Caught live on stg: every wardriving session showed airtimeMs=0, even
for senders with confirmed multi-hop relay paths. The in-memory store
is memory-bounded (maxMemoryMB/maxPackets) and had already evicted
these transmissions despite them still being well within the SQL
24h/7d window — AirtimeForTransmissions treated "not found in memory"
the same as "found, zero relays", so ok=true with a silent 0 came back
indistinguishable from a genuinely never-relayed message.
Now ok=false when NONE of the requested IDs are held in memory
(nothing knowable → omit the field). A partial match still returns
ok=true with the known subset's total, matching how every other
airtime metric already tolerates eviction rather than going dark
entirely once retention exceeds the in-memory window.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds AirtimeMs to WardrivingSession: total network airtime consumed
relaying a session's messages, using the exact same formula as the
Overview tab's "Relay Airtime Share" (issue #1768) — ToA(payload_bytes)
x COUNT(DISTINCT resolved repeater in path), summed per message.
That formula needs the in-memory store's resolved-path index
(resolvedPubkeyReverse), which db.go's SQL-only wardriving code
doesn't have — so buildWardrivingSessions now also tracks each
session's transmission IDs (unexported, json:"-"), and the route
handler calls the new PacketStore.AirtimeForTransmissions(txIDs) to
fill in AirtimeMs after the fact. Omitted entirely (nil) in DB-only
mode rather than shown as a misleading zero.
Frontend adds an Airtime column to the Sessions table (ms/s
formatting, dash when unavailable).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found in production: at least one wardriving client appends plaintext
"<lat>,<lon>" after the standard token (e.g.
"MM:c3e_zJ1rUA:55.59743,13.00128"), confirmed against live traffic.
The generic anomaly framing (flagging non-standard payload length,
showing a raw hex dump) was based on the wrong assumption that this
was an undocumented binary format — it's actually deliberate,
human-readable coordinate sharing by that sender's client.
Replaces WardrivingAnomaly/detectWardrivingAnomalies/
StandardPayloadCount with WardrivingGPSShare/detectWardrivingGPSShares:
detects the specific "<token>:<lat>,<lon>" suffix (with range
validation) and reports the most recent position per sender. The
per-message drill-down gets Lat/Lon instead of a standard/anomaly
flag. Frontend replaces the "Payload Anomalies" section with "GPS
Sharing", linking each position to the live map
(#/map?lat=..&lon=..&zoom=15, an existing deep-link format).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New GET /api/analytics/wardriving/sender-messages returns one sender's
individual messages (most-recent-first, capped at 200): resolved
entry-point path (path[0] first), per-observer SNR/RSSI, and payload
standard/anomaly classification. Pass since+until to scope to one
session's exact range; otherwise it covers the sender's whole window.
Frontend makes sender names in Top Senders and Sessions clickable —
clicking toggles an inline expansion row with that sender's messages,
resolving every distinct path prefix in one /resolve-hops call (same
unique_prefix-only discipline as the aggregate Entry Points table).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
decodeGrpTxt (cmd/ingestor/decoder.go) builds decoded_json.text as
"<sender>: <message>", so a wardriving message's real text is
"<sender>: MM:<base64>", not a bare "MM:<base64>" at the start of the
string. detectWardrivingAnomalies's HasPrefix(text, "MM:") check never
matched anything live, silently returning standardPayloadCount=0 and
no anomalies — caught after deploying idea 6 to stg and seeing 0/0
against traffic that should have had ~160 standard-format messages.
Fixed to match the exact "<sender>: MM:" prefix. The Go test fixtures
shared the same wrong assumption and are fixed too.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Extends GET /api/analytics/wardriving with a payload-anomaly detector:
every "MM:<base64>" message is checked against the standard 7-byte
anonymous session token (confirmed empirically against live traffic).
Non-standard lengths or undecodable payloads are grouped per sender
as a candidate signal that MeshMapper's optional "Broadcast My
Coordinates" mode is active — surfaced as a raw hex dump, never
interpreted as lat/lon, since that mode's byte format is undocumented.
Messages without the "MM:" prefix (plain channel chat) are ignored
entirely rather than polluting either bucket.
Frontend adds a Payload Anomalies table plus a stat card, completing
all 6 originally-proposed wardriving analytics ideas.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Extends GET /api/analytics/wardriving with each sender's messages
grouped into distinct sessions/runs — a gap over 15 minutes starts a
new one. Each session reports duration, message count, and how many
distinct entry-point repeaters/observers it touched. Frontend adds a
Sessions table (most-recent-first) and a session-count stat card.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Extends GET /api/analytics/wardriving with avgSnr/avgRssi over the
same time buckets as the activity series, plus overall averages.
Frontend adds two min/max-scaled line charts (SNR has no natural
zero floor, RSSI is negative dBm, so these can't reuse the 0-baseline
message-volume chart) and two stat cards.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Surfaced by the new Wardriving Coverage-by-Observer table: every
observer on this Danish mesh showed a "—" location because iataCoords
only had US/global codes. Same root cause as issue #1786, just never
extended to the codes this specific mesh's observers actually use.
Added the 11 codes I could verify as real airport IATA codes; QXV,
KRP, MRW look non-standard and are left unresolved.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
New GET /api/analytics/wardriving endpoint plus an Analytics tab
covering the three requested angles: activity over time + top
senders, entry-point repeaters (path[0] tally, unique_prefix-only
name resolution), and per-observer coverage using observers' known
IATA coordinates. MeshMapper's on-air ping is an anonymous session
token by default, not the sender's live GPS, so sender position
itself isn't tracked — documented in the tab and OpenAPI description.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Scoped/Unscoped/Unknown counts answered "how much" but not "which
regions" — a channel spanning multiple hashRegions areas looked
identical to one using just one. New "Regions Used" column lists the
distinct scope_name values seen on each channel's scoped messages,
most-used first.
Backend: GetChannelScopeAdoption now also queries per-(channel,
scope_name) counts and attaches them as ChannelScopeAdoption.Regions.
Non-fatal on query failure — adoption counts still render without it.
[dijkstra] doc-noted that `total` in handleNodes' filtered path reflects
only the returned page-slice, not an all-pages count — matches
pre-existing semantics, not a regression, but worth being explicit that
fetchAllNodes (public/app.js) never reads it for pagination (relies on
page length instead).
[munger] the pagination compensation loop's maxIterations=50 safety cap
now logs a WARN breadcrumb when it's actually hit, instead of silently
returning a possibly-short page. New test seeds 51 consecutive
hidden-prefix nodes at limit=1 to force the cap and asserts both the
empty result and the log line.
MAJOR (carmack): handleConfigClient and nodeListPostFilters both read
s.cfg.GeoFilter unlocked, but the field is cfgMu-guarded — PUT
/api/config/geo-filter swaps it under Lock(). Both sites now go through
s.getGeoFilter() like the other existing read sites.
MAJOR (dijkstra): window.MC_GEO_FILTER is set asynchronously by roles.js's
/api/config/client fetch, but nodes.js read it synchronously during
loadNodes(). On a cold page load where the Nodes page renders before that
fetch resolves, every node was silently classified as domestic until the
user clicked a filter chip. loadNodes() now awaits window.MeshConfigReady
(which never rejects) before the geo-scope filter runs. New regression
test proves the race with a manually-controlled promise, and fails
against the pre-fix code (verified via git stash).
