
+
+
+def _make_messages(user_content: str = "Hello") -> list[dict[str, str]]:
+    return [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": user_content},
+    ]
+
+
+class TestCacheDisabled:
+    @pytest.mark.asyncio
+    async def test_no_cache_by_default(self):
+        """Cache is disabled by default — requests always hit provider."""
+        gateway = LLMGateway()
+        provider = MockProvider()
+        gateway.register_provider("test", provider)
+
+        msgs = _make_messages()
+        await gateway.chat(msgs, "test/model")
+        await gateway.chat(msgs, "test/model")
+
+        assert provider.call_count == 2
+
+
+class TestCacheEnabled:
+    @pytest.mark.asyncio
+    async def test_first_request_is_miss(self):
+        """First request is a cache miss — provider is called."""
+        config = LLMConfig(cache=CacheConfig(enabled=True, backend="memory"))
+        gateway = LLMGateway(config=config)
+        provider = MockProvider()
+        gateway.register_provider("test", provider)
+
+        msgs = _make_messages()
+        response = await gateway.chat(msgs, "test/model", temperature=0.0)
+
+        assert provider.call_count == 1
+        assert response.content == "Mock response"
+
+    @pytest.mark.asyncio
+    async def test_second_request_is_hit(self):
+        """Second identical request is a cache hit — provider NOT called."""
+        config = LLMConfig(cache=CacheConfig(enabled=True, backend="memory"))
+        gateway = LLMGateway(config=config)
+        provider = MockProvider()
+        gateway.register_provider("test", provider)
+
+        msgs = _make_messages()
+        await gateway.chat(msgs, "test/model", temperature=0.0)
+        response = await gateway.chat(msgs, "test/model", temperature=0.0)
+
+        assert provider.call_count == 1  # Not called again
+        assert response.content == "Mock response"
+
+    @pytest.mark.asyncio
+    async def test_cache_hit_usage_has_zero_cost(self):
+        """Cache hit records usage with cost=0."""
+        config = LLMConfig(cache=CacheConfig(enabled=True, backend="memory"))
+        gateway = LLMGateway(config=config)
+        provider = MockProvider()
+        gateway.register_provider("test", provider)
+
+        msgs = _make_messages()
+        await gateway.chat(msgs, "test/model", agent_name="agent1", temperature=0.0)
+        await gateway.chat(msgs, "test/model", agent_name="agent1", temperature=0.0)
+
+        usage = gateway.get_usage(agent_name="agent1")
+        # First request has cost, second (cache hit) has cost=0
+        assert usage.total_cost == 0.0  # No cost config, so both are 0
+        assert len(usage.records) == 2
+
+    @pytest.mark.asyncio
+    async def test_different_messages_are_miss(self):
+        """Different messages produce cache misses."""
+        config = LLMConfig(cache=CacheConfig(enabled=True, backend="memory"))
+        gateway = LLMGateway(config=config)
+        provider = MockProvider()
+        gateway.register_provider("test", provider)
+
+        await gateway.chat(_make_messages("Hello"), "test/model", temperature=0.0)
+        await gateway.chat(_make_messages("World"), "test/model", temperature=0.0)
+
+        assert provider.call_count == 2
+
+
+class TestCacheConfig:
+    def test_config_from_dict(self):
+        """CacheConfig can be loaded from dict."""
+        config = LLMConfig.from_dict({
+            "cache": {
+                "enabled": True,
+                "backend": "memory",
+                "exact_ttl": 7200,
+            }
+        })
+        assert config.cache is not None
+        assert config.cache.enabled is True
+        assert config.cache.backend == "memory"
+        assert config.cache.exact_ttl == 7200
+
+    def test_config_from_dict_no_cache(self):
+        """No cache section in config → cache is None."""
+        config = LLMConfig.from_dict({})
+        assert config.cache is None
+
+    def test_config_from_dict_embedding(self):
+        """Embedding config is loaded correctly."""
+        config = LLMConfig.from_dict({
+            "cache": {
+                "enabled": True,
+                "embedding": {
+                    "provider": "xinference",
+                    "model": "bge-m3",
+                    "base_url": "http://localhost:9997/v1",
+                },
+            }
+        })
+        assert config.cache.embedding_provider == "xinference"
+        assert config.cache.embedding_model == "bge-m3"
+        assert config.cache.embedding_base_url == "http://localhost:9997/v1"
+
+    def test_gateway_creates_cache_when_enabled(self):
+        """Gateway creates cache instance when cache.enabled=True."""
+        config = LLMConfig(cache=CacheConfig(enabled=True, backend="memory"))
+        gateway = LLMGateway(config=config)
+        assert gateway._cache is not None
+        assert isinstance(gateway._cache, InMemoryLLMCache)
+
+    def test_gateway_no_cache_when_disabled(self):
+        """Gateway has no cache when cache is disabled."""
+        config = LLMConfig(cache=CacheConfig(enabled=False))
+        gateway = LLMGateway(config=config)
+        assert gateway._cache is None
+
+    def test_gateway_no_cache_when_no_config(self):
+        """Gateway has no cache when cache config is absent."""
+        gateway = LLMGateway()
+        assert gateway._cache is None

