The first 0.0.17 cut (run 687) failed verify-pins in CI for two reasons. The asset generator embedded admin/dist/.src-hash, a freshness stamp that CI's artifact copy does not carry; it now skips dotfiles. And the Arch zig package emits different code than the ziglang.org tarball that CI installs, so the cut downloads the pinned tarball (ZIG_TARBALL_SHA256 in gates.yml, the full digest keys the cache) and builds the release with it. flake.nix is re-pinned to the bytes both now produce.
The saturated-primary pool test gates its holders on a semaphore instead of sleeps and releases every spawned holder on the way out, so a loaded runner cannot flake it. The package job uploads the payload before the pin check and runs the check when the version or flake.nix changed against the parent. The verify-a-release recipe clones the tag first and builds with the official zig.
Two errors the test build never saw, because the test binary analyses only what the tests reference: `Io.Dir.makePath` does not exist in 0.16.0 (`createDirPath` does), and the tag push passed the optional authorization where the checked token was meant. The test step now depends on the cut compile as well.
flake.nix fetches the release tarballs and carries their SRI hashes in a generated block. The cut tool builds the release locally with the toolchain gates.yml pins, in a normalized nine-variable environment, writes the hashes into flake.nix, and commits it with build.zig.zon as the single bump commit. The package job verifies the pins on the bump commit and the publish job verifies them again on the tag, before anything is uploaded.
The tarballs are written by dist_stage (std.tar.Writer, flate gzip) instead of the runner's tar and gzip, and -ffile-prefix-map keeps checkout paths out of the C objects; two checkouts at different absolute paths produce byte-identical archives. nxdns version, /api/version and the admin footer report the version only: the bump commit cannot know its own sha.
The v0.0.16 cut pushed the tag and then reported a failure because the release.yml run lost one gate to a timing-flaky test while its guard had passed and its publish job never ran. Rerunning the run by hand published the release; the tool had exited 1 and could not be run again, since a tag on origin meant "derive the next version".
The release stage now classifies a failed run from forge state: retryable when the conclusion is failure, the run is on its first attempt, every commit-status context in a failure state is a gate, and no release is published for the tag. A retryable run is rerun once through the actions API and the same run id is watched again, restarted only once the forge reports a higher run_attempt, so a concluded previous attempt is never read as the result. The rerun decision parses statuses strictly and treats anything it cannot read as terminal. A tag already on origin that peels to HEAD with no published release resumes at the release stage after the origin tag is verified (annotated, tool convention, pinned signing fingerprint). Both decisions are pure, table-tested functions. Spec bullets in specs/release-cut.md.
The test "overlapping exchanges each report the entry that answered that call" failed once on a loaded release runner and passed on rerun. It started two exchanges back to back and relied on a 50 ms failing stall to keep the first in flight until the second arrived; nothing checked that the two calls were ever in flight together, and the failing schedule was never observed.
Two Fake behaviours, hold and hold_fail, wait on a test-owned std.Io.Semaphore and then reply or fail. The test holds both calls inside the first entry, waits on the entry's in_flight count for one and then two, releases the failing call so it fails over while the other is still held, then releases the other. The fake draws its behaviour before it raises in_flight, so a count of one also fixes which call holds which gate. Each gate has a deferred post registered after its task's deferred await, so any early return releases the held call instead of deadlocking. No sleep remains in the test; every assertion is kept. No pool bug was found. Addendum in specs/milestone-31.md.
The Overview page takes the decided visual language (specs/ui-visual-redesign.md): four centred totals with their Activity links, a smoothed area chart of total and blocked queries with point hover and a tooltip centred beside the point, a stacked client chart in eight distinct hues plus one Other band that is always a series, and a card row with the cache hit rate, the query types as a single-hue ramp ring, and the upstream breakdown. The count axis grows its margin with the widest grouped tick and draws whole-number ticks only.
