Skip to content

D92 — shell: a meeting made after a stop is told once, naming it, and only where no window of Kitokey has the focus - #824

Open
kitokey-marc[bot] wants to merge 22 commits into
mainfrom
d92-one-notification-when-the-meeting-is-made
Open

kitokey-marc[bot] wants to merge 22 commits into
mainfrom
d92-one-notification-when-the-meeting-is-made

Conversation

@kitokey-marc

@kitokey-marc kitokey-marc Bot commented Nov 14, 2026 •

Copy link
Copy Markdown
Contributor

D92 — shell: a meeting made after a stop is told once, naming it, and only where no window of Kitokey has the focus

Refs #769 — D92 closes once its four criteria hold; this pull request takes criterion 3.


Module

modules/shell — exactly one.

  • This pull request touches a single module
  • ⚠️ Cross-module — the cross-module label is required, justification below

Deliverable

Deliverable: D92 of cycle 10 (docs/project/cycles/10-kitokey-ready-when-you-come-back.md), criterion 3: « Given a meeting made after a stop, when its last step ends — written or failed —, then exactly one notification is sent for it, and only when no window of Kitokey has the focus — none when one has —, naming the meeting, a click opening its page; « Transcription prête » is no longer sent on its own when a summary follows ». Criterion 4 merged with #807.

What it changes

  • One notification for a meeting, at its last step.
    • The transcript, when no minutes follow.
    • Otherwise the minutes, written or failed.
    • A transcript made again with its minutes to follow tells nothing: the minutes alone say it.
  • None while Kitokey has the focus. window_focused asks the toolkit whether the main or the small window has it. When one does, nothing is sent (shell.made_not_told in the log): the page says it there. This follows GNOME's HIG, as the cycle cites it: a notification is for an event while the user is in another application.
  • The meeting named first. The body starts with « Pour … . » / "For … ." and the title is kept as it was.
    • The name is the meeting's title, else its hour on the day it was recorded, else its day and hour: « la réunion de 12:00 », "the meeting of Oct 18 at 12:00 PM".
    • Days and hours come from one formatter in words.rs, with CLDR's abbreviated months, as the window's Intl writes them.
  • Minutes that cannot start (nobody named, no provider, an agreement asked, no whole transcript): their D25 notice is then the meeting's notification. It names the meeting and says what to do, and, as criterion 3 says, it is not sent while a window of Kitokey has the focus: the window says it there (Settings opened, or the agreement's page).
  • A transcript that cannot start after a stop (NoModel, NoTranscriber): the same path, the meeting named, nothing sent while Kitokey has the focus.
  • Two notifications for one meeting, accepted (the coordinator's decision, 2026-11-14, for the decider at acceptance): a notice that asks for a gesture — an agreement to send the transcript (Minutes::Waiting), Ollama waited for (OllamaWaits) — then, once given, the minutes'. They are two events: the first asks the user to act, the second says the meeting is made.
  • A click opens the page (D44), as before.
  • Words. The two new sentences (« Pour {}. » / "For {}.") carry D94's comment in words.rs. for_meeting is named in D94's title test and each notification is read again through it: the title is unchanged, so it still holds 50 characters.

Oracle

  • The success criterion was written before the implementation: scenario BN in scripts/e2e_window.py (lettered AL until 6e8d557, a letter main's D48 scenario holds too) (3934563 for steps 1 to 3, then step 4).
    1. Kitokey's window has the focus: the stop, the transcript and its minutes made, nothing told.
    2. Another application's window is in front and Kitokey's window is still seen: one notification, the minutes', naming the meeting, with no « Transcription prête » before it. Its click opens the page.
    3. The minutes fail: one notification, the failure's, naming the meeting.
    4. The minutes cannot start: one notification saying what to do for them, naming the meeting.
    • Since c167d8b, each step is a real stop (transcribe), made by the page's own click, which moves no focus; the meeting before it is moved out of the list. One redo (context_redo) is kept, for its own branch: the minutes after it, their notification alone. Added: a titled meeting (title.txt) named by its title, no model, no transcription program (each with what to do), and step 4 again with Kitokey in front — nothing told, Settings opened.
  • It was seen failing, for the right reason:
    • At 3934563 (steps 1 to 3, the code before the change), from the coordinator's run:
      • the setup passed three times: the focus given; another window in front, Kitokey's still seen; the click opening the page;
      • the assertion failed three times:
        • focused, a Notify was sent anyway;
        • elsewhere, no minutes' notification named the meeting;
        • failed, no failure's notification named it.
      • The audio defaults were unchanged and no source-output was opened.
    • At 397a654 (step 4 alone, on a throwaway branch from bad30e4, the code before 51e52f3): the three checks failed with a text that names the meeting. The monitor's text did not separate the count from a transcript's notification.
    • 0a5dd00 therefore lists each Notify seen when a check fails: its sender, whether it was relayed, its title and its body. Its AL run is queued; the coordinator moves it to this head if it has not run.
  • It passes now: BN at 6e8d557, the coordinator's run.