MINOR: stale test-file header comment (still described the filter as
using the `foreign` flag, predating 6012ed0's lat/lon rewrite); loose
interface{} != float64 comparison in the geoFilter config test replaced
with explicit type assertions so a shape change fails loudly instead of
comparing unequal to the wrong type.
NIT: trimmed an 8-line struct-field doc comment; added
role="group"/aria-label to both the Status and Domestic/Foreign
filter-button groups on the Nodes page (bot noted Status already lacked
one too).
geo_filter, node blacklist, and hidden-name-prefix were all applied to a
page AFTER the SQL LIMIT already fixed its size, so a page that was
genuinely full at the DB layer could come back shorter than requested.
fetchAllNodes (public/app.js) treats "page shorter than requested" as
"this was the last page" and stops — so a single filtered-out row
anywhere in a page silently truncated everything after it.
Found on stg.meshview.dk: the live map showed ~450 of ~1300 GPS-valid
nodes because a hidden-name-prefix node happened to land in the first
500-row page.
handleNodes now loops, re-fetching and re-filtering additional DB pages,
until it has collected a full page of post-filter results or the DB
itself runs dry (a raw page shorter than requested = genuine end of
data) — restoring the "short page really means end of data" contract
fetchAllNodes relies on. Regression test reproduces the exact scenario
and fails against the old code (verified via git stash).
The `foreign` flag only reflects nodes whose ADVERT was decoded and
classified AFTER geo_filter was configured. On stg.meshview.dk this made
the just-shipped filter nearly useless: only 27 of 1503 nodes were
flagged foreign, while 570 actually have GPS placing them outside the
configured box (543 of those simply never got re-flagged).
Exposes the geo_filter box/polygon via /api/config/client (new
GeoFilter field on ClientConfigResponse), and adds nodePassesGeoFilter
+ helpers to app.js — a faithful JS port of internal/geofilter's
PassesFilter/PointInPolygon/DistToSegmentKm, cross-checked line-for-line
against the real Go implementation for bbox, polygon, and buffer cases.
Nodes with no GPS fix (or (0,0)) count as domestic, matching both the
Go semantics and the user's explicit choice for this feature.
nodes.js's Domestic/Foreign filter-group (added in f99df27) now calls
nodePassesGeoFilter(n.lat, n.lon, window.MC_GEO_FILTER) instead of
reading n.foreign.
Bot review on PR #1852 (munger, MAJOR-adjacent→MINOR): "bounded by the
1-byte channel hash space" only holds for encrypted channels (256
enc_%02x buckets) — plain-text CHAN channels use a free-form name
string (ingestor/db.go), so the bound isn't as hard as the comment
claimed. Correctness of the uncap itself is unaffected; this is a
doc-only fix.
GetChannelScopeAdoption was capped at the top 30 channels by message
volume, silently hiding quieter channels from the Scopes tab. Channel
cardinality is bounded by the 1-byte hash space, so the cap wasn't
protecting against anything real — removed it.
Client gets a filter (All/Unencrypted/Encrypted) based on the 'enc_'
channel_hash prefix db.go already uses for channels the ingestor
couldn't decrypt, so a visitor can separate plain channels from
encrypted ones without hunting through a long unfiltered list.
- MAJOR: config.example.json was missing geoFilterExemptNodeList
entirely — every other Config field in this repo gets a sibling
_comment_<field> entry (established, actively-used convention
throughout the file, ~20 existing examples), documented here to
match. Also corrected the adjacent geo_filter _comment, which
predates this PR and claimed it "restricts ingestion" — it doesn't;
that's foreignAdverts.mode=drop's job, unrelated to this flag.
- MINOR: ?geoFilter=<v> only recognized the literal string "1" to
enable and treated anything else (including "true") as disabling —
a real footgun since "true" reads as an obviously valid boolean
value. Now accepts 1/true and 0/false as explicit overrides; any
other value (including absent) falls through to the deployment
default instead of silently flipping it off.
- Tightened the Foreign Traffic tab's stop-hook test: it previously
only asserted stop() doesn't throw, which a no-op stub would also
pass. Sandbox's setInterval/clearInterval are now real spies
tracking live/cleared ids, so the test verifies render() registers
exactly one interval and stop() actually clears that exact id
(idempotently on a second call).
- NIT: relays.sort now wraps unscoped_relay_count_24h in Number(...)
defensively (AGENTS.md § Type Safety — cast at the boundary).
- NIT: foreignNodes.sort now parses last_seen once per node
(decorate-sort-undecorate) instead of twice per comparator call.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
7cc5aae's GeoFilterAppliesToNodeList had the polarity backwards: it
defaulted to false, meaning geo_filter would NOT apply to the node
list unless a deployment explicitly opted in — but every existing
deployment that already had geo_filter configured (predating this
field entirely) would decode that field as false too, since it's
simply absent from their config.json. They'd silently get the new
"show everything" behavior on upgrade with no way back short of
reading release notes and adding a new config key — exactly the
breaking change the opt-in gate was meant to prevent, just one layer
deeper.
Renamed to GeoFilterExemptNodeList (still defaults false): geo_filter
now applies to the node list by DEFAULT when configured, matching the
long-standing #730 behavior exactly, for every deployment predating
this field. Only a deployment that explicitly sets
geoFilterExemptNodeList=true (i.e. one adopting geo_filter fresh,
purely for foreign_advert classification/analytics) gets the
non-filtering default. ?geoFilter=0/1 still overrides either default
for a single request.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Per bot review on PR #1852 (comment 5012514227): the ?geoFilter=1
opt-in change (ffb2c84) flips a public API's default behavior with no
migration path. On THIS deployment geo_filter was never configured
before this session, so nothing was ever relying on the old filtered
behavior here — but the PR targets the shared upstream codebase, and
any other deployment that already had geo_filter configured (the
pre-existing #730 declutter feature) would silently start getting an
unfiltered node list after upgrading to this code, with no way back.
Adds Config.GeoFilterAppliesToNodeList (default false, matching the
new non-surprising behavior for anyone configuring geo_filter fresh).
A deployment that intentionally relies on the old always-on filtering
sets it to true in config.json to keep exactly the behavior it had
before this PR.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Configuring geo_filter alone was silently changing what GET /api/nodes
returned — every node outside the polygon and not yet foreign_advert
-tagged (which only happens on that node's next ADVERT after
geo_filter is set) vanished from every view built on this endpoint,
including the live map. There was no way to see them again short of
waiting for each one to re-advertise.
geo_filter's own purposes — the ingestor tagging foreign adverts, and
the explicit prune-geo-filter admin flow — are untouched; this only
gates the passive node-list declutter view behind an explicit query
param, so turning geo_filter on to build analytics (e.g. the Foreign
Traffic tab) doesn't have the side effect of hiding nodes from the map.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Working through the outstanding non-blocker findings across all five
automated review passes on the PR:
- MAJOR: GetChannelMessages' rows.Scan() error was silently discarded
— a schema mismatch would produce zero-valued messages instead of a
visible error. Now returns the error, matching sibling query loops.
- Extracted the hashRegions name-normalization rule (trim, "#"-prefix,
dedupe) shared between cmd/ingestor's loadRegionKeys and cmd/server's
region-utilization diff into a new internal/regions package, so the
two can no longer drift apart on the rule independently.
