Fixes
- Turn summaries now include token usage from worker and nested subagents, grouped by model and applied reasoning effort. GPT-6-Luna appears in the breakdown when it was used by a subagent.
- Matching model and effort rows are merged using the existing usage ledger, so repeated reports are counted once.
Validation
- 20 usage accounting tests and 3 summary rendering tests passed.
- Release installer tests passed.
Full Changelog: v2.0.20...v2.0.21