feat(training): release V0.8 自调参 Agent

training_server/tuning/study.py

@@ -0,0 +1,55 @@
+"""Lazy Optuna study integration used for durable trial history and pruning metadata."""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+
+class OptunaStudies:
+    def __init__(self, root: Path):
+        self.path = (root / "optuna.sqlite3").resolve()
+        self.url = f"sqlite:///{self.path}"
+
+    def _study(self, session_id: str):
+        import optuna
+
+        optuna.logging.set_verbosity(optuna.logging.WARNING)
+        return optuna.create_study(
+            study_name=f"reward-tuning-{session_id}",
+            storage=self.url,
+            direction="maximize",
+            load_if_exists=True,
+            pruner=optuna.pruners.SuccessiveHalvingPruner(
+                min_resource=300, reduction_factor=3, min_early_stopping_rate=0
+            ),
+        )
+
+    def record(
+        self,
+        session_id: str,
+        reward_config: dict[str, Any],
+        score: float,
+        eligible: bool,
+        rung: int,
+    ) -> int:
+        """Record an externally proposed Agent config through Optuna ask/tell."""
+        import optuna
+
+        study = self._study(session_id)
+        trial = study.ask()
+        trial.set_user_attr("reward_config", reward_config)
+        trial.set_user_attr("eligible", eligible)
+        trial.set_user_attr("rung", rung)
+        state = optuna.trial.TrialState.COMPLETE if eligible else optuna.trial.TrialState.PRUNED
+        study.tell(trial, score if eligible else None, state=state)
+        return trial.number
+
+    def summary(self, session_id: str) -> dict[str, Any]:
+        study = self._study(session_id)
+        completed = [trial for trial in study.trials if trial.value is not None]
+        return {
+            "studyName": study.study_name,
+            "trialCount": len(study.trials),
+            "bestValue": max((trial.value for trial in completed), default=None),
+        }