Deep Context & Ultra-Economy Settings

Context budget by mode

Mode	Token budget	Compression	Notes
plan	32 000	standard	One-shot, structured JSON output
agent	32 000	standard (configurable)	Interactive; approval gates
full	128 000	off	Full context — no compression
exec	32 000	aggressive	Harness loop; maximise iterations

full mode uses a 4× larger budget and disables all compression, trading token cost for maximum fidelity. It is the right choice for complex multi-file tasks where losing context mid-turn would require backtracking.

Configuring compression level

The default compression level for agent/plan modes is read from ~/.umbra/settings.json:

{
  "compression": {
    "level": "standard"
  }
}

Valid values: off / lite / standard / aggressive / ultra. The exec harness always uses aggressive regardless of this setting.

Default runtime mode

Set via launch flag --mode when starting Umbra, or switch mid-session using /agent, /full, or /plan in the TUI. The active mode is shown in the session status bar. Persisted across launches via --mode flag or runtime-preferences.json → defaultMode.