soothfast

Annotate a function, get a checked benchmark. The output below is real, produced by this exact code on every CI run:

use soothfast::measured;

#[measured(group = "indicators", alloc = 0)]
pub fn rsi() -> f64 {
    let closes = [
        44.0, 44.25, 44.5, 43.75, 44.65, 45.1, 45.4, 45.8, 46.0, 45.6,
    ];
    let deltas: Vec<f64> = closes.windows(2).map(|w| w[1] - w[0]).collect();
    let gain: f64 = deltas.iter().filter(|d| **d > 0.0).sum();
    let loss: f64 = deltas.iter().filter(|d| **d < 0.0).map(|d| -d).sum();
    100.0 - 100.0 / (1.0 + gain / loss)
}

fn main() {
    println!("rsi={:.2}", rsi());
}

rsi=70.51

Docs and performance work the same way. Both are written from the code and checked against it on every build. Adding soothfast costs one runtime dependency (linkme); everything else lives in the separate cargo-soothfast binary.

How it works

cargo soothfast measure records perfcnt, callgrind, walltime and alloc counts into a baseline. cargo soothfast gate fails CI as soon as a fresh run regresses past a threshold. The <!-- soothfast:bind --> and <!-- soothfast:claim --> markers tie markdown prose to that same run, so a stale README fails CI instead of going unnoticed.

There is more than benchmarks here. Spec reconciliation, SDK generation, trend charts, changelog drafts and an agent-facing MCP server are all covered in the full guide.

Quick start

[dev-dependencies]
soothfast = { version = "0.1.0", features = ["runner"] }

[[bench]]
name = "soothfast"
harness = false

// benches/soothfast.rs
soothfast::bench_main!();

mod mylib {
    pub fn sorted(input: &[f64]) -> Vec<f64> {
        let mut v = input.to_vec();
        v.sort_by(|a, b| a.partial_cmp(b).unwrap());
        v
    }
}

#[soothfast::fixture]
fn values_n(n: usize) -> Vec<f64> {
    (0..n).map(|i| (n - i) as f64).collect()
}

// Checked claims: complexity verified by size sweep, alloc count exact.
#[soothfast::bench(group = "sort", setup_sized = values_n, sizes(1024, 4096, 16384),
                 complexity = "n log n", alloc = 2, covers = "mylib::sorted")]
fn bench_sorted(input: &[f64]) { soothfast::keep(mylib::sorted(soothfast::keep(input))); }
// Route constant inputs through soothfast::keep or LLVM const-folds them away.

$ cargo soothfast measure -p mylib --save-baseline base   # backends probe: perfcnt/callgrind/walltime/alloc
$ cargo soothfast gate -p mylib                           # vs baseline; exit 1 on regression
$ cargo soothfast gate -p mylib --against-ref origin/master --ratchet v1.0  # merge-base + ratchet
$ cargo soothfast docs gen-tests -p mylib                 # markdown blocks -> tests + capture examples
$ cargo soothfast docs capture -p mylib                   # run examples, splice real output into docs
$ cargo soothfast docs check -p mylib                     # binds + claims + generated-artifact manifest
$ cargo soothfast docs accept -p mylib                    # re-lock soothfast.lock after verifying prose
$ cargo soothfast spec check -p myserver                  # #[route] vs OpenAPI/GraphQL/MCP specs
$ cargo soothfast report render -p mylib                  # perf tables, trend SVGs, badges, llms.txt
$ cargo soothfast mcp -p mylib                            # agent-facing server on stdio

See the guide for gating thresholds, ratchets, the full bind/claim/capture syntax, and spec reconciliation.

<!-- soothfast:bind mylib::sorted -->
Prose describing `sorted`. CI fails if the code changes under it.
<!-- /soothfast:bind -->

<!-- soothfast:claim mylib::checksum.perfcnt.instructions < 25000 -->
Numbers in prose become gated facts.

Dogfood

CI runs soothfast on soothfast. Seven crates each carry a bench target measured into a fresh baseline on every build, and this README is checked against it. The sentences below are checked claims rather than comments:

Constant inputs and results route through soothfast::keep, the black_box equivalent. Without it LLVM const-folds the measured body away, which is the easiest way to measure nothing at all.

Registry fingerprinting (FNV-1a over the normalized token stream) never allocates.

Summary statistics, meaning median and MAD over a sample set, cost a handful of allocations per call. Eight is the gated ceiling.

Workspace