Raft Ward

Raft Ward checks the integrity of closed SQLite databases, frozen exports, and backup folders. It detects missing files, corruption, suspicious shrinkage, permission changes, and patterns consistent with bulk encryption or renaming. It does not repair, restore, quarantine, or modify the files it watches.

Version 0.1.1 is free software under Apache-2.0, for Linux with Python 3.10 or newer. Its only runtime dependency is tomli on Python 3.10.

Posture

• Local, read-only target access; no telemetry. No network or child processes by default. Configured alert delivery is the only network feature.
• Every watched target has configured owning applications. A running app or background worker skips its own target, with its name and reason in the report. Other targets continue. No metadata, hashing, or integrity work starts for a target whose owner is already observed running.
• Unknown process visibility or any configured nonlocal machine also defers the entire run. A stale local status file never counts as remote closure.
• Owner matching is generic and case insensitive across Linux /proc process names, executable paths, and command lines. Cursor and Codex are defaults, not special cases. There is no owning-process allowlist.
• Targets are opened only with read-only descriptors, no locks. SQLite is never connected to an original. Source symlinks, unsupported files, foreign-owned paths, a visible same-UID foreign target/sidecar descriptor, transfer locks, and running configured transfer programs defer the run.
• Writes are restricted to private application state, an explicit new report file, and disposable private SQLite copies under $XDG_RUNTIME_DIR/raftward. Report/state/runtime locations cannot overlap configured target trees.
• Persistent state contains local paths, keyed identifiers, file metadata, SHA-256 file digests, entropy measurements, findings, and timestamps. It contains no database rows, message text, file excerpts, full process arguments, or conversation data. SQLite copies contain source bytes briefly and are deleted in cleanup on success, deferral, interruption, or exception. On the next run, stale private copies are removed from the runtime directory.

Keep all involved applications and workers closed for the full run. A polling guard cannot establish an atomic interlock with an unrelated application. Raft Ward reaches one checkpoint after each 64 KiB source chunk, throughout directory walks, and every 100 SQLite virtual-machine instructions. Repeated process enumeration is throttled to a 100 ms interval measured from the end of the previous enumeration at these checkpoints. Mandatory checks bypass that interval before target metadata/open, initial source copying, private-copy creation, SQLite connection/integrity work, and completion. A slow process enumeration, bounded retry, or blocked filesystem call can extend the wall-clock delay; this is not a real-time deadline. Once a running owner is observed, that target's result is discarded while other targets continue; cleanup closes descriptors and removes its private copy. An app can start between a check and a syscall or entirely between checks. No zero-race protection is claimed.

Install and configure

From a reviewed checkout:

python3 -m pip install .
raftward --help

Copy and edit examples/config.toml into $XDG_CONFIG_HOME/raftward/config.toml (default ~/.config/raftward/config.toml). Create that configuration with mode 0600. There are no automatically scanned targets. Every target has a unique simple name, path, class, and nonempty owners list. If owners is omitted it defaults to all configured owners. The default owners are cursor and codex.

machines = ["local"]
alert_channels = ["stdout"]

[owners]
notes = ["notes-app", "notes-worker"]

[[targets]]
name = "notes-export"
path = "~/backups/notes.sqlite"
class = "frozen"
sqlite = true
owners = ["notes"]

live tolerates ordinary hash and mtime changes. frozen flags changes in hash, size, or mtime. folder recursively checks mixed files and evaluates bulk changes since the last completed run. sqlite = true identifies an expected database even if its header is already damaged; it defaults true for live. Recognizable SQLite files found in any class also get integrity checks. Prefer explicit sqlite = true for frozen database exports.

File and directory globs support *, ?, and bracket patterns. Recursive ** globs defer; use a folder target for recursive scanning. Symlinks are never followed. Defaults are 10,000 files and a 20-file mass-change threshold. A folder containing other users' files defers rather than silently skipping them. Broad owner substrings may yield conservative false positives.

Multi-machine operation: list every involved machine in machines, including local. v0.1 cannot prove current remote app state and therefore always defers if this list includes another machine. There is no SSH, remote helper, status-file override, or network probe. Do not omit an involved machine to get around this gate. Coordinated cross-machine inspection remains a review/release gap.

Commands

raftward guard
raftward baseline
raftward check --quick --max-gb 2
raftward check --target notes-export --format html --output /tmp/raft-report.html
raftward accept notes-export --reason "verified closed export"
raftward baseline export --format json
raftward status
raftward report --since 7d
raftward report --since 24h --send
raftward targets list
raftward targets add packets ~/backups/packets --class folder --owner notes
raftward targets remove packets
raftward print-timer --interval 30m

All commands accept --format text|json|md|html and --output NEW_FILE. Global --config FILE and --state DIR options precede the command. Report destinations must already have a parent directory; existing files are never overwritten. Text output is a readable metadata record; JSON is structured data. Markdown and HTML escape all interpolated values and load no external assets.

guard inspects process state and configured transfer-lock metadata only; it does not stat or open watched targets. status, report, targets list, and baseline export read stored application metadata only and do not scan sources. A guard result is valid at that checkpoint, not a promise about a later check.

baseline captures all targets or --target NAME; replacing an existing baseline requires --force. accept TARGET re-baselines one reviewed change. Neither will bless missing targets, a lost expected SQLite header, failed integrity, or future mtime. An accept records time, target, and an HMAC fingerprint of the optional reason; it intentionally does not retain free text, which might contain credentials or personal conversation. targets add/remove writes a private complete target list in state that takes precedence over the TOML target list; owner definitions still come from TOML.

