lossless-hermes-py/plugin.yaml at main · mssteuer/lossless-hermes-py · GitHub

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
name: "lossless-hermes"
version: "1.0.0"
description: "Lossless Context Management (LCM) engine with DAG-based summarization"
author: "Michael Steuer / MAKE"
type: "context_engine"

config:
  enabled: true
  context_threshold: 0.75
  fresh_tail_count: 64
  fresh_tail_max_tokens: null
  leaf_chunk_tokens: 20000
  leaf_target_tokens: 2400
  condensed_target_tokens: 2000
  leaf_min_fanout: 8
  condensed_min_fanout: 4
  condensed_min_fanout_hard: 2
  incremental_max_depth: 1

  # LLM settings
  summary_provider: ""  # e.g. "openai" for litellm proxy, or leave empty to use host default
  summary_model: ""  # e.g. "gemini-2.5-flash" via litellm, or leave empty to use host default
  summary_timeout_ms: 60000
  custom_instructions: ""

  # Circuit breaker
  circuit_breaker_threshold: 5
  circuit_breaker_cooldown_ms: 1800000

  # Cache-aware compaction
  cache_aware_compaction:
    enabled: true
    cache_ttl_seconds: 300
    max_cold_cache_catchup_passes: 2
    hot_cache_pressure_factor: 4
    hot_cache_budget_headroom_ratio: 0.2
    cold_cache_observation_threshold: 3

  # Dynamic chunk sizing
  dynamic_leaf_chunk_tokens:
    enabled: true
    max: 40000

dependencies:
  - "sqlite3"

tools:
  - name: "lcm_grep"
    description: "Search conversation history using full-text search"
  - name: "lcm_describe"
    description: "Get LCM state and statistics"
  - name: "lcm_expand"
    description: "Expand specific content from history"