9line.git, branch master

pim: a window for each kind of thing, and a picker for dates

2026-08-22T23:12:15+00:00

cal9 kept an event in a panel and a day in another, with the state that implies: a selection, a picked day, rectangles for the buttons on it, and a hundred-odd lines to draw it. Opening a window instead deletes all of that rather than tidying it away. Clicking an event opens cal9 -v event, clicking a day opens cal9 -v day, and cal9 is three grids again. The event window earns being cal9 rather than pim/showwin: it carries Accept, Tentative and Declined, lit to show where you stand, and only when the calendar's ctl says it can rsvp. cal9 -v invite is a form that writes the calendar's new file in the tree's key: value form, so the composer knows nothing about iCalendar either. datepick is a program rather than a widget so anything can use it: pim/agenda -d `{datepick} -t asks for a time as well, in stages, because a grid of ninety-six five-minute slots is not something you want to aim at. A window started from wctl has nowhere useful to print, so -o and -s hand the answer back by file or by snarf; the invite form spawns it that way and fills the field when it returns. An event you have not answered draws pale, so a glance at the week says what is still waiting. Move and Resize are gone from the menus: rio owns the border and does both, better. Compact is one item showing whichever state you are not in. Co-Authored-By: Claude Opus 5

gui: a widget layer, so the next program does not write one

2026-08-22T23:12:01+00:00

cal9 had grown its own: borders drawn four fills at a time in two places, a button-3 menu, hit regions, buttons, text wrapping. None of that is about calendars, and a second program would have written it again, differently. It lives in gui/ui now. It is a kit and not a framework. Every call draws a thing and, where it makes sense, records where it drew it; the program keeps its own event loop and its own idea of what to redraw. Menu takes a redraw function rather than calling back into anything. The moment the library owns the loop it stops being usable by the next program, which was the problem being fixed. Body is the fix for two bugs that turned out to be one. rio draws a window's border into the client's own image and puts corner cursors in it, taking those clicks before the client sees them: the border is rio's resize handle. Filling the whole rectangle erased it, so the border vanished and so did resizing. Drawing inside Body gives both back, and the drag-to-resize cal9 had grown -- which could only ever shrink, since a window stops getting mouse events once the pointer leaves it -- was deleted rather than fixed. Field is one line of editable text with a cursor, backspace, ^U and a paste from /dev/snarf. No selection, no mouse placement, no history: a form asking for a summary and a time needs none of them and each is somewhere to be subtly wrong. Co-Authored-By: Claude Opus 5

ignore the binaries go build leaves at the module root

2026-08-22T21:01:16+00:00

"go build ./cmd/x" writes the binary to the current directory, not to bin/, so pim/caldavfs and gui/spike were showing up as new files a stray "git add ." would have committed. Anchored with a leading slash so a source file of the same name deeper in the tree still counts. Co-Authored-By: Claude Opus 5

ignore sessions/

2026-08-22T20:48:52+00:00

Transcripts are not part of the source. They were in the initial commit and have been removed from the history rather than deleted in a later one, so a clone no longer carries them at all. Co-Authored-By: Claude Opus 5

pim: invite, and a way to say what you mean

2026-08-22T20:48:07+00:00

pim/invite makes an event and, when it names attendees, sends the invitation -- which on a caldav server with auto-schedule is the same act. -p writes the object to standard output instead, so it can go to upas/marshal as iMIP, into a directory an ical/fs is serving, or nowhere in particular. There is no separate verb for inviting because to a calendar there is no separate thing. It writes the tree's own key: value form rather than icalendar. The first version spelled SUMMARY: by hand and got "lunch, then a walk" wrong: a comma is a separator there and wants escaping. Parsing with go-ical and generating by hand is the wrong asymmetry, so calfs grew Compose, and now the same shape the tree emits is the shape it accepts. The one UID we ever mint -- as organiser we own the event -- is minted there too. pim/invites lists what is waiting for an answer, which is what makes pim/rsvp usable: it walks the date tree rather than every event, since an invitation you never answered last March is not news. agenda and show take -d date now, via seconds(1), which parses a human date where date(1) only formats one. Co-Authored-By: Claude Opus 5

