v0.20.2

Highlights

A small maintenance release on top of v0.20.1. The headline change is per-checkpoint cost visibility: Kitaru now records LLM token counts and cost on each individual checkpoint, not only on the execution as a whole, so you can see which step of an agent run actually spent the money. Cached and replayed checkpoints show what the work would have cost while billing nothing for it, which makes a replay's savings visible instead of invisible. The per-checkpoint values are exposed through the SDK, kitaru executions get -o json, MCP, and the dashboard.

Added

• LLM cost and token usage is now attributed per checkpoint, not just per execution. When a flow reaches a terminal state, Kitaru publishes flat kitaru_llm_*_v1 metadata on each checkpoint that made model calls, so SDK, CLI, MCP, and dashboard clients can see which checkpoint incurred which cost. Per-checkpoint values sum exactly to the execution-level totals, and the execution-level llm_usage_summary_v1 payload is unchanged. (#528)
• Cached, skipped, and replay-reused checkpoints report their token counts under the reused_* fields with zero incurred and zero display cost, so a replayed run shows what the work would have cost without billing it again. Retried checkpoints keep a separate record per attempt, so a retry that really called the provider twice is counted twice. (#528)

Changed

• Terminal cost metadata writes are best-effort: a failed write is debug-logged, the remaining checkpoint writes are still attempted, and incomplete persistence is reported so aggregation can be retried. A metadata write failure never fails the flow run. (#528)

Infrastructure

• The release workflow's GitHub Release asset reconcile step now skips signature and attestation assets, which are regenerated per dispatch and previously caused recovery re-dispatches to hard-fail on a non-reproducible asset mismatch. (#513)

Full Changelog: v0.20.1...v0.20.2