hermes-agent/hermes_cli/sqlite_safe_read.py
teknium1 fbd5e5772b fix(state): stop cancelling our own POSIX locks on live SQLite databases
`hermes sessions optimize` could corrupt state.db. Root cause is Hermes,
not the SQLite WAL-reset bug (#69784).

close() on ANY file descriptor for a SQLite database cancels every POSIX
advisory lock the process holds on that file, including a running VACUUM's
EXCLUSIVE lock (sqlite.org/howtocorrupt.html section 2.2). Hermes byte-probed
live databases in several hot paths: the zeroed-state.db detector runs on every
SessionDB construction (and the gateway builds those constantly), and kanban's
post-commit invariant check ran after every COMMIT. While VACUUM rewrote the
file, those probes dropped its lock and let other processes write into it.

A/B against the real code, only variable being the raw read:

  SQLite 3.50.4, VACUUM + concurrent writers, DELETE mode
    raw open/close during VACUUM   8 vacuums, 319 vacuum errors, 2/2 corrupt
    no raw read (control)        229 vacuums,   0 vacuum errors, 0/2 corrupt

  SQLite 3.53.1 (WAL-reset FIXED) reproduces identically: 2/2 corrupt.
  After this change: 0/4 corrupt, 0 vacuum errors.

Because the upgraded runtime corrupts too, replacing the embedded SQLite does
not fix this class; and because DELETE is where it reproduces, #70055's
"force DELETE on vulnerable builds" mitigation steered users into the failing
mode. That gate is reverted here: vulnerable builds get WAL again and still
warn so operators can upgrade.

- add hermes_cli/sqlite_safe_read.py: read page_count via PRAGMA over the
  existing connection instead of open()+seek(28); byte-level probes are
  restricted to before any connection exists and refused once one is live,
  with an explicit force= escape for offline artifacts (snapshots, archives)
- track live connections in SessionDB and kanban's connect so that guard is
  enforced rather than merely documented
- kanban's torn-extend check now only applies under a rollback journal; in WAL
  a committed page may still legitimately sit in the -wal file
- revert the force-DELETE WAL gate and update the tests that pinned it

Regression tests assert the behavioural contract (an external process stays
locked out across Hermes' inspection calls) and were verified to fail when the
old raw-open behaviour is restored.
2026-07-25 21:44:43 -07:00

