Accounts for the complete transformed provider context before compaction. Reported token counts are reused only for the same model, prompt, tool, operation, and activation revision. Agent run results include cumulative typed usage and nominal-cost summaries.
Marks prepared tool-call arguments explicitly, so policy events and execution share one typed argument value without running wire preparation twice.
Represents idle, running, and compacting phases as S7 values. Agent behavior no longer branches on phase-name strings.
Adds the structural SessionJournal contract and an in-memory implementation.
Assistant partials remain in the active turn and only terminal messages enter
the append-only journal; reset is represented by a committed entry instead of
erasing history.
Adds asynchronous session synchronization and compare-and-append positions.
An idle agent can rebuild its projection from an existing journal, while a
stale writer receives RhoSessionConflictErrorValue before journal mutation.
Uses only counted usage observations when projecting context usage; an unavailable provider observation causes a complete-context estimate.
Gives the provider-operation agent-loop fixture enough time on slower build hosts while retaining an explicit test timeout.
Adds the experimental asynchronous multi-turn agent loop, ordered lifecycle events, tool scheduling, queues, cancellation, and policy generics.
Adds append-only session entries, projected provider context, semantic compaction, typed skip and failure outcomes, threshold compaction, and one provider-input recovery attempt selected by an open S7 generic.