# Compaction Model Example
#
# This example demonstrates how a model can delegate session compaction
# (summary generation) to another, cheaper/faster model via the
# `compaction_model` field.
#
# When a session is compacted (manual `/compact`, the proactive ~90% threshold
# trigger, or post-overflow recovery), the whole conversation is fed to a model
# to be summarized. That summary call is the slowest and most expensive call in
# a session, and it only runs when the context is already large. There is no
# need to spend a heavyweight reasoning model on it: point the primary model's
# `compaction_model` at a small model and docker-agent will use it for the
# summary call only. Everything else still runs on the primary model. When
# `compaction_model` is omitted, compaction reuses the agent's own model.
#
# The field can also be set directly on an agent, or on a provider definition
# as a default for every agent whose model uses that provider. Resolution
# priority: the agent-level `compaction_model` wins, then the model-level one,
# then the provider-level default.
#
# Note: if the compaction model has a SMALLER context window than the primary,
# docker-agent triggers compaction against the smaller window so the summary
# call can always ingest the conversation it must compact. Pair the primary
# with a compaction model whose context window is at least as large to keep the
# proactive trigger aligned with the primary's window.

models:
  primary:
    provider: anthropic
    model: claude-sonnet-5
    # Summarize the conversation with the cheaper Haiku model instead of Sonnet.
    compaction_model: fast
  fast:
    provider: anthropic
    model: claude-haiku-4-5
  budget:
    provider: my_anthropic
    model: claude-sonnet-4-6

providers:
  my_anthropic:
    provider: anthropic
    # Provider-level default: agents whose model uses this provider compact
    # with Haiku unless the model or the agent overrides it.
    compaction_model: anthropic/claude-haiku-4-5

agents:
  root:
    model: primary
    description: An assistant that compacts its sessions with a cheaper model.
    instruction: You are a helpful assistant.
  reviewer:
    model: anthropic/claude-sonnet-5
    description: An assistant configuring its compaction model at the agent level.
    instruction: You are a helpful code reviewer.
    # Agent-level value: wins over any model- or provider-level setting.
    compaction_model: fast
  frugal:
    model: budget
    description: An assistant inheriting its compaction model from the provider.
    instruction: You are a helpful assistant.