open-swe/agent/reviewer_diff.py
Johannes du Plessis 9a8b2d9984
feat: Inject PR title and body into reviewer context (#1368)
* Inject PR title and body into reviewer context

The reviewer agent previously received the PR url, number, and SHAs but not
the PR title or description, so it sometimes missed the original intent of
the PR. Fetch the title/body fresh from the GitHub API on every run (never
cached) so edits to the title/description are reflected on re-reviews, and
inject them into the first-review, re-review, and finding-reply contexts as
an untrusted-data block (author-controlled text, guarded against prompt
injection).

* Harden PR overview escaping against whitespace-padded closing tags
2026-06-01 21:06:19 +00:00

325 lines
11 KiB
Python

"""Diff utilities for the reviewer agent.
The reviewer needs three things from a PR diff:
1. The set of (file, line) tuples that are part of the diff, so ``add_finding``
can validate at creation time rather than at GitHub-publish time.
2. The hunk text relevant to a given (file, start_line, end_line) range, so we
can stash it on the Finding (``diff_hunk``) for rendering in the future UI
without re-fetching from GitHub or the (evictable) sandbox.
3. A way to compute the diff in the sandbox between two SHAs, used both on
first review (``base_sha..head_sha``) and on watched re-review
(``last_reviewed_sha..new_head_sha``).
"""
from __future__ import annotations
import asyncio
import logging
import re
from dataclasses import dataclass
from typing import TYPE_CHECKING, Literal
if TYPE_CHECKING:
from deepagents.backends.protocol import SandboxBackendProtocol
DiffSide = Literal["LEFT", "RIGHT"]
logger = logging.getLogger(__name__)
_DIFF_FILE_HEADER_RE = re.compile(r"^diff --git a/(?P<a>.+?) b/(?P<b>.+?)$")
_HUNK_HEADER_RE = re.compile(
r"^@@ -(?P<old_start>\d+)(?:,(?P<old_count>\d+))? "
r"\+(?P<new_start>\d+)(?:,(?P<new_count>\d+))? @@"
)
@dataclass(frozen=True)
class DiffHunk:
"""One hunk for one file in a unified diff.
``new_start``/``new_end`` are inclusive 1-based line numbers in the
post-PR (RIGHT side) file. ``body`` is the raw hunk text including the
``@@`` header — what gets stored on a Finding's ``diff_hunk``.
"""
file: str
new_start: int
new_end: int
old_start: int
old_end: int
body: str
@dataclass(frozen=True)
class FileDiff:
"""All hunks for one file in a unified diff."""
file: str
hunks: tuple[DiffHunk, ...]
def parse_unified_diff(diff_text: str) -> list[FileDiff]:
"""Parse a unified diff into per-file hunk records.
Skips ``--- ``/``+++ `` and binary file markers. Returns one ``FileDiff``
per file with at least one hunk; files with no hunks (e.g., pure renames)
are dropped.
"""
files: list[FileDiff] = []
lines = diff_text.splitlines()
i = 0
while i < len(lines):
header_match = _DIFF_FILE_HEADER_RE.match(lines[i])
if not header_match:
i += 1
continue
file_path = header_match.group("b")
i += 1
# Skip metadata lines until first hunk or next file header
hunks: list[DiffHunk] = []
current_hunk_lines: list[str] = []
current_meta: tuple[int, int, int, int] | None = None
while i < len(lines) and not _DIFF_FILE_HEADER_RE.match(lines[i]):
line = lines[i]
hunk_match = _HUNK_HEADER_RE.match(line)
if hunk_match:
if current_meta is not None and current_hunk_lines:
hunks.append(
DiffHunk(
file=file_path,
old_start=current_meta[0],
old_end=current_meta[1],
new_start=current_meta[2],
new_end=current_meta[3],
body="\n".join(current_hunk_lines),
)
)
old_start = int(hunk_match.group("old_start"))
old_count = int(hunk_match.group("old_count") or "1")
new_start = int(hunk_match.group("new_start"))
new_count = int(hunk_match.group("new_count") or "1")
# End line is inclusive; if count is 0 (deletion-only), end == start
old_end = old_start + max(old_count - 1, 0)
new_end = new_start + max(new_count - 1, 0)
current_meta = (old_start, old_end, new_start, new_end)
current_hunk_lines = [line]
elif current_meta is not None:
current_hunk_lines.append(line)
i += 1
if current_meta is not None and current_hunk_lines:
hunks.append(
DiffHunk(
file=file_path,
old_start=current_meta[0],
old_end=current_meta[1],
new_start=current_meta[2],
new_end=current_meta[3],
body="\n".join(current_hunk_lines),
)
)
if hunks:
files.append(FileDiff(file=file_path, hunks=tuple(hunks)))
return files
def compute_diff_line_set(diff_text: str) -> dict[str, dict[str, set[int]]]:
"""Return ``{file: {"RIGHT": {new_lines}, "LEFT": {old_lines}}}`` for the
lines covered by the diff.
Inline GitHub review comments anchor either to a new-side line (``side =
RIGHT``, the default — additions and context) or to an old-side line
(``side = LEFT`` — deletions). ``add_finding`` and ``publish_review``
validate a finding's ``(file, start_line..end_line, side)`` against the
matching set so deleted-line bugs aren't wrongly rejected.
"""
out: dict[str, dict[str, set[int]]] = {}
for file_diff in parse_unified_diff(diff_text):
sides = out.setdefault(file_diff.file, {"RIGHT": set(), "LEFT": set()})
for hunk in file_diff.hunks:
for line in range(hunk.new_start, hunk.new_end + 1):
sides["RIGHT"].add(line)
for line in range(hunk.old_start, hunk.old_end + 1):
sides["LEFT"].add(line)
return out
def extract_diff_hunk(
diff_text: str,
file: str,
start_line: int | None,
end_line: int | None,
