go vet says nothing about import order, so when the move to git.ryuvia.com
rewrote every import path without re-sorting them -- the new path sorts before
github.com/..., where the old one sorted after -- both repos went through a
green CI run and a release unformatted.
Added to the release workflow as well as CI, so the two keep running the same
checks; ci.yaml's header claims exactly that, and a check in one but not the
other would quietly make it false.
The step handles gofmt's two failure modes separately because they do not look
alike: a misformatted file is listed on stdout with exit 0, so the failure has
to be raised by hand, while a file that does not parse prints nothing to stdout
and exits 2 -- which a plain emptiness test reads as success. Verified against
all three cases (clean, misformatted, unparseable) before committing.
Assigning an on-call week panicked with "index out of range [-1]" on an
ordinary schedule. The index came from scheduleTable.Cursor().
The cursor is not ours. bubbles' SetRows clamps it down when rows shrink
(`if m.cursor > len(rows)-1`) but never back up, so setting zero rows drives it
to -1 and filling the table afterwards leaves it there -- -1 is not greater
than len-1, so nothing corrects it. Every table in this package is rebuilt from
empty exactly once, when the first WindowSizeMsg arrives before any fetch has
returned, so every cursor started at -1 and stayed there until the user pressed
up or down. Pressing a direction key first is why this was survivable at all.
setRows restores the invariant the rest of the package already assumes: a table
with rows has a usable cursor. Every rebuild goes through it.
Five call sites also bounds-checked only the top of the range, and are now
consistent with their siblings, which already had `i < 0 ||`. They were the
same latent panic: deleting a schedule entry, deleting a user, editing a topic,
opening the API key menu, and the week assignment that actually fired.
The regression test deliberately never calls SetCursor. That is what the
existing schedule tests do, and SetCursor clamps, which is exactly how this got
past them. It drives the real order instead: size, then data, then keys.
Also carries a gofmt pass, which is why untouched files appear in the diff.
The move to git.ryuvia.com rewrote import paths without re-sorting them, and
the new path sorts before github.com/charmbracelet/..., where the old one
sorted after. go vet does not look at import order, so CI had nothing to say.
The module path, the CI pipeline and the self-updater all named GitHub. They now
name the Gitea instance everything else already runs on.
The workflows are rewritten rather than translated, for the reason recorded in
ci.yaml: Gitea's runner image is ubuntu:22.04, whose nodejs is Node 12, so no JS
action runs there -- actions/checkout@v4 dies with a SyntaxError before doing
anything. Every step is shell and checkout is a plain clone, which this public
repo needs no credential for. upload-artifact/download-artifact are JS actions
too, and there is no artifact store here, so the job that builds the binaries is
the job that publishes them.
internal/updater keeps its release and asset types unchanged: Gitea's release
payload carries the same tag_name, and its attachments the same name and
browser_download_url, so only the URL, the Accept header and one error string
move. The asset naming in release.yaml is load-bearing for that matching.
This does strand already-installed binaries, which still poll api.github.com.
The GitHub repository is left in place and untouched, so they report themselves
up to date rather than erroring; its last release is the bridge, and crossing it
is a one-time manual download.
Assigning over a day somebody else held did nothing but flash a 409 for
three seconds. The server holds one person per date and refused any that
was taken, all-or-nothing, so pressing W on a week where a single day was
already assigned placed none of the other six either. The only way
through was d on each day first — seven delete-and-confirm cycles to move
one week.
The clash is already on screen, so it is found before the request rather
than read back out of an error: the picker hands off to a confirmation
naming who loses the days and how many there are, and accepting sends the
whole selection with replace, which terdut-server v0.8.0 added. One
question to move a week, and nobody's shift moves without somebody being
asked. A day nobody holds still assigns with no prompt at all.
Reassigning somebody to a day they already hold raises no prompt, since
it takes nothing from anyone, but it does send replace: the server
rejects any date that exists, so without it a harmless no-op would fail.
terdut-server pages the on-call person through ntfy, but none of it was
reachable from here. A user's topic could only be set with curl, so a
new user silently got no pages and quietly fell back to the shared
fallback topic — which carries no Acknowledge button. And nothing said
whether anybody had been paged at all.
The Users section grows an Ntfy Topic column and t to edit it,
prefilled with the current value. Submitting an empty field clears the
topic rather than being rejected as a mistake: clearing is how somebody
is taken off their own topic, and it is what the server means by an
empty string. Nil and empty arrive as the same thing, because the
server stores a blank topic as NULL, so User.Topic flattens the two
instead of leaving every caller to.
The incident timeline renders the server's notified and notify_failed
events. No new fetch — the timeline endpoint already carried them, and
unknown types already fell through to a generic label; this is about
saying something useful. An event with no user means the fallback
topic, not "the server acted", which is the difference between somebody
having been paged and the rota having been empty.
Both need terdut-server v0.6.0 or later, and the timeline entries a
server newer than that. Against an older one the column stays empty and
editing a topic reports the server's 404, which is the honest answer.
truncate measures and slices by byte, so a string cut inside a
multi-byte rune both mis-measures the fixed-width column it is being
laid out against and emits a broken character. Everything it is handed
is server-supplied — alert names, label values, annotations — and none
of that is guaranteed to be ASCII.
