fix: surface auto-compression handoff

[dso2ng]

Thinking Path

#2477 was split into small slices. Slice A (elapsed timer) has shipped via #2512, and fallback/rate-limit warning transparency has shipped via #2505. The remaining narrow WebUI gap is the compressed-session handoff: during automatic compression the backend can rotate from the original browser session id to a new continuation id before the compressed SSE event is emitted.

Before this PR, that completion event used the rotated s.session_id. The frontend listener still compared d.session_id against the original activeSid, so the active browser stream could drop the completion event and keep showing a stale running compression card.

Refs #2477 (Slice B only).

What Changed

• Keep the compressed SSE event correlated to the origin stream session:
  • session_id / old_session_id now carry the pre-compression origin session id.
  • new_session_id / continuation_session_id carry the compressed continuation id.
• Update the frontend compressed listener to correlate on old_session_id before falling back to session_id.
• Keep the continuation id as transient compression-card metadata and show it in the expanded automatic-compression detail line:
  • Continued in compressed session: <id>
• Add regression coverage for the rotated-session completion path.
• Add an Unreleased changelog entry.

Why It Matters

A long WebUI run should not look stuck right after auto-compression succeeds. The completion event must still reach the active browser stream even if the backend session id has already rotated to the compressed continuation session.

This keeps the completion card visible without inserting fake transcript messages or changing model context.

Scope Boundaries

This PR is Slice B only.

In scope:

• auto-compression compressed SSE payload metadata
• frontend correlation of the completion event
• transient completion-card detail text

Out of scope:

• elapsed timer behavior from fix: show auto-compression elapsed time #2512
• fallback/provider warning transparency from fix: clear fallback streaming warnings #2505
• title grouping / sidebar family grouping
• agent-side compression behavior
• durable transcript/model-context changes

Visual Evidence

Before/after illustration (not added to the code branch):

Verification

Targeted local verification on ea978a198931:

• node --check static/ui.js
• node --check static/messages.js
• $HOME/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_auto_compression_card.py -q
  • 36 passed
• $HOME/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_auto_compression_card.py tests/test_streaming_max_tokens_quota.py tests/test_inflight_stream_reuse.py tests/test_compression_snapshot_runtime_clear.py tests/test_session_lineage_collapse.py tests/test_streaming_session_sidebar.py -q
  • 73 passed
• git diff --check
• public diff hygiene:
  • non_ascii_added_lines=0
  • secret_like_added_lines=0

Full suite on the PR branch:

• $HOME/.hermes/hermes-agent/venv/bin/python -m pytest tests/ -q
  • 5968 passed, 6 skipped, 3 xpassed, 8 subtests passed, 4 failed

The 4 failures are existing upstream baseline failures in tests/test_issue1527_lmstudio_base_url_classification.py; the same file fails on clean origin/master@718a4c76:

• $HOME/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_issue1527_lmstudio_base_url_classification.py -q
  • clean origin/master: 4 failed, 1 passed

Risks / Follow-ups

Low risk. The patch keeps the event shape backward-friendly by adding explicit metadata rather than removing fields. The frontend still ignores events for unrelated active sessions; it simply correlates completion using the origin id after compression rotation.

The continuation detail is transient UI only. It does not push into S.messages, does not alter durable transcript history, and does not change model-facing context.

Model Used

AI-assisted: OpenAI GPT-5.5 via Hermes Agent, with repository tools, local pytest verification, and an independent reviewer subagent.

[nesquena-hermes]

Phase 0 review — narrow Slice B of #2477, contract-clean; ready for batch

Pulled the PR head (ea978a19) and read the three-file diff against origin/master. This is the right shape for the remaining handoff gap — backward-compatible event payload, correlation on the origin id, transient UI only.

The bug shape this closes

Before this PR, the compression-side-effects block in api/streaming.py:4348-4435 would rotate session_id to the agent's new id (_agent_sid), persist the new id under SESSION_AGENT_LOCKS / SESSION_AGENT_CACHE, and then emit the completion event with 'session_id': s.session_id — i.e. the rotated id. The browser's listener (static/messages.js:1812) compared d.session_id !== activeSid and bounced the event for the original sid, leaving the running auto-compression card alive until the next /api/sessions poll healed it.

Server-side payload — clean

if _compression_continuation_session_id is None:
    _compression_continuation_session_id = s.session_id
put('compressed', {
    'session_id': _compression_origin_session_id,
    'old_session_id': _compression_origin_session_id,
    'new_session_id': _compression_continuation_session_id,
    'continuation_session_id': _compression_continuation_session_id,
    'message': 'Context auto-compressed to continue the conversation',
    'usage': _live_usage_snapshot(),
})

Two things I like here:

• session_id is now the origin id (not the rotated one). That keeps the legacy contract intact for any consumer that only knows the original session_id field — they get the correlation they expected, just under different semantics. Subtle but right.
• Both new_session_id and continuation_session_id are populated, so consumers that picked either name as their canonical handoff key continue to work.

The fallback at _compression_continuation_session_id is None: ... = s.session_id handles the "compression detected via compressor state, not via sid rotation" path (api/streaming.py:4412-4416). That's the case where _compressed = True was set but no actual rotation happened — s.session_id is unchanged, so the continuation id matches the origin id, which is correct: there is no handoff to surface, the card just transitions to done.

Client correlation — symmetric

const eventSid=d.old_session_id||d.session_id||activeSid;
if(eventSid!==activeSid) return;
const continuationSid=d.new_session_id||d.continuation_session_id||'';

Preferring old_session_id first is the correct order: with this PR, session_id is already the origin id, but if a future change re-points session_id to the continuation id (e.g. for symmetry with other event types), the listener still correlates correctly via old_session_id.

Transient UI only — verified

static/ui.js:5018-5021:

if(running)return elapsedLabel?`Elapsed: ${elapsedLabel}`:base;
const continuation=String(state&&state.continuationSessionId||'').trim();
const handoff=continuation?`Continued in compressed session: ${continuation}`:'';

continuationSessionId lives in the state object handed to setCompressionUi(state) — that's transient compression-card metadata, not a transcript message. window._compressionUi=null immediately follows (static/messages.js:1837) once appendLiveCompressionCard anchors the card to the turn, so renderMessages() won't duplicate it. Scope boundaries claimed in the PR body match what the code actually does.

Verification matches

Full suite on the PR branch reports the 4 known tests/test_issue1527_lmstudio_base_url_classification.py upstream baseline failures, which I confirm also fail on clean origin/master. Not introduced by this PR.

Lock migration interaction — confirmed safe

The migration block at api/streaming.py:4400-4410 runs before the put('compressed', ...) call at :4430, so by the time the event is emitted, the new sid is already aliased to the same _agent_lock. No window where a subsequent caller would race against a stale lock entry for the origin id.

LGTM for batch as Slice B of #2477.