pi-para 0.5.2

package/scripts/gepa/program.py

@@ -0,0 +1,99 @@
+"""DSPy Module proxy for prompt optimization.
+
+Each optimization target is wrapped in a lightweight proxy module.
+The Signature's docstring (instruction) IS the prompt being optimized.
+GEPA evolves the instruction to maximize the metric.
+
+One LLM call per eval (not 10-50 like the full agent loop).
+"""
+
+from __future__ import annotations
+
+import dspy
+
+
+class WikiTaskOutput(dspy.Signature):
+    """You are a wiki knowledge management agent for a PARA-structured knowledge base."""
+
+    task_context: str = dspy.InputField(
+        desc="Context describing the wiki task to perform"
+    )
+    output: str = dspy.OutputField(
+        desc="Structured wiki output following PARA conventions: "
+        "Topic, Key Facts, Insights, Connections with [[wikilinks]], Sources"
+    )
+
+
+class PromptProxy(dspy.Module):
+    """Lightweight proxy for prompt/instruction optimization.
+
+    The Signature's instruction text is what GEPA evolves.
+    forward() runs a single dspy.Predict call (1 LLM call).
+
+    Usage:
+        proxy = PromptProxy.from_instruction(current_prompt_text)
+        # GEPA optimizes proxy → evolved instruction = optimized prompt
+    """
+
+    def __init__(self, instruction: str | None = None):
+        super().__init__()
+
+        if instruction:
+            # Create a signature class with the instruction as docstring
+            class CustomSig(WikiTaskOutput):
+                pass
+            CustomSig.__doc__ = instruction
+            self.predict = dspy.Predict(CustomSig)
+        else:
+            self.predict = dspy.Predict(WikiTaskOutput)
+
+    def forward(self, task_context: str) -> dspy.Prediction:
+        return self.predict(task_context=task_context)
+
+    @classmethod
+    def from_instruction(cls, instruction: str) -> "PromptProxy":
+        """Create a proxy from a prompt/instruction string."""
+        return cls(instruction=instruction)
+
+    def get_instruction(self) -> str | None:
+        """Get the current instruction text."""
+        sig = getattr(self.predict, "signature", None)
+        if sig:
+            inst = getattr(sig, "instructions", None)
+            if inst and not inst.startswith("str(object"):
+                return inst
+            doc = getattr(sig, "__doc__", None)
+            if doc and not doc.startswith("str(object"):
+                return doc
+        return None
+
+
+def extract_evolved_instruction(optimized_proxy: PromptProxy) -> str | None:
+    """Extract the GEPA-evolved instruction from an optimized proxy.
+
+    After GEPA optimization, the proxy's Predict signature contains
+    the evolved instruction text. This extracts it.
+    """
+    # Check predict.signature first
+    if hasattr(optimized_proxy, "predict"):
+        pred = optimized_proxy.predict
+        for attr_name in ("extended_signature", "signature"):
+            sig = getattr(pred, attr_name, None)
+            if sig:
+                for field in ("instructions", "__doc__"):
+                    text = getattr(sig, field, None)
+                    if text and len(text) > 50 and not text.startswith("str(object"):
+                        return text
+
+    # Fallback to named_predictors
+    try:
+        for name, predictor in optimized_proxy.named_predictors():
+            sig = getattr(predictor, "signature", None)
+            if sig:
+                inst = getattr(sig, "instructions", None) or getattr(sig, "__doc__", None)
+                if inst and len(inst) > 50 and not inst.startswith("str(object"):
+                    return inst
+    except Exception:
+        pass
+
+    return None

package/scripts/gepa/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "pi-para-gepa"
+version = "0.1.0"
+description = "GEPA optimizer for pi-para prompts, tools, and skills"
+requires-python = ">=3.11"
+dependencies = [
+    "dspy>=2.6.0",
+    "anthropic>=0.52.0",
+    "openai>=1.30.0",
+    "pyyaml>=6.0.0",
+]
+
+[tool.hatch.build.targets.wheel]
+packages = ["lm_providers", "wiki_reader", "dataset", "program", "metric", "optimize"]
+include = ["*.py"]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"

package/scripts/gepa/run-all.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+# Run GEPA optimization one target at a time.
+# Default: Sonnet for task/judge (fast, cheap), Opus for reflection (smart mutations).
+# Each target invocation re-reads auth.json for fresh token.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+PROJECT_DIR="$(cd "$SCRIPT_DIR/../.." && pwd)"
+MODEL="${1:-anthropic/claude-sonnet-4-20250514}"
+REFLECTION_MODEL="${2:-anthropic/claude-opus-4-6}"
+AUTO="${3:-light}"
+
+echo "[gepa-all] Model: $MODEL | Reflection: $REFLECTION_MODEL | Budget: $AUTO"
+echo "[gepa-all] Started at $(date)"
+
+# Get target list
+TARGETS=$(cd "$PROJECT_DIR" && node dist/cli.js gepa targets 2>/dev/null | awk '{print $1}')
+
+TOTAL=$(echo "$TARGETS" | wc -l)
+CURRENT=0
+SUCCEEDED=0
+FAILED=0
+
+for TARGET in $TARGETS; do
+  CURRENT=$((CURRENT + 1))
+  echo ""
+  echo "================================================================"
+  echo "[gepa-all] [$CURRENT/$TOTAL] Optimizing: $TARGET"
+  echo "[gepa-all] $(date)"
+  echo "================================================================"
+  
+  cd "$PROJECT_DIR"
+  if node dist/cli.js gepa optimize \
+    --target "$TARGET" \
+    --model "$MODEL" \
+    --reflection-model "$REFLECTION_MODEL" \
+    --auto "$AUTO" 2>&1; then
+    SUCCEEDED=$((SUCCEEDED + 1))
+    echo "[gepa-all] ✅ $TARGET done"
+  else
+    FAILED=$((FAILED + 1))
+    echo "[gepa-all] ❌ $TARGET failed"
+  fi
+  
+  # Brief pause between targets to avoid rate limits
+  sleep 5
+done
+
+echo ""
+echo "================================================================"
+echo "[gepa-all] ALL DONE at $(date)"
+echo "[gepa-all] Succeeded: $SUCCEEDED / $TOTAL, Failed: $FAILED"
+echo "================================================================"