- Batched GetRepeaterNamesByKeys' SQL IN (...) clause in chunks of 500
— an unbounded clause risks SQLITE_MAX_VARIABLE_NUMBER on very large
deployments' byPathHop candidate sets.
- handleScopeStats' remaining silently-swallowed err==nil branches
(GetMatchedRegionNames, GetNodesByDefaultScope,
GetChannelMessageScopeStats, GetChannelScopeAdoption) now log a WARN
breadcrumb on failure instead of failing invisibly.
- Scope Adoption table was rendering 3 of the 4 ChannelScopeAdoption
fields (Unscoped omitted) — the visible numbers didn't reconcile to
the message total without doing the subtraction by hand. Added the
column.
- Removed a duplicate `vertical-align: middle` declaration on
.badge-transport (dead, not a behavior change).
- Deleted ChannelMessageResp — flagged for interface{} vs *string/*int
typing, but turned out to be completely unused dead code; removing
it resolves the finding more directly than retyping something
nothing constructs.
Left two NIT/MINOR items as-is with reasoning:
- renderRegionNodeGroups' inline styles match the same pattern used in
15+ other places in analytics.js — "fixing" only this one function
would make it less consistent with the file, not more.
- TransmissionResp's interface{} fields (a different struct than the
one just removed) are an established, actively-used pattern for
nullable SQL-scanned values across that whole response type;
retyping it is a much larger, unrelated refactor.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
channelKeys and regionKeys (derived from config.json's hashChannels/
hashRegions) were loaded exactly once in main() and held in plain
maps for the process lifetime — adding a new hashtag channel or
region required a full ingestor restart to take effect, which also
resets the in-memory relay/scope-attribution state built up since the
last restart (see the Repeaters-by-Region cold-start discussion).
Wraps both maps in atomic.Pointer-backed hotKeys (same pattern already
used for Config.NodeBlacklist, just actually wired to a trigger this
time) and adds a SIGHUP handler that re-reads config.json and
atomically swaps in freshly-derived keys:
kill -HUP <ingestor-pid>
On a malformed config during reload, the previous keys are kept and
the error is logged — a bad edit must not blank out a working
ingestor. The MQTT message-handler closure now pulls a fresh snapshot
from the atomic pointers per message instead of closing over a static
map, so no other call sites needed to change.
Adds channelScopeAdoption to /api/scope-stats: the existing
ChannelMessages aggregate (scoped/unscoped/unknown for channel chat)
broken down PER CHANNEL — which specific channels (#test,
#wardriving, ...) actually use region scoping vs which never do.
Ordered by message volume, capped at the top 30 channels.
Frontend renders a compact table under "Channel Messages": channel
name, total messages, scoped count+%, unknown count.
Adds hourlyActivityByRegion to /api/scope-stats: each region's message
counts bucketed by hour-of-day (0-23 UTC), aggregated across every day
in the window — answers "when during a typical day is this region
active" rather than "how did volume change over the window" (that's
the existing chronological TimeSeries chart).
Frontend renders a compact heatmap: one row per region, 24 hour
columns, color intensity normalized per-row (each region's own busiest
hour) so a quiet region's daily shape stays visible next to a loud
one instead of being crushed toward zero.
Adds bridgeRepeaters to /api/scope-stats: RepeatersByRegion inverted
into pubkey -> regions, keeping only repeaters that have relayed
traffic for MORE than one region. These are the mesh's literal
backbone nodes connecting otherwise-separate regional communities —
losing one is a more consequential failure than losing a
single-region repeater.
Computed inline while building RepeatersByRegion (reuses the same
byRegion map and role-filtered names lookup, no extra queries).
Frontend renders a small table under "Repeaters by Region": repeater
name (linked to its node detail page), region count, and the region
list.
Adds channelMessages to /api/scope-stats: the same scoped/unscoped/
unknown question as the main Summary, but restricted to payload_type=5
(channel chat) instead of all observed traffic. Most channel chat is
plain FLOOD rather than transport-scoped, so this can read very
differently from the all-traffic numbers — answers "how many of our
actual channel messages carry a region scope" directly instead of
requiring the reader to infer it from the broader stats.
New GetChannelMessageScopeStats() mirrors GetScopeStats' query shape
but scopes TotalMessages to ALL route types for payload_type=5 (not
just route_type 0/3), since restricting to transport routes would
answer a different question than "how many channel messages, period".
Frontend renders a small "Channel Messages" stat-card row under the
main summary cards, window-scoped like the rest of the tab.
Root-caused via a real report: "Repeaters by Region" showed a hyper-
local scope (#dk-fyn-middelfart) as transported by repeaters spread
across the whole country, which shouldn't be possible — MeshCore
firmware only relays a TRANSPORT_FLOOD/DIRECT packet when the
repeater's OWN configured region matches the packet's transport code
(examples/simple_repeater/MyMesh.cpp allowPacketForward, gated by
RegionMap::findMatch against the repeater's local region list).
Confirmed against the meshcore-dev/MeshCore source.
The bug: both TransportedScopes computation paths (bulk
computeRepeaterRelayInfoMap and per-node GetRepeaterRelayInfo) fold a
full pubkey's byPathHop entries together with its matching 1-byte
raw-prefix bucket — an intentional, existing fallback for RelayCount/
LastRelayed ("this node is probably active") that tolerates the
1-byte hash's inherent ambiguity (any node sharing that first byte
gets folded in). Applying the SAME fold to TransportedScopes asserted
something far more specific than the ambiguous signal can support,
and something the protocol itself wouldn't allow.
Scope accumulation now only happens on the exact-key (resolved,
unambiguous) pass; RelayCount/LastRelayed/RelayActive keep the
prefix-bucket fold unchanged, since those remain intentionally
approximate. Added relayEntry.fromPrefix to carry this distinction
through the per-node path, and rewrote the test that had pinned the
old (incorrect) folding behavior.
Adds originatingNodesByRegion to /api/scope-stats: nodes whose OWN
default_scope (#899) is a given region, complementing the existing
repeatersByRegion (transported_scopes) breakdown. The distinction
matters — a repeater can relay traffic for a region it isn't itself
configured with, so "who runs this region" and "who has carried this
region's traffic" are different, both useful questions.
Frontend: refactored the per-region collapsible-list rendering (used
by both breakdowns) into a shared renderRegionNodeGroups() helper
instead of duplicating the HTML-building logic, and added a "Nodes
Running This Region" section alongside "Repeaters by Region".
byPathHop indexes both full pubkeys and short hex-prefix "bucket" keys
used internally for ambiguous-hop resolution — GetRepeaterRelayInfoMap
returns TransportedScopes for every one of those keys indiscriminately.
The first deploy of repeatersByRegion iterated the raw map and fell
back to showing the bucket key itself when no name matched, so a
handful of short internal keys were counted as "repeaters" (stg showed
594 "repeaters" for #dk — every active repeater plus every 2-6 char
bucket key that ever touched a #dk packet).
GetNodeNamesByKeys is now GetRepeaterNamesByKeys and filters
`role IN ('repeater','room')` in the SQL itself, so a key only survives
if it's a real node — bucket keys never match a nodes.public_key row
and are dropped rather than falling back to the raw key.
