RealtimeViz k3 a
kimi/k3 · k3-a
The whole session-so-far's shape at a glance, filling in live — the minimap as an always-on HUD.
The owner's minimap as a live HUD: five semantic lanes fill in stratum-by-stratum under a head marker, flanked by a NOW card, pressure gauges, live delegate/job/watch boards, and a transcript tail with streaming in-flight chips.
Realtime Comprehension — While the session runs: what the agent is doing right now (NOW card with current action/tool/elapsed), what just changed (event feed), whether it's looping or about to fold (alerts + context gauge with fold threshold), state of every delegate/job/watch — plus the accumulating whole-session shape as strata with fault lines and anomaly diamonds.
Long-Session Scaling — Past strata are immutable and only shrink in height (278 turns ≈ 2px each); all live cost concentrates at the head marker and O(actors) rail panels, so glanceability holds at 250+ turns and beyond. Click any past stratum to pin it.
Interactive mockup (raw HTML — download & open locally)
A wall-clock swimlane trace that grows as the session ages: context-pressure band cliffs live at each compaction, in-flight delegate/job bars extend until they land, NOW cursor at the edge.
Realtime Comprehension — What is running in parallel and for how long — delegate and job bars lengthen in real time (animated while active, red on failure), the context band climbs and cliffs at each fold, a watch runaway visibly packs into a barcode; hover any bar for elapsed time and status.
Long-Session Scaling — Lane count grows with actors, not turns; the x-axis starts at a 20-minute window and rescales as the session ages; completed blocks compress into texture while anomalies stay distinct at full 6h42m / 278-turn length.
Interactive mockup (raw HTML — download & open locally)
An ops console for the running agent: loop/runaway/stall/jobs verdicts, throughput sparklines, and projected time-to-fold and time-to-budget-wall, over a clickable whole-session strip.
Realtime Comprehension — Is it healthy? Four explicit verdicts (loop detector, watch runaway, stall check, jobs healthy) with one-line reasons, a big in-turn timer, tokens/min + tools/min sparklines that show a breakage episode as a dip and a step, and rate-based ETAs for the next fold and the budget wall.
Long-Session Scaling — Every panel is fixed-size: sparklines compress from a once-per-second sample buffer, the session strip is one pixel-column per turn, verdicts recompute from O(1) sliding windows — degrades least of the set at thousands of turns.
Interactive mockup (raw HTML — download & open locally)