Contract for a complete native integration

jev_prune/gate.py is a tested state machine intended for a native host adapter, not a hidden replacement for any vendor runtime.

A host must implement a consistent snapshot/proposal function; a revision covering appends, pending input, model, policy and destructive edits; a durable atomic compare-and-swap pruning commit; complete model-aware next-request and scoped budget accounting; provider history/lineage rebasing; and a direct delegate to its unchanged native compactor. Compaction events must describe real compaction only. The host owns its authorization, pending task and safe scheduler boundary.

Gate.begin(episode_id, original_trigger) creates one waiting episode. The UI chooses Choice.PRUNE, Choice.COMPACT or Choice.PAUSE. Choosing Compact calls the original compactor without Jev. Choosing Prune proposes and commits a validated projection, then uses the host budget to decide READY vs WAITING. WAITING permits Compact or Pause, not another automatic prune pass. Failure is never implicit consent to summarize. Missing/unknown budget cannot yield READY. A stale episode closes and can be replaced by a new one. Old responses cannot operate on a new revision. No UI response should be injected as an LLM message.

manual_prune() returns APPLIED, not READY: it must not launch an otherwise idle agent. manual_compact() directly delegates with its existing behavior. The host should add cancellation and serialized access around the state machine; this Python example is synchronous and is not a concurrency lock by itself.

Native implementation targets still required

• Codex fork: add a real command/protocol operation plus TUI/app-server choice; integrate upstream of local/remote auto-compaction dispatch; implement active projection persistence without clearing security review/authorization state; reconcile transport lineage and token counters. Do not install the earlier private-compactor-snapshot patch as though it implements this contract.
• Claude Code: public skill and PreCompact blocking alone do not expose all host responsibilities above. Do not modify live transcript files as a workaround.
• OpenClaw: compose with its selected context engine and session lifecycle, preserving the native compact command and audit/source transcript.
• Hermes/OpenCode/Pi: promote the request-only adapters to host-owned projection, complete accounting and scheduler transitions before claiming full behavior.

Required tests beyond the current package: native next-inference capture; no compaction calls on successful prune; authoritative budgets; crash/replay/fork; retained security evidence; pending tools and user input; cancellations and client disconnect; every auto-trigger path including smaller-model switch and overflow; server-side lineage/cache consistency; UI snapshots; shared policy enforcement.

No benchmark or semantic-retention improvement is established by passing the local contract tests. Evaluate real task success and retained constraints as well as bytes/tokens, cache effects, Jev cost, latency and compaction frequency.