Repository navigation
v0.14.0
evnx v0.14.0
Read this first: what changes for scripts and CI. Each item has its own
entry below with the details and how to migrate.
- CI needs
EVNX_KEY_BUNDLEfor the cloud vault. Create it with
evnx auth token createorevnx auth key-bundle; it is required once the
server stops serving key material to tokens. Set it besideEVNX_TOKENand
EVNX_MASTER_PASSWORD. - A vault id typed in upper case is stored in lower case. Versions pushed
with an upper-case id open only with the same upper-case id, which only
0.13.0 and earlier send: pull them with 0.13.0 and push again. migrate --to vercelprints the environment of the file it read
(development,previeworproduction); it printedproductionfor
every file. If you ran it for a development or preview file, check Vercel
production for development values.- Vercel secrets are
sensitiveindevelopmenttoo, inconvert --to vercelandmigrate --to vercel. - Backups no longer pile up:
<file>.bakis replaced on each write, and
old<file>.bak.<digits>copies are removed on the next write. convertoutput shapes changed:vercelanddopplerprint the
shapes those tools read,terraformkeeps the case of names (it lowercased
them), andherokuis oneheroku config:setcall.- More
scanfindings: names are checked in Dockerfiles, shell scripts
and Python modules; notebooks are read cell by cell;ghs_tokens in the
new form are found; a secret under a public prefix (NEXT_PUBLIC_,VITE_,
...) is one severity level higher. A--severity highgate may start
failing.validatehas a new error,public_secret, and a new warning,
partial_presence, for projects with several environment files. @evnx/clihas nopostinstall:bin/evnxis a Node launcher.- Each release carries a CycloneDX SBOM,
evnx-<version>.cdx.jsonand its
.sha256.
Added
-
evnx auth resend-verification(N14-5). Sends the verification email again
for the signed-in account. It asks the server who you are first, so an
already-verified account is told so and nothing is sent (the server's resend
endpoint answers 202 for everything, by design). Rate limited (3 per hour per
address) it exits 15 and says how long to wait; not signed in exits 10; an
API token is refused. The "email not verified" error andauth statusnow
point at this command. -
Each release carries a CycloneDX SBOM (SEC-2). A new
sbomjob in
release.ymlrunscargo-cyclonedx(0.5.10, installed with--locked) with
all features on and attachesevnx-<version>.cdx.jsonand its.sha256to the
GitHub release next to the archives. The SBOM gets a build provenance
attestation like the archives;gh attestation verifyworks on it. The
README's verify section says so. -
CI:
EVNX_KEY_BUNDLEreplaces fetching key material with an API token.
Create it withevnx auth token createorevnx auth key-bundle; it is
required once the server stops serving key material to tokens. CI now sets
EVNX_TOKEN,EVNX_MASTER_PASSWORDandEVNX_KEY_BUNDLE(or
EVNX_KEY_BUNDLE_FILE, first line of a file). The bundle (evnxkb1_…) is
your sealed keypair and Argon2 salt, with no plaintext secret in it; the
master password still opens it.token createprints it with the token
(--no-key-bundleskips it;-qprints the token only). Without a bundle a
token still asks an older server once; a server that refuses gets exit 11 and
a message saying how to make one. A bundle made before a master-password
change says to regenerate it.auth statusunder a token shows the token's
own info instead of "refused" when the server will not describe the account,
and the post-quantum key backfill at login is skipped under a token. The
high-water warning now also names an admin deleting the latest version. -
evnx platform diff|pull --to vercel: compare and pull a Vercel project's
environment variables (N14-10, phase 1;pushis below,--prunecomes later).
diffprints a plan (+ add,~ update,= unchanged,! remote only,
! type differs) for a local.env.<env>against a Vercel environment, with
--format json; exit 0 with no differences, 1 with some.pullwrites
Vercel's Config values into the local file, keeping comments, order and
formatting. A Secret is write-only on Vercel: one whose key has a local
value keeps it, one the file lacks is written asKEY=and listed, anddiff
can say only that it exists. Local-only keys are kept.pullshows the plan
and asks first (--yeswithout a terminal, exit 4 on a no,--dry-run), and
keeps the old file as.bak. Values are never printed. The project comes
from--project/--team, thenVERCEL_PROJECT_ID/VERCEL_ORG_ID, then
.vercel/project.json; the token fromVERCEL_TOKENor--token-file(there
is no--tokenflag).--env-name production|preview|development(default
development) also picks the file (.env.<env>if present, else.env).
