Use when UniFi Wi-Fi is slow, unstable, or being tuned: "my wifi is slow but speedtest on the router is fast", "great signal, terrible speed", "should I use 80MHz or 40MHz", "channel planning", "co-channel interference", "DFS channels", "my APs keep picking the same channel", "audit my SSIDs", "hidden SSID", or diagnosing throughput that collapses under load. Also roaming: "my phone stays stuck on the far AP", "sticky clients", "improve roaming", "should I enable fast roaming", "minimum RSSI", "TX power tuning". Covers the diagnostic ladder for slow Wi-Fi, channel width and DFS tradeoffs, safe radio writes, cell sizing and roaming persuasion, SSID hygiene, and the in-wall AP port trap. Assumes unifi-connect. Not for firewall policy between networks (unifi-firewall), wired port and client operations (unifi-clients).
SKILL.md
UniFi Wi-Fi
Most "slow Wi-Fi" is not slow Wi-Fi. It is a WAN problem, a backhaul problem, a
client problem, or a bandwidth cap somebody set two years ago. The diagnostic
ladder below exists to find that out cheaply, in that order, before touching
radios. Radio changes are the most disruptive and the most often wrong.
The diagnostic ladder
Run it top to bottom. Each rung eliminates a layer, and the failure that motivated
this skill was only visible at the last rung.
1. Rule out the WAN. Trigger a gateway-side speedtest, which measures the
gateway to the internet with no Wi-Fi involved:
udm raw POST /proxy/network/api/s/default/cmd/devmgr '{"cmd":"speedtest"}'
# wait, then read the RESULT from stat/device, not stat/health
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
s = d.get("speedtest-status")
if s: print(d.get("name"), s.get("xput_download"), s.get("xput_upload"),
"status_summary=", s.get("status_summary"))'
The result lands on the gateway's device object, under the hyphenated key
speedtest-status (not speedtest_status), with throughput in the nested
xput_download and xput_upload fields. status_summary tells you whether the
run finished; a fresh gateway that has never run one reports zeros. stat/health
is a different number entirely and will send you in circles.
Installs
1
2. Check the AP's uplink. An AP on a 100 Mbps link, or meshed rather than
wired, caps every client behind it:
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
up = d.get("uplink") or {}
print(d.get("name"), d.get("type"), up.get("speed"), up.get("type"))'
3. Check the client's own view. Signal, negotiated rate, satisfaction:
udm clients --json | python3 -c '
import json,sys
for c in json.load(sys.stdin):
print(c.get("hostname"), c.get("signal"), c.get("tx_rate"), c.get("satisfaction"))'
This rung is where the interesting case appears. A client at -57 dBm with an
866 Mbps PHY rate has an excellent link. If it measures 32 Mbps, the radio is
fine and something is contending. Signal is not throughput, and confusing the
two is why people replace working access points.
4. Rule out a bandwidth cap. Somebody, possibly you, may have set one:
udm raw GET /proxy/network/api/s/default/rest/usergroup
udm wlans --json # check per-SSID rate limits
5. Compare every AP's channel, width, and channel utilization. The last rung,
and the one that finds co-channel contention:
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
for r in d.get("radio_table_stats") or []:
print(d.get("name"), r.get("radio"), "ch", r.get("channel"),
"bw", r.get("bw"), "cu_total", r.get("cu_total"),
"sat", r.get("satisfaction"))'
Field names matter here and are not the obvious ones. Channel width is bw,
not channel_width. The names are per object, not global: on a client record
the same value really is channel_width. Check the object you are reading rather
than carrying a field name across from another endpoint. cu_total is channel utilization, the percentage of
airtime in use, and it is the closest thing to a direct measurement of the
problem this skill is about. cu_self_tx and cu_self_rx break out how much of
that is this AP's own traffic: high cu_total with low cu_self_* means
somebody else is using your airtime, which is either a neighbor or, more often,
your own AP two rooms away on the same channel.
satisfaction reports -1 when a radio has no clients on it. That is "no data",
not "terrible".
Read radio_table_stats, the operational state, not radio_table, the config.
They disagree more often than you would like.
The 80 MHz trap
The failure that motivated this skill, because it is a change that looks free.
Setting every 5 GHz radio to 80 MHz doubles the theoretical per-client rate, so
it reads like a pure upgrade. It is not, because width is bought with
channels. In the US regulatory domain, 80 MHz with DFS channels disabled leaves
exactly two non-overlapping blocks:
UNII-1: channels 36 / 40 / 44 / 48
UNII-3: channels 149 / 153 / 157 / 161
Two blocks. If you have more than two 5 GHz APs with overlapping coverage, some
of them are sharing an 80 MHz block and taking turns transmitting. Wi-Fi degrades
gracefully in signal terms and catastrophically in contention terms: everything
looks healthy right up until two clients are busy simultaneously.
The observed case: four APs, three of them in-wall units on adjacent walls, all
piled into UNII-1 after a global switch to 80 MHz. Result was roughly 32 Mbps
on a gigabit connection, with an 866 Mbps PHY link at -57 dBm and a confirmed
1276 Mbps gateway speedtest. Neither ISP, backhaul, nor signal was the cap.
Do not disable DFS channels when auditing a channel plan. It is the change
that creates the pileup. DFS is the difference between two usable 80 MHz blocks
and six.
Pick a block, not a channel
At 80 MHz a radio occupies its whole block, so two APs on different channels in
the same block are fully co-channel with each other. This is the mistake to
avoid when spreading APs out: moving an AP from 48 to 104 accomplishes nothing if
another AP already sits on 100.
US 80 MHz blocks:
Block
Channels
DFS?
UNII-1
36, 40, 44, 48
no
52, 56, 60, 64
yes
100, 104, 108, 112
yes
116, 120, 124, 128
yes, and see below
132, 136, 140, 144
yes
UNII-3
149, 153, 157, 161
no
Map what you have to blocks before choosing a target, because the channel numbers
hide the collision:
udm devices --json | python3 -c '
import json,sys
B=([36,40,44,48],[52,56,60,64],[100,104,108,112],[116,120,124,128],[132,136,140,144],[149,153,157,161])
blk=lambda c: next(("%d-%d"%(b[0],b[-1]) for b in B if c in b), "?")
for d in json.load(sys.stdin):
for s in d.get("radio_table_stats") or []:
if s.get("radio")=="na" and s.get("channel"):
print(d.get("name"), "ch", s.get("channel"), "-> block", blk(s["channel"]))'
Prefer 52-64, 100-112, and 132-144 when you need free blocks. The 116-128
block overlaps terminal doppler weather radar (TDWR, 5600-5650 MHz) in the US,
where a CAC can run ten minutes instead of one, and the channels are sometimes
unavailable outright. An 80 MHz block anchored at 116 spans 116/120/124/128 =
5570-5650 MHz, so it swallows the whole TDWR sub-band: at 80 MHz you cannot
sit near 116 without sitting in it.
If an AP keeps falling off a DFS channel, stop restarting it
An AP assigned a DFS channel can run there for a week or two, hit a radar
detection (real or false), quietly move to a non-DFS channel, and keep reporting
the assigned channel in its config the whole time. Nothing alarms, nothing logs
it, and rest/device echoes back whatever you last wrote. One deployment sat like
this for twelve days: config said 116, the radio was on 149, stacked on top of
two APs already sharing the 149-161 block, at 33% and 30% channel utilization
while every other AP sat at 1-2%. Every check had read the config.
Restarting puts it back and buys another week or two, which is treating the
symptom. If a radio keeps getting pushed off the same block, the block is wrong
for that site. Retire it and take an extra co-channel share instead. Moving
that AP to a non-DFS block took it from 33% to 7% utilization and its neighbour
from 21% to 12%. A share you chose beats a channel the AP cannot hold, so pair
whichever APs are furthest apart physically: different floors beats different
rooms.
Audit live-vs-config, not either one alone, on a schedule rather than only
after a write. A radio that drifts once will drift again.
