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.
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.
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.
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 card for a tool without one of its own names what it works on: the place
of an assignment, the day, the note, the booking; its arguments sit under
the labels the other cards use, a new order reads as the names in order.
28 frequent tools carry their own action name in every language.
- The server's own steps without a card still refuse, as before.
- Tests: the rate limit is asked one request after another, the mention list
closes on the caret move a browser reports, the place factory hands its
category over as a plain id, eleven files formatted.
The assistant offered 25 hand-picked tools. It now offers every tool the MCP
registry holds, but a short denylist with a reason per tool (deleting the
trip, share, invite and feed links, membership changes, fetching an address
the model picks, documents to the provider, instance-wide categories and
holidays). The scopes widen to every addon but trips:delete, the share
scopes, file contents and plugins.
- A tool annotated read-only runs at once, every other one waits as a change
card; reads that fetch a planted address wait too.
- A turn starts with the trip's core tools; use_addon_tools loads tours, road
trip, journey, vacay, atlas, collections, trip chat, files, account or
plugin tools for the rest of the turn.
- A writing tool without a card of its own gets one from its arguments:
add, change or remove, the part of TREK it works in, the trip's days,
places and people named.
- The boot check fails on a tool that shares, invites or publishes and is
neither denied nor known to only take back.
- agent loop and decisions without await in loops: steps as recursion,
calls and applied changes in order through one promise chain, the
independent reads and claims in parallel; part mapping in its own helpers
- place card, header and weather strip without refs or components created
in render; long hooks and components split; state reset during render
- trip notes only in a turn's tool session, never an undo's; map view
reported only while the assistant can run
- place strip scrolls through the strip's own smooth scroll, no physical side
- tests follow the raised limits and the new card fields; toasts read through
a mocked useToast instead of a window global; baselines lowered where the
counts fell
Adds an in-app assistant that reads and plans the open trip. Users connect
their own provider (Sign in with ChatGPT, or an Anthropic, OpenAI, Gemini or
OpenRouter key) or use the one the admin set up, including a custom
OpenAI-compatible server. It is an addon and only shows while a connection
is usable.
The chat runs on the existing MCP tools, pinned to the trip. Every change is
held as a change card the user applies or discards, and applied changes can
be undone. On the desktop it is a floating window that can be moved and
resized; on the phone it is fullscreen. Search results show as place cards
that can go onto a day, weather lookups as a strip, proposed places as
ghost markers on the map with a before/after switch, and an answer can be
shared into the trip chat. Booking files can be dropped in and are read
into change cards.
Maurice