TheUnboundDeveloper/AM-Reaper v3.3.1-beta-GT-BE98_PRO

Reaper v3.3.1-beta — GT-BE98_PRO

on GitHub

v3.3.1 — a minimum channel width on Auto, one factory state on every model, the setup box finishes the main network

• Minimum width on Auto (Wireless › Settings, 5 and 6 GHz). With Channel bandwidth on Auto the
  closed channel picker (acsd2) chooses the width as well as the channel, and it narrows to 40 or
  20 MHz when it judges part of a wide channel busy; preamble puncturing exists only at 80 MHz and
  wider, so each such narrowing removed the one state puncturing could have handled. A per-radio
  switch keeps the picker at 80 MHz or wider: 80/160 MHz on 5 GHz and 80/160/320 MHz on
  6 GHz name the widths it may still choose, and each client uses what it supports within the width
  the radio is on (160 MHz on 5 GHz still needs Enable 160 MHz). rc/reaper_bwfloor.c appends the
  radio's 20 and 40 MHz chanspecs (wl chanspecs) to wlN_acs_excl_chans at the end of
  set_acs_ifnames(), the list stock rebuilds at boot and every wireless restart; records them
  under /jffs/reaper_bwfloor/ (the list lives on /jffs and survives a reboot); strips exactly
  those entries before stock reads the list; and adds nothing when no channel at 80 MHz or wider
  would remain, because acsd2 discards an all-excluded list, DFS exclusions included. No closed
  code touched. Needs Auto channel; inactive on a fixed channel or width, an AiMesh node and
  2.4 GHz. Off by default (wlN_bwfloor). With the floor on, the puncturing rows unlock on Auto
  bandwidth (Follow channel lists the slices of the width the radio is on, taken from the
  watcher's report; Dynamic follows the radio); without it they stay locked there, and the
  hover reason says which of the two would open them.
• The floor lifts itself after radar. A radar hit vacates the channel, and with the floor armed
  the picker may have no wide channel to return to for the non-occupancy period. A watcher
  (/sbin/reaper_bwfloor, every 15 s) lifts a radio's floor when it has run narrower than 80 MHz
  for 120 s, or has been off the air for 180 s outside a DFS availability check (up to 11 minutes)
  and outside a Wi-Fi schedule: it removes its own entries and restarts acsd2, for 30 minutes,
  doubling on a repeat within two hours (four hours at most), then re-arms. One reaper_bwfloor
  log line per lift and re-arm; the Settings cell reads Holding, Lifted, Not active or Not applied;
  the diagnostics report (v1.3.25, section 7) adds a min width: line and flags a lifted floor or
  a stopped watcher. The rwatch tick restarts the watcher if it dies.
• Dynamic puncturing: the confirmation scan reads more than busy time. The passive scan that
  resolves a candidate already returns, per 20 MHz channel, the neighbouring networks heard there
  (channel, width and signal; names and addresses are not kept) and the PHY figures (noise floor,
  carrier-sense glitches, bad PLCP headers). Both count now: a neighbour's home channels (its 20 or
  40, or the 80 MHz block holding its control channel) take their counter group's persistent busy
  as a prior, weighted by signal, so a 110 ms snapshot that caught a gap no longer hides where the
  interference lives; a noise floor above the scan's quietest channel or a stream of PHY errors
  counts as busy on that channel even when carrier sense reads low. Neither ever lowers a reading.
  On a scan source a slice is put back only after a fresh scan (150 s) of its channels agrees,
  taken at most once per scan cooldown; the counters alone no longer restore there (Passive is
  unchanged). The Wireless Quality card lists the scanned channels (carrier sense / PHY impairment
  / neighbour signal) and the networks heard; the rpunctd scan line in the log carries the same.
  Decision core host-tested (test_punct_dyn.py, 38 checks).
• Dynamic puncturing: two passive inputs, and no scan may cost a client. (1) Every tick the
  controller reads the radio's own PHY figures (wl chanim_stats: glitches and bad PLCP headers
  per second, background noise in dBm) - the energy below carrier sense that never reads busy. The
  noise term is the rise above the radio's own quiet baseline on its channel (it follows a quieter
  reading at once and a louder one over about an hour): the whole-channel figure grows with the
  width, and a 320 MHz channel at rest reads about -79 dBm where a 160 MHz one reads -93. A
