A companion, a room server, or a sensor does not relay a trace with the
firmware defaults (allowPacketForward in Mesh.cpp: client repeat is off,
and allowed only on 433.000, 869.495, and 918.000 MHz; disable_fwd is on).
A walk that turned at one of them never came home: the history held 36
companion traces and no reply from a real one.
The route to such a node is the same. Only the walk changes: it turns at
the last relay of the route, and the node is not in the spec. The route
out of a walk that comes home is still the full route to the node, so
the Device route row can offer it for DMs and admin.
- relays_traces() in models states the rule. render_forced_spec takes
relays=False, and forced_outbound reads the route back from a walk.
- services/trace_plan holds the route choice that the trace screen made
inline (radio route, outbound paths, best observed, last trace,
direct), so the CLI uses the same one.
- The trace screen names the relay where the walk turns, in a line under
the route. With no relay to turn at, Trace is dim. A neighbour that
does not relay gets "no relays" offered on the Device route row, with
the packets heard directly as its proof.
- The trace session turns a caller's spec (node page, Message paths) at
the last relay, so the callers do not need the rule. A sender that is
not a contact gets its type from its stored adverts.
- The composer and Explore paths turn the walk at the last hop.
- `meshterm trace -t NAME` with no --path walks the menu's route and says
on stderr where it turns. With no known relay, it is a usage error.
- The connection turns a device-routed trace at the last learned relay.
The simulator now stops a walk at a node that does not relay, as the
mesh does, and the tests that walked through Alice walk repeaters.
The picker keeps one list with no mark, sorted as before (JP's choice).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Enter on an arrival in the Message paths dialog (^P in a chat) opens a trace
of the route that the row draws, walked from our node:
- A received message is traced back to its sender: the relays in reverse,
out to the sender, and back on the mirror (Trace target).
- A sent direct message is traced out to its recipient on the relays in
order (Trace target).
- A sent channel message came back to our node, so its relays are the full
walk (Trace path).
- A sender with no key turns the trace at the first relay (Trace path).
- A direct-routed copy carries only the rest of its route, so it offers no
trace, and the hint does not show Enter there.
The trace screen opens idle and armed. Nothing transmits until the user
starts a trace. The dialog stays pushed under it as a hub, so Esc from the
trace comes back to the same dialog and highlight. In Trace target, a trace
that comes home lights the Device route row, so the route a message took can
become the radio's route to that contact.
A sender that is not a contact opens on its key, and open_trace takes a
label so that the screen shows its name instead of the hex.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Explore paths → Probe all walks each candidate once. When the user adopts a
candidate, that trace already measured exactly the adopted route, so the
screen now starts the route with it instead of with no traces. The Device
route row then offers a route that came home at once, with no second
transmission to prove it again.
The session keeps the probe traces by spec (seed_traces). Any later change to
a probed spec starts with them too, because they measured that route.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Trace target gets a Device route row after Explore paths, for a target that
the radio has as a contact. It says what the radio holds now, and Enter opens
the node page's Device route popup.
A trace that came home walked its route out and back, the best proof that
MeshTerm can have for a route. When its route out is not the radio's route,
the row turns green: "Use this route for DMs and admin — 4/5 came home…".
Nothing interrupts the session. The popup then shows the traced route with
its proof, and offers "Use the traced route" and, with a site, "Use and keep
it" before the usual rows. The popup changes no path of the trace screen, so
its traces and medians stay.
The gallery's Trace specimen draws the green row at both widths, and the
features guide says what it does.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Routes tab ends on a "Device route · …" row that says what the radio
holds now: none (it floods), or the hop count and how the radio got it
(learned, set by you, kept at a site). Enter opens one popup that turns its
page:
- The first page draws the route and the kept route (with why a kept route is
not in the radio: its first hop went silent), then the actions: Set the
route…, Keep this route (or Stop keeping it), Forget the route, Discover
routes…. With no site, Keep says that it needs one and opens Outbound paths.
- Set the route… turns to a picker: the suggested route first (an outbound
path, then the best way on), the routes heard, no relays, and Compose…, which
opens the path composer. A picked route replaces a kept route too.
- Discover routes… waits for the cooldown, sends one path discovery, and shows
the route each way. "Use this route" sets the route out.
After a change, the Routes tab is drawn again in place, with the highlight on
the row. Gallery specimens for both pages, and the glossary and features guide
name device, kept, and saved routes and path discovery.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
meshterm route works on the route that the radio holds to each contact,
with a plain face and a --json face. It has no menu row: the menu works
on a route from the node page of the contact.
- list and show: the spec, the origin (kept, set, learned, flood), the
kept spec, where it is kept (site or fixed device), the drawn route
from our node to the contact, and a kept route skipped because its
first hop is silent.
- set (--keep), forget (--yes), keep, unkeep: each prints the route as
it is afterwards. keep without a site fails with exit 1 and names
meshterm outbound use.
- discover (--use): one path discovery, the route each way.
- sync: the reconcile of the connection, with its counts.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The companion holds one device route per contact and replaces it with every
path return, good or bad. MeshTerm can now write that route, keep it, and
make the routes follow the user from site to site.
- Device.set_route reads the radio's own record and writes it back with only
the route changed, at the radio's hash width. Device.discover_route sends one
path discovery and returns the route each way. The radio's PATH_UPDATE is a
new hub event, EventKind.ROUTE.
- The outbound store keeps device routes per site and per fixed device (kept),
and the routes a site learned (saved). Keep needs a site or a fixed device.
- services/device_routes.RouteKeeper (ctx.routes): forget, set, keep, unkeep,
discover, and the status the pages show. At each connection and after each
site change it reconciles: routes that start on another site's first hop are
saved for that site and cleared, the new site's saved routes come back, and
kept routes are written back. A silent first hop is skipped. A route change
that the radio pushes gets the kept route written back.
- A site change on the Outbound paths page, its suggestion, and
`meshterm outbound use` go through the reconcile and say what changed.
- New preference auto_route (off): before a DM, set the suggested route when
the radio has none, or one from another site.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The companion keeps one route for each contact and never forgets a dead
one, so each retry on it was lost with no error. Now ^R on a direct
message that went along the learned route asks: Cancel, On route, or By
flood. By flood forgets the route (nothing is transmitted), flushes the
contact cache, and sends the one retry. A message that already went by
flood keeps the plain ^R. The admin login offers Try by flood the same
way after a silence from a node that has a route, one time for each
flow. Nothing floods unless the user chooses it.
Refs: #34
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The user names the ways out of their node, one or more hops long, in order.
Each site of this computer (Home, Foulab) has its own list, one site is
active, and a fixed device keeps its own list.
- Explore > Outbound paths: the sites with the active mark, the fixed
devices, and a staged editor for each list. Each path shows how well its
first hop hears us (out) and how well we hear it (in). Measure sends one
trace out along a path and back, after the cooldown.
- The path composer has an outbound mode, and puts the active paths first
while the insertion slot is at our node (trace target, trace path, the
TX sweep route).
- MeshTopology.outbound_scenarios: one route through each path. It scores
from the link after the path, lets the user's order win a near tie, and
skips a path whose first hop was silent for outbound_silent_hours. These
routes lead Explore paths, the node page's Routes tab, and the trace auto
route, ahead of a device route that starts on another hop.
- A first link out of our node is scored from what the far end measured,
when a trace gave such readings.
- The site suggestion: when the radio hears the first hop of another site
three times, the main menu asks to use that site and says what it heard.
"Never suggest" sets the new suggest_site preference off.
- CLI: meshterm outbound [show|sites|use|set|remove|fix|unfix], plain and
--json.
Refs: #35
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
core/outbound_store.py keeps, in outbound.json, the sites of this computer
with their outbound paths in order, the active site, and the fixed devices
with their own paths. lookup() gives the fixed device's paths, else the
active site's, else none. The glossary gains outbound path, site, and fixed
device.
Refs: #35
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A link pooled the readings of its two directions, and received packets far
outnumber trace first hops. Link.heard_at now keeps each reading under the
node that measured it, and MeshTopology.readings(sender, receiver) gives one
direction. A neighbour report is the repeater's own reading, so it goes in
as the reverse pair. The pooled ranking does not change.
Refs: #35
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Trace path with no path starts its highlight on Compose path, because
Trace cannot act there. Adopting a route in Explore paths moves the
highlight to Trace, the same as the composer, so Enter walks it.
In the composer, "Use this path" is now the last row, after Auto.
