pdd-cli 0.0.45__py3-none-any.whl → 0.0.90__py3-none-any.whl

pdd/llm_invoke.py

@@ -5,6 +5,8 @@ import os
 import pandas as pd
 import litellm
 import logging # ADDED FOR DETAILED LOGGING
+import importlib.resources
+from litellm.caching.caching import Cache  # Fix for LiteLLM v1.75.5+
 
 # --- Configure Standard Python Logging ---
 logger = logging.getLogger("pdd.llm_invoke")
@@ -24,6 +26,16 @@ litellm_logger = logging.getLogger("litellm")
 litellm_log_level = os.getenv("LITELLM_LOG_LEVEL", "WARNING" if PRODUCTION_MODE else "INFO")
 litellm_logger.setLevel(getattr(logging, litellm_log_level, logging.WARNING))
 
+# Ensure LiteLLM drops provider-unsupported params instead of erroring
+# This prevents failures like UnsupportedParamsError for OpenAI gpt-5-* when
+# passing generic params (e.g., reasoning_effort) not accepted by that API path.
+try:
+    _drop_params_env = os.getenv("LITELLM_DROP_PARAMS", "true")
+    litellm.drop_params = str(_drop_params_env).lower() in ("1", "true", "yes", "on")
+except Exception:
+    # Be conservative: default to True even if env parsing fails
+    litellm.drop_params = True
+
 # Add a console handler if none exists
 if not logger.handlers:
     console_handler = logging.StreamHandler()
@@ -69,7 +81,7 @@ import json
 # from rich import print as rprint # Replaced with logger
 from dotenv import load_dotenv
 from pathlib import Path
-from typing import Optional, Dict, List, Any, Type, Union
+from typing import Optional, Dict, List, Any, Type, Union, Tuple
 from pydantic import BaseModel, ValidationError
 import openai  # Import openai for exception handling as LiteLLM maps to its types
 from langchain_core.prompts import PromptTemplate
@@ -112,6 +124,22 @@ def _is_wsl_environment() -> bool:
         return False
 
 
+def _openai_responses_supports_response_format() -> bool:
+    """Detect if current OpenAI Python SDK supports `response_format` on Responses.create.
+
+    Returns True if the installed SDK exposes a `response_format` parameter on
+    `openai.resources.responses.Responses.create`, else False. This avoids
+    sending unsupported kwargs and triggering TypeError at runtime.
+    """
+    try:
+        import inspect
+        from openai.resources.responses import Responses
+        sig = inspect.signature(Responses.create)
+        return "response_format" in sig.parameters
+    except Exception:
+        return False
+
+
 def _get_environment_info() -> Dict[str, str]:
     """
     Get environment information for debugging and error reporting.
@@ -186,16 +214,80 @@ if PROJECT_ROOT is None: # Fallback to CWD if no method succeeded
 
 ENV_PATH = PROJECT_ROOT / ".env"
 # --- Determine LLM_MODEL_CSV_PATH ---
-# Prioritize ~/.pdd/llm_model.csv
+# Prioritize ~/.pdd/llm_model.csv, then a project .pdd from the current CWD,
+# then PROJECT_ROOT (which may be set from PDD_PATH), else fall back to package.
 user_pdd_dir = Path.home() / ".pdd"
 user_model_csv_path = user_pdd_dir / "llm_model.csv"
 
+def _detect_project_root_from_cwd(max_levels: int = 5) -> Path:
+    """Search upwards from the current working directory for common project markers.
+
+    This intentionally ignores PDD_PATH to support CLI invocations that set
+    PDD_PATH to the installed package location. We want to honor a real project
+    checkout's .pdd/llm_model.csv when running inside it.
+    """
+    try:
+        current_dir = Path.cwd().resolve()
+        for _ in range(max_levels):
+            if (
+                (current_dir / ".git").exists()
+                or (current_dir / "pyproject.toml").exists()
+                or (current_dir / "data").is_dir()
+                or (current_dir / ".env").exists()
+            ):
+                return current_dir
+            parent = current_dir.parent
+            if parent == current_dir:
+                break
+            current_dir = parent
+    except Exception:
+        pass
+    return Path.cwd().resolve()
+
+# Resolve candidates
+project_root_from_cwd = _detect_project_root_from_cwd()
+project_csv_from_cwd = project_root_from_cwd / ".pdd" / "llm_model.csv"
+project_csv_from_env = PROJECT_ROOT / ".pdd" / "llm_model.csv"
+
+# Detect whether PDD_PATH points to the installed package directory. If so,
+# don't prioritize it over the real project from CWD.
+try:
+    _installed_pkg_root = importlib.resources.files('pdd')
+    # importlib.resources.files returns a Traversable; get a FS path string if possible
+    try:
+        _installed_pkg_root_path = Path(str(_installed_pkg_root))
+    except Exception:
+        _installed_pkg_root_path = None
+except Exception:
+    _installed_pkg_root_path = None
+
+def _is_env_path_package_dir(env_path: Path) -> bool:
+    try:
+        if _installed_pkg_root_path is None:
+            return False
+        env_path = env_path.resolve()
+        pkg_path = _installed_pkg_root_path.resolve()
+        # Treat equal or subpath as package dir
+        return env_path == pkg_path or str(env_path).startswith(str(pkg_path))
+    except Exception:
+        return False
+
+# Selection order
 if user_model_csv_path.is_file():
     LLM_MODEL_CSV_PATH = user_model_csv_path
     logger.info(f"Using user-specific LLM model CSV: {LLM_MODEL_CSV_PATH}")
+elif (not _is_env_path_package_dir(PROJECT_ROOT)) and project_csv_from_env.is_file():
+    # Honor an explicitly-set PDD_PATH pointing to a real project directory
+    LLM_MODEL_CSV_PATH = project_csv_from_env
+    logger.info(f"Using project-specific LLM model CSV (from PDD_PATH): {LLM_MODEL_CSV_PATH}")
+elif project_csv_from_cwd.is_file():
+    # Otherwise, prefer the project relative to the current working directory
+    LLM_MODEL_CSV_PATH = project_csv_from_cwd
+    logger.info(f"Using project-specific LLM model CSV (from CWD): {LLM_MODEL_CSV_PATH}")
 else:
-    LLM_MODEL_CSV_PATH = PROJECT_ROOT / "data" / "llm_model.csv"
-    logger.info(f"Using project LLM model CSV: {LLM_MODEL_CSV_PATH}")
+    # Neither exists, we'll use a marker path that _load_model_data will handle
+    LLM_MODEL_CSV_PATH = None
+    logger.info("No local LLM model CSV found, will use package default")
 # ---------------------------------
 
 # Load environment variables from .env file
@@ -225,6 +317,7 @@ if GCS_HMAC_SECRET_ACCESS_KEY:
     GCS_HMAC_SECRET_ACCESS_KEY = GCS_HMAC_SECRET_ACCESS_KEY.strip()
 
 cache_configured = False
+configured_cache = None  # Store the configured cache instance for restoration
 
 if GCS_BUCKET_NAME and GCS_HMAC_ACCESS_KEY_ID and GCS_HMAC_SECRET_ACCESS_KEY:
     # Store original AWS credentials before overwriting for GCS cache setup
@@ -238,12 +331,13 @@ if GCS_BUCKET_NAME and GCS_HMAC_ACCESS_KEY_ID and GCS_HMAC_SECRET_ACCESS_KEY:
         os.environ['AWS_SECRET_ACCESS_KEY'] = GCS_HMAC_SECRET_ACCESS_KEY
         # os.environ['AWS_REGION_NAME'] = GCS_REGION_NAME  # Uncomment if needed
 
-        litellm.cache = litellm.Cache(
+        configured_cache = Cache(
             type="s3",
             s3_bucket_name=GCS_BUCKET_NAME,
             s3_region_name=GCS_REGION_NAME, # Pass region explicitly to cache
             s3_endpoint_url=GCS_ENDPOINT_URL,
         )
+        litellm.cache = configured_cache
         logger.info(f"LiteLLM cache configured for GCS bucket (S3 compatible): {GCS_BUCKET_NAME}")
         cache_configured = True
 
@@ -268,15 +362,22 @@ if GCS_BUCKET_NAME and GCS_HMAC_ACCESS_KEY_ID and GCS_HMAC_SECRET_ACCESS_KEY:
         elif 'AWS_REGION_NAME' in os.environ:
             pass # Or just leave it if the temporary setting wasn't done/needed
 
