Skip to content

Visualizing a session

Render events.jsonl as a chat-style HTML page. Easier to skim than jq-ing the raw log.

This page is the practical how-to; for the bigger picture — why a fully replayable run matters and what the wider observability surface gives you — see Hyper-observability.

How to visualize

uv run tilth visualize                # most recent session
uv run tilth visualize <session_id>   # or name one explicitly

The pre-Phase-3 flag form --visualize still works for one minor version.

Writes sessions/<id>/chat.html — a single self-contained file (inline CSS, no JS) that renders the log as a conversation:

  • the seed context panel above the timeline when sessions/<id>/seed-meta.json exists — TL;DR, open questions, blockers, and scope notes from the prep-feature interview (with a head strip showing the interviewer model, total tokens, and start→end times), so the reviewer sees them before scrolling through per-task events,
  • a seed-prepared marker pinning the interview moment in the timeline (session_prepared),
  • model calls (with collapsible reasoning blocks where the model emitted any),
  • tool calls and results,
  • validator runs,
  • evaluator verdicts (accept / reject, with the rejection category, concern, evidence, and next-step on rejects),
  • proposed learnings,
  • commits,
  • and stops,

…all grouped by task.

The visualizer is read-only and runs over the saved events.jsonl, so it's safe to invoke against a finished or in-progress session.

What the output looks like

Sample chat.html render: session header, task divider, model-call meta-strip with an expanded reasoning fold-out, tool call and result bubbles

Diagram suggestion — annotated screenshot pointing to: (1) the session header strip, (2) per-task dividers, (3) a model-call meta strip with a collapsible reasoning block, (4) a tool-call/result pair, (5) a validator card, (6) an evaluator verdict card. Useful for orienting first-time readers of a long chat.html.

When to use it

  • After a clean run, for a quick scan of how the agent solved each task.
  • After a failed run, to see exactly where the loop diverged before a cap or an error.
  • Mid-run, to peek without disturbing the live process — chat.html is regenerated on demand from the still-growing events.jsonl.