perf(webui): bounded transcript DOM for long sessions

[jkyberneees]

What

Long WebUI sessions no longer slow down as the transcript grows.

The old 80-item cap only counted top-level rows. One turn's stream, holding hundreds of tool steps and each step's eagerly built result view, was never trimmed. Opening a large session also rendered every message before pruning.

• Transcript budget (js/transcript.js, new)
  • Caps the transcript at 300 rows, counting the rows inside turn streams.
  • Live work is never trimmed: the streaming answer, open reasoning, running sub-agents, and tool steps still waiting for a result.
  • While you are scrolled up reading, trimming waits until 600 rows. It then removes only rows wholly above the screen, and the view stays put. Past 1,200 rows the oldest rows go regardless.
  • A note at the top of the transcript says when rows are hidden.
• Windowed session load
  • Opening a session renders its newest ~150 rows.
  • Show earlier renders the previous window above, in order, with a split turn rejoined, and the visible rows don't move.
  • Rows of a loaded message are trimmed together, and the backlog resumes at the first remaining message.
  • Rows trimmed from a live turn come back via Reload earlier history once the turn ends. The reload re-renders only the transcript; the connection, queue and plan are untouched.
  • Tool results outside the window still reach the Activity/evidence panel, in transcript order.
• Lazy tool results: a step's result view (toolbar, search, line rows) is built on first expand.
• Fold measurement: older answers are measured once instead of on every new reply. Previously each new reply forced a layout per assistant bubble.
• Row caps: reasoning blocks keep their newest 400 lines and sub-agent chip logs their newest 200. Older lines are summarized in a note.
• Activity list: repaints once per frame, keeps a running byte total and caches each result's kind.

Measured (headless Chromium, odek serve, synthetic session)

	main	this PR
Open a 600-turn session (3,601 messages)	18.5 s, 10,855 nodes	0.43 s, 2,606 nodes
Live turn, 1,000 tool steps: transcript nodes	67,001	3,004
Live turn, frame-paced step time, steps 100 → 1,000	16 → 30 ms (below 60 fps)	flat ~16 ms

Browser checks:

• Show earlier while scrolled mid-transcript: a visible row stays at the same offset (1981 → 1981 px).
• Trim at the reading ceiling: the row being read stays at the same offset (383 → 383 px).
• Message order after paging back is complete and ascending, older long answers fold, and there are no console errors.

Tests

• New cmd/odek/ui/js/transcript.test.js (32 tests). Every regression test was RED on the code before its fix; the one exception is a live-work invariant that already held.
• Full JS suite: 379/379 pass. go test ./cmd/odek WebUI tests pass, including the JS↔CSS class contract.
• Line coverage:

File	Before	After
render.js	85.5%	89.8%
workspace.js	47.9%	74.5%
utils.js	68.1%	74.1%
sessions.js	67.1%	68.3%
transcript.js	new	99.0%

• Three adversarial review passes (an independent first-commit review, /code-review high, /code-review medium). Every confirmed finding is fixed in commits 2–4, each with a regression test.
• Two findings were declined:
  • Ctrl-F cannot see unexpanded tool output, but collapsed .tb-body was already display:none, so this is unchanged.
  • The parsed session is kept for Show earlier. That is bounded by the session-file cap and needed to recover trimmed rows.
• Every browser scenario above was re-checked after each round.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	odek	d910baa	Commit Preview URL Branch Preview URL	Oct 17 2026, 08:20 PM