Usage, cost & context

Live meters

The engine emits a context meter with every model step, so the TUI footer moves while a turn runs, not only between turns:

• context percentage (the ctx NN% gauge)
• session cost in USD
• cache-hit percentage
• reasoning effort, whenever the model exposes reasoning control

Pricing honesty

ka refuses to display numbers it cannot stand behind:

• Curated placeholder prices are flagged priced = false; the footer and ka models never show costs computed from unverified rows.
• The models.dev overlay (cargo xtask models-sync in the repository) carries real published pricing.
• Subscription plans (coding-plan models) appear as unpriced plan rows; you pay the subscription, not per token.
• With an unpriced model active, guarded-mode approval prompts omit the cost estimate instead of inventing one.

Guards

Two thresholds, both off by default, ask once per session when crossed:

[guards]
spend_usd   = 2.0   # session total cost
context_pct = 90    # context meter percentage

Answering stop aborts the current turn cleanly; the strand stays intact.

What happens as context fills

The context survival ladder kicks in, cheapest first:

1. Caps + spill: oversized tool output becomes a spill:// pointer on disk before the model sees it.
2. Prune: blank tool results outside the protected window are elided.
3. Digest: a same-model summary (/compact [focus] triggers one manually) that never separates a tool call from its result.
4. Promotion: with [context] promote = true (the default), an overflowing turn can be re-dispatched to a bigger-context sibling of the same family.

/context in the TUI shows where the window stands.