+# Check if caching is disabled via environment variable
+if os.getenv("LITELLM_CACHE_DISABLE") == "1":
+    logger.info("LiteLLM caching disabled via LITELLM_CACHE_DISABLE=1")
+    litellm.cache = None
+    cache_configured = True
+
 if not cache_configured:
     try:
-        # Try SQLite-based cache as a fallback
+        # Try disk-based cache as a fallback
         sqlite_cache_path = PROJECT_ROOT / "litellm_cache.sqlite"
-        litellm.cache = litellm.Cache(type="sqlite", cache_path=str(sqlite_cache_path))
-        logger.info(f"LiteLLM SQLite cache configured at {sqlite_cache_path}")
+        configured_cache = Cache(type="disk", disk_cache_dir=str(sqlite_cache_path))
+        litellm.cache = configured_cache
+        logger.info(f"LiteLLM disk cache configured at {sqlite_cache_path}")
         cache_configured = True
     except Exception as e2:
-        warnings.warn(f"Failed to configure LiteLLM SQLite cache: {e2}. Caching is disabled.")
+        warnings.warn(f"Failed to configure LiteLLM disk cache: {e2}. Caching is disabled.")
         litellm.cache = None
 
 if not cache_configured:
@@ -314,29 +415,49 @@ def _litellm_success_callback(
         cost_val = litellm.completion_cost(completion_response=completion_response)
         calculated_cost = cost_val if cost_val is not None else 0.0
     except Exception as e1:
-        # Attempt 2: If response object failed (e.g., missing provider in model name),
-        # try again using explicit model from kwargs and tokens from usage.
-        # This is often needed for batch completion items.
+        # Attempt 2: Compute via tokens and model mapping. If LiteLLM mapping is
+        # missing or API differs, fall back to CSV rates in _MODEL_RATE_MAP.
         logger.debug(f"Attempting cost calculation with fallback method: {e1}")
         try:
-            model_name = kwargs.get("model") # Get original model name from input kwargs
+            model_name = kwargs.get("model")
             if model_name and usage:
-                prompt_tokens = getattr(usage, 'prompt_tokens', 0)
-                completion_tokens = getattr(usage, 'completion_tokens', 0)
-                cost_val = litellm.completion_cost(
-                    model=model_name,
-                    prompt_tokens=prompt_tokens,
-                    completion_tokens=completion_tokens
-                )
-                calculated_cost = cost_val if cost_val is not None else 0.0
+                in_tok = getattr(usage, 'prompt_tokens', None)
+                out_tok = getattr(usage, 'completion_tokens', None)
+                # Some providers may use 'input_tokens'/'output_tokens'
+                if in_tok is None:
+                    in_tok = getattr(usage, 'input_tokens', 0)
+                if out_tok is None:
+                    out_tok = getattr(usage, 'output_tokens', 0)
+
+                # Try LiteLLM helper (arg names vary across versions)
+                try:
+                    cost_val = litellm.completion_cost(
+                        model=model_name,
+                        prompt_tokens=in_tok,
+                        completion_tokens=out_tok,
+                    )
+                    calculated_cost = cost_val if cost_val is not None else 0.0
+                except TypeError:
+                    # Older/newer versions may require input/output token names
+                    try:
+                        cost_val = litellm.completion_cost(
+                            model=model_name,
+                            input_tokens=in_tok,
+                            output_tokens=out_tok,
+                        )
+                        calculated_cost = cost_val if cost_val is not None else 0.0
+                    except Exception as e3:
+                        # Final fallback: compute using CSV rates
+                        rates = _MODEL_RATE_MAP.get(str(model_name))
+                        if rates is not None:
+                            in_rate, out_rate = rates
+                            calculated_cost = (float(in_tok or 0) * in_rate + float(out_tok or 0) * out_rate) / 1_000_000.0
+                        else:
+                            calculated_cost = 0.0
+                        logger.debug(f"Cost calculation failed with LiteLLM token API; used CSV rates if available. Detail: {e3}")
             else:
-                # If we can't get model name or usage, fallback to 0
                 calculated_cost = 0.0
-                # Optional: Log the original error e1 if needed
-                # logger.warning(f"[Callback WARN] Failed to calculate cost with response object ({e1}) and fallback failed.")
         except Exception as e2:
-            # Optional: Log secondary error e2 if needed
-            # logger.warning(f"[Callback WARN] Failed to calculate cost with fallback method: {e2}")
             calculated_cost = 0.0 # Default to 0 on any error
             logger.debug(f"Cost calculation failed with fallback method: {e2}")
 
@@ -354,14 +475,108 @@ def _litellm_success_callback(
 # Register the callback with LiteLLM
 litellm.success_callback = [_litellm_success_callback]
 
+# --- Cost Mapping Support (CSV Rates) ---
+# Populate from CSV inside llm_invoke; used by callback fallback
+_MODEL_RATE_MAP: Dict[str, Tuple[float, float]] = {}
+
+def _set_model_rate_map(df: pd.DataFrame) -> None:
+    global _MODEL_RATE_MAP
+    try:
+        _MODEL_RATE_MAP = {
+            str(row['model']): (
+                float(row['input']) if pd.notna(row['input']) else 0.0,
+                float(row['output']) if pd.notna(row['output']) else 0.0,
+            )
+            for _, row in df.iterrows()
+        }
+    except Exception:
+        _MODEL_RATE_MAP = {}
+
 # --- Helper Functions ---
 
-def _load_model_data(csv_path: Path) -> pd.DataFrame:
-    """Loads and preprocesses the LLM model data from CSV."""
-    if not csv_path.exists():
-        raise FileNotFoundError(f"LLM model CSV not found at {csv_path}")
+def _is_malformed_json_response(content: str, threshold: int = 100) -> bool:
+    """
+    Detect if a JSON response appears malformed due to excessive trailing newlines.
+
+    This can happen when Gemini generates thousands of \n characters in a JSON string value,
+    causing the response to be truncated and missing closing braces.
+
+    Args:
+        content: The raw response content string
+        threshold: Number of consecutive trailing \n sequences to consider malformed
+
+    Returns:
+        True if the response appears malformed, False otherwise
+    """
+    if not content or not isinstance(content, str):
+        return False
+
+    # Check if it starts like JSON but doesn't end properly
+    stripped = content.strip()
+    if not stripped.startswith('{'):
+        return False
+
+    # If it ends with }, it's probably fine
+    if stripped.endswith('}'):
+        return False
+
+    # Count trailing \n sequences (escaped newlines in JSON strings)
+    # The pattern \n in a JSON string appears as \\n in the raw content
+    trailing_newline_count = 0
+    check_content = stripped
+    while check_content.endswith('\\n'):
+        trailing_newline_count += 1
+        check_content = check_content[:-2]
+
+    # If there are many trailing \n sequences, it's likely malformed
+    if trailing_newline_count >= threshold:
+        return True
+
+    # Also check for response that looks truncated mid-string
+    # (ends with characters that suggest we're inside a JSON string value)
+    if not stripped.endswith('}') and not stripped.endswith(']') and not stripped.endswith('"'):
+        # Could be truncated in the middle of an escaped sequence
+        if stripped.endswith('\\'):
+            return True
+
+    return False
+
+
+def _load_model_data(csv_path: Optional[Path]) -> pd.DataFrame:
+    """Loads and preprocesses the LLM model data from CSV.
+    
+    Args:
+        csv_path: Path to CSV file, or None to use package default
+        
+    Returns:
+        DataFrame with model configuration data
+    """
+    # If csv_path is provided, try to load from it
+    if csv_path is not None:
+        if not csv_path.exists():
+            logger.warning(f"Specified LLM model CSV not found at {csv_path}, trying package default")
+            csv_path = None
+        else:
+            try:
+                df = pd.read_csv(csv_path)
+                logger.debug(f"Loaded model data from {csv_path}")
+                # Continue with the rest of the function...
+            except Exception as e:
+                logger.warning(f"Failed to load CSV from {csv_path}: {e}, trying package default")
+                csv_path = None
+    
+    # If csv_path is None or loading failed, use package default
+    if csv_path is None:
+        try:
+            # Use importlib.resources to load the packaged CSV
+            csv_data = importlib.resources.files('pdd').joinpath('data/llm_model.csv').read_text()
+            import io
+            df = pd.read_csv(io.StringIO(csv_data))
+            logger.info("Loaded model data from package default")
+        except Exception as e:
+            raise FileNotFoundError(f"Failed to load default LLM model CSV from package: {e}")
+    
     try:
-        df = pd.read_csv(csv_path)
         # Basic validation and type conversion
         required_cols = ['provider', 'model', 'input', 'output', 'coding_arena_elo', 'api_key', 'structured_output', 'reasoning_type']
         for col in required_cols:
@@ -434,11 +649,26 @@ def _select_model_candidates(
         # Try finding base model in the *original* df in case it was filtered out
         original_base = model_df[model_df['model'] == base_model_name]
         if not original_base.empty:
-             raise ValueError(f"Base model '{base_model_name}' found in CSV but requires API key '{original_base.iloc[0]['api_key']}' which might be missing or invalid configuration.")
-        else:
-             raise ValueError(f"Specified base model '{base_model_name}' not found in the LLM model CSV.")
-
-    base_model = base_model_row.iloc[0]
+            # Base exists but may be misconfigured (e.g., missing API key). Keep erroring loudly.
+            raise ValueError(
+                f"Base model '{base_model_name}' found in CSV but requires API key '{original_base.iloc[0]['api_key']}' which might be missing or invalid configuration."
+            )
+        # Option A': Soft fallback – choose a reasonable surrogate base and continue
+        # Strategy (simplified and deterministic): pick the first available model
+        # from the CSV as the surrogate base. This mirrors typical CSV ordering
+        # expectations and keeps behavior predictable across environments.
+        try:
+            base_model = available_df.iloc[0]
+            logger.warning(
+                f"Base model '{base_model_name}' not found in CSV. Falling back to surrogate base '{base_model['model']}' (Option A')."
+            )
+        except Exception:
+            # If any unexpected error occurs during fallback, raise a clear error
+            raise ValueError(
+                f"Specified base model '{base_model_name}' not found and fallback selection failed. Check your LLM model CSV."
+            )
+    else:
+        base_model = base_model_row.iloc[0]
 
     # 3. Determine Target and Sort
     candidates = []
@@ -449,9 +679,10 @@ def _select_model_candidates(
         # Sort remaining by ELO descending as fallback
         available_df['sort_metric'] = -available_df['coding_arena_elo'] # Negative for descending sort
         candidates = available_df.sort_values(by='sort_metric').to_dict('records')
-        # Ensure base model is first if it exists
-        if any(c['model'] == base_model_name for c in candidates):
-            candidates.sort(key=lambda x: 0 if x['model'] == base_model_name else 1)
+        # Ensure effective base model is first if it exists (supports surrogate base)
+        effective_base_name = str(base_model['model']) if isinstance(base_model, pd.Series) else base_model_name
+        if any(c['model'] == effective_base_name for c in candidates):
+            candidates.sort(key=lambda x: 0 if x['model'] == effective_base_name else 1)
         target_metric_value = f"Base Model ELO: {base_model['coding_arena_elo']}"
 
     elif strength < 0.5:
@@ -668,6 +899,378 @@ def _format_messages(prompt: str, input_data: Union[Dict[str, Any], List[Dict[st
     except Exception as e:
         raise ValueError(f"Error formatting prompt: {e}") from e
 
+# --- JSON Extraction Helpers ---
+import re
+
+def _extract_fenced_json_block(text: str) -> Optional[str]:
+    try:
+        m = re.search(r"```json\s*(\{[\s\S]*?\})\s*```", text, flags=re.IGNORECASE)
+        if m:
+            return m.group(1)
+        return None
+    except Exception:
+        return None
+
+def _extract_balanced_json_objects(text: str) -> List[str]:
+    results: List[str] = []
+    brace_stack = 0
+    start_idx = -1
+    in_string = False
+    escape = False
+    for i, ch in enumerate(text):
+        if in_string:
+            if escape:
+                escape = False
+            elif ch == '\\':
+                escape = True
+            elif ch == '"':
+                in_string = False
+            continue
+        else:
+            if ch == '"':
+                in_string = True
+                continue
+            if ch == '{':
+                if brace_stack == 0:
+                    start_idx = i
+                brace_stack += 1
+            elif ch == '}':
+                if brace_stack > 0:
+                    brace_stack -= 1
+                    if brace_stack == 0 and start_idx != -1:
+                        results.append(text[start_idx:i+1])
+                        start_idx = -1
+    return results
+
+
+def _looks_like_python_code(s: str) -> bool:
+    """
+    Heuristic check if a string looks like Python code.
+
+    Used to determine if we should attempt Python syntax repair on a string field.
+    """
+    if not s or len(s) < 10:
+        return False
+    # Check for common Python patterns
+    code_indicators = ('def ', 'class ', 'import ', 'from ', 'if __name__', 'return ', 'print(')
+    return any(indicator in s for indicator in code_indicators)
+
+
+def _repair_python_syntax(code: str) -> str:
+    """
+    Validate Python code syntax and attempt repairs if invalid.
+
+    Sometimes LLMs include spurious characters at string boundaries,
+    especially when the code contains quotes. This function attempts
+    to detect and repair such issues.
+
+    Args:
+        code: Python code string to validate/repair
+
+    Returns:
+        Repaired code if a fix was found, otherwise original code
+    """
+    import ast
+
+    if not code or not code.strip():
+        return code
+
+    # First, try to parse as-is
+    try:
+        ast.parse(code)
+        return code  # Valid, no repair needed
+    except SyntaxError:
+        pass
+
+    # Try common repairs
+    repaired = code
+
+    # Repair 1: Trailing spurious quote (the specific issue we've seen)
+    for quote in ['"', "'"]:
+        if repaired.rstrip().endswith(quote):
+            candidate = repaired.rstrip()[:-1]
+            try:
+                ast.parse(candidate)
+                logger.info(f"[INFO] Repaired code by removing trailing {quote!r}")
+                return candidate
+            except SyntaxError:
+                pass
+
+    # Repair 2: Leading spurious quote
+    for quote in ['"', "'"]:
+        if repaired.lstrip().startswith(quote):
+            candidate = repaired.lstrip()[1:]
+            try:
+                ast.parse(candidate)
+                logger.info(f"[INFO] Repaired code by removing leading {quote!r}")
+                return candidate
+            except SyntaxError:
+                pass
+
+    # Repair 3: Both leading and trailing spurious quotes
+    for quote in ['"', "'"]:
+        stripped = repaired.strip()
+        if stripped.startswith(quote) and stripped.endswith(quote):
+            candidate = stripped[1:-1]
+            try:
+                ast.parse(candidate)
+                logger.info(f"[INFO] Repaired code by removing surrounding {quote!r}")
+                return candidate
+            except SyntaxError:
+                pass
+
+    # If no repair worked, return original (let it fail downstream)
+    return code
+
+
+def _smart_unescape_code(code: str) -> str:
+    """
+    Unescape literal \\n sequences in code while preserving them inside string literals.
+
+    When LLMs return code as JSON, newlines get double-escaped. After JSON parsing,
+    we have literal backslash-n (2 chars) that should be actual newlines for code
+    structure, BUT escape sequences inside Python strings (like print("\\n")) should
+    remain as escape sequences.
+
+    Args:
+        code: Python code that may have literal \\n sequences
+
+    Returns:
+        Code with structural newlines unescaped but string literals preserved
+    """
+    LITERAL_BACKSLASH_N = '\\' + 'n'      # Literal \n (2 chars)
+
+    if LITERAL_BACKSLASH_N not in code:
+        return code
+
+    # First, check if the code already has actual newlines (mixed state)
+    # If it does, we need to be more careful
+    has_actual_newlines = '\n' in code
+
+    if not has_actual_newlines:
+        # All newlines are escaped - this is the double-escaped case
+        # We need to unescape them but preserve \n inside string literals
+
+        # Strategy: Use a placeholder for \n inside strings, unescape all, then restore
+        # We detect string literals by tracking quote state
+
+        result = []
+        i = 0
+        in_string = False
+        string_char = None
+        in_fstring = False
+
+        # Placeholder that won't appear in code
+        PLACEHOLDER = '\x00NEWLINE_ESCAPE\x00'
+
+        while i < len(code):
+            # Check for escape sequences (both actual and literal)
+            if i + 1 < len(code) and code[i] == '\\':
+                next_char = code[i + 1]
+
+                if in_string:
+                    # Inside a string - preserve escape sequences
+                    if next_char == 'n':
+                        result.append(PLACEHOLDER)
+                        i += 2
+                        continue
+                    elif next_char == 't':
+                        result.append('\\' + 't')  # Keep \t as-is in strings
+                        i += 2
+                        continue
+                    elif next_char == 'r':
+                        result.append('\\' + 'r')  # Keep \r as-is in strings
+                        i += 2
+                        continue
+                    elif next_char in ('"', "'", '\\'):
+                        # Keep escaped quotes and backslashes
+                        result.append(code[i:i+2])
+                        i += 2
+                        continue
+
+            # Check for string delimiters
+            if not in_string:
+                # Check for triple quotes first
+                if i + 2 < len(code) and code[i:i+3] in ('"""', "'''"):
+                    in_string = True
+                    string_char = code[i:i+3]
+                    # Check if preceded by 'f' for f-string
+                    in_fstring = i > 0 and code[i-1] == 'f'
+                    result.append(code[i:i+3])
+                    i += 3
+                    continue
+                elif code[i] in ('"', "'"):
+                    in_string = True
+                    string_char = code[i]
+                    in_fstring = i > 0 and code[i-1] == 'f'
+                    result.append(code[i])
+                    i += 1
+                    continue
+            else:
+                # Check for end of string
+                if len(string_char) == 3:  # Triple quote
+                    if i + 2 < len(code) and code[i:i+3] == string_char:
+                        in_string = False
+                        in_fstring = False
+                        result.append(code[i:i+3])
+                        i += 3
+                        continue
+                else:  # Single quote
+                    if code[i] == string_char:
+                        in_string = False
+                        in_fstring = False
+                        result.append(code[i])
+                        i += 1
+                        continue
+
+            result.append(code[i])
+            i += 1
+
+        intermediate = ''.join(result)
+
+        # Now unescape all remaining \n (these are structural)
+        LITERAL_BACKSLASH_R_N = '\\' + 'r' + '\\' + 'n'
+        LITERAL_BACKSLASH_T = '\\' + 't'
+
+        intermediate = intermediate.replace(LITERAL_BACKSLASH_R_N, '\r\n')
+        intermediate = intermediate.replace(LITERAL_BACKSLASH_N, '\n')
+        intermediate = intermediate.replace(LITERAL_BACKSLASH_T, '\t')
+
+        # Restore placeholders to \n (as escape sequences in strings)
+        result_code = intermediate.replace(PLACEHOLDER, '\\n')
+
+        return result_code
+    else:
+        # Mixed state - some actual newlines, some literal \n
+        # This means the JSON parsing already converted some, but not all
+        # The literal \n remaining are likely in strings, so leave them alone
+        return code
+
+
+def _unescape_code_newlines(obj: Any) -> Any:
+    """
+    Fix double-escaped newlines in Pydantic model string fields.
+
+    Some models (e.g., Gemini) return JSON with \\\\n instead of \\n in code strings,
+    resulting in literal backslash-n text instead of actual newlines after JSON parsing.
+    This function recursively unescapes these in string fields of Pydantic models.
+
+    Also repairs Python syntax errors in code-like string fields (e.g., trailing quotes).
+
+    The check uses literal backslash-n (2 chars) vs actual newline (1 char):
+    - '\\\\n' in Python source = literal backslash + n (2 chars) - needs fixing
+    - '\\n' in Python source = newline character (1 char) - already correct
+
+    Args:
+        obj: A Pydantic model, dict, list, or primitive value
+
+    Returns:
+        The same object with string fields unescaped and code fields repaired
+    """
+    if obj is None:
+        return obj
+
+    def _process_string(s: str) -> str:
+        """Process a string: unescape newlines and repair Python syntax if needed."""
+        result = s
+        # Smart unescape that preserves \n inside string literals
+        if _looks_like_python_code(result):
+            result = _smart_unescape_code(result)
+            result = _repair_python_syntax(result)
+        else:
+            # For non-code strings, do simple unescape
+            LITERAL_BACKSLASH_N = '\\' + 'n'
+            LITERAL_BACKSLASH_R_N = '\\' + 'r' + '\\' + 'n'
+            LITERAL_BACKSLASH_T = '\\' + 't'
+            if LITERAL_BACKSLASH_N in result:
+                result = result.replace(LITERAL_BACKSLASH_R_N, '\r\n')
+                result = result.replace(LITERAL_BACKSLASH_N, '\n')
+                result = result.replace(LITERAL_BACKSLASH_T, '\t')
+        return result
+
+    # Handle Pydantic models
+    if isinstance(obj, BaseModel):
+        # Get all field values and process strings
+        for field_name in obj.model_fields:
+            value = getattr(obj, field_name)
+            if isinstance(value, str):
+                processed = _process_string(value)
+                if processed != value:
+                    object.__setattr__(obj, field_name, processed)
+            elif isinstance(value, (dict, list, BaseModel)):
+                _unescape_code_newlines(value)
+        return obj
+
+    # Handle dicts
+    if isinstance(obj, dict):
+        for key, value in obj.items():
+            if isinstance(value, str):
+                obj[key] = _process_string(value)
+            elif isinstance(value, (dict, list)):
+                _unescape_code_newlines(value)
+        return obj
+
+    # Handle lists
+    if isinstance(obj, list):
+        for i, item in enumerate(obj):
+            if isinstance(item, str):
+                obj[i] = _process_string(item)
+            elif isinstance(item, (dict, list, BaseModel)):
+                _unescape_code_newlines(item)
+        return obj
+
+    return obj
+
+
+def _has_invalid_python_code(obj: Any) -> bool:
+    """
+    Check if any code-like string fields have invalid Python syntax.
+
+    This is used after _unescape_code_newlines to detect if repair failed
+    and we should retry with cache disabled.
+
+    Args:
+        obj: A Pydantic model, dict, list, or primitive value
+
+    Returns:
+        True if there are invalid code fields that couldn't be repaired
+    """
+    import ast
+
+    if obj is None:
+        return False
+
+    if isinstance(obj, str):
+        if _looks_like_python_code(obj):
+            try:
+                ast.parse(obj)
+                return False  # Valid
+            except SyntaxError:
+                return True  # Invalid
+        return False
+
+    if isinstance(obj, BaseModel):
+        for field_name in obj.model_fields:
+            value = getattr(obj, field_name)
+            if _has_invalid_python_code(value):
+                return True
+        return False
+
+    if isinstance(obj, dict):
+        for value in obj.values():
+            if _has_invalid_python_code(value):
+                return True
+        return False
+
+    if isinstance(obj, list):
+        for item in obj:
+            if _has_invalid_python_code(item):
+                return True
+        return False
+
+    return False
+
+
 # --- Main Function ---
 
 def llm_invoke(
@@ -677,6 +1280,7 @@ def llm_invoke(
     temperature: float = 0.1,
     verbose: bool = False,
     output_pydantic: Optional[Type[BaseModel]] = None,
+    output_schema: Optional[Dict[str, Any]] = None,
     time: float = 0.25,
     use_batch_mode: bool = False,
     messages: Optional[Union[List[Dict[str, str]], List[List[Dict[str, str]]]]] = None,
@@ -693,6 +1297,7 @@ def llm_invoke(
         temperature: LLM temperature.
         verbose: Print detailed logs.
         output_pydantic: Optional Pydantic model for structured output.
+        output_schema: Optional raw JSON schema dictionary for structured output (alternative to output_pydantic).
         time: Relative thinking time (0-1, default 0.25).
         use_batch_mode: Use batch completion if True.
         messages: Pre-formatted list of messages (or list of lists for batch). If provided, ignores prompt and input_json.
@@ -810,6 +1415,16 @@ def llm_invoke(
     # --- 3. Iterate Through Candidates and Invoke LLM ---
     last_exception = None
     newly_acquired_keys: Dict[str, bool] = {} # Track keys obtained in this run
+    
+    # Initialize variables for retry section
+    response_format = None
+    time_kwargs = {}
+
+    # Update global rate map for callback cost fallback
+    try:
+        _set_model_rate_map(model_df)
+    except Exception:
+        pass
 
     for model_info in candidate_models:
         model_name_litellm = model_info['model']
@@ -820,6 +1435,9 @@ def llm_invoke(
             logger.info(f"\n[ATTEMPT] Trying model: {model_name_litellm} (Provider: {provider})")
 
         retry_with_same_model = True
+        # Track per-model temperature adjustment attempt (avoid infinite loop)
+        current_temperature = temperature
+        temp_adjustment_done = False
         while retry_with_same_model:
             retry_with_same_model = False # Assume success unless auth error on new key
 
@@ -834,7 +1452,8 @@ def llm_invoke(
             litellm_kwargs: Dict[str, Any] = {
                 "model": model_name_litellm,
                 "messages": formatted_messages,
-                "temperature": temperature,
+                # Use a local adjustable temperature to allow provider-specific fallbacks
+                "temperature": current_temperature,
             }
 
             api_key_name_from_csv = model_info.get('api_key') # From CSV
@@ -847,7 +1466,14 @@ def llm_invoke(
             if is_vertex_model and api_key_name_from_csv == 'VERTEX_CREDENTIALS':
                 credentials_file_path = os.getenv("VERTEX_CREDENTIALS") # Path from env var
                 vertex_project_env = os.getenv("VERTEX_PROJECT")
-                vertex_location_env = os.getenv("VERTEX_LOCATION")
+                # Check for per-model location override, fall back to env var
+                model_location = model_info.get('location')
+                if pd.notna(model_location) and str(model_location).strip():
+                    vertex_location_env = str(model_location).strip()
+                    if verbose:
+                        logger.info(f"[INFO] Using per-model location override: '{vertex_location_env}' for model '{model_name_litellm}'")
+                else:
+                    vertex_location_env = os.getenv("VERTEX_LOCATION")
 
                 if credentials_file_path and vertex_project_env and vertex_location_env:
                     try:
@@ -861,14 +1487,23 @@ def llm_invoke(
                         if verbose:
                             logger.info(f"[INFO] For Vertex AI: using vertex_credentials from '{credentials_file_path}', project '{vertex_project_env}', location '{vertex_location_env}'.")
                     except FileNotFoundError:
+                        # Still pass project and location so ADC can work
+                        litellm_kwargs["vertex_project"] = vertex_project_env
+                        litellm_kwargs["vertex_location"] = vertex_location_env
                         if verbose:
-                            logger.error(f"[ERROR] Vertex credentials file not found at path specified by VERTEX_CREDENTIALS env var: '{credentials_file_path}'. LiteLLM may try ADC or fail.")
+                            logger.warning(f"[WARN] Vertex credentials file not found at '{credentials_file_path}'. Using ADC with project '{vertex_project_env}', location '{vertex_location_env}'.")
                     except json.JSONDecodeError:
+                        # Still pass project and location so ADC can work
+                        litellm_kwargs["vertex_project"] = vertex_project_env
+                        litellm_kwargs["vertex_location"] = vertex_location_env
                         if verbose:
-                            logger.error(f"[ERROR] Failed to decode JSON from Vertex credentials file: '{credentials_file_path}'. Check file content. LiteLLM may try ADC or fail.")
+                            logger.error(f"[ERROR] Failed to decode JSON from Vertex credentials file: '{credentials_file_path}'. Using ADC with project '{vertex_project_env}', location '{vertex_location_env}'.")
                     except Exception as e:
+                        # Still pass project and location so ADC can work
+                        litellm_kwargs["vertex_project"] = vertex_project_env
+                        litellm_kwargs["vertex_location"] = vertex_location_env
                         if verbose:
-                            logger.error(f"[ERROR] Failed to load or process Vertex credentials from '{credentials_file_path}': {e}. LiteLLM may try ADC or fail.")
+                            logger.error(f"[ERROR] Failed to load Vertex credentials from '{credentials_file_path}': {e}. Using ADC with project '{vertex_project_env}', location '{vertex_location_env}'.")
                 else:
                     if verbose:
                         logger.warning(f"[WARN] For Vertex AI (using '{api_key_name_from_csv}'): One or more required environment variables (VERTEX_CREDENTIALS, VERTEX_PROJECT, VERTEX_LOCATION) are missing.")
@@ -887,9 +1522,16 @@ def llm_invoke(
                     
                     # If this model is Vertex AI AND uses a direct API key string (not VERTEX_CREDENTIALS from CSV),
                     # also pass project and location from env vars.
-                    if is_vertex_model: 
+                    if is_vertex_model:
                         vertex_project_env = os.getenv("VERTEX_PROJECT")
-                        vertex_location_env = os.getenv("VERTEX_LOCATION")
+                        # Check for per-model location override, fall back to env var
+                        model_location = model_info.get('location')
+                        if pd.notna(model_location) and str(model_location).strip():
+                            vertex_location_env = str(model_location).strip()
+                            if verbose:
+                                logger.info(f"[INFO] Using per-model location override: '{vertex_location_env}' for model '{model_name_litellm}'")
+                        else:
+                            vertex_location_env = os.getenv("VERTEX_LOCATION")
                         if vertex_project_env and vertex_location_env:
                             litellm_kwargs["vertex_project"] = vertex_project_env
                             litellm_kwargs["vertex_location"] = vertex_location_env
@@ -903,13 +1545,36 @@ def llm_invoke(
             elif verbose: # No api_key_name_from_csv in CSV for this model
                 logger.info(f"[INFO] No API key name specified in CSV for model '{model_name_litellm}'. LiteLLM will use its default authentication mechanisms (e.g., standard provider env vars or ADC for Vertex AI).")
 
-            # Add api_base if present in CSV
+            # Add base_url/api_base override if present in CSV
             api_base = model_info.get('base_url')
             if pd.notna(api_base) and api_base:
+                # LiteLLM prefers `base_url`; some older paths accept `api_base`.
+                litellm_kwargs["base_url"] = str(api_base)
                 litellm_kwargs["api_base"] = str(api_base)
 
-            # Handle Structured Output (JSON Mode / Pydantic)
-            if output_pydantic:
+            # Provider-specific defaults (e.g., LM Studio)
+            model_name_lower = str(model_name_litellm).lower()
+            provider_lower_for_model = provider.lower()
+            is_lm_studio = model_name_lower.startswith('lm_studio/') or provider_lower_for_model == 'lm_studio'
+            is_groq = model_name_lower.startswith('groq/') or provider_lower_for_model == 'groq'
+            if is_lm_studio:
+                # Ensure base_url is set (fallback to env LM_STUDIO_API_BASE or localhost)
+                if not litellm_kwargs.get("base_url"):
+                    lm_studio_base = os.getenv("LM_STUDIO_API_BASE", "http://localhost:1234/v1")
+                    litellm_kwargs["base_url"] = lm_studio_base
+                    litellm_kwargs["api_base"] = lm_studio_base
+                    if verbose:
+                        logger.info(f"[INFO] Using LM Studio base_url: {lm_studio_base}")
+
+                # Ensure a non-empty api_key; LM Studio accepts any non-empty token (e.g., 'lm-studio')
+                if not litellm_kwargs.get("api_key"):
+                    lm_studio_key = os.getenv("LM_STUDIO_API_KEY") or "lm-studio"
+                    litellm_kwargs["api_key"] = lm_studio_key
+                    if verbose:
+                        logger.info("[INFO] Using LM Studio api_key placeholder (set LM_STUDIO_API_KEY to customize).")
+
+            # Handle Structured Output (JSON Mode / Pydantic / JSON Schema)
+            if output_pydantic or output_schema:
                 # Check if model supports structured output based on CSV flag or LiteLLM check
                 supports_structured = model_info.get('structured_output', False)
                 # Optional: Add litellm.supports_response_schema check if CSV flag is unreliable
@@ -918,18 +1583,87 @@ def llm_invoke(
                 #     except: pass # Ignore errors in supports_response_schema check
 
                 if supports_structured:
-                    if verbose:
-                        logger.info(f"[INFO] Requesting structured output (Pydantic: {output_pydantic.__name__}) for {model_name_litellm}")
-                    # Pass the Pydantic model directly if supported, else use json_object
-                    # LiteLLM handles passing Pydantic models for supported providers
-                    litellm_kwargs["response_format"] = output_pydantic
+                    if output_pydantic:
+                        if verbose:
+                            logger.info(f"[INFO] Requesting structured output (Pydantic: {output_pydantic.__name__}) for {model_name_litellm}")
+                        # Use explicit json_object format with response_schema for better Gemini/Vertex AI compatibility
+                        # Passing Pydantic class directly may not trigger native structured output for all providers
+                        response_format = {
+                            "type": "json_object",
+                            "response_schema": output_pydantic.model_json_schema()
+                        }
+                    else: # output_schema is set
+                        if verbose:
+                            logger.info(f"[INFO] Requesting structured output (JSON Schema) for {model_name_litellm}")
+                        # LiteLLM expects {"type": "json_schema", "json_schema": {"name": "response", "schema": schema_dict, "strict": true}}
+                        # OR for some providers just the schema dict if type is json_object.
+                        # Best practice for broad compatibility via LiteLLM is usually the dict directly or wrapped.
+                        # For now, let's assume we pass the schema dict as 'response_format' which LiteLLM handles for many providers
+                        # or wrap it if needed. LiteLLM 1.40+ supports passing the dict directly for many.
+                        response_format = {
+                            "type": "json_schema",
+                            "json_schema": {
+                                "name": "response",
+                                "schema": output_schema,
+                                "strict": False
+                            }
+                        }
+                    
+                    litellm_kwargs["response_format"] = response_format
+
+                    # LM Studio requires "json_schema" format, not "json_object"
+                    # Use extra_body to bypass litellm.drop_params stripping the schema
+                    if is_lm_studio and response_format and response_format.get("type") == "json_object":
+                        schema = response_format.get("response_schema", {})
+                        lm_studio_response_format = {
+                            "type": "json_schema",
+                            "json_schema": {
+                                "name": "response",
+                                "strict": True,
+                                "schema": schema
+                            }
+                        }
+                        # Use extra_body to bypass drop_params - passes directly to API
+                        litellm_kwargs["extra_body"] = {"response_format": lm_studio_response_format}
+                        # Remove from regular response_format to avoid conflicts
+                        if "response_format" in litellm_kwargs:
+                            del litellm_kwargs["response_format"]
+                        if verbose:
+                            logger.info(f"[INFO] Using extra_body for LM Studio response_format to bypass drop_params")
+
+                    # Groq has issues with tool-based structured output - use JSON mode with schema in prompt
+                    if is_groq and response_format:
+                        # Get the schema to include in system prompt
+                        if output_pydantic:
+                            schema = output_pydantic.model_json_schema()
+                        else:
+                            schema = output_schema
+
+                        # Use simple json_object mode (Groq's tool_use often fails)
+                        litellm_kwargs["response_format"] = {"type": "json_object"}
+
+                        # Prepend schema instruction to messages (json module is imported at top of file)
+                        schema_instruction = f"You must respond with valid JSON matching this schema:\n```json\n{json.dumps(schema, indent=2)}\n```\nRespond ONLY with the JSON object, no other text."
+
+                        # Find or create system message to prepend schema
+                        messages_list = litellm_kwargs.get("messages", [])
+                        if messages_list and messages_list[0].get("role") == "system":
+                            messages_list[0]["content"] = schema_instruction + "\n\n" + messages_list[0]["content"]
+                        else:
+                            messages_list.insert(0, {"role": "system", "content": schema_instruction})
+                        litellm_kwargs["messages"] = messages_list
+
+                        if verbose:
+                            logger.info(f"[INFO] Using JSON object mode with schema in prompt for Groq (avoiding tool_use issues)")
+
                     # As a fallback, one could use:
                     # litellm_kwargs["response_format"] = {"type": "json_object"}
                     # And potentially enable client-side validation:
                     # litellm.enable_json_schema_validation = True # Enable globally if needed
                 else:
+                    schema_name = output_pydantic.__name__ if output_pydantic else "JSON Schema"
                     if verbose:
-                        logger.warning(f"[WARN] Model {model_name_litellm} does not support structured output via CSV flag. Output might not be valid {output_pydantic.__name__}.")
+                        logger.warning(f"[WARN] Model {model_name_litellm} does not support structured output via CSV flag. Output might not be valid {schema_name}.")
                     # Proceed without forcing JSON mode, parsing will be attempted later
 
             # --- NEW REASONING LOGIC ---
@@ -944,7 +1678,9 @@ def llm_invoke(
                             # Currently known: Anthropic uses 'thinking'
                             # Model name comparison is more robust than provider string
                             if provider == 'anthropic': # Check provider column instead of model prefix
-                                litellm_kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
+                                thinking_param = {"type": "enabled", "budget_tokens": budget}
+                                litellm_kwargs["thinking"] = thinking_param
+                                time_kwargs["thinking"] = thinking_param
                                 if verbose:
                                     logger.info(f"[INFO] Requesting Anthropic thinking (budget type) with budget: {budget} tokens for {model_name_litellm}")
                             else:
@@ -962,10 +1698,32 @@ def llm_invoke(
                         effort = "high"
                     elif time > 0.3:
                         effort = "medium"
-                    # Use the common 'reasoning_effort' param LiteLLM provides
-                    litellm_kwargs["reasoning_effort"] = effort
-                    if verbose:
-                        logger.info(f"[INFO] Requesting reasoning_effort='{effort}' (effort type) for {model_name_litellm} based on time={time}")
+
+                    # Map effort parameter per-provider/model family
+                    model_lower = str(model_name_litellm).lower()
+                    provider_lower = str(provider).lower()
+
+                    if provider_lower == 'openai' and model_lower.startswith('gpt-5'):
+                        # OpenAI 5-series uses Responses API with nested 'reasoning'
+                        reasoning_obj = {"effort": effort, "summary": "auto"}
+                        litellm_kwargs["reasoning"] = reasoning_obj
+                        time_kwargs["reasoning"] = reasoning_obj
+                        if verbose:
+                            logger.info(f"[INFO] Requesting OpenAI reasoning.effort='{effort}' for {model_name_litellm} (Responses API)")
+
+                    elif provider_lower == 'openai' and model_lower.startswith('o') and 'mini' not in model_lower:
+                        # Historical o* models may use LiteLLM's generic reasoning_effort param
+                        litellm_kwargs["reasoning_effort"] = effort
+                        time_kwargs["reasoning_effort"] = effort
+                        if verbose:
+                            logger.info(f"[INFO] Requesting reasoning_effort='{effort}' for {model_name_litellm}")
+
+                    else:
+                        # Fallback to LiteLLM generic param when supported by provider adapter
+                        litellm_kwargs["reasoning_effort"] = effort
+                        time_kwargs["reasoning_effort"] = effort
+                        if verbose:
+                            logger.info(f"[INFO] Requesting generic reasoning_effort='{effort}' for {model_name_litellm}")
 
                 elif reasoning_type == 'none':
                     if verbose:
@@ -997,6 +1755,166 @@ def llm_invoke(
                     logger.debug("NOT ENABLING CACHING: litellm.cache is None at call time")
 
