devcouncil 0.1.1 → 0.3.0

package/src/devcouncil/llm/cache.py

@@ -1,38 +1,50 @@
 import json
 import hashlib
+import logging
 from pathlib import Path
 from typing import Optional
 from devcouncil.llm.provider import LLMResponse
 
+logger = logging.getLogger(__name__)
+
 class LLMCache:
     def __init__(self, project_root: Path):
         self.cache_dir = project_root / ".devcouncil" / "cache" / "llm"
         self.cache_dir.mkdir(parents=True, exist_ok=True)
 
-    def _get_key(self, model: str, messages: list, temp: float, json_mode: bool) -> str:
+    def _get_key(self, model: str, messages: list, temp: float, json_mode: bool, provider_fingerprint: str = "") -> str:
         data = {
             "model": model,
             "messages": messages,
             "temp": temp,
-            "json_mode": json_mode
+            "json_mode": json_mode,
+            # Provider-specific knobs that change the output for an identical prompt
+            # (e.g. Ollama's num_ctx / base_url). Empty for providers without such knobs,
+            # so their cache keys are unchanged.
+            "provider": provider_fingerprint,
         }
         s = json.dumps(data, sort_keys=True)
         return hashlib.sha256(s.encode("utf-8")).hexdigest()
 
-    def get(self, model: str, messages: list, temp: float, json_mode: bool) -> Optional[LLMResponse]:
-        key = self._get_key(model, messages, temp, json_mode)
+    def get(self, model: str, messages: list, temp: float, json_mode: bool, provider_fingerprint: str = "", cache_key: Optional[str] = None) -> Optional[LLMResponse]:
+        # Hashing the JSON payload is non-trivial; let callers compute the key once
+        # (via ``_get_key``) and pass it to both get() and set() to avoid recomputing.
+        key = cache_key if cache_key is not None else self._get_key(model, messages, temp, json_mode, provider_fingerprint)
         cache_file = self.cache_dir / f"{key}.json"
         if cache_file.exists():
             try:
                 with open(cache_file, "r") as f:
                     data = json.load(f)
+                    logger.debug("LLM cache HIT model=%s key=%s", model, key[:12])
                     return LLMResponse(**data)
-            except Exception:
-                pass
+            except Exception as e:
+                logger.warning("LLM cache read failed for key=%s: %s", key[:12], e)
+        logger.debug("LLM cache MISS model=%s key=%s", model, key[:12])
         return None
 
-    def set(self, model: str, messages: list, temp: float, json_mode: bool, response: LLMResponse):
-        key = self._get_key(model, messages, temp, json_mode)
+    def set(self, model: str, messages: list, temp: float, json_mode: bool, response: LLMResponse, provider_fingerprint: str = "", cache_key: Optional[str] = None):
+        key = cache_key if cache_key is not None else self._get_key(model, messages, temp, json_mode, provider_fingerprint)
         cache_file = self.cache_dir / f"{key}.json"
         with open(cache_file, "w") as f:
             json.dump(response.model_dump(), f)
+        logger.debug("LLM cache STORE model=%s key=%s", model, key[:12])

package/src/devcouncil/llm/model_defaults.yaml

@@ -0,0 +1,44 @@
+openrouter:
+  spec_writer: anthropic/claude-sonnet-4.6
+  prompt_enhancer: anthropic/claude-sonnet-4.6
+  planner_a: anthropic/claude-sonnet-4.6
+  planner_b: google/gemini-2.5-pro
+  critic_a: openai/gpt-5.5
+  critic_b: anthropic/claude-opus-4.8
+  arbiter: openai/gpt-5.5
+  native_agent: anthropic/claude-sonnet-4.6
+  implementation_reviewer: openai/gpt-5.5
+  live_reviewer: openai/gpt-5.5
+vertexai:
+  spec_writer: google/gemini-2.5-flash
+  prompt_enhancer: google/gemini-2.5-flash
+  planner_a: google/gemini-2.5-flash
+  planner_b: google/gemini-2.5-flash
+  critic_a: google/gemini-2.5-flash
+  critic_b: google/gemini-2.5-flash
+  arbiter: google/gemini-2.5-flash
+  native_agent: google/gemini-2.5-flash
+  implementation_reviewer: google/gemini-2.5-flash
+  live_reviewer: google/gemini-2.5-flash
+doubleword:
+  spec_writer: deepseek/deepseek-v4
+  prompt_enhancer: deepseek/deepseek-v4
+  planner_a: deepseek/deepseek-v4
+  planner_b: deepseek/deepseek-v4
+  critic_a: deepseek/deepseek-v4
+  critic_b: deepseek/deepseek-v4
+  arbiter: deepseek/deepseek-v4
+  native_agent: deepseek/deepseek-v4
+  implementation_reviewer: deepseek/deepseek-v4
+  live_reviewer: deepseek/deepseek-v4
+ollama:
+  spec_writer: qwen2.5-coder:7b
+  prompt_enhancer: qwen2.5-coder:7b
+  planner_a: qwen2.5-coder:7b
+  planner_b: qwen2.5-coder:7b
+  critic_a: qwen2.5-coder:7b
+  critic_b: qwen2.5-coder:7b
+  arbiter: qwen2.5-coder:7b
+  native_agent: qwen2.5-coder:7b
+  implementation_reviewer: qwen2.5-coder:7b
+  live_reviewer: qwen2.5-coder:7b