The narrowing walked each layer twice -- once for _layer_head to read its
name, once to decode it -- which is free when there is a lot to skip and a
straight loss when there isn't. Measured against a cache spanning zooms
2-14: low-zoom tiles carry no buildings, housenumbers or POIs at all
(a z10 tile is 98% drawn layers, a z2 tile 100%), so the whitelist saved
them nothing while the extra walk cost 16-18%.
The name is field 1 and in practice the first thing written, so the walk
already knows by then whether the rest is worth keeping; from there an
unwanted layer is stepped over rather than copied out. Field order isn't
guaranteed though, so the flag is only an optimisation -- the decision is
the check after the loop, which is right however the fields arrived and
also covers a layer declaring no name at all. _layer_head goes away.
On the regular platform this turns a 0.97x into a 1.03x on low-zoom
tiles and 1.00x on the largest, while urban z14 tiles keep their 3.19x;
the whole 293-tile cache goes 1.4x -> 1.6x. On the Lyra the median tile
is 3.24x and the largest 4.85x, both up.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The layer narrowing cut what a tile costs to parse; this stops us paying
it twice. Decoding is a pure function of the bytes and the layer set, but
the cache only held the bytes, so every session re-parsed the same tiles
from scratch -- and parsing is the single most expensive thing this app
does.
A decoded/ sidecar now sits beside the raw tiles holding the parsed
layers, via marshal: stdlib, C-speed, and it understands the plain
tuples and scalars a decoded tile reduces to. Deliberately not pickle --
this reads a file off disk, and marshal cannot be talked into importing a
module or calling a constructor.
The sidecar is derived data throughout, and the design leans on that. It
is never what makes a tile "known", so losing it costs only the decode it
was saving. Every way it can fail to be what this build expects -- bumped
wire version, a Python whose marshal differs, a truncated file -- reads
as a cache miss rather than an error. It carries a stamp of the layer set
it was decoded under, so widening DRAWN_LAYERS re-decodes instead of
serving a narrower blob. And because it runs ~2.4x the size of the bytes
it came from, it is the half of the cache that gets a budget: 64 MB, LRU
by mtime, swept once per 8 MB written rather than per tile, since the
walk is slow on the SD card the PicoCalc runs from.
Measured on the device over the 50 real cached tiles: a first session
costs 126 s and a later one 13 s, 9x, with all 50 tiles restoring
identically. Disk goes 7.8 MB -> 7.8 + 19.0 MB. Serialising adds ~13% to
the first decode.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Profiling on the Lyra found vector-tile decode dominating everything else
in the app by about two orders of magnitude: a median cached tile took
1236 ms of pure CPU, the largest 3.4 s, against 16 ms of SD read for a
whole viewport. Decoded tiles are memoised for the session, so this is a
first-pan cost -- but it runs on a GIL-bound worker thread, so it starves
the render loop for those seconds rather than running beside it.
Roughly 40% of that work was thrown away. A planet tile ships building,
housenumber, poi and mountain_peak; _draw_tile looks up ten layer names
and none of them is those. So decode_tile now takes the set of layers the
caller actually wants, reads each layer's identity with a cheap header
walk, and decodes the features only for the ones asked for. Unwanted
layers are still returned, named and empty -- which matters, because the
tile source decides whether a cache entry is real by whether the bytes
parsed into layers, and a tile whose every layer we skip must not read as
blank and get its perfectly good cache entry pruned and re-downloaded
every session. That test moves off "has features" and onto "parsed into
layers", which is sound because the other two ways a tile isn't real --
empty and corrupt -- decode to nothing and raise respectively.
Only the decode narrows; whole tiles are still cached, so drawing more
later costs a re-decode and never a re-download.
Measured on the device across the 46 real cached tiles: 183 s -> 51 s,
3.6x, with the drawn layers byte-identical on every one. The median tile
is 1259 -> 353 ms, the largest 3545 -> 759 ms.
_Reader.skip stopped slicing out the bytes it steps over while it was
here, which is most of what makes the header walk cheap.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A platform's ceiling may change a rule's resolution, never the rule. P3
had picocalc colour node names by *type* instead of by key, so a name's
colour there meant something different than it does on the desktop — the
node's kind, not its identity. Both platforms now run one rule: the key's
first byte picks the hue. The console just spends fewer of them.
node_style gains the binding the heat scale already had. Regular keeps
the full 256-hue wheel; picocalc snaps the same hue to its sixth of it —
the six chromatic bright slots, ~43 first bytes each. It returns the
slot's own hex rather than a color(N), so every downsample on the way out
(Rich's, and the fold's quantizer) lands on that exact slot instead of
guessing a neighbour, and the callers that read an RGB back out of a
style string keep working.
The type marks needed the opposite treatment: their hues are fixed, not
derived, and a naive nearest-RGB match drained the repeater's violet to
grey — a repeater read as an unknown node. They move to the theme's
type.* entries so each platform picks its slot deliberately, and a
braille raster resolves the same entry through theme.mark_rgb, so the
canvas marker and its Rich-drawn twin can't drift.
With names off the type registry, nothing reads it: Platform.name_colour
and core/nodetypes.py go, along with the three reception-layer feeders.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every repaint re-rasterized every row and re-ran every layout; every
screen open re-scanned the whole history. Now work is done once per
change instead of once per frame:
Render pipeline — a content-keyed ANSI cache in render_to_ansi (every
Text/str row app-wide), memoized corner glow, and panel-wrap memos for
the base frame and stacked dialogs, all dropped on a platform switch or
emoji recalibration. Idle-frame compose falls 5-100x per screen
(contacts 5.6->0.1 ms, record card 8.5->0.07 ms, trace 7.6->0.6 ms).
Repository — an observed_at index turns every since-window query into a
range walk (recent_observations 7.5->1 ms); heard_nodes and first_seen
aggregate in one streaming pass by ISO-string comparison instead of
hydrating dataclasses and sorting; node_names groups in SQL (NOT
INDEXED past a slow planner choice).
Per packet — livefeed no longer repaints on events that drew nothing,
the channel list's stats TTL rises above the tick so the repository is
not queried per frame, watch-store alerts batch through the throttled
flush instead of a full-state disk write per new node, and the header's
360-bucket histogram/flags tuples are memoized per minute and packet.
Per open — the resolvers, topology canonical()/display_name(), and
Yen's route search memoize over their frozen inputs; timemachine, chat,
and livefeed drop duplicate queries; the config editor, device actions,
info, advert, and tx-optimize reuse the devstate session cache instead
of re-asking the radio.
