The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
"""Helpers for reasoning text that self-hosted models leave in message content."""
from __future__ import annotations
import json
import os
THINK_OPEN = "<think>"
THINK_CLOSE = "</think>"
_TRUTHY = {"1", "true", "yes", "on"}
def prefilled_reasoning_from_env(name: str) -> bool:
"""Read an explicit reasoning-format flag from the environment."""
return os.environ.get(name, "").strip().lower() in _TRUTHY
def strip_reasoning(
content: str, *, leading: bool = False, prefilled: bool = False
) -> str:
"""Drop a leading reasoning block from model output.
Some deployments put reasoning in ``content`` instead of a separate field.
This helper makes no assumptions about a hosted API or model name.
Content is preserved by default, including complete leading XML documents.
Enable ``leading=True`` only for a deployment whose declared format is
``<think>...</think>answer``. Unterminated reasoning remains unchanged.
``prefilled=True`` is for deployments whose chat template prefills the
opening tag in the generation prompt, so the model only ever emits
``...</think>answer``. Only enable it when the model is known to behave this
way: everything up to the first closing tag is then treated as reasoning.
"""
if not leading and not prefilled:
return content
stripped = content.lstrip()
if stripped.startswith(THINK_OPEN):
close = stripped.find(THINK_CLOSE, len(THINK_OPEN))
if close == -1:
return content
return stripped[close + len(THINK_CLOSE) :].lstrip()
if not prefilled:
return content
close = content.find(THINK_CLOSE)
if close == -1 or THINK_OPEN in content[:close] or _is_json_document(content):
return content
return content[close + len(THINK_CLOSE) :].lstrip()
def _is_json_document(content: str) -> bool:
"""Whether ``content`` is one complete JSON object or array.
Reasoning followed by a closing tag is never valid JSON, so a structured
answer that carries the tag inside a string value is left whole. Input
nested too deeply for the parser to finish is kept whole as well: reasoning
does not open with that many brackets, so cutting it would only risk
truncating a real answer.
"""
stripped = content.strip()
if not stripped.startswith(("{", "[")):
return False
try:
json.loads(stripped)
except RecursionError:
return True
except ValueError:
return False
return True