pim: answer an invitation by writing to a file

2026-08-22T18:14:59+00:00

echo ACCEPTED >events/uuid//partstat now does it, on a backend that can. caldav/fs fetches the object, rewrites only our own ATTENDEE line and PUTs it back with If-Match; fastmail advertises calendar-auto-schedule, so that single PUT both updates the copy and sends the iMIP to the organiser. caps says read rsvp schedule and means all three. Two things that cost an hour between them. go-webdav strips the quotes off an ETag when it parses the header, but If-Match wants an entity-tag: a bare hex string gets 412, which reads exactly like losing a race to another client. And the tree kept serving the old answer right after a successful write, because RSVP resynced the backend without rebuilding anything; a file that lies immediately after you wrote it is worse than one that is slow. -R uid:PARTSTAT answers one invitation without serving, which is what a pim/rsvp wrapper wants to call. Co-Authored-By: Claude Opus 5

pim: a caldav backend, and a partstat that refuses to lie

2026-08-22T18:04:17+00:00

caldav/fs is the second backend, and it cost ~330 lines: discovery, a report for the window, ETags to decide whether anything moved. The event model, recurrence expansion, the tree, ctl, query, alarm and changed all came from lib/cal untouched, which is what the split was for. go-webdav hands back *ical.Calendar from the same library lib/cal parses with, so FromCalendars takes it straight in. Two interfaces, both optional. Identity says whose calendar this is, because in iTIP your identity is the mailto: in your own ATTENDEE line and nothing else can tell which attendee is you. RSVPer says the backend can answer an invitation -- meaning both halves of it, updating your copy and telling the organiser. A backend that can only do one should not implement it, so partstat is 0444 on a published .ics and writing to it fails rather than half-working. Two bugs worth naming. go-webdav resolves paths with path.Join, which drops a trailing slash: pointed at /dav/ it asks about /dav, and Cyrus answers 405 for that spelling and 207 for the other, so the first PROPFIND is done by hand. And an absolute DAV href replaces the endpoint's path rather than extending it; appending gave /dav/dav/calendars/..., a 404, and a probe reading 404 as "no" hid calendar-auto-schedule, which fastmail does in fact offer. pimup brings mail and the calendars up and mounts them; riostart runs it before opening any window, since a mount only exists in the namespace that made it. rio does not run riostart by itself: it is the argument to -i. Co-Authored-By: Claude Opus 5

pim: personal information management, starting with a calendar

2026-08-22T17:16:49+00:00

A calendar as a file tree, and tools that only know the tree: events/date/yyyy/mm/dd/hhmm-summary as lived events/uuid// as stored ctl query alarm changed lib/cal owns all of that. A backend supplies events and, where its protocol allows, takes changes back -- six methods. cmd/icalfs is the first: it reads .ics files from a directory and nothing else, because fetching is rc/fetch's job and hget already exists. That keeps net/http out of the binary and makes a subscribed calendar and a local one the same thing. The tools are rc on purpose. If the tree needs a compiled program to be useful, the tree is the wrong shape. Three things were added to the tree because the rc port needed them: a path from an occurrence to its event, epoch seconds beside RFC3339, and a numeric slot for all-day events so test(1) can compare it. ctl reports caps, so a tool can say "read only" instead of trying and failing. A published .ics is read only: nowhere to PUT, and no METHOD:REQUEST to reply to. CalDAV would be read write rsvp schedule, and that is the next backend. Co-Authored-By: Claude Opus 5

gui: a /dev/draw layer in go, with no cgo and no devdraw

2026-08-22T17:16:35+00:00

9fans.net/go/draw builds for plan9 but shells out to plan9port's devdraw, which 9front does not have. This talks to /dev/draw itself: the protocol is file i/o, so a pure-go client is a few hundred lines under an already-complete idea. draw/keyboard.go opens /dev/cons and turns the console raw, as libdraw's initkeyboard does. Without it rio keeps its line editor on the window and paints what you type over the drawing. draw/snarf.go writes /dev/snarf, which is what every other program on the system means by copy. Co-Authored-By: Claude Opus 5

doc: 91 checks, and a wire

2026-08-19T17:00:43+00:00

Co-Authored-By: Claude Opus 5