tests/unit/test_llm_cache.py

@@ -0,0 +1,604 @@
+"""Unit tests for LLM Cache Core (U1).
+
+Tests cover:
+- CacheKey generation (deterministic, component isolation)
+- InMemoryLLMCache (exact match, semantic match, TTL, LRU, stats)
+- RedisLLMCache (same tests with mocked Redis)
+- Factory function (backend selection, fallback)
+"""
+
+import json
+import time
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from agentkit.llm.cache import (
+    CacheEntry,
+    CacheResult,
+    InMemoryLLMCache,
+    RedisLLMCache,
+    create_llm_cache,
+    _serialize_response,
+    _deserialize_response,
+    _serialize_entry,
+    _deserialize_entry,
+)
+from agentkit.llm.cache_key import generate_cache_key
+from agentkit.llm.protocol import LLMResponse, TokenUsage, ToolCall
+
+
+# ---------------------------------------------------------------------------
+# Fixtures
+# ---------------------------------------------------------------------------
+
+
+def _make_response(
+    content: str = "Hello",
+    model: str = "gpt-4o",
+    prompt_tokens: int = 10,
+    completion_tokens: int = 20,
+    tool_calls: list[ToolCall] | None = None,
+) -> LLMResponse:
+    return LLMResponse(
+        content=content,
+        model=model,
+        usage=TokenUsage(
+            prompt_tokens=prompt_tokens,
+            completion_tokens=completion_tokens,
+        ),
+        tool_calls=tool_calls or [],
+        latency_ms=100.0,
+    )
+
+
+def _make_messages(user_content: str = "Hello") -> list[dict[str, str]]:
+    return [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": user_content},
+    ]
+
+
+def _make_embedding(base_val: float = 1.0, dim: int = 128) -> list[float]:
+    """Create a unit vector for similarity testing."""
+    vec = [base_val] * dim
+    magnitude = sum(x**2 for x in vec) ** 0.5
+    return [x / magnitude for x in vec] if magnitude > 0 else vec
+
+
+def _make_similar_embedding(base: list[float], noise: float = 0.01) -> list[float]:
+    """Create a vector similar to base with small noise."""
+    vec = [x + noise for x in base]
+    magnitude = sum(x**2 for x in vec) ** 0.5
+    return [x / magnitude for x in vec] if magnitude > 0 else vec
+
+
+def _make_different_embedding(dim: int = 128) -> list[float]:
+    """Create a vector with very low cosine similarity to _make_embedding()."""
+    # _make_embedding(1.0) is all-positive unit vector.
+    # Negate first half to create near-orthogonal vector.
+    half = dim // 2
+    vec = [-1.0] * half + [1.0] * (dim - half)
+    magnitude = sum(x**2 for x in vec) ** 0.5
+    return [x / magnitude for x in vec] if magnitude > 0 else vec
+
+
+# ---------------------------------------------------------------------------
+# CacheKey Tests
+# ---------------------------------------------------------------------------
+
+
+class TestCacheKey:
+    def test_deterministic(self):
+        """Same inputs produce same key."""
+        msgs = _make_messages()
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0)
+        key2 = generate_cache_key("gpt-4o", msgs, 0.0)
+        assert key1 == key2
+        assert len(key1) == 64  # SHA-256 hex
+
+    def test_different_model(self):
+        """Different model produces different key."""
+        msgs = _make_messages()
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0)
+        key2 = generate_cache_key("gpt-3.5-turbo", msgs, 0.0)
+        assert key1 != key2
+
+    def test_different_temperature(self):
+        """Different temperature produces different key."""
+        msgs = _make_messages()
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0)
+        key2 = generate_cache_key("gpt-4o", msgs, 0.7)
+        assert key1 != key2
+
+    def test_different_messages(self):
+        """Different messages produce different key."""
+        key1 = generate_cache_key("gpt-4o", _make_messages("Hello"), 0.0)
+        key2 = generate_cache_key("gpt-4o", _make_messages("World"), 0.0)
+        assert key1 != key2
+
+    def test_different_tools(self):
+        """Different tools produce different key."""
+        msgs = _make_messages()
+        tools1 = [{"type": "function", "function": {"name": "f1"}}]
+        tools2 = [{"type": "function", "function": {"name": "f2"}}]
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0, tools=tools1)
+        key2 = generate_cache_key("gpt-4o", msgs, 0.0, tools=tools2)
+        assert key1 != key2
+
+    def test_none_tools_same_as_no_tools(self):
+        """None tools and no tools produce same key."""
+        msgs = _make_messages()
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0, tools=None)
+        key2 = generate_cache_key("gpt-4o", msgs, 0.0)
+        assert key1 == key2
+
+    def test_system_prompt_extracted_from_messages(self):
+        """System prompt is extracted from messages[0] with role=system."""
+        msgs = [
+            {"role": "system", "content": "Be concise"},
+            {"role": "user", "content": "Hello"},
+        ]
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0)
+
+        msgs2 = [
+            {"role": "system", "content": "Be verbose"},
+            {"role": "user", "content": "Hello"},
+        ]
+        key2 = generate_cache_key("gpt-4o", msgs2, 0.0)
+        assert key1 != key2
+
+    def test_max_tokens_affects_key(self):
+        """Different max_tokens produce different key."""
+        msgs = _make_messages()
+        key1 = generate_cache_key("gpt-4o", msgs, 0.0, max_tokens=2000)
+        key2 = generate_cache_key("gpt-4o", msgs, 0.0, max_tokens=4000)
+        assert key1 != key2
+
+
+# ---------------------------------------------------------------------------
+# InMemoryLLMCache Tests
+# ---------------------------------------------------------------------------
+
+
+class TestInMemoryLLMCache:
+    @pytest.mark.asyncio
+    async def test_exact_match_hit(self):
+        cache = InMemoryLLMCache()
+        key = "test_key_1"
+        response = _make_response("Cached answer")
+
+        await cache.put(key, response)
+        result = await cache.get(key)
+
+        assert result.hit is True
+        assert result.match_type == "exact"
+        assert result.response.content == "Cached answer"
+
+    @pytest.mark.asyncio
+    async def test_exact_match_miss(self):
+        cache = InMemoryLLMCache()
+        await cache.put("key1", _make_response())
+
+        result = await cache.get("key2")
+        assert result.hit is False
+        assert result.response is None
+
+    @pytest.mark.asyncio
+    async def test_semantic_match_hit(self):
+        cache = InMemoryLLMCache(similarity_threshold=0.9)
+        emb1 = _make_embedding(1.0, dim=64)
+        emb_similar = _make_similar_embedding(emb1, noise=0.001)
+
+        await cache.put("key1", _make_response("Cached"), query_embedding=emb1)
+        result = await cache.semantic_search(emb_similar, threshold=0.9)
+
+        assert result.hit is True
+        assert result.match_type == "semantic"
+        assert result.response.content == "Cached"
