Plan step

common/llm_client.py

@@ -1,12 +1,66 @@
+import hashlib
 import json
+import logging
+from pathlib import Path
+from typing import Optional, Type
+
 from openai import OpenAI
+from pydantic import BaseModel
+
+class NoSchema(BaseModel): ...
+
+
+logger = logging.getLogger("llm client")
+
+class ResponseCache:
+    """File-backed cache: prompt-hash -> validated-model JSON. Makes runs idempotent."""
+
+    def __init__(self, path: Optional[Path] = None) -> None:
+        self.path = path
+        self._mem: dict[str, str] = {}
+        if path and path.exists():
+            try:
+                self._mem = json.loads(path.read_text("utf-8"))
+            except Exception:  # noqa: BLE001
+                logger.warning("Corrupt cache at %s; starting fresh", path)
+
+    @staticmethod
+    def key(system: str, user: str, schema: Type[BaseModel]) -> str:
+        h = hashlib.sha256()
+        h.update(schema.__name__.encode())
+        h.update(b"\x00")
+        h.update(system.encode())
+        h.update(b"\x00")
+        h.update(user.encode())
+        return h.hexdigest()
+
+    def get(self, k: str) -> Optional[str]:
+        return self._mem.get(k)
+
+    def put(self, k: str, v: str) -> None:
+        self._mem[k] = v
+        if self.path:
+            self.path.parent.mkdir(parents=True, exist_ok=True)
+            tmp = self.path.with_suffix(self.path.suffix + ".tmp")
+            tmp.write_text(json.dumps(self._mem, indent=2, sort_keys=True), "utf-8")
+            tmp.replace(self.path)
+        else:
+            logger.warning("No path specified for cache!")
+
 
 class LLMClient:
-    def __init__(self, base_url, api_key: str | None = None, model: str = "gpt-4o-mini"):
+    def __init__(self, base_url, api_key: str | None = None, model: str = "gpt-4o-mini", cache: Optional[ResponseCache] = None):
         self.model = model
         self.client = OpenAI(base_url=base_url, api_key=api_key or "ollama")
+        self.cache = cache
 
     def chat_with_schema(self, prompt: str, schema: dict, system: str | None = None) -> str | None:
+        # --- idempotency: serve from cache when prompt+system+schema match ---
+
+        cache_key = ResponseCache.key(system or "", prompt, NoSchema) if self.cache else None
+        if cache_key and self.cache and (hit := self.cache.get(cache_key)) is not None:
+            return json.loads(hit).get('response', "")
+
         assert_strict_mode_clean(schema)
         kwargs = {
             "model": self.model,
@@ -19,6 +73,10 @@ class LLMClient:
             "response_format": schema,
         }
         resp = self.client.chat.completions.create(**kwargs)
+
+        if cache_key and self.cache:
+            self.cache.put(cache_key, json.dumps({"response": resp.choices[0].message.content or ""}))
+
         return resp.choices[0].message.content
 
     def chat(self, system: str, user: str, json_mode: bool = False, images: list[str] | None = None) -> str | None: