UniFi Channel Planning
Use this skill to produce a deterministic channel plan from MCP output.
Primary intent:
- Reduce overlap and co-channel contention.
- Keep assignments reproducible and explainable.
- Provide safe change rollout with validation.
Required MCP Inputs
| Tool |
Purpose |
list_sites |
Select target site |
list_devices_by_type |
Get managed AP inventory (type=uap) |
list_site_internal_ap_neighbors_v2 |
Build internal AP-to-AP graph for one site in one call |
Conditional / Optional MCP Inputs
| Tool |
When to use |
list_neighboring_aps |
Optional external RF tie-breaker or explanation; do not fetch by default |
list_ap_neighbors_v2 |
Per-AP drill-down, debugging, or partial-result validation |
get_device_details |
When radio configuration, current channel, width, or regulatory capabilities are not available from inventory |
Optional execution tool:
| Tool |
Purpose |
set_ap_radio_channel |
Apply proposed channel per AP/radio |
Data Contract For Planning
Expected fields from neighbor observations:
ap_mac: AP that observed the neighbor
bssid or mac: observed AP/BSSID identity
channel: observed neighbor channel number
signal: RSSI (dBm)
last_seen: freshness indicator
radio or band: required radio/band dimension (ng/2.4 GHz or na/5 GHz)
channel_width or ht: channel width, required when available for 5 GHz
If any field is missing:
- Drop records without
ap_mac, channel, or signal.
- Do not use records without a band/radio dimension for channel assignment.
- Keep records without width only for a conservative 20 MHz fallback and mark
the affected recommendation medium-confidence.
- Keep processing with partial data and log dropped count.
Correct Data Source Split (Critical)
Use two distinct datasets and do not mix them:
- Internal AP-to-AP graph (for channel assignment):
- Primary source:
list_site_internal_ap_neighbors_v2(site_id, start_ms, end_ms, min_rssi)
- The tool already executes v2 per-AP neighbor queries and keeps only rows where neighbor
mac is a managed AP MAC from the same site.
- Optional fallback/debug:
list_ap_neighbors_v2 per AP.
- This is the authoritative input for "which site AP sees which other site AP and how strong".
- External RF pressure (optional tie-breaker):
- Source:
list_neighboring_aps (stat/rogueap)
- Represents foreign BSSIDs around the site.
- Use only as additional congestion context, not as the internal AP graph.
Data Volume Controls
Keep the MCP interaction bounded and return summaries rather than raw payloads:
- Use a fixed planning window, normally 24 hours; shorten it when the graph or
response is too large.
- Pass
min_rssi during collection (-85 by default, -80 for dense sites)
instead of collecting all weak observations and filtering after output.
- Prefer
list_site_internal_ap_neighbors_v2 because it aggregates the
per-AP v2 calls server-side. Do not repeat list_ap_neighbors_v2 for every
AP unless debugging a specific AP or validating a partial result.
- Request AP inventory with an explicit
limit and offset when supported;
continue in pages until the managed AP set is complete.
- Call
list_neighboring_aps only when external RF pressure is needed for a
tie-breaker or explanation. It can contain hundreds of rows and must never
be dumped into the planning response by default.
- Normalize, deduplicate, score, and sort before presenting results. Return
metadata, counts, top conflict drivers, and a compact per-AP summary instead
of raw JSON.
- Cap displayed neighbors per AP to the strongest relevant observations and
report the total count plus omitted-row count. Offer per-AP drill-down only
on request.
- If a response is still too large, process APs or pages in deterministic
batches and merge the normalized results before optimization; never silently
truncate the graph.
Required collection metadata:
start_ms, end_ms, and min_rssi
- managed AP count and collected edge count
- dropped, skipped, paged, and omitted-row counts
- whether external rogue-AP data was collected
When physical AP locations are unavailable:
- Do not block planning on a floor plan, coordinates, or manually supplied
zones.