Exit codes followcloud: 10 token refused, 11 forbidden, 13 unreachable, 15
Vercel error or a rate limit that did not clear after three tries. A pull
records.evnx/platform/vercel.json(each variable's id, targets, type and
updatedAt; no values, no hashes) for the laterpush; add.evnx/platform/
to.gitignore. Needs thenetfeature; a build without it listsplatform
in the missing-features table. -
evnx platform push --to vercel: create and update Vercel variables from the
local file (N14-10, phase 2). It prints the plan (+ create,~ update,
= unchanged,! type differs,! remote only; never a value) and asks first
(--yeswithout a terminal, exit 4 on a no,--dry-runsends no request that
carries a value). A new key is created in one bulkPOSTwith
upsert=false: Vercel's upsert replaces a variable's whole target list, so it
is never used. A Config whose value differs is changed by id withPATCH;
a Secret that exists on both sides cannot be compared, so it counts as an
update (--only-newskips those). A variable shared with other environments
leaves this environment first (PATCHits targets) and gets its own entry of
the same type, so the others are untouched. The type of a new key is the
classifier's; a public prefix (NEXT_PUBLIC_,VITE_, …) is always Config;
--type config|secretoverrides it for new keys only. An existing variable
keeps its type. Keys only Vercel has are kept (--pruneis a later phase). A
key with an empty value is never pushed, sopullfollowed bypushcannot
blank a Secret.failed[]is read on every create: any entry is exit 1 naming
the keys and Vercel's error code, and the rest stays applied. A 429 waits for
the reset and gives up after three tries (exit 15). Before it writes,push
compares each variable it will change with.evnx/platform/vercel.json; if
Vercel changed it since the lastpullorpush, nothing is written and the
exit code is 14 (no state file: a warning, and it goes ahead). A push that
wrote anything refreshes the state file.--format jsonprints one document.
Response bodies are never printed (create and delete answers echo ciphertext). -
evnx platform push --prune, custom environments and--git-branch(N14-10,
phase 3).--prunelists the variables only Vercel has in this environment as
deletions and deletes them by id after the confirmation (--yesis required
without a terminal, checked before anything is sent;--dry-runsends nothing).
A remote-only variable shared with other environments is not deleted: this
environment is taken out of its targets instead (shown asdetach). A key
that is in the local file with an empty value (KEY=) is never pruned, and the
delete answer, which echoes ciphertext, is never printed.--env-name <slug>
now accepts a custom environment: the slug is looked up with
GET /v9/projects/<id>/custom-environments, its variables are the entries whose
customEnvironmentIdsinclude it, a new one is created with
customEnvironmentIdsand notarget, and the local file is.env.<slug>else
.env.--git-branch <branch>(only with--env-name preview, else exit 2)
selects that branch's preview entries, falls back to a key's all-branches entry
for reading (labelled in the plan, never changed or deleted by the run; a
differing value creates a branch entry beside it), and creates withgitBranch
set.diffandpulltake the same two flags. The state file keeps a separate
record for a branch run (preview@<branch>) and for a custom environment. -
SSL_CERT_FILEandSSL_CERT_DIRadd trusted root certificates, so evnx
works behind TLS-inspecting corporate proxies (RM-68). The bundled roots stay;
the extras are added to them.SSL_CERT_FILEis a PEM bundle;SSL_CERT_DIR
is a directory (or a path-separator list of them) of*.pem,*.crtand
hashed (a1b2c3d4.0) files. A missing path or a file with no certificate is an
error naming the variable and the path (exit 13 forevnx cloud); an
unreadable file inside the directory is a warning. An unknown-issuer failure
now says: set SSL_CERT_FILE to your organisation's CA bundle (PEM). Applies to
cloud,update --check,migrate --to github-actionsandscan packs. -
validateandspec checkreport a name some environment files have and
others lack (N14-12). Apartial_presencewarning ("EXTRA is in .env,
.env.production but not in .env.staging"), until now visible only through
sync --checkanddiff. A warning, so the exit code is unchanged unless
--fail-on warning; silence it with--ignore partial_presenceor
[validate] ignore. It needs two environment files (.envand each
.env.<NAME>; the template and.bakcopies do not count), so a project
with one sees no new finding and--env -/--input-format jsonare never
compared. With a contract, a file is expected to lack a name declared
required = falseor limited byenvironments, and a namemissing_variable
already reports for the checked file is not repeated.spec checkprints the
same finding as a warning on stderr and keeps its verdict.