 
+                # Route OpenAI gpt-5* models through Responses API to support 'reasoning'
+                model_lower_for_call = str(model_name_litellm).lower()
+                provider_lower_for_call = str(provider).lower()
+
+                if (
+                    not use_batch_mode
+                    and provider_lower_for_call == 'openai'
+                    and model_lower_for_call.startswith('gpt-5')
+                ):
+                    if verbose:
+                        logger.info(f"[INFO] Calling LiteLLM Responses API for {model_name_litellm}...")
+                    try:
+                        # Build input text from messages
+                        if isinstance(formatted_messages, list) and formatted_messages and isinstance(formatted_messages[0], dict):
+                            input_text = "\n\n".join(f"{m.get('role','user')}: {m.get('content','')}" for m in formatted_messages)
+                        else:
+                            # Fallback: string cast
+                            input_text = str(formatted_messages)
+
+                        # Derive effort mapping already computed in time_kwargs
+                        reasoning_param = time_kwargs.get("reasoning")
+
+                        # Build text.format block for structured output
+                        # Default to plain text format
+                        text_block = {"format": {"type": "text"}}
+
+                        # If structured output requested, use text.format with json_schema
+                        # This is the correct way to enforce structured output via litellm.responses()
+                        if output_pydantic or output_schema:
+                            try:
+                                if output_pydantic:
+                                    schema = output_pydantic.model_json_schema()
+                                    name = output_pydantic.__name__
+                                else:
+                                    schema = output_schema
+                                    name = "response"
+
+                                # Add additionalProperties: false for strict mode (required by OpenAI)
+                                schema['additionalProperties'] = False
+
+                                # Use text.format with json_schema for structured output
+                                text_block = {
+                                    "format": {
+                                        "type": "json_schema",
+                                        "name": name,
+                                        "strict": True,
+                                        "schema": schema,
+                                    }
+                                }
+                                if verbose:
+                                    logger.info(f"[INFO] Using structured output via text.format for Responses API")
+                            except Exception as schema_e:
+                                logger.warning(f"[WARN] Failed to derive JSON schema: {schema_e}. Proceeding with plain text format.")
+
+                        # Build kwargs for litellm.responses()
+                        responses_kwargs = {
+                            "model": model_name_litellm,
+                            "input": input_text,
+                            "text": text_block,
+                        }
+                        if verbose and temperature not in (None, 0, 0.0):
+                            logger.info("[INFO] Skipping 'temperature' for OpenAI GPT-5 Responses call (unsupported by API).")
+                        if reasoning_param is not None:
+                            responses_kwargs["reasoning"] = reasoning_param
+
+                        # Call litellm.responses() which handles the API interaction
+                        resp = litellm.responses(**responses_kwargs)
+
+                        # Extract text result from response
+                        result_text = None
+                        try:
+                            # LiteLLM responses return output as a list of items
+                            for item in resp.output:
+                                if getattr(item, 'type', None) == 'message' and hasattr(item, 'content') and item.content:
+                                    for content_item in item.content:
+                                        if hasattr(content_item, 'text'):
+                                            result_text = content_item.text
+                                            break
+                                    if result_text:
+                                        break
+                        except Exception:
+                            result_text = None
+
+                        # Calculate cost using usage + CSV rates
+                        total_cost = 0.0
+                        usage = getattr(resp, "usage", None)
+                        if usage is not None:
+                            in_tok = getattr(usage, "input_tokens", 0) or 0
+                            out_tok = getattr(usage, "output_tokens", 0) or 0
+                            in_rate = model_info.get('input', 0.0) or 0.0
+                            out_rate = model_info.get('output', 0.0) or 0.0
+                            total_cost = (in_tok * in_rate + out_tok * out_rate) / 1_000_000.0
+
+                        # Parse result if Pydantic output requested
+                        final_result = None
+                        if output_pydantic and result_text:
+                            try:
+                                final_result = output_pydantic.model_validate_json(result_text)
+                            except Exception as e:
+                                # With structured output, parsing should succeed
+                                # But if it fails, try JSON repair as fallback
+                                logger.warning(f"[WARN] Pydantic parse failed on Responses output: {e}. Attempting JSON repair...")
+
+                                # Try extracting from fenced JSON blocks first
+                                fenced = _extract_fenced_json_block(result_text)
+                                candidates: List[str] = []
+                                if fenced:
+                                    candidates.append(fenced)
+                                else:
+                                    candidates.extend(_extract_balanced_json_objects(result_text))
+
+                                # Also try the raw text as-is after stripping fences
+                                cleaned = result_text.strip()
+                                if cleaned.startswith("```json"):
+                                    cleaned = cleaned[7:]
+                                elif cleaned.startswith("```"):
+                                    cleaned = cleaned[3:]
+                                if cleaned.endswith("```"):
+                                    cleaned = cleaned[:-3]
+                                cleaned = cleaned.strip()
+                                if cleaned and cleaned not in candidates:
+                                    candidates.append(cleaned)
+
+                                parse_succeeded = False
+                                for cand in candidates:
+                                    try:
+                                        final_result = output_pydantic.model_validate_json(cand)
+                                        parse_succeeded = True
+                                        logger.info(f"[SUCCESS] JSON repair succeeded for Responses output")
+                                        break
+                                    except Exception:
+                                        continue
+
+                                if not parse_succeeded:
+                                    logger.error(f"[ERROR] All JSON repair attempts failed for Responses output. Original error: {e}")
+                                    final_result = f"ERROR: Failed to parse structured output from Responses API. Raw: {repr(result_text)[:200]}"
+                        else:
+                            final_result = result_text
+
+                        if verbose:
+                            logger.info(f"[RESULT] Model Used: {model_name_litellm}")
+                            logger.info(f"[RESULT] Total Cost (estimated): ${total_cost:.6g}")
+
+                        return {
+                            'result': final_result,
+                            'cost': total_cost,
+                            'model_name': model_name_litellm,
+                            'thinking_output': None,
+                        }
+                    except Exception as e:
+                        last_exception = e
+                        logger.error(f"[ERROR] OpenAI Responses call failed for {model_name_litellm}: {e}")
+                        # Remove 'reasoning' key to avoid OpenAI Chat API unknown param errors
+                        if "reasoning" in litellm_kwargs:
+                            try:
+                                litellm_kwargs.pop("reasoning", None)
+                            except Exception:
+                                pass
+                        # Fall through to LiteLLM path as a fallback
+
                 if use_batch_mode:
                     if verbose:
                         logger.info(f"[INFO] Calling litellm.batch_completion for {model_name_litellm}...")
@@ -1004,6 +1922,16 @@ def llm_invoke(
 
 
                 else:
+                    # Anthropic requirement: when 'thinking' is enabled, temperature must be 1
+                    try:
+                        if provider.lower() == 'anthropic' and 'thinking' in litellm_kwargs:
+                            if litellm_kwargs.get('temperature') != 1:
+                                if verbose:
+                                    logger.info("[INFO] Anthropic thinking enabled: forcing temperature=1 for compliance.")
+                                litellm_kwargs['temperature'] = 1
+                                current_temperature = 1
+                    except Exception:
+                        pass
                     if verbose:
                         logger.info(f"[INFO] Calling litellm.completion for {model_name_litellm}...")
                     response = litellm.completion(**litellm_kwargs)
@@ -1061,13 +1989,12 @@ def llm_invoke(
                                     retry_response = litellm.completion(
                                         model=model_name_litellm,
                                         messages=retry_messages,
-                                        temperature=temperature,
+                                        temperature=current_temperature,
                                         response_format=response_format,
-                                        max_completion_tokens=max_tokens,
                                         **time_kwargs
                                     )
-                                    # Re-enable cache
-                                    litellm.cache = Cache()
+                                    # Re-enable cache - restore original configured cache (restore to original state, even if None)
+                                    litellm.cache = configured_cache
                                     # Extract result from retry
                                     retry_raw_result = retry_response.choices[0].message.content
                                     if retry_raw_result is not None:
@@ -1085,21 +2012,66 @@ def llm_invoke(
                                 logger.error(f"[ERROR] Cannot retry - batch mode or missing prompt/input_json")
                                 results.append("ERROR: LLM returned None content and cannot retry")
                                 continue
-                        
-                        if output_pydantic:
+
+                        # Check for malformed JSON response (excessive trailing newlines causing truncation)
+                        # This can happen when Gemini generates thousands of \n in JSON string values
+                        if isinstance(raw_result, str) and _is_malformed_json_response(raw_result):
+                            logger.warning(f"[WARNING] Detected malformed JSON response with excessive trailing newlines for item {i}. Retrying with cache bypass...")
+                            if not use_batch_mode and prompt and input_json is not None:
+                                # Add a small space to bypass cache
+                                modified_prompt = prompt + " "
+                                try:
+                                    retry_messages = _format_messages(modified_prompt, input_json, use_batch_mode)
+                                    # Disable cache for retry
+                                    original_cache = litellm.cache
+                                    litellm.cache = None
+                                    retry_response = litellm.completion(
+                                        model=model_name_litellm,
+                                        messages=retry_messages,
+                                        temperature=current_temperature,
+                                        response_format=response_format,
+                                        **time_kwargs
+                                    )
+                                    # Re-enable cache
+                                    litellm.cache = original_cache
+                                    # Extract result from retry
+                                    retry_raw_result = retry_response.choices[0].message.content
+                                    if retry_raw_result is not None and not _is_malformed_json_response(retry_raw_result):
+                                        logger.info(f"[SUCCESS] Cache bypass retry for malformed JSON succeeded for item {i}")
+                                        raw_result = retry_raw_result
+                                    else:
+                                        # Retry also failed, but we'll continue with repair logic below
+                                        logger.warning(f"[WARNING] Cache bypass retry also returned malformed JSON for item {i}, attempting repair...")
+                                except Exception as retry_e:
+                                    logger.warning(f"[WARNING] Cache bypass retry for malformed JSON failed for item {i}: {retry_e}, attempting repair...")
+                            else:
+                                logger.warning(f"[WARNING] Cannot retry malformed JSON - batch mode or missing prompt/input_json, attempting repair...")
+
+                        if output_pydantic or output_schema:
                             parsed_result = None
                             json_string_to_parse = None
 