Refs #29
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A MeshCore node keeps its name in a 32-byte field, but the sender and mention
patterns stopped at 20 characters. A sender with the name YUL-H7H-SiO-C-Heltec1
(21 characters) showed with no name and no key. The patterns of the channel
sender, the mention, and the compose chip now use one limit, MAX_SENDER_NAME.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The reporter of #28 asked for a way to clear one or all of the stored admin
passwords. Del in the node picker forgets one. This adds the other: a
"Forget passwords" row in the Managed nodes section of the menu, behind the
red typed confirm (type FORGET), which names the nodes it will forget. It
also forgets the password of a node that is no longer a contact, which no
list shows. With nothing stored, it says so and asks nothing.
On the CLI it is `forget-admin-passwords --yes`: without --yes it is a usage
error, and with it the answer lists the nodes it forgot (NAME, LAST_LOGIN,
KEY). It exits 5 when nothing was stored, and works with no device.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A stock repeater sends no reply for a wrong password: handleLoginReq returns 0,
and the companion pushes LOGIN_FAILED only when a reply arrives that is not "OK".
The simulator refused a wrong password instead, so --mock and the tests never
walked the path of #28. Now a wrong password gets NO_REPLY from the simulator,
and only a node in its new _refusing set refuses, as other firmware can.
The comments that said the firmware sends a refusal now say what it does. A new
test walks #28 from end to end on the simulator: the stored password gets
silence, the user selects "New password…", and the repeater accepts the new one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A script cannot get the "New password…" offer of the menu. When a saved password
gets no reply, `repeater-admin` and `tx-optimize` now say that the password may
have changed, because a node does not answer a wrong password, and that a new run
with --password repairs it: a password that the node accepts replaces the saved
one. With an explicit --password, the error says only that it can be wrong (#28).
The CLI manual describes both cases.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The menu had no way to make MeshTerm forget an admin password (#28). Now Del on a
managed node in the node picker asks first (amber, Cancel or Forget), then forgets
the password. The list refreshes in place: the node leaves "Managed nodes" and the
highlight stays on it. The footer shows "Del forget" only on a managed row, and the
PicoCalc lane shows its Delete chip.
The first page of the TX optimize dialog has the same rows, so it takes the same
Del. A wizard page can now carry a Del key and an on_delete handler, and a turned
page brings its own Del key, so the second page shows none.
The feature guide also describes the new-password offer of the previous commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A stock repeater or room server does not answer a wrong password, so a password
that the owner changed comes back as no reply, the same as a node that is down.
MeshTerm kept the old password after the silence and tried it again at each
visit with no prompt, so the user could never type the new one (#28).
Now the dialog says that a changed password can be the cause, and offers Cancel
or "New password…". The new password replaces the stored one only when the node
accepts it, so a node that is down keeps its password. Repeater admin, the TX
sweep, and the neighbour fetch of the trace screen share this flow
(ui/admin_login.py). The bulk clock sync names the same two causes on its line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rows of "Archive contacts — n to archive" read NAME · HEARD · PKTS · PLACE.
MSGS and HOPS were almost always empty for a sweep candidate; the scores still
use both signals. PLACE is text, so its word and values align left, and it is
last, so ←→ brings it back on a narrow terminal.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The places were found once with the real finder (tiles, then Nominatim) and
frozen in world.PLACES. The history stores them, so the demo shows places at
once, its place job sends no request, and each run is the same. The demo home
also clears places.json at each start.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An archived contact is no longer on the device, so the job never saw its
location unless an advert was heard. The contact store keeps that location;
the scan now reads the archived contacts of the device from it. On JP's data,
290 of 422 archived contacts had a location and only 68 had a place.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A text lane padded its label to the lane width before the triangle, so the
triangle landed at the far end of the wide PLACE lane, often past the screen
edge. The label is now unpadded, as NAME's is; the columns do not move.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The CLI sorts on the same ring as the Contacts screen (name, heard, packets,
place, hash). The ring moved to ui/widgets.py so the CLI loads no TUI code, and
ordered_contacts sorts by place (no place last) and by key.
The Archived list reads NAME · ARCHIVED · PLACE · KEY, sorts by place, cuts no
value and pans with ←→, like the Contacts list. The place fill and its live
refresh moved from ContactsScreen into ContactListScreen, so any list that
declares the PLACE lane gets them from places=.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MapMarker carries a place, open_map reads it for each marker, and both the screen
and the renderer match label or place through MapMarker.matches.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Each screen opens with node names and keeps its own choice. Node detail acts only on
the Routes tab (F1). The packet viewer changes its via line and its from/to chips (F1).
