AI-Profile-Router/platform/hermes/config.yaml

_config_version: 38

model:
  default: "qwen-medium"
  provider: "custom:router"
  base_url: "http://router:8081/v1"
  api_key: "${ROUTER_API_KEY}"
  context_length: 160000
  # Applies to visible text, tool calls and hidden reasoning together.  Without
  # a cap a local reasoning model can spend tens of thousands of tokens before
  # producing its first useful action.
  max_tokens: 8192
  api_mode: "chat_completions"

# A named provider is required by current Hermes releases.  A bare `custom`
# endpoint is canonicalized to a host-derived identity and can lose its key
# association in resumed sessions.  The stable identity below always resolves
# its credential from the container environment.
providers:
  router:
    name: "Athena Profile Router"
    api: "http://router:8081/v1"
    key_env: "ROUTER_API_KEY"
    transport: "chat_completions"
    default_model: "qwen-medium"
    discover_models: true

# Commands run in an isolated, persistent workspace. Host and Unraid changes
# use the audited operator/MUA MCPs instead of a Docker socket or host mount.
terminal:
  backend: "local"
  cwd: "/workspace"
  timeout: 600
  home_mode: "profile"
  persistent_shell: true
  lifetime_seconds: 1800

web:
  # Keenable is bundled with Hermes and provides both search and extraction
  # through its keyless tier. This remains available after Hermes moved to
  # Unraid, where the former Athena-local SearXNG no longer exists.
  search_backend: "keenable"
  extract_backend: "keenable"
  extract_char_limit: 15000
  keyless_fallback: true
  keyless_rescue: true

# Hermes discovers MCP servers in the background.  Athena has several remote
# servers and the Home Assistant relay can finish just after the short upstream
# default, leaving its tools absent from the first agent turn until a manual
# reload.  Wait long enough to build the initial tool snapshot completely;
# completed discovery returns immediately, so this does not add a fixed delay.
mcp_discovery_timeout: 15.0
mcp_single_query_discovery_timeout: 30.0

agent:
  # A single turn may be substantial, but must not silently consume an entire
  # context window in a no-progress loop. Long builds continue from a compact
  # checkpoint in a fresh turn instead of receiving an effectively unlimited
  # 500-step budget.
  max_turns: 64
  # Ordinary turns run without hidden reasoning. Users can enable it explicitly
  # per session with /reasoning medium or /reasoning high when a task needs it.
  reasoning_effort: false
  gateway_timeout: 3600
  session_stall_timeout: 600
  tool_loop_guardrails:
    warnings_enabled: true
    hard_stop_enabled: true
    warn_after:
      exact_failure: 2
      same_tool_failure: 3
      idempotent_no_progress: 2
    hard_stop_after:
      exact_failure: 5
      same_tool_failure: 8
      idempotent_no_progress: 5
    loop_caps:
      max_web_searches: 8
      max_subagents: 8

# Keep ample room for long agent work. Compression starts late; deterministic
# tool-result pruning and rolling micro-compaction keep it from getting there
# during normal jobs.
# router profile is selected. Keep the result compact enough that a local
# model does not spend many minutes generating the handoff.
compression:
  enabled: true
  # Fold one older assistant/tool exchange into the rolling summary every five
  # completed turns. The selected 27B profile remains the summarizer; there is
  # no separate low-fidelity compression model.
  micro_compact: true
  micro_compact_every_n_turns: 5
  micro_compact_defrag_threshold_tokens: 2000
  progress_notices: true
  threshold: 0.95
  target_ratio: 0.15
  max_attempts: 1
  tail_mode: "lean"
  protect_last_n: 20
  protect_first_n: 0
  proactive_prune_tokens: 32000
  proactive_prune_min_result_chars: 2000
  proactive_prune_min_reclaim_tokens: 2048
  context_timeout_seconds: 45
  # A failed local summarizer must not block an interactive client for ten
  # minutes. Continue without dropping messages after two minutes.
  context_total_ceiling_seconds: 120

auxiliary:
  compression:
    provider: "main"
    reasoning_effort: false
    timeout: 120
    max_concurrency: 1
  # Session names are cosmetic and used to create a second concurrent LLM
  # request after every first reply. On a single inference slot this blocks the
  # actual chat, so keep the original timestamp/session id instead.
  title_generation:
    enabled: false

# The full hermes-cli preset injects several large, overlapping schemas on
# every turn. Athena already exposes browsing, orchestration and host services
# through its MCPs; retain the generally useful local primitives and load MCP
# tools through Hermes' deferred discovery. This setting is shared by every
# model profile.
platform_toolsets:
  cli: [web, terminal, file, skills, todo, memory, vision, tts]

# German is the platform default. The display setting localizes the static
# messages Hermes currently supports; agent replies are governed by SOUL.md.
display:
  language: "de"
  show_reasoning: false

# Speech input stays local and private. The fixed German hint helps with
# short utterances and product names.
stt:
  enabled: true
  provider: "local"
  language: "de"
  prompt: "Hermes, Athena, Qwen, OpenWebUI, Unraid, Home Assistant, Sonarr, Radarr, Navidrome"
  local:
    model: "base"
    language: "de"

# Reuse Athena's OpenAI-compatible Qwen3-TTS route.
tts:
  provider: "openai"
  speed: 1.0
  openai:
    model: "qwen3-tts"
    voice: "alloy"
    speed: 1.0
    base_url: "http://router:8081/v1"

voice:
  auto_tts: true
  client_direct: false

skills:
  creation_nudge_interval: 20

plugins:
  enabled: ["browser-browser-use", "web-keenable"]

timeouts:
  tools:
    concurrent_batch: 900
    sequential_call: 900

# BEGIN MANAGED MCP SERVERS
mcp_servers: {}
# END MANAGED MCP SERVERS