Spaces:
Sleeping
Sleeping
Moving more things over into agent_server
Browse files- agent_server/formatting_reasoning.py +98 -0
- proxy.py +2 -96
agent_server/formatting_reasoning.py
ADDED
|
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import json
|
| 2 |
+
import re
|
| 3 |
+
import typing
|
| 4 |
+
|
| 5 |
+
from agent_server.sanitizing_think_tags import scrub_think_tags
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def _format_reasoning_chunk(text: str, tag: str, idx: int) -> str:
|
| 9 |
+
"""
|
| 10 |
+
Lightweight formatter for reasoning stream. Avoid huge code fences;
|
| 11 |
+
make it readable and incremental. Also filters out ASCII/box-drawing noise.
|
| 12 |
+
"""
|
| 13 |
+
text = scrub_think_tags(text).rstrip("\n")
|
| 14 |
+
if not text:
|
| 15 |
+
return ""
|
| 16 |
+
noisy_prefixes = (
|
| 17 |
+
"OpenAIServerModel",
|
| 18 |
+
"Output message of the LLM",
|
| 19 |
+
"─ Executing parsed code",
|
| 20 |
+
"New run",
|
| 21 |
+
"╭",
|
| 22 |
+
"╰",
|
| 23 |
+
"│",
|
| 24 |
+
"━",
|
| 25 |
+
"─",
|
| 26 |
+
)
|
| 27 |
+
stripped = text.strip()
|
| 28 |
+
if not stripped:
|
| 29 |
+
return ""
|
| 30 |
+
# Lines made mostly of box drawing/separators
|
| 31 |
+
if all(ch in " ─━╭╮╰╯│═·—-_=+•" for ch in stripped):
|
| 32 |
+
return ""
|
| 33 |
+
if any(stripped.startswith(p) for p in noisy_prefixes):
|
| 34 |
+
return ""
|
| 35 |
+
# Excessively long lines with little signal (no alphanumerics)
|
| 36 |
+
if len(stripped) > 240 and not re.search(r"[A-Za-z0-9]{3,}", stripped):
|
| 37 |
+
return ""
|
| 38 |
+
# No tag/idx prefix; add a trailing blank line for readability in markdown
|
| 39 |
+
return f"{stripped}\n\n"
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def _extract_final_text(item: typing.Any) -> typing.Optional[str]:
|
| 43 |
+
if isinstance(item, dict) and ("__stdout__" in item or "__step__" in item):
|
| 44 |
+
return None
|
| 45 |
+
if isinstance(item, (bytes, bytearray)):
|
| 46 |
+
try:
|
| 47 |
+
item = item.decode("utf-8", errors="ignore")
|
| 48 |
+
except Exception:
|
| 49 |
+
item = str(item)
|
| 50 |
+
if isinstance(item, str):
|
| 51 |
+
s = scrub_think_tags(item.strip())
|
| 52 |
+
return s or None
|
| 53 |
+
# If it's a step-like object with an 'output' attribute, use that
|
| 54 |
+
try:
|
| 55 |
+
if not isinstance(item, (dict, list, bytes, bytearray)):
|
| 56 |
+
out = getattr(item, "output", None)
|
| 57 |
+
if out is not None:
|
| 58 |
+
s = scrub_think_tags(str(out)).strip()
|
| 59 |
+
if s:
|
| 60 |
+
return s
|
| 61 |
+
except Exception:
|
| 62 |
+
pass
|
| 63 |
+
if isinstance(item, dict):
|
| 64 |
+
for key in ("content", "text", "message", "output", "final", "answer"):
|
| 65 |
+
if key in item:
|
| 66 |
+
val = item[key]
|
| 67 |
+
if isinstance(val, (dict, list)):
|
| 68 |
+
try:
|
| 69 |
+
return scrub_think_tags(json.dumps(val, ensure_ascii=False))
|
| 70 |
+
except Exception:
|
| 71 |
+
return scrub_think_tags(str(val))
|
| 72 |
+
if isinstance(val, (bytes, bytearray)):
|
| 73 |
+
try:
|
| 74 |
+
