TranscriptViz k3 b
kimi/k3 · k3-b
Who/what caused what — delegation trees, tool→file→edit chains, error cascades, trigger graphs. “Why did the session end up here?”
The session as a provenance DAG — click any node to flood-trace why it happened (amber ancestors) and what it caused (teal descendants), with semantic zoom into the retry cascade's internal chain.
Comprehension Value — Answers “why did the session end up here?” directly: click any entity and its upstream chain floods amber while consequences glow teal, sidebar spelling out the chain (GOAL COMPLETE ← CI green ← job_ci rerun ← T147 fix residue). Turns the 287-turn log into a queryable cause-and-effect structure.
Long-Session Scaling — 287 turns collapse to ~16 key entities; graph size tracks structural complexity, not turn count. Semantic zoom expands the retry cascade in place; dimming keeps large graphs legible during traces.
Interactive mockup (raw HTML — download & open locally)
Actor swimlanes over wall-clock time — every actor (user, root, each delegate, each job, watches, context pressure) gets a lane; tool calls are density currents and cross-lane arcs carry the causality.
Comprehension Value — Answers “who was doing what when, and which events crossed actor boundaries to cause others?” The watch storm reads as three converging red arcs from job_fuzz into ROOT's error diamonds; the CONTEXT lane shows pressure climbing to 92% then dropping at each fold — compaction causation made visible.
Long-Session Scaling — Rendering cost is O(minutes × lanes), independent of turn count: 2,146 tool calls render as density currents; the overview-strip brush zooms from 6h12m down to the 90-second watch storm (verified), macro shape and micro incident one drag apart.
Interactive mockup (raw HTML — download & open locally)
A forensic ledger of 14 auto-clustered episodes (cause→mechanism→outcome strips with evidence chips and causal link pills) under a blast-radius file×time heatmap — the chain of custody for the whole session.
Comprehension Value — Answers “what were the ~14 things that actually happened, what caused each, and what did the session touch?” Non-adjacent causality is explicit (E8 retry cascade traces to E2's T44 diff); the heatmap shows which files absorbed edits, where failures landed, where folds sit.
Long-Session Scaling — Ledger grows with auto-detected episodes (~14 for 287 turns), not turns — a 2,000-turn session stays a one-screen scroll. Heatmap is files × 48 time buckets regardless of length.
Interactive mockup (raw HTML — download & open locally)