+
+    @pytest.mark.asyncio
+    async def test_semantic_match_miss(self):
+        cache = InMemoryLLMCache(similarity_threshold=0.9)
+        emb1 = _make_embedding(1.0, dim=64)
+        emb_different = _make_different_embedding(dim=64)
+
+        await cache.put("key1", _make_response("Cached"), query_embedding=emb1)
+        result = await cache.semantic_search(emb_different, threshold=0.9)
+
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_semantic_match_empty_cache(self):
+        cache = InMemoryLLMCache()
+        result = await cache.semantic_search(_make_embedding(dim=64))
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_ttl_expiry_exact(self):
+        cache = InMemoryLLMCache(exact_ttl=1)  # 1 second TTL
+        await cache.put("key1", _make_response())
+
+        # Wait for expiry
+        time.sleep(1.1)
+        result = await cache.get("key1")
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_ttl_expiry_semantic(self):
+        cache = InMemoryLLMCache(semantic_ttl=1)
+        emb = _make_embedding(dim=64)
+        await cache.put("key1", _make_response(), query_embedding=emb)
+
+        time.sleep(1.1)
+        result = await cache.semantic_search(emb)
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_lru_eviction(self):
+        cache = InMemoryLLMCache(max_entries=3)
+
+        for i in range(4):
+            await cache.put(f"key{i}", _make_response(f"Response {i}"))
+
+        # key0 should be evicted (oldest)
+        result = await cache.get("key0")
+        assert result.hit is False
+
+        # key1-key3 should still be present
+        for i in range(1, 4):
+            result = await cache.get(f"key{i}")
+            assert result.hit is True
+
+    @pytest.mark.asyncio
+    async def test_lru_access_refreshes(self):
+        cache = InMemoryLLMCache(max_entries=3)
+
+        await cache.put("key0", _make_response("R0"))
+        await cache.put("key1", _make_response("R1"))
+        await cache.put("key2", _make_response("R2"))
+
+        # Access key0 to move it to most-recently-used
+        await cache.get("key0")
+
+        # Adding key3 should evict key1 (now LRU)
+        await cache.put("key3", _make_response("R3"))
+
+        result = await cache.get("key1")
+        assert result.hit is False
+
+        result = await cache.get("key0")
+        assert result.hit is True
+
+    @pytest.mark.asyncio
+    async def test_invalidate_all(self):
+        cache = InMemoryLLMCache()
+        await cache.put("key1", _make_response())
+        await cache.put("key2", _make_response())
+
+        count = await cache.invalidate()
+        assert count == 2
+
+        result = await cache.get("key1")
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_invalidate_pattern(self):
+        cache = InMemoryLLMCache()
+        await cache.put("abc_1", _make_response())
+        await cache.put("abc_2", _make_response())
+        await cache.put("xyz_1", _make_response())
+
+        count = await cache.invalidate("abc_*")
+        assert count == 2
+
+        result = await cache.get("xyz_1")
+        assert result.hit is True
+
+    @pytest.mark.asyncio
+    async def test_stats(self):
+        cache = InMemoryLLMCache()
+        await cache.put("key1", _make_response())
+        await cache.get("key1")  # hit
+        await cache.get("key2")  # miss
+
+        stats = await cache.stats()
+        assert stats["total_entries"] == 1
+        assert stats["total_hits"] == 1
+        assert stats["total_misses"] == 1
+
+    @pytest.mark.asyncio
+    async def test_tool_calls_cached(self):
+        cache = InMemoryLLMCache()
+        tool_calls = [
+            ToolCall(id="call_1", name="search", arguments={"query": "test"})
+        ]
+        response = _make_response(tool_calls=tool_calls)
+
+        await cache.put("key1", response)
+        result = await cache.get("key1")
+
+        assert result.hit is True
+        assert len(result.response.tool_calls) == 1
+        assert result.response.tool_calls[0].name == "search"
+
+    @pytest.mark.asyncio
+    async def test_put_without_embedding(self):
+        cache = InMemoryLLMCache()
+        await cache.put("key1", _make_response(), query_embedding=None)
+
+        # Exact match should still work
+        result = await cache.get("key1")
+        assert result.hit is True
+
+        # Semantic search should return miss (no embeddings)
+        result = await cache.semantic_search(_make_embedding(dim=64))
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_put_updates_existing_key(self):
+        cache = InMemoryLLMCache()
+        await cache.put("key1", _make_response("Old"))
+        await cache.put("key1", _make_response("New"))
+
+        result = await cache.get("key1")
+        assert result.hit is True
+        assert result.response.content == "New"
+
+
+# ---------------------------------------------------------------------------
+# Serialization Tests
+# ---------------------------------------------------------------------------
+
+
+class TestSerialization:
+    def test_serialize_deserialize_response(self):
+        response = _make_response(
+            content="Test",
+            model="gpt-4o",
+            prompt_tokens=5,
+            completion_tokens=10,
+            tool_calls=[ToolCall(id="c1", name="tool1", arguments={"k": "v"})],
+        )
+        serialized = _serialize_response(response)
+        deserialized = _deserialize_response(serialized)
+
+        assert deserialized.content == "Test"
+        assert deserialized.model == "gpt-4o"
+        assert deserialized.usage.prompt_tokens == 5
+        assert deserialized.usage.completion_tokens == 10
+        assert len(deserialized.tool_calls) == 1
+        assert deserialized.tool_calls[0].name == "tool1"
+
+    def test_serialize_deserialize_entry(self):
+        entry = CacheEntry(
+            response=_make_response(),
+            query_embedding=[0.1, 0.2, 0.3],
+            created_at=12345.0,
+            hit_count=5,
+        )
+        serialized = _serialize_entry(entry)
+        deserialized = _deserialize_entry(serialized)
+
+        assert deserialized.response.content == "Hello"
+        assert deserialized.query_embedding == [0.1, 0.2, 0.3]
+        assert deserialized.created_at == 12345.0
+        assert deserialized.hit_count == 5
+
+    def test_serialize_response_no_tool_calls(self):
+        response = _make_response()
+        serialized = _serialize_response(response)
+        assert serialized["tool_calls"] == []
+
+        deserialized = _deserialize_response(serialized)
+        assert deserialized.tool_calls == []
+
+
+# ---------------------------------------------------------------------------
+# RedisLLMCache Tests (with mocked Redis)
+# ---------------------------------------------------------------------------
+
+
+class TestRedisLLMCache:
+    def _make_mock_redis(self):
+        """Create a mock Redis client that simulates basic operations."""
