Prototype the dashboard UI: fleet overview and instance detail #227

Activity

1. lbacik

   added

   wayfinder:prototypeDisposable prototype requiring human evaluationDisposable prototype requiring human evaluation

   on Oct 14, 2026
2. lbacik

   added a parent issue on Oct 14, 2026

   • Agent fleet dashboard: map #222
3. lbacik

   marked this as blocked on Oct 14, 2026

   • Define the telemetry snapshot contract #224
4. lbacik

   marked this as blocking on Oct 14, 2026

   • Define the dashboard project: repo, layout, stack and delivery slices #228
5. lbacik

   unmarked this as blocked on Oct 14, 2026

   • Define the telemetry snapshot contract #224
6. lbacik

   marked this as blocking on Oct 14, 2026

   • Define the telemetry snapshot contract #224
7. lbacik

   self-assigned this

   on Oct 14, 2026
8. lbacik commented on Oct 14, 2026

   

   lbacik

   on Oct 14, 2026

   OwnerAuthor

   More actions

   Resolution

   Decided by reacting to a throwaway UI prototype with three structurally different variants (ops table + detail drawer, attention-sorted budget lanes, inbox + instance cockpit) on a mocked fleet covering: soft threshold crossed with handoff in progress, recovery hold, consecutive infrastructure errors, unreachable instance with a stale snapshot, exited container, and an older agent without telemetry (version skew).

   Prototype (primary source): branch prototype/fleet-dashboard-ui, file prototype/fleet-dashboard-ui.html (open in a browser; ?variant=A|B|C). The variant chosen is A; the decisions below adjust it.

   Layout: variant A (ops table + detail drawer)

   Fleet overview is a dense table, one row per managed instance, with columns:

   1. instance (container name + target repository)
   2. status headline: one sentence of what the instance is doing or needs (e.g. "Recovery hold on token_estimate_degraded fires on the main thread when usage arrives after the tool result #233: push rejected", "Soft threshold crossed, handoff delivering on #318", "2/3 consecutive infrastructure errors", "Unreachable: docker exec timed out after 5 s", "Waiting for a ready issue"). Borrowed from variants B/C: without it the operator has to assemble the picture from several columns.
   3. reach (live / stale / unreachable / not running), with snapshot age when not live
   4. intake
   5. attempt (issue number + phase, or the recovery hold)
   6. budget (bar from 0 to the hard ceiling with a soft-threshold tick; hatched when partly estimated) + used / hard ceiling
   7. main context (+ number of running subagents)
   8. turns
   9. cache hit rate
   10. consecutive errors (count / max)

   Default sort: attention first (borrowed from variant B): unreachable / recovery hold / consecutive errors / handoff in progress, then approaching the soft threshold, then working, then idle, then offline. Not alphabetical.

   Instance detail is a right-hand drawer, sections in this order:

   1. current attempt (first: the section operators reach for most): budget used with measured/estimated split, soft threshold value + rule + crossed-at, hard ceiling, turns (+ continuation index), cache hit rate, budget bar, tokens by category (detail only, never in the table), main context vs window with peak, per-subagent contexts, provider rate-limit headroom shown as "unavailable in v1".
   2. control: repository, agent version, intake, active attempt (issue number, issue title, branch, phase, started), stop plan, handoff (phase + model deadline), recovery hold (reason + ready-to-copy next action), consecutive errors, last finished attempt (issue, outcome, tokens, finished at, PR).
   3. recent events: last N log events with level, event name and one-line text.

   Signals cut from v1

   • USD cost: dropped. It is a CLI estimate; the token budget is the real limit.
   • Charts / sparklines: dropped (variant A has none; history and trends stay in the map's fog).

   Signals kept

   • Per-figure provenance markers (measured / estimated / configured), following the provider-limits research. Unavailable figures render as "unavailable", never as zero.

   Requirements passed to "Define the telemetry snapshot contract" (#224)

   Fields the chosen UI needs that today's agentctl status does not carry:

   • schema_version and a snapshot timestamp (the dashboard also measures age on its side).
   • active_attempt.issue_title.
   • Current-attempt telemetry: budget used / measured / estimated, soft threshold value + rule + crossed-at, hard ceiling, tokens by category, main context, peak main context, context window (configured), subagent contexts, turns, continuation index, cache hit rate. No USD figure.
   • Provenance per figure (measured / estimated / configured) and an explicit "unavailable" for missing figures.
   • recent_events (bounded list: timestamp, level, event, one-line text).
   • last_finished attempt (issue, outcome, tokens, finished at, PR).
   • recovery.next_action as a ready-to-copy command.
   • Enough state to derive the status headline and attention rank dashboard-side (recovery, handoff phase, consecutive errors, intake, phase, budget vs soft threshold). The headline itself is computed by the dashboard, not sent by the agent.

   Dashboard-side only (not in the snapshot): container name/id, running/exited, exec ok/error, snapshot age, attention rank and status headline.

   No further fields were requested beyond the above. No new tickets or fog surfaced.
9. lbacik

   closed this as completedon Oct 14, 2026