Follow-up note: this is a pretty painful user-facing gateway bug, not just cleanup.

In Telegram DM topics, a restart could make the gateway synthesize a fresh auto-resume turn in topics where the useful work was already over or where the previous recovery attempt had already failed. The visible symptom is noisy/phantom assistant activity after restart — for example repeated rate-limit messages or the agent continuing in an old topic without new user input.

This PR now hardens the recovery gate in a few ways:

• Use durable SQLite transcript timestamps for SessionStore.load_transcript() so DB-backed histories are not treated as legacy timestamp-less/fresh forever.
• Clear resume_pending instead of scheduling auto-resume when the transcript already ends in a final assistant answer.
• Clear resume_pending when the tail is only a gateway-generated recovery note from a previous empty auto-resume attempt, preventing restart loops.
• Treat timed-out / undelivered clarify tool results as a user-wait boundary, not unfinished tool work to auto-process on startup or on the next unrelated user message.

Local verification:

• python -m pytest -o addopts='' tests/gateway/test_restart_resume_pending.py -q
• python -m pytest -o addopts='' tests/gateway/test_clean_shutdown_marker.py tests/gateway/test_restart_resume_pending.py tests/gateway/test_telegram_topic_mode.py -q
• python -m py_compile gateway/run.py gateway/session.py hermes_state.py tests/gateway/test_restart_resume_pending.py
• ruff check gateway/run.py gateway/session.py hermes_state.py tests/gateway/test_restart_resume_pending.py
• git diff --check

Follow-up high-priority fix for Telegram DM topic status spam:

• Root cause: queued/interrupt follow-up recursion kept the outer _run_agent per-turn background helpers alive. Each nested run started its own long-running "Still working..." notifier, so several timers could send to the same topic seconds apart with different elapsed clocks.
• Fix: add per-session notifier ownership tokens and cancel the outer run's auxiliary tasks before recursing into a queued follow-up.
• Regression: tests/gateway/test_long_running_notifications.py verifies stale outer notifiers do not send after a queued follow-up starts.

Local verification:

• python -m pytest -q -o 'addopts=' tests/gateway/test_long_running_notifications.py tests/gateway/test_run_cleanup_progress.py tests/gateway/test_queue_consumption.py tests/gateway/test_restart_resume_pending.py
• python -m py_compile gateway/run.py tests/gateway/test_long_running_notifications.py
• git diff --check
• python -m ruff check gateway/run.py tests/gateway/test_long_running_notifications.py

Follow-up fix pushed: eedb4af fix(gateway): suppress killed process completion turns.

Root cause from the new repro was not startup resume_pending: sent the requested final after killing the process, then the process watcher still injected [IMPORTANT: Background process ... completed] for the killed notify_on_complete process, creating a fresh internal turn and a second visible assistant reply.

Fix:

• Treat process(action=kill) / already-exited kill as completion consumption, same as wait/poll/log.
• For notify_on_complete watcher completions, skip all delivery when the process completion is already consumed, including the plain text notification fallback.
• Added regressions in tests/tools/test_notify_on_complete.py and tests/gateway/test_background_process_notifications.py for the killed notify process not re-entering the session.

Verification run locally:

• uv run --extra dev pytest -o addopts='' tests/tools/test_notify_on_complete.py tests/gateway/test_background_process_notifications.py tests/gateway/test_internal_event_bypass_pairing.py tests/gateway/test_duplicate_reply_suppression.py tests/gateway/test_long_running_notifications.py -q -> 87 passed
• uv run --extra dev --with ptyprocess pytest -o addopts='' tests/tools/test_process_registry.py -q -> 58 passed
• python3 -m pytest -o addopts='' tests/gateway/test_restart_resume_pending.py tests/gateway/test_telegram_topic_mode.py -q -> 116 passed
• uv run --extra dev ruff check gateway/run.py tools/process_registry.py tests/tools/test_notify_on_complete.py tests/gateway/test_background_process_notifications.py -> passed
• python3 -m py_compile ... and git diff --check -> passed

Follow-up fix pushed in 3939180: fix(gateway): drop stale watch-pattern continuations.

New root cause from session 20260519_180814_588b4293: this was not another startup resume_pending loop. At 2026-05-21 00:02 the completed turn drained queued process watch_pattern matches from the Gold Apple login probe, formatted them as gateway process notifications, and re-injected them through adapter.handle_message() into the same Telegram DM Topic. Because the original final answer had already been delivered, that synthetic internal message created a fresh agent turn in the finished topic.

Fix:

• propagate per-session run_generation through gateway.session_context into terminal background process sessions and watch events;
• drop watch notifications whose generation is stale for their session_key;
• drop same-turn watch notifications during the completing run instead of re-entering the session;
• keep existing cross-session/current watch notification behavior intact.

Regression coverage added:

• same-turn watch_match does not call adapter.handle_message();
• stale-generation watch_match is dropped;
• watch matches carry run_generation from the process session;
• session context exposes HERMES_SESSION_RUN_GENERATION.

Verification:

• HERMES_HOME=/tmp/tmp.OS8JucTGYC python3 -m pytest -o addopts='' tests/tools/test_watch_patterns.py tests/gateway/test_background_process_notifications.py tests/gateway/test_duplicate_reply_suppression.py tests/gateway/test_long_running_notifications.py tests/gateway/test_internal_event_bypass_pairing.py tests/tools/test_notify_on_complete.py tests/gateway/test_restart_resume_pending.py tests/gateway/test_telegram_topic_mode.py tests/gateway/test_session_env.py tests/gateway/test_session_hygiene.py tests/gateway/test_slash_access_dispatch.py tests/gateway/test_approve_deny_commands.py tests/gateway/test_reload_skills_command.py tests/gateway/test_running_agent_session_toggles.py tests/gateway/test_unknown_command.py tests/gateway/test_steer_command.py tests/gateway/test_status_command.py -q -> 348 passed, 1 warning
• uv run --extra dev pytest -o addopts='' tests/gateway/test_background_process_notifications.py::test_same_turn_watch_match_does_not_reenter_session tests/gateway/test_background_process_notifications.py::test_stale_watch_match_generation_is_dropped tests/tools/test_watch_patterns.py::TestCheckWatchPatterns::test_match_carries_run_generation tests/gateway/test_session_env.py::test_set_session_env_sets_contextvars -q -> 4 passed
• uv run --extra dev ruff check ... -> passed
• python3 -m py_compile ... -> passed
• git diff --check -> passed

Superseded by the narrower restart-resume recovery PR: #30030. This older branch accumulated unrelated gateway/runtime/test changes, so #30030 keeps the review scope to restart resume freshness, Telegram reply anchors, and failed-turn goal continuation gating.