Kind of oracle: end-to-end, the window's e2e on played engines, reading the session bus's notifications. It needs a window, which this session's sandbox does not open.


Review budget

  • Lines changed (excluding generated): 436 (383 added, 53 removed), 192 of them the e2e's scenario and 90 the unit tests; about 250 of code and tests outside the e2e
  • Files touched: 4
  • Within budget
  • ⚠️ Over budget — 436 against ~400: the e2e's six cases, each through a real stop, asked by the verifier's review; the code itself is about 130 lines

Contracts

  • No contract affected

Definition of Done

  • Oracle green: BN, above
  • Lint, format, types: cargo fmt --check, cargo clippy --all-targets -D warnings, python3 -m py_compile scripts/e2e_window.py
  • Unit tests: tests/words.rs.
    • A day and an hour as the window says them: twelve months, a first of the month, midnight and noon.
    • A meeting named by its title, else its hour that day, else its day and hour. A title of blanks counts as none, and a start that cannot be read gives « la réunion ».
    • The notification named first keeps its title and ends with its own body.
    • D94's title test reads for_meeting. It was seen failing with the function left out of NOTIFYING.
    • The shell's whole suite passes except tests/migrate.rs::a_model_written_from_home_is_named_where_it_went. It writes in $HOME, which this sandbox refuses; the same happens on main, and shell: the test of a model written from the home uses a home of its own #819 fixes it.
  • Affected documentation updated: no page describes notifications; the sentences are in words.rs
  • Diff self-reviewed line by line

Test sheet

Verifier: session 92ded11e-58ef-492d-ba56-c820b9011de8

# Given · when · then Kind Result Evidence Commit Confirmed
S1 BN, step 1: given a recording stopped, Kitokey's window focused, when its transcript and minutes are made, then no notification is sent automated not verified —
S2 BN, step 2: given a recording stopped, another application in front, when its minutes are written, then exactly one notification is sent: the minutes', naming the meeting, with no « Transcription prête » before it, and its click opens the meeting's page; made again (context_redo), the same automated not verified —
S3 BN, step 3: given a recording stopped and given a title, another application in front, when its minutes fail, then exactly one notification is sent: the failure's, naming the meeting by its title automated not verified —
S4 BN, its last steps: given a recording stopped, another application in front, when its transcript cannot start (no model, no program) or its minutes cannot (no one chosen), then exactly one notification is sent, saying what to do and naming the meeting; with Kitokey's window focused, none, Settings opened automated not verified —
S5 tests/words.rs: given a meeting's title, or its start only, when it is named, then: the title in quotes; else « la réunion de 12:00 » that day; else « la réunion du 18 oct. à 12:00 »; in both languages; the notification's title unchanged automated passed Own clone at the head, cargo test --locked --test words: 51 passed; 0 failed. Five mutants, each applied alone (git diff --numstat 1 1 or 1 2), then restored, each killed by one test: for_meeting leaving the body as it was (the_notification_that_ends_a_meeting_names_it_first_and_keeps_its_title); « that day » inverted in French and a title of blanks kept (a_meeting_is_named_by_its_title_else_its_hour_that_day_else_its_day_and_hour); the English hour on one digit, %l (a_day_and_an_hour_are_said_as_the_window_says_them); for_meeting taken out of NOTIFYING (a_notification_s_title_holds_50_characters). The window's formatter, Intl.DateTimeFormat with hour: "2-digit" and month: "short" (node, ICU 78.3): 09:42 AM, 05:31 PM, 12:05 AM, 18 oct., 00:05, as words::day and hour say them 6c3c6be —

Verifier's remarks: read at 6c3c6be; S1 to S4 wait for the coordinator's AL run at the head.