Per frame, per screen — dashboard folds five window scans into one
digest; walk caches everything derived from its frozen graph (the
onward-count quadratic included); packet viewer, record dialog, trophy
rows, trace stats/log, and the node-detail route fan memoize on their
frozen inputs, invalidated by revision counters where the input can
move.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The first P3 (a84abf3) ran on the wrong model and its palette never
reached the TUI: MESH_THEME_16 was written in 256-colour cube indices
instead of the 16 VT slots, and the platform theme pick was wired into
the scripted-CLI console while the real rasterizer stayed hardcoded to
truecolor. Replace it end to end:
- theme._VT_SLOTS: the canonical slot->RGB design (conventional hue
families kept; slots 5/6 become the track/faint slates), MESH_THEME_16
rewritten against color(0..15) with the VT's bold-as-brightness and
dim-bank-background rules pinned by tests. /etc/vtrgb + a boot oneshot
install it (console-font script), OSC fallback when kbd is absent.
- platforms.on_platform: set_platform now re-runs registered bindings,
so theme, rasterizer and widgets swap implementations once per switch
and hot paths read module globals - no per-call get_platform().
- The render-boundary fold (theme.fold_text): one str.translate pass
(NFKD accent folds - the Terminus base has no accented Latin - plus
symbol stand-ins and the width-padded emoji map), an embedded-SGR
quantizer for canvas truecolor, and a ?-at-own-width catch-all that
makes the no-wide-glyph short-circuit true by construction. Applied at
the only two places every visible byte passes: render_to_ansi and
MapCanvas.to_ansi_lines.
- ui/fontset.py freezes the installed font's inventory (dumped live);
the gallery now asserts rendered picocalc ANSI carries no truecolor/
256-colour SGR and no character outside it - the parity gate the
first attempt faked by asserting a dataclass field.
- Font build: nine new 6x12 marks (ellipsis, warn, backspace, shift,
gear, refresh, clock, crosshair, padlock) drawn into donor slots, with
donor misses now fatal and a keeper canary (donating pi once erased
Cyrillic pe - shared slot). Table re-dumped on-device: identical to
fontset.py, 519 = 519.
- Type-based name colouring: registry moved to core/nodetypes (services
no longer import ui), fed from monitor observations, heard_nodes and
merge_contacts. Heat gradient quantizes to heat.* steps on picocalc.
- ui/marks.py: the P2-deferred constant relocation - widgets no longer
drags map_render/mapcanvas/pathgraph for tuples and aliases.
- meshterm specimen: the visual language on one card through the real
funnels - the on-device acceptance screen (JP's request).
- tools/chat: delivered/no-ack state uses the standard check/query
marks, not the banned emoji pair.
Verified on the Lyra: font+palette installed, specimen framebuffer
captures in .claude/plans/. 1259 tests green, 13 xfailed (P6 list).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PicoCalc plan phase P2, driven by measurements taken on the Lyra as a
paired A/B (git archive HEAD vs the working tree, run side by side
against the same venv).
The headline is the database. Every packet the monitor hears is one
durable commit, and on the SD card each was costing 33.78 ms of fsync.
db._tune() opens in WAL with temp_store=MEMORY: 0.59 ms per commit, a
54x cut. synchronous=NORMAL is requested only once WAL is confirmed
engaged - the pragma returns the mode it settled on rather than raising
when a filesystem refuses it - because under WAL it risks the last
transactions to a power cut, and under the rollback journal it risks the
file itself. This device runs on a battery pack.
Startup: bleak/dbus-fast were already lazy; the real finds were two
cheap *values* dragging heavy subtrees. core/connection.py imports one
constant from core/channels.py, which imported PyCryptodome at module
level (224 ms of CPU-feature probing per run) - now deferred into the
two functions that decrypt. And tools/map.py's register_cli, which runs
for every tool at boot, reached into ui/map_screen.py for a --fraction
default, dragging the basemap's urllib.request -> http.client -> ssl
behind it; DEFAULT_VIEW_FRACTION moves to core/geo.py (pure arithmetic,
already on the boot path) and is re-exported so existing importers are
untouched. 26 modules and 316 ms of import self-time gone.
Per-frame: the corner glow only ever recolours truecolor foregrounds, so
on a 16-slot console it scanned every line of every frame to hand back
the identical list - skipped outright where effects are off, along with
the battery blink and the braille spinner cycle. _has_wide_glyph answers
from the platform when no emoji can be drawn instead of scanning the
composed frame (6.5-9.1 ms per repaint on its own). The header built its
segments twice to measure the compact separator, on exactly the narrow
terminals that take that branch every frame; it now builds once and
measures arithmetically. Together: 101.2 -> 73.2 ms per frame at 53x26,
151.8 -> 110.4 at 53x40, and 72x24 unregressed at 91.7.
That finally gave numbers to set the clocks against. refresh_interval
becomes per-platform tick_s (2.0 s on picocalc: a 1 Hz idle tick alone
was 7-11% of a core doing nothing, and only the header's per-minute
pulse drifts, since screens push their own repaints via invalidate).
The 0.12 s spinner was untenable - 84% of a core at 53x26, and at 53x40
it did not fit in its own interval at all. Five duplicate
_SPINNER_INTERVAL = 0.12 constants across the UI collapse into one
platform-aware spinner_interval() beside the Spinner it paces.
The header refactor was verified equivalent over 540 renders, plain text
and style spans both. The span check earned its keep: the first cut
passed style= to the Text constructor, which makes it the object's base
style and blankets everything appended after, so the version string
would have rendered brand-under-muted. test_header.py only asserted on
.plain and never saw it; it now pins the styling too.
tests/test_performance.py pins each contract as behaviour rather than
timing, including a subprocess guard that fails if Crypto, the network
stack, or the map screen ever creep back onto the startup import graph.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A soldered SoC UART like the Luckfox Lyra's /dev/ttyS1 is never enumerated by
pyserial's comports(), so a serial [profiles.*] entry pointing at one could only
ever be reached via `meshterm --port` on the command line -- it never surfaced in
the startup picker, which listed only scanned devices (and TCP profiles). Fold
configured serial profiles into the picker too, mirroring the existing TCP-profile
folding, de-duplicated by port so a port pyserial *does* enumerate keeps its richer
scanned row.
