Spark Scope 0.1.4

Added

• Apple Silicon Mac nodes, collected locally or over SSH with the built-in macOS tools and no sudo (#45):
  • GPU load, shared system memory, swap and compressed memory, disk and the inference process;
  • thermal state, MacBook system power and GPU memory in use in the slots for GPU temperature, GPU power and clock;
  • memory warnings that follow the OS memory pressure level.
• oMLX, read from its status API (#47):
  • running and waiting requests, cache hit since start, and mean prefill and decode speeds of completed requests;
  • polling never loads a model or resets oMLX's idle timer.
• Optional engine API key, sent as a Bearer header on every engine read (#47):
  • SPARK_SCOPE_API_KEY for a single server, or apiKeyEnv per server in topology.json, naming an environment variable that holds the key;
  • the key never reaches the pages or the logs. Nothing changes if your engine needs no key.
• llama.cpp, from llama-server --metrics (#42, #46):
  • live output speed from /slots while requests run, cache hit rate, mean decode time and context used;
  • an idle server is never woken by polling.
• Strata, with TTFT and TPOT p95 from its histograms and context used (#43, #46). A server that has unloaded its model reads as idle.
• GPU memory for machines with their own VRAM, such as a workstation next to the Sparks (#41):
  • "GPU memory" on the node card and "VRAM" on the mini window and the rack panel;
  • gpu.memory in /api/state; memory stays system RAM;
  • npm run demo -- --discrete adds one to the made-up data.
• Calendar: a switch between Output, New input + output and Total tokens, remembered in the browser (#38).
• Mini window in Safari and on phones: the page itself switches to the mini view, and a phone opens in it (#35):
  • a back arrow returns to the full dashboard;
  • in a window at least 720 px tall, Glance, Scope and Runs stack instead of sitting behind tabs;
  • the separate popup window for other browsers is gone; /mini/ still works.
• README: a prompt for installing with a coding agent, and the dsh-spark-scope plugin for the DeepSeek Harness sidebar.

Changed

• The engine panel, the mini window and the rack panel hide the fields an engine never reports instead of showing unknown (#46):
  • /api/state adds reported and metricKinds to each inference reading;
  • TensorFold 1.0 (native server) counts live and finished output for decode, and reads TPOT and prefill from its histograms;
  • vLLM and SGLang hide speculative acceptance while speculative decoding is off.
• Up to eight nodes and several model servers (#44):
  • the node grid fills its rows without an empty cell;
  • the model servers share one strip under the status line;
  • the mini window lists each server with the total;
  • the rack panel shows bays for the first four nodes, and its band still counts every node.
• Clearer wording on the pages and the rack panel, in English and Korean (#36), for example "Turn orange when", "Cache hit rate" and "3 kernel errors".
• npm run check holds the repository rules, and one "CI passed" check gates merges (#39).

Fixed

• A failed system service whose unit file was gone, such as the mount of a removed snap revision, kept the "failed system service" warning until a reboot (#37).
• On engines without latency histograms, TTFT and TPOT changed to "no requests" after the second poll (#46).
• The mini window's GPU power chip read 0 W when no node reported power (#45).
• npm run demo -- --mode fault failed with 7 or 8 nodes (#44).

Spark Scope 0.1.3

Added

• Model servers: nodes that serve in separate groups (two nodes used one by one, four as 2 + 2, three as 2 + 1) are listed under servers in topology.json, each with its own API (#32, #33):
  • a row per server above the chart, and its server on each node card;
  • All at once (a line and an engine panel per server, the total on top) or One at a time, in the settings;
  • a chip per server on the rack panel's band, ?server= to follow one;
  • one token ledger for every server;
  • npm run demo -- --servers 2 shows them on made-up data.
• Keyboard shortcuts: S Scope, L token ledger, M mini window, , settings, ? the list of them (#33).

Fixed

• The mini window's M did nothing while a Korean keyboard layout was on (#33).

Spark Scope 0.1.2

Added

• TensorFold alongside vLLM and SGLang: its metrics, and on CUDA the /health counters for live output speed, cache hit rate and prefill rates (#29).

Changed

• SPARK_SCOPE_API_URL must be an http:// or https:// URL without a user name or password; otherwise the server stops with a message (#30).

Fixed

Fixed in #30:

• Settings: the Readings dropdowns were pushed to the right in a narrow dialog, and on a short screen below 1050 px the choices slid behind the preview.
• Kernel errors from before a reboot were not counted, and a journal the account cannot read showed 0 errors instead of "unavailable".
• Token ledger:
  • a clock that went back booked a restarted run again on every poll;
  • the page could open on the viewer's month instead of the server's and did not follow the month change;
  • the input chart was empty when the engine does not split cache read and new input.
• Rack panel:
  • the band jumped back and forth and drew dips that never happened;
  • the footer cut the power reading short;
  • the low-memory reason ignored mem=gb.
• Web page:
  • the trend legends kept old values while the server was down;
  • a custom node colour did not update its row until the focus moved;
  • dragging a selection out of the settings closed them.

