Repository navigation
Releases
Releases · mechubsec/mecmcp
v0.34.0
Security
- deps: russh 0.65.1 for GHSA-9264-jvgf-6w2f and GHSA-736f-c93q-9fwx (#627).
kex_doneran the shared-secret callback before server host-key verification
(>=0.55, <0.65.1), andtry_signcould sign after an agent lock (0.65.0). The
workspace floor rises to>=0.65.1, because the bound, not the lockfile, is
what protects a consumer that runscargo update. Now 0.65.2 (#630). - deps: rustix 1.1.6 for RUSTSEC-2026-0486 (#625). Lockfile only, in
fuzz/Cargo.lock.
Changed
- mecmcp-oidc: jsonwebtoken 11.1.0 -> 12.0.0 (#631). Breaking for code
that implementsKeySourceor callsJwksCache::keys: both use
jsonwebtoken::jwk::JwkSet, so a consumer must move its ownjsonwebtoken
to 12 together with this release. A consumer still on 11 resolves two
versions and fails with E0053 onfetch_jwks(rustjunosmcp#661). Verification
behaviour is unchanged. jsonwebtoken 12 rejects an empty HMAC key, so
EncodingKey::from_secretis now fallible; only test code called it here.
Dependencies
- rmcp 3.9.0 -> 3.10.0, ureq 3.5.1 -> 3.5.2, clap 4.6.7 -> 4.6.8,
rustls-webpki 0.103.16 -> 0.103.17, rand 0.10.3 -> 0.10.4, trybuild
1.0.122 -> 1.0.125 (#630); arc-swap 1.9.3 -> 1.9.5, uuid 1.29.0 -> 1.29.2,
rustls-pki-types 1.15.1 -> 1.16.0 (#628).
CI
- RUSTSEC-2026-0543 (chrono unmaintained) is ignored in
deny.toml(#632) and
.cargo/audit.toml(#633). chrono cannot be dropped: rmcp and oauth2 require
it. Without the second file,rustsec/audit-checkfailed every push to
mainwhile pull requests stayed green. allowed_signersnow lists the ArmyGreen release-signing key, so
git verify-tagverifies v0.33.0 onward offline (#626).
v0.33.2
Fixed
- ci:
reusable-release-image.ymlfrom 0.33.1 published no image whenever
Docker Hub was configured. GitHub drops any job output that contains a secret's
value (Skip output 'metadata' since it may contain secret), and the Docker Hub
usernamemechubis a substring of everymechubsecimage path. So the plan
job's tags and labels never reached the build and merge jobs: both platforms
built and pushed untagged with no labels, then the merge job refused with
docker metadata produced no tags. No tag moved. The build and merge jobs now
each run the metadata step, and the only job outputs left are the platform
matrix, a boolean and the digest. The first real run (rustproxmoxmcp v0.12.2)
caught it.
v0.33.1
Changed
- ci:
reusable-release-image.ymlpublisheslinux/amd64andlinux/arm64by
default (rustjunosmcp#648). Apple Silicondocker pullfailed withno matching manifest for linux/arm64/v8. Each platform builds on a native runner
(ubuntu-24.04,ubuntu-24.04-arm) and is pushed untagged by digest. A merge job
then tags one manifest list per registry and signs and verifies that list's own
digest. The QEMU step is gone and any other platform is refused, because emulated
builds took most of an hour per release. The smoke test runs on every platform.
Thedigestoutput is now the GHCR manifest-list digest. A private caller must
passplatforms: linux/amd64, since the arm64 runner is free only for public repos.
v0.33.0
Security
- deps: rmcp 3.9.0 for RUSTSEC-2026-0462, unbounded response-body buffering in
rmcp's Streamable HTTP clients (#606). Lockfile only. - deps: pageant 0.2.5 for GHSA-662x-8wvj-2276 (#607). pageant reaches the graph
through russh on Windows targets only, so no shipped Linux binary contained it.
Fixed
- mecmcp-changeset: a transaction whose
validate()writes into the candidate
can now say so withDeviceTransaction::validate_rewrites_candidate()(#612).
