openmls 3.3.0
openmls: ^3.3.0 copied to clipboard
Dart wrapper for OpenMLS — a Rust implementation of the Messaging Layer Security (MLS) protocol (RFC 9420)
3.3.0 - 2026-10-26 #
For Users #
✨ Highlights
- Judge a commit before it takes effect —
inspectMessagereports what a commit would do, with who proposed each change and whose leaves it removes, and applies nothing. - Merge an own commit only once the delivery service took it —
FlexibleCommitOptions.keepPending. - Group-context extensions OpenMLS defines are passed typed — so an
application's own extension can be required with
RequiredCapabilities. - On macOS,
dart run,dart testandflutter testwork from a deep project path — the native library now leaves room for any path in its install name; see Fixed. - openmls v0.9.2 — unchanged this release.
- openmls_frb v2.4.0 — Rust FFI bindings
Added
MlsEngine.inspectMessage— processes a message likeprocessMessageWithInspectand reports the same, without merging a commit, storing a proposal, advancing a ratchet or writing to storage. MLS lets any member commit any Add or Remove and leaves who may to the application; with this, an application judges a commit by its own policy and processes it only when it accepts it. A refused commit is never processed by the member that refuses it — which keeps only that member in the current epoch: if others accept it, the refuser is the one left behind (seeSECURITY.md).StagedCommitInfo.removeCredentials— the credential of each removed member, read before the merge, one perremoveIndicesentry: after the merge those leaves are gone.StagedCommitInfo.proposals(StagedProposalInfo) — every proposal the commit applies, with its own sender and whether it came by reference. A commit can carry other members' proposals, so the committer is not necessarily the author of each change.StagedCommitInfo.committerCredential— the committer's new leaf credential when the commit has an update path, which can replace it.StagedCommitInfo.groupContextExtensions— every extension of the group context the commit leads to, the types OpenMLS interprets itself (RequiredCapabilities,ExternalSenders, …) included, in wire form.StagedCommitInfo.externalSenders(MlsExternalSenderInfo) — the external senders that group context admits, decoded: each may propose adds and removes without being a member.decodeExternalSendersdecodes the same from an extension's data.- Per-proposal detail in
StagedProposalInfo—sender(MlsSenderInfo: member, external sender, new member),leaf(the credential, signature and encryption key an Update or Add introduces),affectedMember(the member a Remove takes out or an Update replaces, as this client's tree has it) andextensions(what a GroupContextExtensions proposal sets). OpenMLS does not check that an Update keeps the proposer's identity; withleafnext toaffectedMemberthe application can. ProcessedMessageInspectResult.senderand.proposalInfo— the authenticated sender of any inspected message, external senders included (senderIndexisnullfor those), and for a proposal message the same detail as a commit's proposals, before it is stored.FlexibleCommitOptions.keepPending(defaultfalse) — leaves the commit pending instead of merging it, formergePendingCommitonce the delivery service accepted it orclearPendingCommitwhen another member's commit won the epoch. Processing that member's commit also replaces it.
Changed
- Extensions are decoded as OpenMLS decodes received ones
(
extensions_from_mls) — a type OpenMLS defines (RequiredCapabilities0x0003,ExternalSenders0x0005, …) becomes its typed variant instead of an unknown extension, wherever the API takesMlsExtensions: group creation, key packages, leaf nodes, group-context updates and proposals, flexible commits. OpenMLS looks for the typed variants: an application's group-context extension used to be refused for want of aRequiredCapabilitieslisting it, which could not be given. A malformed extension of such a type is now anInvalid extension 0x….error at the call. StagedCommitInfohas five more fields,ProcessedMessageInspectResulttwo — all optional in their constructors (empty lists,null), so code that builds one itself, a test double say, compiles unchanged.- Own messages that cannot be read back say so more helpfully — a
commit of this member's fanned back while it is pending now points to
mergePendingCommitunder the ciphertext wire format too. The hint is wording only: the content type it reads is not authenticated before decryption, and any member can send a message that names this one as its sender, so nothing acts on it and the error says the message proves nothing — merge on the delivery service's confirmation, not on it. flexibleCommitrefuses to run while a commit is pending — with or withoutkeepPending, which is what lets a commit stay pending. OpenMLS's commit builder replaced a pending commit silently; had the first already reached other members and been replaced, the group would have forked onto incompatible epoch secrets. Merge or clear the pending commit first.- The native library's dependencies move to their latest compatible
releases (
rust/Cargo.lock, #72) — Dependabot's 47 updates, none of them an openmls crate and none adding or removing one: among them the post-quantumml-kem0.3.2 → 0.3.3 andx-wing0.1.0 → 0.1.1,sha20.11.0 → 0.11.1,der0.8.3,spki0.8.1,rand0.10.3,zerocopy0.8.63,tokio1.54.1, andcc1.4.4 → 1.10.0, which compiles the bundled SQLCipher and OpenSSL. CI built and tested the library with them on Linux, macOS and Windows, cross-compiled it for Android, type-checked it for iOS, and passedcargo audit,cargo denyand the fuzz targets. Takes effect with the nextopenmls_frbrelease.
Fixed
- On macOS,
dart run,dart testandflutter testno longer fail from a deep project path (rust/build.rs) — each copies the native library into the project and rewrites its install name to the copy's absolute path (install_name_tool -id), and the released libraries left room in their Mach-O header for about 180 characters of it: 183 for the arm64 build, 199 for x86_64. A project whose copy landed deeper failed with "larger updated load commands do not fit".openmls_frbis now linked with-headerpad_max_install_nameson Apple targets, which leaves room for any path, and the release workflow refuses a library without that room. Measured on a local build: room for 3,255 characters instead of 183, a library no larger, and consumers at a 186-character path, which failed under bothdart runandflutter test, now pass, as do ones at 400. Flutter app builds for macOS and iOS were never affected: they install the library as a framework under a short@rpath/name. Takes effect with the nextopenmls_frbrelease.
Documentation
- Commit refusal, forks and Welcome ordering —
inspectMessage,CommitResult.welcome,FlexibleCommitOptions.keepPendingandSECURITY.md(item 7, which said there was no inspect-then-reject path) now state that a refusal is local, that a delivery service can fork a group andgroupEpochAuthenticatordetects it, that reported credentials are claims to verify, that inspecting an application message consumes nothing, and that a Welcome goes out only after the delivery service accepted its commit. - A commit's GroupInfo waits for the delivery service too —
CommitResult.groupInfoandFlexibleCommitOptions.keepPendingsay so: a party joining by external commit from a GroupInfo published early enters an epoch that may never happen. AndMlsSenderInfo.indexof an external sender says which list it counts in: theExternalSendersthe message was verified against, which for a commit's proposals is the one before the commit, notStagedCommitInfo.externalSenders.
For Contributors #
Added
- Tests that hold #57's guarantees from the outside
(
test/inspect_guarantees_test.dart) — written when it was merged, apart from its own tests:inspectMessageleaves the database file byte for byte as it was, for a commit, a proposal and an application message, and the same bytes then process, whileprocessMessageWithInspectdoes apply the commit it reports;decodeExternalSendersrefuses every truncation of a valid body, lengths that claim more than follows and trailing data, and answers boundary mutations and 2,000 random inputs without a panic; no call that makes a commit replaces a pending one, which then merges as the commit the others applied; and anExternalSenderschange proposed on its own shows in the inspected proposal and in the commit that applies it.
Changed
- Copier template v4.18.0 → v4.19.0 —
repair-buildrefuses a named run built from a fork, by the run'shead_repository, instead of checking out the fork's tree beside the agent's provider key; an AI provider key holding anything but printable ASCII is refused when it is read, by the name of its variable, sodart:io's header error can no longer carry it into a template-update pull request; an upstreamRELEASE_NOTES.mdorCHANGELOG.mdthat could not be fetched (a rate-limited anonymous request answers 403, andcurl -sexits 0) is reported as unknown instead of as "No release notes were published";finalizeChangelogrewrites only the footer's[Unreleased]:link and refuses a second one;publish.ymlpasses the release notes behind a random delimiter;CONTRIBUTING.mdnames the bot branches the signing ruleset leaves out; andverify_release_artifacts.pyrefuses an Apple library with less than 1,024 characters of install-name room. One conflict, inCONTRIBUTING.md: the template places its Releasing and Repository rulesets sections elsewhere than this file does, so this file's layout was kept with the template's one changed line.
3.2.2 - 2026-10-26 #
For Users #
✨ Highlights
- A peer can no longer crash the application with nested extensions — GHSA-gc79-23g3-8g52, fixed upstream in openmls 0.9.1; every earlier release of this package is affected.
- A member an external commit removes no longer stays in the group — GHSA-94v4-rg25-x625, fixed upstream in openmls 0.9.2; every earlier release of this package is affected, and upgrading does not repair a group it already damaged; see Security.
- A malformed X-Wing key from a peer no longer panics — RUSTSEC-2026-0331, on the 0x004D suite only.
- The P-256 and P-384 suites carry an open constant-time advisory —
RUSTSEC-2026-0387 (
crypto-bigint0.5.5), accepted until OpenMLS moves top256/p3840.14; the other ten suites are unaffected. See Security. - Async calls no longer stop answering when another isolate shuts down — native platforms, also when an application initializes several flutter_rust_bridge packages, in any order; see Fixed.
- On the web,
Openmls.init()throws when the WASM module is not served — instead of waiting forever; see Fixed. hooksandcode_assets2.x are admitted beside 1.x — each Flutter resolves the major it supports; see Changed.- openmls v0.9.2 — two patch releases: 0.9.1 with the GHSA-gc79-23g3-8g52 fix, 0.9.2 with the GHSA-94v4-rg25-x625 and GHSA-p276-g3f3-wwgh fixes, and nothing else this package's API exposes.
- openmls_frb v2.3.2 — Rust FFI bindings
Changed
- The build hook uses a local build only when
make buildmade it from the current crate version and lock file (hook/build.dart,Makefile) — a library inrust/target/release/orrust/target/debug/took priority over the released one on the sole condition that it existed, so one left over from an earlier crate loaded silently, andrustContentHash, which hashes only the names of the bridged functions, could not notice a moved wire format.make buildnow writesrust/target/release/.crate-version, asmake build-webalready did beside the WASM module, and both leave.cargo-lockbeside the stamp, a copy of therust/Cargo.lockthe build was made from: an upstream update moves the dependencies but not the crate version, so a build from before it passed the version check and ran the previous upstream code. The hook refuses a missing or stale stamp, and a copy that differs fromrust/Cargo.lock, with aHookException, and no longer readsrust/target/debug/.make buildalso drops the hook's cached result for this package (.dart_tool/hooks_runner/openmls/), which keptdart teston a downloaded library after a host build. Only a local build is affected — the published package carries norust/target/— and one built before this change is refused untilmake build(ormake build-web) runs again. hooksandcode_assetsadmit 2.x as well as 1.x (pubspec.yaml) —^1.0.0held every application on 1.x, and pub.dev took 10 of its 160 points for it on 3.2.1 ("Support up-to-date dependencies"). The range is now>=1.0.0 <3.0.0, so each Flutter resolves the major it supports: Flutter 3.38 to 3.44 pinmetabelow the 1.19.0 thathooks2.1 needs, and an application on 3.38.4 resolveshooks2.0.2 withcode_assets1.2.1 (measured onexample/), while this package resolved on its own takeshooks2.2.0 withcode_assets2.1.0.hook/build.dartneeds no change: it analyzes clean and its tests pass with 2.2.0/2.1.0, 2.0.2/1.2.1 and 1.0.1/1.0.0, and the whole suite passes with 2.2.0/2.1.0 and with 1.0.1/1.0.0. A 3.x is a migration again, and Dependabot holds it back.
Security
- A peer can no longer crash the application with nested extensions
(GHSA-gc79-23g3-8g52) — fixed in openmls 0.9.1 (
rust/Cargo.toml,rust/Cargo.lock). openmls 0.9.0 decoded an extension's body before checking that its type is allowed where it appears, so aratchet_treeextension inside a leaf node was decoded in full, its leaf nodes and their extensions included, one stack frame deeper per level of nesting. Input nested deeply enough ran the stack out, which is not a panic, so noPanicExceptionstands in for it. In most environments the process ends; through a C FFI, as here, or in WebAssembly, the advisory says the outcome depends on the environment and linker configuration, at worst memory corruption. Neither was measured here. Every call that decodes a message,Welcome,GroupInfo, ratchet tree or key package from a peer was exposed, whatever the ciphersuite: incoming messages (processMessage,processMessageWithInspectand the synchronousmlsMessageExtractGroupId,mlsMessageExtractEpochandmlsMessageContentType), theWelcome,GroupInfoand ratchet tree of a join, aWelcomethat is only inspected (inspectWelcome,exportWelcomeSecret), and key packages (addMembers,proposeAdd,keyPackageLifetimeand the rest). 0.9.1 checks the type first, so the same input fails with an ordinary error. Where that error carries the decoder's message (ratchet trees, key packages, theGroupInfoof an external join, unencrypted handshake messages), it now names the extension type that is out of place, where 0.9.0 said "Found duplicate extensions". The fix is upstream's (advisory: high, CVSS 4.0 8.7, CWE-674). Every native library released before this one carries an affected openmls,openmls_frb2.3.1 (package 3.2.1) included;SECURITY.md(Known Limitations #14) covers them. - A member an external commit removes no longer stays in the group
(GHSA-94v4-rg25-x625) — fixed in openmls 0.9.2 (
rust/Cargo.toml,rust/Cargo.lock). When an external commit removed a member and the removal truncated the ratchet tree, adding the joiner's own leaf extended the tree again with the removed leaf, and some parent nodes, restored from before the commit. The committer and every member processing the commit computed the same tree, so nothing diverged, but the removed leaf kept its encryption key and later commits kept encrypting path secrets to it: a party that holds the removed client's state and keeps processing the group's handshake messages can derive the secrets of every later epoch until the leaf is removed again. In this package the commit is a resync —joinGroupExternalCommitandjoinGroupExternalCommitV2remove any member that has the joiner's signature key — or another client's external commit that includes a SelfRemove (leaveGroupViaSelfRemove) by reference, and every member that processes either withprocessMessageorprocessMessageWithInspectcomputes the same tree. Measured on 0.9.1, with members at leaves 0 to 3, leaf 2 removed and the member at leaf 3 rejoining by either call:groupMembersstill listed index 3, whichprocessMessageWithInspecthad reported inremoveIndices, and while two leaves shared a signature key no new member could join, by Welcome or by external commit. 0.9.2 leaves the regrown part of the tree blank, as RFC 9420 requires; the same steps list three members. Only groups that accept external commits are affected, and upgrading does not repair a tree the bug already wrote: a group stays affected until a member removes the restored leaf withremoveMembers, an ordinary commit every version accepts (SECURITY.md, Known Limitations #16). Per the advisory, an external commit that meets the bug is rejected between clients on 0.9.2 and clients on an affected openmls, this package up to 3.2.1 included, with a path length or confirmation tag mismatch, rather than leaving them on different trees (advisory: medium, CWE-459). - A malformed X-Wing key from a peer no longer panics (RUSTSEC-2026-0331)
— up to
openmls_frb2.3.1, an X-Wing (0x004D) public key shorter than 1184 bytes made the next encryption to it panic insidelibcrux-kem(0.0.9 fromopenmls_frb2.1.0, 0.0.7 before it, with the same defect): aPanicExceptionon native, and on the web an abort that left the database unusable until the page was reloaded. The key is peer data — the init key of a key package being added, another member's tree node, theGroupInfoof an external join. The native library now checks its length before libcrux sees it, and the operation fails with the ordinary error the fixedlibcrux-kem0.0.10 returns. Nothing leaked and nothing could be forged (CVSSVC:N/VI:N/VA:H), and no other suite touches libcrux. The crate itself stays at 0.0.9: OpenMLS pins it — 0.9.2 too — and the fix needs an OpenMLS release that does not exist yet. It is recorded in.cargo/audit.tomlandrust/deny.tomltogether with RUSTSEC-2026-0329 (libcrux-hmac-drbg) and RUSTSEC-2026-0330 (libcrux-kem), which nothing in the dependency graph can reach;SECURITY.md(Known Limitations #13) covers the releases that still panic. - The post-quantum suites' constant-time helper is fixed
(RUSTSEC-2026-0385) —
ctutils0.4.2 → 0.4.5 (rust/Cargo.lock). RustCrypto's ML-KEM decapsulation selects its result with it, and ML-DSA's signing arithmetic compares and selects with it, on secret values, so every post-quantum suite except X-Wing (0x004D, which runs on libcrux) relies on it. Before 0.4.4 its masks could compile to a branch on commonly used platforms, and itsu128comparisons did on 32-bit targets; a branch on a secret value can leak it through a side channel. 0.4.4 fixes both, and 0.4.5 also hardensChoicecomposition and Thumb-1 comparisons.openmls_frb2.1.0 to 2.3.1 carry 0.4.2; nothing was measured here (advisory). - The P-256 and P-384 suites carry an open constant-time advisory
(RUSTSEC-2026-0387) —
crypto-bigint0.5.5, in every release since 1.0.0, has operations meant to be constant-time that could compile to branches on secret values, depending on the compiler and CPU. Its fix, 0.7.6, cannot be taken:elliptic-curve0.13 requires 0.5, and only an OpenMLS release onp256/p3840.14 lifts that. Reviewed before it was accepted: P-256's scalar arithmetic is crypto-bigint's, so every signature onmls128DhkemP256Aes128GcmSha256P256(0x0002) andmls128Mlkem768Aes256GcmSha384P256computes with the private key and the nonce through its modular addition and subtraction, which turn a borrow into a mask with no optimization barrier — the construct the fixed release guards. Onmls192Mlkem1024Aes256GcmSha384P384the arithmetic is fiat-crypto, and crypto-bigint only range-checks, zero-checks and encodes the secrets; the other ten suites do not run it on secret data. Whether the compiled code branches was not measured. The advisory is accepted in.cargo/audit.tomlandrust/deny.toml, andSECURITY.md(Known Limitations #15) says which suites to prefer (advisory). - GHSA-p276-g3f3-wwgh did not reach this package — every openmls this
package has shipped advanced the claimed sender's ratchet before decrypting
a message's content and verifying its signature, and stored it before the
signature check, so a member could send a message in another member's name
that every receiver rejected and that still made the genuine sender's
messages — commits included — fail for the rest of the epoch. Here a failed
call writes nothing back: every operation works on a snapshot of the stored
state, which
processMessageandprocessMessageWithInspecthave dropped on any error since 1.0.0. Measured on 0.9.1 through both calls: after a message at generation 6 that decrypted and then failed signature verification, the sender's messages 0 to 5 all decrypted; withprocessMessagepatched to write the snapshot back on error, messages 0 and 1 were lost as too old. 0.9.2 fixes it upstream by applying the ratchet step only once a message is processed successfully (advisory: low, CVSS 3.1 3.7, CWE-460).
