CLIs that respect your terminal.

Yes, with the honesty a 0.8 deserves. Bracket ships 4,408 CLIs in production today, including tools at ngrok, Fly.io and Railway, and the parser and codegen layers have been API-stable since 0.6. We're pre-1.0 only because the plugin manifest format may still get one more breaking revision before we freeze it. Everything else follows semver, every release runs against a 1,900-case conformance suite across all three languages, and we publish a migration note for any change that touches user code.

Because real teams already are. A typical infrastructure team ships a Rust core for performance, a Go agent for portability and a Node wrapper because that's what their users npm install. Without Bracket, that's three CLI codebases with three flag conventions, three help styles and three sets of completion bugs. Bracket gives each language a native, idiomatic API — derive macros in Rust, struct tags in Go, decorators in TypeScript — that all compile down to the same command model, the same help renderer and the same completion engine. One spec, three native binaries.

Plugins are resolved and linked at build time, never downloaded or executed at runtime — so there is no auto-update surface to exploit. Each plugin declares its capabilities (filesystem, network, environment) in a manifest, and bracket plugin add shows that capability set before it writes anything to your lockfile. A plugin that never declared network access cannot open a socket; the generated code simply doesn't expose the API. Plugin versions are pinned by content hash, and our advisory feed lets you fail a build if a linked plugin has a known CVE.

All argument parsing is generated at build time as a flat match table — there is no reflection, no schema construction and no allocation on the hot path. On the Rust and Go targets a no-op invocation returns in roughly 40ms median and stays under 200ms at p99 on commodity hardware. The Node target carries V8 startup, so it lands closer to 90ms median; still well inside budget. We treat any regression past the 200ms p99 line as a release blocker, and every CI run prints the cold-start histogram so a slow change can't merge unnoticed.

Test the three layers separately. Parsing is pure — call Cmd::parse_from(&["deploy","--region","ord"]) and assert on the resulting struct, no process needed. For command logic, write your run functions to take parsed input and return a value, so they're ordinary unit tests. For end-to-end behavior, Bracket ships a bracket::test harness that captures stdout, stderr and exit code, and supports golden-file snapshots of help text so a docs regression fails loudly. Always assert against --json output in CI — it's stable across color and terminal-width changes.