- Use the normalized internal v2 graph as the measurable overlap proxy:
observed AP pairs, RSSI, freshness, band, channel, and width.
- Do not claim that two APs are physically distant or in separate areas unless
that is reported by UniFi or supplied by the operator.
- Lower confidence for conclusions that depend on unobserved coverage or
client roaming between APs.
- Prefer no-op or small staged changes when the graph is sparse; the planner
may still produce a deterministic plan from the telemetry that is available.
Planning Rules
- Scope by band:
- 2.4 GHz candidates: channels 1-14 (practical non-overlap: 1, 6, 11)
- 5 GHz candidates: local policy set (for example 36/40/44/48 and DFS set if allowed).
- Plan each band/radio independently. Never score a 2.4 GHz observation against
a 5 GHz candidate or use one band to fill missing data in the other.
- Derive occupied channel ranges from primary channel and width. For 20 MHz,
use one channel block; for 40/80 MHz, include every overlapping 20 MHz block
in the configured channel range.
- Before scoring, convert every UniFi radio configuration into a canonical
occupied_20mhz_channels set.
- Do not infer occupied blocks solely from a numeric primary channel. Use the
UniFi-reported primary/control/center channel and width representation.
- If UniFi does not provide enough information to determine occupied 20 MHz
blocks unambiguously, mark the radio medium-confidence and use the
conservative supported interpretation.
- Filter weak observations:
- Default threshold: keep neighbors with
signal >= -85.
- For dense environments, use
>= -80.
- Build per-AP conflict score:
- For each AP and candidate channel, sum weighted conflicts from observed neighbors on same or overlapping channels.
- Stronger RSSI contributes higher penalty.
- Internal AP-to-AP penalties must be weighted higher than external rogue BSSID penalties.
- Deterministic assignment:
- Sort APs by descending neighbor pressure (most constrained first).
- Evaluate assignments against the complete internal AP graph, not each AP in
isolation. Re-score all graph edges after every proposed assignment so a
locally good choice cannot create a global co-channel cluster.
- Use deterministic greedy global optimization: assign the next most
constrained AP, then recompute the complete graph score after each assignment.
- Select the candidate with the lowest total graph penalty. Use descending
pressure and ascending channel order as deterministic tie-breakers.
- Tie-breaker order must be fixed (for example ascending channel number).
Optimization order is strict:
- Minimize the complete RF conflict score.
- Among equivalent or near-equivalent scores, minimize channel concentration
and prefer a balanced population such as
3/3/2 for eight APs on three
2.4 GHz channels.
- Minimize the number of changes.
- Use ascending channel order as the final deterministic tie-breaker.
Default score tolerance:
- Relative tolerance:
2%
- Absolute tolerance:
0.01
- A candidate is near-equivalent when
candidate_score <= best_score * 1.02
or candidate_score - best_score <= 0.01.
For a global 2.4 GHz assignment, score every internal edge using the proposed
channels of both APs:
- Same channel:
rssi_weight(signal)
- Adjacent 2.4 GHz channel:
0.5 * rssi_weight(signal)
- Non-overlapping channel:
0
For 5 GHz, score overlap between the occupied channel ranges. A shared 20 MHz
block receives the same base penalty, scaled by the fraction of overlapping
blocks. An 80 MHz AP therefore contributes more spectrum pressure than a 20
MHz AP when the ranges overlap.
Use deterministic RSSI weight classes instead of assuming a linear RF effect:
signal >= -55: weight 4.0
-65 <= signal < -55: weight 3.0
-75 <= signal < -65: weight 2.0
-85 <= signal < -75: weight 1.0
Multiply same-channel or overlap penalties by this weight. Keep the raw RSSI
in the output reason so the score remains explainable.
Before optimization, normalize the internal graph:
- Treat AP-A seeing AP-B and AP-B seeing AP-A as one undirected physical pair.
- Deduplicate repeated observations within the planning window.
