Update install.sh

.test/src/skill_test/agent/executor.py

@@ -197,8 +197,7 @@ def _load_mcp_config() -> dict[str, Any]:
                 resolved_cfg[key] = val.replace("${CLAUDE_PLUGIN_ROOT}", str(repo_root))
             elif isinstance(val, list):
                 resolved_cfg[key] = [
-                    v.replace("${CLAUDE_PLUGIN_ROOT}", str(repo_root)) if isinstance(v, str) else v
-                    for v in val
+                    v.replace("${CLAUDE_PLUGIN_ROOT}", str(repo_root)) if isinstance(v, str) else v for v in val
                 ]
             else:
                 resolved_cfg[key] = val
@@ -283,7 +282,11 @@ def _get_agent_env() -> dict[str, str]:
 
     # 2. Env vars with known prefixes override settings file values
     # Skip internal Claude Code vars that would confuse the subprocess
-    _skip_keys = {"CLAUDE_CODE_SSE_PORT", "CLAUDE_CODE_ENTRYPOINT", "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY"}
+    _skip_keys = {
+        "CLAUDE_CODE_SSE_PORT",
+        "CLAUDE_CODE_ENTRYPOINT",
+        "CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY",
+    }
     for key, value in os.environ.items():
         if key in _skip_keys:
             continue
@@ -510,7 +513,7 @@ async def run_agent(
     # Pass Databricks auth env vars to MCP server processes
     if mcp_config:
         mcp_env = {k: v for k, v in env.items() if k.startswith(("DATABRICKS_",))}
-        for server_name, server_cfg in mcp_config.items():
+        for _server_name, server_cfg in mcp_config.items():
             if "env" not in server_cfg and mcp_env:
                 server_cfg["env"] = mcp_env
 

.test/src/skill_test/optimize/agent_evaluator.py

@@ -189,15 +189,9 @@ def __init__(
         )
 
         # --- Field-based judges (fallback — when mlflow_trace is None) ---
-        self._field_correctness_judge = create_correctness_judge(
-            skill_guidelines, judge_model=judge_model
-        )
-        self._field_completeness_judge = create_completeness_judge(
-            judge_model=judge_model
-        )
-        self._field_guideline_judge = create_guideline_adherence_judge(
-            skill_guidelines, judge_model=judge_model
-        )
+        self._field_correctness_judge = create_correctness_judge(skill_guidelines, judge_model=judge_model)
+        self._field_completeness_judge = create_completeness_judge(judge_model=judge_model)
+        self._field_guideline_judge = create_guideline_adherence_judge(skill_guidelines, judge_model=judge_model)
 
         self._regression_judge = create_regression_judge(judge_model=judge_model)
 
@@ -300,8 +294,7 @@ def _evaluate(
         facts_str = "\n".join(f"- {f}" for f in facts) if facts else "None specified"
         patterns_str = (
             "\n".join(
-                f"- {p}" if isinstance(p, str) else f"- {p.get('description', p.get('pattern', ''))}"
-                for p in patterns
+                f"- {p}" if isinstance(p, str) else f"- {p.get('description', p.get('pattern', ''))}" for p in patterns
             )
             if patterns
             else "None specified"
@@ -321,8 +314,12 @@ def _evaluate(
         # Circuit breaker: after first failure, skip trace judges entirely
         # to avoid wasting API calls on a model that can't handle them.
         def _judge_with_fallback(
-            trace_judge, field_judge, *,
-            mlflow_trace, response_text, judge_name,
+            trace_judge,
+            field_judge,
+            *,
+            mlflow_trace,
+            response_text,
+            judge_name,
         ) -> JudgeFeedback:
             """Try trace-based judge, fall back to field-based on failure."""
             with self._cache_lock:
@@ -357,7 +354,8 @@ def _judge_with_fallback(
 
         # Correctness: WITH + WITHOUT (WITHOUT cached)
         correctness_with_fb = _judge_with_fallback(
-            self._trace_correctness_judge, self._field_correctness_judge,
+            self._trace_correctness_judge,
+            self._field_correctness_judge,
             mlflow_trace=with_result.mlflow_trace,
             response_text=with_response,
             judge_name="correctness_with",
@@ -366,7 +364,8 @@ def _judge_with_fallback(
             need_correctness_baseline = baseline_key not in self._baseline_correctness_cache
         if need_correctness_baseline:
             fb = _judge_with_fallback(
-                self._trace_correctness_judge, self._field_correctness_judge,
+                self._trace_correctness_judge,
+                self._field_correctness_judge,
                 mlflow_trace=without_mlflow_trace,
                 response_text=without_response,
                 judge_name="correctness_without",
@@ -379,7 +378,8 @@ def _judge_with_fallback(
 