                             try:
-                                # Attempt 1: Check if LiteLLM already parsed it
-                                if isinstance(raw_result, output_pydantic):
+                                # Attempt 1: Check if LiteLLM already parsed it (only for Pydantic)
+                                if output_pydantic and isinstance(raw_result, output_pydantic):
                                     parsed_result = raw_result
                                     if verbose:
                                         logger.debug("[DEBUG] Pydantic object received directly from LiteLLM.")
 
                                 # Attempt 2: Check if raw_result is dict-like and validate
                                 elif isinstance(raw_result, dict):
-                                    parsed_result = output_pydantic.model_validate(raw_result)
+                                    if output_pydantic:
+                                        parsed_result = output_pydantic.model_validate(raw_result)
+                                    else:
+                                        # Validate against JSON schema
+                                        try:
+                                            import jsonschema
+                                            jsonschema.validate(instance=raw_result, schema=output_schema)
+                                            parsed_result = json.dumps(raw_result) # Return as JSON string for consistency
+                                        except ImportError:
+                                            logger.warning("jsonschema not installed, skipping validation")
+                                            parsed_result = json.dumps(raw_result)
+                                    
                                     if verbose:
                                         logger.debug("[DEBUG] Validated dictionary-like object directly.")
 
@@ -1107,26 +2079,59 @@ def llm_invoke(
                                 elif isinstance(raw_result, str):
                                     json_string_to_parse = raw_result # Start with the raw string
                                     try:
-                                        # Look for first { and last }
-                                        start_brace = json_string_to_parse.find('{')
-                                        end_brace = json_string_to_parse.rfind('}')
-                                        if start_brace != -1 and end_brace != -1 and end_brace > start_brace:
-                                            potential_json = json_string_to_parse[start_brace:end_brace+1]
-                                            # Basic check if it looks like JSON
-                                            if potential_json.strip().startswith('{') and potential_json.strip().endswith('}'):
-                                                if verbose:
-                                                    logger.debug(f"[DEBUG] Attempting to parse extracted JSON block: '{potential_json}'")
-                                                parsed_result = output_pydantic.model_validate_json(potential_json)
-                                            else:
-                                                # If block extraction fails, try cleaning markdown next
-                                                raise ValueError("Extracted block doesn't look like JSON")
+                                        # 1) Prefer fenced ```json blocks
+                                        fenced = _extract_fenced_json_block(raw_result)
+                                        candidates: List[str] = []
+                                        if fenced:
+                                            candidates.append(fenced)
                                         else:
-                                             # If no braces found, try cleaning markdown next
-                                            raise ValueError("Could not find enclosing {}")
-                                    except (json.JSONDecodeError, ValidationError, ValueError) as extraction_error:
+                                            # 2) Fall back to scanning for balanced JSON objects
+                                            candidates.extend(_extract_balanced_json_objects(raw_result))
+
+                                        if not candidates:
+                                            raise ValueError("No JSON-like content found")
+
+                                        parse_err: Optional[Exception] = None
+                                        for cand in candidates:
+                                            try:
+                                                if verbose:
+                                                    logger.debug(f"[DEBUG] Attempting to parse candidate JSON block: {cand}")
+                                                
+                                                if output_pydantic:
+                                                    parsed_result = output_pydantic.model_validate_json(cand)
+                                                else:
+                                                    # Parse JSON and validate against schema
+                                                    loaded = json.loads(cand)
+                                                    try:
+                                                        import jsonschema
+                                                        jsonschema.validate(instance=loaded, schema=output_schema)
+                                                    except ImportError:
+                                                        pass # Skip validation if lib missing
+                                                    parsed_result = cand # Return string if valid
+
+                                                json_string_to_parse = cand
+                                                parse_err = None
+                                                break
+                                            except (json.JSONDecodeError, ValidationError, ValueError) as pe:
+                                                # Also catch jsonschema.ValidationError if imported
+                                                parse_err = pe
+                                                try:
+                                                    import jsonschema
+                                                    if isinstance(pe, jsonschema.ValidationError):
+                                                        parse_err = pe
+                                                except ImportError:
+                                                    pass
+
+                                        if parsed_result is None:
+                                            # If none of the candidates parsed, raise last error
+                                            if parse_err is not None:
+                                                raise parse_err
+                                            raise ValueError("Unable to parse any JSON candidates")
+                                    except (json.JSONDecodeError, ValidationError, ValueError, Exception) as extraction_error:
+                                        # Catch generic Exception to handle jsonschema errors without explicit import here
                                         if verbose:
-                                            logger.debug(f"[DEBUG] JSON block extraction/validation failed ('{extraction_error}'). Trying markdown cleaning.")
-                                        # Fallback: Clean markdown fences and retry JSON validation
+                                            logger.debug(f"[DEBUG] JSON extraction/validation failed ('{extraction_error}'). Trying fence cleaning.")
+                                        # Last resort: strip any leading/trailing code fences and retry
                                         cleaned_result_str = raw_result.strip()
                                         if cleaned_result_str.startswith("```json"):
                                             cleaned_result_str = cleaned_result_str[7:]
@@ -1135,35 +2140,166 @@ def llm_invoke(
                                         if cleaned_result_str.endswith("```"):
                                             cleaned_result_str = cleaned_result_str[:-3]
                                         cleaned_result_str = cleaned_result_str.strip()
-                                        # Check again if it looks like JSON before parsing
-                                        if cleaned_result_str.startswith('{') and cleaned_result_str.endswith('}'):
+                                        # Check for complete JSON object or array
+                                        is_complete_object = cleaned_result_str.startswith('{') and cleaned_result_str.endswith('}')
+                                        is_complete_array = cleaned_result_str.startswith('[') and cleaned_result_str.endswith(']')
+                                        if is_complete_object or is_complete_array:
                                             if verbose:
-                                                logger.debug(f"[DEBUG] Attempting parse after cleaning markdown fences. Cleaned string: '{cleaned_result_str}'")
-                                            json_string_to_parse = cleaned_result_str # Update string for error reporting
-                                            parsed_result = output_pydantic.model_validate_json(json_string_to_parse)
+                                                logger.debug(f"[DEBUG] Attempting parse after generic fence cleaning. Cleaned string: '{cleaned_result_str}'")
+                                            json_string_to_parse = cleaned_result_str
+
+                                            if output_pydantic:
+                                                parsed_result = output_pydantic.model_validate_json(json_string_to_parse)
+                                            else:
+                                                loaded = json.loads(json_string_to_parse)
+                                                try:
+                                                    import jsonschema
+                                                    jsonschema.validate(instance=loaded, schema=output_schema)
+                                                except ImportError:
+                                                    pass
+                                                parsed_result = json_string_to_parse
+                                        elif cleaned_result_str.startswith('{') or cleaned_result_str.startswith('['):
+                                            # Attempt to repair truncated JSON (e.g., missing closing braces)
+                                            # This can happen when Gemini generates excessive trailing content
+                                            # that causes token limit truncation
+                                            if verbose:
+                                                logger.debug(f"[DEBUG] JSON appears truncated (missing closing brace). Attempting repair.")
+
+                                            # Try to find the last valid JSON structure
+                                            # For simple schemas like {"extracted_code": "..."}, we can try to close it
+                                            repaired = cleaned_result_str.rstrip()
+
+                                            # Strip trailing escaped newline sequences (\\n in the JSON string)
+                                            # These appear as literal backslash-n when Gemini generates excessive newlines
+                                            while repaired.endswith('\\n'):
+                                                repaired = repaired[:-2]
+                                            # Also strip trailing literal backslashes that might be orphaned
+                                            repaired = repaired.rstrip('\\')
+
+                                            # If we're in the middle of a string value, try to close it
+                                            # Count unescaped quotes to determine if we're inside a string
+                                            # Simple heuristic: if it ends without proper closure, add closing
+                                            is_array = cleaned_result_str.startswith('[')
+                                            expected_end = ']' if is_array else '}'
+                                            if not repaired.endswith(expected_end):
+                                                # Try adding various closures to repair
+                                                if is_array:
+                                                    repair_attempts = [
+                                                        repaired + '}]',  # Close object and array
+                                                        repaired + '"}]',  # Close string, object and array
+                                                        repaired + '"}}]',  # Close string, nested object and array
+                                                        repaired.rstrip(',') + ']',  # Remove trailing comma and close array
+                                                        repaired.rstrip('"') + '"}]',  # Handle partial string end
+                                                    ]
+                                                else:
+                                                    repair_attempts = [
+                                                        repaired + '"}',  # Close string and object
+                                                        repaired + '"}\n}',  # Close string and nested object
+                                                        repaired + '"}}}',  # Deeper nesting
+                                                        repaired.rstrip(',') + '}',  # Remove trailing comma
+                                                        repaired.rstrip('"') + '"}',  # Handle partial string end
+                                                    ]
+
+                                                for attempt in repair_attempts:
+                                                    try:
+                                                        if output_pydantic:
+                                                            parsed_result = output_pydantic.model_validate_json(attempt)
+                                                        else:
+                                                            loaded = json.loads(attempt)
+                                                            try:
+                                                                import jsonschema
+                                                                jsonschema.validate(instance=loaded, schema=output_schema)
+                                                            except ImportError:
+                                                                pass
+                                                            parsed_result = attempt
+
+                                                        if verbose:
+                                                            logger.info(f"[INFO] Successfully repaired truncated JSON response")
+                                                        json_string_to_parse = attempt
+                                                        break
+                                                    except (json.JSONDecodeError, ValidationError, ValueError):
+                                                        continue
+
+                                                if parsed_result is None:
+                                                    raise ValueError("Content after cleaning doesn't look like JSON (and repair attempts failed)")
                                         else:
-                                            # If still doesn't look like JSON, raise error
-                                            raise ValueError("Content after cleaning markdown doesn't look like JSON")
+                                            raise ValueError("Content after cleaning doesn't look like JSON")
 