Live feed changes the endpoint chips of an addressed packet (F1). Trace changes its
route lane and its per-hop links (F1). Each hint and chip shows only where a line has
a named hop.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
fields.node adds place (null when unknown) to its machine object, and the plain
face gets a PLACE column before KEY. The command reads stored places only.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Each screen keeps its own PlaceToggle and opens with node names. The record page
and the board rows, the sweep route line, the walk trail, and the composer's route
preview go through placemode.placed. The hint and the lane chip show ^N only where
a path line has a node that can change.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A PLACE lane sits before KEY. The Contacts list sizes its lanes to their content, and the
table (header and rows) pans with the left and right arrows while the mark column and the
name stay. The place sorts, and the find-as-you-type filter matches it.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Place mode changes a finished path line's hop labels to each node's place
(ui/placemode.py); a hop keeps its key and colour, a named hop with no place
reads (Name), an unnamed hop keeps its hash, our ★ stays. The node resolver
now also tells which node it named, so a shared one-byte hash finds the same
node's place. Message paths is the first screen: ^N, its hint, and a lane chip.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A place row sits above where on the Info tab, read at each paint so a place
the finder answers after the page opened still appears. Set place… types a
place that wins over the found one (blank clears it, marked · typed).
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A background job finds the place of each node that advertises a location
(Villeray, Montréal; a village; a peak near its summit). The map tiles give
the first answer offline; Nominatim corrects it when online (place_source,
default tiles then Nominatim; nominatim_url). Places are stored in the new
node_places table, a typed place in places.json wins over a found one.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
"Sync their clocks…" now opens a dialog that lists every managed
repeater and room from the start: ○ waiting, a spinner with what it
waits for (the transmit cooldown, the login, the command), then ✓ with
the node's reply or a red ✗ with the reason. Long reasons wrap under
the name. Esc stops the run after the node in progress; the OK chip
closes the dialog once the run is done.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The bulk row now shows whenever the page has a managed repeater or room
server, and it syncs each one whatever its offset: a repeater with no
battery restarts from its build date at each power cycle, and the page
only learns that at its next advert. Managed companions still have no
clock sync command, so they are never included.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The Managed nodes section ends on "Sync their clocks…" when a managed
repeater or room server is out of sync or has no sample. One confirm,
then each node in turn: a cancellable transmit-cooldown wait, a quiet
login with its stored password, and "clock sync"; one summary names
each result. "Adverts refused" moves to the end of the page.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
"All clocks" now opens the page, so the shape of every clock comes
before the lists of single nodes; Managed nodes is the first list under
it, then Adverts refused, Out of sync, and Senders with no key.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A room stamps posts with its own clock and sends only posts newer than
the companion's sync_since, so after its clock goes back (often a
reboot) it sends nothing new. When a room's newest stored post is later
than its clock now (from its clock offset), its node page shows
"posts blocked" and offers "Ask for all its posts again…". The repair
removes the room contact and adds the same raw record back, since the
firmware gives only a new contact a sync_since of zero, then logs in
with the stored password. Device.reset_room_sync tries the add twice and
says plainly if the room is left off the device.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A node whose admin password MeshTerm stores is now a "managed node": a
🔑 beside the type mark in the Contacts list, the admin picker (its
first section, renamed from "Remembered admins"), and the node page
header. The menu's "Other nodes" section is now "Managed nodes".
Contacts carry adverts_refused_until: the device's stored advert time
when it is in the future, so the firmware's replay check drops every
new advert. The node page shows it and offers "Accept its adverts
again…", which rewrites the device's own record with only last_advert
lowered (Device.accept_adverts_again). A managed repeater or room gets
"Sync its clock…". The Clock drift page lists refused contacts.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Managed nodes are now the first section and are always listed: a node
with no sample in the window shows its newest offset from all history,
or "no sample". Each table has a header line under its section heading
(NAME / OFFSET / LAST, and OFFSET / # over the histogram, now under
"All clocks"). Older offsets do not count in the histogram or title.