- Use the most recent valid RSSI per undirected pair, retaining band and width
separately. A configurable freshness window may retain the strongest valid
observation as secondary context, but must not replace the current value.
- Never infer the observer's current channel from the
channel field: it is
the observed neighbor/BSSID channel only.
Do not sum an observer's neighbor rows against a candidate while leaving all
other APs at their current channels. That produces misleading results such as
moving several APs to channel 11 simply because each AP was evaluated alone.
- Guardrails:
- Avoid changing all APs at once.
- Keep at least one stable channel anchor per area.
- With enough APs to use the practical 2.4 GHz set, keep every candidate
channel represented unless measured RF data clearly justifies abandoning one.
- When topology and RF scores permit, prefer a balanced channel population. For
eight APs on three channels, prefer counts of
3/3/2 (in deterministic
channel order) over a 4/3/1 or more concentrated distribution.
- Treat a mathematically lower score that moves many APs or abandons a channel
as a candidate for staged review, not as an automatic rollout.
- Respect hardware/regulatory constraints reported by UniFi.
Example Scoring Heuristic
For each observed neighbor relation:
- Same channel penalty:
p_same = rssi_weight(signal)
- Adjacent channel penalty (2.4 GHz):
p_adj = 0.5 * rssi_weight(signal)
- Non-overlapping channel penalty:
0
Total candidate score:
score(ap, ch) = sum(penalties from all neighbors seen by ap on/near ch)
Lower score is better.
Output Contract
Return a plan table with:
ap_mac
current_channel
proposed_channel
band
confidence (high|medium|low)
reason (top 1-3 conflict drivers)
Also return summary metrics:
- Estimated co-channel conflict delta
- Number of APs changed
- APs skipped with reason
Safe Rollout Workflow
- Discovery:
- Run
list_sites and select site_id.
- Build managed AP set (
type=uap).
- Pull internal graph with
list_site_internal_ap_neighbors_v2 over a fixed window (default 24h).
- If needed, inspect problematic APs with
list_ap_neighbors_v2.
- Optionally pull
list_neighboring_aps(site_id, min_rssi=-85) for external context.
- Plan:
- Generate per-AP candidate ranking and proposed channels.
- Mark risky APs (sparse data, stale
last_seen, extreme noise).
- Stage:
- Apply to small batch first with
set_ap_radio_channel.
- Rollout defaults: first wave maximum 3 AP radios; subsequent waves maximum
25% of remaining changed radios; never change more than 50% of site radios
without explicit confirmation.
- Persist the original channel, width, and power settings for every change.
- Wait validation window and compare client health/events.
- If AP instability or client disconnects exceed the configured threshold,
restore the persisted settings before continuing.
- Expand:
- Roll out remaining APs in waves.
- Re-run neighbor scan and adjust.
Validation Checks
After each rollout wave, verify:
- AP online state is stable.
- Client disconnect spikes did not increase.
- Channel reuse distance improved in dense zones.
- No AP is assigned unsupported channel width/channel pair.
Pre-plan data-quality checks:
- Confirm each managed AP has at least one internal neighbor observation in the selected window, otherwise mark low-confidence.
- Treat
channel=auto as unknown current state. Infer it only when the v2
graph provides a fresh observation of that AP as a neighbour. Mark the
inferred current state and resulting recommendation medium-confidence. If
the AP is not observed as a neighbour, keep it unchanged and mark
low-confidence.
- Distinguish the observed neighbor's channel from the observer's channel:
the
channel field on a v2 edge describes the observed neighbor/BSSID.
- Read the observer's current channel from AP radio configuration or telemetry;
never infer it from a row where that AP is the observer. An AP configured as
Auto may be inferred from a fresh row where its MAC is the observed
neighbour, and may be changed to the proposed fixed channel when the plan is
explicitly accepted.
- Reject impossible RSSI outliers for planning (for example values above -20 dBm or below -95 dBm unless intentionally configured).
- If internal graph is empty, do not derive channel plan from
rogueap alone.
