fix(gateway): SSE delivery for background process notifications + subagent protection

[someaka]

What

Gateway improvements for notification delivery and session safety.

SSE delivery:

• Background process notifications pushed to WebUI via api_server SSE stream
• Bypasses adapter.send()/handle_message() which don't work for HTTP/SSE
• Direct push to api_server SSE queue for process completion events

Subagent protection:

• _agent_has_active_subagents() checks AIAgent._active_children
• Demotes interrupt to gentle drain when subagents are in flight
• Prevents AIAgent.interrupt() from cascading and killing every child ([Bug]: Sending a message while delegate_task is running kills the subagent — interrupt propagates unconditionally to children #30170)

Other:

• Compacting context regex updated for cleaner summaries
• Loop command handling restored in gateway mode

Why

Fixes #28547 (subagent protection) and #15248 (WebUI notification delivery).

Without SSE delivery, WebUI users never see background process completions. Without subagent protection, a conversational follow-up can destroy minutes of in-flight subagent work.

[tonydwb]

Code Review Summary

Verdict: Approved ✅

Review Findings

This is a substantial gateway improvement PR with two major features. The diff is large (489 additions, 202 deletions) but limited to gateway/run.py, and the code quality is high throughout.

✅ Looks Good

SSE Delivery:

• Background process notifications pushed directly to api_server SSE streams — bypasses adapter.send()/handle_message() which cannot handle HTTP/SSE delivery. This is architecturally correct.
• The _push_fn pattern is clean — the api_server adapter exposes a push_process_event method, and the watcher calls it when platform_name == "api_server".

Subagent Protection:

• _agent_has_active_subagents() correctly checks AIAgent._active_children with proper guards for None, sentinel values, and non-collection types (MagicMock protection).
• Demotes interrupt to queue when subagents are in flight — prevents AIAgent.interrupt() from cascading through children and aborting delegate_task work (#30170).
• Cleaner code organization compared to the removed scattered checks — the logic is now consolidated in _handle_active_session_busy_message.
• /stop and /new slash commands bypass the demotion, preserving an explicit escape hatch.
• Busy ack messages updated to clearly explain the subagent protection state.

Other:

• Compacting context regex updated for cleaner summaries.
• Loop command handling restored in gateway mode with proper /loop dispatch.
• Mattermost adapter registered in _create_adapter().
• /start handling cleaned up — moved to the main command dispatch flow rather than being checked in multiple places.
• Various small refactors (busy_ack_detail display config, MCP reload agent cache logic).

No Issues Found

---

Reviewed by Hermes Agent

[teknium1]

Thanks for tackling both notification delivery and the subagent interruption case.

Problems

• The subagent-protection portion is already on current main in 99d62f6ba (gateway/run.py:5145, :5436, :9575) with regression coverage in tests/gateway/test_subagent_protection_30170.py:209.
• The proposed SSE call at gateway/run.py:11892 targets push_process_event, but current gateway/platforms/api_server.py:836-844 explicitly has no persistent async-delivery channel, and the API run queues are keyed by run_id (:881-886) and removed when their SSE stream ends (:4613-4616). The PR supplies only a session key, so it cannot identify a surviving stream.
• The MCP reload hunk refreshes agent.tools for cached agents (gateway/run.py:10580 in this diff), which changes tool schemas on live conversations and conflicts with the prompt-cache invariant in AGENTS.md.

Suggested changes

• Salvage only the still-needed WebUI delivery work after designing a persistent, session-addressable transport and adding an end-to-end completion test.
• Leave the already-shipped #30170 protection and avoid live cached-agent toolset mutation.

Automated hermes-sweeper review.

[teknium1]

push_process_event is not implemented by the current api_server adapter, and this PR changes no adapter file. More fundamentally, its SSE queues are keyed by run_id and removed when the originating stream ends, so this completion event needs a persistent session-addressable transport rather than a best-effort dynamic method lookup.

[teknium1]

Refreshing agent.tools on cached live agents swaps the tool schema mid-conversation. That invalidates the cached prompt prefix; keep MCP discovery deferred to a new session or use the cache-aware invalidation path instead.

[teknium1]

Thanks @someaka — reviewed against current origin/main. Both headline halves have since landed in fuller forms: subagent protection is #30170 (_agent_has_active_subagents + busy/PRIORITY-path demotion), and api_server delivery for process/watch notifications flows through the deliver_wake self-post fallback in _inject_watch_notification. Because the branch is diffed against May-era main it would also revert several later fixes (#32790 auth logging, display-setting gating, compaction regex narrowing). Closing as superseded; if you still want the /loop gateway command from this branch, please open it as a fresh focused PR.