Strict Oxlint and Oxfmt Standard v5.3: Moving to tsrs and Cutting Agent Friction
Over the past few weeks I rolled out v5.2 of my strict Oxlint and Oxfmt standard across more than a dozen TypeScript repositories — monorepos, Cloudflare Workers apps, React dashboards, Node CLIs, and Pi extensions.
Strict type-aware linting works remarkably well as guardrails for coding agents. It catches floating promises, unsafe any leaks, dead conditionals, and subtle async bugs before code review ever starts.
It also exposed a handful of places where v5.2 created friction instead of quality. Today I am publishing v5.3 of the Strict Oxlint and Oxfmt Engineering Standard on FitchHub, along with a dedicated raw Markdown version built specifically so coding agents can read it without sifting through HTML.
Why I built an agent-first version
Link to Why I built an agent-first versionWhen you tell a coding agent to apply a standard to a repo, handing it an HTML webpage wastes context on navigation bars, styling markup, and page chrome.
Every published standard on FitchHub now lives in two synchronized forms generated from the same Markdown source:
- For humans:
https://fitchhub.com/standards/oxlint-oxfmt(rendered with table of contents and copyable code blocks) - For coding agents:
https://fitchhub.com/standards/oxlint-oxfmt.md(pure Markdown, zero HTML) - Pinned release:
https://fitchhub.com/standards/oxlint-oxfmt/v5.3.md - Automatic content negotiation: requesting
/standards/oxlint-oxfmtwithAccept: text/markdownreturns the raw Markdown directly, and/llms.txtlists every agent-readable endpoint.
If you want your coding agent to adopt or upgrade a repo right now, you can hand it something like this:
Read https://fitchhub.com/standards/oxlint-oxfmt.md and integrate it into this repository.
If this repo already adopted an earlier version of the standard, diff our existing config,
scripts, and exception records against v5.3 and apply the delta cleanly. If any mandated
rule or guidance causes real regressions in this repo, stop and raise it with me along
with the exact rule and evidence.
The standard itself is written as a standalone final specification for agents: it covers greenfield adoption, delta upgrades from older versions, and an explicit rule requiring the agent to raise concerns to you whenever a mandated rule actually causes regressions.
What changed in v5.3 (and why)
Link to What changed in v5.3 (and why)Every change in v5.3 came directly from auditing real overrides, directives, and failure modes across active TypeScript repositories.
1. Replacing oxlint-tsgolint with tsrs
Link to 1. Replacing oxlint-tsgolint with tsrsOxc recently introduced tsrs, a Rust port of TypeScript's type checker for Oxlint's type-aware rules. Installing oxlint@latest and tsrs@latest switches Oxlint automatically to tsrs with the same CLI flags and .oxlintrc.json config.
In our own matched local benchmarks on a 550+ file TypeScript repo, switching from oxlint-tsgolint to tsrs cut lint time nearly in half and dropped peak memory by more than 750 MiB — and Oxc reports up to 5–7× speedups and ~70–75% lower memory on larger codebases. v5.3 drops oxlint-tsgolint completely and makes tsrs the baseline backend.
v5.3 also leans on current Oxc CLI features that agents benefit from immediately: oxlint --format=agent for compact single-line diagnostics, --debug=timings, --print-config, and optional TypeScript configs (oxlint.config.ts / oxfmt.config.ts) when extending shared npm configs.
2. Inspect real installed code instead of stale memory
Link to 2. Inspect real installed code instead of stale memoryA surprising amount of agent lint friction comes from agents (and humans) operating on outdated assumptions about what TypeScript, Node, or Oxlint actually do today.
Two concrete examples we hit while verifying v5.3 against live oxlint@1.94.0, tsrs, and @types/node@24:
- Dynamic
import(specifier): You do not need anoxlint-disabledirective or wrapper. Writingconst mod: unknown = await import(specifier);satisfiestypescript/no-unsafe-assignmentcleanly out of the box, then forces proper narrowing before use. node:testregistrations: Inspecting@types/node/test.d.tsdirectly showed thatbefore,after,beforeEach, andafterEachreturnvoid(they never triggerno-floating-promises), whiledescribeis re-exported fromsuite. Including"suite"alongside["test", "it", "describe"]inallowForKnownSafeCallsfixes aliased imports likeimport { describe as nodeDescribe } from "node:test"automatically — and makesnode:testa standard test-profile runner mapping instead of a custom exception ceremony.
v5.3 promotes this into a core engineering principle: keep languages and dependencies current, read the actual installed source, .d.ts files, and current-date docs, and lean on native features before writing custom wrappers or shims.
3. No more stock-rule proof tests
Link to 3. No more stock-rule proof testsv5.2 required qualification fixtures proving that built-in rules like typescript/no-floating-promises, eqeqeq, or max-lines still fired. In practice, repositories ended up maintaining hundreds of lines of synthetic tests that merely re-tested Oxlint and TypeScript themselves.
v5.3 removes stock-rule proof suites. Verification now focuses strictly on what your repository owns: real acceptance gates (format:check, lint, typecheck, --report-unused-disable-directives), native coverage checks (--print-config / --debug=files), generated-artifact freshness, and tests for any custom plugins or scripts your repo actually writes.
4. Banning harmful compliance shims
Link to 4. Banning harmful compliance shimsWhen an agent hits a strict rule and tries too hard to make CI green, it sometimes invents code that makes the software worse: throwing synthetic TypeErrors on already-typed parameters, inserting fake await Promise.resolve() yields, or using Reflect.get / as unknown as to dodge type checks.
v5.3 bans compliance shims explicitly. If a mandated rule genuinely collides with a real contract or causes regressions, the right move is a narrow documented exception — or raising the concern directly to the person driving the session.
5. Calibrated defaults where v5.2 was too tight
Link to 5. Calibrated defaults where v5.2 was too tightAuditing real repos also pointed to a few thresholds that produced noise without catching real bugs:
max-lines(production): raised from500to800(tests remain exempt).max-params(production): raised from4to5, which eliminated 32 of 35 real override sites across audited repos (including 5-argument extension tool callbacks) while still catching 6+ argument parameter bags.unicorn/no-useless-undefined: setcheckArguments: falseandcheckArrowFunctionBody: falseso requiredT | undefinedarguments and typed() => T | undefinedcallbacks stop breaking, while uselesslet x = undefinedand default= undefinedremain caught.import/no-cycle: enabledignoreExternal: trueso cycles inside third-partynode_modulespackages do not fail a repository's own lint gate.no-param-reassign: documented the standard file-scopedignorePropertyModificationsForpattern (keepingprops: true) for Immer drafts, middleware, and SDK mutation hooks.
Try it
Link to Try itRead the human version at /standards/oxlint-oxfmt, or point your coding agent straight at /standards/oxlint-oxfmt.md.