Host-Native Governance Migration

The preferred new execution entry point is governed-execution. Existing loop-execution, executing-plans, subagent-execution, and resuming skills—and existing Claude Code commands—remain supported compatibility profiles.

Compatibility guarantees

• Existing workflow Markdown remains canonical and needs no rewrite.
• Existing state and reports remain readable; new policy fields are additive.
• Generic hotl-rt execution remains available when native capability is absent, unknown, disabled, or unsuitable.
• Native drivers are opt-in and never change host permissions or settings.
• Native continuation features provide scheduling and liveness only; HOTL controller ownership, runtime limits, effect reconciliation, and receipts remain authoritative.
• New driver-managed runs require ownership. Existing state remains readable and can be claimed or explicitly taken over during verify-first resume.
• Successful finalization now pauses at ready_to_finish; explicit finish disposition is required for completed.
• No execution entry point is removed in this roadmap release.

Adoption

Run scripts/hotl-adoption-report.sh inside a project to summarize local driver, executor, and receipt evidence. The command performs no upload. Teams may consider a native default only after representative successful receipts and human review; sparse data keeps fallback as the recommendation.

Recurring Codex automations or Claude plugin monitors remain manually installed and explicitly enabled. Adoption reporting does not create or modify an automation.

Future deprecation criteria

An entry point can be deprecated only after its behavior is covered by the shared conformance suite, migration documentation exists, local adoption evidence shows a replacement is used successfully, and maintainers approve a release-noted compatibility window. Removal requires a separate human-approved design.

Memory

scripts/hotl-memory-proposal.sh <run-id> --fact <text> produces a review-required proposal linked to receipt evidence. It never writes memory. A human must verify durability, scope, and sensitivity before accepting any proposed fact.