From 7cbba7bd0a7cfd21e0b3ea824dfd6d3d0846cb9d Mon Sep 17 00:00:00 2001
From: Joao Moura <joaomdmoura@gmail.com>
Date: Sat, 8 Aug 2026 21:24:04 -0700
Subject: [PATCH] fix(agents): close two heredoc bypasses in the guard

A body piped into a shell is executable, not data, so it is no longer
stripped: bash <<'EOF' containing a blocked command is matched again.

The opener's own line is now preserved, so a redirect written after the
delimiter is still seen. Only the body between delimiters is replaced.

Both found by review on the previous commit.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ASfWmW3RGy4qAQm6s8U9jH
---
 .claude/hooks/guard.py      | 43 ++++++++++++++++++++++++++-----------
 .claude/hooks/test_guard.py | 25 +++++++++++++++++++++
 2 files changed, 55 insertions(+), 13 deletions(-)

diff --git a/.claude/hooks/guard.py b/.claude/hooks/guard.py
index d2764071b..44caaca17 100644
--- a/.claude/hooks/guard.py
+++ b/.claude/hooks/guard.py
@@ -94,21 +94,38 @@ def deny(reason: str) -> None:
     )
 
 
-def strip_heredocs(command: str) -> str:
-    """Replace heredoc bodies with a placeholder.
+#: A heredoc opener, its same-line remainder, and the body up to the delimiter.
+HEREDOC = re.compile(
+    r"(<<-?\s*[\"']?(\w+)[\"']?)([^\n]*)\n.*?^[ \t]*\2\b",
+    re.DOTALL | re.MULTILINE,
+)
 
-    A heredoc body is data — a commit message, a file being written, a PR body —
-    not a command. Matching rules against it produces false denials, such as
-    blocking a commit whose message merely discusses `rm -rf docs/images`. The
-    redirection and the delimiter stay on the command line, so a heredoc that
-    genuinely writes into a protected path is still caught.
+#: Commands that execute their heredoc body instead of consuming it as data.
+SHELL_INTERPRETER = re.compile(r"\b(?:bash|sh|zsh|dash|ksh|eval)\b")
+
+
+def strip_heredocs(command: str) -> str:
+    """Replace data heredoc bodies with a placeholder.
+
+    A heredoc body is usually data — a commit message, a file being written, a
+    PR body — not a command, and matching rules against it produces false
+    denials, such as blocking a commit whose message merely discusses a
+    protected path.
+
+    Two things keep that from becoming a bypass. A body piped into a shell
+    (`bash <<'EOF'`) is executable, so it is left intact and still matched. And
+    the opener's own line is preserved in full, so a redirection written after
+    the delimiter (`cat <<'EOF' > docs/v1/x.mdx`) is still seen.
     """
-    return re.sub(
-        r"(<<-?\s*[\"']?(\w+)[\"']?).*?^\s*\2\b",
-        r"\1 HEREDOC_BODY",
-        command,
-        flags=re.DOTALL | re.MULTILINE,
-    )
+
+    def replace(match: re.Match[str]) -> str:
+        line_start = command.rfind("\n", 0, match.start()) + 1
+        opener_line = command[line_start : match.start()] + match.group(1)
+        if SHELL_INTERPRETER.search(opener_line):
+            return match.group(0)
+        return f"{match.group(1)}{match.group(3)}\nHEREDOC_BODY\n{match.group(2)}"
+
+    return HEREDOC.sub(replace, command)
 
 
 def bash_violation(command: str) -> str | None:
diff --git a/.claude/hooks/test_guard.py b/.claude/hooks/test_guard.py
index 52937a20b..4963c4ddc 100644
--- a/.claude/hooks/test_guard.py
+++ b/.claude/hooks/test_guard.py
@@ -136,6 +136,31 @@ def test_a_heredoc_writing_into_a_frozen_snapshot_is_still_blocked() -> None:
     assert "frozen release snapshots" in reason
 
 
+def test_a_redirect_after_the_heredoc_delimiter_is_still_seen() -> None:
+    """The opener's own line must survive stripping, redirect included."""
+    command = "cat <<'EOF' > docs/v1.15.0/index.mdx\nsome new content\nEOF\n"
+    reason = guard.bash_violation(command)
+    assert reason is not None
+    assert "frozen release snapshots" in reason
+
+
+@pytest.mark.parametrize(
+    ("command", "expected"),
+    [
+        ("bash <<'EOF'\nrm -rf docs/images/a.png\nEOF\n", "docs/images/"),
+        ("sh <<'EOF'\npip install ruff\nEOF\n", "do not use pip directly"),
+        ("bash -s <<'EOF'\ngit commit --no-verify -m x\nEOF\n", "--no-verify"),
+    ],
+)
+def test_a_heredoc_piped_into_a_shell_is_executable_not_data(
+    command: str, expected: str
+) -> None:
+    """A body a shell will run must still be matched, not stripped as data."""
+    reason = guard.bash_violation(command)
+    assert reason is not None, f"expected {command!r} to be blocked"
+    assert expected in reason
+
+
 def test_a_command_after_a_heredoc_is_still_inspected() -> None:
     command = "cat <<'EOF' > notes.txt\njust notes\nEOF\npip install ruff\n"
     reason = guard.bash_violation(command)