48 lines
1.4 KiB
YAML
48 lines
1.4 KiB
YAML
|
# Hermes config for gerhard — evaluation instance
|
||
|
|
#
|
||
|
|
# LLM calls go to festinger, which sits in front of the real inference
|
||
|
|
# providers. The agent identity is encoded in the base URL so festinger
|
||
|
|
# can route to the right model and build the per-agent knowledge graph.
|
||
|
|
#
|
||
|
|
# festinger routing: http://festinger:11434/gerhard/v1/chat/completions
|
||
|
|
# → festinger looks up agent_name="gerhard" in agent_models table
|
||
|
|
# → if found: routes to the registered provider/model
|
||
|
|
# → if not found: falls through to upstream_openai in config.yaml
|
||
|
|
|
||
|
|
model:
|
||
|
|
provider: "custom"
|
||
|
|
base_url: "http://festinger:11434/gerhard/v1"
|
||
|
|
# No api_key needed — festinger does not authenticate OpenAI-compat requests.
|
||
|
|
# Set a placeholder so the OpenAI SDK doesn't complain.
|
||
|
|
api_key: "festinger"
|
||
|
|
# Model name is what gets sent in the request body. festinger can override
|
||
|
|
# this per-agent via the agent_models table. Set something reasonable as
|
||
|
|
# a default (festinger's upstream_openai fallback will use whatever it supports).
|
||
|
|
default: "claude-opus-4-6"
|
||
|
|
|
||
|
|
terminal:
|
||
|
|
backend: "local"
|
||
|
|
cwd: "."
|
||
|
|
timeout: 180
|
||
|
|
lifetime_seconds: 300
|
||
|
|
|
||
|
|
agent:
|
||
|
|
max_turns: 60
|
||
|
|
verbose: false
|
||
|
|
reasoning_effort: "medium"
|
||
|
|
|
||
|
|
memory:
|
||
|
|
memory_enabled: true
|
||
|
|
user_profile_enabled: true
|
||
|
|
|
||
|
|
compression:
|
||
|
|
enabled: true
|
||
|
|
threshold: 0.50
|
||
|
|
target_ratio: 0.20
|
||
|
|
protect_last_n: 20
|
||
|
|
|
||
|
|
session_reset:
|
||
|
|
mode: both
|
||
|
|
idle_minutes: 1440
|
||
|
|
at_hour: 4
|