Skip to content

Personal AI, On Personal Devices

Python
11.3 KB
Raw Blame
"""SkillOptimizer — per-skill DSPy/GEPA optimization wrapper (Plan 2A).

Buckets traces by skill name, runs the underlying optimizer on each skill's
bucket, and writes the result as a sidecar overlay file in
``~/.openjarvis/learning/skills/<skill-name>/optimized.toml``.

The actual DSPy/GEPA invocation is done in ``_run_dspy`` / ``_run_gepa``,
which are isolated for easy mocking in tests.  In Plan 2A these are
deliberately minimal — they call the existing optimizer modules and extract
the description + few-shot examples.  Plan 2B will measure the impact via
benchmarks.
"""

from __future__ import annotations

import logging
from dataclasses import dataclass, field
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Dict, List, Optional

from openjarvis.core.paths import get_config_dir
from openjarvis.core.types import Trace, TraceStep
from openjarvis.skills.manager import SkillManager
from openjarvis.skills.overlay import SkillOverlay, write_overlay

LOGGER = logging.getLogger(__name__)


@dataclass(slots=True)
class SkillOptimizationResult:
    """Result of optimizing a single skill."""

    skill_name: str
    status: str  # "optimized" | "skipped" | "error"
    trace_count: int = 0
    overlay_path: Optional[Path] = None
    error: str = ""


@dataclass(slots=True)
class _OptimizerOutput:
    """Internal: extracted output from the underlying DSPy/GEPA optimizer."""

    description: str = ""
    few_shot: List[Dict[str, str]] = field(default_factory=list)


class SkillOptimizer:
    """Per-skill optimization wrapper around DSPyAgentOptimizer / GEPA.

    Parameters
    ----------
    min_traces_per_skill:
        Minimum trace count for a skill to be eligible for optimization.
    optimizer:
        ``"dspy"`` or ``"gepa"``.  Determines which underlying optimizer is
        called inside ``_run_dspy`` / ``_run_gepa``.
    """

    def __init__(
        self,
        *,
        min_traces_per_skill: int = 20,
        optimizer: str = "dspy",
    ) -> None:
        self._min_traces = min_traces_per_skill
        self._optimizer = optimizer

    # ------------------------------------------------------------------
    # Public API
    # ------------------------------------------------------------------

    def optimize(
        self,
        trace_store: Any,
        skill_manager: SkillManager,
        *,
        overlay_dir: Optional[Path] = None,
    ) -> Dict[str, SkillOptimizationResult]:
        """Run the per-skill optimization loop.

        Returns a dict mapping skill name to result.  Writes overlay TOML
        files for each skill that produced output.
        """
        if overlay_dir is None:
            # Try config first; fall back to the default tree.
            try:
                from openjarvis.core.config import load_config

                cfg = load_config()
                cfg_dir = getattr(
                    getattr(cfg.learning, "skills", None),
                    "overlay_dir",
                    None,
                )
                if cfg_dir:
                    overlay_dir = Path(cfg_dir).expanduser()
            except Exception:
                pass
            if overlay_dir is None:
                overlay_dir = Path(
                    str(get_config_dir() / "learning" / "skills")
                ).expanduser()
        overlay_dir = Path(overlay_dir).expanduser()

        traces = trace_store.list_traces(limit=10000)
        buckets = self._bucket_traces_by_skill(traces)

        results: Dict[str, SkillOptimizationResult] = {}
        for skill_name, skill_traces in buckets.items():
            if len(skill_traces) < self._min_traces:
                results[skill_name] = SkillOptimizationResult(
                    skill_name=skill_name,
                    status="skipped",
                    trace_count=len(skill_traces),
                )
                continue

            try:
                if self._optimizer == "gepa":
                    output = self._run_gepa(skill_name, skill_traces)
                else:
                    output = self._run_dspy(skill_name, skill_traces)
            except Exception as exc:
                LOGGER.warning("Skill optimizer failed for '%s': %s", skill_name, exc)
                results[skill_name] = SkillOptimizationResult(
                    skill_name=skill_name,
                    status="error",
                    trace_count=len(skill_traces),
                    error=str(exc),
                )
                continue

            overlay = SkillOverlay(
                skill_name=skill_name,
                optimizer=self._optimizer,
                optimized_at=datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
                trace_count=len(skill_traces),
                description=output.description,
                few_shot=list(output.few_shot),
            )
            path = write_overlay(overlay, overlay_dir)
            results[skill_name] = SkillOptimizationResult(
                skill_name=skill_name,
                status="optimized",
                trace_count=len(skill_traces),
                overlay_path=path,
            )

        return results

    # ------------------------------------------------------------------
    # Bucketing
    # ------------------------------------------------------------------

    def _bucket_traces_by_skill(self, traces: List[Trace]) -> Dict[str, List[Trace]]:
        """Group traces by the skill names they invoked.

