ROMA/sentient.yaml at main · Atubulated/ROMA · 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
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
115
116
# Framework-Level Configuration ONLY
# (Your sophisticated agents are already configured in agents/configurations.py)

# LLM Infrastructure (used by your AgnoAgents)
llm:
  provider: "openrouter"  # Since you're using OpenRouter
  api_key: "your-openrouter-key"  # Or OPENROUTER_API_KEY env var
  timeout: 300.0  # Increased to 5 minutes for complex code execution and reasoning
  max_retries: 3

# Cache System (for your sophisticated agent responses)
cache:
  enabled: true
  cache_type: "file"  # Persist between sessions
  # cache_dir is now automatically set to runtime/cache/agent
  ttl_seconds: 7200  # 2 hours (longer for complex reasoning)
  max_size: 500  # Fewer but larger responses

# Execution Framework
execution:
  max_concurrent_nodes: 10  # Increased from 3 for better parallelization
  max_parallel_nodes: 8     # New: Maximum nodes to process in parallel batches
  max_execution_steps: 500  # More steps for deep reasoning
  rate_limit_rpm: 30  # Reduced from 60 to avoid Gemini rate limits
  enable_hitl: false
  hitl_timeout_seconds: 1200.0  # 10 minutes for human review
  node_execution_timeout_seconds: 2400.0  # 40 minutes max for overall task execution

  # New: State management optimization
  state_batch_size: 50      # Batch size for state updates
  state_batch_timeout_ms: 100  # Max time before flushing state batch
  enable_state_compression: true  # Compress large state objects

  # New: WebSocket optimization
  ws_batch_size: 50         # Batch size for WebSocket messages
  ws_batch_timeout_ms: 100  # Max time before sending WebSocket batch
  enable_ws_compression: true  # Compress WebSocket payloads
  enable_diff_updates: true    # Send differential updates only

  # Immediate slot filling for better concurrency
  enable_immediate_slot_fill: true  # Fill slots as soon as they become available

  # Enhanced timeout strategy with escalation
  timeout_strategy:
    # Warning threshold (warn about stuck nodes)
    warning_threshold_seconds: 60.0   # 1 minute - catch hanging executors faster
    # Soft timeout (attempt recovery)
    soft_timeout_seconds: 180.0       # 3 minutes - faster recovery
    # Hard timeout (force transitions)
    hard_timeout_seconds: 300.0       # 5 minutes - more aggressive
    # Maximum recovery attempts per node
    max_recovery_attempts: 3
    # Enable aggressive recovery for deadlock situations
    enable_aggressive_recovery: true

  # CHANGED: Default to root plan only - only review the initial high-level plan
  hitl_root_plan_only: false  # Set to true to only review root node's plan

  # NEW: Force root nodes to always plan (skip atomizer) - prevents atomic classification of complex queries
  force_root_node_planning: true  # Set to false to allow atomizer to decide for root nodes

  # NEW: Skip atomization entirely - forces hierarchy/depth-based execution decisions
  skip_atomization: true  # Set to true to bypass atomizer and use max_recursion_depth rules

  # These are ignored when hitl_root_plan_only is true
  hitl_after_plan_generation: false
  hitl_after_modified_plan: false
  hitl_after_atomizer: false
  hitl_before_execute: false

# Logging - Clean and readable
logging:
  level: "INFO"
  enable_console: true
  enable_file: true
  # file_path is now automatically set to runtime/logs/sentient.log
  file_rotation: "10 MB"     # Rotate by size
  file_retention: 3          # Keep only 3 files (must be number, not string)

  # Console style options:
  # - "clean": Just messages with smart spacing and colors (default)
  # - "timestamp": Messages with timestamps
  # - "detailed": Full details with time, level, module, and message
  console_style: "clean"

  # Filter noisy logs to separate files or suppress them
  module_levels:
    # Suppress or redirect noisy components
    "sentientresearchagent.server.services.broadcast_service": "WARNING"  # Only warnings and errors
    "sentientresearchagent.server.services.project_service": "WARNING"    # Reduce auto-save noise
    "sentientresearchagent.core.project_manager": "WARNING"              # Reduce save/broadcast noise
    "sentientresearchagent.hierarchical_agent_framework.graph.graph_serializer": "WARNING"  # Reduce serialization noise

    # Keep important execution flow at INFO level
    "sentientresearchagent.hierarchical_agent_framework.orchestration": "INFO"
    "sentientresearchagent.hierarchical_agent_framework.graph.cycle_manager": "INFO"
    "sentientresearchagent.hierarchical_agent_framework.node": "INFO"
    "sentientresearchagent.hierarchical_agent_framework.node_handlers": "INFO"

# Experiment Configuration
experiment:
  base_dir: "experiments"  # Base directory for all experiment outputs
  results_dir: "results"  # Subdirectory for results
  emergency_backup_dir: "emergency_backups"  # Subdirectory for emergency backups
  configs_dir: "configs"  # Subdirectory for experiment configs
  retention_days: 30  # Keep experiment results for 30 days
  auto_cleanup: true  # Automatically clean old results
  timestamp_format: "%Y%m%d_%H%M%S"  # Format for timestamped directories

# Environment
environment: "research"

# Default Agent Profile
default_profile: "general_agent"

# NO AGENT CONFIGURATION HERE - Your agents are already perfectly configured!