+        mock = AsyncMock()
+        mock._data = {}
+        mock._sets = {}
+
+        async def mock_get(key):
+            return mock._data.get(key)
+
+        async def mock_set(key, value, ex=None):
+            mock._data[key] = value
+
+        async def mock_mget(keys):
+            return [mock._data.get(k) for k in keys]
+
+        async def mock_sadd(key, *members):
+            if key not in mock._sets:
+                mock._sets[key] = set()
+            mock._sets[key].update(members)
+
+        async def mock_smembers(key):
+            return mock._sets.get(key, set())
+
+        async def mock_scard(key):
+            return len(mock._sets.get(key, set()))
+
+        async def mock_delete(*keys):
+            for k in keys:
+                mock._data.pop(k, None)
+
+        async def mock_srem(key, *members):
+            if key in mock._sets:
+                mock._sets[key] -= set(members)
+
+        mock.get = mock_get
+        mock.set = mock_set
+        mock.mget = mock_mget
+        mock.sadd = mock_sadd
+        mock.smembers = mock_smembers
+        mock.scard = mock_scard
+        mock.delete = mock_delete
+        mock.srem = mock_srem
+
+        # Pipeline mock — collects commands and executes them on execute()
+        class MockPipeline:
+            def __init__(self):
+                self._commands = []
+
+            def set(self, key, value, ex=None):
+                self._commands.append(("set", key, value, ex))
+
+            def sadd(self, key, *members):
+                self._commands.append(("sadd", key, members))
+
+            def delete(self, *keys):
+                self._commands.append(("delete", keys))
+
+            def srem(self, key, *members):
+                self._commands.append(("srem", key, members))
+
+            async def execute(self):
+                for cmd in self._commands:
+                    if cmd[0] == "set":
+                        _, key, value, ex = cmd
+                        mock._data[key] = value
+                    elif cmd[0] == "sadd":
+                        _, key, members = cmd
+                        if key not in mock._sets:
+                            mock._sets[key] = set()
+                        mock._sets[key].update(members)
+                    elif cmd[0] == "delete":
+                        _, keys = cmd
+                        for k in keys:
+                            mock._data.pop(k, None)
+                    elif cmd[0] == "srem":
+                        _, key, members = cmd
+                        if key in mock._sets:
+                            mock._sets[key] -= set(members)
+
+        mock.pipeline = MagicMock(return_value=MockPipeline())
+
+        return mock
+
+    @pytest.mark.asyncio
+    async def test_exact_match_hit(self):
+        cache = RedisLLMCache()
+        mock_redis = self._make_mock_redis()
+        cache._redis = mock_redis
+
+        key = "test_key"
+        response = _make_response("Cached")
+        entry = CacheEntry(response=response, created_at=time.monotonic(), hit_count=0)
+        entry_json = json.dumps(_serialize_entry(entry))
+
+        # Simulate Redis already has the data
+        mock_redis._data[f"{cache.KEY_PREFIX}{key}"] = entry_json
+
+        result = await cache.get(key)
+        assert result.hit is True
+        assert result.match_type == "exact"
+        assert result.response.content == "Cached"
+
+    @pytest.mark.asyncio
+    async def test_exact_match_miss(self):
+        cache = RedisLLMCache()
+        mock_redis = self._make_mock_redis()
+        cache._redis = mock_redis
+
+        result = await cache.get("nonexistent_key")
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_redis_failure_returns_miss(self):
+        cache = RedisLLMCache()
+        mock_redis = AsyncMock()
+        mock_redis.get = AsyncMock(side_effect=Exception("Connection refused"))
+        cache._redis = mock_redis
+
+        result = await cache.get("any_key")
+        assert result.hit is False
+
+    @pytest.mark.asyncio
+    async def test_put_stores_data(self):
+        cache = RedisLLMCache()
+        mock_redis = self._make_mock_redis()
+        cache._redis = mock_redis
+
+        key = "test_key"
+        response = _make_response()
+        emb = [0.1, 0.2, 0.3]
+
+        await cache.put(key, response, query_embedding=emb)
+
+        # Verify data was stored
+        assert f"{cache.KEY_PREFIX}{key}" in mock_redis._data
+        assert f"{cache.EMB_PREFIX}{key}" in mock_redis._data
+        assert key in mock_redis._sets.get(cache.INDEX_KEY, set())
+
+    @pytest.mark.asyncio
+    async def test_invalidate_all(self):
+        cache = RedisLLMCache()
+        mock_redis = self._make_mock_redis()
+        cache._redis = mock_redis
+
+        # Pre-populate
+        mock_redis._sets[cache.INDEX_KEY] = {"key1", "key2"}
+        mock_redis._data[f"{cache.KEY_PREFIX}key1"] = "data1"
+        mock_redis._data[f"{cache.KEY_PREFIX}key2"] = "data2"
+
+        count = await cache.invalidate()
+        assert count == 2
+
+    @pytest.mark.asyncio
+    async def test_stats(self):
+        cache = RedisLLMCache()
+        mock_redis = self._make_mock_redis()
+        cache._redis = mock_redis
+        mock_redis._sets[cache.INDEX_KEY] = {"k1", "k2", "k3"}
+
+        stats = await cache.stats()
+        assert stats["total_entries"] == 3
+
+
+# ---------------------------------------------------------------------------
+# Factory Tests
+# ---------------------------------------------------------------------------
+
+
+class TestCreateLLMCache:
+    def test_memory_backend(self):
+        cache = create_llm_cache(backend="memory")
+        assert isinstance(cache, InMemoryLLMCache)
+
+    def test_auto_backend_fallback(self):
+        """When redis package is not available, auto falls back to InMemory."""
+        with patch.dict("sys.modules", {"redis.asyncio": None}):
+            # Force ImportError by making redis.asyncio unimportable
+            cache = create_llm_cache(backend="auto")
+            assert isinstance(cache, InMemoryLLMCache)
+
+    def test_redis_backend_with_redis_available(self):
+        """When redis.asyncio is available, auto/redis returns RedisLLMCache."""
+        cache = create_llm_cache(backend="redis")
+        assert isinstance(cache, RedisLLMCache)
+
+    def test_auto_backend_with_redis_available(self):
+        cache = create_llm_cache(backend="auto")
+        assert isinstance(cache, RedisLLMCache)
+
+    def test_custom_parameters(self):
+        cache = create_llm_cache(
+            backend="memory",
+            max_entries=500,
+            exact_ttl=7200,
+            semantic_ttl=172800,
+            similarity_threshold=0.95,
+        )
+        assert isinstance(cache, InMemoryLLMCache)
+        assert cache._max_entries == 500
+        assert cache._exact_ttl == 7200
+        assert cache._semantic_ttl == 172800
+        assert cache._similarity_threshold == 0.95

