Thingiverse's Cloudflare now challenges every request whose TLS handshake
isn't a browser's, whatever the headers or cookies, so file downloads,
preview images and API calls all failed. Route all thingiverse.com
requests through impit, which impersonates Chrome.
Retry downloads once with a delay when blocked by a 429 or Cloudflare
challenge, then report the failure so the caller can handle it. Limit
Thingiverse download concurrency to 2 to reduce triggering rate limits.
* fix: improve context menu interaction and update check display
Add reopenContextMenuBelow utility to close context menus when clicking
outside and forward the right-click event to underlying elements. This
prevents clicks from being lost when users right-click beside a menu.
Replace UpdateCheckSection status text with UpdateCheckAnimation, an
animated component showing orbiting models during checking and colored
status indicator bubbles after completion.
* feat: add admin pages to search
* feat: add internet check to system health
* feat: expand storage usage reporting with disk breakdown
* feat: add providers admin view
* feat(storage): add background reorganization and admin browser
Adds background-driven storage reorganization for moving model files
when the storage template changes, with an admin interface to preview
changes, check impact, monitor progress, and browse the storage tree
built from plate records rather than filesystem walks.
* feat: new auth admin settings page + OIDC support
* test: fix failing test
The test was timing-dependent: it checked for a 409 while the first run
is still going, but completion depended on when the mock fetch resolved.
Replace the instant mock with a promise the test controls, making the
assertion deterministic. Also save/restore the Thingiverse access token
to prevent test isolation issues.
Admins can now make members admins and delete members from the Users page.
The owner, the first admin, can also demote and delete other admins; nobody
can demote or delete the owner, or themselves. A migration adds the owner
flag and gives it to each existing instance's longest-standing admin, and
the first sign-up of a fresh instance takes an advisory lock so concurrent
sign-ups can't make two owners.
Deleting a user removes their models with every file, their tag mosaics and
storage folder, then the account. Roles are now read from the database on
each request, so a demotion or deletion takes effect at once. Both actions
are logged under the acting admin.
Each tag tile now loads one WebP built by the backend from its newest models'
pictures, instead of up to ten images laid out in the browser. Mosaics are
made on first request and cached under storage/tag-mosaics; their URL carries
a hash of the pictures in them, so tagging, untagging or a new render gives a
new URL with nothing to invalidate, and existing libraries need no backfill.
The summary joins pictures only for the requested page's tags, a single
tag's mosaic looks up just that tag's newest models, and the Tags page asks
for enough tags to fill the screen. Deleting a tag, or all of a user's
models, removes its mosaics.
When the last AI model is deleted, categorization is now forced
off server-side, preventing it from silently restarting when a
new model is added. The frontend distinguishes between loading
and no connections, hiding the feature toggle instead of
disabling it, and removes the unused "Add a model first" message.
Provider prompts now identify category candidates by label/path instead
of numeric index, with duplicate-path and fuzzy-match handling. Add
reasoning_effort toggle with automatic fallback for providers that
don't support it.
Replace blocking single-print recategorize with background-queued flow
backed by notifications. Extend bulk-run scope to include folders and
manual categories alongside existing filters.
* fix(import): stop flagging models that mention robots as a MakerWorld CAPTCHA, and let a manual retry try again
isCaptchaChallenge looked for "captcha" or "robot" in every top-level string of a reply,
design titles and summaries included, so a model named "Robot ..." failed as a CAPTCHA on
every retry and started the 2h cooloff for every other import. Only a captcha key or the
reply's own error, message or msg text counts now.
The cooloff also failed a retry straight away after the user solved the puzzle in the
browser, which is what the error message tells them to do. Starting or retrying a queue job
by hand, a bulk start or retry, and POST /import/jobs/:id/retry now lift it. Automatic
collection syncs still respect it.
* feat(extension): offer "Add to the queue" when MakerWorld blocks an import
A CAPTCHA or the daily download limit won't clear by clicking Import again, so the error
panel lets an admin park the link in the queue with the same collection and profile scope.
* fix(import): let a queue job cut off by a restart be started or retried
A restart marks a running job ERROR with its links still PENDING or RUNNING. The queue page
only offered Start for PAUSED jobs and Retry for jobs with FAILED links, and the API agreed,
so such a job could never run again. The same held for a job whose runner crashed.
Start now takes an ERROR job with links waiting, and a retry reruns links an interrupted
run never reached even when none of them failed.
* fix(ai): leave room for a full answer and explain a cut-off one
max_tokens 300 cut answers off once the reason filled its 300-character
limit, and the run showed JSON.parse's position instead of the cause.
Raise it to 1024 and turn a parse failure into a readable error, naming
the token limit when the provider reports finish_reason length.