177 lines
6.3 KiB
Python

"""Lock-safe inspection of SQLite database files.
Why this module exists
----------------------
POSIX advisory locks are cancelled **process-wide** by ``close()`` on *any*
file descriptor for that file::
the close() system call will cancel all POSIX advisory locks on the
same file for all threads and all file descriptors in the process
-- https://sqlite.org/howtocorrupt.html#_posix_advisory_locks_canceled_by_a_separate_thread_doing_close_
So a bare ``open(db_path, "rb") ... close()`` on a **live** database silently
drops every lock SQLite holds on it from this process -- including the
EXCLUSIVE lock a ``VACUUM`` is holding while it rewrites the whole file, and
the RESERVED lock an in-flight ``BEGIN IMMEDIATE`` is holding. Other processes
are then free to write into a file that a writer still believes it owns, which
is the documented route to "database disk image is malformed".
Hermes is exactly the topology this hits: gateway, dispatcher, dashboard,
TUI, CLI, cron and kanban workers all open the same ``state.db`` /
``kanban.db``, and several code paths used to byte-probe those files while
connections were live.
The rules
---------
1. **Never** ``open()`` a database file that may have live connections in this
process. Ask SQLite instead -- :func:`page_count_bytes` reads the same
header field via ``PRAGMA``, over the existing connection, taking no new
descriptor.
2. Byte-level probes are only safe **before any connection exists** for that
path (first-open validation). Route those through
:func:`read_header_bytes_preopen`, which refuses once a connection has been
registered for the path.
:func:`track_connection` / :func:`untrack_connection` maintain the registry of
paths with live connections, so rule 2 can be enforced rather than merely
documented.
"""
from __future__ import annotations
import logging
import os
import sqlite3
import threading
from pathlib import Path
from typing import Optional
logger = logging.getLogger(__name__)
SQLITE_HEADER_MAGIC = b"SQLite format 3\x00"
# Offset of the 4-byte big-endian page-count field in the SQLite header.
_HEADER_PAGE_COUNT_OFFSET = 28
_live_lock = threading.RLock()
# resolved path -> number of live connections opened by this process
_live_connections: dict[str, int] = {}
def _key(path: Path | str) -> str:
try:
return str(Path(path).resolve())
except OSError:
return str(path)
def track_connection(path: Path | str) -> None:
"""Record that this process now holds a connection to *path*."""
key = _key(path)
with _live_lock:
_live_connections[key] = _live_connections.get(key, 0) + 1
def untrack_connection(path: Path | str) -> None:
"""Record that one connection to *path* has been closed."""
key = _key(path)
with _live_lock:
remaining = _live_connections.get(key, 0) - 1
if remaining > 0:
_live_connections[key] = remaining
else:
_live_connections.pop(key, None)
def has_live_connection(path: Path | str) -> bool:
"""Whether this process currently holds any connection to *path*."""
with _live_lock:
return _key(path) in _live_connections
def page_count_bytes(conn: sqlite3.Connection) -> Optional[int]:
"""Logical database size in bytes, read through *conn*.
``page_count * page_size`` is the same quantity the 4-byte header field at
offset 28 carries, but reading it via ``PRAGMA`` opens no new file
descriptor and therefore cannot cancel this process's POSIX locks.
Returns ``None`` when the pragmas cannot be read.
"""
try:
page_count = conn.execute("PRAGMA page_count").fetchone()[0]
page_size = conn.execute("PRAGMA page_size").fetchone()[0]
except (sqlite3.Error, TypeError, IndexError) as exc:
logger.debug("page_count/page_size unavailable: %s", exc)
return None
try:
return int(page_count) * int(page_size)
except (TypeError, ValueError):
return None
def file_length_matches_header(conn: sqlite3.Connection) -> Optional[bool]:
"""Whether the file on disk is at least as long as the header claims.
Detects the "torn extend" shape (file shorter than its own page count)
without ever opening the database file: the header side comes from
``PRAGMA page_count`` over *conn*, and the on-disk side from ``stat()``,
which takes no descriptor and cannot break locks.
Returns ``None`` when the check is not applicable (in-memory database,
unreadable pragmas, or a stat failure).
Note: in WAL mode a freshly committed page may still live in the ``-wal``
file, so the main file legitimately lags. Callers must treat this as
advisory unless the database is in a rollback journal mode.
"""
try:
row = conn.execute("PRAGMA database_list").fetchone()
except sqlite3.Error:
return None
if row is None:
return None
path_str = row[2] if len(row) > 2 else ""
if not path_str:
return None # in-memory or unnamed database
logical = page_count_bytes(conn)
if not logical:
return None
try:
actual = os.path.getsize(path_str)
except OSError:
return None
return actual >= logical
def read_header_bytes_preopen(
path: Path | str,
*,
length: int = 100,
force: bool = False,
) -> Optional[bytes]:
"""Read the first *length* bytes of *path* -- only when no connection is live.
This is the ONLY sanctioned byte-level read of a database file, and it is
restricted to first-open validation (is this file a real SQLite database,
is it zeroed, has it been overwritten by something else). Once any
connection to *path* exists in this process, the read is refused and
``None`` is returned, because the ``close()`` would cancel that
connection's POSIX locks.
Set ``force=True`` only for genuinely offline files (quarantined copies,
snapshot artifacts, archives) that no live connection can reference.
"""
if not force and has_live_connection(path):
logger.debug(
"refusing byte-level read of %s: a live connection exists in this "
"process and close() would cancel its POSIX locks",
path,
)
return None
try:
with open(path, "rb") as handle:
return handle.read(length)
except OSError:
return None