fix(agent): terminal activity stamp to prevent stuck last_activity_at

[yingliang-zhang]

Problem

The Desktop UI depends on last_activity_at in the session DB to detect session freshness and trigger message list reloads. When this timestamp gets stuck, the UI stops showing new messages — the user sees "output swallowed" even though the messages are correctly persisted in the DB.

Root Cause

The turn finally block called _reset_activity_labels_after_turn() which cleared last_activity_description and last_activity_provenance but never force-persisted last_activity_at.

Mid-turn heartbeats via _touch_activity() are rate-limited to one write per 60s (SESSION_ACTIVITY_HEARTBEAT_MIN_INTERVAL_SECONDS), and each write has a 0.5s patience budget (_ACTIVITY_WRITE_PATIENCE_S) to avoid blocking the response-critical path. Under DB contention (large state.db, many read connections), the 0.5s budget is exhausted and the write is silently dropped (fail-open by design).

When ALL intermediate heartbeats in a turn are either rate-limited or dropped, and the turn finally doesn't force-persist, last_activity_at stays at the last successful heartbeat — potentially many messages behind.

Fix

Replace _reset_activity_labels_after_turn() in the turn finally block with the new _finalize_activity_after_turn():

SessionDB.finalize_session_activity() (new, hermes_state.py)

• One atomic UPDATE that stamps last_activity_at AND clears labels in the same write
• Uses elevated write patience (2.0s vs 0.5s) — the turn is over, no response-critical path to protect
• Monotonic guard (WHERE last_activity_at < ?) prevents backwards movement

AIAgent._finalize_activity_after_turn() (new, run_agent.py)

• Does NOT call _touch_activity — avoids the kanban worker bridge (inject_new_comments_from_env) which could trigger a spurious continuation turn on a finished turn
• Does NOT bump in-memory _last_activity_ts — preserves Cached agent reuse silently disables the inactivity timeout, producing a Still working iteration 0/60 (cached) loop after user interrupt #15654 interrupt-recursive watchdog continuity
• Clears in-memory labels after the DB write

_reset_activity_labels_after_turn() — retained without a production caller

Once this PR lands, _reset_activity_labels_after_turn() has no production callers (compression's label handling goes through _touch_activity(..., force_persist=True) and clear_session_activity_labels, not through this function). It is retained for test-only usage and as a utility for future callers that need label-only clearing without timestamp stamping.

Tests

New test file tests/run_agent/test_finalize_activity_after_turn.py (8 tests):

• DB stamp fires even when heartbeat persist window is fresh (rate-limiter bypass)
• In-memory _last_activity_ts is NOT bumped (Cached agent reuse silently disables the inactivity timeout, producing a Still working iteration 0/60 (cached) loop after user interrupt #15654 watchdog continuity)
• In-memory labels are cleared
• DB errors are swallowed (fail-open)
• No-op without session_id / session_db
• Uses existing _last_activity_ts value (not a fresh time.time())
• Existing _reset_activity_labels_after_turn still works unchanged

All 22 tests pass (8 new + 14 existing in test_session_activity_persist.py).

[spfcraze]

This was generated by AI during triage.

Summary:
The description says _reset_activity_labels_after_turn() is still used by compression, but compression's label handling goes through _touch_activity(..., force_persist=True) and clear_session_activity_labels — the turn finally replaced here is its only production call site.

Problems:

• agent/conversation_compression.py:1456 stamps the compression terminal label via _touch_activity(..., force_persist=True); agent/conversation_compression.py:3430 clears the archived parent's labels via clear_session_activity_labels — neither path calls _reset_activity_labels_after_turn().
• At main, _reset_activity_labels_after_turn() has one production caller, the turn finally in run_agent.py that this PR replaces with _finalize_activity_after_turn(); the only other reference is a test file.

Solution:
As a wording that matches the code: _reset_activity_labels_after_turn() is retained without a production caller once this lands, and compression's label paths are _touch_activity(..., force_persist=True) and clear_session_activity_labels.

---

Checked against de26441 — the tip of fix/turn-end-activity-stamp when this was written — and 427584b, main at the same moment.

[yingliang-zhang]

Thanks @spfcraze for the accurate triage — good catch. I've updated the PR description to reflect the actual code structure: _reset_activity_labels_after_turn() is retained without a production caller once this PR lands. Compression's label paths go through _touch_activity(..., force_persist=True) and clear_session_activity_labels, not through this function. The function is kept for test-only usage and as a utility for future callers that need label-only clearing without timestamp stamping.

[yingliang-zhang]

Rebased onto latest main (bccb7972df).