* feat(backend): add a shared logger with levels and greppable fields
One line per event with a level, a scope and key=value fields, or one
JSON object per line with LOG_FORMAT=json. LOG_LEVEL drops lines below
it. Errors log their message, status and code, plus the stack at error
level, and bearer tokens are redacted. Nothing outside it changes yet.
* feat(ai,import): log categorization and import runs through the logger
A bulk AI run kept only its last error in memory, so which models failed
and why was lost. Each provider answer it can't use now logs the model,
finish_reason, completion and reasoning tokens and the end of the answer;
each model's outcome and the run's start and totals are logged too.
Collection, Thingiverse, Printables, profile and zip imports dropped a
failed model without a trace; they now log it with its reason, plus the
job's start and totals. The links runner's console lines move to the
logger.
* fix(ai): read an answer wrapped in markdown or prose
Xiaomi MiMo ignores response_format and sent the object in bold,
`**{...}**`, which failed as not valid JSON although the answer was
complete. When the whole text doesn't parse, try the outermost braces.
MakerWorld can leave a design's defaultInstanceId pointing at a print profile that no longer
exists. pickMakerworldInstanceId (extension) and makerworldInstanceIdFromNextData (backend)
returned it unchecked, so the download-link request went to the dead profile and MakerWorld
answered 403 "The client does not have access rights to the content." Use the default only when
it is still one of the design's instances, otherwise the first, as selectMakerworldProfiles
already does.
Fixes#14
Claude-Session: https://claude.ai/code/session_01KGCb65Zhzwd3daF862ZE6R
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Every collection's Sync configuration now works before it's synced: a
Browse tab lists the user's own collections on MakerWorld (through
their cookie, private ones included), Printables and Thingiverse (by a
username asked once), and takes any collection's link. Each provider's
collections are drawn in its own colours and manner.
On Printables and Thingiverse the collection's current models are
imported by the sync's first check. MakerWorld rate-limits bulk
imports, so there only models added from now on come in. Settings can
be picked before turning sync on.
Add POST /collection/:id/items. Thingport Grab uses it to file a
MakerWorld collection's already-imported models instead of one request
per model, falling back on older instances.
The collection panel sent one /import/status request per model, all at
once. When any failed it counted the model as new, so a collection
already imported could offer to import almost everything again.
Add POST /import/status/batch and use it, falling back to paced
per-model checks on older instances. A failed check now shows an error
instead of guessing.
The route replaced every provider error with a bare "AI provider failed",
so a stopped Ollama looked the same as a timeout or a bad answer. Pass on the
reason (the API key is already redacted) and log it.
A base URL with no path, like Ollama's http://host:11434, gets /v1 added.
URLs that already have a path are kept, so OpenRouter (/api/v1) and Gemini
(/v1beta/openai) keep working. Saved values are normalized when read.
Move weather endpoints to the server with caching and rate-limiting. Prioritize
3D viewer renders over background thumbnails. Query featured models from database
instead of loading all into memory.
* feat: categorize models with an OpenAI-compatible AI provider
Models can now be sorted into the user's categories by any OpenAI-compatible
chat endpoint (llama.cpp, Ollama, vLLM, OpenAI, OpenRouter...), configured once
by an admin under settings: mode off/suggest/auto, confidence threshold,
categorize on import, concurrency, optional preview image.
Every print now records where its category came from (manual, site rule, AI,
or legacy for rows that predate this). AI never changes a manual choice: runs
and the import hook skip them, and a single-model request needs force. Folders
are never touched. Category changes go through one setPrintCategory path, so
the name check and file relocation stay shared, and AI writes are a
compare-and-set against what was read before the provider call, so an edit
made meanwhile wins.
Users get per-model recategorize, accept/reject of pending suggestions (a
rejected category is not offered again), and a background "recategorize all"
with scope flags, progress and cancel that stops after repeated provider
failures.
* feat: AI categorization settings, suggestions and bulk runs in the UI
Admin settings get an AI categorization section with a connection test and a
privacy note. The model menu offers "Recategorize with AI" (asking before it
overrides a manual category), the side panel shows a pending suggestion with
accept/reject and marks AI-set categories, and the models page has
"Recategorize all with AI" with scope, progress and cancel, plus a pending
suggestions list with bulk accept/reject.
* fix: keep suggestion actions from overlapping the reason
MUI's secondaryAction is absolutely positioned over a fixed padding, so the
two buttons covered a long reason. They now sit in the row, and wrap under
the text on narrow screens.
* fix: put the clear API key action inside the key field
A full-width outlined button under the field read as an empty input in an
error state. It is now a small action at the end of the field, and the
helper text says the key goes away on save once it's clicked.
* refactor: update AI categorization admin part
* refactor: move ai categorization buttons to categories manager modal
* fix: multiple smaller problems with AI categorization
---------
Co-authored-by: Tautvydas Derzinskas <tautvydasuk@gmail.com>