  second trigger (Conservative 35 %, Balanced 25 %, Aggressive 20 % on the scan's impairment scale,
  held like the busy trigger) may only request a confirmation scan of the puncturable channels,
  under the same cooldown and idle gate; a slice is chosen only where that scan reads it at trigger
  level. Its edges are logged with the raw figures for calibration. (2) Every 30 s (every tick
  while a change is being judged) it reads the station table (wl bs_data -noreset: a plain read
  clears the counters). Another reader resets that table every few seconds, so each read is taken
  as its own short window and the windows are summed; the per-packet width (txbw / acked) gives
  each client's operating width.
  Clients narrower than the radio break ties between qualifying slices (the slice outside their
  block wins), and 120 s after each change every client's rate and retries are compared with the
  interval before it: when at least two and at least half of the compared clients got worse
  (rate down a quarter and retries up five points), the slice comes off and is not retried on that
  channel; after a restore the same finding only lengthens the next restore hold. A confirmation
  scan that loses a client now stops scanning on that radio (Passive from then). The scan's
  impairment figure takes glitch and bad-PLCP counts as the per-second counts the vendor header
  documents; the r4 image scaled them by 9. The status line repeats only when its verdict
  changes, not when a busy figure moves a point. Host-tested (test_punct_dyn.py, 76 checks,
  including the station parser and noise baseline run on the router's real output; five
  mutants caught).
• One factory state on every model (field, GT-BE98). Units with a printed Wi-Fi key reset to
  that key with the Wi-Fi setup flagged done, so the first-boot banner showed only the password row
  and the setup box - and the main-network reconcile it runs - never ran; the Network page then
  showed an unclickable main card until an unrelated apply. restore_defaults_wifi() no longer
  applies the label key, label SSID or the done flag: every reset is admin/admin with open template
  Wi-Fi, secured through the setup box. The banner also treats a key equal to the printed key as
  factory, and the RSEC_5/9/13 texts say the factory Wi-Fi is open.
• The setup box reboots into a working main network. The main network's SDN rows are created
  only at boot (init_apg, gated on setup being complete), so the v3.2.9 in-place reconcile could
  never create them. The box now marks setup complete, saves and reboots (the page waits about
  three minutes); the result is one all-band main card. The Network page shows a notice pointing
  to the box while no main rows exist on a router or access point.
• The WAN pill follows the primary WAN. The shell's WAN pill, shown on every page but the
  dashboard, read the first WAN only and said Disconnected after a Dual WAN fail-over. It follows
  wan_primary now (either line up in load balance), as the dashboard's Internet card has since
  v3.2.8.
• AiMesh node cards count clients on a box that is not routing (field, GT-BE19000). In Access
  Point, repeater or media-bridge mode every node card on the AiMesh page read 0. The card is stock:
  it counts get_clientlist() rows whose isOnline is set and whose parent-node MAC is the node,
  and networkmap (a closed binary) derives that flag from DHCP leases and conntrack, which a
  bridging box does not have. Same cause as the v3.1.6 dashboard tiles. reaper_dev.cgi now names
  the node each device hangs off (via, from cfg_mnt's per-node wireless and wired client lists,
  this router otherwise), and on a non-routing box the topology page overlays that store after
  every client-list rebuild, read asynchronously and cached 30 s. Rows networkmap already placed are
  kept; MLO links and nodes are folded as on the dashboard. A routing box does not take the path.
  Tripwire: test_apmode_surfaces.py section 5.
• IPv6 Stateful also offers SLAAC (field, GT-BE98). In Stateful mode dnsmasq's dhcp-range
  carried no slaac flag, so the router advertisement had no autonomous prefix: hosts with a
  DHCPv6 client got an address, SLAAC-only hosts (most Linux VMs and containers, Android) got
  none. rc/services.c now adds slaac to the Stateful range (and its constructor: form) when
  the prefix is a /64, so the advertisement sets the A flag while DHCPv6 still hands out leases.
