Global agent configuration

Global agent configuration provides shared defaults for your agents in ~/.skiller/settings/agent.json. Authentication saves your provider and model selection here; you can customize the other settings as needed.

Configuration file

This example selects a model and shows the main configuration blocks. Choose a provider and model available in your LLM provider catalog.

{
  "llm": {
    "provider": "minimax",
    "model": "MiniMax-M3.1-Flash",
    "effort": 3
  },
  "loop": {
    "max_turns": 30,
    "max_tool_calls": 10
  },
  "context": {
    "compaction": {
      "compaction_trigger_ratio": 0.8,
      "compaction_target_ratio": 0.5,
      "keep_last_blocks": 5
    }
  },
  "tools": {},
  "debug": {
    "log_request": false,
    "log_streaming": false,
    "log_override_file": true
  },
  "event_output": {
    "truncate": {
      "enabled": true,
      "max_text_chars": 600,
      "max_json_chars": 4000,
      "max_array_items": 20
    }
  }
}

llm

Selects the provider, model, and optional reasoning effort. See Model selection for fields, defaults, and global or per-agent examples. Connection settings and credentials belong in the provider catalog.

loop

Limits model turns and tool calls per response within an agent step. See Harness loop limits for fields, defaults, and what happens when a limit is reached.

context

Controls the context window and when conversation history is compacted. See Context settings for fields, defaults, and global or per-agent examples.

tools

Stores configuration and permissions for registered tools. See Harness tools for enabling tools and configuring Shell and Files permissions.

debug

Controls LLM request and response diagnostics. See Debug settings for fields, log files, and defaults.

event_output

Limits previews in runtime events without changing the agent’s context. See Event output for fields and defaults.