Loadouts

Three Concepts

Implicit Loadouts

For every configured provider profile, every registry model that provider serves is automatically a usable loadout. With an OpenAI key configured, gpt-5.6-terra is already a loadout name. Nothing to set up, nothing stored.

You only create a loadout to override something:

// ~/.kollab/config.json -> kollabor.llm.loadouts
{
  "fable-slim": {
    "provider_profile": "anthropic",
    "model": "claude-fable-5",
    "temperature": 0.3,
    "effort": "high",
    "max_tokens": 32000,
    "description": "Fable with a small output budget"
  }
}

Only provider_profile and model are required. An explicit loadout may shadow an implicit name; the explicit one wins.

The /llm Command

/llm (aliases /loadout, /ld) opens a fullscreen picker: saved loadouts first, then the implicit models grouped per provider, with context window and pricing notes from the registry. The active one is marked [*].

Picker Keys

• Type to filter. Esc clears an active filter first, closes second.
• Enter activates the highlighted loadout.
• N creates a new loadout, pre-filled from the highlighted row.
• E edits an explicit loadout; on an implicit row it branches a new one.
• D deletes an explicit loadout (press twice to confirm).

The create and edit form arrives pre-filled: name suggested, provider and model locked from the row you picked, defaults for temperature and effort, and a 16,384-token interactive output reserve validated against the model's ceiling. Ctrl+S saves and activates.

Direct Activation

/llm fable-slim     # exact name
/llm terra          # unique substring works too

A miss lists suggestions instead of guessing.

Launching With A Loadout

kollab --llm fable-slim   # explicit loadout
kollab --llm terra        # implicit (unique match against the catalog)

--llm resolves in order: existing profile name, then loadout (exact, then unique substring). A launch flag applies the loadout's fields in memory only; it never rewrites your saved config. Activating from inside the app does persist.

--model and --effort override individual fields on top of whatever --llm resolved to, and work on their own against the active profile:

kollab --llm fable-slim --effort ultra        # loadout's model, ultra effort
kollab --llm openrouter --model x-ai/grok-4.5 # profile's connection, this model
kollab --model gpt-5.6-terra                  # active profile, this model

Precedence: explicit flag, then loadout field, then provider profile value.

Reasoning Effort

/model effort shows the current level; /model effort high sets one. The levels are low, medium, high, xhigh, max and ultra. Providers reject anything else, so the CLI rejects it first.

Relation To /setup

/setup owns provider connections (keys, endpoints, OAuth); /llm owns everything after that. The old /profile command is gone. Provider profiles still exist under the hood (kollabor.llm.profiles holds the connections); Azure or fully custom endpoints are added there by hand.