fix: show auto-compression elapsed time

[dso2ng]

Thinking Path

• Auto-compression during long WebUI runs appears stuck and hides compressed-session handoff #2477 identifies that long WebUI auto-compression can look frozen between the compressing SSE event and the later compressed event.
• The issue's maintainer guidance splits this into independently mergeable slices; this PR implements Slice A only: a persistent running indicator with elapsed time.
• The existing automatic-compression card already provides the right calm UI surface, so this keeps that card and adds a small elapsed timer instead of introducing a new backend event contract or sidebar behavior.
• The state layer changed here is browser transient compression UI (window._compressionUi plus the live compression card DOM), not durable transcript history or model context.

What Changed

• Stamp the automatic-compression running state with startedAt when the compressing SSE event arrives.
• Render running automatic-compression cards as:
  • preview/header: Auto-compressing context... · 00:23
  • expanded body: Elapsed: 00:23
• Once the 5-minute Slice A cap is reached, show 5+ min instead of freezing at 05:00.
• Start a lightweight 1s timer while the running card is visible.
• Update the existing card's preview/detail text in place; the timer does not call renderMessages() each tick, so it avoids disrupting live stream DOM.
• Stop the timer when compression completes/clears, when the session no longer has a matching compression state, or after the 5-minute Slice A cap.
• Preserve the existing live-card path by carrying data-compression-started-at / data-compression-message on anchored live compression cards.
• Add regression coverage in tests/test_auto_compression_card.py.
• Add an Unreleased changelog entry.

Why It Matters

Long compression can take minutes. Before this PR, the user saw the same generic running card the entire time, which made the browser feel stalled. The elapsed label gives a visible heartbeat while preserving the existing quiet compression-card design.

This intentionally does not implement #2477 Slice B/C:

• no compressed-session handoff link,
• no provider fallback/warning contract changes,
• no backend SSE payload change,
• no sidebar/session-list behavior change.

Visual Evidence

Verification

Targeted local verification on the updated head:

node --check static/ui.js
node --check static/messages.js
python -m pytest \
  tests/test_auto_compression_card.py \
  tests/test_compression_snapshot_runtime_clear.py \
  tests/test_session_lineage_collapse.py \
  tests/test_streaming_session_sidebar.py \
  tests/test_issue734_message_windowing.py \
  tests/test_issue1617_tps_message_header.py \
  -q

Result:

74 passed

Hygiene:

git diff --check -> clean
non_ascii_added_lines=1

The one intentional non-ASCII added line is the existing public UI separator · in the elapsed preview.

Full local suite was also run on the updated head:

python -m pytest tests/ -q

Result:

5934 passed, 6 skipped, 3 xpassed, 8 subtests passed, 4 failed

The 4 failures are existing base failures in tests/test_issue1527_lmstudio_base_url_classification.py about prefixed model ids. I confirmed the same file fails the same way on clean origin/master@e6be01c4:

python -m pytest tests/test_issue1527_lmstudio_base_url_classification.py -q
# 4 failed, 1 passed

Risks / Follow-ups

• Page reloads during an in-flight compression are not resume-on-reload. The timer state is browser-transient; after a reload the UI waits for any new compressing SSE state rather than reconstructing elapsed time from server history.
• The timer is intentionally presentation-only and capped at 5 minutes per the Slice A guidance; it does not infer provider progress or compression phases.
• If fix: clear fallback streaming warnings #2505 lands first, this may need a small rebase because both PRs touch tests/test_auto_compression_card.py, but the implementation layers are separate.
• Slice B/C from Auto-compression during long WebUI runs appears stuck and hides compressed-session handoff #2477 remain open follow-ups for compressed-session handoff surfacing and fallback transparency.

Model Used

AI-assisted with Hermes Agent using OpenAI Codex provider, model gpt-5.5, with terminal/file/browser tooling for implementation, tests, and visual evidence capture.

[nesquena-hermes]

Summary

Reading static/ui.js:1970-2010, 4905-4935, 5050-5070 on this branch against origin/master, the diff cleanly extends the existing auto-compression card with a 1-second-tick elapsed timer, capped at 5 minutes. It addresses the real complaint that an auto-compress card with no time signal looks frozen when the summarizer LLM call takes longer than expected. The implementation reuses the existing _formatActiveElapsedTimer helper, stamps startedAt on first SSE event, and persists the start time on the DOM so the live card stays in sync after renderMessages(). Solid shape; a few details worth tightening.