Confidence must be deterministic:
high: complete internal data, fresh observations, known capabilities, and
a clear score improvement or a validated no-op.
medium: partial or stale data, inferred Auto state, missing width, or a
marginal score difference.
low: no or very few observations, skipped APs, unknown capabilities, or an
unknown current channel.
Failure Handling
- If neighbor dataset is empty for a site, do not auto-assign.
- If an AP has no observations, keep current channel and mark
low confidence.
- If an AP is configured as
Auto and its current channel is inferred from a
fresh neighbour observation, preserve the original Auto setting for
rollback and allow the accepted plan to set an explicit fixed channel.
- On per-AP apply failure, continue with others and report partial status.
- If
list_site_internal_ap_neighbors_v2 reports skipped_aps, keep those APs unchanged and mark low confidence.
- If only
rogueap is available and internal AP graph is missing, return "insufficient internal data" instead of proposing deterministic AP-to-AP assignments.
- If the current configuration is already within the accepted score tolerance,
return a valid no-op plan with
0 APs changed; do not manufacture changes.
Example Prompt Pattern
"For site X, use the internal AP neighbor graph to produce a low-interference
channel plan per AP for 2.4 GHz first. Use channels 1/6/11 only, use minimum
RSSI -85, minimize same-channel conflicts, and return deterministic sorted
results with reasons. Fetch external rogue-AP data only if needed for context."
1---2name: unifi-channel-planning3description: >4 Build an AI-driven low-interference WiFi channel plan from UniFi MCP neighbor5 observations. Use when the task is RF optimization, channel conflict6 reduction, co-channel interference mitigation, or per-AP channel assignment.7metadata:8 source: https://skillmd.com/skills/enuno/unifi-channel-planning9---1011# UniFi Channel Planning1213Use this skill to produce a deterministic channel plan from MCP output.1415Primary intent:1617- Reduce overlap and co-channel contention.18- Keep assignments reproducible and explainable.19- Provide safe change rollout with validation.2021## Required MCP Inputs2223| Tool | Purpose |24|---|---|25| `list_sites` | Select target site |26| `list_devices_by_type` | Get managed AP inventory (`type=uap`) |27| `list_site_internal_ap_neighbors_v2` | Build internal AP-to-AP graph for one site in one call |2829## Conditional / Optional MCP Inputs3031| Tool | When to use |32|---|---|33| `list_neighboring_aps` | Optional external RF tie-breaker or explanation; do not fetch by default |34| `list_ap_neighbors_v2` | Per-AP drill-down, debugging, or partial-result validation |35| `get_device_details` | When radio configuration, current channel, width, or regulatory capabilities are not available from inventory |3637Optional execution tool:3839| Tool | Purpose |40|---|---|41| `set_ap_radio_channel` | Apply proposed channel per AP/radio |4243## Data Contract For Planning4445Expected fields from neighbor observations:4647- `ap_mac`: AP that observed the neighbor48- `bssid` or `mac`: observed AP/BSSID identity49- `channel`: observed neighbor channel number50- `signal`: RSSI (dBm)51- `last_seen`: freshness indicator52- `radio` or `band`: required radio/band dimension (`ng`/2.4 GHz or `na`/5 GHz)53- `channel_width` or `ht`: channel width, required when available for 5 GHz5455If any field is missing:5657- Drop records without `ap_mac`, `channel`, or `signal`.58- Do not use records without a band/radio dimension for channel assignment.59- Keep records without width only for a conservative 20 MHz fallback and mark60 the affected recommendation medium-confidence.61- Keep processing with partial data and log dropped count.6263## Correct Data Source Split (Critical)6465Use two distinct datasets and do not mix them:66671. Internal AP-to-AP graph (for channel assignment):6869- Primary source: `list_site_internal_ap_neighbors_v2(site_id, start_ms, end_ms, min_rssi)`70- The tool already executes v2 per-AP neighbor queries and keeps only rows where neighbor `mac` is a managed AP MAC from the same site.71- Optional fallback/debug: `list_ap_neighbors_v2` per AP.72- This is the authoritative input for "which site