-
v2.0.0-rc4
Pre-releasereleased this
2026-09-29 22:50:19 +00:00 | 8 commits to master since this releaseThis release is what came of running the code the test suite had never
executed. Three of the defects it found could reach a live account: an
unreadable API token was destroyed by the next save, an import could create a
zone the file never named, and a destructive confirmation could fire twice.Added in 2.0.0-rc4
- A record of which lines of this application the test suite has never
executed.scripts/line_coverage.pyreports every executable line that no
test run reached, and everything below is what it found. - A replace or a delete across the account leaves a restore point. Search
and Replace was the one interface here with no version history; it now files
one entry per zone it wrote to, naming the records that really changed, with
any refusals named beside the count.
Fixed in 2.0.0-rc4
- A rate limit of 0, which means no limit, was the one setting the 429 backoff
could never touch, so deSEC could refuse request after request without the
application ever slowing down. From unlimited it now starts pacing at the
application default and halves from there, while a limit deliberately set
below the floor is still left alone. - A rate-limit refusal froze the window while the backoff waited on the lock
that paces requests. It no longer takes that lock. - An error response whose
detailfield was not a string escaped the API
client as an unhandled exception on a queue worker thread, instead of being
reported. - A rate-limit refusal with an empty body put an empty message on screen.
paramswas silently dropped on every request verb but GET, contradicting
the documented behaviour.- An API token that could not be decrypted was destroyed by the next unrelated
save. One launch under a different home directory, or aconfig.json
restored without its salt file beside it, was enough. The stored token is kept
byte for byte now, so a later ordinary launch can still migrate it. - A failed token migration said nothing useful. The error named no reason and
fired on successful migrations as well, so an upgrade that lost the key looked
exactly like a credential revoked at deSEC; those need different actions and
are reported differently now. - The salt file's permissions are corrected on a file that already exists,
not only on one this application creates. - A clock that had gone backwards left deleted records on screen. A
timestamp ahead of the current time read as fresh, so a timezone change, a
daylight saving transition or an NTP correction could pin deleted records in
the pane and let the delegation check trust a parent copy it had been told not
to trust. - An import could create a zone in the account that the file never named. A
file that does not name its zone fell back to a real, registrable domain name
and created it; that is refused now, before any request goes out. - A djbdns import could create the wrong zone, taking the zone to be the
first record's name minus its first label, so a file whose first line is the
apex ofexample.comproduced the zonecom. - A merge import that wrote records reported writing none. One record
missing its TTL abandoned the counts for the whole file; a bad record is named
and counted now, and the rest of the file still goes in. - Exporting a zone stripped the timestamps out of the copy held in memory.
One export with metadata unticked left every record in that zone with no
creation or change dates until the next sync. - A BIND file this application exported could not always be read back. Six
defects: multi-value record sets came back as duplicates, types longer than
six characters were dropped outright, fully qualified owners of the kinddig
and an AXFR produce were refused, a parenthesised SOA was shredded into bogus
records, inline comments ended up inside record values, and a legal
three-field line was dropped in silence. - A djbdns export could not be imported at all. The reader understood two of
the six line types the exporter writes, and colons in an AAAA address and in
every SPFinclude:broke a colon-delimited format; all six round-trip now. - A line no parser could read is named with its line number rather than
dropped in silence, in all four formats, and the count reported is what the
file actually holds. - The import documentation was wrong in three places: it required an
$ORIGINthe BIND reader no longer needs, it listedsubnameas required for
JSON and YAML when the apex is deliberately the empty string, and it said
nothing about unreadable lines being reported. - The lines a file could not be read from are shown before you confirm the
import, not after, three named in full and then a count. - The preview's "Original Zone (from file)" line could never appear. It shows
what the file said against the target you chose now, so a deliberate retarget
can be told from having opened the wrong file. - Every destructive confirmation could be answered twice. Deleting a record,
a zone, a token, a policy, a bulk selection, a whole account's worth of
matched records or the cache all go through one drawer, and it stayed live for
the length of its closing animation. One confirmation is one action now. - A confirmation asked for without a second step had no gate at all, and
every click fired. - Escape after confirming made the drawer answer twice, cancelled as well as
confirmed, which put the Import page back to idle over a run it had already
started. - Deleting the same token twice while the list refreshed reported a failure
for a token that had already gone. Token deletes carry the per-item guard that
policy deletes were given, keyed per item, so two partly overlapping
selections still delete every token in them. - Search and Replace's destructive buttons had no re-entrancy guard, so a
second click enqueued the same records again, turned rows red for deletes that
had in fact succeeded, and overwrote the first run's summary. - Clearing the Queue page's history could swallow a callback that had not
fired yet, which left whatever was waiting on it waiting for ever. The
dispatch no longer looks an item up in a table that the clear prunes. - Retry on the Queue page left the item in the state that had failed. It no
longer starts with its retries spent, no longer shows the previous failure's
response in the detail panel, and no longer comes back after a restart. - The Pause button on the Queue page followed only its own click, so it
still offered to pause a queue the window had already paused for offline mode
or a rate-limit cooldown. - A typo in the TTL filter widened a Search and Replace to the whole account.