Adds repeatersByRegion to /api/scope-stats: for every region that has
ever matched a transmission, which distinct repeaters/rooms have
relayed traffic carrying that scope. Sourced from the same 5-min
background-recomputed bulk relay-info cache the Nodes page already
uses (GetRepeaterRelayInfoMap / TransportedScopes, #1751) — no new
expensive computation, just an inversion + name lookup.
Frontend renders a collapsible per-region repeater list (name links
to the node detail page) under a new "Repeaters by Region" section,
explicitly framed as a coverage/redundancy signal: a region carried
by only one repeater is a single point of failure for that area.
Adds configuredRegions/unusedRegions to /api/scope-stats: an all-time
(not window-scoped) diff between the operator's configured hashRegions
list and the set of scope_name values that have actually matched a
transmission still in retention. Surfaces how much of the region list
is dead weight — directly actionable evidence for pruning, which is
also the real fix for the HMAC-collision noise (fewer configured
regions -> lower birthday-collision probability per packet).
Server config now parses hashRegions (previously ingestor-only, same
config.json key) purely to read the configured names — no HMAC key
derivation happens server-side.
Frontend: a "Region Utilization" section on the Scopes tab shows
used/unused counts and a collapsible list of the unused region names.
The packet detail pane already showed scope (with an "unknown scope"
fallback for empty scope_name), but the packets table itself gave no
at-a-glance signal — the existing transportBadge() "T" marker only
encoded route type.
transportBadge() now takes an optional scopeName argument: a resolved
region enriches the tooltip, an empty scope_name (transport-eligible
but unmatched/ambiguous) renders as "T?" with a distinct muted badge
style instead of the confident amber, so it's visually distinguishable
from a resolved scope without relying on color alone. Existing callers
that don't pass scopeName (live.js) are unaffected.
QueryGroupedPackets (SQLite + in-memory) didn't select scope_name at
all — added it, since the Packets tab defaults to the grouped view.
GetChannelMessages already returned an empty scope string for
transport-eligible packets whose region couldn't be determined (no
configured region matched, or matchScope now reports an HMAC
collision as unknown), but the UI treated empty scope the same as
"not applicable" and rendered nothing — indistinguishable from a
plain FLOOD/DIRECT message that never carries a scope at all.
Adds route_type to the channel-message payload (both SQLite and
in-memory paths, plus the decrypt-candidate and live WS paths) so the
frontend can tell "not transport-scoped" (routeType 1/2, no tag) apart
from "transport-scoped but unresolved" (routeType 0/3 with empty
scope, now shown as "Scope: unknown").
transport_code_1 is only 16 bits, so with enough configured hashRegions
a different, unrelated region's HMAC can coincidentally also match a
packet's real code1 (expected rate ~= len(regionKeys)/65534 per
packet). matchScope used to return the first match found while
iterating the region-key map, whose order Go randomizes per process —
so an ambiguous packet's assigned scope was effectively a coin flip
that could even change across ingestor restarts.
Confirmed on stg.meshview.dk (1098 configured regions): a live GRP_TXT
message's code1=C417 matched both "#dk" (the sender's true region) and
"#dk1906" (an unrelated collider), and was observed resolving to
either name across sends.
matchScope now scans every configured region and only returns a name
when exactly one matches; two or more matches are reported as
unknown-scoped ("") rather than guessed, with a log line for operator
visibility. Doesn't fix the root collision-probability cause (that
needs fewer/coarser configured regions) but stops confidently
displaying a wrong single answer.
The WS packet broadcast (IngestNewFromDB / IngestNewObservations in
cmd/server/store.go) builds its own map independent of the REST
response helpers, so scope_name was missing there even after it was
added to GetChannelMessages and txToMap. Adds it to both broadcast
paths and wires channels.js's live-append handler to read it, so
brand-new messages show their scope immediately instead of waiting
for the next periodic REST refresh.
Adds scope_name to GetChannelMessages (both SQLite and in-memory
store paths) and to the general packet response shape, then renders
it as "Scope: <name>" in the channel chat meta line so operators can
see which region scope a message was transported under.
Fixes#1838
## Problem
`/api/scope-stats` reported 100% scoped whenever any region was
configured. Reporter noticed on a scopeless instance that "unscoped" was
always zero — the pie visual is misleading to operators deciding on
`denyf *`.
## Root cause
`cmd/server/db.go:22` restricted the entire scope-stats denominator to
`route_type IN (0, 3)`. Per firmware `docs/packet_format.md § Route
Types`:
- `0` = `TRANSPORT_FLOOD`
- `1` = `FLOOD`
- `2` = `DIRECT`
- `3` = `TRANSPORT_DIRECT`
Only routes 0 and 3 carry `transport_code_1` (transport-level scope).
Routes 1 and 2 are inherently unscoped by protocol. The existing SQL was
correct for the "how many transport-scopable routes are actually scoped"
question, but the denominator was silently promoted to "all traffic" in
the UI. Bonus: the comment on `routeTypeTransportSQL` labelled routes
0+3 as "FLOOD (0) and DIRECT (3)" — wrong on both counts.
## Fix
- `cmd/server/db.go` — corrected the `routeTypeTransportSQL` comment;
added `routeTypeNonTransportSQL = "route_type IN (1, 2)"` alongside it.
- `GetScopeStats` runs a second `COUNT(*)` over `route_type IN (1,2) AND
first_seen >= ?` and folds that count into `Summary.Unscoped`. Same
index path as the existing query — one extra scan per `/api/scope-stats`
call (cached 30s per triage's carmack finding).
- `public/analytics.js` — Scopes tab header explains the denominator
(all observed transmissions) and which route types carry scope. Card
notes now render `X% of all traffic` for Scoped/Unscoped and `X% of
scoped` for Unknown Scope so the pie's denominator is explicit.
## TDD
- Red: `5554ffe4` — extended `TestGetScopeStats` +
`TestHandleScopeStats` with `route_type=1` and `route_type=2` rows and
asserted `Unscoped = 3` (1 transport-NULL + 2 non-transport). Ran the
tests and confirmed assertion failure (`Unscoped = 1, want 3`).
- Green: `ebbb9253` — implementation + label copy. Full `go test
./cmd/server/...` passes (54s).
## Preflight overrides
- check-branch-clean: justified — cross-stack fix by design (backend
semantics change + matching frontend label copy). All 4 files are
exactly the surface the triage comment identified.
## Verification
- `go test ./cmd/server/...` — 54s, all pass.
- Firmware confirmation: `firmware/docs/packet_format.md:20-24` (route
type table).
## Files touched
- `cmd/server/db.go` — comment fix + second COUNT query.
- `cmd/server/db_test.go` — extended fixture.
- `cmd/server/routes_test.go` — extended fixture + isolate from seed
data.
- `public/analytics.js` — labels and header copy.
---------
Co-authored-by: corescope-bot <bot@corescope.dev>
## Summary
Phase A of #1828: extract the 5 aggregate builders in
`handleObserverAnalytics` into pure helpers in a new
`cmd/server/observer_analytics.go`. Handler becomes a snapshot + filter
+ 5 composed calls.
Also adopts the `byTxID` direct-read in `buildPacketTypes` (issue body's
core observation): the payload-type histogram no longer allocates a full
`enrichObs` map + interface-boxed fields just to read `tx.PayloadType`.
That's the ~90% perf win the triage called out.
