Miroir Local Sync -- WordPress 插件

开发者	tommybordas
更新时间	2026年9月9日 21:23
PHP版本:	8.1 及以上
WordPress版本:	7.1
版权:	GPLv2 or later
版权网址:	版权信息

1.5.1

Nothing is ticked when you open the comparison — ever. Since 1.3.2 the screen re-ticked the scope of your last sync on that site; after a "send everything" you came back to four ticked boxes you had not chosen, which reads like a trap. Sending is now an explicit choice each time: the button stays inactive until you tick something. The per-user memory of the last scope is removed and its leftovers are cleaned up on update.

1.5.0

The comparison screen has been redesigned around one question: "if I send, what changes, and is anything at risk?". Each category (database, themes, plugins, media) is now a single line with a plain-language verdict — "✓ Identical on both sides", "● 2 changes: files at the root, miroir-local-sync", "⚠ PROD has more recent content: Posts" — and its details fold away behind a "Details" toggle. Nothing is expanded by default; what you open stays open next time.
Everything that used to be shown up front has moved into the details: the content-type table (now only the types that differ, the identical ones under a fold), per-item actions (Preview, Sync, File details), the media folders (five first, the rest under a fold) and the environment line (WordPress, WooCommerce and theme versions).
Jargon is gone from the summary: no more "Δ = LOCAL minus PROD", "same latest change on both sides", "protected by default" or "INCLUDED" badges. The state of each item is said in words: identical, modified on PROD, 3 more on LOCAL, kept on PROD.
The whole thing reads correctly in both directions: pulling says "to fetch", and the side names follow the direction selector.

1.4.5

The Media card now says WHERE the difference is. It could only give a total — "PROD has 1,254 more files" — a correct number you cannot act on. Media are now counted folder by folder (the year/month folders WordPress creates, plus those added by plugins), largest gap first, with the folders that match on both sides summarised in one line. This costs nothing: the numbers come from the directory walk the comparison already did, and no file is hashed.
Each differing folder has a "File details" button, like themes and plugins: it lists the files of that folder only, computed on demand from both sides. A whole media library is still never hashed — only the folder you open.
Both sides need version 1.4.5 for this: update the Agent on the paired site, otherwise the card keeps showing the overall total, which stays correct.
Fixed along the way: the REST parameter carrying a component name was sanitised as a file name, which erased slashes, brackets and spaces. The "(root files)" pseudo-component detail was therefore asking the remote site for a name it had mangled.

1.4.4

Fixed: on some shared hosts, a database push failed at the import step with "mysql import failed (-1):" and no explanation, and the rollback then failed the same way — leaving an operation that looked half-done although nothing had been written. The native mysql client on those hosts exits with a status PHP cannot even read, and says nothing on its error output. That is not an SQL error: the plugin now recognises it, falls back to its own PHP import engine (which works there), and only treats a failure as fatal when the client actually reported something. Warnings such as "Using a password on the command line interface can be insecure" no longer count as failures either.

1.4.3

Fixed, and it could damage a production site: the content merge ("Sync pages", "Sync posts", and the grouped sync) wrote the source site's addresses into the target. A page built in Gutenberg carries absolute URLs inside its blocks — column background images, galleries, internal links — and they were copied verbatim. A page pushed from a local site landed on production asking for https://your-site.local/wp-content/uploads/...: the file was there under the right domain, but the address pointed at a domain that only exists on the developer's machine, so nothing displayed — and on an HTTPS site the browser blocked the request as mixed content. Full database push and pull have always rewritten these addresses; the content merge did not. It now applies the same rewriting, in both directions, before writing anything, and the report says how many addresses were rewritten.

1.4.2

Fixed: on the confirmation screen, a scope carried over from the comparison (shown as chips) was sent empty when clicking Send or Fetch, and the operation was refused with "Choose at least one thing to sync" although four chips were displayed. The scope is now read correctly whatever form it takes.
Ticking "Database" on the comparison screen now ticks and greys out every content type below it: they are included in the database sync, so the grouped content sync is disabled and says so. Unticking the database gives the boxes their previous state back.

1.4.1

Snapshots have their own tab, next to Sites, Logs and Settings, with a site switcher when several sites are paired. They were only reachable from the "⋯" menu of a site card, where they got lost among connection and maintenance actions.
That menu is now grouped: connection actions first, then the lock release, then deletion.