tests/unit/test_semantic_router.py

@@ -0,0 +1,219 @@
+"""Unit tests for Semantic Router (U3)."""
+
+import pytest
+
+from agentkit.chat.semantic_router import (
+    SemanticRouteResult,
+    SkillEmbeddingIndex,
+    SemanticRouter,
+)
+from agentkit.memory.embedder import MockEmbedder
+
+
+def _make_embedding(base_val: float = 1.0, dim: int = 128) -> list[float]:
+    """Create a unit vector for similarity testing."""
+    vec = [base_val] * dim
+    magnitude = sum(x**2 for x in vec) ** 0.5
+    return [x / magnitude for x in vec] if magnitude > 0 else vec
+
+
+class MockSkill:
+    """Mock skill for testing."""
+
+    def __init__(self, name: str, description: str = "", keywords: list[str] | None = None, capabilities: list[str] | None = None):
+        self.name = name
+        self.config = MockSkillConfig(
+            name=name,
+            description=description,
+            keywords=keywords or [],
+            capabilities=capabilities or [],
+        )
+
+
+class MockSkillConfig:
+    """Mock skill config for testing."""
+
+    def __init__(self, name: str, description: str = "", keywords: list[str] | None = None, capabilities: list[str] | None = None):
+        self.name = name
+        self.description = description
+        self.intent = MockIntentConfig(keywords=keywords or [])
+        self.capabilities = [MockCapabilityTag(tag=t) for t in (capabilities or [])]
+
+
+class MockIntentConfig:
+    def __init__(self, keywords: list[str] | None = None):
+        self.keywords = keywords or []
+
+
+class MockCapabilityTag:
+    def __init__(self, tag: str):
+        self.tag = tag
+
+
+class MockSkillRegistry:
+    """Mock skill registry for testing."""
+
+    def __init__(self, skills: list[MockSkill] | None = None):
+        self._skills = {s.name: s for s in (skills or [])}
+
+    def list_skills(self):
+        return list(self._skills.values())
+
+    def get(self, name: str):
+        if name not in self._skills:
+            raise KeyError(f"Skill '{name}' not found")
+        return self._skills[name]
+
+
+class TestSkillEmbeddingIndex:
+    @pytest.mark.asyncio
+    async def test_build_from_registry(self):
+        embedder = MockEmbedder(dimension=64)
+        index = SkillEmbeddingIndex(embedder)
+
+        skills = [
+            MockSkill("content_gen", description="生成文章内容", keywords=["写作", "文章"], capabilities=["content"]),
+            MockSkill("data_analysis", description="数据分析与可视化", keywords=["分析", "数据"], capabilities=["analytics"]),
+        ]
+        registry = MockSkillRegistry(skills)
+        await index.build(registry)
+
+        assert index.size == 2
+
+    @pytest.mark.asyncio
+    async def test_search_returns_results(self):
+        embedder = MockEmbedder(dimension=64)
+        index = SkillEmbeddingIndex(embedder)
+
+        skill = MockSkill("content_gen", description="生成文章内容")
+        await index.update_skill("content_gen", skill)
+
+        # MockEmbedder produces deterministic embeddings based on text hash
+        # Different text → different embedding
+        query_emb = await embedder.embed("生成文章")
+        results = await index.search(query_emb)
+
+        assert len(results) >= 1
+        assert results[0][0] == "content_gen"  # skill_name
+        assert results[0][1] > 0.0  # similarity
+
+    @pytest.mark.asyncio
+    async def test_search_empty_index(self):
+        embedder = MockEmbedder(dimension=64)
+        index = SkillEmbeddingIndex(embedder)
+
+        query_emb = await embedder.embed("test")
+        results = await index.search(query_emb)
+
+        assert results == []
+
+    @pytest.mark.asyncio
+    async def test_remove_skill(self):
+        embedder = MockEmbedder(dimension=64)
+        index = SkillEmbeddingIndex(embedder)
+
+        skill = MockSkill("test_skill", description="Test")
+        await index.update_skill("test_skill", skill)
+        assert index.size == 1
+
+        index.remove_skill("test_skill")
+        assert index.size == 0
+
+    def test_build_source_text_with_description(self):
+        skill = MockSkill("test", description="A test skill", keywords=["test"], capabilities=["testing"])
+        text = SkillEmbeddingIndex._build_source_text(skill)
+        assert "A test skill" in text
+        assert "test" in text
+        assert "testing" in text
+
+    def test_build_source_text_fallback_to_name(self):
+        skill = MockSkill("my_skill", description="", keywords=[], capabilities=[])
+        text = SkillEmbeddingIndex._build_source_text(skill)
+        assert "my_skill" in text
+
+
+class TestSemanticRouter:
+    @pytest.mark.asyncio
+    async def test_high_confidence_match(self):
+        """When similarity > 0.85, return high confidence."""
+        embedder = MockEmbedder(dimension=64)
+        router = SemanticRouter(embedder, similarity_high=0.5, similarity_low=0.3)
+
+        # Add a skill with known embedding
+        skill = MockSkill("content_gen", description="生成文章内容")
+        await router.update_skill("content_gen", skill)
+
+        # Query with same text should produce very similar embedding (MockEmbedder is hash-based)
+        # With low thresholds, even moderate similarity will be "high"
+        result = await router.route("生成文章内容")
+        # MockEmbedder may or may not produce high similarity for different text
+        # Just verify the result structure
+        assert result.confidence in ("high", "medium", "low")
+        assert isinstance(result.similarity, float)
+
+    @pytest.mark.asyncio
+    async def test_low_confidence_empty_index(self):
+        """Empty index returns low confidence."""
+        embedder = MockEmbedder(dimension=64)
+        router = SemanticRouter(embedder)
+
+        result = await router.route("任何查询")
+        assert result.confidence == "low"
+        assert result.skill_name is None
+        assert result.similarity == 0.0
+
+    @pytest.mark.asyncio
+    async def test_medium_confidence_zone(self):
+        """Test medium confidence zone (0.6-0.85)."""
+        embedder = MockEmbedder(dimension=64)
+        router = SemanticRouter(embedder, similarity_high=0.99, similarity_low=0.01)
+
+        skill = MockSkill("content_gen", description="生成文章内容")
+        await router.update_skill("content_gen", skill)
+
+        # With very high similarity_high and very low similarity_low,
+        # most matches will be "medium"
+        result = await router.route("生成文章")
+        # The result should be medium (since threshold is 0.99)
+        assert result.confidence in ("medium", "low", "high")
+
+    @pytest.mark.asyncio
+    async def test_embedder_failure_graceful(self):
+        """Embedder failure returns low confidence."""
+        class FailingEmbedder(MockEmbedder):
+            async def embed(self, text):
+                raise RuntimeError("Embedding API failed")
+
+        router = SemanticRouter(FailingEmbedder(dimension=64))
+        result = await router.route("test query")
+        assert result.confidence == "low"
+        assert result.skill_name is None
+
+    @pytest.mark.asyncio
+    async def test_build_index_from_registry(self):
+        """Build index from skill registry."""
+        embedder = MockEmbedder(dimension=64)
+        router = SemanticRouter(embedder)
+
+        skills = [
+            MockSkill("skill_a", description="Skill A"),
+            MockSkill("skill_b", description="Skill B"),
+        ]
+        registry = MockSkillRegistry(skills)
+        await router.build_index(registry)
+
+        assert router._index.size == 2
+
+    @pytest.mark.asyncio
+    async def test_chinese_query(self):
+        """Chinese query works with semantic router."""
+        embedder = MockEmbedder(dimension=64)
+        router = SemanticRouter(embedder, similarity_high=0.01, similarity_low=0.001)
+
+        skill = MockSkill("geo_optimizer", description="地理内容优化", keywords=["优化", "SEO", "地理"], capabilities=["optimization"])
+        await router.update_skill("geo_optimizer", skill)
+
+        result = await router.route("帮我优化内容")
+        # With very low thresholds, should match
+        assert result.confidence in ("high", "medium")
+        assert result.skill_name == "geo_optimizer"

