Repository navigation
Releases
v0.22.1
Fixed
- vega login now sends x-vega-capabilities: r2-grant on every enroll call. The control plane had no way to tell this build from one too old to know an r2Grant field exists, so an incapable client's enroll call silently consumed a pending direct-R2 grant and discarded it. The control plane now withholds a reveal from any request that does not declare this.
- stdenv.isLinux replaced with stdenv.hostPlatform.isLinux in flake.nix.
Added
- CI now runs nix flake check on every push and pull request, failing the build on a deprecation warning rather than only surfacing one at release time through the GHCR publish workflow.
v0.22.0
An operator can now grant a trusted tenant direct R2 write access to its own dedicated bucket, with the credential revealed to that account exactly once, on its next vega login.
Added
vega loginsaves and announces a direct-R2 tenant write grant when an operator has issued one for that account (spec 0006), the one time it is ever shown: the control plane erases its own copy the instant the response leaves it. Saved to~/.config/vega/r2-grant.jsonat mode 0600, written atomically (a temp file renamed over the target, never truncated in place), and saved before the owner credential so that losing the one secret that cannot be re-revealed never depends on a later write succeeding. The secret itself never reaches stdout; a save that still fails prints it once to stderr with a loud warning rather than losing it silently. No flag, no config: an account either has a pending grant or it does not.
v0.21.0
Large or unreliable-connection uploads no longer have to restart from zero: a NAR of 64 MiB or more now uploads in resumable parts.
Added
- NARs of 64 MiB or more upload in parts through the control plane's multipart
routes. Each part is retried on its own under the same inactivity timeout as a
single PUT, so a stall costs one part. A job re-run after an interrupted
upload resumes from the parts the control plane has recorded rather than from
zero, with no state kept on the runner. A session that expires, is aborted or
is being completed by an overlapping run is waited for or restarted within the
same run. NARs past R2's 4.995 GiB single-PUT limit can be uploaded, up to the
control plane's multipart limit. R2 checks each part against its MD5 only; the
control plane verifies the assembled NAR's SHA-256 before accepting it. A
control plane without multipart support still receives a single PUT.
v0.20.2
Every fetch now refuses a response that declares a compressed encoding before reading its body, closing the Brotli decoder RCE ahead of a patched Node release.
Security
- Mitigated the Brotli decoder out-of-bounds write (CVE-2026-111497) ahead of
a patched Node release.fetch()decompresses a response based on its own
Content-Encodingheader, including Brotli, regardless of what was
requested, so every cache, upstream, device-flow, OIDC and control-plane
fetch now refuses a response that declares a non-identity encoding before
reading its body. A deliberate refusal is no longer retried as a dropped
connection.
v0.20.1
A failed reproduction names the derivation it blames again, and can no longer be made to blame the wrong one.
Fixed
- The agent reads the failing derivation from nix's own failure line and recognised only
builder for '<drv>' failed. Nix 2.34 and 3.21, the version the reproducer pins, sayCannot build '<drv>'., so the field was empty on every real failure and per-derivation grouping had nothing to group by.
Security
- A failing build could name a different derivation as the one that failed. The older wording was matched anywhere in the captured output, and a build controls what it writes, so printing that phrase against another derivation was enough to be believed. On current nix, where nix itself no longer writes that phrase, every such match would have come from the build. A regular expression returns its earliest match and nix quotes a build's log lines before its own error, so a forged line would have been preferred to the real one.
- Both wordings are now taken only from the start of nix's own
error:line, which quoted build output cannot begin.
v0.20.0
A failed reproduction now says why it failed, and vega diagnose reports what is blocked.
Added
- A failed reproduction reports one diagnosis, the derivation nix named as failing and a bounded, redacted excerpt. A candidate that can never build is visible instead of being re-dispatched for ever.
vega diagnosereports the reproductions blocked for an owner, and exits non-zero for a blocking one under--fail-on-blocking.
Fixed
- Capturing the build log put the agent in the middle of a stream the kernel used to throttle, so a console that could not keep up left the unwritten output queued in memory without bound. The build now waits for the console.
- The retained tail of a log began wherever its byte bound landed. A cut inside a line takes away the prefix redaction matches on, leaving the rest of a secret in an excerpt as ordinary text. The tail starts at the first whole line.
- The failure classifier short-circuited on exit code 126, which nix never returns.
Security
- The reusable reproduction workflow ran dependency install scripts in the job that holds the OIDC token gating signing, and never set
VEGA_ATTEMPT. It now installs without them, and a test derives the required variables from the agent source so it cannot drift from them again.