Scope is exactly Phase A per the second triage comment. Phase B
(sub-endpoints, caching, SQL migration) is deferred to a follow-up.
## Byte-identical output
- Timeline / NodesTimeline: same key set, same sort, same labels.
- PacketTypes: same keys/counts. Both legacy
(`enriched["payload_type"].(int)`) and new (`tx.PayloadType == nil`
guard) skip obs whose tx is missing or `PayloadType` is `nil`.
- SnrDistribution: same 2-unit floor bucketing (negative-side rounding
preserved), same ascending sort.
- RecentPackets: still the first 20 enriched observations (`enrichObs`
kept only here, where the extra fields are actually needed).
## TDD
- Red commit: `9dc62f43` — 7 unit tests fail on assertions (not build
errors) against stubs.
- Green commit: `8d41011d` — implementations + handler rewire. All new
tests + existing `TestObserverAnalytics*` handler tests pass.
## Preflight
`bash ~/.openclaw/skills/pr-preflight/scripts/run-all.sh origin/master`
→ clean (all 8 hard gates + 3 warnings pass).
## Non-goals
- No new endpoints.
- No SQL migration.
- No public API signature change.
- Snapshot count unchanged (still one under RLock, per #1481 P0-2).
Fixes#1828.
---------
Co-authored-by: fix-1828-bot <bot@corescope.local>
Co-authored-by: clawbot <bot@corescope>
Adds, per node, how many distinct FLOOD adverts it originated in the
last 7 days. Zero-hop adverts (route_type DIRECT) are excluded, so a
nearby observer hearing a node's cheap local adverts does not inflate
the number - the existing advert_count mixes both kinds and cannot tell
a chatty flooder (mesh-wide airtime) from
the recommended 240-minute zero-hop cadence (local only).
Consumers (the ArcScope repeater advisor) rate advert hygiene against
the community practice of one flood advert every ~49h; with the mixed
total, a correctly configured repeater looked chatty whenever an
observer sat within zero-hop range.
Implemented like the relay-liveness fields: a pure, unit-tested counter
over (first_seen, route_type, hash) entries with the same timestamp
parsing and hash dedup, fed by a from_pubkey-indexed query capped at the
2000 most recent advert rows. The flood route-type constant is named
advertRouteTypeFlood so this merges independently
of the open unscoped-relay PR (#1823).
---------
Co-authored-by: Waydroid Builder <build@waydroid.local>
Follow-up to #1823: TestRepeaterUnscopedRelayCount and its _Bulk twin
were ~30 verbatim lines apart (DB, node insert, store seeding,
assertions), differing only in the lookup under test - seeding changes
had to land twice. Both now use a shared seedUnscopedRelayFixture +
assertUnscopedCounts and contain only their
respective lookup call. No behaviour change; the relay-liveness suite
passes.
Co-authored-by: Waydroid Builder <build@waydroid.local>
## What
Adds a per-node API field `unscoped_relay_count_24h` on repeater/room
nodes: the
number of the node's 24h relay-hops that were unscoped floods
(route_type == ROUTE_TYPE_FLOOD). A strict subset of relay_count_24h.
## Why
A well-configured repeater runs `flood.max.unscoped 0` and should not
rebroadcast
unscoped floods — each one is re-sent by every repeater that hears it,
so one
packet turns into mesh-wide traffic. Exposing this lets clients (the
ArcScope
repeater advisor) detect and flag that base-config problem from observed
packets.
## How
Computed like relay_count_24h in both paths (bulk /api/nodes + per-node
detail)
with a route_type==FLOOD filter; reuses the byPathHop index, no
migration. Wired
into both handlers + OpenAPI schema + unit tests (per-node and bulk).
Co-authored-by: Waydroid Builder <build@waydroid.local>
## Fixes#1741
`TestBoundedLoad_OldestLoadedSet` (and any test building a 5000-row
fixture) hung/timed out, blocking reliable `go test ./cmd/server` and
CI.
## Root cause
The four test-DB builders in `cmd/server/bounded_load_test.go`
(`createTestDBAt`, `createTestDBWithObs`, `createTestDBWithAgedPackets`)
inserted rows in a loop with no `BEGIN`/`COMMIT`. With the pure-Go
`modernc.org/sqlite` driver every `Exec` auto-commits → one fsync per
row → ~2N fsyncs for N transmissions (tx + obs). At
`numTx=5000` that's ~10k fsyncs and the fixture blows past the test
timeout. Sibling tests with `numTx<=3000` happened to stay under the
timeout, so only the 5000-row cases visibly hung.
## Fix
Wrap each insert loop in a single `BEGIN`/`COMMIT` so the whole fixture
build becomes one commit. Fixtures now finish in well under a second
regardless of `numTx`; the tests' actual assertions (`oldestLoaded` set,
newest-first ordering, bounded load) are exercised instead of the
timeout masking them. Also made the
prepared-statement `Exec` calls check their error (previously discarded)
so a failed insert surfaces instead of silently leaving the DB short.
No production code changed — test infrastructure only.
## Verified
- `TestBoundedLoad_OldestLoadedSet`: **0.18s** (was: 30s timeout /
FAIL).
- Full `TestBoundedLoad*` + retention group: passes in ~1.2s.
- `go test ./...` in `cmd/server`: exit 0 (no longer blocks on this
test).
Co-authored-by: Waydroid Builder <build@waydroid.local>
Co-authored-by: Claude <noreply@anthropic.com>
## Fixes#1729
The firmware-default **Public** channel (channel-hash byte `0x11` = 17)
was rendered as an opaque **"Encrypted (0x11)"** row at the bottom of
the analytics Channels tab, despite the key being well-known and
builtin.
## Root cause
`computeAnalyticsChannels` applied the #978 rainbow-table validation
(`SHA256(SHA256("#name")[:16])[0]`, the **hashtag** hash scheme) to
every decoded channel name. The Public channel is a **PSK** channel
whose hash byte is key-derived (`SHA256(key)[0]` = 17), not
hashtag-derived (`186` for `#Public`). So the ingestor-decoded name
`"Public"` failed the hashtag check and was discarded, the row forced to
`encrypted=true, name="ch17"`.
## Fix
Trust the ingestor's `decryptionStatus`. The ingestor already persists
`decryptionStatus:"decrypted"` when it decoded a packet with a real key
(PSK), and `"no_key"` / `"decryption_failed"` otherwise. When the packet
is `decrypted`, skip the hashtag hash check and keep the name — it came
from a key-based decryption, not a
rainbow-table lookup. The #978 mismatch rejection still applies to
non-decrypted packets, so rainbow-table collisions are still caught.
Frontend needs no change: `encrypted=false, name="Public"` lands in the
"Network" group (top), not "Encrypted".
## Tests
- `makeGrpTx` gains `makeGrpTxWithStatus` companion to set
`decryptionStatus`.
- `TestComputeAnalyticsChannels_PublicChannelPreserved`: hash 17 /
"Public" / `decrypted` → name stays `"Public"`, `encrypted=false`.
- `TestComputeAnalyticsChannels_UndecryptedNameStillValidated`: a
non-`decrypted` name failing the hashtag check is still downgraded to
`ch17` (#978 regression guard).
All channel-analytics tests pass; `go build ./...` clean.
