Repository navigation
RocketRide Server v3.6.0
·
1024 commits
to develop
since this release
Added
- typescript: add
SignupProperties,PkceCredentialandnormalizeSignupPropertiesfor optional validated signup attribution resolved immediately before PKCE authentication; credentials without context retain their existing bytes (#2890). - shell: freeze additive API v8 with consent-gated
signupPropertiesinstartOAuthandshell:loginRequest, kept only in the PKCE credential and out of native telemetry facts; recheck privacy signals, consent and expiry after redirects and before authentication (#2890). - typescript:
SignupPropertiesgains the optionalconsent_state: 'undecided', which marks a cookieless context from a visitor who has not chosen on the consent banner;normalizeSignupPropertiesaccepts no other state, andconsent_versionstays 1, so consumers checkconsent_statebefore treating a context as consented (#2978). - shell: carry an undecided visitor's cookieless signup context in the OAuth
stateas a{ v: 1, signup }envelope: nothing in site storage, readable and unsigned, kept in browser history and the identity provider's request logs. It is rechecked when built, on the callback and before authentication, and dropped once the visitor accepts or rejects, or sends DNT/GPC (#2978). - nodes: the vector-store tools gain
soft_deleteandrestoreon all nine store nodes (rocketride_vector,store_astra,store_atlas,store_chroma,store_milvus,store_pinecone,store_postgres,store_qdrant,store_weaviate): they move documents to the trash and back over the drivers' existing soft-delete support, are offered besideupsertanddeleteonly while writes through tools are allowed, and reportmatched_chunkswhere the store counts the rows it marks (the two Postgres stores and Chroma).rocketride_vectorandstore_postgresalso offerlist_documents(one entry per document with its parent path and chunk count, paged and ordered by object ID, active documents or the trash) and reportdocument_countinstats; both come from an optional driver capability (DocumentStoreBase.listDocuments,countDocuments,supports_document_listing) any other driver can adopt (#2587) - shell, TypeScript client: the shell reports sign-in starts and completions as product-telemetry facts for the active app to forward to its own analytics:
shell:telemetry({ event, props, ts, id }, typed asTelemetryFactinrocketride/app-sdk) carriesauth:complete(is_new,waitlisted,duration_ms) live, andconnectionManager.drainTelemetry()hands over, once, the facts held while no app listened (up to 32, for 30 minutes, in the tab's sessionStorage),auth:startalways among them since the OAuth redirect follows it at once.shell:loginRequestgainsauto, set by the shell's own auth gate, andConnectResultgainsisNewUser: true or false when the account provider knows whether the sign-in created the account, null when it does not. On the waitlist screen the shell runs the home app's./telemetryentry once per page load (#2858) - nodes:
rocketride_vectorcan expose its store to agents and apps as tools (agents call<node id>.search, apps callclient.toolwith the tool name and the node id):search,objdir,getandstatswhen the new "Expose as tools" setting is on,upsertanddeleteonly when "Allow writes through tools" is on as well. Both are off by default, since the table lives in the team's managed database. The node now sits on the shared store base classes instead of a copy of their lane code (#2587) - builder: keep test server output, show it when tests fail (#2630)
- builder, server, nodes: standalone c++ node build (#2644)
- client-typescript: share app scaffold logic between the CLI and the VS Code extension (#2849)
- nodes: add tool_masky node for Masky AI avatars (#2731)
- shell: serve the Cookie Policy and the Marketplace Publisher Terms as public legal pages (#2900)
- shell, client: switch or leave organizations; invites pre-select the org's development team (#2871)
Changed
- nodes: on all nine vector-store nodes,
getanswers an object ID the store does not hold, or one in the trash, with{"success": false, "error": ...}instead of an empty page;upsertnumbers the chunks sent for one object ID in the order given instead of storing them all as chunk 0, and rejects an explicitchunkIdthat repeats within a call;upsert,delete,soft_delete,restoreandgetrefuse the reserved object IDschema(the store's control row);deleteechoes the distinct IDs it was given and counts them indeleted_count;statscarriesdocument_count, null where the store cannot count documents. On the two Postgres storesobjdirpages count parent paths rather than documents, so a page holdslimitentries while more follow, and lists a document whatever chunk IDs it has (#2587, #2668) - nodes: clarification of the 3.5.0 entry, which named