Code reference

The start-time stamping inside setCompressionUi:

const nextState={...state};
if(nextState.automatic&&nextState.phase==='running'&&!_compressionElapsedStartedAt(nextState)){
  nextState.startedAt=Date.now()/1000;
}
window._compressionUi=nextState;
if(nextState.sessionId) _setCompressionSessionLock(nextState.sessionId);
if(nextState.automatic&&nextState.phase==='running') _startCompressionElapsedTimer();
else _clearCompressionElapsedTimer();

The DOM-anchored fallback used after renderMessages() strips window._compressionUi:

function _compressionLiveCardState(){
  const node=_compressionLiveCardNode();
  const started=Number(node&&node.getAttribute('data-compression-started-at'));
  if(!node||!S.session||!Number.isFinite(started)||started<=0)return null;
  return {sessionId:S.session.session_id,phase:'running',automatic:true,
          message:node.getAttribute('data-compression-message')||'Auto-compressing context...',
          startedAt:started};
}

The cap at _COMPRESSION_ELAPSED_MAX_SECONDS=5*60 and the _compressionElapsedExpired check stop the timer from ticking forever if the compressed SSE event never arrives.

Diagnosis / Recommendation

Two correctness things and one small UX thing:

1. Page-reload behavior. appendLiveCompressionCard stamps the start time onto the DOM node, but a full page reload rebuilds the DOM from renderMessages() reading server-side history — the live card hasn't been persisted yet (the compression card is rendered transiently from window._compressionUi, not from s.messages). So if a user reloads mid-compression, the elapsed counter resets to wherever the new SSE state arrives. That's probably acceptable, but worth noting in the PR description so reviewers don't expect resume-on-reload.
2. 5-minute cap is the right ceiling but inaccurate framing. Once the counter hits 5:00, the timer clears via _clearCompressionElapsedTimer() but the card still displays 05:00 as the frozen value. If a real compression hangs past 5 minutes, the user sees a non-ticking "Elapsed: 05:00" and may think the elapsed display itself is broken. Consider switching the suffix at the cap from a frozen number to a textual fallback like Elapsed: 5+ min, or kicking off an "auto-compression appears stuck" inline notice. Related to issue Auto-compression during long WebUI runs appears stuck and hides compressed-session handoff #2477, which is already open.
3. Detail vs preview duplication. _autoCompressionDetailText produces "${detail}\nElapsed: ${elapsedLabel}" (two lines), while _autoCompressionPreviewText produces "${detail} · ${elapsedLabel}" (single line). The preview is what shows in the card header; the detail is in the <pre> body. Functionally fine, but slightly redundant when a user expands the card — they see "Auto-compressing context... · 00:42" up top and "Auto-compressing context...\nElapsed: 00:42" below. Either drop the elapsed from the preview (header), or drop the redundant detail line from the body and only show Elapsed: ... there.

Test plan

The three new test cases in tests/test_auto_compression_card.py pin: the SSE handler stamps startedAt before calling setCompressionUi; the cards render _compressionElapsedLabel; and the live-card path persists data-compression-started-at + data-compression-message for DOM-anchored timer refresh. That covers the structural contracts. Worth adding one more: a state with startedAt set 6 minutes in the past should produce the capped label (05:00) and not call _startCompressionElapsedTimer on render.

Manual verification: trigger an auto-compression (paste a large message that exceeds the model's context, send a turn) and confirm the card shows a ticking Elapsed: MM:SS that stops at 5 minutes if the compression never completes.

CHANGELOG entry covers it. Ready to land once #2 (cap framing) is reconciled with the spirit of issue #2477.

[dso2ng]

Thanks for the detailed pass — I pushed d72b3382 to address the concrete items:

• Page reload behavior: updated the PR description to explicitly call out that this is browser-transient state, not resume-on-reload state reconstructed from server history.
• 5-minute cap framing: changed the capped label from a frozen 05:00 to 5+ min, so if compression exceeds the Slice A cap the UI no longer looks like a broken stopped timer.
• Preview/detail duplication: kept the elapsed heartbeat in the card header/preview and changed the expanded body to show only Elapsed: ... for running cards.
• Regression coverage: added tests for the non-frozen capped label and for avoiding duplicate message text in the expanded body.
• Visual evidence: updated the PR body screenshot to show the new detail-body shape plus the 5+ min cap example.

Targeted local verification on the updated head:

node --check static/ui.js
node --check static/messages.js
python -m pytest \
  tests/test_auto_compression_card.py \
  tests/test_compression_snapshot_runtime_clear.py \
  tests/test_session_lineage_collapse.py \
  tests/test_streaming_session_sidebar.py \
  tests/test_issue734_message_windowing.py \
  tests/test_issue1617_tps_message_header.py \
  -q
# 74 passed

git diff --check -> clean
non_ascii_added_lines=1  # intentional public UI separator: ·

Full local suite on the updated head still has only the existing base-failure subset:

python -m pytest tests/ -q
# 5934 passed, 6 skipped, 3 xpassed, 8 subtests passed, 4 failed

The 4 failures are the same tests/test_issue1527_lmstudio_base_url_classification.py prefixed-model-id failures previously reproduced on clean origin/master@e6be01c4.