1.4.0 A performance release: nothing moves on screen, but what was slow is now fast, and the operations that used to give up on large sites finish. Update the Agent on every paired site to get compressed dumps and 4 MB transfer chunks. A local site on 1.4.0 keeps working with a remote site still on 1.3.x: it simply falls back to the smaller chunks and to uncompressed dumps. Speed

Miroir no longer loads on pages that have nothing to do with it. Every page of your site, front end included, was compiling 474 KB of Miroir's code, and every admin screen just over 1 MB — for code that only runs on Miroir's own screens and on the REST routes the Agent answers. That is now 55 KB, loaded where it is actually used.
A database transfer sends less, in far fewer round trips. The dump is compressed before it leaves, and the two sides agree on the largest chunk the target's PHP will accept — typically 4 MB, where it used to be a fixed 512 KB whatever the server could take. Pushing a 100 MB database went from 202 requests to 6, pulling one from 201 to 14. A 100,000-row table that produced a 10.8 MB dump now produces 4.6 MB, or 0.5 MB once compressed.
Search-replace and export walk each table by its primary key instead of counting rows from the beginning. On a large wp_posts, every slice made MySQL go through again everything it had already passed, so each slice was slower than the one before; the cost is now flat, and an interrupted run resumes exactly where it stopped.
Content sync and the comparison screen ask the database far less. Matching an editorial item on the target cost up to eight full scans of the posts table and now costs none — the identifiers are read once, in a single pass. Table lists, column lists and the manifest are computed once per request instead of once per use.
The activity log no longer keeps a full copy of every report. A report carrying hundreds of entries is stored as a summary, and its detail row says the list was truncated instead of presenting it as complete.

Long operations

A long import no longer looks dead while it works. Streaming a dump into mysql, or walking thousands of files, could go minutes without a sign of life — long enough for the screen to declare the operation stalled while it was in fact running. Both ends now report progress as they go, and the lock is refreshed throughout.
Comparing a large media library no longer recomputes every checksum. Hashes are kept between runs, and when the cache fills, the ones left unused longest go first.
A file list cut short for size says so. The file-by-file diff used to show the first few hundred files as if that were all of them.

Fixes

A dump exported through a pipe was announced as "0 B" on the transfer screen although it was complete.
A gzip-compressed dump arriving on a server without the zlib extension is now refused before anything is written, instead of failing half-way through the import. A compressed dump truncated in transit is reported as incomplete, rather than as an import that succeeded.
The scope of the last sync, on a site card, is recorded as data instead of as a finished sentence: it now reads in your own language, whatever language the sync ran in.
"Loading file details…" and the whole WP-CLI output stayed in English whatever the site language. They are translated.

Under the hood

A build script that refuses to package a release whose three version numbers disagree, a PHP_CodeSniffer ruleset the code passes clean, a GitHub Actions pipeline, and an integration suite that exercises export, import, search-replace, snapshots and the REST loop against a real WordPress and a real MySQL.

1.3.2

The comparison screen no longer pre-ticks "Database": nothing is selected until you choose, or the last scope you launched on that site is restored. The database card stays open regardless, so the comparison is always visible.
Launching with nothing selected is now refused with a clear message, instead of silently falling back to a full database sync.

1.3.1

Each theme or plugin row now has a "File details" button that lists, right under the row, which files would be sent, which extra files on the target would be deleted when you deploy, and which differ. The file-by-file diff existed but nothing led to it; the row also says the gap at a glance ("36 more files in LOCAL"). The confirmation of a deployment now announces the real number of deletions.
The "(root files)" pseudo-component is named for what it is — files at the root of the folder, outside any theme or plugin — and its details now render instead of coming back empty.
The content-merge preview button is labelled "Preview" instead of a bare "i".

1.3.0 Ten things a user expects to see at a glance, found by re-reading every screen through a user's eyes.

Each site card now says when you last synced, in which direction, and whether it succeeded — not only when the connection was last tested. Failed operations are now filed under their site in the log.
The confirmation screen says what will happen before you click: how many tables of the target are replaced and their size, which groups are kept as they are by the live-data protection, whether files are sent, and that the site stays reachable meanwhile.
After a successful sync, a "Roll back to the previous state" button opens the Snapshots view with that restore point already selected.
The database card no longer claims "93 items modified on the target" while every row says both sides are identical: it only counts rows where the target really is newer, and names the types. A legend explains the Δ column.
The comparison table now says, for each content type, which side holds the most recent version — amber when the target is newer (syncing would overwrite it), green when the source is, grey when both match — with the date of the last change under each counter.
One sentence on the database card explains that including the database replaces it entirely, while the per-row "Sync" buttons only merge. Users and comments are marked "protected by default".
An (i) button on each content row previews what a merge would do — how many items to create or update, and which — without changing anything.
The media card says what sending would do ("0 files to send · PROD has 5 more, sending never deletes anything") instead of two contradictory verdicts.
The log shows the cause of a failure (step and error) in an expandable row, and exports both to CSV.
Shortcuts list what they deploy and show their last run and, when scheduled, the next one.
The Snapshots view says how many are kept and what each one was taken before (push or pull, and its scope). A snapshot just taken can no longer be pruned by the operation that created it.
Notices from other plugins no longer pile up at the top of the Miroir screen.
The comparison table now says, for each content type, which side holds the most recent version — "PROD has the most recent version" (amber: syncing would overwrite it) or "LOCAL has the most recent version" (green: safe). Each counter shows when that side last changed. The previous badge only said that the target had changed, and the explanation was hidden in a tooltip.
Notices from other plugins no longer pile up at the top of the Miroir screen.