rocketride_vectoronly: the vector-store tool limits apply to every store node on the shared base, the nine named above (getreturns at most 100 chunks per call,objdirat most 1000 paths, pagination values must be integers), andStoreGlobalBasereads theexpose_toolsandallow_writeconfig keys on every one of them; onlyrocketride_vectordeclares them as settings, the other eight default both on (#2668) - nodes: add tool_tenki - disposable Tenki Sandbox Linux VM for agents: run commands and code, edit files, work in git repositories, private ones through Tenki secret policies (#2302)
Fixed
- nodes:
rocketride_vectorand PostgreSQL vector stores share a search base that continues HNSW scans past soft-deleted and metadata-excluded neighbours on pgvector 0.8+, using transaction-local settings. Searches with no qualifying matches retry exactly so hidden active documents remain searchable; nonempty results stay approximate and avoid a full scan merely to fill the requested limit. Older pgvector versions also retry underfilled results. Unindexed tables avoid the redundant retry. Exact retries can be expensive on large tables. PostgreSQL store connections serialize operations and use explicit transactions for chunk replacement and iterative searches, releasing them after success or failure. - parse: RAR4 and RAR5 archives extracted nothing since #2440, silently and on every platform: the move of the Tika code into the parse node left
RarSevenZipParserbehind in the unbuiltpackages/tika, soConfigBuilderreplaced Tika's RAR parser with a class that was not on the classpath. The class now lives in the parse node library, the swap checks that the class resolves before it removes Tika's own parser, and a RAR5 archive yields its member text again (#2678) - parse: parse tasks that started together, or while another parse task was running, broke each other's 7-Zip-JBinding: every task rewrote the shared native library in
<data>/cachein place (16.02-2.01 never matches its own hash), so on Windows the later tasks lost RAR support silently and on Linux a running task could crash with SIGBUS or SIGSEGV. The parse node now extracts the library itself, reuses a complete copy and publishes a new one only by an atomic rename (#2790) - profiler: the call tree repeated a function's callees under every path that reached it, and the Profiler app's table summed the repeats (2–3× on shared callees). A function's copies now split its callee totals and add up to them, the all-threads tree is built over every thread at once, so work spread thinly over a pool is not pruned away, and
rrext_cprofile_report_treereturnsfunctions, the exact per-function totals the table shows (#2786) - nodes:
parseextracts images, audio and video, embedded in documents or as standalone files, only when something is connected to that lane (the optional Aspose PDF parser still extracts images itself and only skips sending them), and reuses its Tika configuration for embedded objects instead of rebuilding it per object; with the stock Tika PDF parser, PDFs drawn as thousands of inline images parse in about a second on text pipelines instead of minutes, and pipelines that consume the images get the same images several times faster on image-heavy documents (#2745). - anonymize: with
--modelserverthe node sent GLiNER one request at a time: every chunk worker queued on a recognizer-widethreading.Lockaroundpredict_entities, held across the model-server round trip, so a document's ~70 windows and the documents in flight beside it were served strictly one after another (91 % of a parse → anonymize → NER → embed pipeline's node time under load). The guard now comes frommake_device_lock()as in the vision nodes: a no-op with the model server, the same lock for in-process inference (#2862) - nodes: vector-store trash operations preserve Qdrant metadata and update every matching Weaviate chunk (a Weaviate update that does not persist now fails the operation instead of looping); on Qdrant they set only the nested flag (
set_payloadwithkey='meta'), which needs a Qdrant server of 1.8 or newer.objdirexcludes soft-deleted documents on Atlas, Chroma, Milvus, Pinecone and Weaviate, and now returns paths on Chroma and Weaviate: Chroma filtered on achunkkey the store never writes, so it returned nothing, and given a parent it sent a two-fieldwhere(with the parent JSON-encoded) that the client rejects; Weaviate subscripted aDocMetadataand failed. On Qdrantobjdirappliesoffset(the scroll offset is a point-ID cursor, so every offset returned the first page) and, likeget, lists a document whoseisDeletedflag is missing; Chromastatscounts with the server's count instead of reading every record ID (#2587) - nodes: Astra directory listing uses the supported SDK cursor, groups active chunk-zero headers before paging, and orders paths consistently; collection counts pass the required SDK bound and report the server's estimate once the exact-count ceiling is passed (#2587).
