claude-evolve 1.9.8 → 1.9.10

package/lib/ai-cli.sh

@@ -145,13 +145,13 @@ $prompt"
       ;;
     glm-openrouter)
       local ai_output
-      ai_output=$(timeout -k 30 600 opencode -m openrouter/z-ai/glm-4.6 run "$prompt" 2>&1)
+      ai_output=$(timeout -k 30 600 opencode -m openrouter/z-ai/glm-4.7 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     glm-zai)
       # GLM -- can be slow sometimes
       local ai_output
-      ai_output=$(timeout -k 30 1800 opencode -m zai-coding-plan/glm-4.6 run "$prompt" 2>&1)
+      ai_output=$(timeout -k 30 1800 opencode -m zai-coding-plan/glm-4.7 run "$prompt" 2>&1)
       local ai_exit_code=$?
       ;;
     deepseek-openrouter)

package/lib/evolve_worker.py

@@ -50,6 +50,7 @@ class Config:
     memory_limit_mb: int = 0
     timeout_seconds: int = 600
     max_candidates: int = 5
+    max_validation_retries: int = 3  # Max attempts to fix validation errors (if validator.py exists)
     # Retry configuration with exponential backoff
     max_rounds: int = 10
     initial_wait: int = 60
@@ -202,6 +203,136 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
         except Exception:
             return False
 
+    def _find_validator(self) -> Optional[Path]:
+        """
+        Auto-detect validator.py in the evolution directory.
+        No config required - if validator.py exists, we use it.
+        """
+        validator_path = Path(self.config.evolution_dir) / "validator.py"
+        if validator_path.exists():
+            return validator_path
+        return None
+
+    def _run_validator(self, candidate_id: str) -> Tuple[bool, Dict[str, Any]]:
+        """
+        Run the validator (fast smoke test) before full evaluation.
+
+        AIDEV-NOTE: Auto-detects validator.py in evolution directory.
+        Returns exit code 0 on success, non-zero on failure.
+        Resilient to any output format - handles JSON, plain text, or nothing.
+
+        Returns:
+            Tuple of (success, error_info_dict)
+            - success: True if validation passed
+            - error_info: Dict with whatever info we could extract from output
+        """
+        validator_path = self._find_validator()
+        if not validator_path:
+            return True, {}  # No validator found, skip
+
+        cmd = [self.config.python_cmd, str(validator_path), candidate_id]
+        log(f"Running validator: {' '.join(cmd)}")
+
+        try:
+            result = subprocess.run(
+                cmd,
+                capture_output=True,
+                text=True,
+                timeout=30,  # Validator should be fast (~3 seconds)
+                cwd=self.config.evolution_dir
+            )
+
+            # Combine stdout and stderr for full context
+            stdout = result.stdout.strip() if result.stdout else ""
+            stderr = result.stderr.strip() if result.stderr else ""
+            combined_output = f"{stdout}\n{stderr}".strip()
+
+            # Try to extract structured info, but be resilient to any format
+            error_info = {'raw_output': combined_output}
+
+            # Try to parse JSON from stdout (validator may output JSON)
+            if stdout.startswith('{'):
+                try:
+                    parsed = json.loads(stdout)
+                    if isinstance(parsed, dict):
+                        error_info.update(parsed)
+                except json.JSONDecodeError:
+                    pass  # Not valid JSON, that's fine
+
+            # If no structured error, use the raw output
+            if 'error' not in error_info and combined_output:
+                error_info['error'] = combined_output
+
+            if result.returncode == 0:
+                log("Validation passed")
+                return True, error_info
+            else:
+                error_type = error_info.get('error_type', 'validation_failed')
+                log_warn(f"Validation failed: {error_type}")
+                return False, error_info
+
+        except subprocess.TimeoutExpired:
+            log_error("Validator timed out")
+            return False, {'error': 'Validator timed out after 30 seconds', 'error_type': 'timeout'}
+        except Exception as e:
+            log_error(f"Validator error: {e}")
+            return False, {'error': str(e), 'error_type': 'exception'}
+
+    def _build_fix_prompt(self, candidate: Candidate, target_basename: str, error_info: Dict[str, Any]) -> str:
+        """
+        Build AI prompt to fix validation errors.
+
+        AIDEV-NOTE: Resilient to any error_info structure - uses whatever is available.
+        """
+        prompt = f"""{get_git_protection_warning()}
+
+The code in {target_basename} failed validation. Please fix the errors and try again.
+
+## Validator Output
+
+"""
+        # Include whatever structured fields we have
+        if error_info.get('error_type'):
+            prompt += f"**Error Type:** {error_info['error_type']}\n\n"
+
+        if error_info.get('error'):
+            prompt += f"**Error:**\n{error_info['error']}\n\n"
+
+        if error_info.get('suggestion'):
+            prompt += f"**Suggested Fix:**\n{error_info['suggestion']}\n\n"
+
+        if error_info.get('traceback'):
+            tb = error_info['traceback']
+            # Truncate if too long
+            if len(tb) > 1500:
+                tb = "..." + tb[-1500:]
+            prompt += f"**Traceback:**\n```\n{tb}\n```\n\n"
+
+        # If we only have raw output (no structured fields), show that
+        if not any(error_info.get(k) for k in ('error', 'error_type', 'suggestion', 'traceback')):
+            raw = error_info.get('raw_output', 'No output captured')
+            # Truncate if needed
+            if len(raw) > 2000:
+                raw = raw[:2000] + "\n... (truncated)"
+            prompt += f"```\n{raw}\n```\n\n"
+
+        prompt += f"""## Instructions
+
+1. Read the file {target_basename} to understand the current code
+2. Identify the issue based on the validator output above
+3. Fix the code to resolve the validation error
+4. The fix should still implement: {candidate.description}
+
+**CRITICAL:** Make sure to actually fix the error. Do not just add comments or make cosmetic changes.
+
+To help debug, you can run the validator yourself:
+```
+python validator.py {target_basename}
+```
+"""
+
+        return prompt
+
     def _run_evaluator(self, candidate_id: str, is_baseline: bool) -> Tuple[Optional[float], Dict[str, Any]]:
         """
         Run the evaluator.
@@ -347,6 +478,52 @@ CRITICAL: If you do not know how to implement what was asked for, or if the requ
                     csv.update_candidate_status(candidate.id, 'pending')
                 return 0  # Will retry
 
