Railgun — state-of-the-art native TypeScript validation

The fastest TypeScript toolchain that exists. Period. One process. One traversal. No spray.

Go-native TypeScript 7 tsc (8–12x vs TS6) + tsgolint (20–40x vs ESLint + typescript-eslint). Full type-aware lint + typecheck of a 776-file workspace in ~6 seconds — while you are still reading this line.

Q3A scoreboard — what green looks like

PERFECT      0 errors | 1 files | 28 LoC        | 2ms | cache: hot
HEADSHOT     1 errors | 42 files | 2100 LoC     | 18ms | cache: hot
IMPRESSIVE   3 errors | 39 files | 11300 LoC    | 47ms | cache: hot
EXCELLENT   20 errors | 66 files | 29599 LoC    | 127ms | cache: hot

0 errors = PERFECT. 1 error = HEADSHOT. 20 errors < 300 ms = EXCELLENT. The instrument panel is the weapon: score line, LoC, milliseconds, cache state — one block, machine-parsable with --agent.

What ships in this package

One package, bodies live in bin/<os>-<cpu>/:

| file | role | | --- | --- | | bin/railgun.js | Node launcher + railgunize workspace migrator | | bin/win32-x64/railgun / railgun-ts | PGO'd Oxlint + Go-native TS7 backend | | bin/linux-x64/railgun / railgun-ts | same, musl x64 | | bin/linux-arm64/railgun / railgun-ts | same, arm64 | | bin/darwin-arm64/railgun-ts | TS7 backend (Oxlint body from CI macos-14 lane) | | skills/SKILL.md | the validation ladder for AI agents |

Why it wins

• One traversal replaces the whole ladder: npx tsc, eslint, oxlint, next lint, per-package loops → one railgun command, type-aware included.
• Persistent named-pipe daemon (\\.\pipe\railgun-<id>): the Go TS program, project index and dirty set stay warm; 300 s idle, cold fallback so the hot loop never pays two timeouts.
• Migration in one shot: npx @cyberluke/railgun init patches 48 manifests of a pnpm+Turbo+Next workspace, merges typegen/check/lint/typecheck into turbo.json, appends the managed AGENTS.md block, drops a customized SKILL.md into .claude/.roo/.kilo, stores baseline + LoC fingerprints in .railgun/. Idempotent: second init changes nothing but timestamps.
• Baseline-aware: railgun baseline capture + status separate your new errors from inherited debt — agents ship on new errors: 0.
• LoC as a first-class metric: ±added/removed per run, staged/unstaged/untracked split. No mystery budgets.

The map

| instead of | use | | --- | --- | | npx tsc --noEmit | railgun check . --type-aware --type-check | | npx tsc -p … / pnpm tsc | railgun typecheck | | npx eslint . / pnpm lint / next lint | railgun lint | | npx oxlint | railgun lint | | per-package check loops | railgun workspace check --affected |

Modifiers: -p|--project, --noEmit, --pretty, --max-diagnostics, --summary-only, --changed, --no-daemon. Machine formats: --agent, --json, --jsonl, --timings. Launcher defaults (injected when not given): --quiet and --agent (-f agent). Explicit flags always win; --json / --jsonl suppress the --agent injection.

Hierarchy: next typegen route types > Railgun native checks > next build final gate. Native TS7 tsc remains the merge gate; the daemon belongs to the agent hot loop.

Reference run: 59 errors | 120 warnings | 776 files | 212572 LoC | 6125ms — cold, type-aware, --no-daemon. Every number on the panel is real.