Checks use nice(10) when allowed. --max-gb limits source bytes for the entire run; exceeding it defers rather than recording a partial success. The runtime directory must be available in XDG_RUNTIME_DIR. --quick requests PRAGMA quick_check; full integrity_check is the default.

Exit codes: 0 no findings/success, 1 findings or rejected unsafe baseline, 2 usage/runtime/alert-delivery error, 3 deferred or partial check/guard/status. A report returns 1 when the requested history contains findings. Timer unit text is printed as report data only. Review the ExecStart executable path before installing units yourself; Raft Ward never installs, enables, or starts timers.

SQLite and transfer boundaries

For a closed source without sidecars, Raft Ward copies bytes through bounded read-only reads into a unique 0700 directory. The copy is created 0600 before any bytes are written. Source device/inode, size, mtime, ctime, and mode are checked again. The copy is opened with mode=ro&immutable=1, query_only=ON, trusted_schema=OFF, and temp_store=MEMORY. Only ok/not ok and page count are retained from SQLite; diagnostic strings may reveal content and are discarded.

Any -wal, -shm, or -journal beside a SQLite target defers the entire run. Raft Ward never assumes immutable mode validates uncheckpointed WAL contents. It does not checkpoint or remove sidecars. A closed WAL-mode main database without sidecars can be checked as a main-file snapshot. WAL reconstruction and live WAL integrity validation are explicitly unsupported.

Configured transfer_commands default to rclone and rsync. Any matching process stops the entire run even when its arguments do not identify a target. This includes permanent rclone mount processes and rsync daemons: they can defer every run indefinitely. The 24-hour deferral notice names this conservative transfer policy. Subcommand/target attribution has not been proven safe and no mount or daemon exemption is applied. transfer_dirs stops work if a process command references a configured directory; transfer_locks and singular transfer_lock are supported. No process is stopped or signaled.

Raft Ward requires a full host /proc view. A restricted hidepid mount defers; a PID namespace containing only a subset of the host is not proof that other owners are closed. Owner/worker and transfer matching uses comm plus cmdline for every visible process, with exe as supplementary identity when readable. Only Raft Ward's own PID is excluded; its parent, sibling processes, and workers are still checked. Same-UID identity/descriptor permission failures get three 50 ms rereads. A persistently unavailable exe may be omitted when identity remains readable. A same-UID process whose /proc/PID/fd is kernel-owned, such as a non-dumpable user service, participates in name matching while its open-file check is reported as unverifiable. The same applies when fd entries remain listable but their readlinks return EACCES/EPERM after retries and the process's CapPrm or CapEff contains capabilities absent from our own. Readable fd links still participate in target-handle checks. Other unreadable same-UID identity or descriptors still defer. process-descriptor-state-unknown distinguishes descriptor visibility failure.

Other-UID exe and descriptor access is normally restricted by Linux. Raft Ward matches those processes using readable comm/cmdline; at least one identity field must be readable. It does not inspect other-UID descriptors and cannot prove that another user (including root) has no target file open. This is an explicit visibility gap, not a closed-state claim. Transient ENOENT/ESRCH means a vanished process/descriptor; readable owner names are retained when only exe disappears. PID reuse/UID changes trigger a reread. No owner allowlist or privilege request is used to make an unknown state pass.

Findings

Renames that retain an inode retain the entropy comparison. A replacement written to a new inode/path cannot be reliably associated with the old file; deletion, extension, and mass-change findings still apply. These are evidence signals, not a malware diagnosis. Entropy is estimated from a sample, not the whole file.

State, privacy, and alerts

State is under $XDG_STATE_HOME/raftward (default ~/.local/state/raftward), mode 0700. Files are created 0600 with os.open; no post-write chmod is used. Unsafe existing state permissions or symlinked parents are refused, not repaired.

• fp.key: 32 random bytes used for HMAC-SHA256 identifiers (fp: plus 16 hex characters). A missing key beside an existing baseline is an error.
• baseline.json, previous.json, last.json: metadata and findings.
• history.jsonl: at most 30 days and 4 MiB of recent metadata (older rows trimmed).
• accepts.jsonl: review records with reason fingerprints, at most 4 MiB.
• targets.json: optional CLI-managed targets; deferred.json: deferral episode; alerts.json: hashed delivery receipts with a 24-hour dedupe period.

Paths appear only locally. Recognized token-shaped substrings in local metadata are HMAC-redacted. Because arbitrary filenames and operator-supplied configuration are not universally recognizable as secrets, keep credentials out of target names and paths. File contents are never rendered or scanned for display.

Critical findings alert during check. Other findings are available in reports; report --send explicitly delivers their aggregate summary. alert_channels accepts stdout, slack, ntfy, or email; an empty list means stdout. Network delivery is opt-in by configuring a non-stdout channel. Payloads contain rule IDs, counts, and severity only, and end with "run raftward report locally for details". They contain no hostname, paths, process arguments, or file content. Delivery is deduped for 24 hours and uses a five-second timeout. Redirects are not followed.

Use environment-variable names, never inline credentials: alert_slack_webhook_env, alert_ntfy_token_env, alert_smtp_password_env. Other channel keys are alert_ntfy_topic, alert_smtp_host, alert_smtp_port, alert_smtp_security (starttls, ssl, or loopback-only none), alert_smtp_username, alert_smtp_from, and alert_email. HTTPS is required except on loopback; SMTP encryption is required off loopback. Errors never echo remote credential-bearing URLs.

Development and review

python3 -m pytest -q

Tests use a temporary HOME, fixture databases, and injected process listings. They do not inspect real application data. No timer is installed. See REVIEW.md for the independent acceptance report and remaining limits.