GET /api/overview takes a client parameter; the scoped read uses idx_query_log_ts and the cache keeps scoped slots. The device selector beside the period selector is URL state, so a scoped view is a link, and the tile links carry the scope into Activity. The route reduces a pasted IPv6 scope to the RFC 5952 spelling the logger stores, mapped addresses included, and drops anything that is not an address. A failed device list says so under the selector with a retry.
All measured quantities go through admin/src/lib/format.ts: grouped counts, two-decimal percentages, one-decimal rates, durations as the two largest nonzero units. Identifiers, configured values and preset labels render as written; the module header states that scope. A sweep test refuses toFixed, toLocaleString, Intl.NumberFormat and padStart anywhere else.
Chrome: one 4px radius from the metrics constants, shared Card with a prominent title and a one-line description on every panel, the settings form sections on the same card with a floated legend, the sidebar grouped into Monitoring and System with a status block (protection, queries per minute on Overview, uptime), keyboard-focusable table scroll wrappers, and the accent darkened to 5.43:1 on its wash.
Not built: the spec's ranked-list primitive, which has no consumer and no API rows. Codex reviewed sessions B to D over five rounds (thirty-three findings fixed, thirteen rejected as non-quantities); the owner skipped a sixth round.
Claude-Session: https://claude.ai/code/session_01VTgx3a1zz1R78o4K55kkwR
selected states stop changing font weight, buttons gain a pressed scale and scoped 120ms transitions with a reduced-motion override, every loading and empty state reserves its height, charts measure before first paint, hit targets rise to the 44px enhanced target where layout permits, long domains clamp to two lines on an unpadded inner span, chips and name cells truncate, tabular figures on counts and time columns, a z-index layer scale replaces magic numbers and fixes the dialog-over-confirm tie, history's empty state gains a clear-filters action, page headings balance, font smoothing and color-scheme land on the html reset.
ci's frontend job timed out three page tests at 5s; the runner paid the barrel's ~1500 icon modules on first render. deep imports load only the icons the app uses; local test import time drops by a quarter.
health strip marks, menu ticks, chip crosses, the search magnifier, the checkbox check and the back arrows all come from @phosphor-icons/react now, pinned exactly and entered in the license ledger. changelog and version bump for 0.0.15, including the bundle budget raise to 900000 bytes.
the freetype client field is gone. the picker is a select-only menu of known clients with multi-select, chips for the active set, and a 32-client cap shared with the server. the api accepts a comma-separated client list and filters any-of with bound parameters. the trigger carets come from phosphor icons, newly adopted. bundle budget rises to 900000 bytes for the picker and the icon dependency.
the five-field form with its apply button is gone. one row holds a search-shaped domain field and a client ip field that debounce into the url with enter flushing at once, result segments and a time menu that commit instantly, and a clear that appears only when a filter is active without reflowing the row. presets freeze both absolute bounds at click time so a bookmark describes the same investigation later, custom ranges apply atomically through set range with inline validation, and echo queues keep in-flight commits from clobbering newer typing or newer picks. live mode renders no toolbar, the availability banner became a footer note, previous rows stay visible during refetch, and every control carries a 44px hit target.
the client address follows its name as visible muted text in the query tables, the config lock indicator prints its reason beside the tag except in table rows where a page-level note explains the lock instead, and the locked delete buttons describe themselves through that one visible note. the chart legend tooltip is deleted because a named client is deliberately not addressed in the chart, and the dead series address field went with it. titles that merely repeat visible copyable text stay.
the upstream, blocklist source, and safe search enable controls mutate the row the moment they move, so they now carry the switch role via a shared drawn rac switch with a 44px hit area, a focus-visible ring, and naming modes made exclusive by a discriminated union. safe search supersedes its one-commit-old checkbox form, and the file-authority guard now names the switch role instead of relying on the bare input selector.
the assigned sources list is a rac checkboxgroup and safe search uses the same drawn checkbox, extracted to a shared ui component with grouped and standalone modes enforced by a discriminated union. the label carries a 44px pointer-target floor on both axes, the focus ring is driven from rac's focus-visible state and guarded by a test, and toggleSource is gone because the group hands back the whole set.