         # Completeness: WITH + WITHOUT (WITHOUT cached)
         completeness_with_fb = _judge_with_fallback(
-            self._trace_completeness_judge, self._field_completeness_judge,
+            self._trace_completeness_judge,
+            self._field_completeness_judge,
             mlflow_trace=with_result.mlflow_trace,
             response_text=with_response,
             judge_name="completeness_with",
@@ -388,7 +388,8 @@ def _judge_with_fallback(
             need_completeness_baseline = baseline_key not in self._baseline_completeness_cache
         if need_completeness_baseline:
             fb = _judge_with_fallback(
-                self._trace_completeness_judge, self._field_completeness_judge,
+                self._trace_completeness_judge,
+                self._field_completeness_judge,
                 mlflow_trace=without_mlflow_trace,
                 response_text=without_response,
                 judge_name="completeness_without",
@@ -401,7 +402,8 @@ def _judge_with_fallback(
 
         # Guideline adherence: WITH only
         guideline_adherence_fb = _judge_with_fallback(
-            self._trace_guideline_judge, self._field_guideline_judge,
+            self._trace_guideline_judge,
+            self._field_guideline_judge,
             mlflow_trace=with_result.mlflow_trace,
             response_text=with_response,
             judge_name="guideline_adherence",
@@ -444,7 +446,10 @@ def _judge_with_fallback(
             reg_val = regression_fb.value
             if isinstance(reg_val, bool):
                 regression_penalty = 1.0 if reg_val else 0.0
-            elif isinstance(reg_val, str) and reg_val.strip().lower() in ("yes", "true"):
+            elif isinstance(reg_val, str) and reg_val.strip().lower() in (
+                "yes",
+                "true",
+            ):
                 regression_penalty = 1.0
 
         # Phase 4: Deterministic fact/pattern assertions (zero LLM cost — static spine)
@@ -481,16 +486,20 @@ def _judge_with_fallback(
         quality_composite = (correctness_with + completeness_with + guideline_adherence_score) / 3.0
         assertion_coverage = 0.5 * fact_score + 0.5 * pattern_score
 
-        final_score = max(0.0, min(1.0,
-            0.25 * effectiveness_delta
-            + 0.20 * correctness_with
-            + 0.15 * completeness_with
-            + 0.15 * guideline_adherence_score
-            + 0.10 * assertion_coverage
-            + 0.05 * execution_success
-            + 0.05 * token_efficiency
-            - 0.05 * regression_penalty
-        ))
+        final_score = max(
+            0.0,
+            min(
+                1.0,
+                0.25 * effectiveness_delta
+                + 0.20 * correctness_with
+                + 0.15 * completeness_with
+                + 0.15 * guideline_adherence_score
+                + 0.10 * assertion_coverage
+                + 0.05 * execution_success
+                + 0.05 * token_efficiency
+                - 0.05 * regression_penalty,
+            ),
+        )
 
         # Build rich side_info for GEPA reflection
         side_info: dict[str, Any] = {}

.test/src/skill_test/optimize/eval_criteria.py

@@ -92,11 +92,7 @@ def filter_by_modules(self, tool_modules: list[str]) -> "EvalCriteriaSet":
 
         Criteria with empty ``applies_to`` are always included (general-purpose).
         """
-        filtered = [
-            s
-            for s in self.skills
-            if not s.applies_to or any(m in s.applies_to for m in tool_modules)
-        ]
+        filtered = [s for s in self.skills if not s.applies_to or any(m in s.applies_to for m in tool_modules)]
         result = EvalCriteriaSet.__new__(EvalCriteriaSet)
         result.skills = filtered
         result._by_name = {s.name: s for s in filtered}
@@ -134,8 +130,7 @@ def _to_markdown(self) -> str:
             lines.append(f"- **{s.name}**: {s.description}")
         lines.append("")
         lines.append(
-            "Use the read_eval_criteria tool to load relevant criteria. "
-            "Use read_eval_reference for detailed rubrics."
+            "Use the read_eval_criteria tool to load relevant criteria. Use read_eval_reference for detailed rubrics."
         )
         return "\n".join(lines)
 
@@ -184,9 +179,7 @@ def discover_eval_criteria(
     if not base.is_dir():
         logger.debug("Eval criteria directory not found: %s", base)
         return EvalCriteriaSet([])
-    paths = sorted(
-        d for d in base.iterdir() if d.is_dir() and (d / "SKILL.md").exists()
-    )
+    paths = sorted(d for d in base.iterdir() if d.is_dir() and (d / "SKILL.md").exists())
     if paths:
         logger.info(
             "Discovered %d eval criteria: %s",

.test/src/skill_test/optimize/judges.py

@@ -87,6 +87,7 @@ def _is_rate_limit_error(exc: Exception) -> bool:
 # AI Gateway support
 # ---------------------------------------------------------------------------
 
+
 def _get_gateway_base_url() -> str | None:
     """Return the AI Gateway base URL if configured, else None.
 
