Repository navigation
Conversation
added 5 commits
September 30, 2026 16:41
`:models` in the command box printed the table `figaro models` prints, with os.Stdout: round the pager's writer, onto the terminal. Nothing on it could be acted on. It is now an itemised live view (itemView, like the form and the queue), so the pit's one picker drives it: ^N/^P and j/k move, the wash and the marker show the choice, `y` yanks the model id, `/` searches, G and gg reach the ends of three hundred models. The subject's current model carries a ●, and each row carries its provider. ENTER SETS THE MODEL, and its provider with it, in one patch: system.model to the chosen id and system.provider to the provider that listed it. A model is only meaningful to its own provider, and choosing copilot's model while the board still said anthropic would send a name Anthropic has never heard of on the next turn. When the provider already matches, the store drops the unchanged key and the write is exactly `figaro set system.model <id>`. The bar says what took; the ● moves. Activate hands off to a goroutine: it runs under the render lock and the set dials. `/` could not reach a provider: the heading is chrome and search walks selectable rows. The picker now searches a row's note as well as its text -- what a reader sees is what `/` finds -- which also makes the tab pit's descriptions searchable. The standalone `figaro models` shares the fetch and now writes to the package's stdout, so it can never paint over a pager again. Verified with the flake build in an isolated .#sandbox shell, real credentials: :models lists every configured provider's models, Enter on claude-opus-5-5 set it (figaro state agrees) and the next turn was answered by it; Enter on a copilot model set model and provider together and the next turn was answered through copilot; Enter back on anthropic restored both.
…g it
Everything the pager said went through one slot, the first cell of the
status bar, which retires after notice_ttl. After that it existed nowhere
but one lastError string and the log file. A `𝄞` glyph for a notifications
pit had been reserved since August, and nothing opened it.
Now every alert is POSTED to a history first, and the bar and a pit are two
views of it:
the ALERT the newest one, in the bar's first slot, retiring as before
the MARK 𝄞 N while warnings or errors are unread, red or yellow by the
worst of them; it is what stays when the alert retires
the PIT `space n` or `:notifications`: the history, newest first, as
the same picker every pit is
A row is time, level (✗ ! ·), source (cli or the aria it concerns) and text,
with ×N when it repeated: the same source, level and text, digits folded,
coalesce into one row that moves to the top and is unread again, as Gluck's
Neovim msg_router coalesces. Enter expands a row with what it could not hold;
`a` attends the row's aria; `y` yanks it; `f` cycles all, warnings and up,
errors; `/` searches. Opening it reads it. `:notifications clear` empties the
pager's copy, as Vim's `:messages clear` does.
Space is a leader: `space n` is `<leader>n`, which is where Gluck's Neovim
opens its notification history. n stays search-repeat. A leader arms and
waits, so space no longer falls through to a panel's any-key dismissal: the
keymap oracles record that, dated.
What is posted: every verb result and refusal (setCommandNoteAt, about 60
sites, levels as they were), every turn report for the aria on screen
(reportAt, with the aria as the source), and the pager process's own slog
WARN and ERROR (a freeze dump written, templates that failed to load,
messages dropped on exit), minus reportAt's echo of every alert. It lives on
the TRANSCRIPT, in the pager's memory: a hop mints a new session status, and
setStatus binds the same history to each.
Found on the way: `[cli] notice_ttl` was documented and never applied (nothing
called setNoticeTTL). It is now, through setStatus, so a hop keeps it.
Phase 2, the daemon's own trouble and turns finishing on arias this session
has shown, is designed in plans/notifications.md and not built here.
scripts/notifications-pty.sh drives the real binary in tmux, hermetically:
14/14 on this build, 5/14 on 0.37.11 (the five are absences).
# Conflicts: # internal/cli/command_router.go
…ry's Merging the two pits showed it: the models pit says "…models" while the list loads, and every routed verb says "…verb" while it runs. With every note now kept, each :ls became two history rows, one a progress indicator. Progress goes to the bar alone (setProgressNote / in.progress); the result that follows is the news.
This was referenced Sep 30, 2026
Two polish passes on the notifications pit, both asked for after living with it (Gluck, 2026-09-30). ENTER. It used to print the full date and the source under a row, which is furniture: the row already carries the time and the aria. It now spells the notification OUT, the way the state pit's Enter opens a value (formView.Activate -> openBranch): the whole text, wrapped to the pane, every line selectable and yanking the whole, so a long single line can be read rather than clipped at the pane edge. Bounded at 200 lines, as an opened form value is at 500, with the same "y yanks all of it". THE LEVEL IS A COLOUR, not only a glyph. It cannot ride in the row's text, which goes through pitText and loses every escape, so a row carries a tone (pitRow.tone) and the picker paints it: the palette's own red for an error, its yellow for a warning, nothing for info. Over the selection wash each is the same colour lifted (term.NoticeOnWash), because autumnRed on mid-grey reads at 1.9:1. A pit that sets no tone renders byte for byte as it did before, which is asserted rather than assumed. Tests, red first: two view cases (the whole text wrapped at width 60, and no date or source anywhere in it) and one picker case (wash plus lifted red on the selected row, the palette's red on an unselected one, an untoned row unchanged byte for byte, no escape at all with colour off). scripts/notifications-pty.sh grows four checks, each of which fails on 654d999 and passes here, plus the premise they rest on: the row really cannot hold the text.
Owner
Author
|
Two polish passes on the notifications pit, pushed as 027fc9c. Enter spells the notification out. It used to print the full date and the source under the row, which is furniture (the row already carries the time and the aria). It now opens the whole text beneath the row, wrapped to the pane, following the state pit's Enter rather than a parallel mechanism: every opened line is selectable and yanks the whole, so a long single line is readable instead of clipped at the pane edge. Bounded at 200 lines with the same "y yanks all of it" the form uses at 500. The level is a colour. A row now carries a tone (
One thing seen while looking and NOT touched: a verb error like |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Both branches merged into one, so there is one thing to validate. Worktree:
~/dev/figaro-qua/pits.space n/:notifications, the𝄞 Nunread mark, and the history pit.:modelspit. Enter setssystem.modelandsystem.provider.The merge
caseto the pager's own verbs in the same place. Both are kept.in.note) and failures (in.noteErr) are now recorded in notifications at the right level.Found by integrating, and fixed (
654d9994): the models pit shows "…models" while it loads, and every routed verb shows "…verb" while it runs. With every note kept, each:lsbecame two history rows, one of them a progress indicator. Progress now goes to the bar only (setProgressNote/in.progress), and the result that follows is what's recorded. #34 had this flaw on its own too.Verified on the flake build of the merged commit
notifications-pty.sh: 14/14.tabpit-pty.sh: 12/12..#sandboxdevshell, real credentials)::models→/claude-sonnet-5-5→ Enter. The bar confirmed it,space nshowed exactly one row for the switch, and the next turn answered (200) with the board onclaude-sonnet-5-5.