the drawer is inline flow content, so disclosure is the honest semantic: rac now owns aria-expanded, aria-controls, and the panel hidden state, while the open guard still unmounts the drawer contents so a closed drawer keeps no second pause control or health poll alive. the disclosure modules cost 3,669 bytes and the assets gate had 1,998 of headroom, so the budget moves from 800,000 to 850,000.
network assignment rows, activity filters, and the settings password pair now mark the offending input with aria-invalid and point it at the error text with aria-describedby. the prefix validator returns the row and field it is about, and any row mutation clears a message that named a position. settings numeric fields carry aria-invalid on an unparseable value.
the streamed-query detail panel is now a rac modal dialog with focus containment and restore. the live/history switch is rac tabs driven by the url, the overview period picker is a rac radio group, and the clients delete flow uses the shared confirm dialog; an authority turn keeps the dialog open and withdraws only the destructive action.
the header row survives only on narrow screens; on desktop its lone occupant, log out, joins pause in the sidebar footer, both full width. pause leaves the query detail page's related actions, where a global control had no business, and its hand-rolled duration dropdown becomes a react-aria menu with real keyboard navigation, dismissal and positioning.
the shared button variants set no cursor, so only components with local one-off styles showed pointer. all nine variants and the six buttons styled outside them now carry pointer, with not-allowed when disabled.
querylog.db carries a schema version; migrations run at startup as one transaction after a vacuumed 0600 backup, and every failure refuses startup (exit 2, no systemd restart loop) instead of starting empty. corruption is the only automatic recreate left. the cut gate now requires a fixture-proven migration or an explicit versioned break with restore instructions, and locks shipped migration files and fixtures byte-for-byte.
waiting for a slot now spends the query budget; truncated attempts that
expire fault the budget, not the upstream, and are never attributed.
admission sweeps in priority order before blocking. forward zones spend
read_timeout_ms once across udp, truncation and tcp. adds
nxdns_upstream_budget_exhausted_total and a 64-upstream validation limit.
the hand-written scale, tick, stacking and arc math is replaced by visx 4.0.0
primitives; rendering, colours and themes stay the app's own. all four charts
share one hover treatment: the client chart gains the tooltip and dimming the
query timeline had, the donuts gain both, an open tooltip follows a data
refresh instead of going stale, and it retires when the window rolls. the
timeline's third series is named allowed instead of other, and the client
chart's other aggregate disappears from a window where it counted nothing.
licenses gain the isc text for the bundled d3 modules.
settings and upstream writes now follow a prepare, commit, publish, retire
contract: candidates are built and validated before the database transaction,
published as infallible pointer swaps, and old generations retire after their
readers drain. per-query policy values snapshot once per query; upstream pool,
cache, rate limiter, sessions, api limiter, log sink, blocklist scheduler and
the query-log queue each gained one named live operation. restart_required
shrinks from every scalar key to the bind keys and web.enabled; the admin ui
drops its restart notices for everything else. file mode is unchanged.
the shell grid grew past the viewport and scrolled the document,
carrying the sidebar with it. the shell is now viewport-height at the
wide breakpoint with main as the sole scroll container; the nav list
scrolls inside the pinned rail; router scroll restoration targets the
inner scroller so navigation resets it and back/forward restores it.
the gate recomputes the previous release tag's ddl fingerprint from the
remote peeled object and compares it against the tree's; a change must
be disclosed by 'resets your query history' in the version's changelog
section. the 0.0.9 reset shipped with an announcement claiming no
schema change; this makes the impact mechanical instead of remembered.
the bisect proved the runner image's qemu 8.2 was never affected; the
hangs this masked were the logger deadlock fixed in f1a85d3. the 10x
slowdown it cost also broke the full-batch flush timing test in run 592.
mokhtar