AP sees which other site AP and how strong".73742. External RF pressure (optional tie-breaker):7576- Source: `list_neighboring_aps` (`stat/rogueap`)77- Represents foreign BSSIDs around the site.78- Use only as additional congestion context, not as the internal AP graph.7980## Data Volume Controls8182Keep the MCP interaction bounded and return summaries rather than raw payloads:8384- Use a fixed planning window, normally 24 hours; shorten it when the graph or85 response is too large.86- Pass `min_rssi` during collection (`-85` by default, `-80` for dense sites)87 instead of collecting all weak observations and filtering after output.88- Prefer `list_site_internal_ap_neighbors_v2` because it aggregates the89 per-AP v2 calls server-side. Do not repeat `list_ap_neighbors_v2` for every90 AP unless debugging a specific AP or validating a partial result.91- Request AP inventory with an explicit `limit` and `offset` when supported;92 continue in pages until the managed AP set is complete.93- Call `list_neighboring_aps` only when external RF pressure is needed for a94 tie-breaker or explanation. It can contain hundreds of rows and must never95 be dumped into the planning response by default.96- Normalize, deduplicate, score, and sort before presenting results. Return97 metadata, counts, top conflict drivers, and a compact per-AP summary instead98 of raw JSON.99- Cap displayed neighbors per AP to the strongest relevant observations and100 report the total count plus omitted-row count. Offer per-AP drill-down only101 on request.102- If a response is still too large, process APs or pages in deterministic103 batches and merge the normalized results before optimization; never silently104 truncate the graph.105106Required collection metadata:107108- `start_ms`, `end_ms`, and `min_rssi`109- managed AP count and collected edge count110- dropped, skipped, paged, and omitted-row counts111- whether external rogue-AP data was collected112113When physical AP locations are unavailable:114115- Do not block planning on a floor plan, coordinates, or manually supplied116 zones.117- Use the normalized internal v2 graph as the measurable overlap proxy:118 observed AP pairs, RSSI, freshness, band, channel, and width.119- Do not claim that two APs are physically distant or in separate areas unless120 that is reported by UniFi or supplied by the operator.121- Lower confidence for conclusions that depend on unobserved coverage or122 client roaming between APs.123- Prefer no-op or small staged changes when the graph is sparse; the planner124 may still produce a deterministic plan from the telemetry that is available.125126## Planning Rules1271281. Scope by band:129130- 2.4 GHz candidates: channels 1-14 (practical non-overlap: 1, 6, 11)131- 5 GHz candidates: local policy set (for example 36/40/44/48 and DFS set if allowed).132- Plan each band/radio independently. Never score a 2.4 GHz observation against133 a 5 GHz candidate or use one band to fill missing data in the other.134- Derive occupied channel ranges from primary channel and width. For 20 MHz,135 use one channel block; for 40/80 MHz, include every overlapping 20 MHz block136 in the configured channel range.137- Before scoring, convert every UniFi radio configuration into a canonical138 `occupied_20mhz_channels` set.139- Do not infer occupied blocks solely from a numeric primary channel. Use the140 UniFi-reported primary/control/center channel and width representation.141- If UniFi does not provide enough information to determine occupied 20 MHz142 blocks unambiguously, mark the radio medium-confidence and use the143 conservative supported interpretation.1441452. Filter weak observations:146147- Default threshold: keep neighbors with `signal >= -85`.148- For dense environments, use `>= -80`.1491503. Build per-AP conflict score:151152- For each AP and candidate channel, sum weighted conflicts from observed neighbors on same or overlapping channels.153- Stronger RSSI contributes higher penalty.154- Internal AP-to-AP penalties must be weighted higher than external rogue BSSID penalties.1551564. Deterministic assignment:157158- Sort APs by descending neighbor pressure (most constrained first).159- Evaluate assignments against the complete internal AP graph, not