Also fix the liveness poll for the same ports: serial_port_present() now treats an
existing /dev character device as present (platform UARTs aren't in comports() but
their node persists exactly as long as the port does; a USB node vanishes on unplug,
so this never masks a real disconnect). Without it the watcher would falsely report
a soldered radio unplugged mid-session.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A contact's last_advert is stamped by the advertising node's own clock;
a mis-set one lands in the future, where the age clamp read it as a
permanent "now" and the heard sort pinned it to the top — for as long
as the bogus stamp stayed ahead of the wall clock, since the firmware
only ever advances the field (its replay filter).
advert_time (models.py) is now THE converter for every last_advert
epoch — the live contact table and the cross-session store both pass
through it — and refuses a stamp more than 5 minutes ahead as unknown.
The devstate contacts fetch then fills the gap from our own recorded
receptions (stamped by our clock), so a refused contact we really do
hear shows an honest age, and one we never heard reads never.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The chat picker's no-contacts line and the CLI's no-serial-devices line
were the last two wearing parentheses; the standard shape is bare prose
with an em-dash explanation (housekeeping H-005).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The NODE lane suited the two classes that name a node and left every
other one reading a dash. On this mesh that is half the traffic: the
direct messages, requests, responses, returned paths, acks and traces
that relay past us naming no origin. So the lane asks each class its own
question instead, and the header says SUBJECT — the word that stays true
down the whole column.
An advert or a telemetry frame is still about its node. A channel text is
about its channel, named by the key whose MAC confirms the frame — never
by the fingerprint alone, which 256 buckets make no proof at all; a
channel we hold no key for reads as that fingerprint and says so. A
direct message, request, response, returned path or anonymous request is
about the two nodes it travels between, drawn sender → recipient on the
hop-sequence widget, so endpoints read exactly as route hops do: resolved
names, key-derived hues, our own node white the moment a frame is
addressed to us. Too wide, and the sender drops to its hash — the frame
is addressed, so the addressee is the half worth the cells. An ack is
about the message it answers, a trace about its tag. And an unsigned
channel post is filed under its channel rather than the bare word
"channel", which named nothing.
Half of that was unreadable before because nobody had opened the frame.
The companion hands every overheard packet to the library, which decodes
the body for exactly two classes — advert and channel text — and leaves
the rest as bytes. Those bytes are not opaque: MeshCore addresses a frame
in the first few of them, and core/frames.py now reads them, class by
class, against the layouts the protocol's own docs/payloads.md publishes.
The trace tag is the little-endian uint32 the firmware writes, so it
reads as the same number a trace reply reports; an ack's checksum keeps
wire order, the form the companion reports its own delivery acks in. Two
checks guard every decode, because a wrong slice yields not an error but
a plausible-looking hash: the frame's payload version must be v1 (the
header reserves two bits for a v2 that widens both hash and MAC), and the
body must be long enough for the layout its class promises. Either
failing yields nothing at all.
The feed opens seeded from history, so any of this that lived only in a
live event's raw payload would leave the opening screen blank — the same
problem the channel-text crypto trio had. Same answer: three columns
(dest/src/tag), written on capture and restored to the very keys the
event used, so a replayed frame and a fresh one read identically all the
way up. Older rows carry NULL until re-heard.
The viewer gained the matching rows: to, from, tag, acks, and a channel
for a datagram — whose body is not text, but whose MAC still names it.
Verified against the MeshCore protocol documentation and firmware source,
then corroborated on the air: every frame captured off the radio has a
body length that decomposes exactly into its documented layout, an
anonymous request's 51 bytes (1 + 32 + 2 + one AES block) included.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The sweep table title used ASCII ->, TraceResult.path_str was an unused
ASCII-arrow renderer, and the trophy browser truncated long walks at the
fold — SelectScreen's own hscroll now slides the highlighted row.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add a "🗑 Purge stale contacts…" action at the foot of the Contacts list. It
offers an age ladder (1 week → 1 year, plus a never-heard bucket), each rung
showing a live count of how many contacts it would sweep, so the impact is
visible before choosing; the pick is gated behind a typed "delete" confirm.
A purge deletes each contact from the device table (new Device.remove_contact,
over the meshcore remove_contact command) and forgets it from the cross-session
contact store so a firmware-less bridge doesn't merge it back — but only the
contact is dropped: each node's reception history in MeshTerm is untouched, and
our own pinned node is never a candidate. The contacts cache is invalidated and
the list rebuilt so swept contacts are gone from it.
ContactListScreen gains a `tail` param (symmetric to `lead`) for the action and
Back rows; age bucketing (stale_past / never_heard) is factored out pure and
tested, alongside MockDevice.remove_contact and the screen's tail composition.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A firmware-less radio bridge holds its contact table only in RAM, so the Contacts
screen and the chat recipient list come up empty (or only as sparse as this
session's adverts) after the bridge restarts, even though you were messaging those
nodes last time. Record every contact MeshTerm reads, keyed by the device's own
public key, and union any the device is no longer reporting back into the list at
the one read chokepoint both lists share (DeviceState.contacts()).
No device write is needed: a direct message addresses a node by a prefix of its
public key, which the remembered contact carries, so a restored contact is
messageable without being written back onto the device. The union is inert where
it isn't needed — a firmware radio reports its whole table, so nothing is added,
no device-type check required — and recording never removes a contact on an empty
read, so a bridge that just restarted doesn't wipe what it knew.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A firmware-less radio bridge holds its configuration only in RAM, so the node
name, radio parameters, and tuning set through MeshTerm are back to firmware
defaults when the bridge process restarts. Record every setting changed through
MeshTerm (editor, config CLI, and TOML restore all funnel through _apply_setting),
keyed by the device's own public key.
Unlike channels — silently replayed because filling an empty slot overwrites
nothing — a setting is a single canonical value, so replaying a stale one could
clobber a change made elsewhere. So the store never writes on its own: on connect
it detects drift between what it remembers and what the device reports, and a
startup offer lets the operator restore the saved values, adopt the device's
current ones, or stop remembering the device (the "don't ask again" exit). The
offer is quiet unless a connected device actually drifts, so a firmware radio you
don't manage through MeshTerm is never interrupted.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A firmware-less radio bridge holds its channels only in RAM, so every one is
lost when the bridge restarts — the Channels page came up empty each session
even after you added channels. MeshTerm read channels live from the device and
never remembered them.