Co-authored-by: Waydroid Builder <build@waydroid.local>
Co-authored-by: Claude <noreply@anthropic.com>
Red commit: `3898dbc5` (verified locally — CI run URL pending)
## Problem
A standalone `mqtt-broker` container (`eclipse-mosquitto:2`) was
provisioned out-of-band on the staging VM. It now owns MQTT, is attached
to external docker network `meshcore-net`, and binds host port `8883`.
The current `docker-compose.staging.yml` still:
- Publishes `1883:1883` on the host (dead weight; conflicts the moment
the broker moves to that port).
- Defaults `DISABLE_MOSQUITTO=false`, so the in-container mosquitto
burns RAM and briefly contests the `mqtt-broker` docker DNS name on cold
start.
- Doesn't join `meshcore-net`, so the ingestor can't resolve
`mqtt-broker:1883` via docker DNS without manual surgery.
## Fix (`docker-compose.staging.yml` only)
1. Remove the `1883:1883` host port publish from `staging-go`.
2. Flip `DISABLE_MOSQUITTO` default from `false` to `true`. Operators
can opt back in with the env var.
3. Attach `staging-go` to both `default` and `meshcore-net`; declare
`meshcore-net` as `external: true` so the file never tries to
create/destroy operator state.
Healthcheck and Caddy/443 plumbing untouched (out of scope).
## Test added (TDD framing: Option A — Go shape-asserts)
`cmd/server/staging_compose_broker_test.go:1` adds four regex-based
assertions on the compose file shape:
- staging-go does **not** bind port `1883` in ANY form (quoted/unquoted
short form, or long-form `target: 1883` / `published: 1883`).
- `DISABLE_MOSQUITTO` uses the interpolated default form
`${DISABLE_MOSQUITTO:-true}` (preserves operator override). Bare literal
`true`, or a later `=false` override in the same env block, is rejected.
- Top-level `networks:` declares `meshcore-net` as `external: true`.
- `staging-go` attaches to `meshcore-net` via a real
`services.staging-go.networks:` sub-key (comment-stripped so an
in-comment example can't masquerade).
Regex (not YAML byte-equality) so cosmetic edits don't break the guard.
No new go module deps. Red commit `3898dbc5` fails all 4 assertions on
master. Green commit `38297ff4` makes them pass. Round-1 hardening
commit `9f7155e2` tightens the regexes (per adversarial + kent-beck
must-fixes) and was verified against master's YAML shape — all 4 tests
fail on `origin/master`'s compose, pass on branch, proving the tightened
regexes still gate a real regression.
## Risk
Low, with one intentional semantic change.
- **Semantic change (v3.7+):** `DISABLE_MOSQUITTO` in
`docker-compose.staging.yml` now defaults to `true`. This is a
**deliberate flip** — the standalone `mqtt-broker` container is now
authoritative on the staging host, and running the in-container
mosquitto alongside it wastes RAM and races the docker DNS name
`mqtt-broker` on cold start. Operators who want the pre-v3.7 shape
(in-container mosquitto + host-published `1883`) must explicitly opt
back in via env override AND re-add the `1883:1883` port mapping
(concrete snippet is inline in the compose file and in `DEPLOY.md` under
"Standalone MQTT broker (staging)"). This intent is called out in a
`SEMANTIC CHANGE (v3.7+)` header comment at the top of
`docker-compose.staging.yml`.
- **Deploy prereq:** the external `meshcore-net` docker network MUST
already exist on the host before `docker compose up`. If it doesn't,
compose refuses to start `staging-go`. This is documented inline in the
compose file (with the `docker network create meshcore-net` one-liner)
and in `DEPLOY.md`.
- **Only takes effect where the standalone broker is deployed** — which
it already is on staging today. The legacy `DISABLE_MOSQUITTO=false`
path remains reachable via env override; the ingestor's upstream config
is untouched.
Partial fix — no tracking issue; follow-up to operator-side broker
provisioning.
---------
Co-authored-by: corescope-bot <bot@corescope.local>
Co-authored-by: openclaw-bot <bot@openclaw.local>
Partial fix for #1809.
Red commit: c9c782b5 (CI runs on PR open; standalone red-branch CI not
configured — local repro proves the gate, see below).
## Problem
Issue #1809: at startup the background fill loader logged `background
load FAILED` within seconds and `backgroundLoadFailed=true` was set,
leaving the coverage gate tripped even though LoadChunked itself
completed normally.
## Root cause
`main.go:225-245` spawned `go store.loadBackgroundChunks()` as soon as
`FirstChunkReady` fired (chunk #1 = 10000 tx). But `s.oldestLoaded` is
only assigned at the end of `LoadChunked` (`chunked_load.go:329-333`),
~tens of seconds later. The bg loader read `oldestLoaded==""` at
`store.go:1462-1466`, broke out immediately, walked zero chunks, and the
coverage gate at `store.go:1543-1554` flipped
`backgroundLoadFailed=true`.
## Fix (initial)
Introduce `PacketStore.RunStartupLoad(chunkSize)` (`chunked_load.go`).
It runs `LoadChunked` first; only on success and only when
`hotStartupHours > 0` does it call `loadBackgroundChunks`. `main.go`
invokes `RunStartupLoad` in the same goroutine pattern as before, so
`FirstChunkReady` still unblocks the HTTP listener bind at chunk #1 —
only the bg loader is gated.
## Round-1 followups (this push)
Reviewer-driven hardening on top of the initial fix:
### Production behavior (commit db5592f6)
- **Steady-state semantics tightened.** `RunStartupLoad` now picks a
terminal state on every branch:
- LoadChunked error → `backgroundLoadFailed=true` with captured error
(was: `done=false, failed=false` indefinite).
- `hotStartupHours == 0` → `backgroundLoadDone=true` immediately,
`progress=100` (was: `done=false` forever → healthz stuck on
`backgroundLoadComplete=false`).
- Successful hot-window path → terminal state is whatever
`loadBackgroundChunks` sets (#1690 semantics, unchanged).
- **Runtime invariant assertion (A7).** `loadBackgroundChunks` panics
when `oldestLoaded==""` and packets exist — a future refactor that
re-introduces the parallel-spawn race fails loudly instead of silently
shipping the same coverage regression.
- **`RunStartupLoad` cleanup.** Inlined the superfluous goroutine +
channel that wrapped `LoadChunked` (direct call is equivalent).
- **Logging.** Added an INFO line between `LoadChunked` completion and
bg-loader start (the #1809 post-mortem needed exactly this signal).
Fixed the lying `"background load will start"` log that fired even on
the `hotStartupHours==0` branch.
- **Immutability documented.** `hotStartupHours` is now explicitly
documented as immutable post-construction, so the lock-free reads in
`LoadChunked` / `RunStartupLoad` / `loadBackgroundChunks` are sound.
### Test coverage (commits db5592f6 + e9e12acf)
- **Tautology fix (B1, commit e9e12acf).** The original
`Test1809_StartupLoad_BgLoaderSeesOldestLoaded` fixture seeded all 100
rows inside the 1h hot window, so `LoadChunked` alone produced
coverage=1.0 — the test passed even if `loadBackgroundChunks` was a
no-op. Rewrote the fixture to spread 100 rows over 14 days with
`hotStartupHours=24`, so only ~7 rows are hot and the remaining ~93 MUST
be loaded by the bg loader for the assertions to hold. Original
red-commit assertions kept intact; added `len(packets) > hot-only cap`
and `oldestLoaded < hot-cutoff - 12h` assertions on top.