each AP in160 isolation. Re-score all graph edges after every proposed assignment so a161 locally good choice cannot create a global co-channel cluster.162- Use deterministic greedy global optimization: assign the next most163 constrained AP, then recompute the complete graph score after each assignment.164- Select the candidate with the lowest total graph penalty. Use descending165 pressure and ascending channel order as deterministic tie-breakers.166- Tie-breaker order must be fixed (for example ascending channel number).167168Optimization order is strict:1691701. Minimize the complete RF conflict score.1712. Among equivalent or near-equivalent scores, minimize channel concentration172 and prefer a balanced population such as `3/3/2` for eight APs on three173 2.4 GHz channels.1743. Minimize the number of changes.1754. Use ascending channel order as the final deterministic tie-breaker.176177Default score tolerance:178179- Relative tolerance: `2%`180- Absolute tolerance: `0.01`181- A candidate is near-equivalent when `candidate_score <= best_score * 1.02`182 or `candidate_score - best_score <= 0.01`.183184For a global 2.4 GHz assignment, score every internal edge using the proposed185channels of both APs:186187- Same channel: `rssi_weight(signal)`188- Adjacent 2.4 GHz channel: `0.5 * rssi_weight(signal)`189- Non-overlapping channel: `0`190191For 5 GHz, score overlap between the occupied channel ranges. A shared 20 MHz192block receives the same base penalty, scaled by the fraction of overlapping193blocks. An 80 MHz AP therefore contributes more spectrum pressure than a 20194MHz AP when the ranges overlap.195196Use deterministic RSSI weight classes instead of assuming a linear RF effect:197198- `signal >= -55`: weight `4.0`199- `-65 <= signal < -55`: weight `3.0`200- `-75 <= signal < -65`: weight `2.0`201- `-85 <= signal < -75`: weight `1.0`202203Multiply same-channel or overlap penalties by this weight. Keep the raw RSSI204in the output reason so the score remains explainable.205206Before optimization, normalize the internal graph:207208- Treat AP-A seeing AP-B and AP-B seeing AP-A as one undirected physical pair.209- Deduplicate repeated observations within the planning window.210- Use the most recent valid RSSI per undirected pair, retaining band and width211 separately. A configurable freshness window may retain the strongest valid212 observation as secondary context, but must not replace the current value.213- Never infer the observer's current channel from the `channel` field: it is214 the observed neighbor/BSSID channel only.215216Do not sum an observer's neighbor rows against a candidate while leaving all217other APs at their current channels. That produces misleading results such as218moving several APs to channel 11 simply because each AP was evaluated alone.2192205. Guardrails:221222- Avoid changing all APs at once.223- Keep at least one stable channel anchor per area.224- With enough APs to use the practical 2.4 GHz set, keep every candidate225 channel represented unless measured RF data clearly justifies abandoning one.226- When topology and RF scores permit, prefer a balanced channel population. For227 eight APs on three channels, prefer counts of `3/3/2` (in deterministic228 channel order) over a `4/3/1` or more concentrated distribution.229- Treat a mathematically lower score that moves many APs or abandons a channel230 as a candidate for staged review, not as an automatic rollout.231- Respect hardware/regulatory constraints reported by UniFi.232233## Example Scoring Heuristic234235For each observed neighbor relation:236237- Same channel penalty: `p_same = rssi_weight(signal)`238- Adjacent channel penalty (2.4 GHz): `p_adj = 0.5 * rssi_weight(signal)`239- Non-overlapping channel penalty: `0`240241Total candidate score:242243- `score(ap, ch) = sum(penalties from all neighbors seen by ap on/near ch)`244245Lower score is better.246247## Output Contract248249Return a plan table with:250251- `ap_mac`252- `current_channel`253- `proposed_channel`254- `band`255- `confidence` (`high|medium|low`)256- `reason` (top 1-3 conflict drivers)257258Also return summary metrics:259260- Estimated co-channel conflict delta261- Number of APs changed262- APs skipped with reason263264## Safe