tests/unit/test_unified_evolution_store.py

@@ -0,0 +1,458 @@
+"""Tests for unified EvolutionStoreProtocol compliance
+
+Verifies that all backends implement the full Protocol interface:
+- InMemoryEvolutionStore
+- PersistentEvolutionStore
+- PostgreSQLEvolutionStore (mocked async session)
+- EvolutionStore (legacy, with NotImplementedError for skill_version/ab_test)
+"""
+
+import os
+import tempfile
+
+import pytest
+
+from agentkit.core.protocol import EvolutionEvent
+from agentkit.evolution.evolution_store import (
+    EvolutionStore,
+    EvolutionStoreProtocol,
+    InMemoryEvolutionStore,
+    PersistentEvolutionStore,
+    create_evolution_store,
+)
+
+
+# ── Fixtures ──────────────────────────────────────────────
+
+
+@pytest.fixture
+def sample_event():
+    """A sample EvolutionEvent."""
+    return EvolutionEvent(
+        agent_name="test_agent",
+        change_type="prompt",
+        before={"prompt": "old prompt"},
+        after={"prompt": "new prompt"},
+        metrics={"accuracy": 0.9},
+    )
+
+
+@pytest.fixture
+def memory_store():
+    return InMemoryEvolutionStore()
+
+
+@pytest.fixture
+def sqlite_store(tmp_path):
+    db_path = str(tmp_path / "test_unified.db")
+    return PersistentEvolutionStore(db_path=db_path)
+
+
+# ── Protocol compliance tests ─────────────────────────────
+
+
+class TestProtocolCompliance:
+    """Verify all stores implement EvolutionStoreProtocol."""
+
+    def test_inmemory_is_protocol(self):
+        assert isinstance(InMemoryEvolutionStore(), EvolutionStoreProtocol)
+
+    def test_persistent_is_protocol(self, tmp_path):
+        db_path = str(tmp_path / "protocol_check.db")
+        assert isinstance(PersistentEvolutionStore(db_path=db_path), EvolutionStoreProtocol)
+
+    def test_pg_store_is_protocol(self):
+        from agentkit.evolution.pg_store import PostgreSQLEvolutionStore
+
+        store = PostgreSQLEvolutionStore(database_url="postgresql+asyncpg://test:test@localhost/test")
+        assert isinstance(store, EvolutionStoreProtocol)
+
+    def test_legacy_evolution_store_is_protocol(self):
+        """Legacy EvolutionStore also satisfies Protocol (has all method signatures)."""
+        from unittest.mock import AsyncMock, MagicMock
+
+        store = EvolutionStore(session_factory=MagicMock(), evolution_model=MagicMock())
+        assert isinstance(store, EvolutionStoreProtocol)
+
+
+# ── InMemoryEvolutionStore: full Protocol ─────────────────
+
+
+class TestInMemoryFullProtocol:
+    """InMemoryEvolutionStore implements all Protocol methods."""
+
+    async def test_record_and_list_events(self, memory_store, sample_event):
+        event_id = await memory_store.record(sample_event)
+        assert event_id is not None
+
+        events = await memory_store.list_events()
+        assert len(events) == 1
+        assert events[0]["agent_name"] == "test_agent"
+        assert events[0]["change_type"] == "prompt"
+
+    async def test_rollback(self, memory_store, sample_event):
+        event_id = await memory_store.record(sample_event)
+        result = await memory_store.rollback(event_id)
+        assert result is True
+
+        events = await memory_store.list_events()
+        assert events[0]["status"] == "rolled_back"
+
+    async def test_rollback_nonexistent(self, memory_store):
+        result = await memory_store.rollback("nonexistent")
+        assert result is False
+
+    async def test_list_events_with_filters(self, memory_store):
+        await memory_store.record(
+            EvolutionEvent(agent_name="a", change_type="prompt", before={}, after={})
+        )
+        await memory_store.record(
+            EvolutionEvent(agent_name="b", change_type="strategy", before={}, after={})
+        )
+
+        events = await memory_store.list_events(agent_name="a")
+        assert len(events) == 1
+        assert events[0]["agent_name"] == "a"
+
+    async def test_record_and_list_skill_version(self, memory_store):
+        vid = await memory_store.record_skill_version("search", "v1", '{"prompt": "v1"}')
+        assert vid is not None
+
+        versions = await memory_store.list_skill_versions("search")
+        assert len(versions) == 1
+        assert versions[0]["version"] == "v1"
+        assert versions[0]["content"] == '{"prompt": "v1"}'
+
+    async def test_skill_version_with_parent(self, memory_store):
+        await memory_store.record_skill_version("search", "v1", '{"prompt": "v1"}')
+        await memory_store.record_skill_version(
+            "search", "v2", '{"prompt": "v2"}', parent_version="v1"
+        )
+
+        versions = await memory_store.list_skill_versions("search")
+        assert len(versions) == 2
+        assert versions[0]["version"] == "v2"
+        assert versions[0]["parent_version"] == "v1"
+
+    async def test_record_and_get_ab_test_result(self, memory_store):
+        rid = await memory_store.record_ab_test_result("t1", "control", 0.8, 5)
+        assert rid is not None
+
+        results = await memory_store.get_ab_test_results("t1")
+        assert len(results) == 1
+        assert results[0]["variant"] == "control"
+        assert results[0]["score"] == 0.8
+        assert results[0]["sample_count"] == 5
+
+    async def test_ab_test_multiple_variants(self, memory_store):
+        await memory_store.record_ab_test_result("t1", "control", 0.8, 10)
+        await memory_store.record_ab_test_result("t1", "experiment", 0.9, 10)
+
+        results = await memory_store.get_ab_test_results("t1")
+        assert len(results) == 2
+
+    async def test_list_skill_versions_empty(self, memory_store):
+        versions = await memory_store.list_skill_versions("nonexistent")
+        assert versions == []
+
+    async def test_get_ab_test_results_empty(self, memory_store):
+        results = await memory_store.get_ab_test_results("nonexistent")
+        assert results == []
+
+
+# ── PersistentEvolutionStore: full Protocol ───────────────
+
+
+class TestSQLiteFullProtocol:
+    """PersistentEvolutionStore implements all Protocol methods."""
+
+    async def test_record_and_list_events(self, sqlite_store, sample_event):
+        event_id = await sqlite_store.record(sample_event)
+        assert event_id is not None
+
+        events = await sqlite_store.list_events()
+        assert len(events) == 1
+        assert events[0]["agent_name"] == "test_agent"
+
+    async def test_rollback(self, sqlite_store, sample_event):
+        event_id = await sqlite_store.record(sample_event)
+        result = await sqlite_store.rollback(event_id)
+        assert result is True
+
+        events = await sqlite_store.list_events()
+        assert events[0]["status"] == "rolled_back"
+
+    async def test_record_and_list_skill_version(self, sqlite_store):
+        vid = await sqlite_store.record_skill_version("search", "v1", '{"prompt": "v1"}')
+        assert vid is not None
+
+        versions = await sqlite_store.list_skill_versions("search")
+        assert len(versions) == 1
+        assert versions[0]["version"] == "v1"
+
+    async def test_skill_version_with_parent(self, sqlite_store):
+        await sqlite_store.record_skill_version("search", "v1", '{"prompt": "v1"}')
+        await sqlite_store.record_skill_version(
+            "search", "v2", '{"prompt": "v2"}', parent_version="v1"
+        )
+
+        versions = await sqlite_store.list_skill_versions("search")
+        assert len(versions) == 2
+        assert versions[0]["version"] == "v2"
+        assert versions[0]["parent_version"] == "v1"
+
+    async def test_record_and_get_ab_test_result(self, sqlite_store):
+        rid = await sqlite_store.record_ab_test_result("t1", "control", 0.8, 5)
+        assert rid is not None
+
+        results = await sqlite_store.get_ab_test_results("t1")
+        assert len(results) == 1
+        assert results[0]["variant"] == "control"
+
+    async def test_ab_test_multiple_variants(self, sqlite_store):
+        await sqlite_store.record_ab_test_result("t1", "control", 0.8, 10)
+        await sqlite_store.record_ab_test_result("t1", "experiment", 0.9, 10)
+
+        results = await sqlite_store.get_ab_test_results("t1")
+        assert len(results) == 2
+
+    async def test_list_skill_versions_empty(self, sqlite_store):
+        versions = await sqlite_store.list_skill_versions("nonexistent")
+        assert versions == []
+
+    async def test_get_ab_test_results_empty(self, sqlite_store):
+        results = await sqlite_store.get_ab_test_results("nonexistent")
