Rust Unsafe Boundaries

SkillDev tools

Isolate and review unsafe Rust behind small, documented, testable boundaries with explicit invariants. Use when writing, refactoring, or reviewing unsafe code, raw pointers, unsafe functions, unsafe traits, MaybeUninit, pointer aliasing, panic safety, Miri checks, or safe abstractions over unsafe internals.

Use Rust Unsafe Boundaries in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Rust Unsafe Boundaries and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Rust Unsafe Boundaries skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Rust Unsafe BoundariesStart free

What this skill tells your AI

The instructions your AI receives, as published by hashgraph-online/awesome-codex-plugins in plugins/LVTD-LLC/skills/skills/rust-unsafe-boundaries/SKILL.md and read by Ahel’s review.

Use this skill to isolate unsafe Rust behind small, documented, testable boundaries. Unsafe code is acceptable only when a safe API cannot express the needed operation with acceptable correctness and performance.

Core Workflow

  1. Try the safe design first. Check standard APIs, ownership restructuring, iterators, synchronization primitives, and existing crates.
  2. State the invariant that safe Rust cannot prove. If the invariant cannot be written down, do not write unsafe code yet.
  3. Keep unsafe blocks tiny. Put runtime checks and setup in safe code before the block.
  4. Add a // SAFETY: comment immediately before each unsafe block explaining why every unsafe operation inside is valid.
  5. Mark a function unsafe fn only when callers must uphold extra conditions. Document those conditions in a # Safety section.
  6. Enable or respect unsafe_op_in_unsafe_fn; unsafe operations inside unsafe functions should still be wrapped in explicit unsafe blocks.
  7. Test normal behavior, boundary cases, panic paths, and drop behavior. Run Miri when the project supports it.

Boundary Rules

Read references/safety-invariants.md before adding or approving unsafe code.

  • Prefer private unsafe internals plus a safe public wrapper.
  • Prefer MaybeUninit<T> over deprecated or ad hoc uninitialized memory patterns.
  • Never create references from raw pointers unless validity, alignment, initialization, aliasing, and lifetime are all proven.
  • Do not use set_len, pointer arithmetic, or from_raw_parts without proving capacity, initialization, and ownership.
  • Make panic safety explicit when partially initialized values, manual drops, or length changes are involved.
  • Avoid static mut; prefer OnceLock, LazyLock, atomics, or locked state.

Documentation Pattern

/// # Safety
///
/// `ptr` must be non-null, aligned for `T`, initialized, and valid for reads
/// for the returned lifetime. No mutable reference may alias the same value.
pub unsafe fn read_ref<'a, T>(ptr: *const T) -> &'a T {
    // SAFETY: The caller guarantees `ptr` satisfies the documented contract.
    unsafe { &*ptr }
}

Review Checklist

  • Every unsafe block has a local SAFETY explanation.
  • Every unsafe fn or unsafe trait has a # Safety contract.
  • Public safe APIs cannot be used to violate internal invariants.
  • Drop, panic, and early-return paths preserve initialization and ownership.
  • Tests or Miri cover the dangerous edge, not only the happy path.

Signals

GitHub stars
1k
Forks
316
Last commit
Oct 2026
Advanced
Item type
skill
Key
rust-unsafe-boundaries
Source
github.com/hashgraph-online/awesome-codex-plugins