Add a durable, per-device channel store (<config_dir>/channels.json, keyed by
the device's public key) that records every channel written through MeshTerm.
On connect, reconcile() replays any remembered channel the device isn't
reporting into a free slot — filling gaps, never overwriting an occupied slot.
The gate is provenance, not device type: only channels you wrote through
MeshTerm are recorded, so a firmware radio you never edit through MeshTerm keeps
an empty record and does a single identity probe on connect (untouched). A
bridge, where every channel is necessarily created through MeshTerm, gets its
whole set back. Every mutation (create/join/import/edit/reorder/clear) now flows
through a single write_channel() boundary that keeps the store faithful.
Verified end to end against the live SPI bridge: write a channel, clear it off
the device to mimic a restart, reconnect in a fresh session, and reconcile
restores it with the right key.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pressing Enter on a contact now opens its whole story on one page instead of
a bare row: identity (name in its hue, key with the routing hash lit, type,
first/last heard, packet tally, reception SNR/RSSI), an inline location
preview centred on the node (a small static basemap via the new MiniMap), the
observed routes to it, and an action list (trace armed with the suggested best
path, open full map, time machine). Contacts owns the show/open/reshow loop;
topology.suggested answers "the best evidence-backed way to reach this node",
wired ahead of the firmware route in the trace arming precedence. Timemachine
and Trace grew the entry points the page opens.
Four fixes fell out of driving it on real hardware:
Slow open. Each located-node page built a fresh BasemapSource and resolved its
TileJSON -- a blocking HTTPS round-trip (~100ms, up to the 12s timeout on a bad
link) -- on every open. Hold one source for the session (ctx.basemap_source),
shared by the map, the picker, and the preview, and warm it behind the menu at
login, so the resolve is paid once, off the navigation path.
Black map. A companion advertising nonsense coordinates (Homestead reports
lat -97, lon -1042) flung the viewport off the world, leaving the map a screen
of empty/water fill that reads as solid black -- both from "open full map" and
from filtering to it and framing. The old guard only caught the 0/0 no-GPS
null island; the new geo.usable_fix() rejects any out-of-range fix too, applied
to the map markers and the location preview alike. Verified on-device: of 265
located contacts, only Homestead drops.
Path to node. The route/suggest rows render through THE path widget with a
1-byte hash in parentheses (a "name (3d)" tag), so they cross-read hop for hop
against the route graph's byte labels.
Route graph. Draw us on the left and the target on the right (us to target,
"how we reach it", left to right) via a local endpoint swap that leaves the
shared route_graph_style untouched for its other callers, and give the fan a
generous lane step and row budget so a well-connected node's candidate paths
spread into separate, legible tracks instead of a cramped tangle -- the height
stays adaptive, so a one- or two-route node still draws compact.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The recipient picker listed every contact, but a courier message is a
direct message — and a DM to a repeater, room, or sensor is meaningless.
Filter the recipient list to companions (an untyped contact gets the
benefit of the doubt), and open it sorted A→Z by name instead of by
heard, with heard/packets/key still a Ctrl+arrow away.
Fold the rule the chat picker already applied inline into one shared
predicate, is_direct_messageable(node_type), so "DMs go to companions
only" is enforced in a single place and both pickers speak it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The "Nodes" menu was a contacts roster mislabeled — it lists the device's
contacts (get_contacts), which upstream meshcore calls "contact" too. Rename
throughout, no CLI alias:
- tools/nodes.py -> tools/contacts.py (ContactsTool, `meshterm contacts`)
- ui/nodes_screen.py -> ui/contacts_screen.py (ContactsScreen/open_contacts)
- ui/nodelist.py -> ui/contactlist.py (ContactListScreen/ContactRow)
- widgets: NodesSort->ContactsSort, nodes_table->contacts_table
Courier's recipient is a contact you address, so its prose and CLI arg
(`courier queue CONTACT ...`) follow; persisted node_key/node_name stay.
The reception/hardware layer keeps "node" — HeardNode, heard_nodes(),
observations.node, trace_hops.node, map/atlas participants, node types, admin
targets, and the Time Machine's reception picker (its "Nodes" heading stands).
CLAUDE.md's lexicon is rewritten to the refined definitions: node = any mesh
radio; contact = a node your device knows (discovered or added).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Watchlist rows now lead with the shared node-type marker in its own
colour (the type is stored at star time; legacy entries resolve
through the contact table) with the name in its key-derived hue —
silence keeps its ⚠ tail instead of tinting the glyph. Unacked alert
labels resolve back to a key for their hue (contacts first, then the
watchlist's own names, so a starred node colours even offline); acked
rows recede to muted whole. The star-a-node picker draws the same
glyph-and-hue rows. Long rows can now be read to the end: SelectScreen
grows an opt-in hscroll flag — ←/→ slide every row under the pinned
pointer via crop_cells, clamped at the widest row, surviving ↑↓ and
resetting on filter edits — enabled only here, with the footer
advertising it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adverts carry the transmitting node's whole public key, but only its 12-hex
prefix was ever stored (observation_from_event truncated it with [:12]). That
prefix is the app's identity/join key — heard_nodes groups on it, and the map,
nodes, and watchtower screens correlate it against contacts' 12-hex key_prefix
— so it must stay short. Instead, record the whole key alongside it in a new
observations.public_key column (schema v12) and keep the prefix as node.
The Time Machine picker prefers the captured full key, then the live contact
resolver, then the bare prefix — so a heard node's key now shows as fully as
the hash lane fits, offline included. HeardNode surfaces the key from whichever
row captured it; the simulator's observations carry it too.
No backfill: the full keys past adverts arrived in were never persisted (the
raw payload wasn't stored) and no local source has them, so old rows can't be
completed. The database was wiped to start fresh under the new schema.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add our own node to the subject picker — first in the Nodes list, always,
matching the Nodes screen's own-node-first order — and give it a full activity
page built from the outbound record, the one subject reception history can't
describe since we never overhear ourselves: Activity (traces + messages sent),
Reach (trace bottleneck SNR), Rhythm (by local time of day), and a Ledger.
Backed by three new repository reads (self_transmissions, self_trace_reach,
self_activity_ledger) and a SelfActivity model.
Make the picker sortable and filterable: Ctrl+left/right pick the column
(name / heard / packets / hash), Ctrl+up/down force direction, a two-colour
header lighting the active sort. Supported by letting a Separator title be a
Rich Text and mapping Ctrl+Up/Down to actions.
