Salastil/odysseus
1
0
Fork 0
mirror of https://github.com/pewdiepie-archdaemon/odysseus.git synced 2026-06-17 02:05:22 -04:00
Code Issues Packages Projects Releases Wiki Activity

fix(search): add download budgets to web_fetch with truncation notice and hard ceiling (#3955)

Browse Source 
* fix(search): add download budgets to web_fetch with truncation notice and hard ceiling

MAX_OUTPUT_CHARS only trims what the agent sees; fetch_webpage_content
buffered and cached the entire response body first, so a large or hostile
URL could pull arbitrarily many bytes into memory and the content cache.

The fetch is now a capped streaming GET (SSRF redirect guard unchanged):
a soft default budget (WEB_FETCH_SOFT_MAX_BYTES, 2 MB), a per-call
override via full/max_bytes on the web_fetch tool, and a hard ceiling
(WEB_FETCH_HARD_MAX_BYTES, 20 MB) that the override can never exceed.
When Content-Length already declares a body over the ceiling the fetch
is refused before any body bytes are buffered. Truncated results carry
truncated/fetched_bytes/total_bytes, the tool output leads with a
partial-content notice telling the model how to re-fetch with full=true,
and the tool schema documents the flag. A truncated PDF is reported as
a budget error since a cut PDF is unparseable. The effective cap is part
of the content-cache key so a truncated fetch is never served to a
full-budget request.

Existing tests that faked httpx.get or the old _get_public_url signature
are adapted to the streaming interface; behavior pins are unchanged.

Fixes #3812

* fix(search): close compressed-body cap bypass and protect the partial notice

Addresses RaresKeY's review on #3955:

- Force Accept-Encoding: identity for the capped fetch. With gzip/deflate the
  wire bytes (and Content-Length) can be a fraction of the decoded body, so a
  tiny compressed response could pass the hard-cap preflight and then expand
  past the ceiling in a single decoded chunk before the streamed cap could
  slice it. Identity makes Content-Length the true body size and keeps each
  streamed chunk bounded by the network read, so the hard ceiling actually
  bounds memory.
- Lead web_fetch output with the partial-content notice and cap the page
  title. The notice is the user-facing contract for partial fetches, but the
  title is untrusted, uncapped page content; placed ahead of the notice a giant
  title could push it past MAX_OUTPUT_CHARS and drop it. The notice now leads
  and the title is capped as a second guard.

Adds regressions: the fetch advertises identity encoding, and a truncated
result with an oversized title still surfaces the partial notice.

* fix(search): reject compressed responses that ignore the identity request

Requesting Accept-Encoding: identity is not enough on its own: a server can
ignore it and still return Content-Encoding: gzip, and httpx.iter_bytes would
decode that, so a tiny compressed body could balloon into one decoded chunk
far past the hard cap before the streamed loop slices it (and Content-Length,
the compressed wire length, makes the preflight and size metadata unreliable).