) -> str | None:
"""Extract the hunk body covering ``file:start_line..end_line``.
Returns ``None`` if no hunk overlaps. For file-level findings (both lines
None) returns the first hunk in the file as best-effort context.
"""
file_diffs = [fd for fd in parse_unified_diff(diff_text) if fd.file == file]
if not file_diffs:
return None
hunks = file_diffs[0].hunks
if not hunks:
return None
if start_line is None or end_line is None:
return hunks[0].body
for hunk in hunks:
if hunk.new_start <= end_line and start_line <= hunk.new_end:
return hunk.body
return None
def is_range_in_diff(
line_set: dict[str, dict[str, set[int]]],
file: str,
start_line: int | None,
end_line: int | None,
side: DiffSide = "RIGHT",
) -> bool:
"""Return True if every line in ``start_line..end_line`` is on the given
side of the diff for ``file``.
File-level findings (both None) are always allowed. ``side`` selects
new-side lines (``RIGHT``, the default — additions/context) or old-side
lines (``LEFT`` — deletions). Pass the finding's recorded ``side``.
"""
if start_line is None and end_line is None:
return True
if start_line is None or end_line is None:
return False
file_sides = line_set.get(file)
if not file_sides:
return False
side_lines = file_sides.get(side)
if not side_lines:
return False
return all(line in side_lines for line in range(start_line, end_line + 1))
async def fetch_pr_diff(
*,
owner: str,
repo: str,
pr_number: int,
token: str,
timeout: float = 30.0,
) -> str | None:
"""Fetch the PR's unified diff (base..head) from the GitHub REST API.
Returns ``None`` if the request fails. This is the same diff GitHub
validates against when posting inline review comments, so it's the right
source for ``add_finding``'s in-diff anchor validation and for
``publish_review``'s 422 retry filter.
"""
import httpx
headers = {
"Accept": "application/vnd.github.diff",
"Authorization": f"Bearer {token}",
"X-GitHub-Api-Version": "2022-11-28",
}
url = f"https://api.github.com/repos/{owner}/{repo}/pulls/{pr_number}"
try:
async with httpx.AsyncClient() as client:
response = await client.get(url, headers=headers, timeout=timeout)
response.raise_for_status()
except httpx.HTTPError:
logger.exception("Failed to fetch PR diff for %s/%s#%s", owner, repo, pr_number)
return None
return response.text
async def fetch_pr_metadata(
*,
owner: str,
repo: str,
pr_number: int,
token: str,
timeout: float = 30.0,
) -> tuple[str, str] | None:
"""Fetch the PR's title and body from the GitHub REST API.
Returns ``(title, body)`` or ``None`` if the request fails. Always
fetched fresh per run (never cached) so an edited title/description is
reflected on every re-review. ``body`` is normalized to ``""`` when the
PR has no description.
"""
import httpx
headers = {
"Accept": "application/vnd.github+json",
"Authorization": f"Bearer {token}",
"X-GitHub-Api-Version": "2022-11-28",
}
url = f"https://api.github.com/repos/{owner}/{repo}/pulls/{pr_number}"
try:
async with httpx.AsyncClient() as client:
response = await client.get(url, headers=headers, timeout=timeout)
response.raise_for_status()
payload = response.json()
except httpx.HTTPError:
logger.exception("Failed to fetch PR metadata for %s/%s#%s", owner, repo, pr_number)
return None
except ValueError:
logger.exception("Failed to parse PR metadata for %s/%s#%s", owner, repo, pr_number)
return None
title = payload.get("title")
body = payload.get("body")
return (title if isinstance(title, str) else "", body if isinstance(body, str) else "")
async def compute_diff_in_sandbox(
sandbox_backend: SandboxBackendProtocol,
work_dir: str,
base_ref: str,
head_ref: str,
*,
merge_base: bool = False,
) -> str:
"""Run ``git diff`` inside the sandbox and return its stdout.
Refs can be SHAs or branch names. Caller is responsible for ensuring both
refs exist locally (e.g., having fetched the PR head).
Args:
merge_base: When ``True``, use three-dot ``base...head`` (the merge-base
diff — what GitHub shows on the PR's "Files changed" tab). Use this
for first review so we don't pick up changes that landed on the
base branch after the PR diverged. When ``False``, use two-dot
``base..head`` — appropriate for re-review deltas where ``base`` is
the previously reviewed SHA and we want exactly the commits added
since.
"""
operator = "..." if merge_base else ".."
cmd = f"cd {work_dir} && git diff --no-color {base_ref}{operator}{head_ref}"
result = await asyncio.to_thread(sandbox_backend.execute, cmd)
exit_code = getattr(result, "exit_code", None)
if exit_code not in (0, None):
output = _stdout_from_result(result)
raise RuntimeError(
f"git diff failed (exit {exit_code}) for "
f"{base_ref}{operator}{head_ref} in {work_dir}. Output:\n{output}"
)
return _stdout_from_result(result)
def _stdout_from_result(result: object) -> str:
"""Best-effort extraction of stdout from a sandbox execute() result.
Different sandbox providers return different shapes; this normalizes them.
"""
if isinstance(result, str):
return result
if isinstance(result, dict):
for key in ("stdout", "output", "text"):
value = result.get(key)
if isinstance(value, str):
return value
stdout = getattr(result, "stdout", None)
if isinstance(stdout, str):
return stdout
text = getattr(result, "text", None)
if isinstance(text, str):
return text
return ""