 
                                 # Check if any parsing attempt succeeded
                                 if parsed_result is None:
+                                    target_name = output_pydantic.__name__ if output_pydantic else "JSON Schema"
                                     # This case should ideally be caught by exceptions above, but as a safeguard:
-                                    raise TypeError(f"Raw result type {type(raw_result)} or content could not be validated/parsed against {output_pydantic.__name__}.")
+                                    raise TypeError(f"Raw result type {type(raw_result)} or content could not be validated/parsed against {target_name}.")
 
-                            except (ValidationError, json.JSONDecodeError, TypeError, ValueError) as parse_error:
-                                logger.error(f"[ERROR] Failed to parse response into Pydantic model {output_pydantic.__name__} for item {i}: {parse_error}")
+                            except (ValidationError, json.JSONDecodeError, TypeError, ValueError, Exception) as parse_error:
+                                target_name = output_pydantic.__name__ if output_pydantic else "JSON Schema"
+                                logger.error(f"[ERROR] Failed to parse response into {target_name} for item {i}: {parse_error}")
                                 # Use the string that was last attempted for parsing in the error message
                                 error_content = json_string_to_parse if json_string_to_parse is not None else raw_result
                                 logger.error("[ERROR] Content attempted for parsing: %s", repr(error_content)) # CORRECTED (or use f-string)
-                                results.append(f"ERROR: Failed to parse Pydantic. Raw: {repr(raw_result)}")
+                                results.append(f"ERROR: Failed to parse structured output. Raw: {repr(raw_result)}")
                                 continue # Skip appending result below if parsing failed
 
-                            # If parsing succeeded, append the parsed_result
+                            # Post-process: unescape newlines and repair Python syntax
+                            _unescape_code_newlines(parsed_result)
+
+                            # Check if code fields still have invalid Python syntax after repair
+                            # If so, retry without cache to get a fresh response
+                            if _has_invalid_python_code(parsed_result):
+                                logger.warning(f"[WARNING] Detected invalid Python syntax in code fields for item {i} after repair. Retrying with cache bypass...")
+                                if not use_batch_mode and prompt and input_json is not None:
+                                    # Add a small variation to bypass cache
+                                    modified_prompt = prompt + "  "  # Two spaces to differentiate from other retries
+                                    try:
+                                        retry_messages = _format_messages(modified_prompt, input_json, use_batch_mode)
+                                        # Disable cache for retry
+                                        original_cache = litellm.cache
+                                        litellm.cache = None
+                                        retry_response = litellm.completion(
+                                            model=model_name_litellm,
+                                            messages=retry_messages,
+                                            temperature=current_temperature,
+                                            response_format=response_format,
+                                            **time_kwargs
+                                        )
+                                        # Re-enable cache
+                                        litellm.cache = original_cache
+                                        # Extract and re-parse the retry result
+                                        retry_raw_result = retry_response.choices[0].message.content
+                                        if retry_raw_result is not None:
+                                            # Re-parse the retry result
+                                            retry_parsed = None
+                                            if output_pydantic:
+                                                if isinstance(retry_raw_result, output_pydantic):
+                                                    retry_parsed = retry_raw_result
+                                                elif isinstance(retry_raw_result, dict):
+                                                    retry_parsed = output_pydantic.model_validate(retry_raw_result)
+                                                elif isinstance(retry_raw_result, str):
+                                                    retry_parsed = output_pydantic.model_validate_json(retry_raw_result)
+                                            elif output_schema and isinstance(retry_raw_result, str):
+                                                retry_parsed = retry_raw_result  # Keep as string for schema validation
+
+                                            if retry_parsed is not None:
+                                                _unescape_code_newlines(retry_parsed)
+                                                if not _has_invalid_python_code(retry_parsed):
+                                                    logger.info(f"[SUCCESS] Cache bypass retry for invalid Python code succeeded for item {i}")
+                                                    parsed_result = retry_parsed
+                                                else:
+                                                    logger.warning(f"[WARNING] Cache bypass retry still has invalid Python code for item {i}, using original")
+                                            else:
+                                                logger.warning(f"[WARNING] Cache bypass retry returned unparseable result for item {i}")
+                                        else:
+                                            logger.warning(f"[WARNING] Cache bypass retry returned None for item {i}")
+                                    except Exception as retry_e:
+                                        logger.warning(f"[WARNING] Cache bypass retry for invalid Python code failed for item {i}: {retry_e}")
+                                else:
+                                    logger.warning(f"[WARNING] Cannot retry invalid Python code - batch mode or missing prompt/input_json")
+
                             results.append(parsed_result)
 
                         else:
-                            # If output_pydantic was not requested, append the raw result
+                            # If output_pydantic/schema was not requested, append the raw result
                             results.append(raw_result)
 
                     except (AttributeError, IndexError) as e:
@@ -1246,10 +2382,40 @@ def llm_invoke(
                     Exception) as e: # Catch generic Exception last
                 last_exception = e
                 error_type = type(e).__name__
+                error_str = str(e)
+
+                # Provider-specific handling for Anthropic temperature + thinking rules.
+                # Two scenarios we auto-correct:
+                # 1) temperature==1 without thinking -> retry with 0.99
+                # 2) thinking enabled but temperature!=1 -> retry with 1
+                lower_err = error_str.lower()
+                if (not temp_adjustment_done) and ("temperature" in lower_err) and ("thinking" in lower_err):
+                    anthropic_thinking_sent = ('thinking' in litellm_kwargs) and (provider.lower() == 'anthropic')
+                    # Decide direction of adjustment based on whether thinking was enabled in the call
+                    if anthropic_thinking_sent:
+                        # thinking enabled -> force temperature=1
+                        adjusted_temp = 1
+                        logger.warning(
+                            f"[WARN] {model_name_litellm}: Anthropic with thinking requires temperature=1. "
+                            f"Retrying with temperature={adjusted_temp}."
+                        )
+                    else:
+                        # thinking not enabled -> avoid temperature=1
+                        adjusted_temp = 0.99
+                        logger.warning(
+                            f"[WARN] {model_name_litellm}: Provider rejected temperature=1 without thinking. "
+                            f"Retrying with temperature={adjusted_temp}."
+                        )
+                    current_temperature = adjusted_temp
+                    temp_adjustment_done = True
+                    retry_with_same_model = True
+                    if verbose:
+                        logger.debug(f"Retrying {model_name_litellm} with adjusted temperature {current_temperature}")
+                    continue
+
                 logger.error(f"[ERROR] Invocation failed for {model_name_litellm} ({error_type}): {e}. Trying next model.")
                 # Log more details in verbose mode
                 if verbose:
-                    # import traceback # Not needed if using exc_info=True
                     logger.debug(f"Detailed exception traceback for {model_name_litellm}:", exc_info=True)
                 break # Break inner loop, try next model candidate
 
@@ -1277,7 +2443,7 @@ if __name__ == "__main__":
         response = llm_invoke(
             prompt="Tell me a short joke about {topic}.",
             input_json={"topic": "programmers"},
-            strength=0.5, # Use base model (gpt-4.1-nano)
+            strength=0.5, # Use base model (gpt-5-nano)
             temperature=0.7,
             verbose=True
         )
@@ -1358,7 +2524,7 @@ if __name__ == "__main__":
             {"role": "system", "content": "You are a helpful assistant."},
             {"role": "user", "content": "What is the capital of France?"}
         ]
-        # Strength 0.5 should select gpt-4.1-nano
+        # Strength 0.5 should select gpt-5-nano
         response_messages = llm_invoke(
             messages=custom_messages,
             strength=0.5,