1.2.0

Fixed: a failed import that used a different table prefix no longer leaves its half-imported tables behind after the rollback. Only tables the import itself created under the source prefix are removed — never anything that existed before, so a second WordPress install sharing the database is safe. This release repairs the safety net a database push relies on, closes the ways an operation could run twice or write outside its own folder, and finishes the screens that had been left half-wired. Update the Agent on every paired site: reliable rollback, file restore and single-pass search-replace all need 1.2.0 on both ends, and Miroir now says so on the operation screen when the remote side is too old.

Safety net

Automatic rollback after a failed push works again. The import renames the storage folder to match the target's own key, so the snapshot taken minutes earlier was no longer where the rollback looked for it: every restore ended on "Snapshot not found" and the target stayed on the half-written database. The Agent now returns the snapshot's absolute path, and the restore falls back to searching the renamed folder.
A failed push now also puts the files back. Deployment takes a snapshot of the components it is about to replace, and the rollback restores them. Until now a theme overwritten halfway through an operation stayed overwritten, with nothing on screen saying so; when the components are too large to archive beforehand, the report says that plainly instead of implying a rollback that will not happen.
An import that stops halfway no longer blocks its own rollback. A dump is replayed statement by statement, and mysqldump writes its data inside LOCK TABLES … WRITE blocks: an import that failed inside one of them left the connection holding the lock, so the restore that followed died on its first DROP TABLE — the safety net gave way at the exact moment it was needed. Table locks are now released whatever happens, and new dumps no longer contain any.
New Snapshots screen, per site: the restore points held on each side, their date and size, with Restore (confirmation by site name) and Delete. Snapshots existed but there was no way to see or use one outside an automatic rollback.
A push whose two sites have nothing to rewrite is refused before the snapshot, not after the import. The operation used to run to the end and leave the target carrying the local addresses.
Search-replace applies every address in a single pass. Replacing pairs one after another could rewrite a URL twice and produce "example.local.local"; the same value was also counted twice in the report.
A search-replace that cannot write a row no longer reports success. Failures are counted, and past a threshold the operation stops and says how many rows still carry the old address — so the rollback can still be triggered.
wp_options is processed in the first slice of a search-replace. Between the end of the import and the end of a resumable search-replace, the target was serving the local site's addresses for minutes; that window is now seconds.
The plugin's own options are left out of the search-replace. They were already kept out of the import, but not out of the rewrite that follows: on a pull the reversed pairs rewrote the remote site's address stored in the profile, so the very next call — the /unlock that ends the operation — went to the wrong host. The pairing came back as an authentication error and the remote site stayed locked until the lock expired. A filter, sumotori_sync_search_replace_skip_options, lets an add-on protect its own options the same way.

Refusing what should never have been accepted

A table prefix or a table list that does not match the expected shape is refused, and the answer says what was wrong. A crafted request could previously name tables outside the current prefix, contradict the live-data protections, or make the import rename tables it should never have touched. Nothing is renamed unless the dump really contains that prefix's options and posts tables.
Every file path used by a transfer is checked against the plugin's own storage folder before anything is read, written or deleted.
Incoming archives are bounded: an archive that announces more data than the disk can take, or whose compression ratio no real set of files could produce, is refused before extraction rather than filling the disk.
Symbolic links are no longer followed when building an archive. A link pointing outside the site could pull an arbitrary part of the disk into a deployment; links are now skipped, counted, and named in the result message.
SFTP transfers carry a sha256 checksum, verified on arrival. A truncated or corrupted deposit is refused, and nothing is imported.
The SSH host key of an SFTP server is pinned on first connection. A key that changes afterwards aborts the transfer and says what it may mean; the profile's SFTP settings offer "Forget the fingerprint" for a legitimate server change.
Stored credentials are re-encrypted with a derived key bound to the plugin version. Existing profiles are migrated transparently the first time they are read — nothing to redo. When the site's WordPress salts have changed since pairing, the message now says the credentials can no longer be decrypted and that the site must be paired again, instead of reporting a connection failure no amount of retrying could fix.
Pairing over http:// is refused, because the application password would travel in clear text. A local address is still accepted, and a checkbox lets you accept the risk explicitly; sites already paired over http keep working and are flagged on the Sites screen.