Anchor the columns to the left: size the name lane to its content, not the
terminal, so the columns stay put as the window widens, and flex the hash lane
to fill the freed width — showing a heard node's full public key when a contact
holds it (new make_key_resolver) instead of the stored 12-hex prefix. Truncate
hashes on a byte boundary (even digit count) in the shared highlighted_hash
widget, floor the picker hash lane at 8 cells.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The viz package existed solely to render one interactive HTML chart for
tx-optimize sweeps, pulling in Plotly — a heavyweight dependency serving a
single artifact while every other timeline renders in-terminal via
braillechart. Drop it with no replacement: delete meshterm/viz, the
--viz/--no-viz flag and its render call, the now-unused output_dir setting
(config field, default, loader, example), the Plotly test, and the README
and pyproject references.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three independent consistency passes:
1. **Message paths: unknown relays as grey full-hash**
The message-paths dialog showed unnamed relay hops as a single lit byte (e.g.
`e8`). Now they render at the device's path-hash-mode width, muted grey, and
annotated with their addressed byte — `e839f2 (e8)` — so the rows stay
readable and cross-reference the route graph's byte labels (matching named
hops' trace presentation). Added `hash_as_name` flag to path_text widget.
2. **Dialogs: multi-layer compositor + backdrop for empty stack**
Opening a confirm over a floating popup used to stretch that popup to
full-frame (the single-float compositor's bug). Now the compositor maintains
a pool of _MAX_DIALOG_LAYERS=8 centered-box floats, each rendering one layer
above the background — so a popup and its confirm both stay their own size.
Also: button_dialog/typed_confirm/autocomplete now float over a blank base
when the stack is empty (via _run_dialog_screen), guaranteeing popups render
as small boxes on an empty stack, not full-frame chrome. Simplified
message_dialog to delegate its backdrop logic to button_dialog.
3. **Data-loss dialogs: red `destructive` tier**
Added a second caution tier to the dialog API: danger=True stays amber
(disruptive: discard edits, reboot), destructive=True is red (irreversible
data loss). Updated: single-record delete, channel "clear slot", courier
"clear finished" (new confirm), and startup "remove device" (splash confirm).
Per UX spec: bulk/irreversible deletions gate behind typed_confirm (also red);
single-record delete is a red Cancel/[verb] button dialog.
All changes are backward-compatible. Full suite: 793 tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add TCP alongside serial and BLE as a first-class companion transport, end to
end: discovery/connection/selection learn a tcp: stable_id and host:port target,
--tcp and TCP config profiles reach a WiFi board or network proxy, and a network
companion -- which isn't discoverable -- is rebuilt into the picker from its
remembered endpoint.
Device picker gains the network-device affordances:
- "Add a network device" prompts for host:port, smoke-tests it, and remembers it.
The row is marked "- experimental" in the cautionary hue.
- Delete on a network row opens a chromeless Cancel/Remove confirm (Cancel left,
Remove right and default: Enter removes, Esc cancels), then DeviceStore.forget
prunes it and re-draws the list. Serial/BLE rows stay inert.
- The "Del remove" footer atom now follows the highlight, surfacing on the box's
bottom border only while the cursor sits on a deletable row; the box is sized
to the fullest footer so it never widens mid-navigation.
Folds in the accompanying courier, channels, and config refinements built along
the way. Screens stay readable at 72 cols; full suite green (745 tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Some channels — an automated #wardriving beacon, a chatty public relay —
carry traffic worth recording but not worth being notified about. Muting
a channel now stops its new messages from raising the unread badge: the
recorder skips the unread bump for a muted channel (its transcript is
still written to history), and muting zeros whatever unread it had so the
badge drops immediately. The mute lives in a new MuteStore
(<config_dir>/mutes.json), keyed by the channel's intrinsic identity so
it follows the channel across slot moves and survives restarts.
The toggle is a state-reflecting row in the channel-detail subscreen
(🔕 Mute / 🔔 Unmute), and the manager list marks a muted channel with a
🔕 in its (then always-empty) unread lane.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The companion protocol reports only a battery voltage, so charging is a
voltage-trend inference. Some devices could instead expose the standard
GATT Battery Service (0x180F) whose Battery Level Status characteristic
(0x2BED, Battery Service 1.1) carries a real, firmware-reported charging
flag. Dumping the GATT table shows no current MeshCore build implements
it — the BLE profile is just the Nordic UART pipe plus DFU — but wire it
up now so it lights up automatically the day one does.
get_hw_charging() reads 0x2BED best-effort over BLE (reaching the bleak
client the forced-teardown path already uses) and prefers its flag; any
failure — service absent, read refused, short value, non-BLE transport —
is a quiet None, so the poller falls back to the inference and every
current device behaves exactly as before. Verified against a live device:
the read resolves and returns a clean, steady None. The 0x2BED bitfield
decode is a pure function unit-tested against the spec's Charge State enum
so the parse is correct ahead of any hardware that populates it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A trace is one packet that must travel the whole path out and back before
its reply reaches us, so a flat 10s wait cut long walks off mid-flight —
the packet completed the circuit on the mesh, but our wait had already
expired, so we logged a phantom "no reply" for a route that worked.
run_trace now sizes the wait to the hop count (base + per-hop, capped)
when the caller doesn't pin a timeout, counting the transmitted path
(return leg included). An unknown path-less flood falls back to the old
flat budget. run_traces and the mock thread the optional timeout through,
so every trace surface — live screen, CLI, probe, TX optimizer — inherits
the scaling.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The channel manager still felt uncached: its capacity read walked all 64
slots one-by-one on firmware that never rejects an out-of-range index
(the Waymarker), on every open, outside the session cache.
Two fixes for the two slot probes:
- read_channel_slots now stops after a run of MAX_CHANNELS consecutive
empty slots. The manager packs channels from slot 0 up, so an unbroken
empty run that long means every configured channel has been seen — safe
against within-capacity gaps, which can't reach that length. This cuts
every (re)probe on never-rejecting firmware from 64 reads to ~11, felt
after each channel edit (a mutation invalidates + reprobes).
- channel_capacity deliberately keeps walking to the firmware's rejection
(it must discover a larger rejecting firmware's real ceiling — the
empty-run guess would truncate an empty 12-slot device to 8). Its only
slow case is never-rejecting firmware, bounded instead by caching: a new
DeviceState.channel_capacity() holds it for the session (a hardware
constant, dropped only on reconnect) and prewarm warms it after login,
so the manager opens from cache like every other screen.
