Multi-stage context-compaction runner.

The Claude Code paper describes Anthropic's compaction as a five-layer pipeline (Budget Reduction → Snip → Microcompact → Context Collapse → Auto-Compact / Summary), each layer cheaper than the next. The cheaper layers run first; the LLM-summary layer fires only when deterministic shrinkage couldn't get the conversation under target.

The pipeline implements the existing ExAthena.Compactor behaviour, so loop.ex doesn't have to change — the kernel still calls compactor.should_compact?/2 then compactor.compact/2. Internally this module dispatches across a list of ExAthena.Compactor.Stage modules.

Configuration

The stage list is read from state.meta[:compaction_pipeline], falling back to ExAthena.Compactor.Stage.default_pipeline/0. Hosts can swap individual stages or replace the list entirely.

Loop.run("hi",
  provider: ...,
  compactor: ExAthena.Compactor.Pipeline,
  compaction_pipeline: [MyApp.MyStage, ExAthena.Compactors.Summary]
)

Reactive recovery

When a mode signals :error_prompt_too_long, the kernel re-invokes the pipeline with force: true, which runs every stage regardless of cost. See run/3.