For CI: a project with several environment files that differ on purpose
gets new warnings invalidate(summary.warnings, and the evnx-action
validate-findingscount if it counts warnings); add the type to the ignore
list to keep the old count. -
evnx remove --check NAME…: is it gone everywhere? (N14-11). Read-only
and made for CI after a removal merges: exit 0 when no named variable (or
quoted glob) is in any environment file, the template or the[vars]
contract; exit 1 listing each file (and line) where one still is; exit 2,
no verdict, when it cannot read a file whole. It always looks everywhere, so
it takes none of the scope flags (--template,--contract,--all,
--env-name) nor--dry-run,--yes,--ignore-missingor--no-backup.
--jsonprintsschema_version,check,clean,names,checkedand
present(name,file,kind, andlinesortable); no value appears. -
evnx sync --prune: remove stale names from the template (N14-11).
synconly ever added; a name that every environment had dropped stayed in
.env.exampleuntil someone edited it by hand.--prunecompares the
template (--example,[defaults] example, else.env.example) with.env
and every.env.<NAME>, lists the names none of them has, and removes them
after asking. It usesremove's editor and writer, so comments, line
endings and everything else in the template stay byte for byte.- A name declared in
.evnx.toml [vars]is never pruned: it is listed as
kept, with theevnx remove NAME --template --contractthat drops both. --dry-runpreviews and writes nothing;--yes(-y, only with
--prune) skips the question; no terminal without--yesis exit 2;
answering no is exit 4. Only the template is written, and no.bakis
kept for it (it is committed, as withremove --template).- Refused, with nothing changed (exit 2): no environment file, every
environment file empty, or one that has a line evnx cannot read (a name
could hide behind it). - It is a mode of its own: clap refuses
--prunewith--check,--format,
--env,--env-name,--direction,--force,--placeholder,
--skip-without-env,--template-configand--naming-policy.
- A name declared in
-
A precision and recall gate for
scan, run bycargo test(RM-70).
tests/scan_corpus/is a labelled corpus of 60 files: 37 whose secrets must be
found, each labelled with its rule id and line, and 23 look-alikes that must
not be (documentation keys, near-miss tokens, hashes, UUIDs, lockfile
digests, base64 images, notebook outputs, placeholders, public keys, config
references).tests/scan_quality.rsruns the realevnx scanover it and
prints precision and recall overall, per family (cloud, vcs, payment, chat,
keys, jwt, ai, saas, database, generic) and per file format. A drop in recall
anywhere fails the build; so does a precision drop of more than 0.02. The
committed numbers aretests/scan_corpus/baseline.json; move them on purpose
withEVNX_UPDATE_SCAN_BASELINE=1 cargo test --test scan_quality -- --nocapture.
No secret-shaped value is committed: the files are templates filled at test
time from parts, and a test fails if one holds a value that matches a rule.
In CI the table is also written to the job summary. -
init'sschema.jsonsays howscansees each secret variable (RM-91).
A newscansection insrc/assets/schema.jsonhas an entry for every
variable the catalogue declares thatscanwould treat as a credential, or
that sits behind a browser-exposed prefix: the id of the rule that reports its
value (stripe-live-secret-key,github-token, ...),heuristicwhen only
the generic name and entropy checks would, orpublicwhen a prefix from
PUBLIC_PREFIXES(NEXT_PUBLIC_,VITE_, ...) puts it in the client bundle.
Additive: the rest of the file is unchanged.schema::loader::scan_view(name)
reads it.tests/schema_scan_view.rsholds the section,rules.tomland the
prefix list in step, and checks each claim against the scanner itself, so a
renamed or removed rule, a dropped variable or a new prefix breaks a test.
It needs no target-format catalogue (RM-61): the rule ids come from the rule
definitions. -
tests/invariants.rs: evnx accepts what evnx generates (F14-8, OW-17).