Junos<commit-check/>gives a newsystem login userauid, so every change
set adding one was refused at commit withcandidate changed since the caller observed it(rustjunosmcp#626). When it returnstrue,validate_operation
re-reads the fingerprint after every validation attempt, passed or failed, and
stores it inrecord.current; the caller passes that to the commit. Only an
implementation that holds an exclusive device lock from stage through commit may
returntrue, because a refresh without one would absorb another session's
edit. The default isfalse, so every other vendor is unchanged. A re-read that
fails leaves the operationIndeterminatewith the lock held. - mecmcp-runtime: a consumer that sets only
aboutand flattens a shared
Argsstruct no longer prints that struct's rustdoc as its--helpdescription
(#610).ToolModeArgs,WebApproverArgs,VerifiedApproverArgs,
OAuthResourceArgs,EvidenceArgsandCliopt out; the consumer's own
aboutor doc comment still wins.
Changed
- mecmcp-changeset: fingerprint refusals now append both fingerprints,
(expected X, found Y)or(recorded X, found Y), and the step that checked,
; checked before commit(#612). The leading phrases are unchanged, so code
matching them withcontainskeeps working; an exact-equality match will not. - mecmcp-runtime: a binary that parses
Cliitself (parse_for) no longer
showsCommon CLI arguments for MCP servers.as its description (#610).
Added
- mecmcp-transport:
test_harness::serve_on_loopback_tls(behind
test-util), and the first test that completes a TLS handshake through
serve_router's listener (#611).
Dependencies
v0.32.0
Fixed
- ci:
reusable-release-image.ymlVerify signatures (hard gate)now retries each
ref'scosign verify(6 attempts, backoff 2/4/8/15/30 s, about 59 s) before failing.
Docker Hub's referrers index is eventually consistent, so verify right after signing could
reportno signatures foundand fail the release (seen on rustsdcmcp v0.6.0). The gate
stays hard: a signature that never appears still fails the job with cosign's last error.
Consumers pin this workflow by SHA, so they only get the fix after bumping that pin (#595).
Added
- mecmcp-http: opt-in explicit egress proxy,
HttpClientConfig::explicit_proxy: Option<ProxyUrl>(#589).None, the default, is unchanged: no proxy.ProxyUrl::parse
accepts only anhttp://orhttps://URL with a host, no path, query, or
fragment, and at most 2048 bytes serialized (hyper-util would otherwise drop an
over-long proxy without an error and go direct); HTTPS origins are tunnelled with
CONNECT, so TLS stays end-to-end. Userinfo is accepted and sent as
Proxy-Authorization: Basic, matching rustsdcmcp'segress_proxy. It never appears
inDebugor errors, andProxyUrlhas noDisplay. Environment proxy variables
(HTTP(S)_PROXY,ALL_PROXY,NO_PROXY) stay ignored with or without it.
Changed
- Breaking for source (#589):
mecmcp_http::HttpClientConfighas a new public
field,explicit_proxy: Option<ProxyUrl>. Code that builds anHttpClientConfigby
full struct literal must addexplicit_proxy: None. rustmistmcp does
(crates/rustmistmcp-core/src/client.rs:187and:290) and will not compile until
it does. Consumers that fill the rest with..Default::default()— mecfeedmcp,
rustopnsmcp, rustunifimcp, rustproxmoxmcp and mecviz — are unaffected. When
rustsdcmcp moves itsegress_proxyontoProxyUrl, a deployed value with a path or
query (which it accepts today and reqwest ignores) will be refused at startup.
v0.31.1
Fixed
- mecmcp-redact:
junos::redact_json_valuenow masksprivacy-key, the one
Junos-onlyREDACT_LOG_KEYSname the generic denylist missed, so
get_junos_config format=jsonno longer returns it in plaintext (#591). Bare generic
words inREDACT_LOG_KEYS(e.g.value) are deliberately not treated as secret JSON
keys: they are ordinary field names in consumers' structured output. - mecmcp-redact:
junos::redact_json_valuestring values now also get the Junos log
vocabulary (redact_log_text), so raw CLI/config text in a field such asvalueis masked
likejunos::redact_text(#591). - mecmcp-redact: a
junos::redact_json_valuestring value that starts with<and parses
as XML (e.g.| display xmloutput in a batchcommands[].value) gets the structural
junos::redact_xml, so it stays well-formed. Other strings, including<-led text that
does not parse, keep the text scan (#591). - mecmcp-redact:
junos::redact_xmlattribute values (not otherwise masked) and comment
bodies get the same scan as its character data: the generic scan with the session shield,
thenredact_log_text(#591). Before, they gotredact_log_textalone, so
note="password: ...",<!-- token=... -->andftp://u:...@h/userinfo passed through.
