Rust Craftsman

Triggers & Scope: Use when writing or substantially editing Rust so it lands idiomatic, gated and installed the first time. Review is the rust-reviewer agent; multi-language scaffolding is polyglot-architect.

If the repository has its own AGENTS.md, it adds the local specifics (which crate to reuse for what, the layering rule, per-binary configuration). This skill is the craft that applies in every repository.

BEFORE WRITING

• Reuse before writing. Check the workspace's shared crates and the dependencies already declared before adding a helper or a dependency. A private helper that fits should be exposed, not rewritten.
• Blast radius of a shared crate. Before changing a public item in a library others consume, list its consumers (a code graph if you have one, otherwise a workspace-wide search). Removing or renaming an export breaks consumers that only rebuild later.
• Match the neighbours: error type, logging macro, module layout, test style.

WRITING

• Errors: Result with thiserror in libraries and anyhow in binaries. No unwrap, expect or panic! outside tests. A security check that cannot evaluate returns "unsafe", never "safe".
• Strings are char-safe: truncate with .chars().take(n), never a byte slice. A byte index panics on a multi-byte boundary. Guard every truncation with a multi-byte test.
• Async: tokio. Every external call gets its own .timeout(...). Blocking work (SSH, std::thread::sleep, synchronous file walks) runs outside the runtime or in spawn_blocking, never on a worker.
• Config: endpoint and secret NAMES in code, values from the environment at runtime (treat a blank value as unset). Nothing host-specific is compiled into a library, or into anything headed for a public repository.
• Output channels: an MCP server's stdout carries only JSON-RPC frames, so logs go to stderr. A context-enrichment hook (session or prompt context) fails open: it exits 0 on every error, because a blocking exit code can erase the user's prompt, and it prints nothing on its skip path. A security or approval guard keeps its blocking exit code, since blocking is its job.
• Logs: send them through the project's logging layer (structured, to the system journal or stderr). Never write a log file under an agent's configuration directory; that directory is harness config, not a log location.
• A corner cut on purpose (global lock, O(n²) scan, heuristic) gets a comment naming its ceiling and the upgrade path.
• Every non-trivial change leaves one test behind: the smallest one that fails if the logic breaks. Real captured fixtures beat hand-written ones. If you run a local coder model, hand it the isolated unit tests and simple scaffolding to conserve cloud context.

DONE MEANS

1. Gate green: cargo fmt --check, cargo clippy --all-targets -- -D warnings and cargo test in the crate, or the repository's own gate script when it has one. Fix a red gate before anything else.
2. rust-reviewer returns PASS. Re-review by messaging the same agent with the delta, not a fresh spawn. For core crates or complex logic, add a cross-harness check on a harness other than the author's, rotating per the delegation skill (cross-agent claude -a rust-reviewer, cross-agent agy, cross-agent codex), to remove single-model confirmation bias.
3. Installed: the gate builds debug, but timers, hooks and MCP servers often run the release binary. Finish with cargo build --release -p <crate> where that applies. An MCP server then needs a reconnect and a timer picks the binary up on its next run. A stale release binary is a live bug: it can keep running old behaviour that the new code deliberately removed.
4. Verified live where the tool has live behaviour: a dry-run mode, a handshake or a smoke test against a throwaway environment.