Release v1.3.7

🚀 Release v1.3.7

UI stability pass on top of v1.3.6 — three independently-reported visual bugs fixed and the install/uninstall docs synced with what the scripts actually drop on disk. No backend or wire-protocol changes; clients from v1.3.6 keep working.

🛠️ Fixes

TCPing chart no longer "reloads itself"

• loadTCPingHistory now returns { __error: true } on transient fetch failures instead of an empty object that looked identical to a "no samples yet" response. Brief 502s from a flaky reverse proxy or a server restart used to tear the existing chart down and flash "No data" before the next 30 s tick repainted; now the live chart is preserved across hiccups and only the initial load surfaces the empty state.
• Added a per-system re-entrancy guard (tcpingIncrementalInFlight) so the per-row 30 s poll and the SSE metric_updated push can't race two fetches onto the same canvas — that race was the most reproducible cause of the "reload" flicker on slow networks.
• cleanupTCPingChart and handleTCPingConfigUpdated clear the guard, so a real config change isn't gated by a stale "fetch in progress" entry.

Admin server list no longer jitters every 3 s

• formatBytes now always emits two decimals — "1.00 KiB" / "1.50 KiB" / "1.55 KiB" — instead of stripping trailing zeros, so the traffic column stays the same glyph count between pushes.
• .traffic-info-container is pinned with font-variant-numeric: tabular-nums and a 110 px min-width; combined with stable formatting that fully eliminates the left/right drift that used to drag the edit / copy buttons sideways.
• editBtn.dataset.system is no longer rewritten on every 3 s tick. A tiny data-edit-key digest of the fields the edit modal actually reads (id / name / secret / tags / ipv4 / ipv6 / os / cpu_model / virtualization_type) gates the rewrite, so volatile traffic counters no longer invalidate the row's CSS cascade — visibly smoother in glassmorphism mode and on lower-end devices.

Public dashboard numeric stability

• Added a scoped tabular-nums rule for every data-bearing cell on both dashboards (.system-row .text-right, .system-row span.font-medium, [data-tcping-{avg,loss}], [data-traffic-tag] .traffic-value, [data-speed-tag] .speed-value, [data-detail-{speed,total}], total-net-{in,out}). Body prose stays proportional so modals and tooltips read naturally.

Admin list motion & glass-mode paint cost (follow-up polish in this tag)

• Removed the per-row stagger entrance and the scale-in animation on newly inserted rows — the outer page already runs .animate-content, and stacking another timed wave on top clashed visually with SSE refreshes.
• .draggable-item now transitions colors only (transition-colors), not transition-all, so routine cell updates don’t constantly re-sample backdrop-filter in glass mode.

Public table first paint

• The data-row-in stagger after the initial HTML paint is gated with a once-per-session latch. Transient empty metric snapshots used to clear the in-memory cache and replay the entire top-to-bottom fade when data returned — fixed.

TCPing chart & narrow screens

• After the debounced window resize handler (and on orientationchange), every live Chart.js TCPing instance receives an explicit chart.resize() so mobile Safari/WebViews keep canvas width aligned with the horizontally scrollable table.
• TCPing target tab buttons use touch-action: manipulation (with tap highlight suppressed) for snappier switching on phones.

TCPing chart open latency — server-side and client-side

• Store.GetTCPingResults now seeks bbolt's cursor directly to the 24h cutoff prefix instead of bucket.ForEach-ing the entire history bucket. The keys are already <unix-seconds>_<client>_<target> (10-digit timestamp, lexicographic = numeric until 2286), so the seek lands on the first in-window record and walks forward from there. On databases with weeks of multi-client tcping history this collapses the per-request cost from "JSON-unmarshal every record ever written" to "unmarshal records in the 24-hour window", typically a 10–100× speedup. The same trick was already used by CleanupOldTCPingResults — the read path now matches.
• The 50 ms setTimeout that gated chart initialization on row expansion is replaced with a single requestAnimationFrame so the fetch starts ~50 ms earlier on average without losing the "let the expand animation paint first" guarantee.
• The per-target chart payload is now built in a single linear pass that parses each result.timestamp exactly once — old code did .filter + .sort + .forEach per target, parsing every timestamp three times. Min/max time are tracked as actual extrema (not "first / last seen") so X-axis range stays correct even when a fork swaps in a backend with weaker ordering guarantees.

