Zombie Killer follows evidence and safety gates, not calendar promises. Items below are direction, not commitments.
- Local engine, CLI, bounded protocol, and standard-library Python SDK.
- Evidence-based Swarm scan plus process-instance-bound action domain primitives and native transport.
- SDK root/fingerprint declarations composed into manifest-scoped Graveyard quarantine and exact restore, including bounded legacy restore-only recovery.
- Registered JSON/JSONL and constrained SQLite Memory Vault scalar repair, verified backups, snapshot history, and planned restore; Chroma and common FAISS flat-header inventory remain read-only.
- Tauri/React desktop workflows, strict native method/result validation, finite headless CLI commands, and synthetic browser demo.
- Persistent one-shot plan authority, mixed-operation activity receipts, local no-overwrite diagnostics, applied revisioned settings, accessibility support, and strict CSP.
- Concurrency and recovery hardening with explicit cooperating-writer limits for JSON and transactional in-place SQLite restore.
- Unsigned native archives preserving the engine resource layout, hidden manifest entries, source legal material, and bundled Python runtime attribution inventory.
- Keep the verified identity-bound CLI containment path covered by PID-reuse, target-changed, and clean-machine release regression tests.
- Keep completed service/broker/desktop/CLI workflows covered by separate-instance and interruption regressions.
- Verify supported local stores and declared resources with synthetic fixtures before using important workloads; maintain independent backups.
- Expand Windows and Linux fixture coverage for permission failures, process trees, mounts, junctions, and service supervisors. Port and qualify macOS native workflows separately before offering macOS artifacts.
- Reconcile interrupted operations at startup and expose clearer degraded-coverage diagnostics.
- Review complete platform binary third-party obligations, compare lock/source and manifest-bound binary SBOMs, and run clean-machine installation tests. The bundled Python legal inventory is attribution evidence, not complete native binary legal approval.
- Signed Windows packages after maintainers control suitable certificates. Any later macOS distribution requires native qualification, signing, and notarization.
- Opt-in adapters for additional local memory formats through a documented capability contract.
- Language-neutral lifecycle SDKs and supervisor integrations.
- Locally configurable retention/purge workflows with dry-run reports and recoverability windows.
- Provider account control, remote provider memory editing, or provider safety/rate-limit bypass.
- Automatic killing based on a heuristic or signature match.
- Secret, prompt, response-body, or process-environment collection.
- Inferred billing or cost estimates without measured evidence.
- Hosted telemetry or a cloud control plane in the core project.
Proposals should open a feature request and explain evidence, privacy impact, failure behavior, and how the feature remains safe when visibility is partial.