v0.55.0

Release Notes — v0.55.0

Major: Read a spreadsheet without reading it

A workbook of forty thousand rows was a blind spot: the indexer skipped it, and the
only way at the numbers was opening the file. Spreadsheets are now documents ogcode
reads like any other. .xlsx, .xlsm, .csv and .tsv are profiled by a new
internal/sheet package — column types (integer, number, date, boolean, text)
resolved from a lattice over the first thousand rows, sample values and ranges taken
from a stride of later ones, the result kept inside a fixed profile budget instead
of the whole file. Each sheet becomes its own page in the project index, so a
three-sheet workbook shows three labelled sheets rather than one line carrying only
the last of them.

That profile is what makes the new query_sheet tool cheap. It loads the sheet into
an in-memory SQLite table typed from the profile, then runs a single read-only
SELECT: more than one statement is refused, so is anything that writes, queries
stop at three seconds, output stops at 200 rows, and the connection goes
query_only once the rows are in. Ask for "total spend per vendor, top ten" and you
get a table back instead of forty thousand lines of context. Plain-delimited files
above 256 MB are refused with a pointer to the workbook form, which streams.

Minor: The model catalogue syncs itself

Ogcode ships 162 models, and a new one could not reach an installed copy without a
release. The compiled-in list now sits on top of a shared JSON catalogue the server
fetches itself: at boot, once a day afterwards, and on demand from a card in
Settings → General that shows when it last landed and what the sync changed. Rows a
user has edited still outrank it, ids the catalogue does not know arrive placed by
collection, implausible entries are skipped rather than merged, and a failed fetch
keeps whatever was already there. Point it elsewhere with OGCODE_CATALOG_URL, or
turn syncing off with OGCODE_CATALOG_SYNC=false.

Minor: Every model row is editable

A model no catalogue answers for was all guesswork: no window, no price, no output
ceiling. Worse, a catalogue that guessed wrong could not be corrected at all. Any
row — one you added, one a provider lists, one in the catalogue — now carries your
own context window, output ceiling, price per million input and output tokens, cache
read and write prices, and whether it takes images. Migrations 070 through 072 hold
them, they survive a restart, and they layer over the listing rather than replacing
it, so an edited catalogue model keeps its name, collection and provider routing. A
window you set is now authoritative: a listing is often a heuristic, and a real
overflow still self-corrects from evidence. The editor rides a pencil on each row, a
catalogue figure sits in the field as its placeholder, and an "edited" tag marks a
listed row carrying your numbers.

Minor: OGX is billed at the gateway's own rates

OGX used to read as "included in your plan": its tokens were counted but never
priced. The gateway publishes rates per model, and those are now what the usage page
bills at. The vendor catalogue figure becomes a yardstick only, an unpublished rate
shows as unpriced instead of free, and a price you set yourself still applies, cache
rates included.

Minor: Tool output is there when you want it

Every tool card in the chat opens now, the agent-navigation ones included, so an
audit of what file_map or update_intent actually returned is one click away.
pdf_index and docx_index render their answer as a page list — page number, then
that page's labels — instead of a wall of quoted JSON, and the file viewer's index
entry lists every page's labels with no truncation.