redact_xml_strand its profile variant are unchanged: attribute values and comments get
shape checks only. - mecmcp-redact:
junos::redact_xmltext and CDATA nodes now also get the Junos log
vocabulary (redact_log_text), so CLI text such asprivacy-key "..."inside an
<output>element is masked while the document stays well-formed (#591). A redacted CDATA
body that would contain]]>is split into adjacent CDATA sections, so a mask can no
longer end the section early. Attributes named by a Junos extra key (privacy-key,
value) are masked, and other attribute values and comment bodies getredact_log_text. - mecmcp-redact:
junos::redact_log_artefactreturns the structuraljunos::redact_xml
result for XML-shaped input instead of rerunningredact_log_textover the serialized
document, which read markup as values and truncated it (e.g.privacy-key "..."lost
</output></rpc-reply>) (#591). - mecmcp-redact: XML redaction (
redact_xml_str, its profile variant andjunos::redact_xml) scans each
run of adjacent text, entity references and CDATA as one string, so a key in text and its
value in a CDATA section (password=<![CDATA[...]]>) are masked together (#591). Such
a run is written back as one escaped text node; a lone CDATA section keeps its form. - mecmcp-redact: XML redaction (
redact_xml_str, its profile variant andjunos::redact_xml,
hencejunos::redact_log_artefact) no longer writes processing instructions and DOCTYPE
declarations out unscanned (#591). A PI the character-data scan would change is dropped.
Behaviour change: every DOCTYPE loses its internal subset (<!DOCTYPE x [...]>becomes
<!DOCTYPE x>, keeping the root name and any external ID verbatim), because entity values
can hide a key behind a character reference. A DOCTYPE whose remaining text the scan would
change is dropped. Other PIs and DOCTYPEs without a subset pass through unchanged.
v0.31.0
Added
- mecmcp-redact:
junos::redact_text, the complete Junos text redactor
(generic scan, then the Junos log vocabulary) that preserves vetted
non-secretsessiontext: counter rows such asValid sessions: 12,
thesession-init/session-close/limit-sessionkeywords, and
show security flow session ...filters with one address or integer
argument (rustjunosmcp#579). - mecmcp-server: the built-in
whoamitool (mecmcp_server::whoami,
WHOAMI_DESCRIPTION). It returns the calling token's own name, auth type
and actor type, pluscreated_atandexpires_atfor a static token
(nullmeans it never expires), so an unattended client sees an expiring
token before the 401. A server opts in by registering the tool (#580). - mecmcp-auth:
WHOAMI_TOOL;ScopeSet::allows_toolpermits it under
every tool scope, including an empty allowlist, unless the server lists a
whoamiof its own inwrite_tools(#580).
Changed
- mecmcp-redact:
junos::redact_json_valuestring values now go through
the same session-preserving scan. Generic and other-vendor paths are
unchanged. - Breaking for source (#580):
mecmcp_auth::CallerCtxhas two new public
fields,token_created_atandtoken_expires_at, both
Option<DateTime<Utc>>. Code that builds aCallerCtxby struct literal
must addtoken_created_at: None, token_expires_at: None. The store fills
them for a static token, and an OAuth caller leaves bothNone. A GitHub
code search (not exhaustive) found such literals in rustjunosmcp
(rust-junosmcp/src/server.rs,rust-junosmcp/src/server/srx.rs),
rustsdcmcp (crates/rustsdcmcp/src/server.rs), rustmistmcp
(crates/rustmistmcp/src/server/mod.rs) and rustunifimcp
(rustunifimcp/src/server/mod.rs).