Chains commands and feeds each one's output to the next: everyinit
component and blueprint mustvalidate(placeholders are the only error),
doctor(no error) andscan(no finding) and parse;convertof a.env
holding quotes,$,#, spaces,=, unicode, an empty value, a multi-line
value and the words YAML reads as something else must read back to the same
keys and values in every format (JSON, YAML, TOML and base64 parsed by their
own readers, the script formats run undershagainst stand-ins for
gcloud,azandheroku); everymigrate --print-onlyscript passes
sh -n; andsync,addandremoveleave files that parse, withdiff
of a file against itself clean. Found one bug, fixed below. -
Known-answer tests for every primitive and for evnx's own envelope (SEC-1,
tests/crypto_kat.rs). The official vectors run against the crate versions in
Cargo.lock: BLAKE3 hash, keyed hash and derive_key (the BLAKE3 team's
test_vectors.json, vendored unchanged), HKDF-SHA256 (RFC 5869), X25519
(RFC 7748), Ed25519 (RFC 8032), ChaCha20-Poly1305 (RFC 8439),
XChaCha20-Poly1305 (draft-arciszewski-xchacha), AES-256-GCM (McGrew and
Viega, Test Case 16) and Argon2id (RFC 9106). Golden vectors generated once
fromevnx-crypto0.2.1 pin evnx's wire format: master-key derivation, the
vault key wrapped under the master key, the blob (nonce || AES-256-GCM
bound byvault_aad) and its BLAKE3 hash, an encrypted private key and the
three public keys derived from it, and a hybrid X25519 plus ML-KEM-768 wrapped
vault key. Changing any of them is a breaking protocol change. The tests live
here becauseevnx-cryptois a separate crate evnx consumes from crates.io.
Adds four dev-dependencies (chacha20poly1305,x25519-dalek,hkdf,
ed25519-dalek) that are already in the build graph. -
Local fuzz targets for the
.envparser, the scan engine and decryption
(SEC-1,fuzz/). A cargo-fuzz project of its own, outside every build and
without a CI job:env_parser(everyParserConfig),scan_engine(one
in-memory file read as.env, text, JSON, YAML, TOML or a Jupyter notebook)
andcrypto_decrypt(evnx-cryptoon arbitrary ciphertext, wrapped keys and
encodings, which must always return an error).fuzz/README.mdsays how to
run them;fuzz/seeds/holds small starting inputs. The directory is excluded
from the published crate.
Changed
-
Repo files: CODEOWNERS, Dependabot grouping for cargo, a PR template
(G5 to G7)..github/CODEOWNERSassigns the workflows, scripts, build script,
manifests,.cargo/, Dockerfile,pyproject.tomlandsrc/assets/to the
owner. The cargo entry independabot.ymlgroups minor and patch updates
(open-pull-requests-limitstays 0, as documented there). The PR template asks
what changed, how it was tested, and whether it touches workflows, scripts,
build.rsor dependencies. -
SECURITY.mdstates the support policy (RM-79). The latest minor gets fixes;
the previous minor gets security fixes for 90 days after the next minor ships;
older versions get none. It also states the minimum supported Rust version
(1.87, therust-versioninCargo.toml, checked by the CIMSRVjob) and
how it may change, and links to the security notices at evnx.dev/security. -
scanreads files in parallel (RM-74). Files are scanned on all cores and
reported in path order, so text, JSON, SARIF and GitHub output, the exit code
and the notes on stderr are the same as before and the same at any thread
count.RAYON_NUM_THREADS=Ncaps the threads (1is the old behaviour).
Measured on a 4,500-file, 51 MB tree on two cores: 9.4 s to 4.9 s with the
built-in rules, 10.9 s to 6.2 s with the gitleaks pack.-vlines and
evnx:allowwarnings are now printed together after the scan rather than as
each file is read.
Fixed
-
convertoutput shapes checked against each tool's documented input
(RM-65, I-092, I-093, I-097).vercelprints the array of
{key, value, type, target}that Vercel's create-environment-variables API
takes, withtypesensitivefor a secret-looking name or value (the same
classifier asplatform).dopplerprints the flat JSON object
doppler secrets uploadreads, not{value, computed}per name.terraform
keeps the case of each name (it lowercased them) and lets--transform
change it;A_Banda_bno longer collide. It still writes.tfvars
values; the new--terraform-declarationswrites a.tfofvariableblocks
instead, withsensitive = trueand no default for a secret-looking
variable, so no secret lands in a file people commit.herokusets
every variable in one shell-quotedheroku config:set A=… B=…call, so the
app restarts once.dopplerandvercelare no longer marked experimental;
railwaystill is, because Railway documents no schema for its JSON import. -
scan -vsays names are checked in Dockerfiles, Containerfiles, shell
scripts and Python modules. They printread as text: value patterns, and names in assignments; other text files keepvalue patterns only, names not checked. The JSONread_asstaystext. -
scan: false positives found by the RM-70 corpus (SCAN-FP).- A PEM or OpenSSH private key is one
private-keyfinding; its base64 body
lines are no longer alsohigh-entropy-string. - The body of a
PUBLIC KEY,RSA PUBLIC KEYorCERTIFICATEblock, and
the blob of anssh-rsaorssh-ed25519public key line, are not
reported: public material is not a secret. CACHE_KEY,ETAG_KEYandIDEMPOTENCY_KEYholding a 32, 40 or 64 digit
hex digest are identifiers, not credentials. Any other word in the name
(CACHE_API_KEY,API_KEY,HASH_KEY) or any other value keeps the check.- A Stripe publishable key (
pk_live_,pk_test_) is public by design and
is not reported by name or by entropy. A secret key underNEXT_PUBLIC_
is still reported. - YAML, JSON and TOML references are not values: a whole-value template
({{ .Values.apiToken }},${{ secrets.TOKEN }}), anything under a
KubernetessecretKeyReforsecretRef, and a lowercase secret name in a
secrets:list (- db_password). - Names ending in
_PASS(RABBITMQ_DEFAULT_PASS,DB_PASS) are password
names, forscanand for masking.BYPASS_CACHEandCOMPASSare not.