One operation at a time

Two operations can no longer run on the same site at once. Starting a second push while one is running is refused, naming the operation that holds the lock and how long it has been running, instead of two imports interleaving on the same database.
The lock records which operation owns it: only that operation releases it, and "Force unlock" now clears both ends. An interrupted operation no longer blocks the site until a transient expires.
The lock survives the import that used to erase it. It is now kept, along with your profiles, settings, shortcuts and logs, when the incoming database replaces wp_options.
A long import keeps the lock alive while it works, so a slow operation is no longer declared dead and restarted underneath itself.

Database engine

Databases of any size import without loading the dump into memory. The dump is read and replayed statement by statement; a 2 GB dump no longer needs 2 GB of memory on a host without mysqldump. Dumps using DELIMITER or MySQL conditional comments are read correctly.
An import that fails on a statement says which one: its rank in the dump and its first characters, instead of a bare SQL error with no way to find the cause.
Note: database views and triggers are not transferred. They were being dumped as if they were tables, which produced a dump that could not be imported at all — and therefore a snapshot that could not be restored. Views are now left out and named in the report, so you can recreate them on the other side.
Note: a table prefix that does not end in an underscore is not reconciled. Tables are then imported under their source names and the operation says so, rather than renaming tables on a guess.

Content merge

An item can no longer be given the type of another. Matching is now constrained to the post type, so a custom post type sharing an identifier with a post is skipped and named, instead of quietly turning one into the other.
Modification dates survive the merge. Items arrived stamped with the time of the sync, which made every later comparison show the target as "recently edited".
Featured images follow. When the media does not exist on the other side, the item is named in the report and you are told to sync media first — instead of arriving silently without its image.
Items whose author does not exist on the target are named, with the account they were reassigned to.
Note: metadata whose key starts with an underscore is still transferred by default. That is where ACF, Meta Box and Yoast keep their field references, and excluding it would break those merges. Settings → Protected metadata offers an allowlist, or full exclusion, for sites that need it.

Screens

An operation that stops answering says so. The progress card used to spin forever, with no way to tell whether the operation was still running: it now reports the silence after 90 seconds, stops polling, and points to the Logs. Losing the network shows an inline error instead of a browser alert box.
The content selection now works in both directions. The checkboxes were displayed on a pull, but no button ever ran the selection.
A batch sync that fails on one type says "Finished with errors: 1 of 3", marks the failing row and re-enables the button, instead of concluding with a check mark.
An operation whose comparison could not be run no longer starts on trust. The screen states that the two sites could not be compared, offers to analyse them there and then, and requires the site name to be typed before running without a loss check.
Confirmation prompts say what is deployed, where, what is lost and what is not, instead of asking a generic question.
Updating the Agent no longer reloads the page: the version badge updates in place and the result message stays readable.
Logs: a Site column, filters by level and by site, pages of 50, and a "Clear the log" action. The CSV export now opens with its accents intact in Excel.
Accessibility: contrast raised on badges and secondary labels, the information button is a 24 px target with a visible focus ring, live regions announce progress without re-reading the whole card, and keyboard focus returns where it was after a refresh.
Step names in reports and e-mails are translated ("Connection", "Search-replace") instead of showing their internal identifier.
On nginx, the system status screen warns that the .htaccess protecting the storage folder has no effect, and gives the location block to add (see the FAQ).
An invalid notification e-mail or webhook URL is now refused with an explanation, instead of being saved as a notification that would never be sent. Deleting a site also disarms any schedule pointing at it.

WP-CLI and scheduled runs

wp miroir push|pull runs the same data-loss check as the screen and asks for confirmation, listing what it found. --yes keeps its meaning (warn, log and continue), --strict makes those warnings fatal, and --no-loss-check skips the check outright.
Command-line and scheduled runs take the same lock as the interface, send the same notifications, and name their steps in plain words.

Entries for versions before 1.2.0 were removed: the plugin directory truncates any changelog longer than 5,000 words, which cut the recent releases — the ones that matter when deciding whether to update.

Miroir Local Sync

标签

下载

详情介绍:

安装:

屏幕截图:

升级注意事项:

常见问题:

更新日志: