Feature Request: Session Recovery on Temporary Provider Outage

Problem or Use Case*

When the LLM provider goes down temporarily (restart / update / network hiccup for 2–3 minutes), the entire Hermes session is destroyed and cannot be recovered — even after the provider comes back online.

In a self-hosted setup (e.g. Manifest on a LAN server), this happens regularly. Hermes performs 3 retries in ~12 seconds, but if the provider takes 2–3 minutes to recover, the session is dead. Over 2 days I've had 17 such API errors, losing multiple sessions with 200–350 messages (~240k tokens each).

Error: HTTP 404: Cannot POST /v1/chat/completions (provider briefly unavailable, but dashboard is still running).

This affects anyone running Hermes with:

• A self-hosted provider (Manifest, Ollama, LiteLLM, vLLM)
• An unstable LAN server
• Any provider with occasional brief outages

Proposed Solution*

Four suggestions that complement each other:

1. Session Pause instead of Kill When all retries are exhausted: put the session in a "paused" state instead of destroying it. The user sees: "Provider unreachable, session paused — try again later."

2. Longer Retry Window Current: 3 retries at ~2s / ~4s / ~5s = ~12s total. Better: Exponential backoff up to 120s (e.g. 5s / 15s / 30s / 60s / 120s). This covers brief provider restarts.

3. Automatic Session Recovery After a successful reconnect: check if the session can be resumed (repair role alternation) and retry the last failed turn.

4. Transparent Provider Failover (optional) When a fallback provider is configured: auto-switch while the primary is down, auto-switch back when it recovers. The session keeps running — the user only notices a slight delay.

Alternatives Considered

• Manual Session Repair: Currently possible via sqlite3 state.db "UPDATE sessions SET model='...'" — but this only fixes model errors, not broken role alternation
• Hard reset (/new): Currently the only reliable option — losing all 240k tokens of context
• Do nothing: 17 errors in 2 days, lost work, frustration

Feature Type*

Scope: Small / Medium — this is a session lifecycle improvement in the agent loop (comp/agent), no new infrastructure needed. The main changes: retry logic in agent/conversation_loop.py, session state handling, and role-alternation repair on reconnect.

Additional Context

Environment:

• Hermes Agent: v0.14.0 (2026.5.16) — self-hosted in Proxmox LXC
• Provider: Manifest (self-hosted on Mac mini, http://192.168.1.245:2099/v1)
• Fallback: Ollama on same Mac mini (http://192.168.1.245:11434)
• Frontend: Hermes WebUI (browser-based)
• Session count in DB: ~650 sessions
• Typical session: 200–350 messages, ~240k tokens

Relevant existing issues:

• #5694 (clearer recovery UX when model failover is exhausted)
• #32411 (per-task fallback_providers config for transparent failover)
• #18452 (ACP adapter does not pass fallback_providers to AIAgent)

Hermes 0.14.0