        A trace is added to a skill's bucket once for each unique skill it
        invoked.  Traces with no skill-tagged tool calls are dropped entirely.
        """
        buckets: Dict[str, List[Trace]] = {}
        for trace in traces:
            skill_names = self._extract_skill_names(trace)
            for name in skill_names:
                buckets.setdefault(name, []).append(trace)
        return buckets

    @staticmethod
    def _extract_skill_names(trace: Trace) -> List[str]:
        """Return the unique skill names invoked in *trace*'s tool steps."""
        seen: List[str] = []
        steps = getattr(trace, "steps", []) or []
        for step in steps:
            metadata: Dict[str, Any] = {}
            if isinstance(step, dict):
                metadata = step.get("metadata", {}) or {}
            elif isinstance(step, TraceStep):
                metadata = step.metadata or {}
            else:
                metadata = getattr(step, "metadata", {}) or {}
            skill_name = metadata.get("skill")
            if skill_name and skill_name not in seen:
                seen.append(skill_name)
        return seen

    # ------------------------------------------------------------------
    # Underlying optimizers (mockable in tests)
    # ------------------------------------------------------------------

    def _run_dspy(
        self,
        skill_name: str,
        skill_traces: List[Trace],
    ) -> _OptimizerOutput:
        """Run DSPyAgentOptimizer on the bucket and extract description + few-shot.

        In Plan 2A this is intentionally simple — it pulls the few highest-
        feedback traces as candidate few-shot examples and uses the agent
        optimizer's output `system_prompt` as the new skill description if
        non-empty.  Plan 2B will measure and refine.
        """
        from openjarvis.core.config import DSPyOptimizerConfig
        from openjarvis.learning.agents.dspy_optimizer import DSPyAgentOptimizer

        cfg = DSPyOptimizerConfig(
            min_traces=max(1, self._min_traces),
            max_bootstrapped_demos=4,
            max_labeled_demos=4,
        )
        optimizer = DSPyAgentOptimizer(cfg)

        # Build a tiny in-memory trace store shim for the underlying optimizer
        class _BucketStore:
            def __init__(self, _traces):
                self._traces = _traces

            def list_traces(self, *, limit=100, **_kwargs):
                return list(self._traces[:limit])

        try:
            optimization_result = optimizer.optimize(_BucketStore(skill_traces)) or {}
        except Exception as exc:
            LOGGER.warning(
                "DSPyAgentOptimizer raised for '%s' (using empty output): %s",
                skill_name,
                exc,
            )
            optimization_result = {}

        updates = optimization_result.get("config_updates") or {}
        description = str(updates.get("system_prompt", "")) if updates else ""
        few_shot_raw = updates.get("few_shot_examples", []) if updates else []
        few_shot: List[Dict[str, str]] = []
        for item in few_shot_raw or []:
            if isinstance(item, dict):
                few_shot.append(
                    {
                        "input": str(item.get("input", "")),
                        "output": str(item.get("output", "")),
                    }
                )

        # Fallback: if DSPy produced nothing, derive few-shot from top traces
        if not few_shot:
            top = sorted(
                skill_traces,
                key=lambda t: t.feedback if t.feedback is not None else 0.0,
                reverse=True,
            )[:4]
            for tr in top:
                if tr.query and tr.result:
                    few_shot.append({"input": tr.query, "output": tr.result})

        return _OptimizerOutput(description=description, few_shot=few_shot)

    def _run_gepa(
        self,
        skill_name: str,
        skill_traces: List[Trace],
    ) -> _OptimizerOutput:
        """Run GEPAAgentOptimizer on the bucket.  Same shape as _run_dspy."""
        from openjarvis.core.config import GEPAOptimizerConfig
        from openjarvis.learning.agents.gepa_optimizer import GEPAAgentOptimizer

        cfg = GEPAOptimizerConfig(min_traces=max(1, self._min_traces))
        optimizer = GEPAAgentOptimizer(cfg)

        class _BucketStore:
            def __init__(self, _traces):
                self._traces = _traces

            def list_traces(self, *, limit=100, **_kwargs):
                return list(self._traces[:limit])

        try:
            updates = optimizer.optimize(_BucketStore(skill_traces)) or {}
        except Exception as exc:
            LOGGER.warning(
                "GEPAAgentOptimizer raised for '%s' (using empty output): %s",
                skill_name,
                exc,
            )
            updates = {}

        description = str(updates.get("system_prompt", "")) if updates else ""

        # Parse GEPA's few_shot_examples output (matches DSPy path semantics)
        few_shot_raw = updates.get("few_shot_examples", []) if updates else []
        few_shot: List[Dict[str, str]] = []
        for item in few_shot_raw or []:
            if isinstance(item, dict):
                few_shot.append(
                    {
                        "input": str(item.get("input", "")),
                        "output": str(item.get("output", "")),
                    }
                )

        # Fallback: if GEPA produced nothing, derive few-shot from top traces
        if not few_shot:
            top = sorted(
                skill_traces,
                key=lambda t: t.feedback if t.feedback is not None else 0.0,
                reverse=True,
            )[:4]
            for tr in top:
                if tr.query and tr.result:
                    few_shot.append({"input": tr.query, "output": tr.result})
        return _OptimizerOutput(description=description, few_shot=few_shot)


__all__ = ["SkillOptimizer", "SkillOptimizationResult"]