Automatic per-merge builds shared one list with deliberately cut releases, and the previous change only made them identifiable rather than separate: both were still prereleases, which is the thing the issue objects to. Labelling was the wrong answer to it.
GitHub has no channel concept, but it has two published states, and there are exactly two kinds of release here. A push-triggered build is now a prerelease; a workflow_dispatch build is an ordinary release and the newest one carries the Latest badge. The Releases page shows two lists instead of one list with two kinds of title in it, and /releases/latest resolves for the first time.
The prerelease flag therefore means "automatic and uncurated" rather than "beta". The product being 内测 is stated in the release notes and in docs/installation.md, which is where a claim about stability belongs; the flag cannot carry that meaning while it is also the only thing separating the channels. docs/installation.md now tells readers to take the Latest badge, describes both channels, and says explicitly that being in the release channel means somebody picked it, not that it is stable.
Already published releases are not touched, so the Latest badge stays empty until the first dispatch after this lands.
Refs MSIME-Windows#167
The Hold-for-review step built its comment with printf and spliced ${{ steps.classify.outputs.reason }} into the command text inside single quotes. That value is the list of file names the engine changed. GitHub substitutes ${{ }} textually before bash parses the line, so a path containing a single quote closes the quoting and the rest runs as a command — on a runner holding a contents:write and pull-requests:write token for this repository. Merge rights in MSIME-Engine became command execution here. Non-adversarially the same splice makes the step die under set -euo pipefail on any engine path containing an apostrophe, so a routine bump gets a red triage job and no review comment.
Every value now reaches the shell through env: and the comment body goes through --body-file. The $GITHUB_OUTPUT heredoc that carries those same file names gets a random delimiter, since a path equal to the literal EOF would otherwise close the block early.
Separately, the classifier could auto-merge a range it had not inspected. With an empty file list, echo emits one blank line, grep -v prints it, and command substitution strips it, so the emptiness test read "nothing suspicious changed" and enabled auto-merge. The compare API returns an empty file list whenever head is already an ancestor of base — an engine force-push, or the pin sitting ahead — which is exactly the case where a bump silently reverts engine source this frontend depends on. It now fails closed on any compare status other than a plain fast-forward, and on an empty file list.
Found while auditing MSIME-Apple, which carried a byte-identical copy of this workflow. Fixed there in metasequoiaime/MSIME-Apple#269.
Dependency review was guarded on the pull_request event, which made it report "skipped" on every other trigger. A required status check that reports "skipped" blocks a merge rather than passing it, and #166 made this one of the nine contexts main requires, so the automatic release could no longer land its own release pull request: land-release-pr.sh watched ci.yml go green and then got "the base branch policy prohibits the merge" (run 34004706430, pull request #173).
The guard cannot simply move to the release path either. ci.yml ignores the release pull request by path, because a pull_request run started by GITHUB_TOKEN is created in action_required and never executes, so the release workflow dispatches ci.yml on the release branch and that dispatch is the only run reporting checks on that commit.
The job now runs on every trigger. A pull_request event supplies both refs on its own; every other trigger compares the default branch against the checked-out commit, which is the same question the pull request asks.
GitHub has three release states and no notion of a channel, so an automatic per-merge build and a release somebody deliberately cut both land in the same Pre-release list that docs/installation.md presents to users as the 内测版 channel. Before every releasable merge produced a release that did not matter; now it does.
Releases published from a push get 「（自动构建）」 in the title and a note saying they were not curated. Releases published from a workflow_dispatch keep the bare tag as their title, which is what release-please already set. docs/installation.md tells readers which is which and why they share one list.
This is deliberately a labelling change and not a second release track. Giving automatic builds their own tag namespace would need a parallel versioning mechanism alongside the one release-please owns, and stacking that on a pipeline that is currently failing and shipping unsigned artifacts is the wrong order.
Refs MSIME-Windows#167
* fix(server): publish candidate pages as immutable snapshots
Global::candidate_ui is a plain global with no lock. The IPC worker rebuilds and destroys its vectors while the UI thread paints from them, and the handoff is an asynchronous PostMessage, so the worker returns to its loop and pops the next task without waiting. A second keystroke, or an async english.db candidate arrival, could free the page_views buffer and its strings while FillItemsFromUi held a reference into it. ShowItemContextMenu reads the same vectors from a mouse callback with no message handoff at all.
The worker now copies each finished page into an immutable snapshot and publishes it at the single point where the rebuild completes; every UI-thread reader loads that snapshot once and reads only its own copy. PostMessage is deliberately left alone: the worker also serves UI-originated tasks, so a synchronous handoff would invite a cross-thread deadlock.
* fix(server): bound the aux and diagnostic pipe reads
Both listeners serve a single pipe instance inline and polled a NOWAIT handle for as long as the process ran, breaking only on a read that succeeded or failed with something other than ERROR_NO_DATA. A client that connected and never wrote held the only instance forever, so the loop never returned to ConnectNamedPipe and no later client was accepted, permanently losing the langbar right-click channel and the terminal-deactivation teardown fallback.
Give both loops the same deadline the hello reads already use and go back to accepting when it expires.
