openworkers-conformance

One suite of guest JavaScript, run unchanged on every OpenWorkers runtime, to measure what each backend actually implements.

A low score is a result, not a failure: the suite is written against the WinterTC (TC55) Minimum Common API and the specs it points at, never against what a runtime happens to do. Guest assertions never set the exit code, so the run stays usable while the runtimes are incomplete.

Layout

tests/ holds plain classic scripts. Each declares its tests with test('name', fn) at column 0, and asserts with the helpers the harness injects (assert, assertEqual, assertType, assertArrayEqual, assertThrows, assertRejects). Nothing in a test file is backend-specific.

area	file	tests
crypto	js/crypto/crypto.js	30
encoding	js/encoding/base64.js	20
encoding	js/encoding/text.js	35
globals	js/globals/abort.js	24
globals	js/globals/blob.js	26
globals	js/globals/microtasks.js	10
globals	js/globals/misc.js	16
globals	js/globals/structured-clone.js	17
globals	js/globals/timers.js	16
headers	js/headers/headers.js	28
request	js/request/request.js	29
response	js/response/response.js	38
streams	js/streams/streams.js	25
url	js/url/search-params.js	35
url	js/url/url.js	33
wintercg	wintercg/minimum-common-api.js	66
	total	448

Every file runs in a fresh worker, and every test runs under a 2s deadline, so one missing global costs its own line and nothing else. The denominator comes from the source, not from the run: a backend that cannot parse a file still scores 0/N against the same N as the others.

Running

cargo run --release --features v8 --bin conformance
cargo run --release --features boa --bin conformance -- --verbose
cargo run --release --features quickjs --bin conformance -- --filter url
cargo run --release --features jsc --bin conformance -- --json

One backend per build, like the rest of the workspace. --verbose lists every failing test, --json emits the whole run, --filter selects by path.

Scoreboard

v8 and nova measured 2026-09-13, the other three 2026-09-12, one commit per runtime, all against openworkers-core v0.15.0.

area	v8	nova	boa	quickjs	jsc
crypto (30)	30	29	10	21	23
encoding (55)	55	54	47	44	33
globals (109)	109	103	80	43	37
headers (28)	28	28	24	24	24
request (29)	29	29	20	16	19
response (38)	38	38	30	21	27
streams (25)	25	24	13	11	12
url (68)	68	68	66	66	66
wintercg (66)	66	56	29	24	27
total (448)	448	429	319	270	268
	100%	95%	71%	60%	59%

Branch names lag the code: every one of these builds against core v0.15.0.

backend	branch	commit
v8	main	6a898ab
nova	main	2857ade
boa	main	be42bf1
surface	openworkers-wintertc main	e7825c2
jsc	feat/core-0.14	2df2560
quickjs	feat/core-0.14	ed14afe
wasm	feat/core-0.14-wasmtime-bump	c97afa3

What the numbers say

No test fails on every backend any more, down from 77 in August. Every one of the 77 that left the set left because v8 took it from openworkers-wintertc: jsc, quickjs and boa score exactly what they scored in August. v8 passes all 448, so the set that measures what the backends share is now the set that measures what they lack.

nova scores 429, against 242 on its own interfaces in August. The surface took it to 368 with no op answered at all: thirteen of the nineteen modules ask nothing of their host. Five ops took it the rest of the way, along with timers, encodeInto and a subtle that computes. Streams 0 to 24 of 25, Request 22 to 29, Response 30 to 38, globals 18 to 103 of 109, crypto 9 to 29 of 30. Of the 19 it still fails, 7 ask for WebAssembly, which nova_vm does not have.

Two of the ops it cannot answer are held up by the same gap: textEncode and the compression codecs hand back bytes, and nova_vm gives an embedder no way to build a typed array. Everything else crosses as text, which is why URL is answered by a parser that replies in JSON and subtle.digest in hex.

The same run found five deviations in the surface, caught by nova's own test suite rather than by this one: a FormData body reached the wire as [object FormData], every method was uppercased where the standard normalizes six, a body-less read marked the body used, a Request built from another took its stream instead of a tee, and a Headers init accepted only an array.

jsc, quickjs and boa write every one of these interfaces themselves.