- A PEM or OpenSSH private key is one
-
scan: names are checked in Dockerfiles, shell scripts and Python
modules (SCAN-NAMES). A file read as text was judged by value alone, so
ENV DB_PASSWORD=<random>(alsoARG, theENV KEY valueform and
continuation lines),export DB_PASSWORD="<random>"orDB_PASSWORD=<random>
in a.sh,.bash,.zshor#!-script, andDB_PASSWORD = "<random>"in
a.pyfile scanned clean. Those assignments now get the name check, the
placeholder, empty-value and reference filters a.envline gets, and are
reported assensitive-config-keyon the right line. Never flagged:ARG X
with no value,$VAR,${VAR},$(cmd), backticks,os.environ[...],
os.getenv(...),settings.X, f-strings, empty strings and placeholders.
Commented-out assignments are left to the value rules. -
scan: thegithub-tokenrule matches stateless GitHub App installation
tokens (SCAN-GHS). Since GitHub finished its rollout on 2026-10-02,ghs_
tokens, the ActionsGITHUB_TOKENamong them, areghs_<appid>_<JWT>(about
520 characters) and the classic 36 to 40 character shape missed them. The
rule now has a second branch for that form and keeps the classic one. It is
reported once, asgithub-token, not also asjwtorhigh-entropy-string;
a trailing.,,or)is not part of the match. -
scanno longer flags the Neon sample URLinit --with neonwrites
(F14-8).postgresql://user:password@ep-example-123456.<region>.aws.neon.tech/...
was reported as a live connection-string credential, so a new project did not
scan clean. A host with anexamplelabel is now a placeholder host, as
localhostandexample.comalready were;user:passwordis still required,
so a real host with those credentials, or a real password on an example host,
stays a finding. -
Monorepos: paths in
.evnx.tomlare read from the folder that holds it, and
doctornotices nested.envfiles (F14-10, I-333, I-331).- A relative path in
.evnx.tomlmeans "from the folder holding this file",
wherever evnx runs below it. That was already so for[defaults] example
and[cloud] file;[scan] excludepatterns with a/now are too, so
apps/web/fixtures/**excludes that folder from a scan run inapps/web.
Patterns as written still match (a union), and a pattern with no/
matches anywhere as before. Flags are relative to where you run. doctorlists env files in subfolders (apps/web/.env, up to five, then
"and N more") as info withevnx doctor apps/webas the hint; it checks
only the folder you gave.node_modules,target,.gitand other
dependency, build and hidden folders are skipped, asinitskips them.
initalready named a nested.envand touches none.
- A relative path in
-
scanreads Jupyter notebooks cell by cell (F14-6, OW-4). A notebook is
JSON, so a key in a cell or an output sat behind string escapes and was
missed. Each cell'ssource(code and markdown) and each output'stext,
data["text/plain"]andtracebackare now read unescaped, and a finding is
located by cell:nb.ipynb:2 (cells[3].outputs[0].text), the line being
within that cell or output. Image output is not read. A notebook that does not
parse is scanned as text, as before. The baseline fingerprint of a notebook
finding now includes its cell path. -
platform pull: no more "already up to date" for a file that does not exist.
When Vercel has nothing for the environment and the local file is absent, pull
saysnothing to pull; <file> not created(exit 0). An existing file with
nothing to change still reads "is already up to date". -
platform push: the plan says when it splits a shared entry. A row now reads
~ KEY secret update (split: Preview, Production keep their own entry; …), the
summary countsN split (shared entry left), and--format jsonhas
summary.splitand, per entry,splitandkeeps_shared_entry. -
platform: the state file keeps one record per environment (PLAT-STATE).