Final pass on data-fetch & first-paint hot path (this build)

• Server cache now stores pre-encoded JSON bytes. tcpingCacheEntry used to hold the full TCPingHistoryResponse struct; every cache hit therefore re-ran json.Encoder.Encode against the same data. The cache now stores the encoded body (with the same trailing \n Encoder.Encode would emit, so the wire format on hit and miss is byte-identical) and serves it via a writeCachedJSON helper that just Writes the bytes — measurably reduces CPU under load, especially on busy multi-client deployments where many users may open the same chart inside the 2-min cache window.
• Incremental update merge dropped from O(N log N) to O(K) per dataset per tick. The previous code did currentDataset.data.filter(within24h).sort(byTime) after each merge — reasonable when datasets are small but it ran every 3–30 s for every open chart. Because the server returns sorted data and we only ever append at the tail and expire from the head, the new code does two short directed walks (find tail-newer-than-last, drop expired-prefix) and splices in place. With a chart that already holds a few thousand points, per-tick cost becomes effectively constant; the array reference is preserved so Chart.js sees the mutation transparently and the post-merge min/max recompute / chart.update('none') / stats refresh remain unchanged.
• Chart.js entrance animation kept at the library default — the data lands on the canvas well before the curve animation finishes, so capping the animation didn't actually shorten the user-visible "first sample on screen" time, only the polish.

📚 Docs

README.md / README_EN.md install & uninstall sections now match exactly what the scripts drop on disk:

• Server uninstall removes /opt/pulse/scripts/ and the three pulse-{backup,restore,migrate} CLI shortcuts in /usr/local/bin/, with separate "keep data" vs "wipe everything" flavours.
• Linux client uninstall stops & removes the pulse-client-update.timer + .service alongside pulse-client.service, with a warning about not clobbering co-located server data.
• macOS client uninstall prose tightened to cover the com.pulse.client.update LaunchDaemon installed by the auto-updater.

📦 Docker Images

• xhh1128/pulse:1.3.7
• xhh1128/pulse:latest

📥 Downloads

Client binaries

Platform	File
Linux amd64	probe-client
Linux arm64	probe-client-arm64
macOS Intel (amd64)	probe-client-darwin-amd64
macOS Apple Silicon (arm64)	probe-client-darwin-arm64
Windows amd64	probe-client.exe
Install Script (Linux + macOS)	install.sh

Server standalone binaries (uploaded by the standalone workflow)

Platform	File
Linux amd64	pulse-server-standalone-linux-amd64
Linux arm64	pulse-server-standalone-linux-arm64

Quick Install (Linux & macOS):

curl -sSL https://raw.githubusercontent.com/xhhcn/Pulse/main/client/install.sh | sudo bash -s -- --id YOUR_ID --server http://YOUR_SERVER:8080

One-click migration to a new host:

PASSWORD="$OLD_ADMIN_PW" TOKEN="$NEW_ADMIN_PW" \
  bash scripts/migrate.sh --from https://old.example.com --to https://new.example.com

📝 Cumulative changes since v1.3.5

v1.3.7 builds directly on v1.3.6, which delivered the one-click migration & backup story (pulse-{backup,restore,migrate} CLI), full macOS client parity (Apple Silicon & Intel CPU model / cores / freq / mem / swap / disk / net), the cross-fade auth overlay, glassmorphism theme polish, the SEO/metadata pass, and the standalone server's Cache-Control parity with the Docker/Nginx build. See the v1.3.6 release notes for the full list — every fix below applies on top of those.