  1. AL's steps 2 to 4 are not made after a stop. They make the meeting again (context_redo), which goes through retranscribe and its minutes_after branch. Only step 1 is made after a stop (transcribe), and with the focus in Kitokey. So the new tell_made calls in transcribe (a transcript with no minutes to follow, after = "transcript", and a transcript failed) run in no scenario with the focus elsewhere. S2 to S4 say « given the same »: if they are run as written, they hold for a meeting made again, not one made after a stop.
  2. D25 notices with the focus in Kitokey. When the minutes cannot start, the notice is sent whatever has the focus, as the description says. Criterion 3 says « none when one has ». Step 4 runs with the focus elsewhere only, so S4 does not cover this case. Which of D25 and D92 wins is a choice for the decider. Step 4's check also counts and looks for the name, but not for what to do: any one notification naming the meeting passes it.
  3. Two notifications for one meeting. When an agreement is asked (Minutes::Waiting) or Ollama is waited for (OllamaWaits), its notice names the meeting, and the minutes written afterwards are told again. That is one meeting with two notifications, against criterion 3's « exactly one ». It may be intended, since the agreement asks for a gesture. No scenario covers it.
  4. A transcript that cannot start after a stop. NoModel (l.4111) and NoTranscriber (l.4129) are sent with plain tell: the meeting is not named, and the focus is not looked at.
  5. A meeting with a title is never notified in a test. of_meeting's title, from named::read or else minutes::title, is not run: AL's meeting has none, and tests/words.rs calls words::meeting directly.
  6. Merging with D94 — shell: the tests read every sentence of words.rs against the voice guide #812 (open, fc3b37e; the merge was probed in a throwaway branch, not pushed). The merge is clean as text, but tests/words.rs then fails twice:

Summary

DONE

  • tell_made: the end of a meeting's making is told once, naming the meeting, only when no window of Kitokey has the focus.
  • A transcript with minutes to follow tells nothing. The minutes' D25 notices name the meeting.
  • words::day, hour, meeting and for_meeting: one way to name a meeting, in both languages.
  • The scenario BN (AL until 6e8d557), its steps 1 to 4, and its list of each notification seen when a check fails.

VERIFIED

ASSUMED

  • The toolkit's is_focused is the desktop's focus. BN holds it on Linux only.

NOT VERIFIED

  • BN green at this head: the coordinator runs it.
  • The cause of 397a654's red, between the count and a transcript's notification: 0a5dd00's listing is there to say it.
  • Focus on macOS and Windows: BN runs on Linux.

RISKS

  • A failure the user does not see in the window while it has the focus is not told elsewhere. The criterion asks for exactly that: the page says it.

Signals to report

  • I had to look at another module's implementation
  • A kernel or playbook rule got in my way for no good reason
  • A quality gate blocked without improving quality
  • A rule applied by hand would be worth automating

🤖 Generated with Claude Code

kitokey-marc Bot and others added 15 commits November 13, 2026 22:55
…the page gives an estimate and Cancel, and one notification says it is ready

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
…(BE)

Scenario BE, criterion 4 of D92: a transcript and minutes made again with a first
model, held by the played programs — each with a child deaf to TERM —, « Annuler »
on the meeting's page: both processes gone and the shell idle within two seconds,
the version already whole kept byte for byte with nothing beside it, the page
saying what was stopped, no notification; then a second model started at once,
the version shown naming it. A first transcript cancelled leaves nothing to be
taken up; the next model makes it whole. Fails today: no « Annuler » on the page.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ole version kept (D92)

The meeting's page offers « Annuler » while its transcript or its minutes are made,
or wait their turn. The program is stopped as a group: TERM, a second's grace, then
KILL to what is left — the summary's program now leads a group of its own, as the
transcript's did. A cancelled making is no failure: nothing told, no notification,
nothing made after it; the page says it stopped, and its button to make it again is
offered. A redo's earlier transcript is put back, the minutes there were never
replaced; a first transcript cancelled leaves nothing to be taken up by another model.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…by a test (D92)

The verifier's review of 6dac438. A mark is now by meeting and kind, set by
« Annuler » only under its kind's lock while that making runs for the meeting, and
taken under the same lock as it ends or begins: a cancel between the start's two
steps is no longer erased, and one pressed after the end no longer marks a later
making — of the other kind least of all, a first transcript's failure taken for a
cancel and its file removed. A meeting taken from a line and not yet begun is said
taken; « Annuler » withdraws it there and its start begins nothing. A test holds
the group told TERM first, a child ending by it within the grace.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ted (D92)

The verifier's second review of the same piece found a third window between two
locks: marks set by « Annuler » and read by a making's start and end could still
be lost or left. The field's convention replaces them — Go's context, tokio's
CancellationToken, .NET's: each making owns one token from the line to its end,
cancelled, or holding its running program, under one lock. « Annuler » cancels
the meeting's tokens — in a line, taken from it, being made —; a start begins
nothing once cancelled, and holds its program, stopped at once if cancelled
since; an end is « cancelled » by its own token alone. No mark, no set by meeting.
tests/cancel.rs holds the token's three orders and a cancel racing the hold.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…llows it (D92)

