Hermes Agent v0.8.0 — The Intelligence Release
Overview
v0.8.0 — The Intelligence Release. Released April 8, 2026. Just 5 days after v0.7.0: 209 merged PRs · 82 resolved issues · hundreds of files changed · dozens of community contributors.
If v0.7.0 was “the resilience release” (swappable memory providers, Camofox anti-detection browser, deep gateway hardening), v0.8.0 is about one thing — getting smarter. Hermes learned to diagnose its own failure modes and patch itself. It learned to notify you when background tasks finish. It learned to track actual tool activity rather than wall-clock time when deciding whether to time out. These aren’t feature bullets. They’re a systematic lift in agent intelligence.
Major Features
1. Background Task Auto-Notifications (notify_on_complete) — No More Polling
Background tasks now automatically notify the agent when they finish. Start a long-running process — model training, test suites, deployments, builds — and the agent picks up the results on completion. No ps, no tail -f, no polling loops. The agent just tells you when it’s done.
# Start a background training job
hermes run "python train.py --epochs 100" --background
# Keep working on other things with the agent
# When training finishes, the agent gets notified and reports results(#5779)
2. Free Xiaomi MiMo v2 Pro on Nous Portal
Nous Portal’s free tier now includes Xiaomi MiMo v2 Pro for auxiliary tasks — compression, vision, summarization. Model selection UI shows pricing alongside each model, and free-tier gating is transparent. If you’ve been holding off on Hermes because your wallet was tired, this week the free tier got meaningfully better.
3. Live Model Switching (/model Command) — Switch Models Mid-Session
Switch models and providers mid-session from CLI, Telegram, Discord, Slack, or any gateway platform. The resolver is aggregator-aware — if you’re on OpenRouter or Nous Portal, it keeps you on the aggregator when the model is available there, only falling through to direct providers when needed. Telegram and Discord get interactive inline-button pickers — tap to select, no typing required.
# Switch anytime during a session
/model claude-sonnet-4-20250514
# Or open the interactive picker
/model4. Self-Optimized GPT/Codex Tool-Use Guidance — The Agent Patched Itself
This is the most technically fascinating PR in v0.8.0. The team built an automated behavioral benchmarking pipeline that systematically tested GPT and Codex models, identified five specific tool-calling failure modes, generated targeted guidance strings to fix them, measured the improvement, and shipped the result as part of the system prompt. The agent diagnosed itself and patched itself — humans sat in the loop as reviewers, not prompt engineers.
Paired with this are execution discipline guidance in system prompts and thinking-only prefill continuation for structured reasoning, all dramatically improving reliability when running Hermes on OpenAI models.
5. Google AI Studio (Gemini) Native Provider
Direct access to Gemini models through Google’s AI Studio API. Integrated with the models.dev registry for automatic real-time context length detection — no more manually specifying token limits.
(#5577)
6. Inactivity-Based Smart Timeouts — Working Agents Never Get Killed
Gateway and cron timeouts now track actual tool activity instead of wall-clock time. Long-running tasks that are actively producing output (running commands, reading/writing files) will never be terminated. Only truly idle agents time out.
7. Approval Buttons on Slack & Telegram — Say Goodbye to Typing /approve
Dangerous command approvals now happen through native platform buttons. Slack preserves thread context, Telegram shows emoji reactions for approval status. Much faster than typing /approve every time.
8. MCP OAuth 2.1 PKCE + OSV Malware Scanning
Full OAuth 2.1 PKCE standards-compliant client for MCP server authentication. Plus, MCP extension packages are automatically scanned against the OSV vulnerability database at install time — catching malicious code before it runs.
9. Centralized Logging & Config Validation
Structured logging to ~/.hermes/logs/ (agent.log + errors.log), with the new hermes logs command for tailing and filtering. Config structure validation catches malformed YAML at startup with actionable error messages — no more “mystery crashes.”
# View recent logs
hermes logs
# Errors only
hermes logs --level WARNING10. Plugin System Expansion
Plugins can now:
- Register standalone CLI subcommands without touching core code
- Receive request-scoped API hooks with correlation IDs
- Prompt for required environment variables during install
- Hook into session lifecycle events (finalize / reset)
11. Matrix Promoted to Tier 1
Matrix gets reactions, read receipts, rich formatting, and room management — reaching feature parity with Telegram and Discord as a first-class platform.
12. Security Hardening Pass
SSRF protections, timing attack mitigations, tar traversal prevention, credential leakage guards, cross-session isolation — a systematic security audit across the entire codebase.
