1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
|
{ config, hermes-agent, ... }:
{
imports = [
hermes-agent.nixosModules.default
];
age.secrets.renard-agent = {
file = ../secrets/renard-agent.age;
owner = "renard";
group = "renard";
mode = "0440";
};
services.hermes-agent = {
enable = true;
# the agent has it's own user on the host under
# which it runs and can change state at.
user = "renard";
group = "renard";
stateDir = "/home/renard";
extraDependencyGroups = [ "messaging" ];
environmentFiles = [
# required secrets:
# * OPENROUTER_API_KEY
# * DISCORD_BOT_TOKEN
# * DISCORD_ALLOWED_USERS
config.age.secrets.renard-agent.path
];
settings = {
model = {
provider = "openrouter";
# my current favorite budget model, with recent price
# changes by far the cheapest for the performance,
# with other chinese models also being close.
# only modality is text, requires auxiliary vision model!
# native price: $0.44/m in, $0.87/m out.
default = "deepseek/deepseek-v4-pro";
};
agent = {
reasoning_effort = "high"; # 'high' rated highest in benchmarks, tune to taste!
max_turns = 90;
};
auxiliary.vision = {
# auxiliary vision model for image comprehension.
# this qwen model is built for this exact use-case and is extremely
# cheap for the performance (though below gemini), however,
# it does not do any reasoning on the visual input!
# possible upgrade path for reasoning: qwen3-vl-30b-a3b-thinking
# native price: $0.13/m in, $1.56/m out.
provider = "openrouter";
model = "qwen/qwen3-vl-32b-instruct";
timeout = 120;
};
toolsets = [ "all" ];
terminal = {
backend = "local";
persistent_shell = true;
timeout = 180;
};
memory = {
memory_enabled = true;
user_profile_enabled = true;
};
# slightly higher limits and targets than default for better
# long session preservation.
compression = {
enabled = true;
threshold = 0.75;
target_ratio = 0.35;
protect_last_n = 40;
};
# progressive message editing, a little silly in discord but looks like
# the classic llm chats.
streaming = {
enabled = true;
};
display = {
streaming = true; # small regression with streaming.enabled, keep both on.
tool_progress = "new"; # only show fresh new tool calls, not every call.
};
discord = {
require_mention = false;
thread_require_mention = false;
auto_thread = false;
reactions = true;
history_backfill = true;
allow_mentions = {
everyone = false;
roles = false;
users = true;
replied_user = true;
};
};
unauthorized_dm_behavior = "ignore"; # ignore strangers
timezone = "Europe/Berlin";
group_sessions_per_user = false;
};
};
}
|