val = val.decode("utf-8", errors="ignore")
|
| 75 |
+
except Exception:
|
| 76 |
+
val = str(val)
|
| 77 |
+
s = scrub_think_tags(str(val).strip())
|
| 78 |
+
return s or None
|
| 79 |
+
try:
|
| 80 |
+
return scrub_think_tags(json.dumps(item, ensure_ascii=False))
|
| 81 |
+
except Exception:
|
| 82 |
+
return scrub_think_tags(str(item))
|
| 83 |
+
try:
|
| 84 |
+
return scrub_think_tags(str(item))
|
| 85 |
+
except Exception:
|
| 86 |
+
return None
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
_FINAL_RE = re.compile(r"(?:^|\\b)Final\\s+answer:\\s*(.+)$", flags=re.IGNORECASE)
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
def _maybe_parse_final_from_stdout(line: str) -> typing.Optional[str]:
|
| 93 |
+
if not isinstance(line, str):
|
| 94 |
+
return None
|
| 95 |
+
m = _FINAL_RE.search(line.strip())
|
| 96 |
+
if not m:
|
| 97 |
+
return None
|
| 98 |
+
return scrub_think_tags(m.group(1)).strip() or None
|
proxy.py
CHANGED
|
@@ -19,6 +19,8 @@ import contextlib
|
|
| 19 |
# Upstream pass-through
|
| 20 |
import httpx
|
| 21 |
|
|
|
|
|
|
|
| 22 |
from agent_server.helpers import normalize_content_to_text, _messages_to_task, _openai_response, _sse_headers
|
| 23 |
from agent_server.openai_schemas import ChatMessage, ChatCompletionRequest
|
| 24 |
from agent_server.sanitizing_think_tags import scrub_think_tags
|
|
@@ -57,102 +59,6 @@ app = fastapi.FastAPI()
|
|
| 57 |
async def healthz():
|
| 58 |
return {"ok": True}
|
| 59 |
|
| 60 |
-
|
| 61 |
-
# ---------- Reasoning formatting for Chat-UI ----------
|
| 62 |
-
def _format_reasoning_chunk(text: str, tag: str, idx: int) -> str:
|
| 63 |
-
"""
|
| 64 |
-
Lightweight formatter for reasoning stream. Avoid huge code fences;
|
| 65 |
-
make it readable and incremental. Also filters out ASCII/box-drawing noise.
|
| 66 |
-
"""
|
| 67 |
-
text = scrub_think_tags(text).rstrip("\n")
|
| 68 |
-
if not text:
|
| 69 |
-
return ""
|
| 70 |
-
noisy_prefixes = (
|
| 71 |
-
"OpenAIServerModel",
|
| 72 |
-
"Output message of the LLM",
|
| 73 |
-
"─ Executing parsed code",
|
| 74 |
-
"New run",
|
| 75 |
-
"╭",
|
| 76 |
-
"╰",
|
| 77 |
-
"│",
|
| 78 |
-
"━",
|
| 79 |
-
"─",
|
| 80 |
-
)
|
| 81 |
-
stripped = text.strip()
|
| 82 |
-
if not stripped:
|
| 83 |
-
return ""
|
| 84 |
-
# Lines made mostly of box drawing/separators
|
| 85 |
-
if all(ch in " ─━╭╮╰╯│═·—-_=+•" for ch in stripped):
|
| 86 |
-
return ""
|
| 87 |
-
if any(stripped.startswith(p) for p in noisy_prefixes):
|
| 88 |
-
return ""
|
| 89 |
-
# Excessively long lines with little signal (no alphanumerics)
|
| 90 |
-
if len(stripped) > 240 and not re.search(r"[A-Za-z0-9]{3,}", stripped):
|
| 91 |
-
return ""
|
| 92 |
-
# No tag/idx prefix; add a trailing blank line for readability in markdown
|
| 93 |
-
return f"{stripped}\n\n"
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
def _extract_final_text(item: typing.Any) -> typing.Optional[str]:
|
| 97 |
-
if isinstance(item, dict) and ("__stdout__" in item or "__step__" in item):
|
| 98 |
-
return None
|
| 99 |
-
if isinstance(item, (bytes, bytearray)):
|
| 100 |
-
try:
|
| 101 |
-
item = item.decode("utf-8", errors="ignore")
|
| 102 |
-
except Exception:
|
| 103 |
-
item = str(item)
|
| 104 |
-
if isinstance(item, str):
|
| 105 |
-
s = scrub_think_tags(item.strip())
|
| 106 |
-
return s or None
|
| 107 |
-
# If it's a step-like object with an 'output' attribute, use that
|
| 108 |
-
try:
|
| 109 |
-
if not isinstance(item, (dict, list, bytes, bytearray)):
|
| 110 |
-
out = getattr(item, "output", None)
|
| 111 |
-
if out is not None:
|
| 112 |
-
s = scrub_think_tags(str(out)).strip()
|
| 113 |
-
if s:
|
| 114 |
-
return s
|
| 115 |
-
except Exception:
|
| 116 |
-
pass
|
| 117 |
-
if isinstance(item, dict):
|
| 118 |
-
for key in ("content", "text", "message", "output", "final", "answer"):
|
| 119 |
-
if key in item:
|
| 120 |
-
val = item[key]
|
| 121 |
-
if isinstance(val, (dict, list)):
|
| 122 |
-
try:
|
| 123 |
-
return scrub_think_tags(json.dumps(val, ensure_ascii=False))
|
| 124 |
-
except Exception:
|
| 125 |
-
return scrub_think_tags(str(val))
|
| 126 |
-
if isinstance(val, (bytes, bytearray)):
|
| 127 |
-
try:
|
| 128 |
-
val = val.decode("utf-8", errors="ignore")
|
| 129 |
-
except Exception:
|
| 130 |
-
val = str(val)
|
| 131 |
-
s = scrub_think_tags(str(val).strip())
|
| 132 |
-
return s or None
|
| 133 |
-
try:
|
| 134 |
-
return scrub_think_tags(json.dumps(item, ensure_ascii=False))
|
| 135 |
-
except Exception:
|
| 136 |
-
return scrub_think_tags(str(item))
|
| 137 |
-
try:
|
| 138 |
-
return scrub_think_tags(str(item))
|
| 139 |
-
except Exception:
|
| 140 |
-
return None
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
# Helper to parse explicit "Final answer:" from stdout lines
|
| 144 |
-
_FINAL_RE = re.compile(r"(?:^|\\b)Final\\s+answer:\\s*(.+)$", flags=re.IGNORECASE)
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
def _maybe_parse_final_from_stdout(line: str) -> typing.Optional[str]:
|
| 148 |
-
if not isinstance(line, str):
|
| 149 |
-
return None
|
| 150 |
-
m = _FINAL_RE.search(line.strip())
|
| 151 |
-
if not m:
|
| 152 |
-
return None
|
| 153 |
-
return scrub_think_tags(m.group(1)).strip() or None
|
| 154 |
-
|
| 155 |
-
|
| 156 |
# ---------- Live stdout/stderr tee ----------
|
| 157 |
class QueueWriter(io.TextIOBase):
|
| 158 |
"""
|
|
|
|
| 19 |
# Upstream pass-through
|
| 20 |
import httpx
|
| 21 |
|
| 22 |
+
from agent_server.formatting_reasoning import _format_reasoning_chunk, _extract_final_text, \
|
| 23 |
+
_maybe_parse_final_from_stdout
|
| 24 |
from agent_server.helpers import normalize_content_to_text, _messages_to_task, _openai_response, _sse_headers
|
| 25 |
from agent_server.openai_schemas import ChatMessage, ChatCompletionRequest
|
| 26 |
from agent_server.sanitizing_think_tags import scrub_think_tags
|
|
|
|
| 59 |
async def healthz():
|
| 60 |
return {"ok": True}
|
| 61 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 62 |
# ---------- Live stdout/stderr tee ----------
|
| 63 |
class QueueWriter(io.TextIOBase):
|
| 64 |
"""
|