Fixed
- Async calls no longer stop answering when another isolate shuts down
(
lib/src/openmls.dart,lib/src/platform/platform_io.dart,lib/src/platform/platform_web.dart) — flutter_rust_bridge sends every Rust-to-Dart message (an async call's result, a store callback) through one process-wideDart_PostCObjectpointer per native library, and replaces it with a no-op when that library's count of live isolates drops to zero. An isolate missing from the count kept its sync calls, while every async call waited for an answer that had been dropped, and flutter_rust_bridge 2.13.0 leaves one out in two ways.RustLib.init()installs the pointer before it counts the new isolate, so the last other counted isolate shutting down in between left the no-op behind. And the watcher that counts an isolate is one per isolate for every flutter_rust_bridge library, so an isolate is counted only by the first library it initializes: when the main isolate had initialized another flutter_rust_bridge package first, a worker that used openmls alone — acompute()orIsolate.runworker included — took openmls's count to zero for the whole process, once the exited worker's watcher was garbage-collected.Openmls.init()now counts its isolate in openmls's own library beforeRustLib.init(), through flutter_rust_bridge'sfrb_create_shutdown_callback, which closes both; the order in which an application initializes its flutter_rust_bridge packages no longer matters for this one. Measured with a copy of the library initialized first in the other package's place: once a worker that used only openmls had exited and been collected, the main isolate's next async call went unanswered for 15 s with the previous code and was answered with this one; openmls alone, five workers in a row, was answered with both. The symbol is internal to flutter_rust_bridge, which this package pins to 2.13.0; a version that renamed it would makeinit()fail rather than hang. Two things stay open: a finalizer on another thread that has counted to zero and not yet written the no-op, which only flutter_rust_bridge can close, and another flutter_rust_bridge package without this change, which still needs the same initialization order in every isolate for itself. The web has neither the pointer nor the count. - On the web,
Openmls.init()throws when the WASM module is not served, instead of waiting forever (lib/src/openmls.dart,lib/src/platform/platform_web.dart) — flutter_rust_bridge loadspkg/openmls_frb.jsthrough a<script>tag and waits for itsloadevent with no error path or timeout, so a missing module — a 404, or the app's own HTML page that a development server sends in its place with a 200 — leftinit()pending with nothing reported.init()now fetches the script first and throws aStateErrornaming it, its HTTP status and content type, and the ways out, when the answer is not a success or is HTML. A request that fails outright, under a Content Security Policy that forbidsfetchsay, throws nothing, and the loader runs as before. Measured in headless Chrome: with no module served,init()threw at once withpkg/openmls_frb.js is not served (HTTP 404, text/plain; charset=utf-8); with it served,init()completed, and a sync call (supportedCiphersuites()) and an async one (MlsEngine.create) answered.
Documentation
- The README says when
web/pkg/can be left empty, and whatinit()then does (README.md, Known Limitations) — aflutter run -d chromethat skips the build hook after a run for another platform is Flutter 3.38 to 3.41's: 3.38.4 runs the hooks through thedart_buildtarget, whose stamp does not include the target platform, while 3.44.8 and 3.47.4 call them directly with the target platform (3.41 not checked here). Where it still happens,init()throws theStateErrorabove; the README no longer says the call stays pending. SECURITY.mdno longer says the tag ruleset stops awritecollaborator — it reserved tags to "Admins/Maintainers", but the ruleset's bypass list also held role 4, which GitHub nameswrite. The list now holds Admin only (.github/rulesets/protect-release-tags.json), and the text says that Admins bypass the ruleset's signature rule as well; the release scripts sign their tags.THIRD_PARTY_NOTICES.txtgives the openmls crates the canonical MIT text — the file ships in every native archive and in the package. With openmls taken from crates.io (see For Contributors), its seven openmls crates carry the MIT text withCopyright (c) OpenMLS Authors, composed from the crate metadata, because a published.cratehas no LICENSE file; until now it quoted the repository's LICENSE, which names the year 2020. The licence terms are unchanged.
For Contributors #
Added
- A regression test for GHSA-94v4-rg25-x625
(
test/group_lifecycle_test.dart) — a resync byjoinGroupExternalCommitand byjoinGroupExternalCommitV2, in a group where the removal it carries truncates the tree, must leave three members and no signature key twice. Both cases fail on openmls 0.9.1, which still lists the removed leaf.
Changed
-
copier template adopted: v4.15.1 -> v4.15.2 — workflow fixes.
.copier-answers.ymlrecords the new template revision.The agent workflows now use
anthropics/claude-code-actionv1.0.261 (.github/workflows/ai-review.yml,.github/workflows/repair-build.yml) — both pins move from v1.0.235: to v1.0.236 with this adoption, to v1.0.240 in Dependabot's #54 and to v1.0.261 with the v4.18.0 adoption below, bringing the generated workflows onto the newer bundled Claude Code and Agent SDK versions.The notices refresh now supplies the repository to its signed commit (
.github/workflows/refresh-notices.yml) — theCommit and pushstep'senv:gainsGH_REPO: ${{ github.repository }}, so the GraphQL mutation and itsverification.verifiedread-back receive a repository instead of failing when$GH_REPOis unset. The obsoleteAPP_SLUGsetting is removed because that step no longer uses it for the Git identity or bot lookup.Large notices inventories now reach
jqwithout exceeding Linux's argument limit (.github/workflows/refresh-notices.yml) — the base64 content is written to$RUNNER_TEMP/notices.b64and passed with--rawfileinstead of being expanded into one--argargument. This lets the signed commit payload handle inventories whose base64 representation is larger than the per-argument limit. -
The App token is kept out of the update-checker checkouts (
.github/workflows/check-openmls-updates.yml,.github/workflows/check-template-updates.yml) — both jobs mint a write-scoped GitHub App token and then run repository code and third-party tooling (make,pip install copier) in the same job. Theactions/checkoutstep no longer receives that token; it uses the job's default, read-scopedGITHUB_TOKEN(the jobs grantcontents: read), andpersist-credentials: falsekeeps even that out of.git/config, so nothing in the worktree can read a write-scoped credential.GITHUB_TOKENis enough to check the repository out whether it is public or private. The pull request is still pushed bypeter-evans/create-pull-requestvia its App-token input, and theghsteps read the App token fromGH_TOKEN. Mirrorsrefresh-notices.yml. -
The update checkers can regenerate their own pull requests (
.github/rulesets/signing-commit.json,.github/rulesets/README.md) —peter-evans/create-pull-requestrefreshes an existing branch by force-pushing it, andSigning commitforbids that on~ALLbranches with no bypass actors, soforce_updatecould not regenerate an open update pull request: measured on 2026-10-20 (run39450303118), the App's signed API commit was created and the ref update refused with "Cannot force-push to this branch". A scheduled run fails the same way when a version's branch outlives its pull request.refs/heads/update-openmls-*andrefs/heads/update-template-v*are now excluded from that ruleset, asrefs/heads/dependabot/**/*already is: a push there may rewrite history and need not be signed, and the App signs regardless.mainkeeps its pull-request gate, its signatures and its no-bypass ban on force-pushes, and the branches stay underDelete branches. -
copier template adopted: v4.15.2 -> v4.17.0 — by hand, superseding the update bot's #55 (v4.16.1).
.copier-answers.ymlrecords the new revision and the new question's answer,upstream_source: git_tag; switching the source is a change of its own.pub.dev's upload limits are measured (
make verify-pub-limits) — pub.dev refuses a README.md, CHANGELOG.md, LICENSE or example over 256 KiB, or a pubspec.yaml over 128 KiB, at upload, after the tag, and the dry-run checks none of them. CI runs it on the Linux x86_64 leg and warns from 200 KiB,make publish-dry-runruns it first, andmake releasemeasures the finalized CHANGELOG before committing.iOS is type-checked on every push and pull request (
make rust-check-ios, theType-check (iOS)job) — the three iOS targets were compiled only by a release tag, so an iOS-only compile error first showed up there.make releaserefuses a stage-1 release whose native sources moved since its tag —rust/(exceptrust/fuzzandrust/deny.toml) andlib/src/rust/are compared with theopenmls_frb-*tag, so any commit there after stage 1 needs a new crate.Every Ubuntu job names its image,
ubuntu-26.04— the Linux release builds move fromubuntu-latest(x86_64) andubuntu-24.04-arm(arm64) toubuntu-26.04andubuntu-26.04-arm; a Linux library records the glibc symbol versions of the image that built it.An upstream bump moves only what the new release requires —
make rust-updateiscargo update --workspace, and the rest of the graph moves in Dependabot's weekly cargo pull request, which now takes transitive crates too (allow: dependency-type: all).Only an Admin can create, move or delete a tag — see Documentation above; on GitHub the ruleset changes when
make setup-repo-protections ARGS="--update"applies it.The agent workflows — OpenCode is pinned at 1.18.35 and makes no session-title call unless
AGENT_OPENCODE_TITLE_MODELnames a model; the AI reviewer is told where each file's hunks start and who wrote each commit; the build repair leaves drafts alone, re-runs a first failure once, leaves a commit to a person after two runs that ended without a verdict, and can run the wasm32 half ofmake rust-doc; the update bot's CHANGELOG entry is written from the lockfile diff and the bound upstream files' patches.make getpassesARGSon, somake get ARGS="--directory=example_cli"resolves that package.The App-token checkout hardening and the update checkers' branch exclusion, both above, now come from the template.
-
openmls comes from crates.io instead of a git tag (
rust/Cargo.toml,rust/Cargo.lock,rust/deny.toml,.github/dependabot.yml,.copier-answers.ymland the files theupstream_source: crates_ioanswer renders) — the update check found upstream releases through GitHub's latest release and accepted only anopenmls-vX.Y.Ztag, so openmls 0.9.1, published on 2026-10-07 under the tagv0.9.1and with no GitHub release, was invisible to it.openmlsis now pinned exactly ("=X.Y.Z") on both of its declarations, the four other openmls crates sit at the requirements that release declares, and the check reads crates.io. The switch, made on 0.9.0 before the bump in Security, moved no code: all nine openmls packages in the lockfile were published from the commit theopenmls-v0.9.0tag names (3a3e35de), and only theirsourceandchecksumlines moved.rust/deny.tomlno longer allows the openmls git repository, and Dependabot ignores the four openmls crates that arrive only through the others, so they move with the openmls release that requires them. -
copier template adopted: v4.17.0 -> v4.18.0 — by hand.