.evnx/platform/vercel.jsonheld a single environment's record, so pulling
production after development made a developmentpushwarn "no state" instead
of stopping (exit 14) on a change made on Vercel. It now holds a record per
project, team and environment, and a push checks only its own. The old
single-record file is still read, and upgraded by the nextpullorpush. -
Backups no longer pile up. Every command that keeps the replaced file as
<file>.bak(platform pull,cloud pull,restore,diff --apply,
validate --fix,remove) now keeps exactly one:<file>.bakis replaced on
each write instead of becoming.bak.2,.bak.3, … Old numbered copies of
that same file (<file>.bak.<digits>, nothing else:.bak.oldand
notes.bak.txtare never touched) are removed on the next write, with one line
saying how many..gitignoregets a single*.bakline, once, under the same
--gitignore/--no-gitignore/--yesrules, instead of a question per copy;
nothing is staged or committed. Because the backup is replaced, a second edit
in a row leaves a copy of the first edit's result, not of the original. -
platform diff: a Secret present on both sides is no longer a difference
(F1). Vercel never returns a Secret's value, so such a variable madediffexit
1 forever and the command was useless as a CI gate. It now prints as
= secret present (write-only, not compared)and counts as unchanged for the
exit code.diff --strictrestores the old reading (an update, exit 1). In
--format jsonthe entry is"action": "unchanged"with"compare": "write_only", the summary gainspresent, and the document gainsstrict; no
existing field was removed. -
platform: a 403 that saysinvalidTokenexits 10, not 11 (F2). The
message is "the token is invalid or expired; create one at
vercel.com/account/tokens" instead of a hint about the project and team. A 403
without that flag stays 11. -
platform diffwith no.env.<env>and no.envcompares against an empty
local side (F3): every Vercel variable is "remote only", and a note says so,
instead of exit 2. A--filethat does not exist is still an error. -
platform pullandpushoffer to add.evnx/platform/to.gitignore
(F4), the same yes/no question the env files get, on a terminal without
--yes.--gitignoreadds it without asking; otherwise the text hint stays. -
evnx spec initwrote[vars.spring.datasource.url]for a variable with dots in its name.
TOML reads that as nested tables, soevnx validatethen reportedMissing required variable: springand unknown keys. Names that are not TOML bare keys (anything butA-Za-z0-9_-) are
now quoted, as[vars."spring.datasource.url"], wherever evnx writes or prints a
[vars.NAME]table (spec init,add,remove --contractmessages and thevalidate
hints). Bare names such asmy-varand9LIVESstay unquoted. A.evnx.tomlalready written
with the unquoted form must be edited by hand: quote the name. -
evnx migrate --to vercelprintedproductionfor a development file. With
--env-name development(orpreview) it read.env.developmentbut every printed
vercel env addsaidproduction, and the dry-run saidEnvironment : production; every
earlier version did this. The environment now follows the file that was read
(.env.development/.env.devisdevelopment,.env.previewispreview, anything else,
including plain.env, isproduction). If you ran it for a development or preview
file, check Vercel production for development values. -
evnx add --helpsaid a declined confirmation exits 0; it exits 4, as
0.13.0 made it (the help text was not updated). It now lists 0, 2 and 4. -
Vercel secrets are now
sensitiveindevelopmenttoo (convert --to vercel,
migrate --to vercel). 0.13.0 kept them out ofdevelopmentbecause Vercel rejected a
Sensitive variable there; Vercel reversed that rule on 2026-08-24 ("Secrets are allowed in
Development", docs,
changelog).
convert --to vercelnow gives a secret-looking variable all three targets (production,
preview,development) and no longer prints the "not set for development" warning;
migrate --to verceladds--sensitivefordevelopmentas for the other environments.