706 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The radio's RX log has recorded every frame's relay path all along;
now a chat message can be asked how it got here. ^P floats a paths
window for the picked (or latest) message: one row per logged arrival
— time, reception SNR, and the relay chain through the shared compact
path widget — so a channel broadcast that reached us three ways reads
as three rows over two distinct paths, resends labelled.
Channel messages are matched by evidence, not guesswork: overheard
GRP_TXT frames decrypt under the channel's own key (fetched from the
device when the conversation didn't carry it) and match the message by
content, tolerating the "Name: " sender-prefix convention on the wire
— our own broadcasts included, since repeaters' rebroadcasts of us are
overheard like anything else. Direct frames are ECDH-encrypted, so the
view offers the window's TXT_MSG frames matched by time alone and says
so plainly. The sender-prefix parser moves to core.channels where the
matcher, the transcript, and the dashboard feed now all share it, and
the repository gains the windowed packet_frames_between query.
Both chat kinds now share one message-pick model: ↑ picks, the pick
walks with ↑↓/PgUp/PgDn/^Home and day-jumps with ^PgUp/^PgDn, ^End
returns to compose. Enter on a pick replies in a channel (unchanged)
and opens the paths in a direct chat, each with its banner cue.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Four small UX gaps closed together:
- Dashboard's footer hint was missing Home/End even though the screen
already handled them.
- The packet viewer's paging keys were reversed relative to the list
that opens it: PgUp/PgDn walked packets while the dashboard feed
walks with up/down. Swapped so up/down page through packets (matching
the feed) and PgUp/PgDn scroll a tall body (matching every other
screen's scroll keys).
- A raw "packet" entry (an overheard RX_LOG_DATA frame) now shows its
parsed payload class and route, and — for a channel-text frame naming
a channel we hold the key for — its decrypted text. The AES/HMAC
logic lives in core/channels.py as pure, unit-tested crypto
(decrypt_channel_text), independent of the vendored meshcore
library's own reader internals; the dashboard fetches the device's
channel secrets once at open time and hands them down to each
PacketViewer it opens. Raw-field dumps now skip fields already folded
into a flavoured row instead of repeating them.
- The Time Machine's whole-mesh charts (packets/day, nodes/day, rhythm)
gained the same mirrored numeric y-axis the dashboard's activity
chart already had. That gutter-drawing logic moved into
braillechart.axis_chart, a shared widget, and the dashboard's own
activity chart was refactored to use it too instead of hand-rolling
its own copy.
meshterm/ui/channels.py's _read_slots is now public (read_channel_slots)
since the dashboard opener needs it too.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The monitor seeds its six-hour activity buckets and its by-kind tallies
from stored observations at session start, so the header pulse, the
dashboard chart, and the traffic panel open mid-story after a restart
instead of counting from zero. A once-per-session sweep prunes
observations past the new history_days retention setting (default 365,
0 = keep forever), keeping the always-recording database bounded.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Queue a direct message for a node that isn't reachable right now and a
session-long background loop delivers it when the recipient is next
heard — or at a scheduled local time ("next 07:00", "in 3 h", any
HH:MM). Each attempt is an ordinary chat send (it lands in the
conversation history with its ack state, exactly as if typed); failures
back off exponentially and, after the first shot, wait until the node
is heard again, so an absent node is never hammered on faith. One
attempt per pass keeps a backlog draining politely, and after five
unacknowledged tries the courier gives up and says so. Outcomes raise
Watchtower alerts, so the header's badge lights when an overnight
delivery finally lands.
The outbox persists in courier.json (a queued message must survive
restarts or the promise is hollow), a Mesh-section screen carries the
queueing flow, per-entry Send now / Cancel, and the finished history,
and a scripted `courier <node> <text…> [--at HH:MM]` queues without
transmitting anything.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Star the nodes that matter (the roof repeater, the far gateway) and
session-long rules watch what the always-on hub already hears: a
silence alarm when a starred node goes quiet past its threshold
(latched per quiet spell, re-armed — with a recovery note — when it
returns), an SNR-sag warning when its receptions' median drops well
below the run before it, and a mesh-wide heads-up the first time a
never-before-seen node id appears (baselined against the whole stored
observation history, announced at most once ever). Nothing transmits.
Alerts persist in watchtower.json alongside the watchlist and rules,
show as a red ▲ badge in the header, and are read, acknowledged, and
configured in a new Mesh-section Watchtower screen. A remote-battery
rule was considered and dropped: observations carry no parsed battery
telemetry, so there is no honest data source for it yet.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A new Device management entry for administering repeaters and room
servers. The shared credentialed-first picker (extracted from TX
optimize) chooses the node, login runs on remembered passwords with a
one-time prompt otherwise (forgotten on rejection), and the editor
lands in the local device-configuration screen's exact lanes — but
speaking the node's text CLI through a new public
Device.send_remote_command, which brings the repeater-only knobs the
local editor never had: TX delay, Direct TX delay, RX delay, airtime
factor, advert intervals, flood max, repeat, guest password.
Reads cost one paced mesh round trip each, so the editor never
bulk-reads on open: values show what the last read said, age-stamped,
from a new per-node remote.json cache; Read settings refreshes them
under an abortable progress dialog and Apply sends the staged sets
the same way, folding confirmed values back into the cache and
keeping refused ones staged. Advert, clock sync, password change, and
reboot run as confirmed popups.
The Command line is a readline-style screen onto the node's raw CLI:
per-node history persisted across sessions (↑↓ recall, adjacent-dupe
suppression, capped), Tab completion over the known verbs plus every
catalog get/set spelling with ghost-text preview, one command in
flight at a time. The simulator grows a full repeater CLI so all of
it drives against --mock, and a scripted one-shot
(repeater-admin <node> <command…>) covers automation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
MeshTerm now announces the connected node by itself: a zero-hop advert
every 1/2/3/4 hours (default hourly) and a flood advert every
3/6/12/24/48/168 hours (default daily), each switchable off. The
cadences are staged and applied from a new Background adverts section
in the device-configuration editor, and persist per device (keyed by
public key) in adverts.json alongside the last-sent mark for each type
— so the countdown survives restarts, and a manual send from any
advert flow resets the same clock the scheduler reads.
