Upgrade GEPA from 0.0.22 to 0.1.4

CHANGELOG.md

@@ -10,6 +10,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added
 
 ### Changed
+- Upgraded GEPA dependency from `gepa==0.0.22` to `gepa==0.1.4`.
+  CodeOptiX continues to use GEPA's `InstructionProposalSignature` for reflective
+  prompt proposals. GEPA 0.1.x renames default template placeholders to
+  `<curr_param>` and `<side_info>` (from `<curr_instructions>` /
+  `<inputs_outputs_feedback>`). Custom `prompt_template` configs must use the
+  new placeholders. The MinimalGEPAProposer LLM wrapper now accepts both string
+  prompts and multimodal message lists to match GEPA 0.1.x LanguageModel.
 
 ### Deprecated
 

pyproject.toml

@@ -39,7 +39,7 @@ dependencies = [
     "numpy>=1.24.0",
     "litellm>=1.80.0",
     "tenacity>=8.0.0",
-    "gepa==0.0.22",
+    "gepa==0.1.4",
     "agent-client-protocol>=0.7.0",
 ]
 

src/codeoptix/evolution/gepa_integration.py

@@ -5,6 +5,13 @@
 
 Note: This is NOT a full GEPA framework integration. We're using GEPA's proven
 instruction proposal mechanism rather than the complete GEPA optimization engine.
+
+GEPA 0.1.x notes:
+- InstructionProposalSignature placeholders are ``<curr_param>`` and ``<side_info>``
+  (renamed from ``<curr_instructions>`` / ``<inputs_outputs_feedback>`` in 0.0.x).
+- Custom ``prompt_template`` values must include those 0.1.x placeholders.
+- LanguageModel may receive either a string prompt or a multimodal messages list;
+  CodeOptiX only feeds text feedback today, so the string path is the common case.
 """
 
 from typing import Any
@@ -25,7 +32,8 @@ def __init__(self, llm_client, config: dict[str, Any] | None = None):
 
         Args:
             llm_client: LLM client compatible with GEPA's LanguageModel interface
-            config: Configuration dictionary
+            config: Configuration dictionary. Optional ``prompt_template`` must use
+                GEPA 0.1.x placeholders ``<curr_param>`` and ``<side_info>``.
         """
         self.llm_client = llm_client
         self.config = config or {}
@@ -118,8 +126,7 @@ def _wrap_llm_client(self):
         """
         Wrap CodeOptiX LLM client to GEPA's LanguageModel interface.
 
-        GEPA expects a LanguageModel Protocol with a __call__ method that takes
-        a prompt string and returns a string response.
+        GEPA 0.1.x LanguageModel accepts ``str | list[dict]`` and returns a string.
         """
 
         class GEPALLMWrapper:
@@ -130,16 +137,20 @@ def __init__(self, llm_client):
                 if hasattr(llm_client, "config") and isinstance(llm_client.config, dict):
                     self.default_model = llm_client.config.get("model", self.default_model)
 
-            def __call__(self, prompt: str) -> str:
+            def __call__(self, prompt: str | list[dict[str, Any]]) -> str:
                 """
                 Generate response using CodeOptiX LLM client.
 
-                This implements GEPA's LanguageModel Protocol:
-                - Takes a prompt string
-                - Returns a string response
+                Implements GEPA's LanguageModel Protocol for both string prompts
+                and OpenAI-style multimodal message lists.
                 """
+                if isinstance(prompt, str):
+                    messages: list[dict[str, Any]] = [{"role": "user", "content": prompt}]
+                else:
+                    messages = prompt
+
                 response = self.llm_client.chat_completion(
-                    messages=[{"role": "user", "content": prompt}],
+                    messages=messages,
                     model=self.default_model,
                     temperature=0.7,
                 )