+        assert results == []
+
+
+# ── PostgreSQLEvolutionStore: mocked Protocol ─────────────
+
+
+class TestPGStoreMocked:
+    """Test PostgreSQLEvolutionStore with mocked async session.
+
+    Since we can't require a running PostgreSQL in unit tests,
+    we mock the async session to verify the logic paths.
+    """
+
+    def _make_pg_store(self):
+        from agentkit.evolution.pg_store import PostgreSQLEvolutionStore
+
+        return PostgreSQLEvolutionStore(database_url="postgresql+asyncpg://test:test@localhost/test")
+
+    async def test_record_with_mock_session(self, sample_event):
+        from unittest.mock import AsyncMock, MagicMock, patch
+        from contextlib import asynccontextmanager
+
+        store = self._make_pg_store()
+
+        # Mock the session factory
+        mock_session = AsyncMock()
+        mock_session.add = MagicMock()
+        mock_session.commit = AsyncMock()
+        mock_session.rollback = AsyncMock()
+
+        @asynccontextmanager
+        async def mock_sf():
+            yield mock_session
+
+        store._session_factory = mock_sf
+        store._initialized = True
+
+        event_id = await store.record(sample_event)
+        assert event_id is not None
+        assert sample_event.event_id == event_id
+        mock_session.add.assert_called_once()
+        mock_session.commit.assert_called_once()
+
+    async def test_rollback_with_mock_session(self):
+        from unittest.mock import AsyncMock, MagicMock, patch
+        from contextlib import asynccontextmanager
+
+        store = self._make_pg_store()
+
+        # Create a mock entry
+        mock_entry = MagicMock()
+        mock_entry.status = "active"
+
+        mock_result = MagicMock()
+        mock_result.scalar_one_or_none.return_value = mock_entry
+
+        mock_session = AsyncMock()
+        mock_session.execute = AsyncMock(return_value=mock_result)
+        mock_session.commit = AsyncMock()
+        mock_session.rollback = AsyncMock()
+
+        @asynccontextmanager
+        async def mock_sf():
+            yield mock_session
+
+        store._session_factory = mock_sf
+        store._initialized = True
+
+        result = await store.rollback("test-event-id")
+        assert result is True
+        assert mock_entry.status == "rolled_back"
+        mock_session.commit.assert_called_once()
+
+    async def test_rollback_not_found(self):
+        from unittest.mock import AsyncMock, MagicMock
+        from contextlib import asynccontextmanager
+
+        store = self._make_pg_store()
+
+        mock_result = MagicMock()
+        mock_result.scalar_one_or_none.return_value = None
+
+        mock_session = AsyncMock()
+        mock_session.execute = AsyncMock(return_value=mock_result)
+        mock_session.rollback = AsyncMock()
+
+        @asynccontextmanager
+        async def mock_sf():
+            yield mock_session
+
+        store._session_factory = mock_sf
+        store._initialized = True
+
+        result = await store.rollback("nonexistent")
+        assert result is False
+
+    async def test_record_skill_version_with_mock(self):
+        from unittest.mock import AsyncMock, MagicMock
+        from contextlib import asynccontextmanager
+
+        store = self._make_pg_store()
+
+        mock_session = AsyncMock()
+        mock_session.add = MagicMock()
+        mock_session.commit = AsyncMock()
+        mock_session.rollback = AsyncMock()
+
+        @asynccontextmanager
+        async def mock_sf():
+            yield mock_session
+
+        store._session_factory = mock_sf
+        store._initialized = True
+
+        vid = await store.record_skill_version("search", "v1", '{"prompt": "v1"}')
+        assert vid is not None
+        mock_session.add.assert_called_once()
+        mock_session.commit.assert_called_once()
+
+    async def test_record_ab_test_result_with_mock(self):
+        from unittest.mock import AsyncMock, MagicMock
+        from contextlib import asynccontextmanager
+
+        store = self._make_pg_store()
+
+        mock_session = AsyncMock()
+        mock_session.add = MagicMock()
+        mock_session.commit = AsyncMock()
+        mock_session.rollback = AsyncMock()
+
+        @asynccontextmanager
+        async def mock_sf():
+            yield mock_session
+
+        store._session_factory = mock_sf
+        store._initialized = True
+
+        rid = await store.record_ab_test_result("t1", "control", 0.8, 5)
+        assert rid is not None
+        mock_session.add.assert_called_once()
+        mock_session.commit.assert_called_once()
+
+
+# ── Legacy EvolutionStore: NotImplementedError tests ──────
+
+
+class TestLegacyEvolutionStoreStubs:
+    """Legacy EvolutionStore raises NotImplementedError for skill_version/ab_test."""
+
+    async def test_record_skill_version_raises(self):
+        from unittest.mock import MagicMock
+
+        store = EvolutionStore(session_factory=MagicMock(), evolution_model=MagicMock())
+        with pytest.raises(NotImplementedError, match="skill_version"):
+            await store.record_skill_version("s", "v1", "content")
+
+    async def test_list_skill_versions_raises(self):
+        from unittest.mock import MagicMock
+
+        store = EvolutionStore(session_factory=MagicMock(), evolution_model=MagicMock())
+        with pytest.raises(NotImplementedError, match="skill_version"):
+            await store.list_skill_versions("s")
+
+    async def test_record_ab_test_result_raises(self):
+        from unittest.mock import MagicMock
+
+        store = EvolutionStore(session_factory=MagicMock(), evolution_model=MagicMock())
+        with pytest.raises(NotImplementedError, match="A/B test"):
+            await store.record_ab_test_result("t1", "control", 0.8)
+
+    async def test_get_ab_test_results_raises(self):
+        from unittest.mock import MagicMock
+
+        store = EvolutionStore(session_factory=MagicMock(), evolution_model=MagicMock())
+        with pytest.raises(NotImplementedError, match="A/B test"):
+            await store.get_ab_test_results("t1")
+
+
+# ── Factory tests ─────────────────────────────────────────
+
+
+class TestCreateEvolutionStoreExtended:
+    def test_create_memory_backend(self):
+        store = create_evolution_store(backend="memory")
+        assert isinstance(store, InMemoryEvolutionStore)
+
+    def test_create_sqlite_backend(self, tmp_path):
+        db_path = str(tmp_path / "factory_test.db")
+        store = create_evolution_store(backend="sqlite", db_path=db_path)
+        assert isinstance(store, PersistentEvolutionStore)
+
+    def test_create_default_backend(self):
+        store = create_evolution_store()
+        assert isinstance(store, InMemoryEvolutionStore)
+
+    def test_create_sql_backend_without_params_falls_back(self):
+        store = create_evolution_store(backend="sql")
+        assert isinstance(store, InMemoryEvolutionStore)
+
+    def test_create_postgresql_without_url_falls_back(self):
+        """PostgreSQL backend without database_url falls back to memory."""
+        # Clear env var if set
+        old_val = os.environ.pop("AGENTKIT_DATABASE_URL", None)
+        try:
+            store = create_evolution_store(backend="postgresql")
+            assert isinstance(store, InMemoryEvolutionStore)
+        finally:
+            if old_val is not None:
+                os.environ["AGENTKIT_DATABASE_URL"] = old_val
+
+    def test_create_postgresql_with_url(self):
+        """PostgreSQL backend with database_url returns PostgreSQLEvolutionStore."""
+        from agentkit.evolution.pg_store import PostgreSQLEvolutionStore
+
+        store = create_evolution_store(
+            backend="postgresql",
+            database_url="postgresql+asyncpg://user:pass@localhost/db",
+        )
+        assert isinstance(store, PostgreSQLEvolutionStore)
+
+    def test_create_postgresql_with_env_url(self):
+        """PostgreSQL backend reads database_url from environment variable."""
+        from agentkit.evolution.pg_store import PostgreSQLEvolutionStore
+
+        old_val = os.environ.get("AGENTKIT_DATABASE_URL")
+        try:
+            os.environ["AGENTKIT_DATABASE_URL"] = "postgresql+asyncpg://user:pass@localhost/db"
+            store = create_evolution_store(backend="postgresql")
+            assert isinstance(store, PostgreSQLEvolutionStore)
+        finally:
+            if old_val is not None:
+                os.environ["AGENTKIT_DATABASE_URL"] = old_val
+            else:
+                os.environ.pop("AGENTKIT_DATABASE_URL", None)