Fixed
- mecmcp-redact:
junos::redact_log_text(and sojunos::redact_text
andjunos::redact_log_artefact) runs in linear time. A long whitespace
run, a line with many secret keys, or a long token with many internal
boundaries (a.a.a…,x=x=…,\"\"…, U+00A0 runs) rescanned the line at
every position: 200 KB of device text took up to 31 s. Output is
byte-for-byte unchanged (mecmcp#584).
v0.30.1
Fixed
token listandtoken set-scopesshowdevice_selectors. The DEVICES column now
reads e.g.* [site=lab-*], andset-scopesprints the selectors it leaves unchanged;
before, a selector-scoped token was listed as if it had its fulldevicesscope.- 0.30.0 upgrade note corrected. An older server (mecmcp ≤ 0.29.4) that merely reads
atokens.jsoncontainingdevice_selectorsignores the field and grants the broad
devicesscope at once — not only when a token is rewritten. Upgrade every binary that
reads or rewrites the store before adding selectors.
v0.30.0
Minor release (breaking for literal construction). New public fields:
TokenEntry.device_selectors,CallerCtx.verified_proposer,Attribution.verified_proposer,
ChangeSetRecord.owner_verified,TokenAction::Add.device_selector,
VerifiedApproverArgs.require_verified_proposer; new variant
VerifiedApproverArgsError::ProposerRequiresIssuer. Behaviour is unchanged unless the new
options are used. Upgrade every binary that rewritestokens.jsonbefore adding
device_selectors: an older binary drops the field ontoken rotate/set-scopes,
widening the token.
Added
- Static
tokens.jsonentries acceptdevice_selectors, the same grammar and semantics as
the OAuth scope map (narrowsdevices;nullrefused; one attribute snapshot per
authentication; an unlabelled device never matches). Opt in with
CallerCtx::from_entry_with_attributes;CallerCtx::from(&entry)grants no devices to a
selector entry.token add --device-selector(repeatable) and
TokenStoreFile::add_with_selectors(#400). - Approver roles are selector-scoped through the same field: approval already checks the
approver's device scope against the change set's one device, so asite=lab-*approver
cannot approve a prod change set. Pinned by an integration test; no new flag (#400). ChangesetCoordinator::approve_change_set_within: approves only if every device in the
change set'stargets()passes the caller's scope check, so asite=lab-*approver cannot
approve a multi-target set that also names a prod device. Servers' approve handlers should
call it instead ofapprove_change_set(#400).Mecmcp-Proposer-Assertionstep-up header (#400): a fresh IdP assertion whose subject equals
the proposing token's boundoidc_subject, bound withbind_proposer, which needs neither a
humantoken nor the approver role. A separate header and a separate field
(CallerCtx::verified_proposer,Attribution::verified_proposer) so it is never approval
evidence; it shares thejtireplay set, so one assertion is good once in either role.
A server with no verifier answers the header 400, as for the approver header. Both
assertion headers are now removed from the request once verified, so the JWT never reaches
tool extensions.--require-verified-proposer(VerifiedApproverArgs, requires--oidc-issuer) and
ChangesetCoordinator::with_require_verified_proposer: a proposal must carry an
IdP-verified proposer (ProposerIdentity, built byfrom_attribution; an OAuth principal
counts), proposed through the newcreate_change_set_as. The change set records
owner_verifiedand approval refuses one without it, so the two-person rule compares two
IdP-verified subjects. A verified record makes the state file version 9; nothing changes
for a deployment that leaves the flag off. Under the mode, both waiver transitions and
apply_change_setalso refuse a record whose owner was not verified at propose time, so a
change set proposed before the mode was on cannot skip it (#400).
Changed
- Every audit line gains
identity_mechanism(token,oauth,oidc-step-up,none), and an
approval's evidence record carriesmetadata.approver_mechanism(token/oidc, plus the
verified issuer and subject), via the newEvidenceRecorder::approval_with_identity. The
plainapprovalis unchanged (#400).
Documentation
docs/OAUTH.mdsection 9: Client ID Metadata Documents are the authorization server's
concern, what the resource server already provides for them, and what to configure at the
IdP. The OIDC human-identity spec's Phases 1-4 carry status notes (#400).
v0.29.4
Added
- OAuth scope-map entries accept
device_selectors(site=,vendor=,tag=,*glob,
AND) that narrowdevicesby inventory attribute, resolved per authentication against a
consumer-suppliedDeviceAttributeSource(OAuthResource::with_device_attributes). New
mecmcp_auth::{DeviceAttrs, Selector, SelectorSet}. Strictly additive: files without
device_selectorsbehave identically (#400 phase 2).