writing-trust

name

writing-trust

description

Write, fix, or convert Rust in a project that uses Trust, the strict Rust dialect. Use when you see `cargo trustc` in build instructions, `[package.metadata.trust] strict = true` (or `[workspace.metadata.trust]`) in a Cargo.toml, a `#![strict]` inner attribute at the top of a .rs file, build errors with R-codes like R0042/R0001/R0014 (`error[R0042]: call to ... must use named arguments`), named-argument call syntax such as `f(width: 1920, height: 1080)`, the pipe operator `|>`, or `requires!(...)` preconditions in function bodies. Also use when asked to convert an existing crate to Trust or scaffold a new Trust project with `trust new`.

Writing Trust

What Trust is

Trust is a strict dialect of Rust: 20 hard lints (R-codes) plus three syntax extensions, lowered to plain Rust before rustc sees it. Build with cargo trustc build (or trust build for single files) — never plain cargo: stock rustc rejects the dialect syntax and skips the lints.

Project setup

• Greenfield: trust new <name> scaffolds a strict project (Cargo.toml with the metadata key, named-arg hello-world, .gitignore, README). Build it with cargo trustc build.

• Existing crate: add two lines to Cargo.toml:

  [package.metadata.trust]
  strict = true
  

• Workspace-wide: [workspace.metadata.trust] strict = true in the root manifest opts in every member. Dependencies are never affected.

• Mixed crates / single files: put #![strict] as an inner attribute at the top of a file to opt in file-by-file. Stock rustc rejects #![strict]; only the Trust toolchain compiles it (the wrapper strips it during lowering).

The syntax you must write

• Named arguments on every call with arity > 1 to a function Trust can see (in-crate, or indexed via TRUST_SIGNATURE_PATH):

  fn make_rect(width: u32, height: u32) -> Rect { /* … */ }
  
  make_rect(width: 1920, height: 1080)   // required — never make_rect(1920, 1080)
  make_rect(height: 1080, width: 1920)   // order is free; names are checked,
                                         // lowering reorders to declaration order
  

  No mixing named and positional in one call. Arity-0/1 calls and calls into unindexed dependencies stay positional. A wrong name is error R3001.

• Pipe operator: e |> f(args) lowers to f(e, args) — the receiver becomes the first argument.

  let s = read_input()? |> normalize(mode: Mode::Strict) |> render();
  // == render(normalize(read_input()?, mode: Mode::Strict))
  

  Works with paths (e |> path::to::f(a)) and named args. Binds lower than ./field/indexing, left-associative. Keep method chains as .method(); the pipe is for free functions.

• Preconditions: requires!(cond) at the top of a strict fn body lowers to debug_assert!(cond, "requires violated: …") — checked in debug builds, gone in release. No ensures!, no solver.

  fn withdraw(balance: u64, amount: u64) -> u64 {
      requires!(amount <= balance);
      balance - amount
  }
  

Important: this syntax only compiles via cargo trustc / trust build. Plain cargo build failing on f(width: 1920, ...) is expected behavior, not a bug — do not "fix" it by deleting the names; switch the build command instead.

The iteration loop

1. cargo trustc build --message-format json — Trust diagnostics arrive on stderr as one JSON document per failing file (same shape as trust check --format json):

   {
     "version": "0.1",
     "file": "src/main.rs",
     "diagnostics": [
       {
         "rule": "R0042",
         "severity": "error",
         "message": "call to `make_rect` must use named arguments (arity 2)",
         "span": {"start": 50, "end": 62, "startLine": 2, "startColumn": 40,
                  "endLine": 2, "endColumn": 52},
         "why": "positional argument ordering is the largest LLM-authored bug class…",
         "help": "rewrite as `make_rect(width: ..., height: ...)`",
         "fix": {"span": {"...": "..."}, "replacement": "(width: ..., height: ...)",
                 "applicability": "hasPlaceholders"}
       }
     ]
   }
   

   Spans are byte offsets plus 1-based line/column. why, help, fix may be null.

2. Apply the fix. applicability: "automatic" fixes are safe to apply unattended; "maybeIncorrect" needs review; "hasPlaceholders" contains ... you must fill. Otherwise follow the help/instead guidance literally.

3. Rebuild. Repeat until clean (exit code is non-zero while any error remains).

• For a file full of R0042 violations, trust fix <file> --write mechanically inserts the name: prefixes (idempotent; only the names are spliced in). Pass - to read stdin.
• trust explain <CODE> explains any rule (why + what to write instead); with no code it prints the whole catalogue. --format json for machine consumption.
• Single-file lint without cargo: trust check foo.rs --format json. Note cargo trustc check resolves to whole-crate cargo check instead.