Spark Scope 0.1.1

Added

• Settings dialog, from the gear button in the header (#14, #18, #19, #21):
  • node card: the four readings and their order, the bars, full or short labels, the levels that turn a reading orange;
  • a colour per node, from the palette or custom;
  • units: °C or °F, GiB or GB, 24- or 12-hour clock;
  • dashboard: panels to show, chart range, refresh interval, pausing hidden tabs, theme;
  • rack panel: band motion and the kiosk URL;
  • a settings link that carries them to another browser.
• Designs: Default, Console and Soft (#23).
• Korean for the web page and the rack panel (?lang=ko) (#16).
• Mini window (M): glance, scope and recorded runs, kept on top in Chrome and Edge (#24).
• Token ledger: statement, calendar and charts, a table by model, CSV export (#26).
• ? explanations next to the figures that need one (#17).
• Rack panel options in its address: temp=f, mem=gb, colors=, motion= (#21).
• npm run demo: the pages on made-up data, without a Spark (#25).
• API: /api/state adds version, messageKey and messageParams, the 5-minute latency fields, prefillUpdatedAt and rackSeenAt; /api/usage adds each day's and the month's models and firstDay (#15, #16, #21, #24, #26).

Changed

• TTFT and TPOT p95 cover the last 5 minutes; the value since the engine started is behind the ? (#15).
• A hidden web tab pauses its updates and catches up when shown (#13).
• The rack kiosk uses about a third of the CPU: the band moves in one-pixel steps and the kiosk turns off Chromium's renderer accessibility. Reinstall kiosk/spark-scope-kiosk to get the second part (#13).
• A shorter README with a Korean version; the reference moved to docs/ (#11, #27).
• CI renders the pages and fails on clipped text, overflow or script errors (#12). A shorter contributing guide; English is enough for a pull request (#20).

Fixed

• "GPU temperature" wrapped onto two lines on narrow cards and pushed the readings out of line (#18).

Spark Scope 0.1.0

The first tagged release. Spark Scope went public on 2026-10-01; this covers everything changed since.

Added

• CPU load, NVMe and ConnectX NIC temperatures, every ACPI thermal zone and the total GPU power, on the web page and the rack panel (#3).
• A favicon and home-screen icon (#5).
• Rack panel: a compact layout for five or more nodes and narrow screens, ?width down to 800, a centred letterbox. Web page: state-coloured badges, a three-way theme button (system, light, dark), a Logical input column in the daily table, eight node colours (#8).
• SPARK_SCOPE_ALLOWED_HOSTS and a Content-Security-Policy on every response (#9).
• CI on x64 and arm64, contributing and security notes, issue templates (#10).

Changed

• /api/state calls the engine metrics inference (vllm is still sent as an alias and will be removed later). It no longer sends the engine URL, the local model path, interface names or raw SSH error text (#9).
• Requests that name another site's host are refused (#9). If you reach the dashboard under a name other than localhost, an IP address or the machine's own hostname, add it to SPARK_SCOPE_ALLOWED_HOSTS.
• Latency p95 is interpolated inside the histogram bucket, as Prometheus does (#9).
• Token counts use three significant digits on both pages (1.5K, 9.55M) (#8).
• The pages poll without history every 2 seconds and fetch it every 30 seconds; responses are gzip-compressed; the rack band moves with a CSS transition instead of a frame loop (#6).
• The topology is looked for in ~/.config/spark-scope/topology.json first (#4).
• The kiosk starts Chromium with --class=spark-scope-kiosk and the labwc rule matches only that window. Change an existing identifier="chromium*" rule to spark-scope-kiosk (#10).

Fixed

• iOS Safari scrolled the page on every poll (#1).
• A hung nvidia-smi or docker emptied the node card and left processes behind. An SGLang restart could drop tokens from the ledger. Metrics an engine does not export were stored as 0. vLLM data-parallel engines were read as one. Interval settings such as 2s were accepted (#2).
• The kiosk hung for host names starting with "rack" and failed to start over SSH. The README quick start missed steps (#4).
• Stale-data, drawing-error and ledger-month handling on the web page and the rack panel (#6).
• SGLang's decode rate read 0 while requests ran and spiked when they finished (#7).
• Long names, rounding at unit boundaries, contrast below AA and a node id at breaking the trend charts (#8).
• Unread engine responses held connections open. A mistyped interface name read as a broken link (#9).