mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-25 17:18:11 +00:00
082bd17122
Mid-turn progress lives only in process memory — the agent flushes to SQLite at turn end — so an app/backend/machine death mid-turn lost the turn entirely: reopening the session showed the recovered partial, but the work never finished and the prompt itself survived nowhere durable. Turns now write a durable marker (bounded per-profile sidecar) when they start running and clear it when they conclude; success, handled error, and interrupt all clear it, so a surviving marker is positive proof of a process death. session.resume reads the marker: a fresh interruption (desktop.auto_continue.freshness_minutes, default 15) is re-submitted automatically as a continuation turn carrying the original prompt in an interruption note, streams live to the client that just resumed, and renders as a "resumed interrupted turn" event row. Stale markers are cleared and the recovered partial speaks for itself; a turn that keeps crashing stops auto-continuing after max_attempts (default 2) — the same freshness + crash-loop-breaker posture as the messaging gateway's restart auto-resume. |
||
|---|---|---|
| .. | ||
| bootstrap-installer | ||
| desktop | ||
| shared | ||