- **New tests (commit db5592f6, `runstartup_load_test.go`)** codify the
new contracts:
- `TestRunStartupLoad_HotStartupHoursZero_SetsDoneImmediately`
- `TestRunStartupLoad_LoadChunkedError_SetsFailedTerminal`
- `TestRunStartupLoad_EmptyDB_SetsDoneTerminal`
- `TestRunStartupLoad_BgLoaderRunsAfterLoadChunkedSets_OldestLoaded`
- `TestLoadBackgroundChunks_PanicsOnOldestLoadedEmpty_Invariant`
### Docs (commit 70fa16f7)
- Package-level doc in `chunked_load.go` now documents `RunStartupLoad`
as the orchestrator entry point alongside `LoadChunked` /
`loadStatusMiddleware` / `OnChunkLoaded`.
### Preflight (commit eec1b48c)
- Test-fixture DDL annotated with `// PREFLIGHT: async=true
reason="unit-test fixture"` so the async-migration gate distinguishes
ephemeral test schema from prod migration paths.
## What's still NOT fixed (left intentionally open)
This PR addresses the startup race specifically. Issue #1809 will be
closed by the operator after observing healthy startup logs (`background
load complete: ... coverage=100.0%`) in prod for a full restart cycle.
Do not auto-close — leave open until the operator verifies.
## Test
Local repro (red branch state, before green commit): `go test
./cmd/server/ -run Test1809_StartupLoad_BgLoaderSeesOldestLoaded` → FAIL
on assertion `backgroundLoadFailed=true ...
oldest="2026-06-30T18:56:09Z"`. After green commit + round-1 followups:
PASS. Full `cmd/server/...` suite: ok in ~70s.
## Risk
Low — startup path only. New behavior gates surfaced by the new tests;
coverage-gate semantics unchanged. Runtime panic is a new failure mode
but only fires on a state (`oldestLoaded=="" && len(packets)>0`) that is
unreachable on the current code path — it exists solely as a refactor
tripwire.
---------
Co-authored-by: mc-bot <bot@corescope.local>
Co-authored-by: openclaw-bot <bot@openclaw.local>
Co-authored-by: meshcore-bot <bot@meshcore>
# Partial fix for #1749 — MQTT watchdog escalation + panic recovery +
tick exposure
Red commit: 9912bbb3e9 (CI run:
https://github.com/Kpa-clawbot/CoreScope/actions/runs?branch=fix%2Fissue-1749)
## Problem
Production CoreScope v3.9.1 (and a more recent prod recurrence on
2026-06-30 with the wcmesh source on `ssl://mqtt2.wcmesh.com:8883`)
showed two distinct watchdog failure modes:
1. **Per-source paho machinery dies silently.** `IsConnectedFn` returns
false; paho's `SetAutoReconnect(true)` never retries. The watchdog's
`processLivenessTransition` deliberately stays silent on
`LivenessDisconnected`, trusting paho to recover — so there is no
escalation path when that trust is misplaced.
2. **Watchdog goroutine death.** Three sources went silent within ~60s
of each other; no `WATCHDOG` log lines for 75 min. The most plausible
single point of failure is a panic inside `emit` (e.g. a blocked log
pipe) killing the loop with no defer/recover.
## Changes
**`cmd/ingestor/mqtt_watchdog.go`**
- Added `disconnectedReconnectMultiplier = 5` constant.
- Added `DisconnectedSinceUnix int64` (atomic) on `SourceLivenessState`.
Stamped on the first tick the source is observed disconnected; cleared
on any non-disconnected tick.
- `processLivenessTransition` now escalates: when `(now -
DisconnectedSinceUnix) > multiplier × threshold`, emits a `WATCHDOG
ESCALATION` WARN and calls `maybeForceReconnect` (subject to existing
`forceReconnectThrottle`). Distinct from the existing `LivenessStalled`
path so operators can grep escalation events independently.
- Added package-level `watchdogLastTickUnix atomic.Int64` +
`WatchdogLastTickUnix()` getter. The loop stamps it BEFORE per-source
processing — a wedged source-handler does not freeze the clock for an
external observer.
- `runLivenessWatchdogLoop` wraps each per-source
`processLivenessTransition` call in `func() { defer recover; ... }()` so
a panic in `emit` (or in any per-source code path) is logged and
skipped, not fatal. The loop continues to the next source and the next
tick.
**`cmd/ingestor/stats_file.go`**
- Added `WatchdogLastTickUnix int64` field on `IngestorStatsSnapshot`
(additive, `omitempty`); populated from `WatchdogLastTickUnix()` each
stats tick.
**`cmd/server/mqtt_status.go`**
- `MqttStatusResponse` gains `WatchdogLastTickUnix int64` (additive,
`omitempty`) sourced from the ingestor stats file; surfaced via `GET
/api/mqtt/status`.
**`config.example.json`**
- No new config field added — the multiplier is a code constant (5×) per
the issue's "N×threshold (e.g. 5×)" recommendation. The active
per-source `threshold` is the 5-minute scan threshold hard-coded at the
sole `runLivenessWatchdog` callsite (`cmd/ingestor/main.go:460`:
`runLivenessWatchdog(60*time.Second, 5*time.Minute)`), so escalation
fires at ~25 minutes (5 × 5min) of continuous disconnect — plus a
deterministic per-source jitter of 0..30s (#1810 round-1, see Taleb #4)
to avoid synchronized escalation across N sources sharing an upstream
broker outage.
## Acceptance (#1749)
- [x] Persistent `LivenessDisconnected` > N×threshold → force-reconnect
+ WARN
- [x] Watchdog goroutine liveness clock exposed (`WatchdogLastTickUnix`
in `/api/mqtt/status`)
- [x] Test: `IsConnectedFn` false for >5×threshold → assert
`ForceReconnectFn` invoked at least once
- [x] Test: panic in `emit` → assert loop recovers and continues ticking
## Test plan
- 4 new tests in `cmd/ingestor/mqtt_watchdog_1749_test.go` (all RED on
master, GREEN on this PR).
- Existing watchdog tests (`mqtt_watchdog_force_reconnect_test.go`,
`mqtt_reconnect_test.go`, r1/r2/m1 suites) continue to pass — the
escalation path is additive.
## Preflight
- TDD: red commit pushed and asserted to fail BEFORE green commit
landed.
- PII grep: clean on diff and PR body.
- Worktree: `_wt-fix-1749` on branch `fix/issue-1749`.
---------
Co-authored-by: corescope-bot <bot@corescope.dev>
Co-authored-by: openclaw-bot <bot@openclaw.local>
Co-authored-by: bot <bot@local>
## Summary
Replaces the three drifted per-surface payload-type label vocabularies
with a single canonical map keyed by firmware enum name.
Per the locked triage comment on #1799
([comment-4823975431](https://github.com/Kpa-clawbot/CoreScope/issues/1799#issuecomment-4823975431)):
> Create `public/payload-labels.js` exporting `{GRP_DATA: {short:'Group
Data', long:'Group data packet', enumId:6}, ...}`. Migrate `packets.js
typeMap`, `packet-filter.js FW_PAYLOAD_TYPES`, `live.js TYPE_COLORS
legend` to consume it. E2E that scrapes each surface and asserts label
equality.
