from __future__ import annotations import json import re from dataclasses import dataclass from typing import Any SECTION_MARKER_RE = re.compile(r"(?im)^[ \t]*(thought|action|final(?: answer)?)\s*:\s*") @dataclass(frozen=True, slots=True) class ParsedModelReply: thought: str | None actions: list[dict[str, Any]] | None = None final_answer: str | None = None @property def is_final(self) -> bool: return self.final_answer is not None def parse_model_reply(text: str) -> ParsedModelReply: """Parse a model turn using the visible Thought + Action/Final protocol. Preferred formats:: Thought: ... Action: {"tool": "read", "arguments": {"path": "data/a.txt"}} Thought: ... Final: Done. The legacy JSON-only tool-call format is still accepted for compatibility. """ action_match = find_section_marker(text, {"action"}) if action_match is not None: action_text = text[action_match.end() :].strip() actions = extract_tool_actions(action_text) return ParsedModelReply(thought=extract_thought(text, stop_match=action_match), actions=actions) final_match = find_section_marker(text, {"final", "final answer"}) if final_match is not None: final_answer = text[final_match.end() :].strip() if not final_answer: raise ValueError("Final section is empty") return ParsedModelReply( thought=extract_thought(text, stop_match=final_match), final_answer=final_answer, ) try: return ParsedModelReply(thought=extract_prefix_thought(text), actions=extract_tool_actions(text)) except ValueError: if is_plain_final_text(text): return ParsedModelReply(thought=None, final_answer=text.strip()) raise def find_section_marker(text: str, names: set[str]) -> re.Match[str] | None: normalized_names = {name.lower() for name in names} for match in SECTION_MARKER_RE.finditer(text): if match.group(1).lower() in normalized_names: return match return None def extract_thought(text: str, *, stop_match: re.Match[str] | None) -> str | None: thought_match = find_section_marker(text, {"thought"}) if thought_match is not None: end = stop_match.start() if stop_match is not None and stop_match.start() > thought_match.end() else len(text) thought = text[thought_match.end() : end].strip() return thought or None if stop_match is not None: prefix = text[: stop_match.start()].strip() return prefix or None return None def extract_prefix_thought(text: str) -> str | None: json_start = find_first_json_start(text) if json_start is None: return None prefix = text[:json_start].strip() return prefix or None def is_plain_final_text(text: str) -> bool: stripped = text.strip() if not stripped: return False if stripped.startswith("```"): return False return stripped[0] not in "[{" def extract_tool_actions(text: str) -> list[dict[str, Any]]: parsed = extract_json_value(text) if isinstance(parsed, dict): if "actions" in parsed: raw_actions = parsed["actions"] elif "tool_calls" in parsed: raw_actions = parsed["tool_calls"] else: raw_actions = [parsed] elif isinstance(parsed, list): raw_actions = parsed else: raise ValueError("tool call JSON must be an object or array") if not isinstance(raw_actions, list) or not raw_actions: raise ValueError("actions must be a non-empty array") actions: list[dict[str, Any]] = [] for index, raw_action in enumerate(raw_actions, start=1): if not isinstance(raw_action, dict): raise ValueError(f"action #{index} must be an object") actions.append(normalize_tool_action(raw_action)) return actions def normalize_tool_action(raw_action: dict[str, Any]) -> dict[str, Any]: if "function" in raw_action and isinstance(raw_action["function"], dict): function = raw_action["function"] return normalize_named_tool_action(function.get("name"), function.get("arguments", {})) if "tool" in raw_action or "name" in raw_action: return normalize_named_tool_action( raw_action.get("tool", raw_action.get("name")), raw_action.get("arguments", {}), ) if "action" in raw_action: return dict(raw_action) raise ValueError("action object must contain 'tool', 'name', 'function', or 'action'") def normalize_named_tool_action(name: Any, arguments: Any) -> dict[str, Any]: if not isinstance(name, str) or not name.strip(): raise ValueError("tool name must be a non-empty string") if isinstance(arguments, str): try: arguments = json.loads(arguments) if arguments.strip() else {} except json.JSONDecodeError as exc: raise ValueError(f"tool arguments are not valid JSON: {exc}") from exc if arguments is None: arguments = {} if not isinstance(arguments, dict): raise ValueError("tool arguments must be an object") action = dict(arguments) action["action"] = name.strip() return action def extract_json_value(text: str) -> Any: stripped = strip_code_fence(text.strip()) try: return json.loads(stripped) except json.JSONDecodeError: return json.loads(find_first_json_value(stripped)) def strip_code_fence(text: str) -> str: if not text.startswith("```"): return text lines = text.splitlines() if len(lines) >= 3 and lines[-1].strip() == "```": return "\n".join(lines[1:-1]).strip() return text def find_first_json_start(text: str) -> int | None: object_start = text.find("{") array_start = text.find("[") starts = [index for index in (object_start, array_start) if index != -1] return min(starts) if starts else None def find_first_json_value(text: str) -> str: start = find_first_json_start(text) if start is None: raise ValueError("no JSON object or array found") opening = text[start] closing = "}" if opening == "{" else "]" depth = 0 in_string = False escaped = False for index in range(start, len(text)): char = text[index] if in_string: if escaped: escaped = False elif char == "\\": escaped = True elif char == '"': in_string = False continue if char == '"': in_string = True elif char == opening: depth += 1 elif char == closing: depth -= 1 if depth == 0: return text[start : index + 1] raise ValueError("unterminated JSON value")