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
cfg={r["radio"]:r for r in d.get("radio_table") or []}
for s in d.get("radio_table_stats") or []:
c=cfg.get(s["radio"],{}).get("channel")
if s.get("channel") and str(c)!=str(s["channel"]) and c!="auto":
print("DRIFT", d.get("name"), s["radio"], "cfg", c, "live", s["channel"])'
The tradeoff is real and worth stating honestly: DFS channels require radar
detection, cost a ~50-60 second silent scan when a radio moves onto one, and some
older or cheap client devices will not associate on them at all. Weigh that
against four APs sharing one block. In a dense deployment, 40 MHz on
non-overlapping channels beats 80 MHz on shared ones, every time.
Changing a radio channel safely
Read and write are different endpoints. This one costs people an hour:
GET /proxy/network/api/s/default/stat/device/<MAC> # the object
PUT /proxy/network/api/s/default/rest/device/<DEVICE_ID> # the write
GET rest/devicereturns an empty list on Network 10.4.57, with or without a
device id appended, exactly like rest/firewallrule in unifi-firewall. It is
not broken and your key is fine: reads live on stat/device. An agent that GETs
rest/device, sees [], and concludes the controller has no devices is on the
wrong endpoint.
The write body that works is the radio table alone, sent complete:
# GET the device, edit ONLY the target field inside radio_table,
# then PUT the WHOLE radio_table array back:
udm raw PUT /proxy/network/api/s/default/rest/device/<DEVICE_ID> \
'{"radio_table": [ <every radio, every field, one value changed> ]}'
Verified on Network 10.4.57: this preserves the radio you did not touch. You do
not need to send the entire device object back, and you should not, because
stat/device hands you a pile of read-only operational fields along with the
config.
Send the full radio_table array. A partial write drops the fields you
omitted, including transmit power and width on the other radio.
channel may read auto in config while radio_table_stats shows a real
number. Those are both correct: the AP is auto-selecting and has landed
somewhere. Writing a concrete channel silently opts that radio out of automatic
selection, which is usually what you want for a deliberate plan, but it is a
second change riding along with the first. Know you are making it.
Moving onto a DFS channel triggers a CAC (channel availability check) and the
radio goes silent for it. Two moves observed on Network 10.4.57 took roughly 45
and 105 seconds, so budget more than a minute and do not conclude the write
failed at the 60-second mark. Clients drop and reconnect once. Do it when nobody
is on a call.
The two directions are not equally hard. Moving off a DFS channel onto a
non-DFS one is easy: one observed move settled about 75 seconds after a
force-provision, no restart, clients reassociating on their own. Moving onto a
DFS channel is where it sticks, and a force-provision can leave the radio on its
old channel indefinitely; an AP restart (cmd/devmgrrestart, ~2.5 minutes
offline) is what finally lands it after CAC. If a non-DFS move needs a restart,
something else is wrong. Go look rather than power-cycling.
Verify from radio_table_stats, never the PUT echo, and read state rather
than just channel:
state
Meaning
RUN
on air and serving
DFS_WAIT
mid-CAC, silent, this is the normal DFS transition
INIT
not on air
A radio still reporting the old channel with state: RUN has not moved yet. One
in DFS_WAIT is moving and needs more time.
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
for s in d.get("radio_table_stats") or []:
print(d.get("name"), s.get("radio"), s.get("channel"),
s.get("state"), "clients", s.get("num_sta"))'
Roaming
Nothing here makes a client roam. The client decides, every time, using logic
its vendor did not document and you cannot change. Every setting below either
changes what the client sees or takes the current AP away from it. Read that again
before spending an evening on thresholds: this is lobbying, not commanding.
A sticky client is usually not a roaming problem. It is a cell-size problem. An AP
loud enough to be heard comfortably three rooms away gives a client that walks
that far a perfectly good signal and no reason to move. Fix coverage overlap
first. The persuasion knobs compensate for a cell plan, they do not replace one.
What the controller exposes, and where
Two different objects, which is the first thing that trips people up.
Every field name above was read off a live controller on Network 10.4.57.
There is no 802.11k neighbor-report toggle in either object on this version.
If a guide tells you to enable 11k in UniFi, confirm it is not describing a
different controller generation.
Order of operations
Work down. Each step is cheaper to undo than the one after it.
1. Size the cells with transmit power. The most effective fix and the least
clever one. In a dense deployment, turning power down improves roaming: a
smaller cell means the client runs out of signal while a better AP is already well
within range.
Set every AP covering the same area to the same mode. One AP left on auto
among manual neighbors resizes its own cell on its own schedule while everything
around it holds still, so the overlap you designed drifts without you touching
anything.
udm devices --json | python3 -c '
import json,sys
for d in json.load(sys.stdin):
for r in d.get("radio_table") or []:
print(d.get("name"), r.get("radio"), r.get("tx_power_mode"),
r.get("tx_power"), "range", r.get("min_txpower"), r.get("max_txpower"))'
Two things observed on real mixed hardware:
min_txpower and max_txpower differ per radio and per model. Ranges of
6..22, 6..26, and 4..20 were seen across three AP models on one network. A
single dBm value applied network-wide is not portable, and a value outside a
radio's range is not a setting.
tx_power does not track tx_power_mode. Radios reporting mode low were
observed carrying tx_power values of both 6 and 20. When the mode is a
named preset, read tx_power as a leftover, not as effective radiated power.
2. Raise the minimum data rate.minrate_na_enabled plus
minrate_na_data_rate_kbps sets a floor. Slow frames from a distant client consume
airtime far out of proportion to the data they carry, and a client that cannot meet
the floor has to go find an AP where it can. Gentler than a disconnect, and it
attacks the contention problem from earlier in this skill at the same time.
3. Enable 802.11r and 802.11v.fast_roaming_enabled shortens the
reassociation handshake; bss_transition lets the AP nominate a neighbor. Neither
forces anything, and 11v suggestions are routinely ignored. 802.11r has a long
history of breaking older clients, so turn it on for one SSID and watch what stops
connecting rather than flipping it everywhere at once.
4. Only then consider a minimum RSSI. Read the next section first.
The min RSSI trap
Minimum RSSI is not a handoff. It is a disconnect. The AP stops answering a
client below the floor, and what happens next is entirely the client's choice.
That includes immediately reassociating to the same AP, which is a disconnect loop
wearing a roam costume.
Two thresholds exist and they are not the same knob: min_rssi is per radio in
radio_table, roaming_assistant_na_rssi is per SSID in wlanconf. Both can be
active simultaneously. Which one wins when they disagree is not established
here, because settling it means writing to a live network and watching clients
drop. Read both before changing either.
min_rssi and min_rssi_enabled are independent, so read the flag, not the
number. A radio will happily report min_rssi: -79 with min_rssi_enabled: False, and that threshold is doing nothing. Observed on Network 10.4.57 with the
flag True on every 5 GHz radio and False on every 2.4 GHz radio of the same
APs. Auditing the values alone tells you a story that is half wrong.
Before changing a floor, find out who it will evict. A tighter number is only
free if nobody is living in the gap:
udm clients --json | python3 -c '
import json,sys
for c in json.load(sys.stdin):
s=c.get("signal") or 0
if c.get("radio")=="na" and s and s <= -75: print(c.get("hostname"), s)'
Set it too aggressive and you have built a coverage hole out of software. A client
evicted at -70 that sees nothing better than -72 has nowhere to go, and the
symptom is not poor roaming, it is a device that will not stay on the network at
all. Anything tighter than roughly -75 needs a specific reason.
Keep the floor consistent across APs covering the same space. Values drifting
between -70 and -79 on one network is the pathological case: the AP with the
tightest floor evicts clients that every neighbor around it is perfectly happy to
serve.
Checking whether any of it worked
The controller will not tell you. Client records carry roam_count, and
comparing it across a few days is a far better signal than reading it once:
udm clients --json | python3 -c '
import json,sys
for c in json.load(sys.stdin):
print(c.get("hostname"), c.get("ap_mac"), c.get("signal"),
c.get("tx_rate"), "roams", c.get("roam_count"))'
Run it, walk the building, run it again. ap_mac changing is a roam. signal
improving after it changed is a good roam. A roam_count that climbs while the
device sits still is thrash, and usually means a threshold is too tight.