.copier-answers.ymlrecords the new revision and the new question's answer,enable_upstream_advisories: true. Three files conflicted, each resolved to the template's side: theclaude-code-actionpin in both agent workflows and a comment in.github/dependabot.yml. What a consumer observes is under For Users: the isolate count ininit(), the web module check, thehooks/code_assetsrange and the build hook's lock-file check.openmls/openmls's own security advisories are checked every day (
make check-upstream-advisories, a job ofcheck-openmls-updates.yml) —make rust-auditandmake rust-denyread RustSec and Dependabot reads GitHub's global database, and GHSA-gc79-23g3-8g52 reached none of them. The check reads the repository's published advisories and holds them against every versionrust/Cargo.lockresolves; on this tree none of the seven covers one.make setup-repo-protectionsconfigures both release environments —native-buildandpub.devget the owner as the one required reviewer, no admin bypass, and a deployment policy that admits onlyopenmls_frb-*andv*tags;publish.yml'spublishjob now runs on a tag only. GitHub changes when the command runs:--dry-runon this repository showsnative-buildletting an Admin skip the review and both environments admitting any ref.The minimum Rust version is checked against the dependency graph (
make check-msrv, on the Linux x86_64 leg) — 1.91 everywhere it is stated, the floor openmls 0.9.2 declares.cargo-deny fails on a yanked crate version and on a wildcard dependency (
rust/deny.toml) —yanked = "deny", and with every crate from crates.iowildcards = "deny", path dependencies allowed;make rust-denypasses.The update tooling —
make update-template ARGS="--version vX.Y.Z --data key=value"answers a question a new version adds, and from the next adoption on reports text a merge put in twice (none in this one: every merged file was compared with a fresh render of both versions); andmake update-evidencewrites down what an upstream update changed, attached to the update pull request, which now also says whether it could be merged unread, as a report only. The build repair reads the openmls crates' source from crates.io, takes the update it repairs from the compare API, may writeCHANGELOG.md,README.mdandSECURITY.md, and asks what moved behind a changed signature; the AI reviewer holds a bot entry's upstream claims against the upstream range; the update bot replaces the upstream Highlights line whoever last wrote it.The macOS and Windows jobs name their images,
macos-26andwindows-2025-vs2026— the release builds included, and the images-latestalready gave them:openmls_frb2.3.1 was built onmacos-26-arm64andwindows-2025-vs2026.Swatinem/rust-cachemoves to v2.9.3 — andclaude-code-actionto v1.0.261 (see the v4.15.2 entry above); both upstream tags dereferenced to the pinned commits. This supersedes Dependabot's #71.Dependabot — the
coverage >= 1.16.0andffigen-major ignores now come from the template, which took them from this repository: its text replaces ours and givesffigena reason of its own, flutter_rust_bridge'sfrb_dartpinning^20.1.1, where ours named hooks 2. Thehooks/code_assetsignore now holds back 3.x.Smaller:
compareVersionsorders prereleases and build metadata as SemVer does; the update-openmls skill no longer describes a crate bump the update never makes; README and CONTRIBUTING say what the update pull request changes;protect-main.jsonnamesrequire_extra_approval_for_unattributed_changes: true, which the live ruleset already holds.
3.2.1 - 2026-09-29 #
For Users #
✨ Highlights
- A failed start on the Web now says why, instead of spinning forever — the example app reports the error and names the usual cause and its fix.
- The
--wasmlimitation now tracks the pull request that fixes it — and says where that fix is: a prerelease, so the limitation still stands here. - openmls v0.9.0 — unchanged this release
- openmls_frb v2.3.1 — Rust FFI bindings
Fixed
-
The example app now reports an initialization failure instead of spinning forever (
example/lib/main.dart) —_initOpenmls()is fire-and-forget frominitState()and caught nothing, so anyOpenmls.init()throw left_isInitializedfalse with the progress indicator running and the exception reaching the console and nowhere else. A hang that names no cause is a much longer diagnosis than the one-line problem behind it.The body now branches three ways — initialized, failed, neither — and the failure path renders the error. The raw message is kept on purpose: any
init()failure lands there, not only the missing-WASM one the hint names, so a message asserting the cause would be wrong as often as right. The hint names the usual cause on web and the command that fixes it.⚠ This does not fix the underlying reason
init()fails on web, which is not fixable from this package:flutter run -d chromeafter a run for another platform reuses that run'sdart_buildstamp — the build directory key does not include the target platform — and skips the build hook abovehooks_runner, soweb/pkg/is never provisioned. Known Limitations names the escape. What changes here is that the app says so.
Documentation
-
The
--wasmlimitation now names the pull request that fixes it, instead of a stale issue about something else (README.md) — Known Limitations tracked flutter_rust_bridge#2575, which was closed in July 2025 and describes a different failure (aDartFnFuturecallback crash under Chrome), not the decoder cast this package actually hits. It now tracks flutter_rust_bridge#3182, the pull request that replaces that cast, merged on 2026-09-05. The quoted error is the verbatim one as well, so that searching for what the console prints finds this section.The limitation still applies and was re-measured rather than assumed. What is new for a reader is where the fix is and why waiting is the only option: it is in
flutter_rust_bridge2.14.0-beta.2, a prerelease, and this package can neither ship a stable release that depends on a prerelease nor be rescued by adependency_overridein the host app, because its own generated bindings carry the same cast. The section now says so.
For Contributors #
Added
-
The resolved cargo feature graph is now a gate, not a comment (
scripts/verify_feature_graph.dart,scripts/src/feature_graph.dart,Makefile,.github/workflows/test-reusable.yml,test/scripts/feature_graph_test.dart,test/scripts/feature_graph_policy_test.dart,CLAUDE.md) —rust/Cargo.tomlcarries a comment explaining thatopenmls'stest-utilsmust stay absent from a shipped binary, because it impliesbacktraceandLibraryError::custom()then formats a symbolized Rust backtrace — build-machine paths, symbol names, crate layout — into an error that travels the ordinary error channel out to the Dart caller, no panic required. Nothing enforced it.A manifest comment cannot: cargo unifies features across the graph, so the feature can be switched on for a crate this manifest never mentions, by a dependency arriving transitively in a version bump nobody here reviewed. The check therefore reads
cargo tree— what openmls is built with — rather than the manifest text.⚠ It is keyed by crate, and a flat list of feature names would be red on a healthy tree. Measured rather than assumed:
openmls_basic_credentialcarriestest-utilsdeliberately — it is what makesSignatureKeyPair::private()reachable, and that crate'stest-utilsimplies no backtrace — andallo-isolate, which arrives underflutter_rust_bridge, enables abacktracefeature of its own. Both sit in the test fixture for exactly that reason.backtraceis forbidden alongsidetest-utilsbecause upstream allows enabling it on its own, which is a shorter path to the same leak.Absence is a failure rather than a pass: if a crate named in the rules is missing from the graph, the check fails instead of reporting clean, so a renamed or dropped dependency cannot silently retire it.
The mechanism has since moved into the copier template, so the two scripts and the mechanism test are now byte-identical to what it renders and a future adoption will not conflict over them. What stays is the part that is this project's rather than the template's: the rules —
openmlsmust carry neithertest-utilsnorbacktrace— andfeature_graph_policy_test.dart, whose fixture is the real graph (openmls_basic_credentialcarryingtest-utilson purpose,allo-isolatecarrying its ownbacktrace) and therefore the evidence that a flat list of feature names would be red on a healthy tree.CLAUDE.mdgains a## Feature Graph Gatesection carrying what the code cannot say: why the rules are keyed by crate, and why--manifest-pathis right here althoughmake rust-clippy-webrequires the opposite. It is a section of its own rather than lines inside the existing blocks, because that file is template-owned — an edit inside a rendered block conflicts at every adoption, while a section the template knows nothing about does not.--edges normal(the graph that ends up inside the shipped library),--target all(a clean union means every target is clean) and--locked(a property of the revision, not of the machine). It resolves without compiling and runs inside the already-required Linux x86_64 test job, so it adds no required context. Verified red as well as green: withjsadded to the rule set the gate fails and names the crate, the feature, the enabled set and thecargo tree --invertcommand that finds the culprit.
Changed
-
copier template adopted: v4.15.0 -> v4.15.1 — the notices refresh now writes signed commits through GitHub's GraphQL API instead of pushing an unsigned commit.
.copier-answers.ymlrecords the new template revision.The notices refresh now creates a mergeable commit (
.github/workflows/refresh-notices.yml) — it previously used an ordinarygit commitandgit pushon Dependabot cargo branches. The exclusion forrefs/heads/dependabot/**/*insigning-commit.jsonallows that push, but it does not make the resulting pull request mergeable intomain, whererequired_signaturesstill rejects the unsigned workflow commit. An approval therefore cannot unblock the pull request.The workflow now uses GraphQL
createCommitOnBranch, which signs the bytes it writes, and passesexpectedHeadOidso a concurrent push causes the mutation to fail instead of silently reverting newer work. This also removes thegit configsetup and/users/{bot}lookup that were needed only to imitate the commit attribution; the mutation credits the token's owner.The workflow reads
verification.verifiedfrom the new commit and fails if it is nottrue. A signing failure is therefore reported during the refresh instead of being left for the pull request's merge check to discover. -
copier template adopted: v4.14.1 -> v4.15.0 — superseded template-update pull requests can now be closed safely, and the generated workflows and documentation catch up with several template fixes.
.copier-answers.ymlrecords the new template revision. The resolved feature-graph gate and its supporting scripts produce no diff in this project; that machinery was already present here.Older template-update pull requests are now managed automatically (
.github/workflows/check-template-updates.yml) — the workflow identifies olderupdate-template-vX.Y.Zbranches and closes their pull requests during scheduled runs, while manual runs default to a dry run. It leaves a pull request open when the branch is not an older template-update branch, carriesagent-repaired, has a commit with arepair-build:trailer, or contains a human-authored commit. Those vetoes matter because a newer update supersedes an untouched older one, but must not discard a branch whose conflicts or repair work a human has addressed.The agent workflows now read their configuration from the default branch (
.github/workflows/repair-build.yml,.github/workflows/ai-review.yml) —OPENCODE_CONFIGpreviously pointed into the checkout being repaired or reviewed, while the workflow and its expected permissions came from another revision. Both workflows now usegh apito read the default-branch config into$RUNNER_TEMP, so a pull request cannot supply the permissions that govern the agent judging it. The repair workflow also reports a deleted update branch as missing instead of displaying the 404 response as a branch head.The generated documentation and ignore rules now describe the current build behavior (
CLAUDE.md,.gitignore) —CLAUDE.mddocuments the.crate-versionstamp required by local Web builds, why an unstamped or stale WASM directory is rejected, and how to rebuild it..gitignorenow excludes Python__pycache__directories, preventing local release-gate runs from adding generated bytecode to commits.The
setup-androidcomments in.github/workflows/test-reusable.ymland.github/workflows/build-openmls.ymlnow match thev4.0.4pin and explain why the explicitpackages: platform-toolsinput remains. The OpenCode setup is still duplicated between the two agent workflows; this adoption records why a shared local composite action is not used, rather than changing that behavior. -
copier template adopted: v4.14.0 -> v4.14.1 —
make build-webnow stamps local WASM builds with the Rust crate version, andhook/build.dartrejects a missing or stale stamp instead of serving an older module.Makefilewritesrust/target/wasm32/.crate-versionfrom the version inrust/Cargo.toml. The web build hook compares that stamp with the current crate before it copies a local module intoweb/pkg/; a directory built before this change has no stamp and is rejected. This matters because a localrust/target/wasm32/build can otherwise outlive a crate bump while still satisfying the old existence-only check. The stamp is also a build dependency, so changing it invalidates the cached hook result.test/hook/build_hook_test.dartadds the regression cases for the missing and mismatched version stamps..copier-answers.ymlrecords the new template revision.The workflow changes in v4.14.1 are byte-identical in this project and produce no diff: the
actions/checkout@v7update in.github/workflows/release.ymland theanthropics/claude-code-actionandandroid-actions/setup-androidpins in the generated workflow files were already present here.
3.2.0 - 2026-09-20 #
For Users #
✨ Highlights
- Past epoch message secrets are now yours to keep and to clear — seven engine methods read and set how many past epochs a group keeps, and delete what is kept. Retaining them is a forward-secrecy trade-off, not a tuning knob; the default still keeps none.
- A second browser tab can no longer overwrite the first — every operation on the Web runs under an exclusive cross-tab lock, so the tabs and workers of an origin queue instead of dropping each other's writes.
- openmls v0.9.0 — unchanged this release
- openmls_frb v2.3.0 — Rust FFI bindings
Added
-
Past epoch message secrets can now be retained and deleted on demand (
rust/src/api/engine.rs) —pastEpochDeletionPolicyreads how many past epochs a group keeps,setPastEpochDeletionPolicyMaxEpochsandsetPastEpochDeletionPolicyKeepAllset it, and fourdelete…methods remove what is kept:deleteAllPastEpochSecrets,deletePastEpochSecretsOlderThan,deletePastEpochSecretsBeforeanddeletePastEpochSecretsWithoutTimestamps.An application message is encrypted under the epoch its sender was in, so a message still in flight when a commit advances the group can only be read from that epoch's secrets.
MlsGroupConfig.maxPastEpochscould already keep a fixed number of them, but only at creation, and it cannot express "keep all" at any number — nor could anything delete them before their turn came.⚠ Retaining them is a forward-secrecy trade-off rather than a tuning knob: the secrets of a past epoch decrypt every message of that epoch, including traffic recorded earlier, for as long as they are stored. The default keeps none, and
keepAllkeeps them forever unless the application deletes them — which is what the fourdelete…methods are for. SECURITY.md says so where it discusses group state.⚠
setConfigurationresets the policy, becauseMlsGroupConfigcarries the same setting asmaxPastEpochsand the struct has no "leave as is". Set the policy after the configuration, not before. Documented on the setter, onsetConfigurationand on the field, and pinned by a test.deleteAllPastEpochSecretsdeliberately takes nomaxPastEpochscap, unlike the other three: OpenMLS applies that modifier only to the selective deletions, so "delete everything but keep a few" would delete everything. The combination is absent from the API rather than documented as a trap. It is also a sweep rather than a switch — underkeepAllthe next commit starts recording past epochs again, so stopping accumulation means setting amaxEpochspolicy, not deleting.setPastEpochDeletionPolicyMaxEpochsrefuses 4294967295 and nameskeepAllin the error. That single value is what OpenMLS writes to mean keep-all whereusizeis 32 bits — the Web and 32-bit Android — so accepting it would make two settings one stored value there and two everywhere else.⚠ Secrets recorded by a version of this package built on OpenMLS 0.8.1 or earlier (that is, before 3.0.0) carry no timestamp, and the two time-based deletions skip them in silence.
deletePastEpochSecretsWithoutTimestampsis the step that clears them. It applies only wheremaxPastEpochswas above zero back then, since the default recorded no past epochs at all. No database migration is involved: OpenMLS reads the older encoding through a serde alias and a defaulted field, and the schema version does not move.Shown by execution on both platforms, because the storage encoding of keep-all is written in terms of
usize::MAXand therefore differs by target. Measured in a browser: a stored keep-all comes back from IndexedDB asMaxEpochs(4294967295)on wasm32 and is normalised so that Dart sees the same answer everywhere. 17 Dart tests cover the native path, and five Rust tests the conversions, one of them driving a real IndexedDB group. Negative controls: removing the normalisation fails three browser tests while the whole Dart suite stays green, and dropping themaxPastEpochscap from its test makes a deleted epoch decrypt again.
Security
-
Every operation on the Web now takes a cross-tab lock (
rust/src/web_lock.rs,rust/src/api/engine.rs) — a second browser tab was an unguarded second writer to the same IndexedDB database. Each tab runs its own WASM instance with its own engine, so the engine-wide mutex that serializes overlapping calls could not see it: two tabs could load the same snapshot, operate on it, and write back in turn, and the later write-back would drop what the first had saved — a merged commit, an advanced ratchet, a stored proposal — desynchronizing the group.Every load → operate → save cycle now runs under an exclusive Web Lock named after the IndexedDB database, so the tabs and workers of an origin queue instead of overwriting each other. They are not refused the way a second native process is, because a tab is opened and closed by the person using the application rather than by the application: a contending operation waits, for at most five seconds, and then fails with "Database is busy" rather than waiting behind a wedged tab forever. That failure is retryable and leaves nothing behind: the lock is taken before any snapshot is read, so an operation that cannot get in has loaded no state and released the engine-wide mutex on its way out.
⚠ The guarantee needs a secure context, which is where
navigator.locksexists at all:https, orhttponlocalhost. Served over plainhttpfrom any other host the API is absent, and operations run exactly as they did before — without the cross-tab guarantee, never with an error. The degradation is deliberate: the alternative is failing every operation on an origin where the Web build used to work.⚠ A panic inside a guarded operation does not release the lock. wasm32 aborts on panic, so
Dropnever runs, while the instance itself keeps working — so every later operation on that database reports "Database is busy" until the page is reloaded. The escape is a reload, not a retry. Native unwinds and releases.Native is unchanged and was never exposed: its sidecar lock file has refused a second opener outright since 2.0.0. This closes the same hole on the one platform whose processes are tabs.