The CLI lists each managed node too, with null fields when it has none.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A new Watch tool, "Clock drift" (meshterm clocks): a histogram of the
clock offsets in log steps on two sides, then the managed nodes, then
each other node out of sync, largest offset first; Enter opens the node
page. Channel senders with no key are listed by name. When most clocks
disagree with ours the same way, a note says our own clock may be wrong.
The CLI lists every sender with offset_s, samples, and flags. A new
clock_report_days preference sets the window. The demo world gets four
wrong clocks.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
ADVERT packets now keep their own time (observations.sent_at, schema 22).
With the arrival times of messages, each advert, DM, and channel message
is a clock sample: the sender time minus ours. A channel sender counts
for the node its name resolves to, or under its name. A node's offset is
the median of its three newest samples. The node page's Info tab shows
"clock 3h 12m behind" in warn past the tolerance, or "in sync".
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Rows stored before the arrival time had only the sender time, so a
wrong clock still put them on the wrong day. A one-time background
backfill decrypts the stored GRP_TXT packets with the device's channel
keys; a packet with the same text and sender time is the message, and
its first copy is the arrival. On a real history it matched 12,071 of
14,197 old Public messages in 1.6 s. DMs cannot be decrypted and keep
their sender time.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
A received message now keeps its arrival time (received_at) beside the
sender time, and whether it waited in the companion's queue. Heard times,
the Messages list order, channel activity, and the packet-log search use
the arrival time, so a DM from a clock a year ahead no longer pins its
contact at "heard now". The transcript shows a corrected time as 12:31?,
and a picked message says what the sender's clock wrote. The CLI history
carries the same mark; its JSON gives both times. clock_tolerance_s
replaces ADVERT_FUTURE_SKEW_S as a preference.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
The spectrum was an HSV wheel at one saturation and value, which is not
one lightness to the eye: blue names sat at OKLab L 0.55 and yellow at
0.93. A chip's fill is its label at CHIP_LIGHTNESS of that lightness, so
chip contrast followed the label, from APCA Lc 19 on blue to 64 on
yellow.
Each byte now keeps its HSV hue angle and chroma at OKLab L 0.72, with
the chroma fitted where the gamut cannot hold it. JP chose 0.72 on a
ladder of wheels from 0.70 to 0.85. Every name reads equally light and
every chip lands at Lc 34 to 41.
The 256 colours are written out as _NODE_SPECTRUM, because computing
them costs about 14 ms here and _IDENTITY_STYLES needs all of them at
import, which would slow every start on the handhelds. A test holds the
table to _spectrum_hex, and checks one lightness, 256 distinct hues, no
console-slot collision, and that each hue angle is kept.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The thin chevron between two chips of one colour took the previous
chip's label colour. It now takes the OKLab midpoint of that label and
the chip's fill (oklab.midpoint), so the line is softer than the text.
The 16-colour console keeps the label, because it has no slot between
a bright hue and its dim twin.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Typing narrows the outbox again, and its footer says so. With the scroll
atom, the footer of an entry row is wider than 72 cells, which JP
accepted. The footer line no longer cuts a hint that is too wide at the
edge, which took Esc with it: it drops atoms instead, as the splash
already did. The outbox drops "Del delete" first.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each row now carries a third column, MESSAGE, after TARGET and TRIGGER
(or OUTCOME). Nothing is shortened to fit: the rows and their headers pan
together with ←→, and the footer says so while the table runs past the
edge. The outbox no longer filters as you type, which keeps the footer
within 72 cells.
A column header that pans now pins with its section heading, as a
header that does not pan always did.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each section of the outbox now has column headers: TARGET and TRIGGER for
waiting entries, TARGET and OUTCOME for finished ones. The rows no longer
show the message. Enter on a waiting entry opens it to edit (the courier
holds the entry while the prompt is open, so the old text cannot go out
meanwhile), and Enter on a finished one shows it. Del deletes an entry
after a red confirm, and ^R sends a waiting entry now; on the PicoCalc
they ride F3 as Send, with Delete behind it. On a narrow screen a long
name gives cells to the trigger and breaks off on its crack.
The message field has the colours of the chat's compose line: a white
prompt mark and white text.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When a row was a chip followed by text (a Courier entry: the recipient,
then what it waits for) and the text ran past the screen, the cut looked
back for a chip fill, found the chip far to the left, and drew a half
block in its colour at the right edge. The crack now forms only inside
the path that the row stamps, so a cut in plain text gets the ellipsis.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>