Persist compression boundary summary for reload UI

[franksong2702]

Summary

This PR resolves #1833.

• Persist compression anchor metadata during manual /session/compress:
  • compression_anchor_summary
  • compression_anchor_visible_idx
  • compression_anchor_message_key
• Persist equivalent compression-anchor metadata during auto-compression in streaming (run_conversation path), so the boundary marker survives session reloads.
• Render compression reference cards from session.compression_anchor_summary when the raw context-compaction message is not present in S.messages.
• Keep behavior unchanged for non-compression flows.

Testing

• /.venv_test/bin/python -m pytest -q tests/test_sprint46.py tests/test_auto_compression_card.py
• /.venv_test/bin/python -m py_compile api/routes.py api/streaming.py api/models.py
• node --check static/ui.js

[nesquena-hermes]

Summary

Reading api/streaming.py:1539-1620, 3082-3093 on the PR branch + the matching _handle_session_compress block in api/routes.py:7505-7728, plus api/models.py:329-413, 575 for the persistence shape, plus static/ui.js:4753-4787 for the render path, the PR correctly closes the gap from #1833: compression-anchor metadata now persists across reloads on both manual /session/compress and the auto-compression path inside _run_agent_streaming.

Code reference

The persistence shape is symmetric across both paths:

# api/models.py:411-413 — METADATA_FIELDS now includes the new field
'compression_anchor_visible_idx', 'compression_anchor_message_key',
'compression_anchor_summary',

And _run_agent_streaming stamps all three immediately after the compression migration:

# api/streaming.py:3082-3093
if _compressed:
    visible_after = _visible_messages_for_compression_anchor(s.messages)
    s.compression_anchor_visible_idx = (
        max(0, len(visible_after) - 1) if visible_after else None
    )
    s.compression_anchor_message_key = (
        _compression_anchor_message_key(visible_after[-1]) if visible_after else None
    )
    s.compression_anchor_summary = _compact_summary_text(
        _compression_summary_from_messages(s.messages)
        or _compression_summary_from_messages(s.context_messages)
    )

The render path correctly falls back to the persisted summary when the raw context-compaction message is no longer in S.messages (the typical reload state):

// static/ui.js:4779-4784
const referenceText=referenceMessage
  ? msgContent(referenceMessage)||String(referenceMessage.content||'')
  : sessionCompressionSummary;
const referenceNode=(!compressionState && !!referenceText && (sessionCompressionAnchor!==null || sessionCompressionAnchorKey || sessionCompressionSummary))
  ? ...
  : null;

Diagnosis / recommendation

The change is well-scoped and the contract holds across the three layers (Session save/load, route response, render path). A few notes:

1. Logic duplication between streaming.py and routes.py. Both files now have local copies of _compact_summary_text, _compression_summary_from_messages, _compression_anchor_message_key, and _visible_messages_for_compression_anchor. The streaming-side helpers (api/streaming.py:1539-1620) and the route-side helpers (api/routes.py:7505-7536 etc.) implement nearly the same logic with slightly different semantics — routes.py:_compression_summary_from_messages filters on role == "assistant" and matches "context compaction|compression" in the body, while streaming.py:_compression_summary_from_messages uses _is_context_compression_marker(m) (which has more aliases — see api/streaming.py:1527-1535). Worth consolidating to a single source of truth in a follow-up to avoid drift; the streaming-side variant is the more general one.

2. Truncation limits differ across the two helpers. Streaming-side caps at 320 chars with …; routes-side also caps at 320 but uses a different ellipsis offset (txt[:314]… vs streaming's txt[: limit - 6]…). 6 vs 6 is the same in this case, but the duplication invites future drift.

3. compression_anchor_summary == None path on the render side. static/ui.js:4763-4765 does String(...).trim() and falls back to ''. The reference card render path at :4781-4784 checks !!referenceText, which correctly excludes null/undefined/''. Good.

4. Session.save() persists None for compression_anchor_summary. That's fine because the loader in Session.__init__ accepts None defaults at models.py:331. New sessions won't carry a stale summary on disk.

Test plan

tests/test_sprint46.py:test_session_compress_roundtrip is a strong pin — it walks the manual compress path, asserts the response payload contains the new field, and reloads via get_session() to confirm persistence. The tests/test_auto_compression_card.py:test_context_anchor_reference_uses_session_summary_fallback static check pins the static/ui.js render fallback. Together they cover both halves.

What's missing: a streaming-path test that drives _run_agent_streaming through compression and asserts s.compression_anchor_summary lands on disk (the equivalent for the auto-compression branch at streaming.py:3082-3093). The current test covers the manual path; the auto path is statically covered by the test file's existence but not behaviorally exercised. Worth a follow-up if not in scope here.

Cross-repo note

This is webui-side display-state only. The agent's compressor (agent/context_compressor.py in hermes-agent) emits the summary as the new context-compaction message at compression time; the webui captures it into compression_anchor_summary for reload survivability. No agent-side change required.

LGTM. Closes #1833 cleanly given the manual-path pin.