- ai: Whisper no longer falls silently back to the CPU on a GPU host: the requirements pin
ctranslate2>=4.8,<5so a stale engine cache cannot keep the 4.7.x build the GPU probe refuses, the probe runs a real transcription on the GPU in a child process before the model is declared GPU-ready, and a fallback is logged at error level with its cause and stamped asdevice_fallbackin the model metadata. The first GPU load in a process runs thetinymodel once when its weights are local (fetched beforehand when the hub is reachable, waiting at most 30 seconds), which adds a few seconds; without them the probe stops at the storage-view check, as before, and the GPU is still used; a probe that runs out of GPU memory is reported and not cached, except that in model-server mode a second out-of-memory result after the storage-view check caches that successful check (#2498) - ocr: with
--modelserverthe OCR node sent one image at a time: every worker thread queued on a node-widethreading.Lockaround the reader, so the model server never had concurrent OCR requests to batch. The guard now comes frommake_device_lock()as in the other vision nodes: a no-op with the model server (+53% to +63% documents per second ondocument-processor.pipe), the same lock for in-process inference (#2820) - embedding_video: with
--modelserverthe node embedded one frame at a time across every video in the task: each frame's CLIP/ViT call was wrapped in a task-widethreading.Lockheld across the model-server round trip. The guard now comes frommake_device_lock()as in the other vision nodes: a no-op with the model server, the same lock for in-process inference; a single video's frames are still embedded in order (#2863) - profiler: the SDK docs promised a
RuntimeErrorfor an active or missing session, which comes back as{'status': 'error'}, andROCKETRIDE_CLI.mdadvised--min-pct 0on short runs; the docstrings, JSDoc and agent docs now give the real outcomes and the CLI's no-data exit codes (#2876) - ai: a bare port in
--modelserver(--modelserver=5590) is normalized tolocalhost:5590, the model server on the engine's own loopback; the Python client read the digits as a hostname and connected tows://5590:5565/models. A model server on the Docker host still needs an address such ashost.docker.internal:5590; remapping for Docker tasks is tracked in #2593 (#2679) - nodes: the
objdirtool ofstore_postgresandrocketride_vectorno longer lists the store's internalschemacontrol row (it appeared as aNonekey) or soft-deleted documents (#2587) - ai: hosted engines never install or import torch, model server or not (#2682)
- ai: say the provider may reject an output limit above the model's, not will (#2692)
- builder: run a bracket's teardown when a step inside it fails (#2777)
- cli: close or continue the browser tab after login (#2852)
- config: ignore editor name metadata in typo warnings (#2959)
- docs: fix the footer Instagram link and add TikTok (#2860)
- engine: intern Python error locations so stored errors stay valid (#2797)
- mcp: require exp, iat and iss on /mcp tokens explicitly (#2796)
- nodes: c++ node build deps (#2779)
- server: Linux crash dumps work on hosts that set
fs.suid_dumpable=2(#2865) - shell: keep /marketplace and /build-publish-earn out of sitemap.xml and llms.txt until they relaunch (#2953)
- sql-ui: discover RocketRide SQL nodes as connections (#2686)
- task: refuse a pipe to an org that is out of credit (#2641)
- task: load placeholder values at each start instead of keeping them (#2794)
- task: a start that fails inside a restart ends the run as stopped (#2834)
- vscode: preserve pipeline files and IDs in the sidebar watcher (#2615)