TranscriptViz glm52 a
lunaroute/glm-5.2-vision · glm52-a
Matrix/heatmap/glyph encodings that make thousands of events scannable as patterns before any clicking or reading.
Every turn becomes one column in a fit-to-width ribbon so bursts, stalls, failure clusters, and compaction boundaries read as visual patterns before any clicking.
Comprehension Value — Answers “what happened across the whole session, turn by turn, and where are the phase boundaries, failures, and compaction folds?” — column hue = turn kind, inner bar = tool-call intensity, amber dividers = compaction, red corner glyphs = failures, amber dots = steering. Delegates and jobs get lanes beneath, sharing the time axis.
Long-Session Scaling — One column per turn: a 263-turn session fits 2 rows × 132 columns; a 1,000-turn session becomes ~8 rows — always one screen, never scroll.
Interactive mockup (raw HTML — download & open locally)
A dense activity heatmap where rows are tool types, columns are time-buckets, and cell intensity encodes call count — phase transitions and tool-mix shifts as spatial patterns.
Comprehension Value — Answers “how did the tool mix shift, which tools dominated each phase, where did cost concentrate, where is the error episode?” Exploration, delegation, and integration phases read as distinct heat regions; the single red cell is the error/retry episode; amber bands are compactions.
Long-Session Scaling — 50 buckets × 10 tool types keeps the grid to one screen regardless of session length; adding buckets or rows only changes cell size, not layout.
Interactive mockup (raw HTML — download & open locally)
The 3-level delegation tree as an icicle with stacked micro-ribbons showing tool-call density, token cost, and wall-clock duration inside each node.
Comprehension Value — Answers “how is cost (tokens, wall-clock, tool calls) distributed across the delegation tree, and which branch owns most of the session?” — the widest purple delegate bar owns the most; its cyan grandchild carries the red error ribbon and a tall duration ribbon showing recovery time.
Long-Session Scaling — One node per session element, width proportional to turn span: a 3-level tree fits one screen regardless of depth; fat nodes with tall ribbons show where tokens and time concentrate.
Interactive mockup (raw HTML — download & open locally)