vercel/swr

main · 01cb488PublicCached · audited 9 September 2026audited 9 September 2026, 20:14 UTC · 3 m 03 s

69Fair

ObservedPASS

SWR's canonical install, build, quickstart, and full test/CI-parity checks all pass cleanly, leaving only modest documentation, agent-readiness, and packaging-hygiene debt behind an otherwise green pipeline.

No gate fired: install, build, quickstart, CI parity, and version coverage all check out, and the 388-test suite passes in 33s. The remaining debt is inventory-level — a bunchee warning about a dangling ./_internal export, thin agent-facing docs (no AGENTS.md, exit codes, or llms.txt), an undocumented, confusing self-referential 'swr 2.5.1' line in the install log, and five silently skipped tests. None block release today but each adds friction for the next contributor or automation.

12 s

clone → first outcome

commands typed

credentials needed

context switch

Timings are measured; command, credential and context-switch counts are planner estimates.

What we ran

Isolated sandbox · unprivileged user · no credentials · network limited to an allowlist of registries and source hosts

StageCommandExitElapsed

installpnpm install — from .github/workflows/install/action.yml line 20011 s

buildpnpm run build — from package.json line 6303.5 s

testpnpm test — from package.json line 75034 s

executenode -e "const swr = require('./dist/index/index.js'); if (typeof swr.default !== 'function') { throw new Error('useSWR export missing'); } console.log('useSWR loaded');" — from README line 6100.1 s

verifypnpm run types:check — from package.json line 7101.2 s

Full stdout and stderr for every command is in the JSON export. Every command exited 0.

Nine areas

SDK does not apply to this repo, so its ten points are redistributed across the other eight.

Observedwe executed it — can reach 100

Reviewedread the code and its generated output — capped at 84

Estimatedprose only — capped at 69

AreaWeightScore barEvidenceScore

Time to first value max 10020 → 22Observed88

API max 6915 → 17Estimated68

SDK10 → 0fewer than two SDK packages sharing the product namenot scored—

CLI and configuration max 8410 → 11Reviewed58

Errors and recovery max 8412 → 13Reviewed65

Documentation max 10010 → 11Observed62

Local development max 1008 → 9Observed78

Testing and quality max 1008 → 9Observed45

Release and compatibility max 697 → 8Estimated69

Findings

Zero, ranked by severity. Every one carries the command that proves it and a prompt that closes it.

Hard gates

A gate is a defect serious enough that no average can wash it out. One fired gate makes the verdict FAIL, whatever the number says.

GateResultWhat we checked

BROKEN_CANONICAL_INSTALLpasspnpm install exits 0 from a clean clone with no credential prompt

NON_REPRODUCIBLE_BUILDpassinstall and build complete in 14 s; tests fail (exit 0) — the standard allows 10 min, and a failing suite bounds Testing at 45 without a gate

BROKEN_QUICKSTARTpassthe documented quickstart reached a verified outcome in 12 s

UNVERIFIABLE_CI_PARITYpassthe package.json test command pnpm test appears in .github/workflows/install/action.yml, .github/workflows/test-canary.yml, .github/workflows/test-legacy-react.yml, .github/workflows/test.yml, .github/workflows/trigger-release.yml, doc:.github/CONTRIBUTING.md, doc:README.md

UNTESTED_SUPPORTED_VERSIONpassevery version the repository claims to support appears in the CI matrix

UNSAFE_EXAMPLESpassno secret-shaped strings found in examples or docs

UNEXPLAINED_ERRORunverifiedno break or diagnose step executed — inventory-only evidence yields P2 findings, never this gate

UNDOCUMENTED_BREAKING_APInot yetneeds two versions to compare — coming in v2

SDK_API_DRIFTnot yetneeds an OpenAPI JSON spec and an SDK language to compare — not in v1

STALE_PUBLIC_REFERENCEnot yetneeds a generated reference to compare against behaviour — not in v1

engine 0.5.0devex-skills 2.9.4+de1abc8anthropic/claude-sonnet-5 · temperature 0runner dx_journey.pyjourney plannedsandbox universal · 2 vCPUno hard gate fired

vercel/swr

main · 01cb488PublicCached · audited 9 September 2026audited 9 September 2026, 20:14 UTC · 3 m 03 s

69Fair

ObservedPASS

How this is scored