Core Agent & Architecture
Provider & Model Support
- Native Google AI Studio (Gemini) provider with models.dev auto-detect for context length
/modelcommand — complete provider+model system overhaul with live switching across CLI and all gateways- Interactive model picker for Telegram and Discord with inline buttons
- Nous Portal free-tier model gating with pricing display in model selection
- xAI (Grok) prompt caching via
x-grok-conv-idheader - MiniMax TTS provider (speech-2.8)
- Non-agentic model warning — alerts when loading Hermes LLMs not designed for tool use
- Ollama Cloud auth, model switch persistence, alias tab completion
- Model pricing display for OpenRouter and Nous Portal
- Auxiliary client payment fallback — retry with next provider on 402
- Auxiliary client resolves named custom providers and ‘main’ alias
- Provider credential reset windows honored in pooled failover
- OAuth token sync between credential pool and credentials file
- Stale OAuth credentials no longer block OpenRouter auto-detect
- Codex OAuth credential pool disconnect + expired token import fixes
- Vision auto-detection tries main provider first
- MiniMax context lengths, thinking guard, aux model, config base_url corrections
- Z.AI endpoint auto-detect via probe and cache
- Community provider/model resolution fixes — salvaged 4 community PRs
Agent Loop & Conversation
- Self-optimized GPT/Codex tool-use guidance — automated behavioral benchmarking identified and patched 5 failure modes
- GPT/Codex execution discipline guidance in system prompts
- Thinking-only prefill continuation for structured reasoning
- Accept reasoning-only responses without retries — set content to “(empty)” instead of infinite retry
- Jittered retry backoff — exponential backoff with jitter for API retries
- Smart thinking block signature management — preserve Anthropic thinking signatures across turns
- Coerce tool call arguments to match JSON Schema types — fixes models that send strings instead of numbers/booleans
- Save oversized tool results to file instead of destructive truncation
- Streaming fallback improved after edit failures
- Codex empty-output gaps covered in fallback + normalizer + auxiliary client
- Codex stream output backfill from output_item.done events
- Stream consumer creates new message after tool boundaries
- Filter transcript-only roles from chat-completions payload
- Sanitize tool_calls for all strict APIs (Fireworks, Mistral, etc.)
- Bridge tool-calls in copilot-acp adapter
Memory & Sessions
- Supermemory memory provider — multi-container, search_mode, identity template, env var override
- Shared thread sessions by default — multi-user thread support across gateway platforms
- Subagent sessions linked to parent and hidden from session list
- Profile-scoped memory isolation and clone support
- Thread gateway user_id to memory plugins for per-user scoping
- Honcho plugin drift overhaul + plugin CLI registration system
- mem0 API v2 compatibility, prefetch context fencing, secret redaction
- RetainDB — API routes, write queue, dialectic, agent model fixes
- Hindsight memory plugin overhaul + memory setup wizard fixes
- OpenViking atexit safety net for session commit, tenant-scoping headers
- ByteRover brv query runs synchronously before LLM call
- Clean user message used for all memory provider operations
Messaging Platforms (Gateway)
Gateway Core
- Inactivity-based timeout replaces wall-clock — active tasks never killed
- Approval buttons for Slack & Telegram + Slack thread context preservation
- Live-stream /update output + forward interactive prompts to user
- Infinite timeout support + periodic notifications + actionable error messages
- Duplicate message prevention — gateway dedup + partial stream guard
- Webhook delivery_info persistence + full session id in /status
- Tool preview truncation respects tool_preview_length in all/new progress modes
- Approval session key isolated per turn
- Active-session guard bypass for /approve, /deny, /stop, /new
- Typing indicator paused during approval waits
- Caption check uses exact line-by-line match instead of substring (all platforms)
- MEDIA: tags stripped from streamed gateway messages, extracted from cron delivery
- Profile-aware service units + voice transcription cleanup
- Thread-safe PairingStore with atomic writes
- Cron static method wrappers to prevent self-binding
Platform Improvements
| Platform | Key Changes |
|---|---|
| Telegram | Group topic skill binding, emoji reaction approvals, duplicate message prevention, command name sanitization, per-platform disabled skills |
| Discord | Channel controls (ignored_channels / no_thread_channels), skills as native slash commands, /approve registered as slash commands, removed unnecessary members intent |
| Slack | Thread auto-engagement, mrkdwn in edit_message, thread replies without @mentions |
| Matrix | Tier 1 upgrade: reactions, read receipts, rich formatting, room management, CJK input, E2EE, reconnect |
| Signal | Full MEDIA: tag delivery |
| Mattermost | File attachment support |
| Feishu | Interactive card approval buttons, reconnect + ACL fixes |
| Webhooks | {__raw__} template token, thread_id passthrough for forum topics |
CLI & User Experience
Interactive CLI
- Defer response content until reasoning block completes
- Ghost status-bar lines cleared on terminal resize
- Normalise \r\n and \r line endings in pasted text
- Native Windows image paste support
--yoloand other flags no longer silently dropped when placed before ‘chat’ subcommand- ChatConsole errors, curses scroll, skin-aware banner, git state fixes
Setup & Configuration
- Config structure validation — detect malformed YAML at startup with actionable messages
