Place folder is archived mid-sync, leaving the mirror empty (ENOENT on rename)

Back to board
## ▼ FIELD 1 — Steps to reproduce **Environment** | | | |---|---| | MCP server / plugin | **2.12.1** on both incident days; FIELD 5 measured live on **2.12.2** | | Measured on | 2.12.2 | | AI client | Claude Code 2.1.226 | | Model | Claude Opus 5 | | OS | Windows 11 Pro (10.0.26200) | | Studio | 0.733.0.7330989 | | Tier | PRO | **Steps** We do not have a deterministic recipe — the archiving is intermittent. The name divergence that precedes it can be observed live on the current version (FIELD 5). What reliably precedes the damage: 1. Work in a Studio session in which `game.Name` is not stable. For one and the same `placeId` we see two folder names alternating under `weppy-project-sync\`: the real place name, and Studio's generic document name of the form `Place<N>` — the counter Studio uses for the documents opened in a session. 2. Start a sync (or let it auto-start after a reconnect). 3. While it is building `explorer_tmp_<uuid>`, the place folder for that `placeId` is moved to `_deleted-place-backups\place_<id>\<UTC-stamp>_place_<id>_<name>\`. 4. The sync finishes its work and renames `explorer_tmp_<uuid>` → `explorer` **at the old path**, which no longer exists. **How to see the evidence after the fact** — this is the part we can hand over completely: ```powershell # every archive event, with its UTC timestamp in the folder name Get-ChildItem "<userDataDir>\weppy-project-sync\_deleted-place-backups\place_<id>" -Directory # any explorer_tmp that never became an explorer Get-ChildItem "<userDataDir>\weppy-project-sync" -Recurse -Directory | Where-Object Name -like 'explorer_tmp*' ``` --- ## ▼ FIELD 2 — Expected behavior A sync session that is writing into a place folder either keeps that folder for the duration of the run, or is aborted cleanly with a clear error. Archiving the directory out from under a running sync should not be possible; if it is, the sync should notice and fail loudly rather than attempting a rename against a path that no longer exists. The mirror should not end up empty as a result. --- ## ▼ FIELD 3 — Actual behavior **3a — The physical proof: the in-flight temp directory is inside the archive folder.** On 2026-08-04 the plugin reported (captured from the in-memory log buffer; paths shortened): ``` [SyncManager] Error: Failed to send sync complete: handleSyncInit: ENOENT: no such file or directory, rename '<syncRoot-B>\explorer_tmp_931937ef-8a88-49da-b2bf-3e99db3e8880' -> '<syncRoot-B>\explorer' ``` That exact directory exists today — **not** at the path in the message, but here: ``` _deleted-place-backups\place_<id>\2026-08-04T15-14-13-808Z_place_<id>_<name-A>\ explorer_tmp_931937ef-8a88-49da-b2bf-3e99db3e8880\ (756 files, mtime 15:14:19Z) ``` So the folder the sync was writing into was moved to the archive at **15:14:13.808Z**, and the temp tree — 756 files — travelled with it. The rename then failed because its parent path was gone. Note the two names: the error message points at folder name **B**, the archive that received it is named **A**, for the same `placeId`. The archive folder contains **no `explorer`** — the mirror was empty afterwards, which is the user-visible damage. **3b — This is not a one-off.** Archive events for a single `placeId`, from the folder names: | Date | Archive events | |---|---| | 2026-08-03 | 1 | | 2026-08-04 | 4 | | 2026-08-06 | **10** | | 2026-08-07 | 2 | | 2026-08-10 | 1 | | **Total** | **18 in 8 days** | Two of them are millisecond pairs — e.g. `15-10-28-528Z_place_<id>_<name-A>` and `15-10-28-529Z_place_<id>_<name-B>`: both name variants of the same `placeId` archived in the same operation, one holding the full mirror (5052 files), the other nearly empty (3 files). **To be precise about what these 18 are:** they are archive *events*, not 18 damage events. Only the two listed in 3c verifiably caught a sync in flight (their `explorer_tmp` tree travelled into the archive). The rest carry a complete `explorer` and did no visible harm — e.g. the most recent one, `2026-08-10T14-30-34-925Z`, contains 5256 files including a full `explorer`. **The defect we are reporting is the race, not the archiving as such.** The 18 matter only as a measure of how often the window is opened. **Version coverage, measured from the Studio log rather than from memory:** on 2026-08-10 the plugin loaded as v2.12.1 at 13:09:59Z and as v2.12.2 at 16:15:06Z. The archive event that day is stamped 14:30:34.925Z, so it fell in the 2.12.1 window — every one of the 18 events above ran on 2.12.1 or earlier. We have seen no archive event since 2.12.2 started, but that is roughly one day of observation on an intermittent fault, so we are reporting it as coverage, not as evidence of a fix. **3c — Leftover temp directories, all inside archives.