• v8 is the only backend with a complete crypto, 30/30, and shares a complete URL with nova, 68/68. Its subtle covers the six digests, AES-GCM round trips, raw HMAC and AES key import and export, and ECDSA key generation; nova has all of that but the ECDSA pair, at 29/30. jsc and quickjs stop after digest and HMAC; boa has no subtle.digest at all, which is most of the distance between its 10/30 and everyone else.
• jsc and quickjs miss URL.canParse and URL.parse and nothing else in url.js; boa takes url.js whole and loses its two points on URLSearchParams. All three back URL with the url crate.
• boa is 129 behind v8 (319 against 448) on the strength of boa_wintertc, while being the weakest on crypto by far.
• The DOM event core exists on v8 and nova, from openworkers-wintertc: Event, EventTarget, CustomEvent, ErrorEvent, MessageEvent, PromiseRejectionEvent, MessagePort and MessageChannel, with AbortSignal built on EventTarget. v8 and nova take js/globals/abort.js whole, 24/24, against 11/24 for boa, 2/24 for quickjs and 0/24 for jsc.
• Headers is complete on v8 and nova, 28/28, on the interface both take from openworkers-wintertc. The other three sort no iteration, reject neither an invalid name nor a value with a newline, and trim no whitespace around a value.
• Request and Response are complete on v8 and nova, 29/29 and 38/38, from openworkers-wintertc. The other three set no Content-Type from the body, so neither a string nor a URLSearchParams arrives with its type, and none enforces the status rules (a body with 204 or 304, a status below 200 or above 599).
• Streams stop at ReadableStream on three of the five: WritableStream, TransformStream, the queuing strategies, TextEncoderStream/ TextDecoderStream, pipeTo and ReadableStream.from are missing on jsc, quickjs and boa, and only v8 and nova make a stream async-iterable. v8 is the only backend with CompressionStream and DecompressionStream, which compress chunk by chunk over a host codec instead of buffering the body, and the codec is an op nova does not answer. The byte tier is on v8 and nova: ReadableByteStreamController, ReadableStreamBYOBReader and ReadableStreamBYOBRequest, with the buffer transfer the standard asks of read(view). The three other backends expose none of them, and asking one for a byob reader still returns a default reader whose read drops the view on the floor.
• structuredClone, Blob and File exist on v8, nova and boa, and multipart/form-data parses on v8 and nova.
• performance exists on v8, nova and quickjs, and quickjs alone lacks the Performance interface the standard names behind it; self exists on v8, nova and boa. Timers are on v8 and nova, 16/16 both, and on no one else.
• The WebAssembly streaming entry points are on v8 alone, and they check the response's status and media type before compiling: nova takes the module, but its engine has no WebAssembly for it to stand on.

Streaming battery

The guest suite asks what a script can observe about itself. It cannot see a body that does not arrive all at once, because only the host is on the other end of the channel. src/streaming/ is that second measurement, and STREAMING.md holds its matrix.

cargo run --release --features v8 --bin streaming
cargo run --release --features jsc --bin streaming -- --filter backpressure
cargo run --release --features wasm --bin streaming -- --json

25 probes on the JavaScript backends, 4 on wasm, each in its own thread with its own deadline, so a runtime that deadlocks mid-stream costs one line and gets reported as hangs rather than taking the run with it.

Backends

One cargo feature per backend, mutually exclusive, the same pattern as openworkers-task-executor.

wasm builds and reports a skip: that runtime executes wasm components, not JavaScript, so this suite has nothing to say about it. Reporting 0/448 there would be a lie, not a measurement.

nova builds with temporal off. nova_vm 1.0 with the feature pulls temporal_rs 0.1.2, written against icu_calendar 2.1, while v8 152 pulls temporal_capi 0.2.3+, which wants 2.2. One lockfile cannot hold both, and this crate needs v8.

SSR oracle

examples/ssr_oracle.rs and fixtures/sveltekit-app/ are a separate, complementary measurement: 17 scenarios of a real SvelteKit app, recorded on v8 and diffed byte for byte by the other runtimes.

cargo run --release --features v8 --example ssr_oracle -- --check