Configuration Reference

Inferoa loads configuration from the user config path unless --config is provided. By default, the user config path is:

~/.inferoa/config.yaml

Use INFEROA_STATE_DIR or --state-dir to change the state directory. The state directory also holds the local secret vault, session database, and other durable state.

Top-Level Shape

The full default configuration is written by inferoa debug init. The defaults shipped in src/config/defaults.ts are:

model_setup:
  mode: direct
  provider: vllm
  base_url: http://localhost:8000/v1
  context_window: 32768

model_retry:
  initial_delay_ms: 1000
  max_delay_ms: 60000
  backoff_factor: 2
  jitter_ratio: 0.2
  request_timeout_ms: 300000

omni:
  enabled: false
  endpoints: {}

permissions:
  mode: full_access

context:
  compression_threshold: 0.8
  context_window: 32768
  protected_recent_loops: 3
  engine:
    provider: auto
    startup: welcome
    require_ready_before_chat: true
    watch: true

skills:
  enabled:
    - coding-workflow
  managed_installs: ask

web_search:
  provider: auto

rtk:
  enabled: true
  delivery: managed
  version: 0.42.3
  auto_download: true

daemon:
  poll_ms: 1000

Key Fields

• model_setup.mode — one of direct, auto, or external. auto delegates model selection to vLLM Semantic Router.
• model_setup.provider — provider id such as vllm, vllm-sr, or an external provider id from the setup wizard.
• omni.endpoints — keyed by capability. Supported keys are vision, image_generation, image_edit, video_understanding, video_generation, audio_understanding, audio_generation, and speech.
• permissions.mode — workspace permission mode. One of full_access, auto_approve, ask, or custom. Override per workspace from the TUI with /access.
• context.engine.provider — auto, codegraph, builtin, or off.
• skills.managed_installs — ask, always, or never.

Environment Overrides

Environment variables override config at startup. They never persist to disk.

Variable	Effect
INFEROA_BASE_URL	Overrides model_setup.base_url
VLLM_BASE_URL	Fallback override for model_setup.base_url
INFEROA_MODEL	Overrides model_setup.model
VLLM_MODEL	Fallback override for model_setup.model
INFEROA_MODE=auto	Sets model_setup.mode=auto and uses the vLLM Semantic Router
INFEROA_RTK	Enables or disables RTK
INFEROA_RTK_PATH	Uses a specific RTK binary path
INFEROA_RTK_AUTO_DOWNLOAD	Enables or disables managed RTK download
INFEROA_OMNI_VISION_URL	Enables and configures the Omni vision endpoint URL
INFEROA_OMNI_IMAGE_URL	Enables and configures image generation
INFEROA_OMNI_IMAGE_EDIT_URL	Enables and configures image editing
INFEROA_OMNI_VIDEO_URL	Enables and configures video generation
INFEROA_OMNI_SPEECH_URL	Enables and configures speech generation

Each Omni URL override also supports a matching INFEROA_OMNI_<KEY>_MODEL variable. The accepted keys are the endpoint names: VISION, IMAGE, IMAGE_EDIT, VIDEO, SPEECH.

Secret Handling

Config files should store api_key_ref, not raw api_key values. The setup wizard writes raw secrets into the local vault and stores only references in the YAML file. The same rule applies to model_setup, omni.endpoints.*.api_key, and web_search.api_key.

inferoa debug setup redacts any *api_key* field before printing. Public docs, progress logs, and evidence artifacts must not include raw keys.