  Behaviour change: on Stateful, DHCPv6-capable hosts may now hold a SLAAC address beside their
  lease. Stateless and SLAAC-only modes are unchanged. Diagnostics section 6 now prints the
  autoconfiguration type, the dnsmasq range mode, the DHCPv6 lease count, ipv6_fw_enable and
  the ip6tables policies, with a finding when the LAN is DHCPv6-only.
• A DNS server on one of the router's VLAN networks is no longer routed out the WAN. On every
  WAN-up stock wan_up() adds a /32 host route via the WAN gateway for each WAN DNS server unless
  it is in the WAN subnet, the main LAN subnet or lan_route. Guest Network Pro / VLAN subnets were
  not checked, so a resolver on such a network (AdGuard Home on a VLAN, set as the WAN DNS server)
  got <dns> via <WAN gateway>, which beats the bridge's own route: the router's queries to it and
  the replies to its upstream lookups left through the WAN and timed out. rc/wan.c
dns_in_sdn_subnet() reads the saved VLAN subnets (get_mtsubnet(), so it is right even when
  the WAN comes up before the bridges) and skips the route for any address inside one.
• Interference mitigation (Wireless › Settings, per radio). A new row sets the radio driver's
  own interference mitigation: Driver default (Reaper writes nothing) or + HW ACI, which
  adds the hardware adjacent-channel mitigation bit (16) to the driver's mask (default 75 on this
  dongle). /sbin/reaper_punct applies it at start-up, after every wireless restart and on save
  without a radio restart; it records the value it found, undoes only its own write, and reports
  applied / driver / refused / waiting in /tmp/reaper_wlmit.state, shown as the row's status.
  The rwatch tick re-runs the applier while the row is set. A change resets dynamic puncturing's
  noise baseline and PHY-impairment state, since the mitigation moves those figures by design.
  In a 2.4 GHz microwave A/B on the RT-BE96U, + HW ACI did not improve client throughput and at
  times reduced it; the help text says so and recommends Driver default. Off by default
  (wlN_rmit). The diagnostics report (section 7) adds an interference: line per radio. Of the
  other driver functions trialled, dynamic BSS width (obss_dyn_bw) was dropped (it switches the
  whole BSS width by channel-switch announcement and cost clients), spatial reuse stays with
  stock's watchdog, and the energy-detect thresholds are regulatory and left alone.
• The floor's re-arm no longer moves the channel. When a lifted floor re-armed while the radio
  was already 80 MHz or wider, the watcher still restarted acsd2, which could pick a new channel
  and drop clients. The re-arm now only restores the exclusions and marks the radio
  picker-pending; acsd2 picks them up at its next restart, and is restarted by the watcher only
  if the radio later runs below the floor.
• Dynamic puncturing: a ceiling on the Active scan source. The Active source scanned whatever
  the radio's own airtime; it now scans only while the radio's own traffic is at or below 30 % of
  airtime (the idle-scan tier keeps its preset level), so a confirmation scan does not land in
  the middle of a busy client's stream. The status line says own airtime above the scan ceiling while it waits. test_punct_dyn.py 78 checks; test_wlmit.py covers the mitigation
  applier and the re-arm.

Images & checksums (GT-BE98_PRO)

Two flashable images: + AI Advisor (default) and Standard (noMCP, all AI components compiled out entirely). Flash the *_nand_squashfs.pkgtb via Administration > Firmware Upgrade.

This is a beta release. Its filename carries _BETA and the router reports the same string on the dashboard and the About page, so you can always tell which channel a flashed image came from. Stable releases carry no marker.

Verify a download against the attached SHA256SUMS-GT-BE98_PRO-Reaper_v3.3.1.txt.

Corresponding source & reproducibility

The GT-BE98_PRO image for v3.3.1-beta is built from this repository at tag v3.3.1-beta-GT-BE98_PRO: the pinned Asuswrt-Merlin base (3006.102.8-beta2, a7ebfa133a) plus the complete patch series. The tag freezes the exact source that produced it.

• Patches: patches/
• Provenance record: provenance/manifest.json
• How to verify: docs/REPRODUCIBILITY.md and docs/SOURCE-AVAILABILITY.md

The auto-attached Source code (zip/tar.gz) asset below is this repository at tag v3.3.1-beta-GT-BE98_PRO (patches + docs).