Bound Hermes prompt and reasoning latency

platform/hermes/config.yaml

@@ -6,6 +6,10 @@ model:
   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`
@@ -52,6 +56,10 @@ agent:
   # checkpoint in a fresh turn instead of receiving an effectively unlimited
   # 500-step budget.
   max_turns: 64
+  # Keep enough deliberation for tool choice while avoiding the provider's
+  # unbounded `auto` reasoning mode on ordinary turns. Users can still raise it
+  # per session with /reasoning.
+  reasoning_effort: "low"
   gateway_timeout: 3600
   session_stall_timeout: 600
   tool_loop_guardrails:
@@ -76,20 +84,38 @@ compression:
   enabled: true
   progress_notices: true
   threshold: 0.65
+  # A common absolute ceiling keeps all router profiles responsive. It also
+  # prevents a long Medium/Large/Ultra chat from becoming impossible to move
+  # back to Fast later.
+  threshold_tokens: 60000
   target_ratio: 0.15
   tail_mode: "lean"
-  protect_last_n: 20
+  protect_last_n: 8
+  protect_first_n: 0
   proactive_prune_tokens: 50000
   proactive_prune_min_result_chars: 4000
   proactive_prune_min_reclaim_tokens: 4096
   context_total_ceiling_seconds: 600
 
 auxiliary:
+  # 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
   compression:
     provider: "main"
     model: ""
     reasoning_effort: "none"
 
+# 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:

platform/hermes/install-profiles.sh

@@ -24,6 +24,17 @@ create_profile() {
   fi
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.default "$model"
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.context_length "$context"
+  # These are platform-wide latency and loop safeguards, not model-specific
+  # tuning. Enforce them on old profiles as well as newly cloned profiles.
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.max_tokens 8192
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set agent.max_turns 64
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set agent.reasoning_effort low
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.title_generation.enabled false
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.threshold_tokens 60000
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.protect_last_n 8
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.protect_first_n 0
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set platform_toolsets.cli \
+    '["web","terminal","file","skills","todo","memory","vision","tts"]'
   [[ $(docker exec "$HERMES_CONTAINER" hermes -p "$name" config get model.default) == "$model" ]] || \
     die "Modellalias von Profil $name konnte nicht verifiziert werden."
   [[ $(docker exec "$HERMES_CONTAINER" hermes -p "$name" config get model.context_length) == "$context" ]] || \