secure-execution-review
Check MVM runtime, backend, policy, networking, package, or lifecycle changes against local authority and evidenced security guarantees. Use before designing or reviewing security-sensitive MVM work.
- License 许可
- Apache-2.0
- 所属仓库
- tinylabscom/mvm
- 最近更新
- 2026-11-15
- Stars
- ★ 47
- Forks
- 6
01 前言 Front Matter
- name
- secure-execution-review
- description
- Check MVM runtime, backend, policy, networking, package, or lifecycle changes against local authority and evidenced security guarantees. Use before designing or reviewing security-sensitive MVM work.
03 SKILL.md 正文
Review a secure execution change
Read AGENTS.md first. Start from the governing issue and the ADRs and
contracts it links. Use Trail to locate the owning production path and its
callers. ADR-001's claims table and backend matrix govern shipped security
claims; proposals, fixtures, and tests for other backends do not extend them.
Do not use this skill to turn a small unrelated change into a repository-wide
audit. For a focused security review, use the security-audit skill in guidance
mode when available; a full codebase audit requires its own request. Rust
changes still follow the repository's Rust-skill rule.
- State the contract. Identify the user-visible operation, selected backend, caller, signed plan or pack, trust boundary, and exact outcome. Classify each involved capability as shipped, preview, or proposed. If ownership, public API, authorization, or the security model changes, obtain the required design/ADR review before implementation.
- Trace the one authority path. Follow the operation from the client
through policy resolution, signed admission, the selected backend, and
host-side enforcement. For network or mailbox work, distinguish canonical
MVM local contracts from optional MVMD cross-host delivery. Check that
each backend either enforces the advertised capability or refuses it
before execution; do not introduce a second gate, guest NIC, SSH path,
silent fallback, or cache-based authorization.
When platforms genuinely implement the same capability differently,
keep the caller on one existing trait or contract and put the implementations
under
platforms/, selecting modules at their boundary rather than scatteringcfg(unix)through callers. Preserve unsupported-platform refusals; do not invent a trait for a single-platform module, a test assertion, or one syscall, or hide module declarations behindinclude!. Keep Windows/WHP as a future target of the same contract, not a current enforcement claim: a Windows adapter needs its own owner-only storage, isolation and refusal witnesses before it can replace a fail-closed platform module. - Name the threat and evidence. Say which actor may be compromised and whether protected data is at rest, in transit, in use, or retained after termination. Check issuer, digest, revocation freshness, epoch, replay, bounds, key custody, and audit where relevant. Do not call a host-held key, best-effort RAM scrub, or simulated exchange a malicious-host or physical-erasure guarantee. State any hardware requirement explicitly.
- Review changed transitions. For launch, build/cutover, mailbox, offline/reconnect, quarantine, replacement, or pack admission, enumerate states, transition owner, durable evidence, deadlines, cancellation, crash recovery, and typed refusal or unknown outcomes. A verified build is not an admitted or ready replacement; mailbox acceptance, delivery, and application acknowledgement are separate.
- Prove behavior at the right boundary. Add a happy-path witness and negative tests for the relevant tampering, wrong principal/key, stale epoch, replay, unavailable enforcement, quota exhaustion, and cleanup failure. Use the actual backend or label mock/fixture evidence as such. Measure cold and warm readiness with admission included, and record image preparation, launch, and safe replacement separately when touched. Do not trade a required check for a faster reported clock.
- Report honestly. Update the owning issue, applicable contract and user documentation with observed status and links to tests. Run the repository's scoped and required gates, name exact source revisions and commands, and call out unrun live/hardware checks. Keep unfinished checkboxes open. Do not claim another repository's design direction is implemented here.