Context Window
Domain: Memory & Context
Motivation
The token capacity of the model route being used. It is a property of the adapter that owns the provider/model route and is exposed through ctx.llm.resolveModelInfo().context; the token meter itself makes no model calls and adds no model-visible content. dsh-compaction-basic reads it to decide when automatic condensation should begin, and a single backend can serve models with different window sizes via per-model overrides.
Concrete Example
In a dsh-compaction-basic config you set thresholdRatio: 0.8 to start condensing at 80% of the window, and a modelPolicies entry can give a small-context route its own thresholdRatio: 0.7 and retainTokens: 2048.
Analogy
The fixed size of a clipboard — the model can only read what fits on it, so a long conversation has to be summarized to stay on it.