Skip to main content

​
Decision

tsrs is a Rust port of the TypeScript 7 compiler built with the TypeScript team and planned to become part of Oxc. It is experimental. tsc (TypeScript 7.0.2) stays the required type-check gate. tsrs runs in two advisory places only:
  1. The pre-push gate defaults to tsc. Explicit --checker tsrs comparisons are advisory and always confirmed by tsc, including a tsrs pass. Resolution requires a planned qualified project; hook mode accepts only an existing TSRS_BIN and never fetches packages. Clean-source parity alone does not qualify detection of seeded errors.
  2. .github/workflows/tsrs-shadow.yaml compares tsrs with tsc on main pushes that touch qualified projects and reports differences as warnings. The job is continue-on-error and is not a required check.
The pin is tsrs@0.18.0-ts7.1.0-dev.20260929 (version string 7.1.0-dev.20260930.4). It is installed outside the workspace (npm install --prefix in CI, the npm exec cache locally) so the shadow never changes bun.lock. Set TSRS_BIN to use a different binary locally.

​
Evaluation (2026-10-23)

Median of three cold runs per checker on a 10-core, 24 GB Apple Silicon machine, -p <project> --noEmit --incremental false --composite false, queued through ttr resources run. Workspace dist declarations were not built (builds are CI-only), so absolute diagnostic counts include missing-module noise; parity is still meaningful because every checker saw the same files. Timings were taken while other sessions shared the machine, so treat them as relative. “Exact” compares the full diagnostic multiset including message text, independent of order. “Location” compares file, line, column and code. Findings:
  • tsrs matches TypeScript 7.1, not 7.0. Every difference from 7.0.2 also appears with tsc 7.1.0-beta. packages/ui and apps/web gain implicit-any diagnostics under 7.1 semantics; apps/tasks and apps/calendar differ only in how long types are printed. Against 7.1-beta, tsrs has location parity on all eight projects and exact parity on seven; apps/web differs in two messages where tsrs prints a truncated property type instead of ....
  • Speed and memory. tsrs is 1.3-6.3x faster (larger projects gain more) with roughly 50-70% less peak memory.
  • TypeScript 7.1 beta is 1.1-1.3x faster than 7.0.2 on packages and apps/tasks, but slower on apps/calendar and apps/web in this run, with lower memory. It is a beta and changes diagnostics, so it is not adopted.
  • Configuration support. All tsconfig features we use type-check the same: package-name extends chains, paths, NodeNext/Bundler resolution, jsx preserve/react-jsx, types, lib including webworker, JSON modules and include globs. --build --dry resolves the same project-reference graph for apps/meet and apps/web. --showConfig is not supported. Emit (declarations, --build writes) was not evaluated because it is a build.
  • Exit status is 2 for diagnostics where tsc 7.0.2 returns 1; compare success versus failure, not exact codes.

​
Shadow job

node scripts/ci/tsrs-shadow.js runs both checkers per project, records wall time and peak RSS (/usr/bin/time), compares diagnostic multisets and writes a step summary, ::warning annotations (up to 20 per side) and a JSON report kept for 30 days. The default project list is the parity-qualified set: packages/types, packages/utils, packages/supabase, packages/internal-api, apps/tasks and apps/calendar. CI builds the workspace declarations first, exactly like the required type-check, so CI parity is the authoritative evidence. Pull requests run the job only when the shadow tooling changes; workflow_dispatch runs it on demand.
The local flag runs tsrs only for projects in TSRS_PROJECTS, even when one workspace has both qualified and unqualified configs. Tsc still checks every project. Exact diagnostic-key multiset differences are printed as advisory tsrs-only / tsc-only output, including when tsc passes; a shadow mismatch never changes the authoritative tsc result.

​
Promotion criteria

Do not replace tsc as the required gate until all of these hold:
  1. Version alignment. The pinned tsrs release tracks the same TypeScript minor as the repository gate (or the repository adopts the matching stable TypeScript minor), and is a non-dev release pinned exactly through the owning workspace’s package manager with a reviewed lockfile change.
  2. Parity over time. At least 30 consecutive shadow runs on main per promoted project show location parity, matching success/failure status and no tsrs crash, with warm declarations. Message-text differences are reviewed and recorded. Each promoted project must also have a seeded-error corpus that both checkers reject, covering real type errors, path/ambient declarations and reference boundaries. Until then, keep tsc confirmation on passes too.
  3. Feature coverage. Every project the required job checks is covered, including project references (apps/meet), worker/boundary configs and declaration-emitting builds, or an explicit tsc fallback is kept for each unsupported feature.
  4. Operational evidence. Linux CI wall time and memory are recorded for the promoted set, and rollback is a single-commit switch back to tsc.
Keep tsrs advisory, with tsc confirming every outcome, until all criteria hold. Then switch only the qualified scopes, preserving explicit tsc fallbacks and emit tasks. Update this page, the project list in scripts/ci/tsrs-shadow.js and the pre-push gate together.