Addresses CodeRabbit review on PR #1765 (pullrequestreview-5069337680).
pihole_monitor.py:
- last_raw is now tracked per source ({"primary": N, "secondary": M}
per device) instead of one combined value. Combining raw totals
across sources before diffing let a counter reset on one instance
silently net out against real traffic on the other - e.g. primary
+2000 (a real spike) and secondary resetting 1000->5 (-995) would
combine into a raw delta of only 1005, hiding most of the primary's
actual spike behind the secondary's unrelated restart.
- New aggregate_source_deltas(): diffs each source independently via
compute_delta(), then sums only the valid deltas. A source with no
valid delta this run (bootstrapping or just reset) contributes
nothing and doesn't block the others; each source keeps its own
reference point going forward.
- State loaded from before this change (last_raw as a plain number,
not per-source) is now tolerated instead of crashing - treated as no
prior reference point, so every source just bootstraps fresh on the
next run.
README.md:
- Fixed a self-contradicting line: a less frequent schedule means
larger per-run deltas, so PIHOLEMON_MIN_BLOCKED may need *raising*,
not lowering as it previously said.
- Corrected PIHOLEMON_HISTORY_DAYS guidance: it's a retention window,
not a detection delay. A new device becomes evaluable on its 3rd
successful run (1st anchors the counter, 2nd records the first
delta, 3rd has a baseline to compare against), not after the full
retention window.
Tests: 54 (up from 48). New coverage: aggregate_source_deltas() unit
tests including the exact dual-source reset-masking scenario, a
main()-level integration test for the same, and a regression test for
tolerating pre-per-source state. Both the reset-masking fix and the
legacy-state guard verified via mutation testing (reverted each,
confirmed the relevant tests fail, restored). 99% line+branch coverage
maintained.
References PR #1765.
Docs:
- Added PIHOLEMON to docs/PLUGINS.md and a new "Approach 4" section in
docs/PIHOLE_GUIDE.md, leading with anomaly detection (the actual
differentiator vs PIHOLEAPI) and explaining when to pick each plugin.
- README/PLUGINS.md/config.json's UI-facing description all reordered
and shortened to lead with anomaly detection instead of device
import, and to drop implementation detail that belongs in the
README, not the Settings page.
- Trimmed the "Why not extend PIHOLEAPI" README section per feedback -
useful context for a maintainer, not for an end user configuring
the plugin.
config.json / pihole_monitor.py:
- RUN defaults to "disabled", matching every other non-core plugin.
- VERIFY_SSL split into PRIMARY_VERIFY_SSL / SECONDARY_VERIFY_SSL -
each instance can be http/https independently. Settings reordered so
each *_VERIFY_SSL sits right under its matching *_PASSWORD.
- GRAPHQL_TOKEN removed; graphql_token now reads the core API_TOKEN
setting instead of a plugin-specific duplicate.
- GRAPHQL_URL replaced with a GET_OWNER boolean - the endpoint is now
derived from this app's own GRAPHQL_PORT (single source of truth)
instead of a URL the user had to keep in sync by hand.
- HISTORY_LENGTH (run count) replaced with HISTORY_DAYS (a real time
window): state now stores [timestamp, delta] samples and
trim_history() drops anything older than the window, so the
baseline means the same thing regardless of schedule - a faster
schedule adds more data points instead of shrinking the window.
- STATE_FILE moved from the log folder to dbFolderPath, so the rolling
anomaly baseline survives NetAlertX upgrades instead of being wiped
with the logs.
- netalertx_device_owner() (1 GraphQL call per device) replaced by
netalertx_device_owners() (1 call per run, batched) - avoids N
blocking round-trips on a large network.
- Fixed a zero-baseline bug: `bool(... and baseline and ...)` silently
exempted a device with an all-zero blocked-query history (0.0 is
falsy in Python) from ever being flagged, even on its first real
spike. Now checks `baseline is not None`.
- Fixed the placeholder-MAC filter: only excluded the literal "ip-::",
not Pi-hole's general "ip-<address>" placeholder pattern. Caught
downstream by is_mac() either way, but now the actual placeholder
check does what it looks like it does.
- Fixed a cumulative-counter bug: Pi-hole's /api/stats/top_clients
returns a count that's cumulative since FTL last started, not a
per-interval or daily-resetting one (confirmed against FTL's own
source and long-standing user reports that it doesn't reset at
midnight). Comparing that raw total directly against a rolling
average made any device's ordinary growing traffic look like an
escalating anomaly. compute_delta() now diffs each run's raw count
against the previous run's (state gained a per-key last_raw
reference point alongside the delta history) - None (not 0) on the
first-ever run for a device or right after a counter reset, so
those runs re-anchor the reference point instead of fabricating or
swallowing a delta.
- RUN_SCHD default changed from every 6 hours to every 5 minutes now
that the baseline window is real days, not run count, so a frequent
schedule only adds data points instead of narrowing the window; also
matches the default most other device-scanner plugins use.
- RUN_SCHD gained the same live cron-validity checkmark ARPSCAN and
other scanner plugins use (a ✓/✗ icon next to the field, validated
client-side against a regex) - reuses the existing generic
validateRegex() widget, nothing plugin-specific to build.
Tests: 48 tests (up from 37), 99% line+branch coverage. Every fix
above verified via mutation testing (deliberately broken, confirmed
the relevant test fails, then restored).
CodeRabbit follow-up on PR #1765
(https://github.com/netalertx/NetAlertX/pull/1765#discussion_r3888374014):
test_main_history_length_never_produces_empty_or_growing_unbounded only
asserted len(history) >= 1, which a mis-clamped history_length (e.g.
keeping 4 items instead of 1) would still pass unnoticed.
