File size: 9,716 Bytes
6172a47 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 | import re
import uuid
from enum import Enum
from typing import Any
from loguru import logger
# Some OpenAI-compatible backends/models occasionally leak internal sentinel tokens
# into `delta.content` (e.g. "<|tool_call_end|>"). These should never be shown to
# end users, and they can disrupt downstream parsing if left in place.
_CONTROL_TOKEN_RE = re.compile(r"<\|[^|>]{1,80}\|>")
_CONTROL_TOKEN_START = "<|"
_CONTROL_TOKEN_END = "|>"
class ParserState(Enum):
TEXT = 1
MATCHING_FUNCTION = 2
PARSING_PARAMETERS = 3
class HeuristicToolParser:
"""
Stateful parser that detects raw text tool calls in the format:
● <function=Name><parameter=key>value</parameter>...
This is used as a fallback for models that emit tool calls as text
instead of using the structured API.
"""
# Class-level compiled patterns (compiled once, not per instance)
_FUNC_START_PATTERN = re.compile(r"●\s*<function=([^>]+)>")
_PARAM_PATTERN = re.compile(
r"<parameter=([^>]+)>(.*?)(?:</parameter>|$)", re.DOTALL
)
def __init__(self):
self._state = ParserState.TEXT
self._buffer = ""
self._current_tool_id = None
self._current_function_name = None
self._current_parameters = {}
def _strip_control_tokens(self, text: str) -> str:
# Remove complete sentinel tokens. If a token is split across chunks it
# will be removed once the buffer contains the full token.
return _CONTROL_TOKEN_RE.sub("", text)
def _split_incomplete_control_token_tail(self) -> str:
"""
If the buffer ends with an incomplete "<|...|>" sentinel token, keep that
fragment in the buffer and return the safe-to-emit prefix.
This prevents leaking raw sentinel fragments to the user when streaming.
"""
start = self._buffer.rfind(_CONTROL_TOKEN_START)
if start == -1:
return ""
end = self._buffer.find(_CONTROL_TOKEN_END, start)
if end != -1:
return ""
prefix = self._buffer[:start]
self._buffer = self._buffer[start:]
return prefix
def feed(self, text: str) -> tuple[str, list[dict[str, Any]]]:
"""
Feed text into the parser.
Returns a tuple of (filtered_text, detected_tool_calls).
filtered_text: Text that should be passed through as normal message content.
detected_tools: List of Anthropic-format tool_use blocks.
"""
self._buffer += text
self._buffer = self._strip_control_tokens(self._buffer)
detected_tools = []
filtered_output_parts: list[str] = []
while True:
if self._state == ParserState.TEXT:
# Look for the trigger character
if "●" in self._buffer:
idx = self._buffer.find("●")
filtered_output_parts.append(self._buffer[:idx])
self._buffer = self._buffer[idx:]
self._state = ParserState.MATCHING_FUNCTION
else:
# Avoid emitting an incomplete "<|...|>" sentinel fragment if the
# token got split across streaming chunks.
safe_prefix = self._split_incomplete_control_token_tail()
if safe_prefix:
filtered_output_parts.append(safe_prefix)
break
filtered_output_parts.append(self._buffer)
self._buffer = ""
break
if self._state == ParserState.MATCHING_FUNCTION:
# We need enough buffer to match the function tag
# e.g. "● <function=Grep>"
match = self._FUNC_START_PATTERN.search(self._buffer)
if match:
self._current_function_name = match.group(1).strip()
self._current_tool_id = f"toolu_heuristic_{uuid.uuid4().hex[:8]}"
self._current_parameters = {}
# Consume the function start from buffer
self._buffer = self._buffer[match.end() :]
self._state = ParserState.PARSING_PARAMETERS
logger.debug(
"Heuristic bypass: Detected start of tool call '{}'",
self._current_function_name,
)
else:
# If we have "●" but not the full tag yet, wait for more data
# Unless the buffer has grown too large without a match
if len(self._buffer) > 100:
# Probably not a tool call, treat as text
filtered_output_parts.append(self._buffer[0])
self._buffer = self._buffer[1:]
self._state = ParserState.TEXT
else:
break
if self._state == ParserState.PARSING_PARAMETERS:
# Look for parameters. We look for </parameter> to know a param is complete.
# Or wait for another <parameter or the end of the text if it seems complete.
# If we see a newline followed by anything other than <parameter or spaces,
# we might be done with the tool call.
finished_tool_call = False
# Check if we have any complete parameters
while True:
param_match = self._PARAM_PATTERN.search(self._buffer)
if param_match and "</parameter>" in param_match.group(0):
# Detect any content before the parameter match and preserve it
pre_match_text = self._buffer[: param_match.start()]
if pre_match_text:
filtered_output_parts.append(pre_match_text)
key = param_match.group(1).strip()
val = param_match.group(2).strip()
self._current_parameters[key] = val
self._buffer = self._buffer[param_match.end() :]
else:
break
# Heuristic for completion:
# 1. We have at least one param and we see a character that doesn't belong to the format
# 2. Significant pause (not handled here, handled by caller via flush if needed)
# 3. Another ● character (start of NEXT tool call)
if "●" in self._buffer:
# Next tool call starting or something else, close current
# But first, capture any text before the ●
idx = self._buffer.find("●")
if idx > 0:
filtered_output_parts.append(self._buffer[:idx])
self._buffer = self._buffer[idx:]
finished_tool_call = True
elif len(self._buffer) > 0 and not self._buffer.strip().startswith("<"):
# We have text that doesn't look like a tag, and we already parsed some or are in param state
# Let's see if we have trailing param starts
if "<parameter=" not in self._buffer:
# Treat the buffer as text (it's not a parameter)
# But wait, we are in PARSING_PARAMETERS.
# If we have " some text", we should emit it and finish tool call.
filtered_output_parts.append(self._buffer)
self._buffer = ""
finished_tool_call = True
if finished_tool_call:
# Emit the tool call
detected_tools.append(
{
"type": "tool_use",
"id": self._current_tool_id,
"name": self._current_function_name,
"input": self._current_parameters,
}
)
logger.debug(
"Heuristic bypass: Emitting tool call '{}' with {} params",
self._current_function_name,
len(self._current_parameters),
)
self._state = ParserState.TEXT
# Continue loop to process remaining buffer (which is empty or starts with ●)
else:
break
return "".join(filtered_output_parts), detected_tools
def flush(self) -> list[dict[str, Any]]:
"""
Flush any remaining tool calls in the buffer.
"""
self._buffer = self._strip_control_tokens(self._buffer)
detected_tools = []
if self._state == ParserState.PARSING_PARAMETERS:
# Try to extract any partial parameters remaining in buffer
# Even without </parameter>
partial_matches = re.finditer(
r"<parameter=([^>]+)>(.*)$", self._buffer, re.DOTALL
)
for m in partial_matches:
key = m.group(1).strip()
val = m.group(2).strip()
self._current_parameters[key] = val
detected_tools.append(
{
"type": "tool_use",
"id": self._current_tool_id,
"name": self._current_function_name,
"input": self._current_parameters,
}
)
self._state = ParserState.TEXT
self._buffer = ""
return detected_tools
|