The two commands agree again.--sensitiveis the legacy spelling of the Secret type and
still works. -
A scan rule's
namesnow beats a built-in rule in YAML and JSON too. A key such as
app: secret_key: sk_live_…reported the built-in Stripe rule instead of the
project's rule, because the text pass (no key) found Stripe and its remediation URL
outranked the named rule;.envwas already right.namesstay case-sensitive regexes
in every format, as in.env: write(?i)or both spellings to coverSECRET_KEYandsecret_key. -
Two
remove(orsync --prune) runs on the same file can no longer undo
each other (E-7). A run now takes an advisory lock file beside each file it
writes (.evnx-<name>.lock, created exclusively, holding its process id)
before it checks the file is unchanged and writes. The second run stops at
once with exit 2, "another evnx (process N) is changing .env", and changes
nothing; a lock left by a crashed run is taken over (the process is gone, or
the lock is over ten minutes old). No locking crate is a dependency and
File::lockneeds Rust 1.89, so a lock file it is. -
removeandremove --checksay when two environment files differ only in
case (.env.Staging,.env.staging): they are separate files except on a
case-insensitive filesystem, and are never merged (G-2). -
A secret under a browser-exposed prefix is worse, and says why (F14-5).
VITE_,NEXT_PUBLIC_,REACT_APP_,EXPO_PUBLIC_and the other prefixes
scanalready knew are inlined into the client bundle.scannow raises such a
finding's severity one level (lowstayslow: a Maps browser key or a Sentry
DSN is meant to ship), andvalidatehas a new error,public_secret, for a
secret-shaped value (a provider key, a private key, a password in a URL) under
one of them;--ignore public_secretskips it. The prefix list is one list,
shared withplatform ... --to vercel(it gainsSTORYBOOK_, and matches in any
case). -
vault sharetakes--useras well as--with, andvault roleand
vault revoketake--withas well as--user(F14-7, OW-16): the member's
email under whichever name you learned first. The aliases were already in the
parser; this adds the test that holds them, so neither flag can be dropped by
accident. -
convert --to jsonand--to vercelwarn, on stderr, when--excludeis used
without--include(F14-9, OW-19).--excludeis a denylist, so a variable you
did not think of is copied into a file that usually ends up in an app build or a
web bundle (ANDROID_KEYSTORE_PASSWORDdoes not match*_SECRET*). The output is
unchanged and the exit code stays 0; the warning says to list what may go there
with--include. -
cloud pushreads the vault's current version after the password prompt and
the Argon2id unlock, not before (CLOUD-409). The version it sends as
base_versionis now read just before encrypting, so the window in which a
teammate's push makes yours answer 409 no longer includes a person typing a
password. A wrong password also stops before any version request. -
The cloud master password is read by
utils::secret_source, asbackupand
restoreread theirs (CLOUD-PW, the cloud half of B-WP3).--password-stdin,
EVNX_MASTER_PASSWORDand the prompt behave as before, and the notice on
stderr is unchanged. New:EVNX_MASTER_PASSWORD_FILEnames a file whose first
line is the password (nothing else is trimmed), read before the variable and
announced without the value; a missing or empty file is an error. Changing a
password (rotate-master-password) andloginstill always ask. Tests cover
each source in order. -
cloudwarns when a vault's "latest" goes backwards (RM-89, a client-only
slice of P6). The CLI remembers the highest version of each vault it has seen
(pushed, fetched, or reported as latest) inhigh-water.jsonin the config
directory: a vault id and a number, no name, key, value or hash. When the
server later reports a lower latest (a version removed, a rollback, a restored
backup),push,pull,run,diffandexportprint a warning on stderr
naming both numbers andevnx cloud history. It is a hint and never stops
the command; it is per machine, andvault deleteforgets the vault. -
A vault id typed in upper case is stored in lower case (C10).
--vault 0B8A…was kept verbatim and bound into the encrypted version's AAD,
while the server's id, which a pull by name uses, is lower case, so a version
pushed by an upper-case id could not be pulled by name. The id is now
lower-cased as soon as it is read, and an id is matched in any case against
vault list. Pushes already made with an upper-case id: they were sealed
with that spelling, so they open only with the same upper-case id, which only
0.13.0 and earlier still send (evnx cloud pull --vault <the same upper-case id>). Pull them once with 0.13.0 and push again with this version; versions
pushed from now on open by id or by name. -
migratekeeps values off the command line where the tool allows it, and
says so where it does not (K8). Foraws-secrets-managerthe printed
commands now write the JSON to a private (umask 077) temp file through a
quoted heredoc and pass--secret-string file://"$f", then remove the file,
as the Azure commands already did: the values no longer show inpsor shell
history, and the stored secret is exactly the JSON.heroku,doppler,
infisicalandrailwaytake values only as arguments, so a one-line warning
on stderr under the commands says they can show inpsand shell history and
should be run on a trusted machine (-qdoes not hide it;> run.shdoes not
capture it). -
Paid-user output gaps (F14-11, review F6, F7, F10, F11).