Identical behaviour for the ASCII case.
Stats was the one full-screen view reached by a key of its own rather
than by tab, and the interface was less coherent for it. It is now a
section sitting third, after Alerts, and behaves like every other one:
tab in, tab out, r to refresh.
Three things fall out of the move. It auto-refreshes for the first time
— the tick handler skips every non-dashboard mode, which is why the
overlay never updated while it was open. Its error path no longer forces
the queue back into view on a failed fetch, an assumption that only made
sense while stats floated above the dashboard. And first-visit loading
keys off a statsLoaded flag rather than slice emptiness, because the
three empty slices a quiet server returns are a real answer, not a
missing one; the loading placeholder is likewise suppressed once
something has been drawn, so a background refresh cannot blank the page
out from under whoever is reading it.
The S key is gone, and with it the ability to peek at statistics from an
open incident and land back on it. That round-trip was the only thing
statsReturnMode bought, and it was the whole reason stats needed a mode.
The release workflow gates a tag, which is the last possible moment: a
commit that breaks the suite stays green on main until somebody decides
to publish.
Runs go vet and go test on pushes to main and on pull requests. push is
scoped to main so a branch pushed as part of a pull request is not
checked twice, and runs for the same ref cancel each other.
An incident with nobody assigned and nobody holding it rendered its
detail view like this:
Assigned: nobody
Acked: not acknowledged
lipgloss pads every line of a styled block out to the width of its widest
line. A trailing newline inside Render therefore produces a second line
made entirely of padding, and the next write to the builder starts after
that padding instead of at the left margin. Twelve call sites put the
newline inside.
Adding a line() helper that keeps the newline outside, and using it
throughout the content builders.
Shipped in v0.4.0 and only visible on the unassigned or unacknowledged
path, which is why it survived the pre-release check: that run had
somebody on call, so the incident was assigned and acknowledged and both
lines took the styled-with-value branch instead.
view_test.go covers the content builders, including the two states that
were broken and an expired snooze not being reported as a snooze.
The repo had no tests at all, which the v0.4.0 rewrite made
uncomfortable: this client speaks terdut-server's REST API directly, and
a wrong path or method is invisible until somebody runs the binary
against a live server. That is exactly how it broke when the server split
alerts from incidents.
The Elm architecture makes most of this cheap to check without a
terminal. Update is (Model, Msg) -> (Model, Cmd), so keypresses can be
synthesised and the resulting model inspected; a nil command is a
readable assertion that the model decided to do nothing.
Three suites:
- client_test.go drives every incident endpoint against an httptest
stub that records method, path, query and body. Also covers the
filter query params, that a server error message survives into the
error the UI shows, that a 404 from the on-call endpoint is not an
error, and that omitted optional fields decode to zero rather than
failing.
- model_test.go covers the pure helpers: filter cycling, the snoozed
pseudo-status, duration formatting, row builders, and the column
width arithmetic that overflowed the terminal once already.
- update_test.go covers the rules worth protecting rather than
coverage for its own sake. Resolve prompts first and cancelling does
not act, since resolution is terminal server-side. A resolved
incident rejects all six workflow keys. Archiving refuses while an
incident is open. The note cursor walks notes only and wraps. Stats
returns to whichever view opened it. Modal states do not auto-refresh
underneath the user.
129 tests, running in about 40ms.
The workflow only built and published, so a tag went straight to
binaries on the releases page with nothing having run against the code
first.
A test job now runs go vet and go test, and build depends on it; release
depends on build, so a tag that fails publishes nothing.
Only runs at release time — a push or pull request is still unchecked.
terdut-server v0.4.0 splits the alerts row into two objects, and the
endpoints this client drove for acknowledgement, comments and archiving
are gone. Pointing the same screens at the new paths would have missed
the point of the split: alerts are now Alertmanager's record, read-only
and carrying no human state, while the incident is the thing anyone
actually works on.
Incidents lead the section list and are what the client opens on. The
queue shows severity, status, assignee and age, and the detail view adds
what only exists server-side now: the group labels Alertmanager
correlated on, the member alerts, and an append-only timeline where
system events and notes are interleaved. That timeline is the whole
history the server keeps — alert rows are still mutated in place — so
rendering it in order matters more than styling it.
Actions all move onto the incident: a/A acknowledge, s assign, z/Z
snooze, c note, d delete note, x archive, R resolve.
Two of those need care rather than a keybinding:
- R, not r, resolves, and it asks first. The server treats a manual
resolve as terminal: a later occurrence opens a new incident instead
of reopening this one, and an alert that never stops firing leaves
the incident closed for good. A stray keypress is not recoverable,
so the prompt says what it means.
- x refuses on an open incident rather than archiving it, since
archiving unresolved work only hides it. Snooze is offered as the
"not now" answer, and the client treats a snoozed_until in the past
as not snoozed, matching the server, which sweeps nothing.