Command quick reference:

trust new <name>                            # scaffold a strict project
cargo trustc build|run|test                  # cargo lifecycle through the shims
cargo trustc build --message-format json     # machine-readable diagnostics on stderr
trust check <file.rs> [--format json]       # lower + lint one file, no rustc
trust fix <file.rs> --write                 # insert named args in place
trust explain [<CODE>] [--format json]      # rule catalogue / single rule
trust lower <file.rs>                       # print the lowered plain Rust
trust index <src-dir> --out sigs.txt        # signature index for a dependency

Rule table

"Test-exempt" = the rule does not fire inside #[cfg(test)] modules / #[test] fns. In project mode, files reachable only through a #[cfg(test)] mod are entirely exempt from all rules (RT-88).

Code	Name	Write this instead	Test-exempt
R0001	no-unwrap	propagate with ?, or .expect("why this can't fail")	yes
R0002	empty-expect	give .expect("…") a real message explaining why it can't fail	no
R0003	no-as-cast	use T::try_from(x)? for fallible casts, or .into() for widening	no
R0004	no-glob-import	import the specific items: use foo::{A, B};	no
R0005	justify-unsafe	precede the unsafe block with a // safety: … comment	no
R0006	justify-allow	precede the #[allow(…)] with a // reason: … comment	no
R0007	no-impl-trait-return	name the type with a type Alias = …; and return the alias	no
R0008	no-user-macros	inline the logic, or opt in with #[strict::macros_ok]	no
R0010	no-todo-macro	finish the implementation, or return a typed Err	yes
R0011	no-panic	return a typed Err and let the caller decide whether to abort	yes
R0012	no-bool-param	replace the bool with a named enum, e.g. enum Mode { On, Off } (private fns also exempt)	yes
R0014	no-bare-index	use .get(i) and handle the Option (slicing v[a..b] is exempt)	yes
R0015	allow-missing-reason	add a reason = "…" argument to the #[allow(trust::…)]	no
R0016	allow-unknown-code	use a real rule code (run trust explain for the catalogue)	no
R0017	no-same-type-params	wrap each in a distinct newtype — trust_std::newtype!(pub Width(u32));	yes
R0018	error-context-dropped	carry the source: `.map_err(	e
R0019	no-unchecked-len-arith	make the choice explicit: .checked_sub(1)?, .saturating_sub(1), or .wrapping_* if wrap is intended	yes
R0020	no-lock-across-await	drop the guard before awaiting (scope it in a block), or use an async-aware lock like tokio::sync::Mutex	no
R0021	no-capacity-as-len	use .len() for element counts; .capacity() only sizes future allocations	no
R0042	no-positional-args	name the arguments — f(width: …, height: …) — or run trust fix	no

All 20 rules are severity error. Codes outside R00xx come from lowering, not the lint runner: R2001 (pipe needs a path-call on the right), R3001 (no parameter with that name).

Pitfalls

• Tests are exempt via #[cfg(test)] — do not rewrite test code into dialect syntax; stock cargo test must still parse it.

• #[allow(trust::R0xxx, reason = "…")] works ONLY in code built exclusively via cargo trustc — stock rustc rejects tool-scoped allows, so published/stock-buildable crates must fix the violation, not suppress it.

• The reason = "…" is mandatory on every trust:: allow (R0015 fires without it); a // reason: … comment must also precede any #[allow(…)] at all (R0006).

• // safety: (R0005) and // reason: (R0006) comment contracts go in the contiguous comment block directly above the site — a blank line breaks the association:

  // safety: ptr is non-null; checked by the guard two lines up
  unsafe { ptr.read() }
  
  // reason: Slab key, not a length-bounded index
  #[allow(trust::R0014, reason = "Slab key, not usize")]
  fn get(&self, key: Key) -> &Entry { /* … */ }
  

• cargo trustc build --message-format json is Trust's flag, not cargo's — cargo's own --message-format takes different values; the wrapper strips the flag and sets TRUST_MESSAGE_FORMAT=json. Setting that env var directly is equivalent.

• R0012 has no #[allow] escape hatch at all: make the fn private, use an enum, or move it under #[cfg(test)].

• Cross-crate calls into unindexed dependencies are positional by design — don't invent names for std/third-party fns unless trust-std or a trust index manifest covers them.

Escalation

If a rule looks wrong for your case, read case-studies/ and the rule's section in docs/SPEC.md before suppressing — most "false positives" have a documented compliant idiom. If suppression is genuinely right, it requires a written reason = "…" (and // reason: comment), and it only survives in code that is never built with stock cargo.