Check fixed_ap_enabled and fixed_ap_mac before diagnosing anything. A
client pinned to a specific AP will not roam no matter what you tune, and the
setting is easy to apply once and forget:
udm clients --json | python3 -c '
import json,sys
for c in json.load(sys.stdin):
if c.get("fixed_ap_enabled"):
print("PINNED", c.get("hostname"), c.get("fixed_ap_mac"))'
Applying the change
wlanconf and radio_table are both full-object PUTs, so the discipline from
unifi-connect applies unchanged: GET the whole object, change one field, PUT the
whole thing back, re-read to confirm. Roaming work is the easiest place in this API
to lose a passphrase or an entire transmit-power plan to a partial write.
Whether a given roaming change needs a force-provision to take effect was not
established here. If a setting reads back correctly but behaves as though it never
applied, udm devices provision <MAC> is the next thing to try.
SSID hygiene
udm wlans --json | python3 -c '
import json,sys
for w in json.load(sys.stdin):
print(w.get("name"), w.get("enabled"), w.get("security"), w.get("networkconf_id"))'
Re-read enabled on every wlanconf touch. An SSID that was deliberately
disabled can come back: a controller update, a config restore, or somebody in
the UI. Deliberate disables do not always stay disabled, so make checking the
state part of the routine rather than assuming your last change held.
Never leave an open SSID on the trusted network. If a device genuinely
cannot do WPA2, and some old handhelds and IoT gear cannot, rebuild that SSID
on the guest or IoT network where an open join reaches nothing. An open SSID
bridged to the trusted LAN is an unauthenticated port on your network with a
radius of about 40 metres.
Hidden SSIDs are not security. They are broadcast in probe requests by every
client configured for them. They mainly make life harder for you.
Confirm each SSID's networkconf_id maps to the network you think it does.
This is how a device ends up on the trusted VLAN despite a careful segmentation
plan: it joined the SSID that was still pointed at the old network.
The in-wall AP port trap
Filed here because it is AP hardware behavior, though it bites during wired
segmentation.
AC-series in-wall APs do not enforce native-VLAN overrides on their pass-through
data ports. Set a port override on the AP, and the controller accepts it and
echoes native_networkconf_id back in port_table as though it applied. The
operational forward value stays all, and untagged traffic from the device
behind that port keeps bridging onto the AP's own network.
Verified with two force-provisions and a forward:'native' variant. The device
behind the port stayed reachable on the old subnet throughout, which is the tell:
the config says isolated, the network says otherwise.
Devices behind in-wall AP data ports cannot be VLAN-jailed. Not a
configuration problem, so do not burn an afternoon on payload variants. Re-cable
to a real switch port, or accept it and document the exception. Once you know
this, it takes ten seconds to check whether a stubborn device is behind one.
A disabled radio may not be disabled
Same hardware family, same shape of lie, and this one has an ongoing cost.
Two In-Wall HD APs on Network 10.4.57, both with a 2.4 GHz radio configured off
and parked on channel 11 at 20 MHz. What they were actually doing:
config (radio_table)
operational (radio_table_stats)
channel
11
1
width
ht 20
bw40
power
tx_power_mode: disabled
state: RUN, cu_self_tx 5-6%
clients
0
Every configured value ignored, and cu_self_tx of 5 to 6 percent means each one
was spending real airtime beaconing to nobody, on the channel where a third AP was
serving every 2.4 GHz device in the building.
num_sta: 0 does not mean a radio is idle. A radio with no clients still
beacons, and at 40 MHz from channel 1 it covers most of the band. Judge a radio by
state and cu_self_tx, never by client count.
Things that did not fix it, so you can skip them:
Force-provision, twice. The stored config was not applied either time.
"enabled": false on the radio entry. The PUT returns rc: ok and the
controller silently strips the field: it does not survive a read-back and it is
not part of the radio_table schema. A clean illustration of why rc: ok is
not evidence, and why a round-trip read is.
An AC In-Wall on the same site, with the same tx_power_mode: disabled, sits at
state: INIT with no channel and zero utilization, which is what genuinely off
looks like. Whether that radio is disabled or simply failed to initialize is not
distinguishable from the API, so treat it as a clue rather than a conclusion.
If you need a radio truly off on this hardware, do it in the UI and then diff.
GET the device's radio_table, toggle the radio in the web interface, GET again,
and compare. That tells you the field the controller actually writes. Do not
invent one: the guess above looked entirely plausible and was silently discarded.