Shown by execution rather than by construction, which on wasm32 is the only kind of evidence that counts — a wasm32 body is a different implementation of the same function, so nothing that runs on the host touches a line of it. Three browser tests (
make test-web) cover exclusivity, the busy error on contention, release on drop, independence of unrelated database names, and the absent-API fallback.
For Contributors #
Added
-
Refresh Dependency NoticesregeneratesTHIRD_PARTY_NOTICES.txton Dependabot's cargo pull requests (.github/workflows/refresh-notices.yml) — the inventory is generated from the resolved cargo graph, and Dependabot editsrust/Cargo.tomlandrust/Cargo.lockwith no way to runmake third-party-noticesafterwards, so every cargo pull request arrived with a stale inventory and failedverify-third-party-notices.That was cosmetic until the ruleset in v4.10.0 required the whole matrix. The step runs
if: matrix.name == 'Linux x86_64', so it fails the job reported astest / Test (Linux x86_64)— now one of twelve required contexts. The same redness that used to be ignorable makes every cargo pull request unmergeable by anyone without a ruleset bypass. The fix follows its cause by one day.It is a separate workflow rather than a step in
test.ymlbecause a workflow triggered by a Dependabot pull request gets a read-onlyGITHUB_TOKENand no Actions secrets — measured, and this repository has no Dependabot secrets either, so such a job can mint no App token and push nothing.scheduleplusworkflow_dispatchrun in the base repository's context instead;pull_request_targetwas rejected, since it would hand a write token to a job checking out pull-request content.An ordinary unsigned
git pushis what lands the commit, and that is only legal becausesigning-commit.jsonexcludes exactlyrefs/heads/dependabot/**/*from the signature requirement — written down in the workflow, because if that exclusion is ever narrowed the commit has to go through the GitHub API instead. Scoped todependabot/cargo/**and to Dependabot's own authorship, checked twice: once when selecting pull requests and again in the job holding the token, since a branch name is not a credential and anyone who can push can createdependabot/cargo/x.The work is split across two jobs for the same reason
repair-build.ymlis, and the author check is what makes the split necessary rather than optional: it proves Dependabot opened the pull request, not that nobody pushed to its branch afterwards — and the signing exclusion above means such a push need not even be signed. So the job that runs branch content — theMakefile, the scripts, and the composite actionsuses: ./…resolves out of the workspace — holds no credential at all and checks out withpersist-credentials: false; the job holding the App token runs nothing from the tree, taking the one file back as an artifact and refusing to commit if anything else in the working tree moved. Checking untrusted files out is safe; executing them beside a credential is not.The revision the inventory was generated from travels with it, because the two jobs check the branch out at different moments and only this workflow's own runs are serialised — nothing stops Dependabot moving the branch in between. The commit job compares, and steps aside when they differ rather than committing an inventory for a lockfile the branch no longer has, which would leave the gate red for exactly as long as it looked fixed.
That hole was found by this repository's own automated reviewer, on the pull request adding the workflow — the first finding of its kind it has produced here that survived checking.
Two things were measured rather than assumed.
.author.loginisapp/dependabotthroughgh --json, notdependabot[bot]as the REST API spells it, so the obviousstartswith("dependabot")predicate selects nothing — for ever, while staying green; the check normalises both forms. It was exercised while being written against seven hand-built cases, including an impostor author on adependabot/cargo/branch, a closed pull request and the two non-cargo ecosystems, with the jq program extracted from the workflow so the exercise could not drift from what runs. That was a measurement taken during development, not a committed test: nothing in the repository reruns it, so a later edit to the predicate is guarded by review alone.And the generator reads licence files out of the cargo registry without fetching them, so the job fetches explicitly: nothing else in it compiles anything, which is why
test-reusable.ymlnever needed the step and this would otherwise have been a laptop-only pass.⚠ Pushing to a Dependabot branch stops Dependabot rebasing it. That is wanted here — a rebase would discard the notices commit — but it means a stale pull request needs a human, and
@dependabot rebaseis the wrong repair.
Changed
-
copier template adopted: v4.13.0 -> v4.14.0 (18 files, plus this entry) — three new gates arrive, each covering something nothing here was checking: the wasm32 half of the crate is now linted, every native library is loaded before it is shipped, and every release archive is checked against what its name claims.
make rust-clippy-webis the lint gate, and.github/workflows/test-reusable.ymlruns it on the Linux x86_64 leg. It is a second gate rather than a thoroughness setting:make rust-clippyruns under the host target, and acfg(target_arch = "wasm32")body is a different implementation of the same function, so the host pass reads none of its lines while reporting green over the file that contains them. The target usescd rustrather than--manifest-path, because cargo discovers.cargo/config.tomlrelative to the working directory and the repository root drops the wasm32 rustflags. The other half of this — the crate actually being clean under that lint — landed separately and is recorded above.scripts/verify_library_loads.pyloads each built library before it is packed.ctypes.CDLLis the whole check and it is more than it looks: it runs the library's initialisers, resolves its dependencies and refuses a wrong architecture. It then looks upfrb_init_frb_dart_api_dl, which flutter_rust_bridge exports from its own binding rather than from generated code, so a library without it is not an FRB library whatever else it is. It runs only on the legs whose runner IS the target — both Linux architectures, macOS arm64 and Windows — since a cross-compiled artefact cannot be loaded by the machine that built it. ⚠ Upstream measuredctypes.CDLLon macOS only; the Linux and Windows legs are inference until this repository's first native build runs them.scripts/verify_release_artifacts.pyrefuses an archive that does not hold what its name says, and runs at both points that can get it wrong: averify-artifactsjob over what the build jobs produced, placed beforecreate-releaseand outside its environment so a bad build is caught without spending a reviewer's approval, and again insidecreate-releaseover the packed archives, before the provenance attestation signs them. ⚠ Checking the architecture would not have caught the likely error: a Linux arm64 and an Android arm64.socarry the same ELF header, and a macOS, an iOS and an iOS-simulator.dylibthe same Mach-O cputype — exactly the pairs a copy-paste slip in the hand-writtentarlist produces. It parses headers instead of callingfile, reading the libc inDT_NEEDED(glibc asks forlibc.so.6, Bionic forlibc.soplusliblog.so) and the platform field ofLC_BUILD_VERSION.Measured here against this project's own published release rather than taken on trust:
make verify-release-artifactsover all twelve archives ofopenmls_frb-2.2.0passes, which is the same callcreate-releasemakes on the archives it packs. That also settles the question the script cannot answer about itself — its platform table is rendered from the copier answers, and nothing had checked that the rendered table matches the archive names THIS repository publishes. It does, and the discriminators do real work on them:linux-arm64reads as glibc andandroid-arm64-v8aas bionic at the samee_machine, andmacos-arm64as macOS againstios-device-arm64as iOS at the same Mach-O cputype..github/workflows/repair-build.ymland.github/agent-prompts/repair-build.mdextend the repair agent to open bot-authored pull requests, committing to their own branches. A red update branch never reachesmain, so nothing watchingmainwould ever see it. Most such branches never reach a model: the commonest failure is that the update workflow could not run the binding generator and left the branch labelledcodegen-failed, which the repair simply runs and clears. Three labels carry state, andagent-repairedis a contract — the repair commits with the same App token as the branch's own commits, so nothing else can tell a branch an agent has worked on from one it has not, and a branch carrying it must not be closed automatically.scripts/src/update_changelog.dartnow returns aChangelogUpdaterather than a model, carrying the one fact only that run can observe: whether[Unreleased]was left naming two upstream versions because a rewritten Highlights line was standing.ciOutputsForemits both keys on both outcomes, since a key that appears only when true cannot be told apart from a script too old to emit it..github/workflows/check-openmls-updates.ymlsurfaces it as a numbered step in the pull-request body. The generator also anchors a created#### Changedon the first####heading, so it no longer lands above#### Added, and the changelog prompt gains two further rules: no attributing a change by the list of files it touched, and no scope file telling a model with no tools to go and fetch something.Smaller:
.github/dependabot.ymlrecords thatincrease-if-necessaryholds only the neighbours of a capped pin and returns the bumped one as a caret;.github/actionlint.yamlcarries a second suppression for the same actionlint bug;CLAUDE.md,CONTRIBUTING.mdand.claude/skills/update-openmls/SKILL.mdfollow..copier-answers.ymlrecords template versionv4.14.0..github/agent-prompts/changelog-scope.mdis updated by hand, because it is the one thing in this release copier structurally cannot deliver: the template writes that file once and never overwrites it, so guidance v4.14.0 adds to its source reaches an existing project only if somebody applies it. The guidance asks that the file name a dependency living in a DIFFERENT repository that still reaches users, and this package has one — the libcrux crates arrive from crates.io throughhpke-rs-libcrux, at versions openmls pins, and are named nowhere in the bound list. They can reach the X-Wing ciphersuite and nothing else. ⚠ The compare the changelog prompt is given covers the openmls repository alone, where such a change leaves one trace — a version number in a manifest — so an entry reports the version move and stops rather than inferring what changed inside. The file was also checked for the other thing that guidance forbids, an instruction telling a model with no tools to go and fetch something; it carries none.Three files conflicted and were resolved by hand.
CONTRIBUTING.mdtook the template'smake rust-clippy-webrow into the commands table this file already maintains, rather than the second table copier tried to place inside the release section; the gate is also named in the pre-push list and the pull request checklist, since it blocks in CI..github/workflows/check-openmls-updates.ymlkept both sides — the template's stacked-highlights step and this project's database-migration step.CHANGELOG.mdconflicted in two places and kept this project's side at both. The first is the file's head: this one has begun at## [Unreleased]since the project was generated, with no title and no Keep a Changelog blurb, so the paragraph v4.14.0 adds above that heading has no context to attach to and is not adopted. Three of the four projects generated from this template start the same way, andCLAUDE.mdalready carries what that paragraph says, at more length. The second is the template's seeded initial release, which this project's real## [1.0.0]section supersedes. The released## [1.0.1]section is left as it shipped: copier had rewritten its headings into the audience split by position, matching its own seeded release rather than anything about this project, and released sections do not change. -
The wasm32 half of the crate is now clean under clippy (
rust/src/encrypted_db.rs) —make rust-clippyruns under the host target only, so everycfg(target_arch = "wasm32")body in the crate had never been linted, here or in CI. A wasm32 body is a different implementation of the same function, so nothing the host lints touches a line of it.Measured with
--target wasm32-unknown-unknown --all-targets -- -D warnings: one finding, acollapsible_ifin the IndexedDB structural migration, now collapsed. Nothing changes at runtime — the two conditions were already evaluated one after the other — but a lint gate for that target cannot be added while the crate fails it, so this is the half of that gate which belongs to this repository rather than to the template that owns the Makefile. -
copier template adopted: v4.12.0 -> v4.13.0 — the generated workflows now use the current Claude action and no longer request the deleted Android SDK Tools package, while the changelog generator gains stricter upstream and highlight handling.
.github/workflows/ai-review.ymland.github/workflows/repair-build.ymlmoveanthropics/claude-code-actionfromv1.0.216tov1.0.222. The pin is updated in both callers so review and automated repair do not drift onto different action releases..github/workflows/test-reusable.ymland.github/workflows/build-openmls.ymlnow passpackages: platform-toolstoandroid-actions/setup-android. Its default also askssdkmanagerfor the obsoletetoolspackage, which Google removed; leaving that name in place fails the Android job before it reaches the pinned NDK installation. The NDK step remains responsible for installing the version the builds need.scripts/src/update_changelog.dartnow normalises a model-supplied list marker before writing a Highlights line, and shares the default highlight text between the prompt and the logic that recognises and supersedes it. Rewritten, multi-line highlights are left intact, while an old generated default can be replaced so successive dependency bumps do not accumulate contradictory native-library versions. The prompt also distinguishes the fetched upstream inputs from facts about a release, preventing an absent release body from being reported as a release-level claim.When release notes are empty, the generator makes a best-effort request for the upstream repository's
RELEASE_NOTES.md. It accepts that fallback only when its first line names the requested tag, avoiding reuse of a file left over from an earlier release. The placeholder for missing notes is now a shared constant, so this case can be handled consistently.The upstream dependency summary now uses the compare response's commits and changed files together.
scripts/src/update_changelog.dartseparates those two inputs and tells the model whether the file list is complete: GitHub's 300-file limit and the generator's 12000-character prompt limit are treated as truncation rather than evidence that no other files changed. With a complete list, the prompt asks for the actual files changed by the bound crates instead of relying on a generic out-of-scope sentence.test/scripts/update_changelog_test.dartadds coverage for these cases, including marker normalisation, release-notes fallback and tag checking, default-versus-rewritten highlight handling, and complete versus truncated upstream file lists..copier-answers.ymlrecords template versionv4.13.0. -
copier template adopted: v4.10.0 -> v4.12.0 — two template releases at once: the notices workflow this project contributed comes back carrying the edits the template made to it, and the dependency-bump changelog prompt is rewritten.
.github/workflows/refresh-notices.ymlis not byte-identical to the copy that went up.actions/upload-artifactmoves v4 -> v7 andactions/download-artifactv4 -> v8; those two lines are the only difference that can change what the job does, and they put the file on the pins every other workflow here already uses, clearing a Node 20 deprecation warning. Measured before moving them: the extraction-layout break indownload-artifactv5 applies to downloads by ID, this one downloads by name, and a v7 -> v8 round trip by name read from a flat path is already green in this repository —ai-review.ymldoes exactly that. Every other difference is a comment: four passages that recorded measurements about this repository by name, reworded to say the same thing without naming it, plus a new block recording why the job deliberately has noSetup Protocstep — nothing in it compiles, so no build script runs, so nothing reaches forprotoc..github/rulesets/README.mdnow says therefs/heads/dependabot/**/*exclusion from required signatures is load-bearing rather than merely convenient: the workflow pushes an ordinary unsigned commit, so restoring signatures on those branches while keeping the force-push and deletion exclusions would put every cargo pull request back to unmergeable with nothing saying why.scripts/src/update_changelog.dartnow gives the dependency-bump changelog prompt a checkable scope argument. Rule 1 asks for the documented bold summary and em-dash format instead of fixing a title as the first line. Rule 2 requires both where an upstream change landed and why that location is outside this package's resolved dependency graph, rejecting a bare out-of-scope verdict or a location without its reason.Rule 4 continues to require the exact no-impact phrase, but the phrase is now shared as
noImpactPhrasewithbreakingContradictsNoImpact, so the prompt and its guard cannot silently drift apart. The prompt also requires the conclusion to appear once in a sentence rather than as a detachedNote:line. Its concrete example has been replaced by an abstract answer shape and an explicit anti-example, making the failure modes inspectable without teaching the model one project-specific response..copier-answers.ymlrecords template versionv4.12.0. -
copier template adopted: v4.9.0 -> v4.10.0 — the generated project now documents the FFI scan boundary and makes the pull-request gate report the full test matrix instead of disappearing behind a path filter.
