Window long-session message rendering

[Michaelyklam]

Thinking Path

• Long Hermes WebUI sessions lag because every visible transcript message is rebuilt into the DOM.
• Issue feat(perf): DOM windowing / message virtualization for long sessions #734 asks for DOM windowing, but the maintainer comment recommends a safer first pass before full IntersectionObserver virtualization.
• The current code already has server-side message pagination for unloaded history; this PR adds an in-memory DOM render cap on top of that.
• The core constraint is preserving the existing render pipeline: markdown, thinking/activity groups, tool cards, timestamps, cache, and live streaming still need to anchor to the correct raw message.
• This PR keeps the session transcript in memory, renders the latest 50 visible messages by default, and provides a top chip to reveal earlier in-memory messages one page at a time.

Closes #734.

What Changed

• Added a default 50-message DOM render window inside renderMessages().
• Added a Load earlier messages (N hidden) chip above the rendered transcript window.
• Expanding the local DOM window preserves scroll position by measuring scrollHeight before/after inserting older messages.
• Preserved the existing server-side _loadOlderMessages() path when the current in-memory session is already fully rendered but older server-paginated messages still exist.
• Updated assistant/tool/thinking anchoring so hidden older tool activity is not accidentally attached to the newest visible assistant turn.
• Included render-window size in the session HTML cache key and rewired the cached load-earlier button after cache restore.
• Added source regression coverage for window size, local load-earlier behavior, scroll preservation, streaming/tool anchoring, cache rewiring, and chip styling.
• Added UI evidence under docs/pr-media/734/message-window-top.png.

Why It Matters

• Long sessions get the primary performance benefit immediately: fewer transcript nodes are built and re-highlighted on initial render.
• The implementation stays within the existing vanilla JS architecture and avoids the bigger risk of full virtual scrolling/unloading.
• Users can still access older transcript messages explicitly without dropping session data from memory or disk.

UI media

Synthetic 121-message browser validation, showing the windowed transcript starting at message 21 with 21 earlier messages hidden:

Verification

• node --check static/ui.js
• node --check static/sessions.js
• /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_issue734_message_windowing.py tests/test_auto_compression_card.py tests/test_1325_user_fenced_code.py tests/test_ui_tool_call_cleanup.py tests/test_turn_duration_display.py tests/test_parallel_session_switch.py -q -> 79 passed
• git diff --check
• Browser QA on isolated local WebUI (127.0.0.1:18734) with synthetic 121-message session:
  • default render showed 50 transcript messages, first raw message index 70, latest raw index 119, and Load earlier messages (70 hidden)
  • clicking the chip expanded to 100 transcript messages, first raw index 20, latest raw index 119, and scroll restoration matched the height delta
  • appending a live assistant message kept a 100-message window, moved the first raw index to 21, and rendered the live assistant turn at raw index 120
  • screenshot stored at docs/pr-media/734/message-window-top.png; raw GitHub URL verified 200 46119
• Full suite attempted twice locally:
  • env -u HERMES_CONFIG_PATH /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/ -q initially failed after 6m32s with unrelated harness/state failures plus source-string drift that this PR then fixed.
  • env -u HERMES_CONFIG_PATH HERMES_HOME=/tmp/hermes-webui-kanban/t_c2343b16/full-test-home HERMES_BASE_HOME=/tmp/hermes-webui-kanban/t_c2343b16/full-test-home /home/michael/.hermes/hermes-agent/venv/bin/python -m pytest tests/ -q ran after fixes and failed after 6m28s with test-server/HERMES_HOME mismatch and connection-refused failures in login/media/session harness tests; targeted affected suites above are green.

Risks / Follow-ups

• This is intentionally not full virtual scrolling: older in-memory messages stay in S.messages, and the user opts into rendering earlier pages with the chip.
• Existing server pagination still uses the previous scroll-top loader; this PR preserves it but prioritizes the explicit chip for already-loaded hidden DOM messages.
• A future larger pass can expose the window size as a preference if maintainers want it in settings.

Model Used

• OpenAI Codex / GPT-5.5 via Hermes Agent.
• Tool use included git/gh CLI, pytest, Node syntax checks, browser QA, and screenshot capture.

[nesquena-hermes]

Closed by the v0.51.1 release in PR #1681 (merged at e23ba59). Massive thanks @Michaelyklam — this is now 19 merged PRs across the v0.50.292–v0.51.1 release window, an extraordinary contribution rate. Each PR was per-claim-vs-diff verified against your description and every security-relevant code path checked under independent review (Opus advisor, 6/6 questions clean). Your closes #N references are all accurate, your Thinking Path / What Changed / Why It Matters body template is consistently helpful, and your test coverage is solid behavioral scope (not source-string scaffolding) on every PR.

Live on production: https://github.com/nesquena/hermes-webui/releases/tag/v0.51.1

🚀