+            # Run validator with retry loop
+            # AIDEV-NOTE: Validator catches structural errors before expensive full evaluation.
+            # If validation fails, we give the AI feedback and ask it to fix the code.
+            validation_passed = False
+            for validation_attempt in range(self.config.max_validation_retries + 1):
+                valid, error_info = self._run_validator(candidate.id)
+
+                if valid:
+                    validation_passed = True
+                    break
+
+                if validation_attempt >= self.config.max_validation_retries:
+                    log_error(f"Validation failed after {self.config.max_validation_retries} fix attempts")
+                    break
+
+                # Ask AI to fix the validation error
+                log(f"Validation failed (attempt {validation_attempt + 1}), asking AI to fix...")
+                fix_prompt = self._build_fix_prompt(candidate, target_file.name, error_info)
+                success, fix_model = self._call_ai_with_backoff(fix_prompt, target_file)
+
+                if not success:
+                    log_error("AI failed to fix validation error")
+                    break
+
+                # Record that we used an additional model call for fixing
+                if fix_model:
+                    with EvolutionCSV(self.config.csv_path) as csv:
+                        current_llm = csv.get_candidate_info(candidate.id).get('run-LLM', '')
+                        new_llm = f"{current_llm}+{fix_model}" if current_llm else fix_model
+                        csv.update_candidate_field(candidate.id, 'run-LLM', new_llm)
+
+                # Re-check syntax after fix
+                if not self._check_syntax(target_file):
+                    log_error("Fix introduced syntax error")
+                    # Don't break - try again if we have retries left
+
+            if not validation_passed:
+                # Validation failed after all retries
+                with EvolutionCSV(self.config.csv_path) as csv:
+                    csv.update_candidate_status(candidate.id, 'failed-validation')
+                    # Store the last error for debugging
+                    if error_info:
+                        error_summary = f"{error_info.get('error_type', 'unknown')}: {error_info.get('error', '')[:100]}"
+                        csv.update_candidate_field(candidate.id, 'validation_error', error_summary)
+                return 1
+
         # Run evaluator
         log("Running evaluator...")
         score, json_data = self._run_evaluator(candidate.id, is_baseline)
@@ -470,6 +647,7 @@ def load_config_from_yaml(config_path: Optional[str] = None) -> Config:
         memory_limit_mb=data.get('memory_limit_mb', 0),
         timeout_seconds=data.get('timeout_seconds', 600),
         max_candidates=data.get('worker_max_candidates', 5),
+        max_validation_retries=data.get('max_validation_retries', 3),
         max_rounds=ideation.get('max_rounds', 10),
         initial_wait=ideation.get('initial_wait', 60),
         max_wait=ideation.get('max_wait', 600)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-evolve",
-  "version": "1.9.8",
+  "version": "1.9.10",
   "bin": {
     "claude-evolve": "bin/claude-evolve",
     "claude-evolve-main": "bin/claude-evolve-main",