Decision-first trace detail
Verdicts, confidence, evidence, and replay — on a frozen trace.
Trace introspection against a frozen bugfix-trace snapshot: observation-kind filter chips, the decide verdict block with confidence bar and collapsible evidence drawer, the deterministic-replay affordance, and the tree / waterfall / state-graph view modes.
Can a run be audited from trace data instead of a chat transcript?
Decision detail, evidence, confidence, replay, and the trace view modes.
The recorded trace is the source of truth for replay, inspection, and tests.
demo video not rendered in this build make demo-feature FEATURE=trace-introspection
Key beats
This is a deterministic bugfix run, loaded from a frozen trace snapshot — no live server, no LLM. Everything you see is derived from the recorded trace, the auditable source of truth. This tour walks the introspection surfaces: observation kinds, decision-first detail, and the three view modes.
Every event carries an observation kind — agent call, decision, host call, world update, narration. Each kind gets a colored dot on its row and a filter chip up here; toggle a chip to fade that category in or out and isolate the part of the run you care about.
Expand a decide call and the verdict leads: accept or reject, the chosen intent, and the runner-up moves that were on the table. The decision the judge reached is the first thing you read — not the last thing you scroll to.
The confidence bar quantifies how sure the judge was. A low bar on an accepted verdict is exactly the kind of soft pass an operator wants to eyeball — it's surfaced, not buried in the raw attrs.
The evidence drawer is collapsed by default so the verdict stays clean — click it to reveal the exact system prompt, prompt, and response the judge saw. Decision first, full provenance one click away.
Full recorded walkthrough (10 steps)
Because the call replays byte-identically from its cassette, the Replay affordance can re-run the recorded decision deterministically — the same prompt, against a chosen operator — with no cost and no drift.
The same trace renders three ways — the default tree, a latency waterfall, and the state graph. Same source of truth, three lenses; switch with these tabs.
The waterfall lays every call out on a time axis — each bar's width is its duration, so the slow agent turn or the long host call stands out at a glance. The offsets come straight from the trace timestamps, never re-derived.
The graph view redraws your route through the story's state machine from the trace — where the run has been, where it is, and where it can go. The same machine the metro stepper and ego graph render, here as the whole picture.
Observation kinds, decision-first detail with confidence + evidence + replay, and three view modes — every surface derived from one recorded, auditable trace. Hit '?' anytime to replay this tour.