tests/unit/test_usage_tracker.py

@@ -5,7 +5,8 @@ from datetime import datetime, timedelta, timezone
 import pytest
 
 from agentkit.llm.protocol import TokenUsage
-from agentkit.llm.providers.tracker import UsageRecord, UsageSummary, UsageTracker
+from agentkit.llm.providers.tracker import UsageSummary, UsageTracker
+from agentkit.llm.providers.usage_store import UsageRecord
 
 
 class TestUsageTrackerRecord:
@@ -23,8 +24,10 @@ class TestUsageTrackerRecord:
             latency_ms=200.0,
         )
 
-        assert len(tracker._records) == 1
-        rec = tracker._records[0]
+        # Verify via get_usage() instead of internal _records
+        summary = tracker.get_usage()
+        assert len(summary.records) == 1
+        rec = summary.records[0]
         assert rec.agent_name == "test_agent"
         assert rec.model == "gpt-4o"
         assert rec.prompt_tokens == 100
@@ -41,7 +44,8 @@ class TestUsageTrackerRecord:
         tracker.record("agent_a", "gpt-4o", usage1, 0.001, 100.0)
         tracker.record("agent_b", "deepseek-chat", usage2, 0.002, 150.0)
 
-        assert len(tracker._records) == 2
+        summary = tracker.get_usage()
+        assert len(summary.records) == 2
 
 
 class TestUsageTrackerGetUsage:
@@ -80,10 +84,11 @@ class TestUsageTrackerGetUsage:
         usage2 = TokenUsage(prompt_tokens=200, completion_tokens=100)
 
         tracker.record("agent_a", "gpt-4o", usage1, 0.005, 100.0)
-
-        # Manually set timestamp of second record to 2 hours ago
         tracker.record("agent_a", "gpt-4o", usage2, 0.010, 200.0)
-        tracker._records[-1].timestamp = now - timedelta(hours=2)
+
+        # Manually set timestamp of second record to 2 hours ago via store
+        store = tracker._store
+        store._records[-1].timestamp = (now - timedelta(hours=2)).isoformat()
 
         # Query last hour only
         summary = tracker.get_usage(start_time=now - timedelta(hours=1), end_time=now + timedelta(hours=1))