Refuse a non-identity Content-Encoding before reading the body. Adds a
regression where the server ignores the identity request and returns gzip;
the fetch is refused before any body is decoded.
This commit is contained in:
Kenny Van de Maele
2026-06-15 19:38:09 +02:00
committed by GitHub
parent 2fab378c6a
commit 074a1e6eff
8 changed files with 422 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/search/content.py
+162 -13
View File
@@ -15,6 +15,8 @@ from urllib.parse import urljoin, urlparse
import httpx
from bs4 import BeautifulSoup
from src.constants import WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES
from .analytics import RateLimitError, error_logger
from .cache import (
CONTENT_CACHE_DIR,
@@ -89,18 +91,128 @@ def _public_http_url(url: str) -> bool:
return False
def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5) -> httpx.Response:
class BodyTooLargeError(Exception):
"""The server declared a body larger than the hard fetch ceiling."""
def __init__(self, url: str, declared_bytes: int):
self.url = url
self.declared_bytes = declared_bytes
super().__init__(
f"response body is {declared_bytes:,} bytes, over the "
f"{WEB_FETCH_HARD_MAX_BYTES:,}-byte hard cap"
)
class _CappedFetch:
"""Result of a size-capped streaming GET.
Carries just what fetch_webpage_content needs from an httpx.Response,
plus the cap bookkeeping: the (possibly truncated) body, whether the
cap cut it short, and the size the server declared via Content-Length
(wire bytes; None when absent).
"""
__slots__ = ("status_code", "headers", "content", "truncated",
"declared_bytes", "encoding", "url")
def __init__(self, status_code, headers, content, truncated,
declared_bytes, encoding, url):
self.status_code = status_code
self.headers = headers
self.content = content
self.truncated = truncated
self.declared_bytes = declared_bytes
self.encoding = encoding
self.url = url
@property
def text(self) -> str:
return self.content.decode(self.encoding or "utf-8", errors="replace")
def raise_for_status(self):
if self.status_code >= 400:
request = httpx.Request("GET", self.url)
raise httpx.HTTPStatusError(
f"HTTP {self.status_code} for {self.url}",
request=request,
response=httpx.Response(self.status_code, request=request),
)
def _get_public_url(url: str, headers: dict, timeout: int, max_redirects: int = 5,
max_bytes: int = None) -> "_CappedFetch":
"""Capped streaming GET with SSRF-guarded manual redirects.
The body is streamed and buffering stops at ``max_bytes`` (default: the
soft cap), so an oversized resource cannot be pulled into memory or the
content cache in full. When Content-Length already declares a body over
the hard ceiling, the fetch is refused before any body bytes are read.
"""
cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
current = url
for _ in range(max_redirects + 1):
if not _public_http_url(current):
raise httpx.RequestError("Blocked private/internal URL", request=httpx.Request("GET", current))
response = httpx.get(current, headers=headers, timeout=timeout, follow_redirects=False)
if response.status_code not in (301, 302, 303, 307, 308):
return response
location = response.headers.get("location")
if not location:
return response
current = urljoin(str(response.url), location)
# Force identity transfer-encoding. With gzip/deflate the wire bytes
# (and Content-Length) can be a small fraction of the decoded body, so
# a tiny compressed response could pass the hard-cap preflight and then
# expand past the ceiling in a single decoded chunk before the streamed
# cap below can slice it. Identity makes Content-Length the true body
# size and keeps each streamed chunk bounded by the network read.
req_headers = dict(headers or {})
req_headers["Accept-Encoding"] = "identity"
with httpx.stream("GET", current, headers=req_headers, timeout=timeout,
follow_redirects=False) as response:
if response.status_code in (301, 302, 303, 307, 308):
location = response.headers.get("location")
if not location:
return _CappedFetch(response.status_code, response.headers, b"",
False, None, response.encoding, str(response.url))
current = urljoin(str(response.url), location)
continue
# A server can ignore the identity request and still return a
# compressed body; httpx.iter_bytes would then decode it, and a tiny
# gzip can balloon into one decoded chunk far past the cap before we
# slice. Refuse a compressed Content-Encoding so the streamed cap
# stays a real memory bound (Content-Length is the compressed wire
# length here, so the preflight and size metadata are unreliable too).
enc = (response.headers.get("content-encoding") or "").strip().lower()
if enc and enc != "identity":
raise httpx.RequestError(
f"Refusing compressed response (Content-Encoding: {enc}) after "
"requesting identity: cannot bound decoded body size",
request=httpx.Request("GET", current),
)
declared = None
raw_len = response.headers.get("content-length")
if raw_len and raw_len.isdigit():
declared = int(raw_len)