Rollout Workflow2652661. Discovery:267268- Run `list_sites` and select `site_id`.269- Build managed AP set (`type=uap`).270- Pull internal graph with `list_site_internal_ap_neighbors_v2` over a fixed window (default 24h).271- If needed, inspect problematic APs with `list_ap_neighbors_v2`.272- Optionally pull `list_neighboring_aps(site_id, min_rssi=-85)` for external context.2732742. Plan:275276- Generate per-AP candidate ranking and proposed channels.277- Mark risky APs (sparse data, stale `last_seen`, extreme noise).2782793. Stage:280281- Apply to small batch first with `set_ap_radio_channel`.282- Rollout defaults: first wave maximum 3 AP radios; subsequent waves maximum283 25% of remaining changed radios; never change more than 50% of site radios284 without explicit confirmation.285- Persist the original channel, width, and power settings for every change.286- Wait validation window and compare client health/events.287- If AP instability or client disconnects exceed the configured threshold,288 restore the persisted settings before continuing.2892904. Expand:291292- Roll out remaining APs in waves.293- Re-run neighbor scan and adjust.294295## Validation Checks296297After each rollout wave, verify:298299- AP online state is stable.300- Client disconnect spikes did not increase.301- Channel reuse distance improved in dense zones.302- No AP is assigned unsupported channel width/channel pair.303304Pre-plan data-quality checks:305306- Confirm each managed AP has at least one internal neighbor observation in the selected window, otherwise mark low-confidence.307- Treat `channel=auto` as unknown current state. Infer it only when the v2308 graph provides a fresh observation of that AP as a neighbour. Mark the309 inferred current state and resulting recommendation medium-confidence. If310 the AP is not observed as a neighbour, keep it unchanged and mark311 low-confidence.312- Distinguish the observed neighbor's channel from the observer's channel:313 the `channel` field on a v2 edge describes the observed neighbor/BSSID.314- Read the observer's current channel from AP radio configuration or telemetry;315 never infer it from a row where that AP is the observer. An AP configured as316 `Auto` may be inferred from a fresh row where its MAC is the observed317 neighbour, and may be changed to the proposed fixed channel when the plan is318 explicitly accepted.319- Reject impossible RSSI outliers for planning (for example values above -20 dBm or below -95 dBm unless intentionally configured).320- If internal graph is empty, do not derive channel plan from `rogueap` alone.321322Confidence must be deterministic:323324- `high`: complete internal data, fresh observations, known capabilities, and325 a clear score improvement or a validated no-op.326- `medium`: partial or stale data, inferred `Auto` state, missing width, or a327 marginal score difference.328- `low`: no or very few observations, skipped APs, unknown capabilities, or an329 unknown current channel.330331## Failure Handling332333- If neighbor dataset is empty for a site, do not auto-assign.334- If an AP has no observations, keep current channel and mark `low` confidence.335- If an AP is configured as `Auto` and its current channel is inferred from a336 fresh neighbour observation, preserve the original `Auto` setting for337 rollback and allow the accepted plan to set an explicit fixed channel.338- On per-AP apply failure, continue with others and report partial status.339- If `list_site_internal_ap_neighbors_v2` reports `skipped_aps`, keep those APs unchanged and mark low confidence.340- If only `rogueap` is available and internal AP graph is missing, return "insufficient internal data" instead of proposing deterministic AP-to-AP assignments.341- If the current configuration is already within the accepted score tolerance,342 return a valid no-op plan with `0 APs changed`; do not manufacture changes.343344## Example Prompt Pattern345346"For site X, use the internal AP neighbor graph to produce a low-interference347channel plan per AP for 2.4 GHz first. Use channels 1/6/11 only, use minimum348RSSI -85, minimize same-channel conflicts, and return deterministic sorted349results with reasons. Fetch external rogue-AP data only if needed for context."