The scheduler stays transmission-shy: at most one advert per 30 s pass
(a flood due together with a direct waits a pass), and a never-sent
device is armed to count from now rather than advertised immediately.
It holds no device subscription — each pass just skips when nothing is
connected — so reconnects need no special handling, and scripted CLI
runs never start it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
An auto (device-routed) Trace target went on the air as a bare
destination hash — an observer confirmed no repeater hops — so any
target past a direct neighbour could never answer. Dumping the
radio's contact table showed why: every contact reports
out_path_len -1 (flood). MeshCore firmware only learns a contact's
out_path from two-way addressed traffic, and tracing repeaters
teaches it nothing, so "device-routed" has nothing to route with.
Auto now resolves a route the session can prove, in order: the
device's learned route when one genuinely exists, else the outbound
leg of the last successful stored walk (_previous_outbound extracts
it from the recorded boomerang), else a bare direct attempt. One
resolver feeds both the wire and the screen, so the planned route
line renders exactly what Trace will transmit, with its provenance
under it — (auto · device route), (auto · last trace · stamp), or
(auto · direct — no known route) — and the summary's path row says
the same instead of claiming a device route that never existed.
The connection layer's path-less fallback keeps the mirrored return
leg for contacts that do carry a learned route (the trace protocol
reflects nothing home), and its docstring now says the firmware
never routes a trace itself. Resolution verified against the real
database and radio: a multi-hop target resolves to its full stored
boomerang, the direct gateway stays a single-hop spec.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A trace is one walked path and "target" is pure UX (hardware-verified),
so the feature now splits along that line: Trace target keeps the
symmetric mirror-and-return flow (picker, Explore paths), while the new
Trace path skips the picker entirely and walks a hand-composed route —
no Auto, no Explore, Trace gated until a path exists, history recorded
under the "(path)" sentinel so composed walks never pollute the target
picker. The composer loses its mode toggle: a pinned target means
mirrored, no target means verbatim, and its preview opens as us → us.
Sampling returns, deliberately: a Sample count action (1/2/3/5/8) runs
that many traces per commit, paced by trace_cooldown_s between sends and
counted off on the in-flight dialog — never a burst. Actions reorder to
compose/explore/width/samples/trace with a Back row; the w/p/x hotkeys
retire.
Display: route nodes drop the underscore (prompt_toolkit draws U+00A0 as
"_", so name (hash) now joins with plain spaces), the composer renders
nodes exactly like the route line, the (previous · …) stamp gets its own
line, the path row ends with the displayed hop count, Explore rows lead
with the route instead of an "observed path" label, and the empty Traces
placeholder is gone.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two long-standing flakes, one root cause each:
Quit with a serial companion often froze until the exit watchdog killed
the process. The meshcore dispatcher's stop() awaits queue.join(), but
its processor task exits after draining at most one event once stopped —
so a routine advert/RX-log burst (two+ events queued at that instant)
deadlocks the join forever (repro confirmed against the installed
library). MeshCoreDevice.disconnect now gives the graceful teardown two
seconds, then abandons it: the dispatcher task is cancelled synchronously
and the raw transport closed directly, so the port is still released and
even the forced path finishes well under the watchdog.
BLE startup often failed until the user re-selected the device. Connecting
by bare address makes bleak re-discover the peripheral with a fresh
internal scan, which on Windows intermittently misses a slow-advertising
companion. The discovery scan's live BLEDevice now rides along
(DiscoveredDevice -> probe/context -> create_ble) so the connect opens the
peripheral directly, and a transport-level ConnectionError is retried once
before surfacing — the manual re-select was never anything but a retry.
PIN/bond rejections still surface on the first attempt, unlooped.
Verified against hardware: BLE info read off the Waymarker connects and
exits cleanly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The topology graph gains its fourth evidence source — the only one not
derived from our own reception. A repeater we hold admin rights on
reports its neighbour table (who it hears directly, at what SNR) when
asked over the mesh, revealing links and whole nodes our radio has
never received anything from: exactly the second vantage point the
composer runs dry without, two hops out.
Probed live against YUL-Cartierville (v1.15.0) through the Waymarker
before building: firmware answers the binary NEIGHBOURS request but
silently ignores guests, so the flow is login-first, tx-optimize style
(stored password, one-time prompt, forget on rejection). An empty table
is a normal answer — the gateway's own table was empty during the probe
(confirmed via its text CLI: '-none-') — so the UI says "ask again
later" instead of erroring.
Standing on a fetchable repeater, the composer shows a "fetch
neighbours" row with the suggestions it extends; the fetch runs under
the abortable spinner dialog, persists as a runs-row-owned snapshot
(schema v8, neighbour_reports), and the composer reopens mid-thought
with the same hops and refreshed suggestions tagged N. Refetching
supersedes the old snapshot per (repeater, neighbour) pair rather than
inflating the link's sample count: a table is current state, not an
event stream.
Verified end-to-end on hardware via MeshCoreDevice over BLE (login +
fetch + clean empty-table return) and in the mock (guest refusal,
discovery of a never-heard node, composer render + hand-off). 441
tests pass; the real database migrated v7 -> v8 cleanly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Everything we receive carries a fragment of the mesh's shape: successful
traces crossed every link in their path (out and back — the reply retraces
the reverse path, so links read both ways), each contact's out_path is a
route the firmware distilled from received floods, and RX-logged packets
report the relay chain they rode in on. A new topology service folds those
into one undirected evidence graph (schema v7 persists packet paths).
The trace screen stops auto-starting on target pick. Enter runs a burst
under a floating tracing dialog (spinner chip, live progress, Abort); "p"
opens a hop-by-hop path composer whose suggestions are the tail's observed
neighbours, strongest link first, with raw hex entry for unseen nodes; "x"
browses ranked path scenarios (device route, direct, best observed
alternatives) and can probe them all — a measured burst per candidate,
recorded as path_candidates rows, reliability-first ranking, winner offered
for adoption.
Composed specs are outbound-only, always ending at the target's hash: the
stored data shows the firmware auto-replies along the reversed path, and
user-composed "there and back" loops (3d,f2,3d) never succeeded. Specs
collapse to a uniform per-hop width the narrowest known hash can honour.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The on/off switch never made sense for a pure listener: recording is
now unconditional, so the toggle submenu, the persisted preference
(MonitorStore/monitor.json), and the header's ON/OFF framing are gone.