# Refuse before buffering anything when the server already tells
# us the body exceeds the absolute ceiling (Content-Length is wire
# bytes; the decompressed body can only be larger).
if declared is not None and declared > WEB_FETCH_HARD_MAX_BYTES:
raise BodyTooLargeError(current, declared)
chunks = []
read = 0
truncated = False
# We requested identity above, so iter_bytes yields the raw body in
# network-read-sized chunks (no decompression expansion); the cap
# therefore bounds what we actually buffer.
for chunk in response.iter_bytes():
read += len(chunk)
if read > cap:
keep = cap - (read - len(chunk))
if keep > 0:
chunks.append(chunk[:keep])
truncated = True
break
chunks.append(chunk)
return _CappedFetch(response.status_code, response.headers,
b"".join(chunks), truncated, declared,
response.encoding, str(response.url))
raise httpx.RequestError("Too many redirects", request=httpx.Request("GET", current))
# PDF extraction (optional dependency)
@@ -222,9 +334,19 @@ def _empty_result(url: str, error: str = "") -> dict:
# ----------------------------------------------------------------------
# Main content fetcher
# ----------------------------------------------------------------------
def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) -> dict:
"""Fetch and extract meaningful content from a webpage with caching."""
cache_key = generate_cache_key(url)
def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0,
max_bytes: int = None) -> dict:
"""Fetch and extract meaningful content from a webpage with caching.
``max_bytes`` raises the download budget per call (clamped to the hard
cap); the default is the soft cap. When the body is cut short the result
carries ``truncated``/``fetched_bytes``/``total_bytes`` so callers can
tell the model the content is partial (#3812).
"""
effective_cap = min(max_bytes or WEB_FETCH_SOFT_MAX_BYTES, WEB_FETCH_HARD_MAX_BYTES)
# The cap is part of the cache identity: a truncated soft-cap fetch must
# not be served to a later full-budget request for the same URL.
cache_key = generate_cache_key(f"{url}#cap={effective_cap}")
cache_file = CONTENT_CACHE_DIR / f"{cache_key}.cache"
# Check cache
@@ -250,15 +372,21 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36",
"Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
"Accept-Language": "en-US,en;q=0.5",
"Accept-Encoding": "gzip, deflate",
# identity so the streamed size cap in _get_public_url stays honest
# (a compressed body can decode to far more than Content-Length).
"Accept-Encoding": "identity",
"Connection": "keep-alive",
}
response = _get_public_url(url, headers=headers, timeout=timeout)
response = _get_public_url(url, headers=headers, timeout=timeout,
max_bytes=effective_cap)
if response.status_code == 429:
raise RateLimitError(f"Rate limit hit for {url} (attempt {retry_attempt})")
response.raise_for_status()
except BodyTooLargeError as e:
error_logger.warning(f"Refused oversized body for {url}: {e}")
return _empty_result(url, f"TooLarge: {e}")
except httpx.HTTPStatusError as e:
error_logger.warning(f"HTTP {e.response.status_code} fetching {url}: {e}")
return _empty_result(url, f"HTTP {e.response.status_code}: {e}")
@@ -269,9 +397,27 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
error_logger.error(str(e))
return _empty_result(url, str(e))
# Size bookkeeping shared by every content branch below. getattr keeps
# plain httpx.Response stand-ins (tests) working without the cap fields.
_size_fields = {
"truncated": getattr(response, "truncated", False),
"fetched_bytes": len(response.content),
"total_bytes": getattr(response, "declared_bytes", None),
}
# PDF handling
content_type = response.headers.get("Content-Type", "").lower()
if "application/pdf" in content_type or url.lower().endswith(".pdf"):
if _size_fields["truncated"]:
# A PDF cut mid-stream is not parseable; unlike text there is no
# useful partial result, so report the budget problem instead.
_declared = _size_fields["total_bytes"]
return _empty_result(
url,
f"TooLarge: PDF exceeds the {effective_cap:,}-byte fetch budget"
+ (f" (size {_declared:,} bytes)" if _declared else "")
+ "; retry with a larger budget if it fits under the hard cap",
)
if pdf_extract_text is None:
logger.error("pdfminer.six is not installed; cannot extract PDF text.")
pdf_text = ""
@@ -295,6 +441,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
"js_message": "",
"success": bool(pdf_text),
"error": "" if pdf_text else "Failed to extract PDF text",
**_size_fields,
}
_cache_result(cache_file, cache_key, result, url)
return result
@@ -329,6 +476,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
"js_message": "",
"success": bool(text_body),
"error": "" if text_body else "Empty response body",
**_size_fields,
}
_cache_result(cache_file, cache_key, result, url)
return result
@@ -391,6 +539,7 @@ def fetch_webpage_content(url: str, timeout: int = 5, retry_attempt: int = 0) ->
"js_message": js_message,
"success": True,
"error": "",
**_size_fields,
}
_cache_result(cache_file, cache_key, result, url)
return result