cloud historynames you asyou (me@example.com)instead ofyou. A push
by an API token is still recorded under its owner, so token versus person
needs the server and is not done here.vault createprints oneNext:line
that stays true whencloud linkis run after it, instead of a--vaulthint
thatlinkthen contradicted. A push into a production vault from the default
.envshows its production warning once: the question when it asks, the!
line only with--yes.auth statusandcloud statusprint
client evnx <version>. -
@evnx/cliworks where install scripts do not run (R6). pnpm 10+ and bun
skippostinstall, so the placeholderbin/evnx(exit 1) was never replaced
byinstall.jsandevnxwas broken there.bin/evnxis now a Node launcher
that resolves@evnx/evnx-<platform>-<arch>/package.json, runs the binary next
to it with the terminal attached, and forwards the exit code and any signal; a
missing platform package gets a clear error naming--omit=optional.
install.jsandpostinstallare gone. The global-install warning that
anotherevnxcomes first onPATH(RM-42) lived ininstall.jsand goes
with it;evnx update --liststill lists every copy.
Installation
Linux / macOS (shell script)
curl -sSL https://raw.githubusercontent.com/urwithajit9/evnx/main/scripts/install.sh | bashHomebrew (macOS and Linux)
brew install urwithajit9/evnx/evnxScoop (Windows, user-local)
scoop bucket add evnx https://github.com/urwithajit9/scoop-evnx
scoop install evnxWinget (Windows, system-wide)
winget install urwithajit9.evnxnpm / Node.js
npm install -g @evnx/cliPython (pipx)
pipx install evnxCargo
cargo install evnx --all-features
cargo installcompiles from source withdefault = [], so plain
cargo install evnxgives you a binary without theauth,
vaultandcloudcommands. Every other channel above ships them
already — the binaries are built with--all-features.
Checksums
SHA256 checksums are provided for each binary. Verify after download:
sha256sum -c evnx-*.tar.gz.sha256Provenance
Every archive has a signed build provenance attestation, which ties it
to the workflow run and commit that built it. Verify one with the
GitHub CLI (signed in with gh auth login):
gh attestation verify evnx-x86_64-unknown-linux-musl.tar.gz -R urwithajit9/evnxSBOM
evnx-0.14.0.cdx.json is a CycloneDX software bill of materials for
this release: every crate in Cargo.lock, with all features on. It
carries a checksum (.sha256) and the same kind of signed attestation:
gh attestation verify evnx-0.14.0.cdx.json -R urwithajit9/evnxWhat's Changed
- ci: stop creating CLI-version releases on evnx-action by @urwithajit9 in #277
- fix: v0.14.0 bug fixes (add help, Vercel environments, scan names in YAML, dotted spec keys) by @urwithajit9 in #278
- feat(net): trust SSL_CERT_FILE and SSL_CERT_DIR roots (RM-68) by @urwithajit9 in #279
- feat: sync --prune, remove --check, partial-presence warning, remove lock by @urwithajit9 in #280
- feat(platform): evnx platform diff and pull --to vercel (sync phase 1) by @urwithajit9 in #281
- feat(platform): push --to vercel, diff --strict, and P1 live-run fixes (N14-10 P2) by @urwithajit9 in #282
- feat(platform): --prune, custom environments, --git-branch; one backup per file by @urwithajit9 in #283
- fix: v0.14.0 small fixes — public-prefix secrets, cloud password and version checks, vault id case, migrate argv, npm launcher by @urwithajit9 in #284
- feat(scan): notebooks, monorepo paths, parallel scan, precision/recall gate, schema scan view by @urwithajit9 in #285
- fix(scan): stateless GitHub tokens, names in Dockerfile/shell/Python, corpus false positives by @urwithajit9 in #286
- test: invariants, crypto known-answer vectors, fuzz targets (F14-8, SEC-1) by @urwithajit9 in #287
- ci: pin apt jobs to ubuntu-24.04; build evnx.exe with a static C runtime by @urwithajit9 in #288
- feat(cloud): CI key bundle (EVNX_KEY_BUNDLE), so API tokens no longer fetch key material by @urwithajit9 in #289
- fix: verified convert outputs, release SBOM, security policy, resend-verification by @urwithajit9 in #290
- chore(release): v0.14.0 by @urwithajit9 in #292
Full Changelog: v0.13.0...v0.14.0
npm / Node.js
npm install -g @evnx/cli # global install
npx @evnx/cli --version # zero-install usage
npm install --save-dev @evnx/cli # project dev dependency