## Changes
- **`public/payload-labels.js`** (new) — canonical map exposed as
`window.PayloadLabels` and `window.PayloadLabelsApi`. Keys are firmware
enum names; values carry `{short, long, enumId}` plus derived
`SHORT_BY_ID` / `FW_PAYLOAD_TYPES` / `TYPE_ALIASES` for legacy callers.
- **`public/packets.js`** — `TYPE_NAMES` + `typeMap` now read from
`PayloadLabelsApi.SHORT_BY_ID`. Literal kept only as a defensive
fallback for the case where the script tag fails to load.
- **`public/packet-filter.js`** — `FW_PAYLOAD_TYPES` + `TYPE_ALIASES`
now sourced from `PayloadLabelsApi`. Literal fallback retained so `node
test-packet-filter.js` still works headlessly.
- **`public/live.js`** — legend `<li>` rows are now generated from
`window.PayloadLabels` in stable order, killing the third-vocabulary
`Message — Group text` / `Direct — Direct message` drift the #1797
review surfaced.
- **`public/index.html`** — `<script src="payload-labels.js">` loaded
before `roles.js` / `packet-filter.js` / `packets.js`.
- **`test-issue-1799-label-vocab-e2e.js`** (new) — Playwright E2E.
Scrapes `#liveLegend` rows and the `/packets` type-filter checklist,
asserts each label matches `window.PayloadLabels[ENUM].short` for
`TXT_MSG`, `GRP_TXT`, `GRP_DATA`. Also verifies `window.PacketFilter`
still recognises the enum names.
- **`.github/workflows/deploy.yml`** — wired the new E2E into the
existing Playwright block.
## TDD trail
- Red commit `eb392d4` — adds the failing E2E only (asserts
`window.PayloadLabels` exists and labels match; both fail).
- Green commit `44e902a` — introduces the canonical map and migrates the
three surfaces.
## Verification
- `node test-packet-filter.js` — 92/92 pass with the new fallback
wiring.
- Preflight: `bash ~/.openclaw/skills/pr-preflight/scripts/run-all.sh
origin/master` — clean.
Browser verified: E2E `test-issue-1799-label-vocab-e2e.js` exercises
`/live` legend + `/packets` type filter against a Playwright headless
Chromium; CI's Playwright block runs it on every push.
E2E assertion added: `test-issue-1799-label-vocab-e2e.js:139` —
`assert(fromLegend === canon, ...)` and `assert(fromPackets === canon,
...)` per enum.
Fixes#1799
---------
Co-authored-by: mc-bot <bot@corescope>
Co-authored-by: openclaw-bot <bot@openclaw.local>
Co-authored-by: clawbot <clawbot@kpa.com>
Co-authored-by: clawbot <bot@clawbot.local>
Two server tests flaked intermittently and reddened CI on unrelated
(frontend)
PRs that merged master:
- TestDistanceConcurrentRequestsDuringBuildReturn202 asserted all 10
concurrent
requests get 202 'during the build window', but the lazy distance build
on the
tiny test DB finishes almost instantly, so on a fast machine some
requests
raced past it and got 200 (~50% flake). Add a nil-by-default
distanceBuildHook
seam on PacketStore (zero overhead in prod) that the test uses to hold
the
build open until all requests have been served — making the window
guarantee
deterministic.
- TestHandleNodePaths_AnchorBiasInconsistency_Issue1278 queried /paths
right
after store.Load(), racing the path-hop index that Load() builds in a
background goroutine (#1008); the membership/canonical result was thus
non-deterministic (rarer flake, worse under suite load). Wait for
PathHopIndexReady() before querying.
Both run 30x green and pass -race. No production behavior change (hook
is nil).
Co-authored-by: Waydroid Builder <claude@michael.arcan.de>
## Summary
Closes the wide-open `/ws` WebSocket upgrader (`CheckOrigin: return
true`) that lets any browser origin scrape live packet data. Replaces it
with an explicit allowlist consulted from `cfg.CORSAllowedOrigins`, plus
an implicit same-origin allowance and an empty-Origin (non-browser
client) allowance.
Fixes#1793.
## Rules (`Hub.checkOrigin`)
- Empty `Origin` header → **allow** (non-browser clients; per-IP
rate/deny gating tracked separately in #1794).
- `Origin` host == request `Host` (case-insensitive) → **allow**
(same-origin).
- `Origin` matches an entry in `cfg.CORSAllowedOrigins` by exact
case-insensitive match → **allow**.
- `"*"` in `cfg.CORSAllowedOrigins` is **deliberately ignored** for
`/ws`. A startup `[ws] WARNING:` is logged once when present.
- Anything else → **reject** (gorilla returns 403).
### Deliberate divergence from CORS XHR
CORS XHR (`corsMiddleware`) still honors `"*"` for read-only
cross-origin GETs. The `/ws` upgrade does NOT, per OWASP's WebSocket
Security Cheat Sheet:
> Use an allowlist, not a denylist. Avoid wildcards or substring
matching.
—
https://cheatsheetseries.owasp.org/cheatsheets/WebSocket_Security_Cheat_Sheet.html
`"*"` on the WS path would re-open the exact CSWSH/scraping vector this
PR closes, so it is rejected with a startup warning rather than silently
honored. This intentional asymmetry is documented in the updated
`_comment_corsAllowedOrigins` in `config.example.json`.
## TDD red → green
- `e5974c6a` **RED** — adds `cmd/server/websocket_checkorigin_test.go`
with five cases; `SetAllowedOrigins` introduced as an enforcement stub
so the test compiles and fails on the assertion (CI fails on this commit
by design).
- `a4791dc3` **GREEN** — implements `Hub.checkOrigin`, wires
`SetAllowedOrigins` from `main.go`, updates the config example. All
tests pass.
## Tests added (`cmd/server/websocket_checkorigin_test.go`)
- `TestCheckOriginRejectsForeignOrigin` — foreign Origin → 403
- `TestCheckOriginAllowsEmptyOrigin` — non-browser client → 101
- `TestCheckOriginAllowsSameHost` — same-origin → 101
- `TestCheckOriginAllowsAllowlistedOrigin` — exact allowlist match → 101
- `TestCheckOriginWildcardDoesNotAllowForeignOrigin` — `"*"` in
allowlist still rejects foreign origin → 403
## Files changed
- `cmd/server/websocket.go` — `Hub.allowedOrigins`, `SetAllowedOrigins`,
`checkOrigin`, wired into `Upgrader.CheckOrigin`.
- `cmd/server/main.go` — `hub.SetAllowedOrigins(cfg.CORSAllowedOrigins)`
at the single call site.
- `cmd/server/websocket_checkorigin_test.go` — new test file.
- `config.example.json` — updated `_comment_corsAllowedOrigins` to
document `/ws` gating and the `"*"` divergence.
## Out of scope (follow-up)
- **#1794** — per-IP rate limit / deny list / connection cap for
non-browser clients (which still bypass Origin because they don't send
one). Layered defense; not in this PR.
## Verification
- `go test ./cmd/server/...` — all server tests pass locally (574s).
- Preflight clean (`bash
~/.openclaw/skills/pr-preflight/scripts/run-all.sh origin/master`).
---------
Co-authored-by: openclaw-bot <bot@openclaw.local>
