Sending back every change called before its group was loaded cost a step
even when the model had the arguments right, and a valid change was no
longer held as the card the user decides on. A blind change is now held as
always when its arguments are of the shape the schema asks for, and sent
back with the schema loaded only when one is missing or of the wrong kind,
such as a number or a list written as text (#2967).
The bringers list moves next to the other bringer helpers so the card file
exports components only, three tests read without non-null assertions or
any, and three files take the formatting their checks expect.
A write from a group the turn had not loaded was held as a change card with
schema-less arguments and failed on apply (#2967); it now goes back to the
model with the group loaded. Calls answered as not run are now listed on
the assistant turn too, so a thread replayed on another provider has no
orphan tool results. The blind-call hint only goes with the blind call.
Refusing every call to a tool of a group not loaded in the turn also stopped calls whose arguments were fine. The call now runs as before; its group is loaded at once, so the next step offers the schema, and when the arguments fail the MCP check the error tells the model the schema is there now and to send the call again.
The model could call a tool it knew from an earlier turn without loading its group in this one. TREK ran it, but the model had never seen its schema in the turn, so numbers, arrays and objects came as strings and required arguments were missing, and the MCP check refused it with -32602 (mostly the Tours tools).
Such a call no longer runs: its group is loaded for the rest of the turn and the call is answered with a tool error asking for it again with the schema the next step now offers.
Closes#2967
A step stopped while its calls ran (the user left the page) kept the
provider's whole record of it but answers only for the calls that ran, so
Anthropic and the Responses API refused that thread with HTTP 400 from then
on. Replaying the thread now answers every recorded call that never ran,
which also mends threads already stuck that way.
Claude Sonnet 5.5 at times ends a turn on suggest_follow_ups alone, without
writing its answer, and the turn failed as empty_answer. That step now goes
back to the model whole, its follow-ups answered that the user saw nothing
yet, and their chips come under the answer it then writes. The prompts no
longer say the call goes alone in its step.
Gemini 3 signs a tool call on its OpenAI-compatible endpoint
(extra_content.google.thought_signature) and refuses the next request
with a 400 when the call comes back without it, so every turn in which
the model called a tool failed. The Chat Completions client now keeps
the signature per call as the step's provider state and sends it back
with the call, to the model that signed it only.
An own server entered with Gemini's endpoint now also gets Gemini's
narrower tool schemas, which it refused before.
remember_for_trip runs at once during a turn, and read_attached_file and
keep_attached_file are answered by the server, but all three are calls the
model made. Rebuilding a thread from its stored steps left them unanswered,
and ChatGPT, the OpenAI reasoning models and Anthropic replay their own record
of each step, call included, so the provider refused the conversation with a
400: the continuation after applied changes failed, and every later message of
the thread with it. A step that kept a note also sent an empty user message.
Every call the model made is now answered as in the live turn, the attached
file's outcome as the answer of its call; threads that broke this way work
again on their next message. A refusal's log line names the refused field's
path, never its message.
Closes#2845
create_day_note and update_day_note described `time` as a clock label and capped
it at 250 characters, while the planner shows it as the Markdown body under the
title and allows 2000. The assistant put long text in the title and had to cut
notes it was asked to edit. Both tools now take text and time from the REST
contract with descriptions naming title and body, and an update still refuses an
empty title. The assistant's change card labels the two fields title and content
instead of text and time.
Closes#2840
TREK's code for Brazilian Portuguese is 'br', which Intl reads as Breton.
The assistant passed it straight to Intl.DisplayNames for the system
prompt's answer language and to the card formatters, so the model was told
to answer in Breton. Convert TREK codes to their BCP-47 tag with
getIntlLanguage before handing them to Intl.
Closes#2800
A stay taken into a journal now has a card of its own, so ASTTOOL-001
looks for its tool; Dawarich's tools are listed only with its addon on,
which the test now turns on beside the others.
