Reading a replay
What a recording holds, how it was made, and how to get from the overview down to a verbatim prompt.
What a recording is
Every driver that runs the real agents writes one at the end of its run, built from the trace the run already keeps (every agent run and step, every enqueue, every model call) and the graph tables. Nothing in it is authored by hand; where attribution is a heuristic, the record says so on each change. The committed recordings are played back on this site.
Five levels, each one click from the next: the episode (what it tests, the commands, the fingerprint, the scenario, the distinct system prompts used); the timeline and the graph building as events land, two arms side by side for a two-arm episode with the paired claims linked; an event (its agent, trigger, cost, what caused it and what it caused, what it read and wrote); a step, verbatim (the prompt step shows the system prompt, the initial messages and the tool definitions; a tool call its full input; a result its full output); and a claim’s lineage (the source’s verbatim text, the Extractor’s proposed form, the Matcher’s decision with the search results it saw, the Steward’s edges and verdicts, its credence over time). Every level has a link.
Detail
How it is made
- Every model call in a tool-using loop or a single-shot completion records a prompt step first: model, effort, token limit, the full system prompt, the initial messages, the tool definitions.
- Every assistant turn and every tool result is recorded as a step, untrimmed, capped only by the trace’s per-step limit, which the recording marks.
- After the drain the exporter reads the window: one event per agent run in time order, harness events for each source submitted, changes to the graph credited exactly where a decision tool named them and by run window otherwise, causality from the enqueue events, cost from the usage rows.
- Two layouts are written: an index with one-line gists and sizes, and one file per event with everything, fetched only when opened.
What it cannot show
Anything the trace did not record: a run traced off yields events with no steps, and the player says so. A recording is one run; it shows what happened, not what usually happens.
Run it
npm run corpus:run -- blackholes --limit=1 # writes runs/<run>/replay.json
npm run corpus:replay -- db --since=<iso> --name=<name> --commit-as=<name>
npx tsx scripts/sync-frontend-content.ts # vendors corpus/replays/ into the site