Replaced with test_main_history_length_clamps_and_trims_exactly,
seeding distinct ordered values and asserting the exact retained
history against each PIHOLEMON_HISTORY_LENGTH boundary. Verified it
actually catches a broken clamp: temporarily reverted the
max(1, ...) fix in pihole_monitor.py, confirmed this test fails
([] == [40]) while the rest of the suite still passes, then restored
the fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CHJAArRiet4GmXUsxnNLdW
Addresses 5 of the 6 actionable comments from CodeRabbit's review of
PR #1765 (netalertx/NetAlertX#1765), plus adds test coverage:
- fetch_top_blocked_clients() returns None on failure instead of {},
so a failed request can no longer be mistaken for "genuinely zero
blocked queries this run" and silently write a false 0 into a
device's rolling history baseline. main() now tracks a
stats_complete flag and skips anomaly evaluation + history
persistence entirely for a run with incomplete blocked-query data.
- fetch_top_blocked_clients() is now called with count=max_clients
(the existing PIHOLEMON_API_MAXCLIENTS setting) instead of a
hardcoded default of 50, so clients beyond the top 50 are no longer
silently dropped from anomaly detection.
- New build_ip_to_mac() derives the IP->MAC identity map from every
gathered device entry instead of from merge_device_entries()'s
by-MAC-deduplicated output, which only kept one IP per device and
silently lost a multi-IP device's other IPs (misattributing their
blocked-query traffic to a bare IP instead of the real MAC).
- PIHOLEMON_HISTORY_LENGTH is clamped to at least 1, so a negative
setting can no longer reach the history[-history_length:] slice
with a nonsensical negative-of-negative length.
- PIHOLEMON_VERIFY_SSL now defaults to true (was false, matching the
official PIHOLEAPI plugin's convention). README documents the
http:// vs https:// credentials trade-off explicitly rather than
forcing https:// - most home Pi-hole setups, including the one this
plugin targets, run over plain HTTP on a trusted LAN.
- Added test/plugins/test_pihole_monitor.py (37 tests, 99% line and
branch coverage of pihole_monitor.py per pytest-cov - only the
`if __name__ == '__main__':` entry-point guard is unreached):
auth and deauth success/failure paths, the None-sentinel-on-failure
contract, fetch_devices()'s own failure path, build_ip_to_mac()'s
multi-IP fix, gather_device_entries()'s skip branches and fake-MAC
fallback, netalertx_device_owner()'s success/failure/no-URL paths,
and main()-level coverage for source aggregation, the
stats_complete gate, the history_length boundary clamp, the
CONSIDER_ONLINE fallback, an unconfigured-sources run, and the
offline-device / invalid-MAC / unknown-IP / owner-lookup branches
together in one run.
Not addressed: CodeRabbit's suggestion to hard-reject http:// URLs in
auth(). Diverges deliberately - it would break the plugin's majority
use case (Pi-hole admin API on a trusted home LAN without TLS), which
this repo's own PIHOLEAPI plugin also targets over plain HTTP.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CHJAArRiet4GmXUsxnNLdW
Does two jobs against the same Pi-hole connection(s), instead of two
separately configured plugins:
1. Device import - same job as the official PIHOLEAPI (pihole_api_scan)
plugin, but supports an optional secondary/failover Pi-hole natively
(accepts two sets of credentials instead of forking the official
plugin, which hardcodes its settings-key prefix and doesn't support
multiple instances).
2. Query anomaly detection - flags a device whose blocked-query count
spikes well above its own recent rolling average (signature of
malware/a compromised device beaconing out), keyed by MAC address
(not IP, which changes under DHCP) and combined across both Pi-hole
instances so a compromised device can't evade detection by switching
resolvers.
Notifications are delegated entirely to NetAlertX's own Watched/Report
on mechanism - the plugin never calls a notification service directly.
Live-tested against a two-Pi-hole home setup (v26.8.5) for several days,
including two real bugs found and fixed during that testing (an
offline-filtered device losing its MAC and falling back to a bare-IP
identifier, and a boolean-expression flake8 style fix).
The whitespace note contradicted itself: it said surrounding whitespace is
trimmed and then that leading or trailing whitespace is invalid. The parser
strips, so trimming is what actually happens.
Also lists the new non-ASCII/newline skip reason and states that skipping is
per entry, and adds the language identifier markdownlint wants on the two
fenced blocks (MD040).
A custom header carrying a non-ASCII character raised UnicodeEncodeError from
inside http.client. That is a ValueError, not a RequestException, so it escaped
both handlers in send() and took the whole publisher down - every notification
lost because of one typo in one header.
build_custom_headers now rejects newlines and non-ASCII the same way it already
rejects malformed and colliding entries: warn, skip that entry, keep the rest.
UnicodeEncodeError is still caught at the request, as a backstop for the
plugin's own headers, since REPORT_DASHBOARD_URL feeds one of them.
Verified against real requests: before, send() raised UnicodeEncodeError; after,
the bad header is dropped and the notification is still posted.
Theme was applied on raw $(document).ready, which fires before the async cacheSettings() call has necessarily populated the settings cache. clearCache() (the header reload button) wipes localStorage immediately before reloading, so that reload could read an empty cache and hard-override the theme to Light via setCookie, even with Dark or System configured. Gated the theme application on callAfterAppInitialized(), which is already used elsewhere for this class of race.
Fixes#943
Jokob @NetAlertX