llama-watch/docs/superpowers/specs/2026-08-14-inflight-activity-design.md

In-Flight Monitor — Count + Activity Spark

Date: 2026-08-14 Status: Approved (design)

Summary

Enhance the In-Flight Monitor action so its key shows the number of in-flight requests for the configured model plus a count-trend spark: a tiny line/area chart of that count over the last ~60s, so activity is visible even during a single long-running request.

Layout (72×72 SVG, rendered via renderInflight)

• Top: model short name (small, white) — existing.
• Center: the status. When ready:
  • count > 0 → the count number, large/bold (e.g. 3), on the red ACTIVE background, with the existing pulse.
  • count === 0 → IDLE on green — existing.
  • loading → LOADING, stopped → OFF, offline → OFFLINE, unconfigured → NO MODEL, unknown → … — all unchanged.
• Bottom strip (~14px): the count-trend spark — a small polyline + faint area fill of the model's in-flight count over the last 60 samples, in the state color. Flat at the baseline when idle; spikes when requests come and go.

Data Flow

• The feed side is unchanged — InflightTracker already exposes exact per-model counts.
• Each key's per-key state (in inflight-monitor.ts, keyed by ev.action.id) gains:
  • history: number[] — a 60-sample ring buffer of the live count.
  • sampler?: ReturnType<typeof setInterval> — a 1s interval that pushes runtime.tracker.count(modelId) into history and re-renders. Started in onWillAppear, cleared in onWillDisappear.
• Live feed events (add/remove/snapshot) still re-render the big number immediately via the existing subscription; the spark refreshes at 1Hz.
• renderInflight gains an optional history?: number[] field. Spark scale: yMax = max(max(history), 1) so a zero line sits at the baseline.

Files

• src/lib/render.ts — renderInflight accepts history and draws the spark; center label shows the count when active.
• src/actions/inflight-monitor.ts — per-key history ring buffer + 1s sampler.
• tests/render.test.ts — new/updated assertions for count rendering and the spark.

Error Handling / Edge Cases

• Spark with <2 samples renders no line (like the GPU chart).
• Sampler cleared on onWillDisappear (no leaks); multi-key safety preserved (buffer + sampler live in the per-key map entry).
• History capped at 60 samples.

Out of Scope

• No change to the GPU Graph action, the feed, or the tracker.
• No throughput-bar variant (count trend was chosen).