kimi/k3 · k3-b

What is causing what, right now — causal structure that forms live as the session runs.

The provenance DAG materializes live — nodes appear with pulsing halos when work starts, causal edges draw themselves at the moment one thing causes another, and a WHY-NOW rail answers "what triggered this?" continuously.

Realtime Comprehension — "Why is the agent doing this exact thing right now?" The WHY-NOW rail walks the current node's upstream chain over edges that exist so far; the alert rail flags a watch runaway at the third fire, not in hindsight; you watch the "bug written @T44" edge land on ROOT CAUSE live.

Long-Session Scaling — The graph accumulates ~21 entities for 287 turns — it tracks causal structure, not turn volume. Completed nodes freeze solid and dim; the eye is pulled to the 1–2 pulsing in-flight nodes. Full 6h12m replay at 128× takes ~3s; seek rebuilds any point.

Interactive mockup (raw HTML — download & open locally)

Actor swimlanes with a sweeping playhead and a growing time axis — lanes fade in as delegates/jobs spawn, jobs render as hatched bars growing until their exit event, watch fires and folds land live.

Realtime Comprehension — Who is active right now, what just fired (watch triangles drop with arcs to their handler turns), whether a runaway/cascade is developing (RIGHT NOW strip flips to blinking red the moment the pattern forms); the CONTEXT lane makes each compaction legible as it happens.

Long-Session Scaling — Rendering cost is O(elapsed minutes × lanes), never O(turns): 2,146 tool calls accumulate as density currents. Lanes appear only when actors exist; the whole 287-turn session stays one screen at 1×–128×.

Interactive mockup (raw HTML — download & open locally)

The forensic ledger writes itself — episodes open as amber-pulsing cards with live action lines, seal with ✓/✗ outcomes, then collapse into one-line custody rows as newer episodes arrive.

Realtime Comprehension — What episode is in flight, what it just caused, is the session healthy? Open cards tick live (elapsed, ~calls, current tool) with "pending — episode in flight…" until the seal event; causal link pills flash in the instant their trigger lands; alert strip badges anomaly onset AND resolution. Multiple cards can be open — parallel work is honest.

Long-Session Scaling — Sealed episodes auto-collapse to one-line custody rows once two newer episodes exist: at T287 the ledger is 12 compact rows + 2 full cards — one screen regardless of length. Heatmap is files × 48 buckets regardless of run length.

Interactive mockup (raw HTML — download & open locally)

---

← Back to index · Other round: retrospective set