Hermes integration: sessions stay 'running' forever — cards never return to idle

[al9000-max]

Summary

Hermes Agent hook events reach CodeIsland correctly, but every session card is created and never returns to an idle/done state. The panel shows "N sessions running" indefinitely even when Hermes is idle — display never matches reality.

Environment

• CodeIsland v1.0.31 (bridge installed 2026-07-23)
• Hermes Agent (Nous Research), running as: macOS desktop app (long-lived serve backend process) + gateway daemon
• macOS

Steps to reproduce

1. Install the Hermes integration (hooks injected into ~/.hermes/config.yaml)
2. Approve the hooks (Hermes consent model — note: unapproved hooks are silently dropped; see related hermes agent 通知不起作用 #226, whose empty hook-events.json had this same root cause on the Hermes side)
3. Run any Hermes session that performs tool calls (desktop app or CLI)
4. Observe: a session card appears and shows as active/running
5. Wait minutes/hours — the card never goes idle, even with zero activity

Root cause

CodeIslands session model assumes a lifecycle: SessionStart → activity → SessionEnd/process exit. Claude Code, Codex, etc. fire events from short-lived CLI processes that exit when the task ends, so CodeIsland knows the session is done.

Hermes breaks that assumption in two ways:

1. Long-lived daemon processes. The desktop app runs a persistent serve backend; gateway chats run in a daemon. The process never exits, so cleanupIdleSessions (dead-PID detection) never removes the session. (Related: CodeIsland repeatedly sends SIGTERM to launchd-managed daemon (Hermes gateway), causing restart loop #243 — the SIGTERM side of this mismatch was fixed, but the "stuck running" display remains.)
2. No idle/done event in the hook set. Hermes fires on_session_start, pre_tool_call, post_tool_call, on_session_end (only on /reset or explicit close), subagent_stop. There is no "response complete" or "idle" event, and on_session_end rarely fires in normal use.

So a card created by the first tool call stays in "running" state forever.

Proposed fixes (any would help; ideally a combination)

1. Map post_llm_call → internal idle/done. Hermes already has a post_llm_call hook event that fires after every LLM response completes. Mapping it to a "done" state (like TaskRoundComplete for Cline) would return the card to idle after each turn. Users would need to add a post_llm_call hook entry to their config — document it in the Hermes setup.
2. Idle timeout per source. For sources whose tracked PID is a persistent daemon (Hermes serve/gateway), transition the session to idle after N seconds without events instead of treating it as perpetually active — without SIGTERM-ing the process (see CodeIsland repeatedly sends SIGTERM to launchd-managed daemon (Hermes gateway), causing restart loop #243).
3. on_session_reset already maps to SessionEnd — good; but also treat a Hermes session as ended when the user closes the conversation in the desktop app (if the app-side can detect it).

Expected behavior

A Hermes session card should show activity only while Hermes is actually processing (tool calls, LLM calls), and settle to an idle state when the turn completes — mirroring how Claude Code sessions behave. The notch should not accumulate permanent "running" cards.

Happy to provide bridge payloads / diagnostics if useful.