@@ -645,7 +646,11 @@ def _call_judge(j):
                 fb = future.result(timeout=timeout)
             except concurrent.futures.TimeoutError:
                 fb_pool.shutdown(wait=False)
-                logger.warning("Fallback '%s' timed out after %ds, trying next", fallback_model, timeout)
+                logger.warning(
+                    "Fallback '%s' timed out after %ds, trying next",
+                    fallback_model,
+                    timeout,
+                )
                 continue
             finally:
                 fb_pool.shutdown(wait=False)
@@ -803,25 +808,15 @@ def create_trace_correctness_judge(
             CLI flag, ``GEPA_JUDGE_LM`` env var, or default.
     """
     criteria_block = eval_criteria.to_prompt(judge_model) if eval_criteria else ""
-    instructions = (
-        _TRACE_CORRECTNESS_INSTRUCTIONS_PREFIX
-        + criteria_block
-        + _TRACE_CORRECTNESS_INSTRUCTIONS_BODY
-    )
+    instructions = _TRACE_CORRECTNESS_INSTRUCTIONS_PREFIX + criteria_block + _TRACE_CORRECTNESS_INSTRUCTIONS_BODY
 
     if skill_guidelines:
-        filtered = [
-            g
-            for g in skill_guidelines
-            if any(kw in g.lower() for kw in _CORRECTNESS_KEYWORDS)
-        ]
+        filtered = [g for g in skill_guidelines if any(kw in g.lower() for kw in _CORRECTNESS_KEYWORDS)]
         if filtered:
             principles = "\n".join(f"- {g}" for g in filtered)
             instructions += f"\n\n## Domain Correctness Principles\n{principles}\n"
 
-    model_uri, inference_params = _to_judge_model_and_params(
-        judge_model or DEFAULT_JUDGE_LM
-    )
+    model_uri, inference_params = _to_judge_model_and_params(judge_model or DEFAULT_JUDGE_LM)
     return make_judge(
         name="trace_correctness",
         model=model_uri,
@@ -844,15 +839,9 @@ def create_trace_completeness_judge(
         judge_model: LLM model for the judge.
     """
     criteria_block = eval_criteria.to_prompt(judge_model) if eval_criteria else ""
-    instructions = (
-        _TRACE_COMPLETENESS_INSTRUCTIONS_PREFIX
-        + criteria_block
-        + _TRACE_COMPLETENESS_INSTRUCTIONS_BODY
-    )
+    instructions = _TRACE_COMPLETENESS_INSTRUCTIONS_PREFIX + criteria_block + _TRACE_COMPLETENESS_INSTRUCTIONS_BODY
 
-    model_uri, inference_params = _to_judge_model_and_params(
-        judge_model or DEFAULT_JUDGE_LM
-    )
+    model_uri, inference_params = _to_judge_model_and_params(judge_model or DEFAULT_JUDGE_LM)
     return make_judge(
         name="trace_completeness",
         model=model_uri,
@@ -879,19 +868,13 @@ def create_trace_guideline_judge(
         judge_model: LLM model for the judge.
     """
     criteria_block = eval_criteria.to_prompt(judge_model) if eval_criteria else ""
-    instructions = (
-        _TRACE_GUIDELINE_INSTRUCTIONS_PREFIX
-        + criteria_block
-        + _TRACE_GUIDELINE_INSTRUCTIONS_BODY
-    )
+    instructions = _TRACE_GUIDELINE_INSTRUCTIONS_PREFIX + criteria_block + _TRACE_GUIDELINE_INSTRUCTIONS_BODY
 
     if skill_guidelines:
         principles = "\n".join(f"- {g}" for g in skill_guidelines)
         instructions += f"\n\n## Required Guidelines\n{principles}\n"
 
-    model_uri, inference_params = _to_judge_model_and_params(
-        judge_model or DEFAULT_JUDGE_LM
-    )
+    model_uri, inference_params = _to_judge_model_and_params(judge_model or DEFAULT_JUDGE_LM)
     return make_judge(
         name="trace_guideline_adherence",
         model=model_uri,