Watching health

tsc --watch never exits, so the only way to know how it is doing is to read its log — and that costs tokens every single turn. A watcher reads it instead and reports transitions.

shell_start command="tsc --watch --noEmit" description="type checker" watch=true
→ watching health (tsc)

… later, on its own:
<shell_health id="sh_9wq2f1ab" title="type checker" status="fail">
tsc: ok → fail
src/auth.ts(42,3): error TS2339: Property 'id' does not exist on type 'User'.
</shell_health>

Nothing is sent while a run keeps producing the same result. A watched process that dies is reported as a failure, so a crashed dev server no longer goes unnoticed.

A rule is three regexes

Not a parser: done (a run ended), fail, ok, plus idleSeconds for tools that never announce the end of a run.

{ "done": "\\d+ (passed|failed)", "fail": "\\d+ failed", "ignoreCase": true }

Rules can be passed to shell_start as watch, or to shell_watch as rule.

Presets

About 35 ship and are picked automatically from the command: tsc, eslint, biome, prettier, mypy, ruff, vitest, jest, mocha, bun test, deno test, pytest, rspec, phpunit, playwright, cypress, vite, next, nuxt, astro, angular, webpack, esbuild, tsup, turbo, metro, storybook, cargo, go, dotnet, gradle, maven, docker compose and terraform.

Add your own in config and they are sent to the daemon as explicit rules, so your presets work without waiting for a release:

{
  "watch": {
    "presets": {
      "e2e": { "done": "\\d+ (passed|failed)", "fail": "\\d+ failed" }
    }
  }
}