The decider's case, as the coordinator settled it: Annuler is pressed to change the
model. A transcript its program wrote all the same is whole and kept, shown as
written — never said stopped —, but no notification is sent and no minutes are
made after it with the model being changed, whatever `after` says. The same for
minutes written after a cancel: kept, nothing told. BE holds it: the played
program finishing on TERM, the settings chaining the minutes, none made and
nothing notified.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uilt (AL)

AL now asks, of a meeting made after a stop: nothing told while a window of
Kitokey has the focus; with another application's window in front, one
notification naming the meeting — the minutes', or the failure's —, never
« Transcription prête » on its own before them; its click opens the page.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ow of Kitokey has the focus (D92)

The end of a meeting's making — the transcript when no minutes follow, else
the minutes, written or failed — is told by one notification, its meeting
named first: its title, else its hour that day, else its day and hour, by one
formatter in words.rs (CLDR's abbreviated months, as the window's Intl). A
window of Kitokey with the focus: nothing told, its page says it. A transcript
made again with its minutes after: theirs alone. AL's precondition met in
app_language too: its notifications awaited with another application in front.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s other gestures say theirs (D92)

tests/silence.rs refused a command that fails without a word: the cancel's
failure — its meeting gone meanwhile — now says it where its making is said,
and the button is given back.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…of a meeting made after a stop (D92)

A transcript written with its minutes to follow is not told: when they cannot
start — nobody named, no provider, an agreement asked, no whole transcript —,
D25's notice is the one notification, and now names the meeting. Told as D25
says, the focus aside. AL's fourth step holds it.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nder, relayed or not, title and body (D92)

At 397a654 the three checks failed with a text that names the meeting:
the tail of the monitor's text did not say whether the count or a
transcript's notification failed them. Each Notify seen is now listed.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
window.js, and #807's own lines of main.rs, words.rs and e2e_window.py, taken from main: #807's
last head. D92 c3's lines kept beside them: the meeting named in words.rs' tests, and the
Settings check of D95 followed by c3's own « away » before the transcript is made again.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… kept (D92)

`for_meeting` returns a notification: named in `NOTIFYING` and each notice read again through it.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tices only where Kitokey is not in front (D92)

D25's notices for minutes that cannot start, and a transcript whose model or program is
missing, go through the one path that names the meeting and holds back while a window of
Kitokey has the focus: criterion 3's « none when one has ».

AL now makes each case through a stop — the page's own click, which moves no focus —, a
redo kept for its own path: the minutes ready and clicked, made again, a titled meeting's
failed minutes named by its title, no model, no transcription program, no one to write the
minutes; that last one again with Kitokey in front, nothing told and Settings opened.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… named by its minutes

Two scenarios said « AL. » on main: meeting_named (D48) and meeting_told_once (D92). The
second takes BN, free (BM is #844's, BL press_on_the_page_left). The print and the docstring
alone: no MANIFEST line names it.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kitokey-marc Bot and others added 6 commits November 15, 2026 09:53
…meeting-is-made

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s reader looks (D92)

`day`, `hour`, `meeting` and `for_meeting`: Shows, When, At most and Placeholders, inside each
body, `for_meeting`'s moved there and its placeholder named as its sentence writes it,
{meeting}. With them, the sentences without their comment are main's own: 600, against the
602 #812 left — `WITHOUT_COMMENT` lowered, as the test asks (red on main as it is).

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…meeting-is-made

The comment ratchet now fails only on an increase: main's WITHOUT_COMMENT kept, 597. Both
sides of tests/words.rs kept otherwise: D92's tests of `day`, `hour`, `meeting`, and
`for_meeting` read in `NOTIFYING` and `every_notice`.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…meeting-is-made

The e2e finds a start's tray item in a second: this branch's attest saves about 20 minutes.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…meeting-is-made

main.rs: a settings file refused, the minutes not started — #813's refusal (`why.detail` to the
line kept for the log, `why.refused` said) told by #824's path, naming the meeting where Kitokey
is not in front. tests/words.rs: both sides' tests kept.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…and BN counts Kitokey's own notifications (D92)

Criterion 3 and its D25 notices: nothing notified while a window of Kitokey has the focus, the
window saying it there. Four older checks asked a notification of a making started from its
page, the window in front: the minutes made again that fail, the cause they name, the redo
through the agreement, the first meeting with no one chosen. They now ask none, the page or
Settings saying it.

BN counted every Notify on the bus: the desktop relays a notification once more, from its own
connection (`x-shell-sender`), which it counted twice. It now counts those Kitokey sent.

Agent-Session: 14f73956-90b9-46a3-9fcc-52dc45c43948
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

over-budget Over the review budget, with its justification in the description

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants