Setup only — do NOT build the chat yet, that is the next task.

Write AGENTS.md at the root of the project with three declared lines,
each at the start of its own line:

stack: <language, framework, storage> e.g. stack: TypeScript, SvelteKit, SQLite
run: e.g. run: npm start
test: e.g. test: npm test

And say, in your own words: why that stack for a chat product with
streaming replies, several model providers and several people; how the
code is laid out; and how you work — where the plan lives, when you
update it, what you treat as done.

Then write TODO.md: the plan for the chat as markdown checkboxes, in
product steps rather than tooling — "reply streams word by word", not
"install the linter". Keep it honest as you go; the next tasks are judged
partly on how you worked.

Finally, make the declared commands real: run: starts something,
test: runs a suite. An empty suite that passes is fine for now.

Build the chat: a person types, the model answers, and the answer appears
word by word. You decide when it is done; a judge panel scores the result.

For now the product talks to one OpenAI-compatible server on the Chat
Completions protocol (POST <address>/chat/completions, streamed). Where
its address, key and model come from is yours — a config file or the
environment is fine. Settings with several providers are the next task.

Scenarios:

Scenario: The reply streams
When someone sends a message
Then the reply appears on screen while it is being written, not once it is complete
And the finished reply stays in the conversation

Scenario: The conversation is remembered
When someone sends a second message
Then the model receives every earlier turn, in order
And after the product restarts, the conversation reads back the same

Scenario: Many conversations
When someone starts a new conversation
Then it appears in a list next to the earlier ones, with a title
And opening an earlier one shows its messages, and continues it

Scenario: The provider fails
When the provider answers an error, or cannot be reached
Then the chat says so where the reply would have been
And the conversation is unchanged, and the next message still works

Everything beyond the scenarios is yours: what a new chat opens with, how
a title is chosen, what the page looks like while a reply is on its way.
Keep TODO.md honest as you go — it is part of how this task is judged.

When you are done, write .ololo/chatgpt-chat-done.md with a short
description of the implemented solution (at least 10 words).