Malbox End-to-End Tracker

[Mjoyufull]

v0.1.0 - Open Issues

A. Task Model / Submission / Scheduler

• A1 - Task plugin selection + config flow broken - Task creation stores plugins = ["0"]. Worker ignores tasks.plugins and iterates the registry snapshot. Worker-side plugin config building is a TODO.
• A2 - Submission API / runtime alignment partial - Task-create API accepts package, module, options, machine, custom, memory, unique - none are meaningfully honored at runtime.
• A3 - Submission-path panic cleanup partial - Request path uses unwrap() / expect() where it should return proper API errors.
• Timeout semantics partial - enforce_timeout is stored, but worker timeout is just "wrap execute_task() in a timeout".

B. Machine Lifecycle / Cleanup / Recovery

• B1 - Timeout-safe machine ownership broken - Once execute_task() is canceled, worker loses machine_id and the machine stays assigned until restart.
• B1 - Post-acquire failure cleanup broken - Later failure paths can bubble out without guaranteed release/repair. Not exception-safe in general.
• B2 - Completed vs. reverted contract partial - Task Completed means "worker reached the release path", not "machine is confirmed clean". Needs an explicit contract.
• B3 - Recovery and reprovision fidelity partial - Recovery is restart-centric, not self-healing while live. Interrupted Provisioning machines get restored to Ready with provider_config = None - lossy.

C. Guest Access / Networking / Trust Model

• C1 - Guest RPC contract + trust/auth broken - Guest sample execution hard-requires gRPC on port 50051. Disabling guest_access doesn't actually disable guest execution semantics. Guest gRPC is a trusted-lab transport assumption, not an authenticated boundary. Registration is reachability-based, not handshake-based.
• C2 - Guest endpoint discovery / IP refresh partial - DHCP lease lookup works for libvirt but the IP is persisted and trusted later without revalidation.
• C3 - Transport config authority partial - ResolvedTransport::Grpc stores an address, but workers still dial http://<db_machine.ip>:50051 directly. network_isolated = false is passed unconditionally. Provider-side endpoint reconstruction hardcodes Windows in places.
• C4 - Guest runtime config parity partial - Windows guest-plugin deploy sets MALBOX_WORK_DIR, but the default guest runtime only reads MALBOX_PLUGIN_PORT - override path isn't wired.
• Task lifetime ownership ambiguous - Sample execution is fire-and-observe. Open decision: should task lifetime be driven by sample exit, observation window, plugin-controlled keepalive, overall timeout, or a mix?
• Sequential guest-plugin ports partial - 50051 + index is brittle if plugin ordering becomes dynamic. First plugin on 50051 also acts as the base guest agent.

D. Plugin Runtime / Ordering / Lifecycle

• D1 - Host plugin reality check broken - Persistent host plugins spawn in Starting, manager acquires Ready/Busy, but the explicit host execution path returns empty results. Event-runtime fallback gives task handlers an empty sample path while examples expect sample bytes.
• D2 - Deterministic plugin ordering + execution contexts broken - Plugin selection and ordering are not task-driven or deterministic. Manifest contexts (exclusive, sequential, parallel) are parsed but not enforced.
• D3 - Guest plugin lifecycle contract partial - Guest plugin lifetime is VM-boot scoped, not task scoped. Worker doesn't drive initialize() / shutdown() in a real lifecycle sense.
• D4 - Event-hook / runtime alignment partial - Event-hook docs and manifests are ahead of reality. Worker only emits a small subset of task events over host IPC; guest event delivery doesn't appear to be used operationally.
• Plugin acquisition tracking partial - Still records Busy { task_id: 0 } instead of the actual task ID.
• Live reconcile wiring partial - Registry pending-change application and manager reconcile exist, but no top-level runtime loop wires plugin watch changes into live reconcile.
• Lifecycle types partial - Persistent and ephemeral are partially real; scoped is not implemented.

E. Results / Reports / Contributor Health

• E1 - Result artifact read/download API partial - Result retrieval is metadata-only.
• E2 - Result model consolidation partial - Two result stories: per-plugin artifact path and older aggregate TaskResult shape. TaskStore::update_task_result() is still a TODO.
• E3 - SQLx contributor workflow + CI broken - task_results repo breaks clean contributor test runs unless DATABASE_URL or prepared SQLx metadata exists.

F. Docs / Config / Operator Surfaces

• F1 - Config / docs drift cleanup doc-lag - Sample config references the removed native provisioner. Some operator-facing comments still reference "native". CLI grouping is aspirational in places.
• F2 - Front-end truthfulness doc-lag - Svelte front end is mock-backed.
• F3 - Security assumptions / lab-mode docs missing - No auth model across HTTP API, CLI, and web UI.
• Windows provisioning shortcuts security - Relies on Administrator / packer, WinRM, insecure cert validation, guest plugin scheduled tasks as SYSTEM.
• Linux provisioning shortcuts security - Base images allow password SSH, malbox user has passwordless sudo.
• Network isolation partial - Isolation controls exist in playbooks as image/provisioning policy, not centrally enforced. Libvirt domain build hardcodes default NAT network + e1000 NIC.

[Mjoyufull]

notes
previsioning is not supposed to fail