fix(telegram): stop typing before command delivery

[vinsew]

Summary

• stop the typing refresh before every final delivery, including slash-command replies
• cover a command send that stalls indefinitely after delivery starts
• keep session ownership and queued follow-up behavior intact

Root cause

The existing pre-delivery typing lifecycle fix excluded slash commands because their replies are normally instant. That assumption fails when Telegram accepts a command reply but the HTTP acknowledgement stalls on a half-closed connection.

In that state:

1. the user can already see the command reply;
2. the command coroutine remains blocked in delivery;
3. the typing refresh task is only stopped at the coroutine tail;
4. Telegram keeps receiving sendChatAction, so the bot appears to type indefinitely.

PR #25210 was closed as superseded by the post-delivery cleanup in #37556, but post-delivery cleanup cannot run while the delivery acknowledgement itself is stuck. This change closes that remaining boundary by stopping typing before the send begins.

Behavior

The change does not add a delivery timeout or alter retries. It does not release the active-session guard early. Messages arriving during delivery remain queued and drain after the current turn.

Two scheduler-sensitive tests now wait for their observable reply with a short bounded loop instead of assuming a fixed number of event-loop ticks; the extra ticks come from the pre-delivery cleanup itself.

Verification

• 139 passed across topic sessions, split-brain command handling, Telegram formatting, and voice notify coverage
• 242 passed, 2 skipped for gateway tests selected by typing or session_reset or slash_command, with an isolated HERMES_HOME
• Ruff passed on changed files
• git diff --check passed

Regression test

test_command_stops_typing_before_a_stuck_delivery holds a /new delivery open forever and asserts that the typing task has already stopped when delivery starts. It fails on current main and passes with this patch.

[teknium1]

Thanks for isolating the normal final-delivery boundary and adding a stalled-send regression.

Problems

• The change does not cover the active-session /new and /reset route. gateway/platforms/base.py:5458 sends that command reply at :5504 before cancelling the existing processing task at :5518; that existing task owns the typing refresh created by _process_message_background (:5796). A stalled command send can therefore still leave typing active.
• The new /new test calls _process_message_background directly, so it bypasses handle_message's active-session dispatch at gateway/platforms/base.py:5586-5601 and cannot cover that route.

Suggested changes

• Stop or pause the existing typing refresh before the active-command sends at gateway/platforms/base.py:5504 and :5615, preserving the documented send-before-cancel ordering.
• Add a handle_message regression with an active turn, a blocked /new send, and an assertion that typing is already stopped.

Automated hermes-sweeper review.

[teknium1]

This only covers the ordinary _process_message_background path. An active-session /new uses _dispatch_active_session_command, which sends at current gateway/platforms/base.py:5504 before cancelling the old typing-owning task at :5518; a blocked command send there still leaves typing active. Please cover that dispatch path too.

[vinsew]

Addressed the active-session /new//reset review gap and rebased the PR onto current main.

The adapter now tracks each session's typing refresh task separately from its processing task. _dispatch_active_session_command() stops that refresh before sending the command acknowledgement, while preserving the required response-before-processing-task-cancel ordering from #18912. Cleanup is owner-aware so an older task cannot remove a newer session's typing task, and shutdown clears remaining tracked refresh tasks.

The new regression test enters through handle_message() with an actually active session and asserts that the old typing refresh is already stopped when the /new acknowledgement send begins.

Validation on the rebased branch:

• 50 passed across command bypass, topic sessions, typing timeout, and gateway shutdown tests
• Ruff passed for the changed production and test files
• git diff --check passed

[vinsew]

Rebased both typing-lifecycle commits onto current and force-pushed the refreshed branch. : 35 tests passed, 0 failures.

[vinsew]

Maintenance refresh: rebased both typing-lifecycle commits onto current upstream main (b359db7) and force-pushed head 7e1264fb47. Verification: tests/gateway/test_command_bypass_active_session.py — 35 passed; Ruff and git diff --check passed.

[vinsew]

Maintenance refresh: rebased onto current upstream main (40643cb) and force-pushed head 2f796ccc58. Verification: 35 active-session command tests; Ruff and git diff --check passed.

[vinsew]

Rebased onto official main f3cbb26 (2026-08-28). No conflicts; gateway typing tests 58 passed / 5 skipped; ruff clean.

[vinsew]

Rebased onto official main (2026-08-28, post-v0.20.6). Scope change: dropped the second commit (per-session typing-task registry for reset replies) — official main has since reworked the command/reset typing path (slash_commands mixin now clears typing when no active agent owns the session), and the carried registry implementation deadlocked test_active_session_text_merge on the current tree (300s CI timeout, deterministic). The remaining two commits carry the still-unabsorbed value: stop typing before slash-command delivery without stalling ephemeral delete scheduling. active_session_text_merge 7 passed; typing/ephemeral/command-bypass focused suite 119 passed / 5 skipped; ruff clean.