Statistics lead with MTTA and MTTR, neither of which was computable
before. The server sends null until something has actually been
acknowledged or resolved, and that renders as — rather than 0: no data
is a different claim from instant.
Alerts keep a tab of their own as the raw feed — useful for asking what
Alertmanager is really sending — with an Incident column replacing Ack
By, and i in the detail view jumping to the incident where something can
be done about it. Archived now holds archived incidents; archiving an
alert is server-side housekeeping and no longer a user action.
BREAKING CHANGE: requires terdut-server v0.4.0 or later. Against an
older server every incident request 404s. Use terdut-tui v0.3.x with
servers before v0.4.0.
The alert list showed only Started, which comes from Prometheus and
never changes for the lifetime of an alert instance. A firing alert
that started 12 days ago looked identical whether Alertmanager
refreshed it 30 seconds ago or went silent a week ago.
terdut-server already tracks this: the webhook upsert sets
received_at on every accepted payload, including the periodic
re-sends issued at repeat_interval, and its archiver treats the
field as a liveness heartbeat. The field was already decoded into
api.Alert.ReceivedAt and simply never rendered.
Add a Last Seen column to the alert tables, rendered with the
existing humanAgo helper. The Alerts and Archived tabs share
alertColumns/alertRows, so both pick it up. The width budget is
re-derived for five columns; the slack constant now accounts for
all of bubbles' per-cell padding, so the table lands exactly on
the terminal width instead of overflowing by two columns as it
did with four.
The detail view gains a matching Last Seen line, with the timeline
labels widened to keep values aligned. Since received_at stops
advancing once an alert resolves, also pull through the server's
resolution_source and show it in the status header
(RESOLVED · alertmanager vs RESOLVED · expiry) so a frozen
timestamp is explained.
Add a fourth tab (Alerts | Archived | Schedule | Users).
Archived alerts are fetched lazily on first visit using the
archived=true query param on GET /api/alerts.
Press x from the Alerts list or detail to archive an alert;
the non-archived list refreshes immediately. Press x from the
Archived list or detail to unarchive; the archived list
refreshes. Ack/unack are disabled in the Archived detail view.
New API methods: ArchiveAlert (POST), UnarchiveAlert (DELETE).
ArchivedAt field added to the Alert type.
Schedule table now shows ISO week number (W21) on the first
row of each week visible in the 14-day window (first row and
every Monday); other rows get a 4-space indent. Date column
widened from 16→18 to accommodate the prefix.
New W key in Schedule section opens the user picker in
week-assign mode, which assigns the selected user to all 7
days (Mon–Sun) of the ISO week containing the cursor. The +
key retains single-day assignment. The user picker header and
footer reflect which mode is active.
Add full user management section (tab to Users):
- User list table with username, email, created date
- n: create user (username + email form)
- d: delete user with confirmation (cascades API keys)
- k: API key management for selected user
- n: create new key with name input; one-time reveal
showing key value and integer key ID prominently
- r: revoke by integer ID (no list endpoint on server)
- c: copy revealed key to clipboard (atotto/clipboard)
API key listing is unavailable server-side, so the reveal
screen displays the key ID prominently for future revocation.
Adds a 14-day schedule list accessible via the Schedule tab.
Keybindings in schedule section:
←/→ (or h/l) shift the 2-week window back/forward by one week
j/k navigate rows
+ open user picker — select a user to assign to the date
d delete the selected day's assignment (y/N confirmation)
r refresh schedule from server
The user picker fetches the user list from the server on first open
(cached for the session). Selecting a user assigns them to the
highlighted date (POST /api/schedule all-or-nothing; 409 conflicts
surface as a status bar error).
The on-call header shows today's scheduled person and the current
window date range. On-call data refreshes on schedule actions.
API additions: GetSchedule, GetCurrentOnCall, AssignSchedule,
DeleteScheduleEntry, ListUsers. ScheduleEntry and User types added.
Press enter on any alert to open a full-screen detail pane (viewport).
Keybindings in detail mode:
a/A acknowledge / unacknowledge (updates immediately)
c compose a comment (text input at bottom)
[/] cycle comment cursor up/down
d delete selected comment (y/N confirmation)
s assign placeholder — shows "not yet supported by server"
S open stats view: top alerts + by-hour/by-day ASCII bar charts
esc back to dashboard
API additions: GetAlert, AcknowledgeAlert, UnacknowledgeAlert,
GetComments, AddComment, DeleteComment, GetTopAlerts,
GetStatsByHour, GetStatsByDay.
Adds a bubbles/table alert list showing Name, Status, Started (relative
time), and Ack By columns. Stats bar shows total/firing/resolved counts.
Filter cycles firing → resolved → all with [f]. Auto-refresh fires on
the configured interval via tea.Tick. Table scrolls with j/k.
Sets up the full project structure following the hactl/gokapi-tui
architecture: strict Elm-pattern Bubbletea TUI, YAML config at
~/.config/terdut-tui/config.yaml, REST API client with Bearer auth,
self-update via GitHub Releases, and a three-section tab placeholder
(Alerts / Schedule / Users) that verifies server connectivity on startup.
Niklas Ye