.claude/skills/frb-patterns/SKILL.mdnow says thatrust/src/api/is a scanned directory, not a general-purpose folder. Becauseflutter_rust_bridge.yamlscans fromcrate::api, a module declared underapi/mod.rsbecomes part of the generated FFI surface. Helpers intended only for the crate therefore belong at the crate root and must be declared fromrust/src/lib.rs; otherwise code generation can silently publish an API and changerustContentHash..github/rulesets/protect-main.jsonnow requires the complete generated check set, includingFRB bindings were regeneratedand the test matrix..github/workflows/test.ymlremoves thepull_requestpath filter that would otherwise make required checks report nothing for documentation-only pull requests; the filter remains onpushfor cache-scope control..github/workflows/codegen-guard.ymlchanges in a comment only — the paragraph explaining why the job'sname:is load-bearing now says that name is one of the required contexts rather than the one. Its job, trigger and steps are byte-identical: measured on the diff, every changed line is a comment line..github/rulesets/README.mdnow explains the required-check model and how to change it safely: context names must be taken from a real pull-request run,make setup-repo-protections ARGS="--update"must be used to apply an existing ruleset edit, and flaky legs should be removed from the required list while remaining in the workflow. It also records the job-levelif:approach for an expensive leg, since a skipped job still reports and can satisfy the gate..copier-answers.ymlrecords the new template version.
3.1.0 - 2026-09-08 #
For Users #
✨ Highlights
- Three bindings taken from openmls 0.9.0 — a secret can now be exported from a Welcome before deciding whether to join, a signature-key rotation can be sent as a proposal instead of a commit, and a key package's validity window can be read out and judged against a clock other than the device's. All three are additive: no existing signature moves.
- openmls v0.9.0 — unchanged this release
- openmls_frb v2.2.0 — Rust FFI bindings
Added
-
exportWelcomeSecret(rust/src/api/engine.rs) —exportSecretone step earlier in the join. It takes the samelabel,contextandkeyLengthand derives from the exporter secret of the epoch the Welcome invites you into, so an invitee can agree a key with the inviter, or prove it can read the epoch, before it accepts. The value is identical to whatexportSecretreturns after joining that epoch, which the tests assert against both the joiner and a member already in the group.Like
inspectWelcome, it commits nothing, and that is load-bearing rather than incidental: processing a Welcome makes OpenMLS delete the key package it was addressed to unless that package is last-resort. Here the delete lands in the call's snapshot and is discarded with it, so a laterjoinGroupFromWelcomeon the same Welcome still finds its key package. A test now pins that behaviour end to end — inspect, export, then join — because a commit added to either function would silently burn the invitation instead of failing loudly.It reads the unverified group info, the same as
inspectWelcome: the confirmation tag is only checked when the Welcome is staged into a group. -
proposeSelfUpdateWithNewSigner(rust/src/api/engine.rs) — the proposal form ofselfUpdateWithNewSigner, filling the one gap in that pair. An Update proposal has to be carried by somebody else's commit, so this is what a member sends when it wants its signature key rotated by the next committer rather than committing itself.Two signers are required because the message and its payload are authenticated against different keys: the envelope by
oldSignerBytes, since the sender's leaf in the tree still carries the old key, and the new leaf inside the proposal bynewSignerBytes, so it verifies against the key it announces. Upstream additionally requires that a credential set in the leaf-node parameters equal the new signer's; this wrapper builds leaf-node parameters fromleafNodeCapabilitiesandleafNodeExtensionsonly and never sets a credential there, so the new signer's credential is always the one folded in and the precondition cannot be violated from the Dart surface.Availability rests on this crate not enabling openmls's
virtual-clients-draftfeature. Upstream gates the function onnot(virtual-clients-draft), its owntest-utils, ortest— and thattest-utilswas deliberately dropped from the shipped binary in 3.0.0, sonot(virtual-clients-draft)is the only arm holding it open. -
keyPackageLifetimeandcheckLifetimeAt(rust/src/api/engine.rs) — two synchronous helpers that expose a key package's validity window and let it be judged against a supplied instant, such as a timestamp from a server, rather than the device's clock.checkLifetimeAtcalls OpenMLS's ownLifetime::validate_with_timerather than re-implementing the comparison, so the boundaries match what a peer will decide about the same package:notAfteris exclusive — an instant equal to it is already expired — whilenotBeforeis inclusive. Neither bound is adjusted here: OpenMLS's hour of clock-skew margin is added byLifetime::newwhen a key package is created (lifetime.rs,not_before = now - 1h), so it is already inside thenotBeforethatkeyPackageLifetimereads back, andcheckLifetimeAtcompares whatever bounds it is handed unmodified — it reaches OpenMLS throughLifetime::init, which upstream documents as "raw lifetime without skew".The device's clock still gates the reading half, and the split of these into two functions is what says so. Getting at the window means validating the key package, and OpenMLS's validation checks signatures, protocol version, extensions and the lifetime, that last one against this device's clock; there is no way to skip it from outside the crate. So
keyPackageLifetimefails on a package this device believes is expired and never yields its bounds — measured here, not inferred, by a test that builds an already-expired key package and asserts on the error OpenMLS reports. The pair can therefore apply an authority stricter than the local clock, which is the useful direction when choosing among a user's key packages, but cannot rescue one the local clock has already rejected.checkLifetimeAttakes the two bounds rather than key package bytes precisely so that half of the check has no local clock in it at all, for a caller that holds bounds from elsewhere.An instant past the year 9999 is an error rather than a verdict, and the cap is explicit rather than left to the platform. The platforms do not agree about what a
SystemTimecan hold:web_time's on wasm32 is a bareDurationsince the epoch and accepts everyu64of seconds, while nativestd::time::SystemTimedoes not — so the same call withu64::MAXwas a verdict on one target and an error on the other, measured on both. The cap is set where no calendar can mean a larger value rather than at any platform's limit, which is what keeps it correct without a survey of the platforms: 9999-12-31T23:59:59Z costs no caller anything, and below it every target this package ships for answers a given argument the same way. It also catches the likeliest way to call this wrongly — Dart has nosecondsSinceEpoch, so a caller reaching forDateTimemeetsmillisecondsSinceEpochfirst, and a present-day millisecond count sits far past the cap, failing loudly instead of returning a confident verdict about a date tens of thousands of years out.
Security
-
The snapshot's read path no longer leaves plaintext copies in freed memory (
rust/src/snapshot_storage.rs) —kv_readhanded out a clone of the stored value and every caller dropped it unwiped, so each read of MLS key material —EpochSecretsandMessageSecretsamong them — left a plaintext copy on the heap that nothing ever overwrote. Its two siblings did not have this: bothkv_writeandkv_deletealready zeroized the value they displaced, and both snapshot maps are zeroized onDrop. The read was the one way out that was not covered.Pre-existing rather than a regression: the diff of this file from
v2.0.1was additions only. The exposure was bounded by the same window as the rest of the snapshot — single-digit milliseconds per operation — but unlike the maps, the copies were never wiped at all, so they persisted until the allocator reused the pages.The fix is a type rather than a convention:
kv_readreturnsZeroizing<Vec<u8>>, so a caller cannot bind the value unwrapped without saying so, and the wipe runs on every path out,?included. The three decoded item lists behind it (append_to_list,read_list,remove_from_list) are wrapped the same way, because deserializing allocates fresh buffers that the wrapper around the value does not reach, andremove_from_listadditionally wipes both the elementVec::removehands back and the needle it was given.Zeroization cannot be asserted from a test — the wipe lands on memory the allocator may already have reused, and reading it back would be undefined behaviour — so a source-level test pins the shape instead, in the same way the release-profile test in
utils.rspins its invariant. It reads this file and fails ifkv_read's return type is widened back, if any decoded item list is bound outside the wrapper, or if the three lists it counts are no longer there. All three of those checks were confirmed to fail on a deliberately broken tree that still compiles.
Documentation
- The README names a Flutter behaviour that leaves
web/pkg/empty — the build hook provisions the WASM module into an app'sweb/pkg/, andflutter build webalways reaches it, butflutter run -d chromereaches it only while Flutter considers itsdart_buildtarget out of date. That target's cache key does not include the platform, so a debug run for another platform leaves a stamp the Chrome run accepts, logsSkipping target: dart_build, and never invokes the hook —RustLib.init()then fails on a 404 forpkg/openmls_frb.js. No hook can defend against it: the skip happens abovehooks_runner, where nothing the hook declares is read. Known Limitations now names the three escapes (flutter build web,rm -f build/*/dart_build.stamp,flutter clean). Behaviour is unchanged — this was always true and was undocumented.
For Contributors #
Changed
-
web-timeis now a direct wasm32 dependency of the native crate (rust/Cargo.toml) —Lifetime::validate_with_timetakes aSystemTime, and openmls'slifetime.rschooses which one bycfg(target_arch):web_time's on wasm32 andstd's everywhere else. They are unrelated types, socheckLifetimeAtcannot name the epoch it adds to without this. Same version openmls already resolves (1.1.0), so no crate enters the graph andTHIRD_PARTY_NOTICES.txtdoes not move — confirmed by the gate, not assumed.Worth knowing for anyone touching wasm32 code here:
make rust-docdoes not catch this class of error. Its--target wasm32-unknown-unknownpass went green on a deliberately wrongSystemTime, because rustdoc does not type-check function bodies.make build-webrejected the same tree withexpected web_time::time::system_time::SystemTime, found SystemTime, so it is the only local gate that compiles wasm32 bodies at all. -
copier template adopted: v4.8.0 → v4.9.0 (23 files, plus this entry) — most of it was written from this project's own findings and has been waiting on the template release; it lands here now. One conflict, in
README.md, where the template inserts the Known Limitations subsection above a heading this project had renamed; resolved by keeping the local headings. Every workflow, composite action and ruleset file is byte-identical to a v4.9.0 render afterwards, and theMakefile's three local hunks (themls_messagefuzz-target examples and theclassical_ops_do_not_init_libcruxreference) survived the merge.Android is cross-compiled on every pull request, all three ABIs, in
test-reusable.yml. Until now nothing outside a release tag cross-compiled Android at all, which is exactly how theopenssl-src3.6 / NDK r26 assembler break sat on main under every green gate and first surfaced as a twice-failed stage 1. Measured before it was written: about four minutes per ABI against this workflow's six-and-a-half-minute critical path, so the wall clock does not move.make verify-android-alignmentmeasures the 16 KB alignment Google Play requires rather than trusting the tool that supplies it. The alignment comes fromcargo-ndk's linker flags, not from the NDK —openmls_frb-2.0.1(r26) and2.1.1(r28) both measurep_align=0x4000— socargo-ndkis now pinned to 4.1.2 in the release job and the pull-request job together, and both jobs verify the bytes that come out. A misaligned.sobreaks no test here; it makes a consumer's app unpublishable.codegen-guardregenerates the bindings instead of only reading a label. A pull request that changes an existing signature was already caught, becausefrb_generated.rsstops compiling — but one that merely ADDS apub fncompiled fine and simply lacked the function on the Dart side. The job now runsmake codegenand refuses drift underlib/src/rust/orrust/src/frb_generated.rs, keeping its name (FRB bindings were regenerated) becauseprotect-main.jsonmatches it as a string.make actionlintand aWorkflow Lint (actionlint)job, pinned by version and by checksum, with the suppressions in.github/actionlint.yamlso a local run reports what CI reports. It found 93 things across this repository's workflows and none of them was a bug: 81 shellcheck findings — 79 fixed, and twoSC2129s (style only, on append-to-$GITHUB_OUTPUTblocks) suppressed inline with a reason — plus 12 false positives from actionlint's stale copy ofactions/create-github-app-token's inputs.anthropics/claude-code-actionmoves from v1.0.213 to v1.0.216 inai-review.ymlandrepair-build.yml, by commit SHA as before. Neither workflow touches the published package; both are pinned by hand, because Dependabot'sgithub-actionsecosystem does not reach the template's copy.make run-example-webclears thedart_buildstamp itself before handing over toflutter run— the working half of the Flutter behaviour described under Documentation above, which otherwise makes a Chrome run after a desktop run serve no WASM at all. A stamp that does not exist cannot be stale, and an unmatched glob is a no-op underrm -f.make verify-frb-pinsreads a sixth source whenrust/fuzz/Cargo.tomlnamesflutter_rust_bridge— a fuzz crate that drifts from the main crate does not merely disagree, it stops resolving, and nothing else notices.⚠
protect-main.jsonnow carries a required status check, and arriving is not the same as being applied. The file is in the tree; GitHub does not read it. Applying it takesmake setup-repo-protections ARGS="--update --yes"as a separate step — and--updateis the load-bearing half: the script is idempotent by ruleset name, so without it an existingProtect main branchis skipped with a message and exit 0, which looks exactly like success while changing nothing. Verified before overwriting that the live rulesets carry no hand-edits the file would clobber:delete-branchesandsigning-commitmatch byte for byte,protect-release-tagsdiffers only in the order of two bypass actors, andprotect-maindiffers only by the new rule plus arequired_reviewers: []default GitHub echoes back. Until it runs,codegen-guardreports nothing and blocks nothing — and it has since been run: the liveProtect main branchruleset now carries theFRB bindings were regeneratedcheck, matching the file. -
The scope file and the README name the new surface (
.github/agent-prompts/changelog-scope.md,README.md) — the scope file is what every future openmls release is classified against, so a binding missing from it is a binding whose upstream changes get reported as invisible to this package's users. It now namesexport_welcome_secret,propose_self_update_with_new_signer, andLifetimetogether withKeyPackageIn::validate— the latter because the whole of that validation (signature, protocol version, extensions, lifetime) is on the pathkeyPackageLifetimetakes and is therefore user-visible. The README's feature table and full API reference list the four new names.
3.0.0 - 2026-09-06 #
For Users #
✨ Highlights
- The published package initialises again —
2.0.1shippedflutter_rust_bridge: ^2.12.0alongside generated bindings that record2.12.0and are compared against the runtime with==. flutter_rust_bridge 2.13.0 was published on 2026-08-23 and landed inside that caret, so from that day on every fresh resolution — there is no committed lockfile to hold it still — threw fromRustLib.init(). The constraint now admits exactly one version. flutter testfinds the native library —flutter_toolsinstalls the hooked library underbuild/native_assets/<os>/, a directory neither of the two paths searched before covered, so a Flutter package depending on this one failed ininit()in its own unit tests while the app itself ran fine.- The post-quantum dependency tree moves off every advisory it was pinned to — the X-Wing path's libcrux crates were held at exact versions by openmls 0.8.1, and 0.9.0 moves all of them. The ignore lists shrink to one entry.
- Nine more ciphersuites, all of them post-quantum (breaking) —
MlsCiphersuitegrows from four values to thirteen, matching what the crypto provider actually runs. The nine were already going out on the wire in every leaf node built from the openmls 0.9.0 bump; they were simply not nameable in the Dart API, which made a group using one impossible to inspect and therefore impossible to decide about joining. - X-Wing is again the only suite delegated to libcrux — three ML-KEM suites had begun routing to a backend validated only for X-Wing, and one of them could not run there at all. Both this and the item above arrived with the 0.9.0 bump and are fixed in the same unreleased version, so no published release was ever affected.
- Rust backtraces stop reaching Dart error strings — the shipped binary
enabled openmls's
test-utilsfeature, and with it openmls formatted a symbolized backtrace into the internal errors that travel the ordinary error channel out to the caller. Dropping the feature also takes a test harness out of the release dependency graph. - openmls v0.9.0 — first upstream release since 0.8.1 (2026-02-13), and it closes an advisory this package had been working around locally.
- openmls_frb v2.1.1 — Rust FFI bindings
Changed (Breaking)
-
MlsCiphersuitegains nine values (rust/src/api/types.rs,lib/src/rust/api/types.dart) — the enum described four ciphersuites while the shipped crypto provider supported thirteen, and it was the provider's list that peers saw.supportedCiphersuites()now returns all thirteen, and every one of them is exercised by a full group lifecycle in both the Rust and the Dart test suites — created, added to, joined via Welcome, and messaged in both directions — rather than merely advertised.The gap was not cosmetic. Enabling
draft-ietf-mls-pq-ciphersuitesfor X-Wing also put nine ML-KEM suites into OpenMLS'sdefault_ciphersuites(), which is whatCapabilities::newfills in when a caller does not pin capabilities — so from the 0.9.0 bump onward, key packages and leaf nodes advertised all thirteen. A peer picking one of the nine produced a group thatinspectWelcomerefused withUnsupported ciphersuite, before the application could decide whether to join, even though the join itself would have worked. 0.8.1's default list held exactly the four the enum named, so 2.0.1 was consistent and the drift never reached a published release — it is introduced and fixed within this one.Ten of the thirteen are experimental suites on provisional code points that are not registered with IANA and may be renumbered or withdrawn; several are pure ML-KEM, with no classical component to fall back on. The README now lists every suite with its code point, its KEM and its signature, and says which are hybrid and which are not.
Action required: an exhaustive
switchoverMlsCiphersuiteno longer compiles. Add the nine new cases, or adefault:. Stored state is untouched — the original four keep their names, their meanings and their positions in the enum, so existing values continue to serialize identically and no stored group or key package has to be migrated.Action required, and this one is on the wire: what a client advertises changes even if you touch no code.
MlsCapabilities.ciphersuitesreads an empty list as "use OpenMLS's defaults", and that default list grew with the same feature — from four entries to thirteen — so a leaf node or key package built without explicit capabilities now advertises all thirteen, nine of them provisional code points not registered with IANA. Nothing breaks: RFC 9420 lets a client advertise suites it is never asked to run, and the group's suite stays the creator's choice. But a peer may now pick an experimental suite for a group you would join. To keep 2.0.1's advertised set, passMlsCapabilitieswith an explicitciphersuiteslist of raw code points —[0x0001, 0x0002, 0x0003]— tocreateGroupWithBuilder,proposeSelfUpdateandcreateKeyPackageWithOptions.
Changed
-
The Android libraries are built with NDK r28 instead of r26 (
.github/workflows/build-openmls.yml,.copier-answers.yml) — the shipped.sofor all three ABIs is now produced by Clang 19 rather than Clang 17. Not a chosen upgrade.rusqlite'sbundled-sqlcipher-vendored-opensslvendors OpenSSL, and the openmls 0.9.0 bump re-resolved that from 3.5.5 to 3.6.3, which ships an SM3 x86-64 assembly implementation using the Intel SM3 instructionsvsm3msg1,vsm3msg2andvsm3rnds2. NDK r26's Clang 17 does not know them, sox86_64-linux-androidfailed to build at all (invalid instruction mnemonic); Clang 18 is the first that assembles them, and r28 is the current stable line.Rolling OpenSSL back was rejected rather than untried:
openssl-src's newest 3.5 packaging is300.5.5+3.5.5, and OpenSSL 3.5.5 carries seven CVEs fixed only in 3.5.6 and later, which that crate does not package. The rollback would have traded a build failure for known vulnerabilities.The gap that let this reach a release tag is that Android is built only on an
openmls_frb-*tag — theTestsworkflow never cross-compiles it — so the breakage sat onmainfrom the 0.9.0 bump until the first tag after it. -
flutter_rust_bridge 2.12.0 → 2.13.0 (
pubspec.yaml,rust/Cargo.toml,Makefile,lib/src/rust/,.copier-answers.yml) — moved in all five places that have to agree, whichmake verify-frb-pinschecks, and the published constraint is>=2.13.0 <2.13.1.The reason is compatibility with sibling packages rather than anything in 2.13.0 itself. Because the constraint admits exactly one version — it has to, see the initialisation fix under Fixed — two flutter_rust_bridge wrappers pinned to different versions cannot resolve together in one application at all.
libsignal_darthas moved to 2.13.0, so staying on 2.12.0 would have published a release known in advance to be unresolvable alongside its next one. Both published packages still agree on 2.12.0 today, which made this the last moment the divergence could be avoided rather than repaired.The move itself is mechanical.
rustContentHashdoes not change (585923240before and after), so the wire signature between Dart and the native binary is unchanged; the whole Dart diff is the@generated bystamp in nine files plus onecodegenVersionstring; the generated Rust gainsstd::result::Result::Okqualification throughout; andCargo.lockmoves the two flutter_rust_bridge crates and nothing else, so the third-party notices change by two version numbers.It does require the native binary this release builds anyway: the WebAssembly module comes out 8.9 KiB smaller (2,549,220 → 2,540,136 bytes), which is change enough to re-run the manual browser check that the crypto has no CI gate for — all four post-quantum lifecycles pass on the rebuilt module, with key package, commit and welcome sizes identical to the previous run.
-
openmls 0.8.1 → 0.9.0 (
rust/Cargo.toml) — the X-Wing ciphersuite is unaffected despite an upstream feature gate. 0.9.0 putsHpkeKemType::XWingKemDraft6andMLS_256_XWING_CHACHA20POLY1305_SHA256_Ed25519behinddraft-ietf-mls-pq-ciphersuites, but nothing was renamed and the code point is still 0x004D, soMlsCiphersuite.mls256Xwing…keeps working exactly as before. Existing databases are unaffected too: the storage format does not move, which was verified by reopening a database written by the published 2.0.1 binary and decrypting a message it had encrypted and never opened. -
Two upstream message states are now named instead of reported as unknown (
rust/src/api/engine.rs) — 0.9.0 splitsOwnPendingCommitandOwnPrivateMessageout of what used to be errors, andprocessMessagewould otherwise have returned "Unknown processed message content type" for both.OwnPrivateMessagereplaces 0.8.1'sCannotDecryptOwnMessage: the same input, the same rejection, reached by a different path, and it now says why — the own sender ratchet is encryption-only, so a message this client authored cannot be read back.OwnPendingCommitcannot occur through this engine at all, because every commit-producing call merges before it returns and no pending commit ever reaches storage; it is handled anyway. No type in the Dart API gains a value. -
Key package lifetime validation errors read differently (upstream) — no code changed here, but the strings these errors produce did, and an application matching on them will stop matching.
LifetimeErrorreplacedRangeTooBig/NotCurrentwithExpired { not_after, now },NotValidYet { not_before, now }andSystemTimeBeforeUnixEpoch, so the reason is now stated with the timestamps instead of being a single opaque case. At the same timeLeafNodeValidationError::LifetimeandKeyPackageVerifyError::InvalidLifetimebecame#[error(transparent)], which drops the wrapper prefix: where 0.8.1 produced "Lifetime is not acceptable.", 0.9.0 produces the inner error's own text. These reach Dart through the ordinary error string, so treat the text as diagnostic and branch on the operation instead. -
rusqlite 0.34 → 0.37 (
rust/Cargo.toml) — not a chosen upgrade. openmls 0.9.0 pullsopenmls_sqlite_storage, which pinsrusqlite = "0.37"; cargo resolves optional dependencies into the lockfile even when the feature is off, andlinks = "sqlite3"admits exactly one. Nothing about the database changes: the SQLCipher amalgamation libsqlite3-sys 0.35 bundles is byte-identical to the one 0.32 bundled (same SHA-256, SQLite 3.46.1), so the on-disk format is untouched.
Security
-
Three ciphersuites had begun using a crypto backend validated only for X-Wing (
rust/src/hybrid_crypto.rs) — this provider delegates one ciphersuite, X-Wing (0x004D), to OpenMLS's libcrux provider, because RustCrypto does not accept it; everything else runs on RustCrypto. The predicate that made that decision matched on the KEM, which identified exactly one suite in openmls 0.8.1. 0.9.0 givesXWingKemDraft6to four suites — X-Wing plus threeMLKEM768X25519variants, which is arithmetically reasonable, since X-Wing is ML-KEM-768 combined with X25519 — so three ciphersuites silently began having their HPKE operations answered by a backend that had never been validated for them. This arrived with the 0.9.0 bump in this same unreleased version, so no published release shipped it.One of the three could not be answered at all:
MLS_128_MLKEM768X25519_CHACHA20POLY1305_SHA384_MLDSA44was advertised in every key package while libcrux rejected it withUnsupportedCiphersuite, having no ML-DSA signatures. RustCrypto implements it, so the routing fix makes that suite work rather than merely stop misrouting.The predicate now matches the whole
HpkeConfigtriple, which is unique per suite and is also the only information the HPKE methods receive — they are handed a config, never a ciphersuite. Two tests pin it, and they are the ones a libcrux advisory's reachability argument should now cite: one enumerates every supported ciphersuite and requires that exactly one reaches libcrux — by the predicate and by running HPKE keygen on a fresh provider per suite — and one keeps the advertised list and the provider in bijection so the set cannot grow unnoticed..cargo/audit.tomland the security-review checklist now name them. The existing guard was not wrong, it was unbounded: it proved that a single hard-coded classical suite stays off libcrux, which said nothing about how many other suites had arrived there. -
openmls no longer formats a Rust backtrace into errors the Dart caller receives (
rust/Cargo.toml,rust/src/api/engine.rs) — the shipped binary enabled openmls'stest-utils, which implies itsbacktracefeature. With that on,LibraryError::custom()builds its message as"Error description: {s}\n Backtrace:\n{…}"— a symbolized Rust backtrace carrying build-machine paths, symbol names and crate layout.LibraryErroris what openmls raises instead of unwrapping when an internal invariant is violated, and it is returned like any other error, so this package formatted it into the error string the caller gets. No panic was required to reach it. The feature is now off on every platform: the" Backtrace:"literal is gone from both shipped artifacts — the native library, 309 KiB smaller, and the WebAssembly one, 84 KiB smaller. Web consumers were affected too, because cargo unifies features across the platform-specific dependency tables, so the wasm32 build had been carryingtest-utilsas well.It could not simply be switched off before. Cargo features are additive, and
MlsGroup::export_group_context()was gated behindtest-utilsalong with thetree_hashandconfirmed_transcript_hashaccessors thatexportGroupContext()reports, so dropping the feature meant dropping fields fromMlsGroupContextInfo— a breaking change for the sake of a hygiene fix. openmls 0.9.0 madeMlsGroup::public_group()public, andexport_group_context()is a one-line wrapper overself.public_group.group_context(), so the replacement is the same call chain rather than an equivalent one: the same six fields, the same values, and regenerating the bindings produces no diff at all. Nothing in the Dart API changes.SECURITY.mddescribed this feature as enabling "accessor methods" with "no test-only code paths activated in production". That was true ofopenmls_basic_credential's feature of the same name — still enabled, still what makesprivateKey()possible — but never of openmls's, and the entry now says which is which. -
Signature private keys are wiped from memory when dropped (upstream,
openmls_basic_credential0.5.0 → 0.6.0) —SignatureKeyPair.privatewas a plainVec<u8>, so a signing key's bytes were left in freed heap memory when the pair went out of scope, and noDropimpl could be added downstream because the type is upstream's. It is now aSecretVLBytes, which isZeroizeOnDrop. This package constructs one of these per signing operation, so it is on a hot path. The fix had been merged upstream since March and unreleased; this package argued for its release on openmls#2116, which 0.9.0 closes. It needs no change here. -
The out-of-bounds parsing bug is now fixed upstream, not only worked around here (
rust/Cargo.toml) — GHSA-rrmv-c79f-cf5r was published on 2026-08-25 withpatched_versions: 0.9.0. It is the bug behind theRead-based decoder this package has carried since 1.4.2: openmls' manualDeserializeBytesimpls sliced the input at the re-serialized length and indexed out of bounds when that exceeded the bytes actually consumed, on bytes that come straight off the network. Upstream now returns the authoritative unconsumed tail. The disclosure window is why 1.4.2's and 2.0.0's entries describe this as hardening in general terms; this is the first release able to name it.The local decoder is gone with it (
rust/src/wire_decode.rs,rust/src/api/engine.rs): all nineteen call sites decode through openmls' owntls_deserialize_exact_bytesagain. That was held back until the fix covered both halves of the problem, because the workaround was never only about the panic —Extension::tls_deserializedid not require a payload to be fully consumed, so on input with trailing bytes the two decoders resumed from different offsets and disagreed about what the message said. 0.9.0 fixes the remainder arithmetic (the advisory) and makes known structured extension payloads reject trailing bytes (upstream #2134), so the two paths now agree and the removal is behaviour-neutral rather than a rollback to the old behaviour. The fuzz target over those decoders is deliberately kept: the types are still what the API parses straight off the network, and they are worth fuzzing whichever decoder is behind them. -
The X-Wing dependency tree moves off every pinned advisory version (
rust/Cargo.lock,.cargo/audit.toml,rust/deny.toml) — 0.8.1 pinned the libcrux crates at versions whose advisories could only be accepted or argued unreachable. 0.9.0 moves the whole tree:libcrux-secrets0.0.5 → 0.0.6, which is the fix version for the one entry that was an accepted availability risk rather than an unreachable one (RUSTSEC-2026-0212, incorrect constant-time swap/select on aarch64);libcrux-ed255190.0.6 → 0.0.9, past the 0.0.7 that fixes RUSTSEC-2026-0075;libcrux-aead0.0.7 → 0.0.9;hpke-rs0.6.1 → 0.7.0. The ignore lists shrink accordingly —.cargo/audit.tomlto nothing,rust/deny.tomlto a single unmaintained build-time proc-macro — verified by deleting the entries and re-running the gates rather than by reading version numbers. -
A panic on the web skips the zeroize that a native panic still performs (
SECURITY.md) —wasm32-unknown-unknowncompiles withpanic = "abort"by target default, so a Rust panic there traps the WebAssembly instance instead of unwinding and no destructor runs: the snapshot's plaintextHashMaps and the database key material are left in the module's linear memory until the page drops it. Native builds unwind and do zeroize —[profile.release]deliberately carries nopanickey — so the hole is web-only and opens only on the panic path; ordinary operation zeroizes on both platforms. Nothing in the shipped package changed: the property was always this way and is now written down, as Known Limitation 12, because a web deployment's threat model depends on it. -
The security policy now covers the post-quantum suites it ships (
SECURITY.md) — the file said nothing about them: no mention of libcrux, X-Wing, or a provisional code point anywhere. That was tolerable while the enum named four suites and one of them was experimental; it is not now that it names thirteen and ten are. Known Limitation 13 states what a reader has to decide about — the code points are unregistered and may be renumbered or withdrawn, several suites have no classical component to fall back on, the implementations are pre-1.0, and an emptyMlsCapabilities.ciphersuitesadvertises all thirteen — and where the mitigation is. It also records the invariant the.cargo/audit.tomlreachability arguments are stated over: exactly one suite reaches libcrux, enforced by three tests. Documentation only; nothing in the shipped code changed.
