Faster skill execution
- Correction snapshots now start at the previous reviewed ledger head, so reviewers receive the correction diff rather than the full plan-base diff.
- The plan checker reports dependency-ready packages through
--completedand--running, respecting concurrency and owner limits. Runtimes without safe individual completions keep wave barriers. record_evidence.py --reuse-identicalcan reuse exact, repeat-safe commands within one invocation while retaining a record and log for every criterion. Reuse is opt-in.- Independent Git validation fixtures run with three bounded workers. Cheap structural checks precede expensive fixtures.
- Scenario recordings retain runtime duration, cost, and token usage.
- Mutation diagnostics use UTF-8 on Windows. A labelled filter supports focused diagnosis without replacing the full mutation suite.
Verification
- 13 performance regression tests passed; all 160 deliberate validator mutations were caught locally.
- All 37 behavior scenarios were freshly recorded for this release and validated.
- CI passed before publication.
- Three scenario assertions now recognize wrapped architecture hand-offs, explicit deploy-time preview recommendations, and explicit refusal to delete an indeterminate gate; their required behavior is unchanged.
Limits
A local sample reduced Git fixture checks from 26.0 to 14.9 seconds. No matched whole-agent benchmark has measured this release, so end-to-end speed gains remain unverified. Dependency-ready scheduling relies on the lead supplying accepted, integrated package state and immutable inputs for work still running.
Includes #47.