A TTL that is not a number was ignored in silence, on a page whose next two
buttons are Apply Changes and Delete Selected; it is reported now. - A rename that could not delete the old record was reported as a clean
rename. It counts as a failure now and names the record left behind. - Deleting a zone left a stale refresh running behind it that came back 404
for the zone just removed. Pending work for a zone is cancelled before the
delete is sent as well as after it, with other deletes left alone. - Confirming the same zone delete twice reported a failure for a zone that had
just been deleted. A delete cancelled by an earlier delete is no longer
rendered as an error, while a cancel asked for from the Queue page still is. - A zone delete on a window with no queue said the zone was deleted and left
it in the list, with Delete Zone still enabled and pointing at it. - A zone the server named differently from the way it was typed never got a
padlock. The name in the create response is what gets registered for the
delegation and DNSSEC check now, so the verdict lands on a row that exists. - Refusing to add a zone because the application is offline was completely
silent. A panel opened while the connection was up and submitted after it
dropped slid away with the typed name and left nothing on screen; it says why
now. - Double-clicking a zone while offline raised "Editing zones is disabled in
offline mode" for a gesture that only reads. - The zone counter dropped its units whenever a filter was typed, reading
"Showing 1 of 2/100" instead of naming zones. - The countdown under a freshly created token was a label, not a clock. A new
token went to the clipboard for thirty seconds under a string that only
claimed to count down. It is two minutes now, ticking once a second in one
line under the token, in amber and then red for the last ten seconds. - Closing the token dialog clears the clipboard early, and the dialog says so
while the token is still on screen. - The clock never stops while the token is still on the clipboard. A
clipboard it cannot read keeps it running, and it ends only on evidence that
the token has gone: other text, or another application holding content it will
not hand over as text. - Copying a token twice started two clocks and the shorter one won, a timer
could fire into a destroyed dialog, and an empty token offered a Copy that
would have cleared an unrelated clipboard. - Deleting a token's last RRset policy makes that token unrestricted, and
nothing said so. It is said in the confirmation, while there is still
something to change, and again from the refreshed list afterwards. - A refused policy delete names which policy it was, where three selected and
one refused gave three identical lines. - The same policy could be deleted twice while its row waited for the
refresh, and the second 404 was reported as a failure over a restriction
that had already gone. - An emptied policy list was not cached, so the next cache-first paint showed
a restriction that no longer existed. - A zone with nothing known about it was treated as settled and left for a
week. A zone whose first check failed comes back for a recheck in hours now,
which is what the unsettled interval is for. - Maximising the Add Zone wizard stretched its four mode tiles down the window
and crammed them into its left half. The grid takes one, two or four columns
according to the room it has, the slack sits at the bottom of each card, and a
line of text stays about the same length however wide the window is. - Ctrl+F could only ever reach the zone search field, never the record one.
Both are in the cycle now.
Technical Improvements in 2.0.0-rc4
- Three more invariants are enforced by detectors that run in the test
suite, numbered I10 to I12: a module that owns state outliving a window may
not ship a member a user can reach that no test has ever executed, one
structure may not be written out as two literals, and a test stand-in may not
be incapable of taking a call the source really makes. - The DNSSEC and delegation status entry is built from one defaults table,
where it had been written out by hand in seven places. - A test file that passed and then died is no longer reported as green.
scripts/test_token_page.pyprinted a clean summary and then exited on a
signal; the runner reads the exit status now and reports a crash as a crash,
and the crash itself is gone, each test closing and deleting its top-level
widgets before the interpreter shuts down. - A wizard run asked for a read and a snapshot of every zone it aimed at,
including the ones the API refused. It uses the set of zones actually
written, which it had already worked out. A stand-in for the API client in
that area could no longer take the call the source makes, and is fixed with
it. - A cache test forces its failure by injection rather than with a read-only
directory, which root is not subject to, so it no longer passes locally and
fails in the CI container. - 2175 tests across 84 files, up from 1634 across 64
Downloads
-
Source code (ZIP)
8 downloads
-
Source code (TAR.GZ)
9 downloads
- A record of which lines of this application the test suite has never