The four new photo link tools join the journal's tools: the journey chat
offers them to editors, the trip chat loads them with the journey group and
the home chat keeps none. A journeyPhotos card shows the photos as
thumbnails with the entries they leave and reach, the first one marked as
the cover in a new order, and a caption old to new; a library add shows the
photos once they landed. A linked photo comes off again, an unlinked one
goes back, moved photos go back and a caption is written back, each while
nothing changed since. A new order of an entry's photos and a library add
have no way back yet. The journal prompt names the photo tools.
The update card silently left out the tags and the pros and cons it was
about to write, so the user approved a change the card did not show. An
entry now shows its first tags and how its verdict leans, and an update
names the tags, each side of the verdict and the coordinates it changes.
A new order of the entries is drawn as the entries that move, in their new
places. A changed entry is written back from what the tool answered it
replaced, and a reorder is taken back to the order before, each while
nothing changed since. A Dawarich stay taken in as an entry gets the
entry card; it has no way back, since deleting the entry would leave the
stay marked as taken in.
Starting a journal, changing its settings, linking or unlinking a trip,
bringing back its suggestions, the user's own view of it and deleting it
fell back to the generic card, which names ids and offers no way back.
They now share a journeyHeader card: the journal with its trips, the one
the change links or unlinks marked, what a new journal brings in and what
a deletion or an unlink takes, and each setting old to new in the
journal's own words. A new journal is deleted again while nothing was
written, added or laid out in it, changed settings are written back from
what the tool answered, a linked trip is unlinked by its owner while none
of its suggestions was written up, and the own switch flips back.
Asked on /journey to create "Japan 2026", the home chat had nothing to
offer: it keeps no write that names a trip or a journey. It now keeps
create_journey, whose trips the tool checks one by one, and delete_journey,
which the journey's own chat leaves out because the conversation goes
with it. For anything inside a journal it offers open_journey_chat, the
twin of open_trip_chat: the journal is read from the user's own journals
and shown as a card whose button goes on in the journal's chat with the
question waiting there. The trip and journal cards share their question
and button.
Asked on a journey page to write an entry, the assistant said it could not:
the only journey chat was the Studio's, which keeps the journey's writes
out. A journey session now holds what either page offers and shows the
model only the asking page's tools. The journal page offers a viewer the
reads, an editor the entry, photo, recorded day and day import writes, and
the owner also the journey's own settings and trips; delete_journey stays
out of the journey's own chat. A few reads that name no journey come along
(place search, the photo libraries, Dawarich stays), the Dawarich journal
entry is pinned through its journalId, and an entry an argument names must
belong to the conversation's journey before the tool runs. The turn reads a
journal card and instructions of its own; the Studio keeps exactly what it
offered before.
The journey's assistant lives in TREK Studio today, and the journal page is
about to get one of its own over the same threads. The chat and decision
bodies take an optional surface, studio or journal, and the journey scope
carries it to the service; left out it reads as the Studio, so today's
client is unchanged. The scope key ignores it, so both pages keep one
thread list per journey.
lint:test-sql counts every .prepare( under tests/ as a raw statement, and the
attachments service's prepare() was called five times in its unit test. The
test holds no SQL; the method takes a name that does not read as one.
The server built the answer through words.text(), words.count() and maps of
plain strings, so the check read all 13 assistant.attach keys as unused. The
helper's lookup is t() now (a numeric count picks the plural form, as the
client's does) and the chip and reason maps hold their keys under *Key
properties.
The chips offered to keep an attached file in Files, but nothing could do it:
the assistant's upload_trip_file stays denied, and the file never reached
Files. A new server write, keep_attached_file, is offered to the model where
the thread holds an attachment and held as a change card naming the file and
the booking, expense or place it goes with.
