settings.json

{
  "session": {
    "affinity_enabled": true,
    "id_header": "X-Session-ID",

    "__ttl_hours_comment": "How long to keep session affinity (in hours)",
    "ttl_hours": 24,

    "__max_sessions_comment": "Max sessions to keep in memory before pruning",
    "max_sessions": 1000
  },

  "context": {
    "__management_mode_options": "Available modes: static, dynamic, reservoir, adaptive, disabled",
    "management_mode": "reservoir",

    "__static_recent_keep_comment": "Number of recent messages to keep verbatim",
    "static_recent_keep": 10,

    "__dynamic_utilization_target_comment": "Target % of context tokens to aim for",
    "dynamic_utilization_target": 0.8,

    "__dynamic_min_utilization_comment": "Minimum % of context tokens to use",
    "dynamic_min_utilization": 0.3,

    "__dynamic_max_boost_comment": "Maximum multiplier for dynamic boost",
    "dynamic_max_boost": 1.5,

    "__reservoir_recent_keep_comment": "Recent messages to keep verbatim in reservoir mode",
    "reservoir_recent_keep": 15,

    "__reservoir_summary_budget_comment": "Token budget for summarization in reservoir mode",
    "reservoir_summary_budget": 400
  },

  "summarization": {
    "enabled": true,
    "mode": "extractive",

    "__max_tokens_comment": "Max tokens for generated summary",
    "max_tokens": 200,

    "__timeout_ms_comment": "Timeout for summarization in milliseconds",
    "timeout_ms": 2000,

    "fallback_to_extractive": true
  },

  "routing": {
    "global_provider_lock": true,
    "__comment": "Ensures only one request per provider at a time, eliminates 429 errors",

    "__use_server_side_system_prompt_comment": "When true (default), the server always injects its STANDARD_SYSTEM_PROMPT + style_directive and ignores the client's system message. When false, the client's messages array is forwarded verbatim, the client's own system message (if any) is respected, and no server-side prompts are added. Set to false for agent frameworks (LangChain, AutoGen, CrewAI) that need full control over their system prompt and message ordering. Can be overridden per-request via the X-Use-ServerSide-System-Prompt header.",
    "use_server_side_system_prompt": true
  },

  "http": {
    "timeout_seconds": 60,
    "__timeout_seconds_comment": "HTTP request timeout in seconds"
  }
}