Fixed
-
The published key package lifetime default was wrong, and four other documents disagreed with the code (
rust/src/api/types.rs,rust/src/api/keys.rs,rust/src/api/engine.rs,rust/src/api/credential.rs,README.md,SECURITY.md,lib/openmls.dart) —KeyPackageOptions.lifetimeSecondsdocumentedNoneas "default (90 days)". openmls 0.9.0 sets that default to3 * 28days, i.e. 84, and it is the default that applies:createKeyPackagebuilds without a lifetime andcreateKeyPackageWithOptionssets one only when the caller passes it. A rotation window sized off the docstring was sized against a number this package never used.Alongside it, six references in published doc comments named Rust spellings that do not exist on the Dart surface —
from_raw()forfromRaw, and the parametersdb_path,encryption_key,process_message,certificate_chain,private_key.make doccannot catch these: the references are in plain backticks, which resolve nothing and so never warn. The README's hardening list namedsigner.serialize()as key material when it carries only the public key and scheme, whereSECURITY.mdnames the two that do carry secrets;SECURITY.mddescribed RustSec ignore justifications in.cargo/audit.toml, whose list is empty and whose header says so; the "Full API reference" omitted eight methods, two of which the security policy tells the reader to use; the architecture table gave iOS one architecture while the release ships and the build hook resolves anx86_64simulator slice (the cell now says so explicitly); andlib/openmls.dart's install snippet still said^1.0.0. -
The way to narrow what a key package advertises was undocumented, and the README said it did not exist (
rust/src/api/engine.rs,lib/src/rust/api/engine.dart,README.md) —createKeyPackageandcreateKeyPackageWithOptionscarried no doc comment at all, and the README stated outright that "createKeyPackagetakes no capabilities argument, so key packages always advertise the full list". The second half is false:KeyPackageOptions.capabilitiesexists and reachesbuilder.leaf_node_capabilities. Since this release is the one that grows the advertised list from four suites to thirteen, that sentence sat exactly where a reader would go looking for the mitigation and told them there was none. Both functions now document what the defaults advertise and which one to reach for; the README points at the option instead of ruling it out. -
Six dead references in the published API documentation (
rust/src/api/types.rs,rust/src/api/engine.rs,lib/src/rust/) — flutter_rust_bridge copies a Rust doc comment into the generated Dart verbatim, and Rust's intra-doc syntax is not Dart's:[`MlsCiphersuite`]puts a code span inside the brackets, which dartdoc reads as a reference named`MlsCiphersuite`and cannot resolve. Each one reached pub.dev as a dead link — the type-level note onMlsCiphersuite, the cross-reference between the two ML-DSA-87 suites, and the mentions ofsupportedCiphersuites,MlsCapabilities.ciphersuitesandMlsEngine.close. They are now written the way this package's Rust API doc comments have to be written: plain backticks around the Dart camelCase name, which resolves on neither side and rots on neither either.Nothing here was noticed by a human: dartdoc reports an unresolved reference as a warning and exits zero, so the links had been dead for as long as they had existed.
make docis a blocking gate now (see the template adoption below), which is what surfaced them. -
RustLib.init()threw for anyone who resolved this package after 2026-08-23 (pubspec.yaml) —flutter_rust_bridgewas declared as^2.12.0, while the committedlib/src/rust/frb_generated.dartrecordscodegenVersion => '2.12.0'and the runtime compares that string to its own with==. flutter_rust_bridge 2.13.0 was published on 2026-08-23 and landed inside the caret, so every fresh resolution from that day on — this repository's CI and every consumer of the published package alike — failed initialisation withcodegen version (2.12.0) should be the same as runtime version (2.13.0).pubspec.lockis deliberately not committed for a library, so nothing held the version still, and the shipped archive carries both halves of the contradiction: the caret in its pubspec and the generated file that fixes the other side. The two pins that were already exact,="2.12.0"inrust/Cargo.tomlandFRB_CODEGEN_VERSIONin theMakefile, were never the ones at risk.The constraint now admits exactly one version, written
>=X.Y.Z <X.Y.Z+1— this release ships>=2.13.0 <2.13.1(see the version move below; the shape is the fix, the number is a separate decision). Nothing wider is safe: the check is string equality, so every version a range admits except the one that generated the bindings fails, and a<next minorbound would only narrow the window — flutter_rust_bridge ships patch releases, and a 2.13.1 would break it identically. One version is also what upstream documents — "all flutter_rust_bridge-related packages will need to have exactly the same version" — and what its ownintegratestep writes withdart pub add.The range form, rather than the bare version, is forced by the release path and not by taste.
dart pub publishwarns that a single-version constraint "should allow more than one version", and it exits 65 on any warning, somake publish-dry-run— which bothmake releaseandpublish.ymlgate on — fails, and the package cannot be published at all. The>=X.Y.Z <X.Y.Z+1form resolves to the same single version and does not trip that check. Measured rather than assumed: four constraint shapes were run throughdart pub publish --dry-run, and only the bare version produced the warning.One consequence for consumers, and it is the intended one. Anyone who also depends on another flutter_rust_bridge wrapper built against a different version now gets a version-solving failure out of
pub get, instead of a successful resolve followed by a throw atinit(). The incompatibility was always there — two sets of generated bindings cannot both equal one runtime version — so what changes is only that it surfaces where it can be acted on.Nothing else moves for this fix:
rustContentHashis unchanged, so it needs no regeneration of its own. Worth stating what that hash does not cover, since this release leans on it nowhere: it is computed over the FFI function signatures, not over enum variants, so the ciphersuite expansion above leaves it unchanged too — bindings naming thirteen suites would load against a native binary that knows four without the runtime check firing. What prevents that is the release order rather than the hash. The build hook resolves its download from the crate version inrust/Cargo.toml, which ships inside the archive, and the stage-2 release refuses to run until the stage-1 native release for that exact version exists. -
The native library was not found under
flutter test(lib/src/platform/platform_io.dart,lib/src/openmls.dart,test/platform/native_asset_search_paths_test.dart) — the build hook registers the library as aCodeAsset, but apackage:asset id is not a path:DynamicLibrary.open()hands it todlopenverbatim, only@Native(assetId:)externals go through the asset mapping, and flutter_rust_bridge needs a library handle — so the file has to be located on disk. The search covered.dart_tool/lib/(dart run/dart test) and../lib/next to the executable (AOT bundles).flutter testuses neither: flutter_tools installs the hooked library underbuild/native_assets/<os>/and never creates.dart_tool/lib/, and on macOS and Linux nothing onflutter_tester's dlopen search path covers that directory — so a Flutter package depending on this one failed ininit()in its own unit tests on a clean tree, while the app itself built and ran fine. A leftover.dart_tool/lib/from an earlierdart testis what made it look intermittent; Windows resolved it by accident, because flutter_tools prepends that same directory to the tester'sPATH. The directory is now probed last: it is relative to the working directory, so ahead of the executable-relative entry it would let a shippeddart build clibinary load whatever happens to sit underbuild/native_assets/<os>/in the directory it was launched from, in preference to the library it shipped with.
For Contributors #
Added
- A test now guards the release profile's panic strategy
(
rust/src/snapshot_storage.rs,rust/Cargo.toml) — the absence of apanickey in[profile.release]is load-bearing:panic = "abort"would skip unwinding, soDropwould never run and the zeroize of both snapshotHashMaps and of the database key material would be silently bypassed, leaving plaintext key material in memory after any panic. Nothing in the tree recorded that, let alone enforced it. It cannot be checked at runtime — Cargo forces unwind for thetestandbenchprofiles and rejects the key on per-package overrides, so the setting that actually ships is invisible from inside a test binary — so the test reads the manifest instead. It goes red onpanic = "abort"in either TOML string form, on the section being renamed away, and — because an exact[profile.release]header match is a thing TOML gives several ways around — on any non-comment line of the manifest that names bothpanicandabort, whatever shape it is written in. Verified by puttingrelease.panic = "abort"somewhere the header match cannot see and watching it go red.
Changed
-
The libcrux routing guard exercises all five HPKE methods, not one (
rust/src/hybrid_crypto.rs) —libcrux_routing_is_limited_to_xwinghas two halves, and the operational half randerive_hpke_keypairalone. Each ofhpke_seal,hpke_open,hpke_setup_sender_and_export,hpke_setup_receiver_and_exportandderive_hpke_keypairmakes its ownroutes_to_libcruxcall, so a wrong backend introduced in one of the other four left both halves green: the predicate is unchanged, and nothing called the method. All five now run per ciphersuite on a fresh provider, with a full seal/open round-trip and an exporter-secret agreement check, and libcrux must still be uninitialised after each — the failure names the method. Confirmed against the defect rather than the patch: routinghpke_openunconditionally to libcrux turns the test red on the first ciphersuite, where before it stayed green. -
The two
unsafe_codeopt-outs are item-level, and the comments describing them are true (rust/src/encrypted_db.rs,rust/src/lib.rs,rust/Cargo.toml) —encrypted_db.rsopened with a module-wide#![allow(unsafe_code)]for a single wasm32-gatedunsafe impl Send + Syncpair, which exempted all 1,500 lines of it, native half included. The allow moves onto the two impls. Both comments said the bridge and this module opt out "each with its own attribute", naming a module-level one here; both now say item-level, which is what they both are. -
CLAUDE.mddescribed a version key that does not exist, and an unsigned release tag — the native library version was documented asopenmls: native_version:inpubspec.yaml; there is no such key, andhook/build.dartreads the crate version out ofrust/Cargo.toml, which is what decides the binary a consumer downloads. The publishing checklist still ended ingit tag -a, which theProtect release tagsruleset rejects for want of a signature and which the two-stagemake release-frb/make releaseflow replaced; it now points at that flow and lists the gates to have green before it, including that stage 1 only warns when local main is ahead of origin. -
copier template adopted: v4.7.0 → v4.8.0 (12 files) — half of it had already been contributed upstream from here, so what actually arrives is one CI gate and three script fixes.
analysis_options.yamljoins the test workflow's path filters, onpushandpull_requestboth. That file decides whatmake analyzereports, so a commit changing only the lint configuration was precisely the one that did not re-run the gate it changes —dartdoc_options.yamlsat one entry above for the same reason and was already listed. This was the one hole the update closed rather than confirmed.make releasenow checks that the stage-1 binary was built from this tree, not merely that a release carrying the same version string exists. The old check was keyed on the crate version alone, and the crate version does not move until stage 1 runs — so running stage 2 on its own passed it and would publish bindings against whatever binary the previous stage 1 left behind. Neither runtime net catches that:rustContentHashcompares the FFI surface, which an FRB upgrade need not move, and the codegen assert compares the bindings against the flutter_rust_bridge runtime package.Two version readers are anchored.
getUpstreamVersion— now split into a testableparseUpstreamTag— andfrbVersionFromGeneratedBindingsboth used an unanchoredfirstMatch, which takes whichever match comes first in the file rather than the live declaration. A commented-out pin, the shape an upgrade leaves behind, therefore outranked the real one below it. Latent here, since this manifest carries no commented-out pin, and the direction that matters is a comment naming a newer tag:check_updates.dartwould then report the dependency as already current and an upstream release, security fixes included, would silently never land, with nothing failing.The manifest test that enforces the panic strategy moves to its template home. It was contributed upstream from this project and comes back in
rust/src/utils.rs, so the local copy inrust/src/snapshot_storage.rsis removed andrust/Cargo.toml's comment now names the file that holds it. The suite still reports 27 tests; the assertion is the same one, and it still goes red both onpanic = "abort"inside[profile.release]and on arelease.panicwritten where an exact section match cannot see it. -
copier template adopted: v4.6.0 → v4.7.0 (34 files) — the release is mostly gates, and two of them close holes this project knew it had.
wasm32 is executed in CI, not merely compiled.
test-reusable.ymlgainsBuild WASMandRust unit tests (browser), andmake test-webruns the crate's#[cfg(target_arch = "wasm32")]tests in headless Chrome through wasm-pack. Until now every wasm32 branch in the crate was covered by nothing:make testis the Dart VM andmake build-webonly compiles, while a wasm32 body is a different implementation of the same function rather than the same code on another host. The manifest gains[target.'cfg(target_arch = "wasm32")'.dev-dependencies] wasm-bindgen-test, which is invisible tomake third-party-notices— it runscargo tree --edges normal,build, which excludes every dev-dependency on every target — andmake verify-third-party-noticesconfirms the inventory is unchanged despite eight new crates inCargo.lock.Two documentation gates now block.
make docpromotesunresolved-doc-referenceto an error through a newdartdoc_options.yaml(.pubignored, so pub.dev's own dartdoc run is not held to it and a future dartdoc release cannot break documentation generation for an already-published version), andmake rust-docruns rustdoc under-D warningson the host and on wasm32. The first was red on adoption — see the six dead references above.Bookkeeping no longer outranks tests.
verify-third-party-noticesandverify-frb-pinsmoved after the test steps, so a stale inventory no longer fails the Linux leg before a single test has run.Answers, not just files.
rust_versionis raised to1.91to match the manifest, soREADME.mdandCONTRIBUTING.mdstop advertising a toolchain that cannot build the crate and the next update cannot render the stale number back over it; the newenable_freezedquestion is answeredfalse, because flutter_rust_bridge needs freezed only for data-carrying enums and structs this API does not have.Also arriving:
codegen-guard.yml(a pull request labelledcodegen-failedfails a required check rather than relying on a reviewer noticing),.github/agent-prompts/repair-build.md,--lockedin the release builds,make rust-geigerandmake run-example-web,lib/src/rust/**excluded from the coverage report (370 lines over 10 files becomes 51 over 4 — the hand-written half — both at 100%, so the badge does not move), four items in the upstream-bump checklist (a feature gate reads asE0599; features are additive and reach the shipped artifacts; advisory ignores are re-earned on every bump; MSRV lives in two files), and a CONTRIBUTING that is a document rather than a stub. -
The
unsafe_code = "deny"comment named an opt-out that does not exist (rust/Cargo.toml,rust/src/lib.rs) — both copies listedsnapshot_storage's "interior-mutability shim" as one of three modules opting out of the deny, but that module contains nounsafeat all. There are two opt-outs: the FRB-generated bridge andencrypted_db's WASMunsafe impl Send + Sync. The parenthetical claiming both go through a module-level#![allow(unsafe_code)]was wrong too — the bridge carries an item-level#[allow].SECURITY.md's own count was already right, which is how the drift stayed invisible. -
MSRV 1.89 → 1.91 (
rust/Cargo.toml) — required by openmls 0.9.0, whose workspace declares it. The 1.89 floor was ours, chosen forstd::fs::File::try_lockso the single-writer lock would not need anunsafelibc::flock; that requirement still holds, it is simply no longer the binding one. No CI change was needed — the MSRV job readsrust-versionout of the manifest and installs that toolchain, so the two cannot drift. -
Dependabot may not raise rusqlite past 0.37 (
.github/dependabot.yml) — for two independent reasons, both written out at the entry. It is pinned rather than chosen, becauseopenmls_sqlite_storagerequiresrusqlite = "0.37"andlinks = "sqlite3"admits exactly one package, so 0.38+ does not resolve at all. And 0.40.x is separately unshippable: it bundles SQLCipher 4.14.0, whosesqlcipher_fprintfallocates on Windows, so a failingVirtualLockundercipher_memory_security = ONlogs a warning, which allocates, which recurses until the stack is gone. Pull requests #19 and #21 both died that way, which is what makes it a ceiling rather than a postponement. -
Dependabot may not propose the
hooks/code_assets2.x migration (.github/dependabot.yml) — it is blocked by the pinned Flutter SDK, not by this repository's code:hook/build.dartneeds a zero-line diff, since hooks 2.0.0's one breaking change (ProtocolExtensionfrom interface to base class) touches none of the three things it uses. What fails is version solving — hooks ≥ 2.1.0 pullsrecord_use ^1.0.0, which needsmeta ^1.19.0, and Flutter 3.38.4 pinsmeta: 1.17.0exactly. The reason to ignore rather than leave a red pull request open is where the damage lands: the root package resolves fine because it has noflutter: sdkdependency, so this repository's own CI understates it, while merging would make the published package unresolvable for every Flutter consumer on an SDK that pins meta below 1.19.0 — anyone on 3.44.x or earlier. Both names are ignored together becausecode_assets 2.0.0requireshooks ^2.2.0. Lift both when the pinned Flutter reaches 3.47.0, the first stable that relaxes the pin tometa: ^1.18.3. -
Two guards now bound what the crypto routing may do (
rust/src/hybrid_crypto.rs,.cargo/audit.toml,.github/workflows/test-reusable.yml,.claude/skills/security-review/SKILL.md) — both were written before the routing fix and confirmed red against it, so they cover the defect rather than the patch.libcrux_routing_is_limited_to_xwingenumerates every supported ciphersuite and requires exactly one to reach libcrux;openmls_defaults_are_executable_and_nameablerequires every suite OpenMLS advertises by default to be one this provider can run and the public enum can name, which is the guard at the source of the drift — that list is what goes on the wire whenever a caller does not pin capabilities.api_list_matches_provider_supportwas also tightened: its second loop usedif let Ok(..), so suites the enum could not name were skipped in silence, which is precisely how nine of them went unnoticed. The workflow comment that named a single test by hand is back to the template's wording, which refers to whichever tests a justification cites. -
openmls_basic_credentialnames the post-quantum feature explicitly (rust/Cargo.toml) — it was reaching the crate only throughopenmls/draft-ietf-mls-pq-ciphersuites, which propagates with?and so applies only whileopenmls/test-utilskeeps the optional dependency alive. Without itSignatureKeyPair::newhas no ML-DSA arms and the four ML-DSA ciphersuites cannot build an identity at all — a coupling that had to be broken beforetest-utilscould be dropped from the shipped binary, which this same release then does (see Security). Feature-only change on its own:Cargo.lockand the third-party notices are untouched by this bullet. -
A test harness no longer resolves into the release dependency graph (
rust/Cargo.toml,rust/Cargo.lock,THIRD_PARTY_NOTICES.txt) — droppingopenmls/test-utils(see Security) takes fifteen crates out of the notices, 279 → 264:wasm-bindgen-testwith its macro and shared crates,minicov,openmls_test,async-trait,cast,itertools0.14,libm,nu-ansi-term,oorandom,same-file,walkdir,winapi-utilandwindows-sys. Nine of them leaveCargo.lockoutright; the rest stay resolved but unreachable.openmls_memory_storageis not among them — it is a plain dependency ofopenmls_rust_cryptoandopenmls_libcrux_crypto, not a test-only one. Neither isbacktrace, which flutter_rust_bridge'sallo-isolatedeclares independently of anything openmls does; what changed is that openmls's own feature of that name is off. -
The example app's post-quantum tab covers three backend paths instead of one (
example/lib/demos/post_quantum_demo.dart) — it now runs the full lifecycle on X-Wing (libcrux), on the ML-KEM-768 + X25519 suite with an ML-DSA-44 signature (RustCrypto, and the suite a routing regression breaks outright), and on a pure ML-KEM-768 suite, after the classical regression check. CI now builds the WebAssembly module and runs the crate's browser tests (see the template adoption above), but that suite is a single test overcurrent_time: no gate exercises a ciphersuite, the IndexedDB store or Web Crypto on wasm32. This tab remains the only runtime check of those, so it has to be re-run by hand after any change to ciphersuites or crypto routing, and it is now worth more when it is. -
The upstream-bump checklist records the feature-gate trap (
.claude/skills/update-openmls/SKILL.md) — a gated ciphersuite variant reports asE0599 ... no variant named XWingKemDraft6, which reads exactly like a removal and cost this bump a wrong diagnosis. The checklist now says to look for a new cargo feature first, and that it has to go on all five openmls crates rather than only onopenmls. A second block covers the opposite direction — which features must stay off in a shipped binary, and how to check the built artifacts for it — so a later bump cannot quietly restore the backtrace described under Security. -
make codegennow uses the pinned generator (Makefile) —FRB_CODEGEN_VERSIONpins the binary thatmake setup-frb-codegeninstalls, butcodegendid not depend on that target and ran whateverflutter_rust_bridge_codegenhappened to be onPATH. Regenerating with a different version rewrites the bindings and thecodegenVersionthey carry — the same drift the three pins exist to prevent, arriving through the one door they did not cover. Where CI already ran the two in sequence nothing changes: the prerequisite only reads--versionwhen the pinned binary is already installed. -
Adopted copier template v4.3.0 → v4.4.0 (
.copier-answers.yml,Makefile,hook/build.dart,scripts/src/check_template_updates.dart,scripts/src/update_template.dart,.github/workflows/check-template-updates.yml,.claude/skills/update-template/SKILL.md) —make update-templatenow runs copier with--skip-tasks. A singlecopier updaterenders the template three times and ran the generation tasks (flutter create,dart pub get,dart format .,rm -rf _templates) in every one of them, including the render into this project, where a task dying between the render and copier's replay of the project's own diff could leave a half-updated tree that still reported_commitas bumped. The update skill's conflict check moves fromfind -name '*.rej'— copier converts those to inline merges and unlinks them, so it always came back empty and read as "no conflicts" — togit status --porcelain | grep '^UU', and the claim that copier's conflicts only ever land in Markdown is corrected in the skill, in this workflow's comments and in the warning it writes into every conflicted pull request — a single real update has since left conflicts inMakefile,pubspec.yaml,rust/Cargo.toml,rust/src/frb_generated.rsand two Dart scripts.get-versionis now declared in.PHONY. -
Adopted copier template v4.4.0 → v4.5.0 (
.copier-answers.yml,.githooks/pre-commit,scripts/src/common.dart,scripts/src/check_template_updates.dart,scripts/src/check_updates.dart,test/scripts/common_test.dart,.github/workflows/check-template-updates.yml,.github/workflows/check-openmls-updates.yml,CONTRIBUTING.md,CLAUDE.md,README.md,.claude/skills/update-template/SKILL.md) — four changes, all of them to how this repository is worked on rather than to what it publishes.The pre-commit hook runs
make rust-checkonly when the commit touchesrust/. Measured here: the three gates together take about five seconds with warm caches, andcargo checkis the one that stops being five seconds — 1s warm against 76s from an emptyrust/target, whichmake cleanproduces outright and which a toolchain update or a dependency bump produce in effect. Most commits touch no Rust at all, including both commits that carried the previous template adoption. The predicate compares the index against HEAD underrust/, falls back to the empty tree on a first commit, fails open when git cannot answer, and announces the skip rather than performing it silently. A staged Rust file still blocks the commit when the crate does not compile, and CI runsmake rust-checkandmake rust-clippyon every push regardless.The update checkers authenticate to the GitHub API, and say what happened when it refuses. Both asked for public releases without a token, which is quota rather than access: anonymous requests are counted at 60/hour per source IP, and hosted runners share theirs — one scheduled run died on a bare
403for that reason. Both now sendGITHUB_TOKEN(orGH_TOKENlocally), the workflows pass the job's own read-only token, and a failure keeps GitHub's own message and thex-ratelimit-*headers instead of reporting the status code alone, which could not tell a spent quota from a missing repository.githubApiGetanddescribeGithubFailureare shared incommon.dartand covered by ten new tests.The template update workflow explains a pull request it could not open. When the App behind
APP_IDlacksWorkflows: Read & write, GitHub refuses the commit withResource not accessible by integrationonPOST /git/trees— naming neither the file nor the permission, and only after every blob has been created. A failure-only step now says so, including thatActionsis a different permission and that a granted permission does nothing until the installation accepts it.Documentation stopped promising a changelog generator that no longer ran. GitHub Models was in its retirement brownout, so the AI step failed on every run and labelled the pull request
changelog-needed; the setup instructions said otherwise in three places, and the update skill told whoever ran it to expect an entry.README.mdalso still described template updates as notification pull requests, which is what they were before the workflow started applying them. The v4.6.0 adoption below replaces the provider outright, so these notices are gone again in the same release. -
Adopted copier template v4.5.0 → v4.6.0 (
.copier-answers.yml, and the files listed below) — five changes arrive with it.The AI changelog works again, against a provider this repository names. GitHub Models was retired on 2026-07-30 and the step had failed on every run since.
AI_MODELSnow holds an orderedprovider/modellist and the first entry that has a key and answers wins, so the next provider change is a repository-variable edit rather than a template release. Anthropic, Google and OpenRouter are each called through their own API overdart:iorather than acurlsubprocess, because the HTTP status decides whether the next entry is tried and a subprocess would put the key in process arguments. The next entry is tried only when a model produced no answer — network failure, an auth/rate-limit/server status, a refusal, or a response cut off at the token limit — never on the content of an answer, and nothing is salvaged from a partial one: a missing field leaves the entry unwritten and the pull request labelledchangelog-needed, which is the path that already existed.This repository is not configured by the adoption itself. There is no default list, deliberately: with
AI_MODELSunset nothing is called, which is also how a project says "no AI here". Until the variable and a provider key are set, update pull requests keep arriving withchangelog-neededexactly as they do today.AI_MODELS_TOKENis now read nowhere and can be deleted.What the entry is judged against is now written down (
.github/agent-prompts/changelog-scope.md) — which crates this package binds, whatMlsEngineactually exposes, and which upstream areas it never touches, including that this package implements its own storage rather than usingopenmls_memory_storage. The prompt classifies every upstream change against that list, and an upstream change that cannot be tied to something named there is invisible to this package's users. The file is written once and never overwritten by a template update, so it is this repository's to keep current.The test suite no longer skips
update-openmls-*pull requests (.github/workflows/test.yml) — the bump whose entire payload is new native code was the only pull request merged without the suite, clippy,rust-test,cargo-deny, the MSRV check orverify-third-party-noticesrunning against it on any platform.make buildruns beforemake testin the reusable workflow and the build hook then findsrust/target/releasewithout downloading, which is what justified the skip and now removes the need for it. This had an immediate consequence: PR #15 (openmls 0.9.0) started running the suite and failed. That failure was real and had been invisible — but the cause recorded here when this entry was written was wrong, and is corrected rather than left standing:HpkeKemType::XWingKemDraft6and theMLS_256_XWING_CHACHA20POLY1305_SHA256_Ed25519ciphersuite were read as removed upstream, and neither was. Both still exist in 0.9.0 with unchanged spelling and the same 0x004D code point, behind the newdraft-ietf-mls-pq-ciphersuitesfeature. A feature-gated variant reports asE0599 ... no variant named X, which is indistinguishable from a removal without checking upstream — the misreading the bump checklist now warns about.make verify-frb-pins, and Dependabot onpubandcargo(Makefile,scripts/verify_frb_pins.dart,.github/dependabot.yml) — five files record the flutter_rust_bridge version and two of them are compared with==at runtime, so the new gate reads all five and rejects both a caret and the unpublishable bare form, with the reason. It reads every occurrence rather than the first, because the first is not always the one that counts: adependency_overridesentry replaces the dependency outright, a second pin in a[target.'cfg(…)']section resolves per target, and make takes a later=over an earlier?=. It runs besideverify-third-party-noticeson the Linux leg; five file reads, no build. Dependabot now watches the published constraints and the native crate's dependencies, ignoringflutter_rust_bridgein both because its version has to move in four places at once and a one-file pull request is wrong by construction. The upstream crates are ignored undercargofor a different reason: Dependabot's cargo updater does follow git refs, so without that it would open its own pull request for the same bump the update workflow exists to make — without codegen, the bindings tripwire, the CHANGELOG entry or the version badge.Also in this adoption:
lintscapped to one minor line becausemake analyze ARGS="--fatal-infos"turns any new info-level lint into a build failure andpubspec.lockis not committed (precautionary — 6.1.0 was measured against this repository with nothing to report);ffigen's floor raised to^20.1.1; protoc added to the template-update workflow's gates upstream, which this repository does not render since it needs no protobuf compiler; and the repair and review agents, both inert untilAGENT_ENGINEnames one. -
Dependabot no longer rewrites constraints it was told to leave alone (
.github/dependabot.yml) —pub's default versioning strategy iswiden, "extend only the upper bound to include the new version", and it applies that across the whole manifest rather than only to what it is updating. The first run here opened a pull request whose four updates wereffigen,lints,code_assetsandhooks— and which also rewroteflutter_rust_bridgefrom">=2.12.0 <2.12.1"to^2.12.0. That is the one constraint in this file that must not float: it is the exact regression that broke every consumer of the published package when flutter_rust_bridge 2.13.0 landed inside that caret.lints,ffigenandhookswere widened past bounds set on purpose as well.ignoreis no defence, and it is worth being precise about why: it stops Dependabot opening a pull request for a dependency, not editing that dependency's constraint while it edits the file for other reasons.flutter_rust_bridgewas ignored and rewritten anyway.versioning-strategy: increase-if-necessaryfixes it — a constraint that already admits the new version is left alone, so a dependency nothing is updating stays untouched.cargoneeds none of this: on the same run it changed exactly the one crate it was bumping and left every pin alone.make verify-frb-pinscaught the rewrite on all four platforms before it could merge — its first real encounter, and what it exists for. -
The
pubandcargogroups take minor and patch only (.github/dependabot.yml) — grouping a major with everything else blocks the rest: one unmergeable entry takes the whole pull request down.It separates more than its name suggests, and the first run after the change is the evidence. Dependabot's commit trailers report a 0.x bump as
version-update:semver-minor, but its grouping applies Cargo's own reading, where a 0.x minor is the breaking bump, and keeps them out of aminor+patchgroup anyway. The six-crate cargo group split into a group of two (logpatch,uuid1.x minor) plus one pull request each forrand,sha2,hkdfandaes-gcm-siv— the four that need a migration. Exactly the intended shape. -
dart-lang/setup-dart's problem matcher is switched off at the source (.github/actions/setup-fvm/action.yml,.github/dependabot.yml) — 1.8.0 added a problem matcher fordart analyzeand registers it with::add-matcher::dart-analyzer.json. The path resolves against the calling action's directory, and this repository calls setup-dart from inside its ownsetup-fvmcomposite action, so the runner looked under.github/actions/setup-fvm/, did not find it, and failed the job fourteen seconds in, before anything was built. It failed that way on all four platforms.The first response, within this same unreleased range, was a Dependabot
ignoreon the 1.8 line. That was the wrong tool twice over. The action is now pinned to 1.8.1 withproblem-matcher: 'false', which is the input the upstream issue points at, so the version is free to move again — and the ignore has been deleted, becauseversions: ["1.8.x"]never matched anything in the first place: Dependabot parses that string throughGem::Requirement, which has no wildcard expansion. It sat in the file looking like protection while the bot went on proposing the bump. The same reading applies to thehooksandcode_assetsmajors, whose ignores are rewritten fromversions: ["2.x"]toupdate-types: [version-update:semver-major]in the same pass.The bump was worth little here in any case: setup-dart exists in that action solely to provide a
dartbinary fordart pub global activate fvmon the next line, and everything that builds and tests this package comes from the FVM-pinned SDK.
Older releases #
2.0.1 and everything before it are in CHANGELOG-ARCHIVE.md.