- Centralized logging to
~/.hermes/logs/(agent.log + errors.log) withhermes logscommand - Doctor diagnostics — sync provider checks, config migration, WAL and mem0 diagnostics
- Timeout debug logging and user-facing diagnostics improved
- Reasoning effort unified to config.yaml only
- Permanent command allowlist loaded on startup
hermes auth removenow clears env-seeded credentials permanently- Bundled skills synced to all profiles during update
hermes updateno longer kills freshly-restarted gateway service- Subprocess.run() timeouts added to all gateway CLI commands
- Docs links added to setup wizard sections
Cron System
- Inactivity-based cron timeout — active tasks run indefinitely
- Pre-run script injection for data collection and change detection
- Delivery failure tracking in job status
- Delivery guidance in cron prompts — stops send_message thrashing
- MEDIA files delivered as native platform attachments
- [SILENT] suppression works anywhere in response
- Cron path traversal hardening
Tool System
Terminal & Execution
- Execute_code on remote backends — Docker, SSH, Modal, and other remote terminal backends
- Exit code context for common CLI tools — helps agent understand what went wrong
- Progressive subdirectory hint discovery — agent learns project structure as it navigates
- notify_on_complete for background processes
- Docker env config — explicit container environment variables via docker_env
- Approval metadata included in terminal tool results
- Workdir parameter sanitized across all backends
Browser
- Switched managed browser provider from Browserbase to Browser Use
- Firecrawl cloud browser provider
- JS evaluation via browser_console expression parameter
- Windows browser fixes
MCP
- MCP OAuth 2.1 PKCE — full standards-compliant OAuth client
- OSV malware check for MCP extension packages
- Prefer structuredContent over text + no_mcp sentinel
- Unknown toolsets warning suppressed for MCP server names
Web & Files
- .zip document support + auto-mount cache dirs into remote backends
- Redact query secrets in send_message errors
Delegation
- Credential pool sharing + workspace path hints for subagents
ACP (VS Code / Zed / JetBrains)
- Aggregate ACP improvements — auth compat, protocol fixes, command ads, delegation, SSE events
Skills Ecosystem
Skills System
- Skill config interface — skills can declare required config.yaml settings, prompted during setup, injected at load time
- Plugin CLI registration system — plugins register their own CLI subcommands without touching main.py
- Request-scoped API hooks with tool call correlation IDs for plugins
- Session lifecycle hooks — on_session_finalize and on_session_reset for CLI + gateway
- Prompt for required env vars during plugin install
- Plugin name validation — reject names that resolve to plugins root
- pre_llm_call plugin context moved to user message to preserve prompt cache
New & Updated Skills
- popular-web-designs — 54 production website design systems
- p5js creative coding
- manim-video — mathematical and technical animations
- llm-wiki — Karpathy’s LLM Wiki
- gitnexus-explorer — codebase indexing and knowledge serving
- research-paper-writing — AI-Scientist & GPT-Researcher patterns
- blogwatcher updated to JulienTant fork
- Claude Code skill comprehensive rewrite v2.0 + v2.2
- Manim CE reference docs expanded — geometry, animations, LaTeX
- Code verification skills consolidated into one
Security & Reliability
Security Hardening
- Consolidated security — SSRF protections, timing attack mitigations, tar traversal prevention, credential leakage guards
- Cross-session isolation + cron path traversal hardening
- Workdir parameter sanitized in terminal tool across all backends
- Approval ‘once’ session escalation prevented + cron delivery platform validation
- Profile-scoped Google Workspace OAuth tokens protected
Reliability
- Aggressive worktree and branch cleanup to prevent accumulation
- O(n^2) catastrophic backtracking in redact regex fixed — 100x improvement on large outputs
- Runtime stability fixes across core, web, delegate, and browser tools
- API server streaming fix + conversation history support
- OpenViking API endpoint paths and response parsing corrected
Notable Bug Fixes
- 9 community bugfixes salvaged — gateway, cron, deps, macOS launchd in one batch
- Batch core bug fixes — model config, session reset, alias fallback, launchctl, delegation, atomic writes
- Batch gateway/platform fixes — Matrix E2EE, CJK input, Windows browser, Feishu reconnect + ACL
- Stale test skips removed, regex backtracking, file search bug, and test flakiness
- Nix flake — read version, regen uv.lock, add hermes_logging
- Lowercase variable redaction regression tests
Testing
- 57 failing CI tests repaired across 14 files
- Test suite re-architecture + CI failure fixes
- Codebase-wide lint cleanup — unused imports, dead code, and inefficient patterns
- browser_close tool removed — auto-cleanup handles it
Documentation
- Comprehensive documentation audit — fix stale info, expand thin pages, add depth
- 40+ discrepancies fixed between documentation and codebase
- 13 features documented from last week’s PRs
- Guides section overhaul — fix existing + add 3 new tutorials
- Salvaged 4 docs PRs — docker setup, post-update validation, local LLM guide, signal-cli install
- Discord configuration reference
- Community FAQ entries for common workflows and troubleshooting
- WSL2 networking guide for local model servers
- Honcho CLI reference + plugin CLI registration docs
- Obsidian Headless setup for servers in llm-wiki
- Hermes Mod visual skin editor added to skins page
Upgrade
hermes updateFor new installations, visit the install guide.
← Hermes Agent Changelog