@martindye/dsh-model-dial

A live dial for the local llama.cpp model, shown in the DSH web app's session header: a 34 px ring that shows what the model is doing right now.

llama-server (llama.cpp) streams no progress over HTTP, but it tells the truth in two places this plugin reads:

• GET http://127.0.0.1:8080/slots — phase (idle / prefill / decode), the task id, prompt and generated token counts, the request's own params.
• the server's own log file (llama-server-*.log, UTF-16) — the print_timing lines carry the prompt-processing progress % and the decode rate.

On top of that, the host half tails the DSH session log (session.v4.jsonl.zstd) for the compaction/start … compaction/end pair, so the minutes the model spends summarising the old context (prompt compaction) are surfaced too — with the summary's live percentage.

What the dial shows

• prompt processing — an amber ring filling to the server's own progress %, the % in the middle.
• generating — a small "generating" label while tokens decode.
• compacting — the harness summarising old context: an amber ring at the summary's live progress (tokens generated over the summarisation budget the request was given), % in the middle; a blind slow spin while no progress is known yet (the first seconds).
• idle / down — a grey ring with a dot; hover for the detail (no active processing, or the server unreachable at 127.0.0.1:8080).

The dial is decorative: nothing it does may ever take the host down, and a dead server simply reads "down".

Layout

• src/index.ts — the host half (runs in the DSH host process): polls /slots every 750 ms, tails the newest llama-server log and the DSH session log, and publishes the modelDial.status Remote.
• src/sessionlog.ts — the zstd-frame walker and the compaction-event parser for the session log.
• src/client/Dial.tsx — the 34 px ring in the session header.
• lib/ — build output (lib/index.js node half, lib/client.js browser bundle). Rebuild with pnpm bundle.

Installing

From npm (the published package ships prebuilt lib/ — no build step):

dsh plugin --profile <profile> add @martindye/dsh-model-dial

for example dsh plugin --profile web add @martindye/dsh-model-dial for the web app. Then restart the web host and open the app: the ring appears in the header of an open session (it is a per-session control, so the landing view has no dial).

Requirements: DSH 0.1.7+ (web app), Node 22+ on the host, and a local llama.cpp server on 127.0.0.1:8080 — the dial is a viewer, it drives nothing.

Configuration (env overrides, read at host start)

• DSH_DIAL_PORT — the llama-server port, default 8080.
• DSH_DIAL_LOG_DIR — the llama-server logs directory (the dial needs the llama-server-*.log files for the prefill %).
• DSH_DIAL_SESSIONS_DIR — the DSH sessions directory, default ~/.dsh/sessions (needed for the compaction indicator).

Changelog (the latest updates)

• 24/10 — The compaction ring shows a live percentage: summary tokens generated over the summarisation budget, read straight off the model slot; the blind spin now only covers the first seconds.
• 22/10 — Compaction indicator: an amber "compacting" ring while the harness summarises the context (the session log's compaction/start … compaction/end pair), and an attach-race fix so a session switch or a host restart mid-compaction still lights it.
• 11/10 — Gauge-only face: a 0–1 progress ring from the server's own progress = P lines (live while a line is younger than 3 s); the speeds and totals live in the app bar, not the dial.
• 10/10 — First version: the 34 px ring in the session header, reading /slots plus a tail of the llama-server log.