Improve AI daily report operations and dedupe observability

ai_daily_report/runner.py

@@ -9,6 +9,7 @@ from .clients import BlogApiClient, OpenAICompatibleClient, fetch_text as defaul
 from .config import load_pipeline_config, load_source_configs
 from .env import load_env, resolve_blog_token, resolve_llm_config
 from .models import SourceConfig
+from .observability import LlmCallObserver, summarize_observed_calls
 from .pipeline import run_stage0_to_stage8
 from .publish import load_published_urls, update_published_urls
 from .sources.registry import get_source_fetcher
@@ -135,15 +136,33 @@ def run_daily_report(
     else:
         raise ValueError("source_mode must be 'mock' or 'live'")
 
+    llm_observability_config = pipeline_config.get("llm_observability", {}) or {}
+    llm_observers: list[LlmCallObserver] = []
+    observe_llm = bool(llm_observability_config.get("enabled", True))
+    prompt_preview_chars = int(llm_observability_config.get("prompt_preview_chars", 500))
+    response_preview_chars = int(llm_observability_config.get("response_preview_chars", 500))
+
+    def maybe_observe(stage: str, call):
+        if not observe_llm:
+            return call
+        observer = LlmCallObserver(
+            call=call,
+            stage=stage,
+            prompt_preview_chars=prompt_preview_chars,
+            response_preview_chars=response_preview_chars,
+        )
+        llm_observers.append(observer)
+        return observer
+
     if llm_mode == "mock":
-        semantic_llm_call = _mock_semantic_llm
-        rewrite_llm_call = _mock_rewrite_llm
-        guide_llm_call = _mock_guide_llm
+        semantic_llm_call = maybe_observe("stage3", _mock_semantic_llm)
+        rewrite_llm_call = maybe_observe("stage4", _mock_rewrite_llm)
+        guide_llm_call = maybe_observe("stage6", _mock_guide_llm)
     elif llm_mode == "live":
         llm_client = llm_client_factory(**resolve_llm_config(env))
-        semantic_llm_call = llm_client.chat
-        rewrite_llm_call = llm_client.chat
-        guide_llm_call = llm_client.chat
+        semantic_llm_call = maybe_observe("stage3", llm_client.chat)
+        rewrite_llm_call = maybe_observe("stage4", llm_client.chat)
+        guide_llm_call = maybe_observe("stage6", llm_client.chat)
     else:
         raise ValueError("llm_mode must be 'mock' or 'live'")
 
@@ -182,6 +201,9 @@ def run_daily_report(
             max_age_days=cross_day_max_age_days,
         )
 
+    llm_observability_report = summarize_observed_calls(llm_observers)
+    result["reports"]["llm_observability"] = llm_observability_report
+
     run_dir = out_dir / run_date
     run_dir.mkdir(parents=True, exist_ok=True)
     (run_dir / "blog_markdown.md").write_text(result["markdown"], encoding="utf-8")