Applied, it runs the checks of a Files upload with the user's rights (the
demo block, the trip, file_upload and file_edit for a link, the instance's
file types, every linked id on the same trip), copies the bytes to a trip
file of their own, and writes the trip file, its expense link and the
attachment's pointer to it in one transaction; the members' planners get
file:created after it. A second keep of the same attachment answers the file
kept the first time. FilesService gains storeBytes and discardBytes so the
storage write stays out of that transaction.
The receipt opens the Files tab; the card's verb, its "Attached to" row and
the receipt's open label are new keys in all 27 locales. Undo is not offered:
no tool takes a trip file back yet.
The assistant could not read a file the user attached: the bytes were gone
after the request, the thread only held the file's name, and read_trip_file
stays denied because it would send any document of the trip to the provider.
A new server tool, read_attached_file, reads only the files attached in this
conversation by this user; any other id is answered "No such attachment in
this conversation.", whoever it belongs to. It is offered in a trip turn whose
thread holds a file with text. The first read of a file is held as a change
card that names the file and the model and host its text would go to; once
the user agrees, the agreement is kept on the attachment and later reads run
at once. The text reaches the model through the history and, for a later
read, as a tool answer only the model gets (AgentToolRun.modelText): no message
part, stream chunk, shared answer or client store ever carries it.
The model is told each kept file's attachment id, and the standing
instructions say to read the file when the user wants to work with it, to
propose a booking found in it through create_reservation and then to offer
keeping the file with it. The card's verb, its "Sent to" row and the step's
name are new keys in all 27 locales.
A booking file without a booking ended the answer on the error "No booking
found in this file." with no action, and the user was left to type. Now the
server answers every booking file itself, without a model: one line that
names the file and what TREK found in it, or that it found nothing, why, and
asks what to do with the file. Under it come up to three chips, picked from
what the user may do on the trip and what the file allows: keep it in Files,
make a booking from its text, attach it to a booking or an expense, ask what
is in it, or keep it with the booking just found. A chip sends its words as
an ordinary chat message.
The line and the chips are written in the language of the request from new
assistant.attach.* keys in all 27 locales, through a small helper the sports
proposals now share. The empty answer finishes as stop; no_bookings_found
stays in the contract for the threads older builds wrote. Route files keep
their answer as it was.
The chat's booking reader kept only the bookings of the import preview and
dropped its warnings, so "AI reading is not set up", "the AI failed",
"a scan without text" and "nothing in it" all came out as the same
no_bookings_found. The reader now reads the reason from the preview (the mode
it ran in and the AI fallback's warnings) and keeps it in the stored read
step, and the note the model is sent names it.
A file attached in the trip chat lived only as a buffer for the one request
that read it, so the assistant could not look at it again and nothing could
keep it in Files. Now its bytes go to storage before the turn writes anything
(under the trip files' category, which is never served by path) and a row in
the new assistant_attachments table keeps its name, type, size, hash and the
text TREK read out of it, written in the turn's opening transaction. The
message's attachment part carries the row's id; old messages without one still
read.
The text is read the way the AI import reads it, through a new public
LlmParseService.readText, and a file whose text cannot be read is kept all the
same. A storage that refuses the bytes refuses the turn before any row exists.
Deleting a thread takes its files along, and the nightly retention sweep drops
files older than seven days and bytes no row names any more.
Every style now has four covers and five double spreads, 108 in all, and the
spreads of each style take different photo counts, from a words-only page to
eight photos, so a book can be laid on them without an empty frame. The five
new styles come with their own themes: Riso (overprinted spot inks), Magazine
(a travel magazine), Retro travel (vintage posters), Midnight (a dark film
look) and Field notes (a traveller's notebook). Each new style has a cover
composer for the design engine and a recipe for the assistant, and its chip,
theme and template names are translated into every language.
The home chat kept every write whose arguments named no trip and no journey,
so the journey group brought update_journey_entry, delete_journey_entry and
reorder_journey_entry_photos (they name only the entry) and the Dawarich
journal entry (it names the journey as journalId). The writes of the
journey's controllers now stay out by their controller, and journalId counts
as a journey, as the home chat's own contract and prompt say.