v0.19.0
Added
-
VEGA_EXCLUDEleaves matching store paths unpublished, as comma-separated
globs over the path NAME (the part after the hash, so a pattern survives a
rebuild):VEGA_EXCLUDE='*.erofs'.VEGA_MAX_NAR_BYTESdoes the same by size,
as a blunt guard against an accidental multi-gigabyte push. Both are exposed on
the Action asexcludeandmax-nar-bytes.This is for output no other machine can use. A host-specific microVM disk image
is one to two gigabytes, changes whenever its guest config does, and nobody
else will ever substitute it; on the deployment that asked for this, such paths
were 40 of 49 publish minutes. An excluded path is neither uploaded nor
attested, so it contributes to no tier, and anything referencing it leaves a
consumer a dangling reference to build themselves. That is right for a
host-specific artifact and wrong for anything another machine might want.Every skip is logged, with a count, the total NAR bytes and each path: a silent
exclusion is indistinguishable from a cache that quietly lost paths. A pattern
that matches nothing warns, as does excluding one of the build's own outputs,
which forfeits its provenance. A malformedVEGA_MAX_NAR_BYTESfails the run
rather than quietly reverting to no ceiling. -
The builder gives its store a free-space floor, so a build collects garbage
instead of taking the host's disk to zero. The image shipped only a periodic
GC, which runs between builds and therefore cannot help when one build cycle
produces more garbage than the disk has headroom: on a shared host that ran to
0 bytes free twice in a day, against roughly 49 GiB of collectable paths per
cycle. Nix's ownmin-free/max-freecollect during a build, and the
entrypoint now writes both into the generatednix.confand prints them at
boot. -
VEGA_MIN_FREEandVEGA_MAX_FREEtune the floor, accepting a byte count or a
binary suffix (VEGA_MIN_FREE=25G).VEGA_MIN_FREE=0disables it, and the
boot line says so rather than leaving the operator to infer it. The defaults
are a fraction of the store's filesystem (10% and 25%, floored at 1 GiB, capped
at 25 GiB and 60 GiB), because a fixed size is wrong at both ends: 25 GiB on a
20 GB runner would collect on every check, and 1 GiB on a large shared host is
a rounding error against one build cycle. -
Booting against a mounted
/etc/nix/nix.confthat sets nomin-freewarns.
That file belongs to the operator and the entrypoint does not edit it, so the
floor is theirs to set; the failure mode is a full host rather than a failed
build, and it is invisible until it happens.
Fixed
vega verifychecks whether Vega has withdrawn a binding. It verified the
signature, the signed tree head and log inclusion, and never asked, so a path
whose shared signature had been revoked still reported a clean pass, which is
the one answer a verifier must not give. The revocation list is fetched from
the cache origin (it is global, not per-tenant) and authenticated with the
user's pinned shared key, never one the same cache served. A revoked binding
now fails by every route, including the tenant and signature-only ones and the
MCP risk gate an agent acts on, whatever key signed the build, since "Vega withdrew this" is a different
answer rather than weaker evidence. A status that cannot be established is
reported as unknown and is not treated as clean, so a cache that withholds the
list is distinguishable from one with nothing to hide. The list carries a
signed timestamp and total, both checked, so a stale replay or a truncated
answer is unknown rather than clean.
v0.18.0
A stalled NAR upload no longer consumes the whole job before a retry can run.
Fixed
- The per-attempt upload deadline was a wall-clock allowance scaled to the payload, so a 5 GB NAR was given roughly 85 minutes and a wedged connection sat there for most of it: two multi-GB uploads went 50 and 88 minutes without moving a byte and the 90 minute step cap killed the job before any retry. Uploads are now bounded by INACTIVITY. The clock restarts on every chunk that leaves the process and once more when the body ends, so a silent upload fails in five minutes and retries, while a slow but progressing one runs as long as it needs. A fixed cap cannot make that distinction, and shortening the old allowance would abort large uploads that are working, since a presigned PUT is not resumable and every retry restarts from zero.
Added
- Stall warnings report bytes transferred and the current rate per path (
upload /nix/store/... (5293s in stage, 350 MiB/5.0 GiB, 0 B/s)), so a wedged upload and a slow one are distinguishable from the job log alone.
Changed
VEGA_UPLOAD_STALL_SECONDS(default 300) sets how long an upload may move no bytes before it is aborted and retried.VEGA_UPLOAD_TIMEOUT_SECONDSno longer has any effect; an absolute per-attempt cap isVEGA_UPLOAD_CAP_SECONDS, defaulting to off. The rename is deliberate: reusing the old name would have turned an existing1800into a hard 30 minute kill no multi-GB upload could survive.
v0.17.3
A publish-crash fix.
Fixed
- The publish crashed at the very end of a large build with
RangeError [ERR_OUT_OF_RANGE]: The value of "delay" is out of range. It must be an integer. The per-attempt upload deadline scales with the payload as(bytes / MiB) * 1000(so a slow but progressing multi-GB PUT is not aborted), which is a float for any NAR that is not a whole number of mebibytes, andAbortSignal.timeoutrejects a fractional delay. A ~1.9 GB closure produced ~1859098.75 ms and the job exited 1 every run, so a self-hosted runner went permanently red even though the build was fine and the cache converged on resume. The deadline is now rounded up where every request deadline is applied. The value itself was correct; only its type was wrong.
Builder image rebuilt on this release, same github-runner 2.336.0.
v0.17.2
A dependency and hardening release; no behavior change.
Changed
@noble/curvesmoved from 1.x to 2.3.0. The 2.x line renamed the subpath export, so the Ed25519 import is now@noble/curves/ed25519.js. The signing API is unchanged, and 1.x and 2.x produce byte-identical public keys and signatures for the same key and message and cross-verify each other, so every existing signature still verifies and the trust root is unaffected.- A known-answer signing vector now pins the exact signature bytes, so a future crypto-library change that silently altered the output fails a test rather than shipping.
The builder image is rebuilt on this release with the same github-runner 2.336.0 as 0.17.1.