Gronk 0.4.0

Two things happened in this release. Sessions stopped being one-at-a-time, and the
transcript stopped living in one file that got slower the bigger it grew. Most of the rest
follows from those.

Before you install this

Do not go back to 0.3.0 or earlier after running this. Conversations move out of
gronk-store.json into one file each, and the store is rewritten without them. Older
versions read the conversation out of the store, so they would show every session as empty.
Nothing is destroyed — the files are still there and this version reads them — but anything
said while downgraded is written where this version no longer looks.

Added

Sessions keep running when you look at another one. Clicking a different session used to
end the work you walked away from. Several can now run at once, the sidebar marks which are
working and which are waiting on you, and each row can be stopped on its own. Coming back to
one shows what it actually says rather than what was half-drawn when you left.

A message typed while the agent is working is queued instead of refused. Send used to be
disabled for the whole turn, so a finished message sat in the box waiting for an end you
cannot see coming. Queued messages are shown in full, each cancellable, and they go in order
as turns finish. Stopping a turn holds them rather than releasing the next one — stopping
usually means you want to say something different — and the button says so.

What you have typed stays with the conversation you typed it in. Leaving the conversation
view used to throw an unsent message away, and switching sessions carried it into the next
conversation, one Enter from the wrong agent. Drafts are kept per conversation, including one
typed before the agent has finished starting. They do not survive quitting the app.

A Changes panel showing what the agent altered in the open folder, and project notes —
plain text, one per folder, kept with Gronk's own settings rather than in the project.

Changed

The transcript store no longer grows without bound. Conversations live in one file each,
so opening a session reads that session rather than everything ever said. A real store had
reached 120 MB, most of it duplicated tool calls re-appended on every reopen, and parsing it
froze the window for about a minute on every read. The duplicates are removed on first launch
and cannot re-accumulate.

The store is read once per operation rather than eighteen times. A single turn was
re-reading and rewriting the whole file repeatedly; the permission audit now has its own file.

Chat and Build have the same shape, and the sidebar says what each one is for. Archived
sessions sit with the sessions rather than in the app footer.

Fixed

A transcript save cannot replace a conversation with a shorter, different one. This is the
fix for real data loss: reopening a session after a restart could overwrite its history with a
partial copy. A save that would drop stored messages is refused, and the surviving history is
kept.

Attached images survive relocating the data folder, and parked attachments are collected
once nothing refers to them — with the collector refusing to act when it cannot be sure.

A stalled browser sign-in no longer disables device-code login. The wait switched off the
exact remedy the screen recommends.

A rejected agent call says which call failed. The CLI reports some failures — including a
spent weekly plan quota — as a bare JSON-RPC error whose standard name is "Internal error",
and that is all the banner used to show. It now names the call and the code, and keeps the
agent's own reason when there is one.

Correction, added after release. This shipped saying a spent weekly plan quota is
reported that way, and the app suggests checking your usage limits when it happens. That
is wrong. The Grok CLI reports rate limits on their own code (-32003, HTTP 429) with
its own copy about upgrading, and classifies -32603 as a server error specifically
not a rate limit. The hint points at the one thing the CLI would already have told you
about clearly. Fixed on main; the wording inside 0.4.0 itself is misleading.

A turn that fails before the agent says anything no longer leaves a blank message in the
conversation, or on disk.

Reading an older part of a conversation is no longer interrupted by the restore finishing.

The mode and model pickers describe the session in front of you rather than what the next
one would start with, and choosing the option already in force no longer restarts the agent —
which used to replace the conversation with an empty session.

Events are attributed to the session they belong to, so a background session cannot
narrate over the one on screen. The account label never surfaces an email address.

---

Verify a download against SHA256SUMS.txt before running it.