The monitor registers its hub subscription at startup — a device-free
step that cannot fail — and its run row opens lazily on the first
observation, so recording follows the hub wherever it opens (eagerly,
or deferred by connect_on_start) and silent sessions leave no empty
runs. The menu entry is now just the read side, "Heard nodes", and the
header shows live heard counters instead of a switch state.
meshterm monitor on the CLI drops --on/--off for a bounded foreground
capture (--seconds N, 0 = until Ctrl-C, mirroring chat listen): it
tails each overheard packet live and summarizes the window's heard
nodes when it ends, while persisting history exactly like a session.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Checked against companion firmware (MyMesh.cpp CMD_SET_TUNING_PARAMS):
rx_delay_base and airtime_factor are floats carried x1000 on the wire
(constrained to 0-20 s and 0-9 respectively), so the settings now speak
real units — previously the raw wire integers were shown and edited,
making "2" mean 0.002.
The TX delay factors added in the previous commit are removed: the
companion handler reads exactly the two fields and ignores trailing
bytes, so they were silent no-ops. They exist only as remote-CLI
settings on repeater/room-server/sensor firmware.
Airtime factor now leads the Tuning section, ahead of RX delay.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New staged settings: contact auto-add bitmask, default flood scope
(hashtag-normalized, 31-byte slot), and the two TX delay factors the
library always zeroed (write-only on hardware, so unknown values fall
back to 0 exactly as before). TX power is now capped by the board's
reported max_tx_power instead of a hardcoded 22, with the prompt hint
and validation sharing one effective_maximum.
Device actions gain Sync clock (drift shown before writing host time;
also a config sync-clock CLI command), and Device info now leads with
a live Status panel: role, model/firmware, battery, storage, clock
drift, uptime, and radio/packet statistics — every read best-effort so
older firmware still renders.
The simulator models all of it, including a seeded clock drift.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Enforce one voice across every screen: column and section headers are
UPPERCASE, menu items are Capitalized buttons, and short descriptions
drop their trailing period (with real ellipses instead of "...").
Tools gain a display `title` separate from their kebab-case CLI name,
so a menu item now reads exactly like the screen it opens ("Device
configuration" opens Device configuration) and result windows inherit
the same name. Every sub-screen backs out via a "Back" row below a
blank separator (the config editor was missing both; its Close is now
Back, spelling out the discard when changes are staged), while the main
menu and device picker keep "Quit". Panel and dialog titles now wear a
brighter tint of their border colour via new title.* theme styles.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A direct message was transmitted once and marked failed the moment its
single delivery-ack window lapsed, leaving Ctrl-R as the only recourse.
Now the send path re-sends an unacknowledged message automatically before
giving up, matching the mesh convention of a few tries.
_deliver_direct loops send_direct_message until an ack lands or the budget
is spent; both the initial send and the manual Ctrl-R resend route through
it. Each attempt already blocks a full ack window before the next, so the
retries are spaced by that window rather than hammering the radio, and a
hard send rejection propagates without retry.
The budget is the new direct_message_soft_retries setting: retries + 1
tries, clamped to 0..2 so a message is never re-broadcast more than twice
(0 = one shot, default 2 = up to three tries).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The quit dialog grows a third button, shown only when the session is on a
Bluetooth link that Windows actually holds a bond for (never serial, an open
PIN-less companion, or a non-Windows platform). Choosing it disconnects the
link and then drops the OS-level pairing during teardown, so the next launch
re-runs the PIN ceremony. MeshTerm's remembered-device record is deliberately
kept — unpairing forgets the credential, not the device identity.
Backing this: static WinRT helpers is_ble_paired / unpair_ble mirroring
_pair_ble_windows (Windows-gated, best-effort), an active_address on the
context, and _can_unpair / _unpair_on_exit in the menu.
Also close the WinRT BluetoothLEDevice handles in every helper. Each
from_bluetooth_address_async returns an IClosable that pins the OS ACL link
open; unpair removes the bond but not the link, so a leaked handle left the
device "unpaired but still connected" and refusing to re-pair until it was
power-cycled. Closing the handle in a finally is what finally tears the link
down.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
bleak's pair() on Windows only performs the CONFIRM_ONLY ("Just Works")
ceremony and never sends a passkey, so a PIN-protected MeshCore companion
bonds without authentication and then rejects the GATT subscribe on its
authenticated UART characteristic with Insufficient Authentication -- a
correct --ble-pin was reported as "rejected" and the device could never
connect.
MeshCoreDevice now runs the WinRT PROVIDE_PIN ceremony itself before opening
the link, creating the ENCRYPTION_AND_AUTHENTICATION bond the characteristic
requires; Windows persists it, so later reconnects need no PIN. A first GATT
auth failure with a supplied PIN also triggers one unpair-and-re-pair retry to
heal a stale unauthenticated bond from an older attempt. The step is a
best-effort no-op off Windows, when no PIN is set, or when already bonded.
Verified on hardware (AA:BB:CC:DD:EE:FF, MeshCore-Homestead): correct PIN
pairs and connects (reads model Seeed Tracker T1000-E), wrong PIN fails clean
with the "rejected" DeviceAuthenticationError, no PIN reports "requires a PIN",
and a bonded device reconnects with no PIN. Full CLI path exercised end to end.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When a companion needs pairing, the interactive picker now opens a centered PIN
popup (its own bordered box under the wordmark, matching the splash) and retries the
connection with the entered code — re-asking with a "rejected" note on a wrong PIN,
and returning to the device list on Esc. The masked field clears its stale error as
soon as the user types a fresh code.
The connect failure is now a distinct DeviceAuthenticationError (a DeviceCommandError
subclass), so the picker can catch "needs a PIN" specifically while the scripted CLI
still catches the base class, prints the remedy (pass --ble-pin), and bails without a
prompt — no interactivity for scripted runs. --ble-pin already feeds the first attempt.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A BLE companion that requires pairing accepts the link-layer connection but rejects
the GATT subscribe with "Insufficient Authentication", which bleak raised as a bare
BleakGATTProtocolError — escaping connect() as an unhandled traceback.
Translate that (and related "Insufficient Encryption"/"not paired" rejections) into
a clean DeviceCommandError that names the fix: pass --ble-pin, and pair once in the
OS Bluetooth settings on Windows. It's classified as its own actionable case, never
as a dropped link, so the session offers a PIN rather than a pointless reconnect.
The startup probe now lets this actionable error through instead of collapsing every
failure to a bare "didn't answer", so the picker shows the real remedy in the splash.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>