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.
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.
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.
- 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.
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 draws the page on request and uploads it once, each provider gets
the image where it expects it, and a refusal falls back to the text
description. No image is stored.
The Anthropic request marked only the instructions and the trip card for
the prompt cache, so every step of a turn paid the whole history again.
It now also marks the end of the history, which the next step starts
from, and the end of what came before the user's question, which every
step of the turn sends unchanged, within the four marks a request takes.
Every provider now reports how much of the input it read from its cache
(Anthropic cache_read_input_tokens, OpenAI cached_tokens), the turn adds
it up, and the usage log line says it, so the saving can be measured.
The three AI addons share one tile in the integration column; a dialog holds
the connection, the assistant's limits and AI parsing, which now reads with
the assistant's connection unless an install keeps its own. Existing parsing
configs stay where they are and keep winning.
The chat sees the map on screen, offers next questions under an answer,
shows its live step in the window head, previews waiting changes inside the
day list, checks whether a day still works, folds a multi-day plan into a
week grid, keeps per-trip notes and can estimate costs. An empty or cut-off
answer gets one push instead of ending silently, and the assistant's own
step and size limits are raised.
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