The open changes of a thread are now the ones still inside their decision
window. A call older than a day already reads as expired in the thread and
can no longer be decided, but it stayed approval_requested until the nightly
sweep, so "+ Tour", "+ Day" and "Plan this" answered decide_first with
nothing left to decide, and a decision had to name the expired call too.
listOpen takes the window's start, and every caller passes the same cutoff.
- design_book takes a brief (energy, decoration, collages, theme), a cover
archetype and a cover lockup, and pages.overrides that give single pages of
the plan another beat or layout. Directives written before parse as they
did, and a design asked for none of it is composed exactly as before.
- The engine composes the seven cover archetypes from the book's own photos,
words, theme faces and colours, lockups, shapes and stickers, varied by the
seed: spine band, poster pop, cut-out, illustrated postcard, grid and
badge, minimal serif and stamp collage. A title the directives give a
display tier is held to it, and the notes say when a band, a badge or a
stamp holds it smaller or a cover leaves a layout out.
- A lively or loud book turns two in five hero and grid spreads into
collages, script overlays, badge grids, stamp rows or two-tone splits, the
first of them always, turns and frames the prints of its other pages, sets
its cover title larger and takes one or two, or three to six, stickers of
its theme's motifs a spread, never over words, a photo's detail or the
fold. A calm book asked for collages gets them and nothing else.
- Every turn the engine gives a print or a sticker stays within what the
brief's energy takes, a calm book's within its theme's own, so a design is
never judged against its own brief.
- New checks name a flat cover, a page without a lead, colours or faces off
the theme and, for a design made with a brief, a book livelier or calmer
than it asked, on the pages the design made and on a page revised in the
draft. Page edits answer with the art checks too, and the assistant's
cards name them in every language.
- list_book_assets lists the shapes, faces, lockups, themes and patterns
beside the stickers, and the Studio's shape groups and photo frames are one
list the assistant reads too.
- The Studio prompt gains an index of themes, archetypes and faces and JSON
recipes of the elements to add and style for each cover and three lively
spreads, and a turn that designs the book may preview sixteen pages, still
three images a request; a viewer is told of none of it.
A GPX file that mixes routes and tracks gives each its own text: the pieces
are matched to the reader's rows by kind, routes first as the reader gives
them, where they were paired by index and a route got a track's points. The
place pictures reach the network through a small repo instead of the api
module, the wanderer parity case and the tool catalog test allow wanderer,
and the knip and theme baselines follow the lower counts.
A third scope, home, next to trip and journey: threads that name neither a
trip nor a journey (the CHECK now only forbids both), reached under
/api/assistant by their owner only. A home turn reads the user's trips
through list_trips into a short card (under way, next, later, behind,
undated), offers every read (a trip's on the id the model names, checked by
the tool with the user's rights) and only the writes that take no trip or
journey (atlas, vacation planner, collections, tour library, settings), all
held as change cards. open_trip_chat answers with a card that goes on in
that trip's own chat with the question waiting in its composer.
On the desktop the window opens on the home chat outside a trip; its head
lists the trips to switch into and a switched trip chat has the way back.
On the phone the home chat opens full screen from the dashboard head and the
dock's More list, and the card opens the trip with its chat. Both run on
useAssistantHome.
P6: insert_tour_stop and remove_tour_stop take a place into a saved hike or
city walk as one stop and out again, each the other's undo, with change
cards in the chat. The planner's insertion rule (poiInsertion) moved to
shared so the planner and the server place a stop the same way. "+ Tour" on
a found place holds it as a stop of the saved tour open in the planner.
P7: GPX, KML and KMZ files attached in the chat are read into their tracks
and each is held as a change that takes it in as a hike through
import_tour_gpx, undoable per track; the GPX text stays on the held call.
The assistant can now search a round hike of about a wished length from one
point and the router's other ways for one leg of a hike, with the planner's
own search, and propose each one as a hike card to compare and take. It can
also search the user's wanderer trails, preview one as a card, import it and
check a wanderer tour against its source.
The round search, its loop geometry, the test whether two lines are one way,
the other ways among a router's offers and the pins that hold a chosen way
move to @trek/shared; the planner and the road trip use them from there.
City walk changes take edit, reverse and close_loop; edits alone keep the
saved way and never ask the router. update_tour turns a hike round or closes
its loop with reshape. New MCP tools mirror the Tours tab's REST writes and
reads: update_tour_details, set_tour_route_color, copy_tour,
classify_ski_track, set_paddle_plan, decide_paddle_obstacles,
restore_recorded_track, get_tour_trail_details and find_horse_routes, with
change cards, undo where the answer keeps the values before, and broadcast
parity cases.
A new element kind, sticker, draws layered vector illustrations from a library in twelve themes, behind its own entry point @trek/shared/stickers. Studio gets a Stickers rail section (themes, localized search, recently used, click or drag to place) and an inspector for colour slots, mirroring, turning and opacity. The assistant gets list_book_assets, add and style of stickers in edit_book_page, stickers in describe_book_page and an overlap warning for a sticker over words. BOOK_SCHEMA_REV goes to 2.
The book catalogue grows from 7 families to 34, all self-hosted through
@fontsource under the SIL OFL 1.1 or the Apache License 2.0: condensed and
poster faces, heavy retro display, elegant display serifs, Montserrat, twelve
scripts and DM Mono. Weights run from 100 to 900, each family offering only
the weights it ships, and the serifs gain real italics.
The catalogue lives in @trek/shared (book-fonts.data.ts) with each family's
roles (display, heading, body, script, mono), weights, italics and letter
width, so the Studio, the findings, the design engine and the assistant read
one list. Page operations and the design engine set a weight a family lacks
at its nearest one, describe_book_page names the role and the weight, and
font_mix no longer counts scripts and mono faces, warning above three.
The Studio gets StudioFontPicker: the current family set in itself, a
searchable list grouped by kind with every family in its own face and the
book's own families first, and a weight slider over the weights the family
has. Faces load only when used; the assistant's page snapshot embeds only the
faces a page draws, the service worker installs only the latin files of the
original seven families, and the faces of an opened book are cached for
offline use. BOOK_SCHEMA_REV goes to 2.
The registry refuses a package with more than two prompt lines at load, which
failed every suite that boots the assistant. The planning line now holds the
tools, the profiles and the hike rules.
The assistant can now plan and save every type the Tours tab plans. A bike
tour or a run is proposed and created through the hike's own plan and save
(propose_tour, create_tour), routed on the server with the costing of its
profile, now kept once in shared and read by the planner as well. A paddle
or ski tour planned with plan_paddle_route or plan_ski_route is saved with
create_paddle_tour or create_ski_tour, and "Plan this" takes those plans,
bike tours and runs too. Every create has its card and its way back.
- Web search: ChatGPT (Responses web_search), Anthropic (web_search server
tool) and OpenRouter (web plugin) search the web themselves when the admin
switch allows it (on as it ships). A search shows as a step ("Searching the
web for ..."), the cited pages as source chips under the answer. Anthropic
answers that searched are replayed with their blocks (new provider state
adapter); the search step never goes back to the model as a call. The
prompt says when to search and that chosen places still go through
search_place.
- Tour writes: assign_tour_track (undo: take that assignment off),
unassign_tour_track, set_tour_activity_excluded (undo: the other switch),
link_tour_packing and unlink_tour_packing (undo from the link before, which
both answers now name) get change cards and their ways back.
- Found places: the strip shows only when the answer names one of them, the
named ones first. A click on a card, or "Add to trip", opens the planner's
place form filled in with the place on the day in focus, its Google details
(photos, hours, description) loading beside it; "+ Day" and the day grid
are gone.