** Scanning the whole sync directory for `explorer_tmp*` returns exactly three, and two of them are the ones that were carried into archives: | Location | Files | mtime | |---|---|---| | archive `2026-08-04T15-14-13-808Z…` | 756 | 2026-08-04 15:14:19Z | | archive `2026-08-06T15-23-05-568Z…` | 235 | 2026-08-06 15:23:24Z | | a different placeId, archive `2026-08-03T16-52-51-738Z…` | 50 | 2026-08-03 16:53:04Z | In every case the temp directory's mtime is a few seconds **after** the timestamp in the archive folder name. **3d — A second, independent ENOENT inside `explorer_tmp`**, from a different day and a different syscall. This one is in Roblox Studio's own log file, so it is persistent and quotable: ``` 2026-08-09T14:43:47.265Z Warning [FLog::CreatorWarning] [SyncManager] Error: Failed to send chunk 10/15 for Workspace: handleSyncInit: SyncWriter.writeProps failed for <syncRoot>\explorer_tmp_f36a0da1-…\Workspace\…\<node>.props.json: ENOENT: no such file or directory, mkdir '<syncRoot>\explorer_tmp_f36a0da1-…\Workspace\…\<node>' ``` Here the write **into** the temp tree fails, not the final rename. Same theme: a path under `explorer_tmp_<uuid>` is not there when the sync expects it. --- ## ▼ FIELD 4 — Already ruled out - **MAX_PATH.** The `explorer_tmp_<uuid>` segment is 41 characters longer than `explorer`, and in the current mirror 201 paths would cross 260 characters with that prefix (longest already 291 without it). We tested it anyway on this exact machine and Node build (`v24.16.0`): `mkdir`, `writeFile` and `readFile` all succeed at path lengths of 267, 309, 351 and 393 characters. The failing path in 3d is 219 characters. **Long paths are not the cause.** - **Not a disk or permission problem.** Both incidents are surrounded by successful syncs on the same day, into the same directory, by the same user. - **Not the EPERM issue we filed separately.** Those are `rename` failures on `observability\changelog\sessions\*.jsonl` with code EPERM and a working retry loop. These are ENOENT, in the mirror tree, and the mirror ends up empty. **What we could not determine, and are not guessing about:** what triggers the archive step, and whether the name alternation is the cause of it or merely happens alongside it. We can see that the same `placeId` carries two folder names, that archive events cluster on days with heavy session churn, and that one archive verifiably swallowed a running sync's temp tree. We cannot see your archiving logic, so we are not going to claim a mechanism for it. **Where it hurts us**, for prioritisation: an empty mirror is silent. Static checks that read the mirror keep reporting success against a stale or absent tree — we lost a working session to that before we understood what had happened. **Suggested direction (yours to judge).** The `placeId` is stable while the folder name is not. Deriving the sync root from the stable identifier, or refusing to archive a folder that has an active sync session, would both remove the window this falls through. --- ## ▼ FIELD 5 — Live check on 2.12.2: one lead tested and dropped, one that survived Added 2026-08-11, after testing a reading of ours rather than shipping it. **Dropped — "the server reports a `syncRoot` that does not exist".** While sync was `idle` after a failed init, `manage_sync status_current_place` returned `syncRoot` = `<root>\place_<id>`, with no name suffix and matching no directory on disk. We were about to file that as a finding. We then re-measured with a **healthy running sync** (`state: syncing`, 5409 instances / 104 scripts, full sync at 09:44:49Z) and got the correct path, `<root>\place_<id>_<name-A>`. **So the suffix-less value is only an idle/uninitialised placeholder, not a wrong path in use. We withdraw it** — please do not spend time on it. **Survived — the two names are live at the same time.** With that same healthy sync running, the two sources still disagree about one `placeId`: | Source | Value | |---|---| | `system_info place_info` → `name` | `Place<N>` — Studio's generic document name | | `manage_sync status_current_place` → `syncRoot` | `<root>\place_<id>_<name-A>` — the real place name | Both readings were taken seconds apart, with sync healthy and a full sync just completed. This is the divergence described in FIELD 1, observable directly rather than inferred from folder names: the active sync root carries `<name-A>`, while what the plugin currently reports as the place name is `Place<N>`. If the root were derived afresh at this moment from the reported name, it would point at a different directory — and `_Place<N>` and `_<name-A>` are exactly the two variants that appear in the archive folder names in 3b, including the millisecond pair. **A third source, from your own failure records.** Later the same day a `failures.jsonl` entry was written for this place, and it carries the place name as a field: ```json { "placeId": <id>, "placeName": "<name-A>", "code": "EBUSY", … } ``` So at the same moment, `place_info` reports `Place<N>` while your failure record independently writes `<name-A>` — the same two variants, from a third code path. Whatever supplies the name is not the same source everywhere. **What this is and is not.** It is a live demonstration that the value the folder name depends on is not stable while the `placeId` is. It is **not** a reproduction of the archiving itself — that remains intermittent, and we still cannot trigger it on demand or see what decides it. --- ## ▼ Feedback Several of our earlier reports have shipped — 2.11.2 sync index recovery, 2.11.1 `get_source` line ranges, 2.12.0 Luau write validation plus failure isolation, 2.11.0 `preflight`. Thanks for that; the failure isolation in particular is why we can now measure things instead of guessing at them.