* fix(server): snapshot the voice config and replace config.toml atomically
GetConfiguredVoiceInput handed out a reference to a global that LoadImeConfig rewrites from the IPC worker thread, clearing and refilling two std::map members and around twenty strings. The voice control thread and the low-level keyboard hook, which copies the struct on every keystroke, walked those nodes as they were freed. LoadImeConfig now parses into a local and publishes it in one step under a process-local shared_mutex, and the getter returns a snapshot by value. The file lock is deliberately not reused for readers: it is a named cross-process mutex with a five second wait, and taking it in a keyboard hook would stall typing system-wide.
WriteConfiguredValue also reopened the live config.toml with ios::trunc, so a crash mid-write destroyed the user settings and could leave the file unparseable for good. It now goes through WriteFileTextAtomically, which this file already uses elsewhere. That makes the directory entry swap atomic; it does not flush, so it is not a durability guarantee.
* fix(ipc): restrict the server pipes to the owning account
The pipe DACL granted Everyone FILE_ALL_ACCESS. That bit set includes FILE_CREATE_PIPE_INSTANCE, so any local account could add an instance to a name this server already owned and answer part of the connections in its place. The names are machine-global with no per-user component, and the TSF client opened them with no SQOS, so a squatter could also impersonate the connecting thread.
Build the descriptor from the process owner SID and admit only LocalSystem, that account, and app containers with connect-only rights so UWP text input still works; fail pipe creation outright when the SID cannot be resolved rather than falling back to something permissive. Pass FILE_FLAG_FIRST_PIPE_INSTANCE on the first instance of each name so a squatted name fails loudly, and request SECURITY_IDENTIFICATION on the client side.
Authenticating the peer process after connect is a larger change to the handshake and is not attempted here.
* fix(tsf): guard the null focus paths and unadvise the compartment sink
_CandidateChangeNotification treated ITfThreadMgr::GetFocus and ITfDocumentMgr::GetTop as if they could only fail through a FAILED HRESULT. Both return S_OK with a null out-pointer when nothing has focus or the document stack is empty, and the results were dereferenced unconditionally, so an external UILess host calling Finalize while focus moved away crashed the host application. The sibling _RequestCancelComposition in the same class already checks both.
CLangBarItemButton::CleanUp only unregistered the compartment inside the SUCCEEDED branch of a CoCreateInstance, then deleted the sink regardless. This TIP registers COM-less activation support, where that call does fail, and the sink was freed while TSF still held it advised. Unadvising never needed the thread manager, so it now runs unconditionally before the deletes.
* fix(ui): drop the cached visuals when the scene is replaced
Window caches focused, hovered and captured Visual pointers into the scene tree, and SetScene destroyed that tree without clearing them. Toggling the input mode rebuilds the floating toolbar scene, so the next mouse move dispatched a virtual call through a freed object.
Clear all three before the assignment, and release the capture only when this window actually holds it: an adopted HWND may never have called SetCapture, and capturedVisual_ is set even when it did not.
* fix(settings): scope external skin CSS through the CSS parser
injectScopedSkinCss concatenated a skin toolbar.css into an @scope block as text. Block nesting is textual, so a leading close brace in the skin ended the block early and the rest styled the whole settings window. The host validates only the skin file name, never the body, and candidatePreviewCss interpolated unvalidated colour strings into rule bodies the same way.
Create the scope rule empty and re-insert every rule parsed out of the skin as a child of it, so nothing lands outside however the braces are balanced, and accept a colour only in the plain notations we actually render.
* fix(release): gate the dictionary source commit on its default branch
verify_published iterated only data["repositories"], but the dictionary block pins a source_commit exactly like they do. The manifest check merely proves the release agrees with the lock about that commit, which a branch commit satisfies just as happily, so a dictionary built from an unmerged branch shipped with the provenance the gate exists to refuse.
Run the dictionary through the same loop under its own name, resolving its repository from the lock. A retired dictionary repository that no longer answers is reported rather than failing the release: its published artifacts are immutable history, and being unable to re-check them is not the same as finding them unmerged.
A push to main carrying a fix or a feat now ends with a published installer, with no human merging the release pull request and no manual workflow_dispatch in between.
release-please only cuts a release for a release pull request that was already merged when it ran, so the prepare job now invokes it twice: once to refresh the pull request against the commits this push brought, and once after land-release-pr.sh has merged it, to turn that merge into the draft the rest of the run builds and publishes. Merging before the first invocation would instead release everything except the push that triggered the run.
land-release-pr.sh is check-release-pr.sh plus the merge it deliberately used to skip. It merges with GITHUB_TOKEN, whose pushes do not start workflow runs, so the merge cannot produce a second Release run queued behind this one on the concurrency group; the release is created and built inside the run that merged it. It merges the exact commit CI proved via --match-head-commit, then waits for that merge commit to be reachable from main so the second release-please invocation does not quietly find nothing to release.
resolve-auto-release.sh picks which of the two invocations reported the release to build, preferring the second when both did, and applies the invariants validate-draft-release.sh gives the manual path: the release is tagged at a full commit SHA, and the tag agrees with the version.
workflow_dispatch stays as a repair path for a run that failed after creating its draft.
This spends one of the signing certificate's limited signatures on every releasable merge, which is what the previous manual design existed to avoid. docs/product-release.md records the trade, including why the queueing this can produce is not answered by cancel-in-progress.
Evan