- Trip notes: remember_for_trip runs at once without a card. The assistant
keeps lasting facts the user mentions (who travels, needs, likes, budget,
pace) on its own and offers the unsure ones as a follow-up; the chat shows
"Noted for this trip: ..." with the notes a tap away, and reads them again
so the dialog never saves over the new note.
- City walks: set_city_walk_visit checks a station off on its day (visited,
skipped, or the mark taken back), with day_edit as in the day plan, a
change card naming the station and its walk, and an undo that restores the
status before while the station is as the change left it.
- Card heads are a step smaller (32px tile, smaller title and count), and the
dot beside "Working on it" in the header is gone.
- propose_hike, create_hike and update_tour take stage_end on the stops a day
ends at, checked by the tours' own stage rules (never at the start or the
end, at most 60 stages) on REST and MCP alike.
- A proposal over several days names its stage ends and sets no single day's
schedule; its card counts the stages and marks the stops they end at.
- New points for a hike in stages keep the keys of the stops they stand on,
so a stage keeps its day; they are refused only where they would drop every
stage end.
- The tours group and the prompt name the stages, how to put each on its day,
and what to offer once a tour is planned: places along it, its packing, the
GPX file.
- The design prompt keeps the print rules and adds energy: calm keeps today's
rules, lively and loud layer photos, tilt, set oversize display type and
block colour. Type sizes come as numbers relative to the page, and the model
answers eight silent questions and previews the cover twice before it
proposes.
- The type scale reaches 2.5 and a display tier sets a cover title up to
200 pt; existing books keep their sizes.
- An overlap is severe only where more than half of a photo or a text is
covered by accident; a title over a photo, a framed inset or a turned photo
is layering.
- The assistant may use every icon of the icon tab, checked by name, and every
enum of the book tools says what its values mean.
A turn may end on suggest_follow_ups: the server answers the call itself, keeps
its questions as data-trek-follow-ups and the chat shows them as chips under
the answer. Without them the chips fall back to what the answer holds, now by
the kind of change: a tour asks about the tour, the day plan's questions only
follow a change to the day plan, and an answer that asks the user something
offers no "tell me more".
The Studio drew the page after a held edit or cover change by running the
card's operations on its own book, which knows nothing of the journey's
photos, words and figures, so a caption from a photo, a text from an entry
or a map from the route came out differently or not at all. The card now
carries the page the server's trial made with the full context, read with
the book's schema and within the card's size, and the Studio draws it while
its book stands at the version the trial ran on. A book that moved on, a
card from before and a page too heavy to carry still run the operations on
the book on screen, the book the change will run on. The receipt leaves the
page out, since an applied change is drawn from the book itself.
A page preview draws the spread the server applied with the journey's photos,
texts and figures instead of applying the operations again without them. The
Studio's welcome, privacy banner and header fit a journey chat: a viewer is
told they can only look, the banner speaks of the book, and the trip notes
button stays in the trip chat. Escape closes the header menu before it stops a
reply, a single page renders at the full long edge, the status says a model
reads pages only while its provider has not refused an image, and pointing at
a page in a design card outlines it in the Studio.
Changes held in one step are applied in the order the chat shows them, so
"undo all" walks them back newest first. A page put back is checked against
the digest of the page as the change left it, so a save to another page no
longer blocks the undo; restore_book_page takes these digests as an option and
keeps the whole book check without them. The book card counts the user's own
photos, the preview tool describes itself per role, a copy into another trip
names no places, a held vacay toggle says up front whether it can be taken
back, and a viewer's Studio prompt names none of the design's tools.
A viewer applying a design gets 403 and an outsider 404 before the draft is
read. The design tools' rights are pinned for outsiders and viewers, the apply
of a new book's pages is covered branch by branch, and a viewer's journey
session no longer offers the design's reads. The colour signals and the grey
statistics drop fallbacks that could never run, and the palette keeps one
fixed order for equal shares.
Maurice