A security release: tenant isolation on every path-tenant handler, the
platform:admin scope, globally unique SMTP usernames, the fail-closed
bootstrap gate, and seven RUSTSEC advisories closed. Three of those break
an existing deployment and the changelog marks each one, the object store
swap included - an upgrade keeps the old minio_data volume while Sentio
talks to an empty store.
cargo update took every compatible bump; all 61 direct dependencies were
already on their latest major line.
Gate 7/7 with 752 tests, e2e 19/19.
Regenerated docs/openapi.json and THIRD-PARTY-NOTICES.md, which the bump
makes stale. utoipa 6 stops writing empty description fields and puts the
$ref before the null in a nullable oneOf; 85 paths, 116 operations and 182
schemas before and after, so nothing moved but the formatting.
These were verified by hand while chasing the spam-scoring bug. Spot
checks rot, so they are assertions now: 19 checks, up from 13.
- GTUBE must score above 10. rspamd answered every request while scoring
every message -0.1, and nothing looked wrong; this is the assertion
that catches that class of fault.
- EICAR must be rejected in-band, which is the one scanner verdict
Sentio enforces itself.
- The webhook signature is recomputed over the raw delivered bytes. A
re-serialised body would verify even if the scheme were wrong, so the
check captures the request with its own listener.
- STARTTLS then AUTH, including that AUTH is absent before the upgrade -
credentials must never be solicited in cleartext. Skipped, not failed,
when no certificate is mounted; the skip path was tested by removing
the cert.
- The rate limiter and DNSBL cache must be writing to the KV store.
Also adds FRESHCLAM_CHECKS=24 so signatures refresh hourly rather than
daily; the image default of 1 can leave the scanner a day behind.
The harness only sent bodies, so nothing exercised the object store. A
20 kB attachment now goes in over SMTP and comes back out through the
API, compared by SHA-256 - an encoding slip or a truncated upload fails
instead of passing quietly. It also asserts the ClamAV scan_status on
the attachment row and that the raw message was archived.
Verified by hand first: MinIO gained two objects (the attachment and the
base64 raw EML), the bytes matched, and the archived message ends with
CRLF, which is the fix from the previous commit holding on a real MIME
message.
Windows compiles green in CI, so the release now carries a third
platform. It needs a real cargo build on a Windows runner rather than
the image-extraction path the Linux binaries use, so it is its own job
feeding the release, and SHA256SUMS is computed once all three are in
place.
Release notes say plainly that Linux is the supported target and the
Windows build is compiled but not exercised in CI.
The banner printed CARGO_PKG_VERSION and the spec hardcoded 1.0.0, so
Cargo.toml said 0.1.0, the spec said 1.0.0 and the tag said 0.1.2. All
three now come from [workspace.package] version, and CONTRIBUTING says
to bump it before tagging.
send-multipart and dns-check are routed and served but were never
registered with utoipa, so the spec omitted them and a generated client
would have lost both. Spec is now 116 operations over 85 paths, matching
the 116 routed handlers.
Also: config table was missing 7 of the 20 sections, the top diagram is
mermaid so GitHub renders it properly, and the callout boxes are plain
prose.
Relicensed so a recipient may take whichever terms suit them. That is
strictly more permissive than either alone: MIT for the simplest terms
and GPLv2 compatibility, Apache-2.0 for the explicit patent grant some
adopters require. It also matches the Rust norm - 255 of the 517
dependencies are exactly this pair.
Distributing binaries is what makes attribution an obligation rather
than a courtesy. The repository alone redistributes nothing, since
dependencies are fetched at build time, but the container image and the
release tarballs contain compiled third-party code, and MIT and BSD both
require their notices to travel with it. 76 crates are MIT-only.
THIRD-PARTY-NOTICES.md is generated by cargo-about from the resolved
graph for the two targets actually shipped, and now lands in
/usr/share/doc/sentio in the image and inside each tarball.
deny.toml existed but nothing ran it, and it would have failed
immediately: the allow list omitted licences already present (0BSD,
MIT-0, Unlicense, CDLA-Permissive-2.0, BSD-1-Clause) and it used the
pre-2.0 schema. Rewritten against what the graph actually contains, and
wired into CI alongside a staleness check on the notices.
Running it surfaced six advisories, none of which appear exploitable
here, all recorded with reasons rather than silenced:
- rsa 0.9 carries Marvin (RUSTSEC-2023-0071), a timing sidechannel in
private-key operations. The crate is used in one place, generating a
DKIM keypair. No decryption, no oracle. There is no fixed release.
- h2 0.3 and rustls-webpki 0.101 come through aws-smithy-http-client's
legacy hyper 0.14 stack, so they sit on the path to the object
store. Sentio's own listeners and delivery client are on rustls
0.23. Getting the AWS SDK off its legacy client is follow-up work.
- rustls-pemfile is unmaintained, not vulnerable.
The quickstart told people to run `up -d --build`, which is a release
build of thirteen crates before they can see anything work. That is a
poor first impression for software whose whole pitch is that it is easy
to run. `docker compose up -d` now pulls ghcr.io/truespar/sentio.
Each architecture is built on its own native runner - amd64 on
ubuntu-latest, arm64 on ubuntu-24.04-arm - and the two are merged into
one manifest afterwards. The obvious alternative, one buildx job with
both platforms, cross-builds arm64 under QEMU, and a Rust release build
under emulation runs for well over an hour. arm64 matters here because
most people trying this on a laptop are on Apple silicon.
Tags: latest on main, sha-<short> always, and semver on v* tags so a
deployment can pin. GHCR needs no credentials beyond GITHUB_TOKEN.
docker-compose.build.yml keeps the from-source path for anyone changing
the code, and the image name is unchanged between the two so nothing
else in the stack cares which produced it. CI's docker job now runs on
pull requests only, since main is covered here.
Verified locally through the build overlay: image builds, stack comes
up, e2e passes 9/9.
The README and the agent guide both said the payload arrives parsed,
with headers, bodies and attachments. It does not. InboundPayload in
crates/sentio-smtp-server/src/routing.rs is POSTed verbatim and carries
identity, envelope, spam_score/spam_action, llm_category/llm_summary,
threading, loop-guard headers, DSN parameters, and raw_eml_key.
No bodies. No attachments. Anyone writing a handler from the old text
would have gone looking for a body field and found nothing, which is the
worst kind of documentation error: confidently wrong about the first
thing a reader touches.
Both pages now show the real shape and point at GET
/v1/messages/{id}/raw and /attachments for the content, and say why the
split exists - a webhook that stays small regardless of a 25 MB
attachment, and a handler that only needs the spam verdict never moving
the body at all.
Two more details taken from the code rather than assumed:
- Route matching is exact, domain, regex or catch_all; first match
wins, and a regex that fails to compile is logged and skipped rather
than failing the delivery.
- Retry classification is 2xx delivered, 5xx/408/429 transient, any
other 4xx permanent, network errors transient - with retries
scheduled through the queue so they survive a restart. The
consequence worth stating for agent authors is that a handler
blocking on a model will time out, be classified transient, and earn
a duplicate.
The capabilities table said "parsed message data" too; corrected.
Sentio SMTP: self-hosted, multi-tenant email infrastructure in Rust.
Gives an AI agent its own real email address. Inbound mail arrives
parsed, authenticated and spam-scored as a webhook; replies go out
DKIM-signed over the same REST API. Full SMTP on both sides, with
DKIM/SPF/DMARC/ARC, MTA-STS, DANE and a three-tier anti-spam pipeline.
Apache-2.0.