Hermes Agent Update v0.8 is HUGE! (Intelligence Release)

Hermes Agent v0.8 ships async task flags, live model failover, and self-patched tool-calling failures.

## Headline Features The release leans on three workflow changes. First, a notify on complete flag on any long-running process lets the agent fire-and-forget instead of polling — it picks up the result when the job lands. Second, live model switching is now exposed across CLI, Telegram, and Discord, and the router stays aggregator-aware: it prefers OpenRouter or Nous Portal, then falls back automatically. The creator specifically calls out that when Opus 4.6 limits are hit, Hermes falls back to Minimax without manual intervention, and notes this is missing in OpenClaw. Third, Hermes ran automated behavioral-match benchmarking against GPT and Codex, surfaced five failure patterns, and patched them inline — including execution-discipline guidance, system-prompt tuning, and thinking_only prefill for structured reasoning.

## Providers and Routing Google AI Studio is now a native provider with automatic context-window detection via the models.dev registry — no hardcoded limits. Inline model pickers on Telegram and Discord replace typed commands. Minimax TTS with the speech-2.8 model is wired in, and Xiaomi Mimo v2 Pro is free on Nous Portal for auxiliary compression and summarization. xAI Grok now gets prompt caching; Ollama Cloud gets auth and alias tab-completion; stale OAuth tokens no longer block OpenRouter auto-detection.

## Tools and Browser execute_code now runs on remote backends (Docker, SSH, Modal) instead of local only. The browser provider migrated from browserbase to browseruse, and Firecrawl is added as a cloud option. JS evaluation via the browser console is live, and terminal results now explain common CLI exit codes instead of returning raw non-zero values.

## Discord, Memory, and Security Discord now supports ignored/no-thread channel config, native /skill command registration, and clickable approval buttons replacing typed /approve. A new supermemory provider ships with per-user identity scoping. Shared thread sessions are now default for multi-user Telegram/Discord gateways — no manual setup. MCP gets OAuth 2.1 PKCE auth, and MCP extension packages are auto-scanned against the OSV vulnerability database.

## Docs The team closed 40+ doc/code discrepancies, documented 13 previously-undocumented features, and added 3 tutorials to the guide section.

## 6 New Skills LLM Wiki (Karpathy's Obsidian-style markdown knowledge graph) is the standout for research workflows; the other five are incremental.