The assistant sent an own server's chat to the address as entered plus
/chat/completions. Ollama, entered as http://host:11434 the way its own docs
write it, answers that with a 404, so every message ended in "something went
wrong" while connecting had worked: the model list is found under the root
at /api/tags. A bare server address now chats under /v1, where Ollama, LM
Studio, vLLM, llama.cpp and LiteLLM all answer; an address with a path, such
as https://api.groq.com/openai/v1, is used as entered.
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
Connecting an own server asked Ollama's /api/tags and gave up on any answer
that was not a 404, and every failure, a refused key included, read as "not
responding". The model list is now looked for at /api/tags, at the address as
entered and at /v1/models, an answer only counts when it is a list, and a
refused key or a server that answered without a list say so (#2808).
AI parsing through an own server always called Ollama's /api/chat, which an
OpenAI-compatible server such as Groq answers with 404. A server without it
now gets the same extraction over /chat/completions at the address as
entered (#2806).
gpt-5.6-sol, gpt-6.1-sol and the other reasoning models refuse function tools
on Chat Completions, so every assistant message on such a model came back as
HTTP 400. An OpenAI connection whose model is from the o-series or GPT-5 and
later now goes through the Responses API, the client a ChatGPT plan already
uses, with the key as its bearer token and no token refresh on a 401. Older
models and the -chat variants stay on Chat Completions.
Closes#2792
Sonnet 5.5 and Opus 5.5 think by default and write what they say between tool
calls into thinking blocks, which they want back unchanged with the rest of
the turn. The client dropped those blocks (and their signatures) and rebuilt
the turn from its text and calls, so after a tool result the next step came
back empty and the assistant gave up with empty_answer. A turn that thought
is now replayed block by block like one that searched the web; the cache mark
before the question skips a thinking block, and replayed blocks are copies so
a mark never sticks to the stored answer.
Closes#2789
The model picker in Admin > Addons did nothing for a ChatGPT connection: the
only way to change the instance's model was saving the connection again, which
needs an API key a plan does not have. A PATCH /api/admin/ai/connection/model
now switches it on its own, to a model the provider offers.
Closes#2783
The output items of a step were replayed as they came, web search calls
included. With store false the provider cannot resolve such a call again and
refused the next step with HTTP 400, which ended the turn with an error right
after a hike was proposed. The search calls are left out of the replay; what
they found is in the answer's text and its sources.
- 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 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.
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.
- Tests that still knew the ChatGPT sign-in as personal only follow the
admin's shared plan: the service takes the caller with their role, the
admin tiles show ChatGPT, the controller tests cover the admin routes.
- The table refused an instance ChatGPT connection with a CHECK, so
sharing the plan failed. A new migration rebuilds ai_connections
without that check; the shipped migration stays as it was.
- useChatgptConnect grew past the function size limit: reading the
clipboard and noticing the return to the tab are helpers of their own.
- The file link factory hands its overrides over apart from the row,
like every other factory, which ends the deep type instantiation.
- Knip: exports only their own file used are private again, and the
booking cap reads MAX_WRITES_PER_STEP instead of an alias.
- lint:i18n-keys reaches the tools' own action names through a key
builder and the connection hints through literal keys.
- MMehrSheet stays in the assistant hints pair only.
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 connection for everyone can now also be the admin's own ChatGPT sign-in.
Whether a plan may be shared that way is the self-hoster's call under their
plan and OpenAI's terms; the form says what it means before the sign-in.
The sign-in starts on the admin route and finishes on the shared one, which
refuses a sign-in for everyone to anyone who is not an admin. The ChatGPT
return reads the clipboard only once the paste handler exists.
OpenAI offers a fixed web redirect for Sign in with ChatGPT only to selected
commercial partners; open-source apps sign in through the 127.0.0.1 loopback
callback. The relay could never be switched on, so it goes again.
Once the TREK project has a client id registered with OpenAI, the sign-in
goes out with the relay page as its redirect address and the install's own
settings in its state; the relay sends the browser back there and the
settings finish the sign-in, nothing to paste. The relay sees the code but
cannot redeem it without the install's PKCE verifier. Without the project's
client id, or for an install without an https address, the loopback page
and the pasted address stay as they are.
relay/chatgpt-callback/index.html is the page to publish.
- the assistant keeps facts for a trip (remember_for_trip) as a change card,
applied into the user's trip notes; existing expenses can be changed
- change card fields as a few icon chips with the rest folded under Details,
the day band marks only what changes, the week grid as compact tiles
- no map button on receipts of things that have no place on the map
- tests follow the composer's accessible name, the receipt's status icon and
the GL map's container; the hanging stream helper honours an early abort
- Postgres probe: qualified columns in the upsert updates; entities restored
to the generator's output; strict ratchet reads the SDK client as untyped
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.