# V4.18.0

> FastGPT V4.18.0 release notes covering configuration changes, upgrade steps, and feature updates.

Source: https://doc.fastgpt.io/self-host/upgrading/4-18/4180
Last Updated: 2026-11-13

## Upgrade guide

**Prerequisite**: Complete the required historical migrations on 4.17.x before upgrading. Version 4.18.x removes the old task registrations and implementations, so they will no longer run; historical status and error records remain in the database.

### Configuration changes

* Added `SERVICE_AXIOS_MAX_CONTENT_LENGTH` (default: `20` MiB) to cap decompressed bodies of buffered responses from the shared server-side HTTP client; requests exceeding the limit are aborted, so custom integrations that need larger responses must raise it. Explicit caller limits take precedence, and streamed downloads remain subject to caller or business quotas.

* The default for `HNSW_EF_SEARCH` / `systemEnv.hnswEfSearch` changes from `100` to `200` for PostgreSQL, Milvus, OceanBase, SeekDB, and openGauss; explicit settings still apply and can be tuned to balance recall and query time.

* `HNSW_MAX_SCAN_TUPLES` / `systemEnv.hnswMaxScanTuples` no longer takes effect, and its admin setting has been removed. PostgreSQL now sets the scan limit automatically from the largest Dataset vector row count involved in the current indexed search, capped at `2147483647`; existing deployments can remove the old setting.

* Added `SHOW_OUTLINK_BRAND`: without `PRO_URL`, it is always `true`; with Pro connected, it defaults to `false` and can be set to `true` to enable branding. The frontend `show_outlink_brand` value comes from the server environment and overrides the same setting in `config.json`.

* The Commercial edition adds `auth.sms.MODEL_DEPRECATION` for the model deprecation SMS template. SMS reminders currently use this Chinese template regardless of the user's language; `auth.sms.MODEL_DEPRECATION_EN` has no effect on SMS delivery. Configure the Chinese template ID in the admin console before enabling SMS reminders.

* **Plugin service deployment**: Docker Compose templates now set `fastgpt-plugin`’s `FASTGPT_BASE_URL=http://fastgpt-app:3000`, point `MONGODB_URI` to a separate `fastgpt-plugin` database, and explicitly set `DB_MAX_LINK=100` to avoid sharing collections with the main database. Custom deployments should use an address for the main service that the Plugin container can reach. If the Plugin container logs `401` during a file upload callback, check this address first, correct it, recreate the Plugin container, and run the Tool again. Before replacing an existing deployment’s template, check which database the Plugin service currently uses and keep its existing dedicated connection; if a database change is needed, back up and migrate the existing Plugin data first.

* Default deployments now use RustFS for object storage; existing deployments can keep their current storage, and switching is covered in [MinIO to RustFS migration](https://doc.fastgpt.io/self-host/docker_minio_to_rustfs).

* **Content moderation Plugin migration**: The legacy built-in Baidu text moderation (`censor.BAIDU_TEXT_CENSOR_CLIENTID` / `censor.BAIDU_TEXT_CENSOR_CLIENTSECRET`), custom censor URL (`censor.customCensorURL`), and per-model `censor` flag in the commercial edition are deprecated and replaced by the `moderation` Plugin system managed by `fastgpt-plugin`.
  * If you used legacy Baidu moderation, go to **Admin Console > Content Moderation > Moderation Policy** to configure and enable the official **Baidu Content Security (`baidu`)** Plugin.
  * If you used a custom moderation URL (`customCensorURL`), install the official **Legacy Custom Moderation (`legacy-custom-censor`)** Plugin to connect your existing HTTP service without code changes, or build a native `moderation` Plugin with `@fastgpt-plugin/sdk-factory`.
  * For a full comparison of legacy and new behavior, field mappings, and Plugin development steps, see the [Content moderation upgrade and migration guide](https://doc.fastgpt.io/self-host/migration/content-moderation).

* **New deployments**: The default MongoDB version in the `main` Compose templates is now `8.0.34`. Startup and health checks use `mongosh`, and dependent services wait for the replica set primary to become ready.

### Image updates

This version is still in development. Image tags will be added at release.

### Upgrade scripts

These migrations run automatically; no manual execution is required.

1. Added **Initialize Dataset vector indexes** (`20261022_initialize_dataset_vector_indexes`), which runs automatically in the background after startup for PostgreSQL without blocking the App: it first verifies and drops the legacy global `vector_index` online, then counts vectors in active Datasets in batches and queues index maintenance jobs. A successful migration means counting and queueing are complete; index builds continue in a globally serial queue. If cleanup or scanning fails, resolve the issue and retry from the upgrade management page to resume from the saved checkpoint. A custom index with the same name but a different definition is preserved and causes an error.
2. PG partial indexes require `halfvec` support, a fixed vector dimension in the `1–4000` range, permission to create and drop indexes online, and sufficient memory, shared memory, and disk space. Datasets whose indexes are not ready or have failed continue to use exact search, which may increase query time. Failed builds do not automatically retry with a lower resource budget; after the issue is resolved, the next actual vector count queues another attempt. A cache hit does not trigger a retry.
3. Added **Backfill historical chat channels** (`20261104_backfill_chat_publish_source`), which runs independently with batched reads and writes in the background after startup without blocking the App or waiting for long-running tasks in the regular background queue. It fills `publishSource` separately for chats and statistics records without changing legacy fields, identities, messages, or counters. Historical API, MCP, and deleted channels may retain only a type without a recoverable channel ID; unrecognized legacy sources are classified as `unknown` without guessing their channel. Records that previously failed source enum validation can be retried directly from the upgrade management page without modifying legacy data or resetting progress; other conflicting references or invalid data remain recorded as failures and can resume from the checkpoint after repair.
4. Added **Backfill API Key types** (`20261104_backfill_openapi_key_type`), which scans all keys in batches, fills missing `type` values as `personal`, and stores valid `appId` values as BSON ObjectId values. Credentials, ownership, quotas, and default target or binding semantics remain unchanged. Repair invalid records and retry from the upgrade management page. Management queries support only normalized data. Legacy-format data present during migration or written by old instances is outside query compatibility support. Legacy records written after the initial scan boundary are not automatically backfilled and require separate normalization.

### Compatibility notes

* **Legacy API keys**: Keep your existing keys. Keys without a `type` are automatically classified as personal keys, with the stored `appId` retained as the default chat target. Their access permissions and personal management remain unchanged. New App keys are limited to their bound App and managed jointly by its administrators.
* **API Key health checks**: `GET /api/support/openapi/health` now reads credentials from the `Authorization: Bearer <apiKey>` header instead of the URL’s `apiKey` parameter. Update existing health probes to send the key in this header and remove it from the URL.
* **System model endpoint authentication**: If an existing system model sets `requestUrl` but relies on the default `AIPROXY_API_TOKEN` for authentication, explicitly configure that endpoint’s `requestAuth` before upgrading. Without it, direct requests through the OpenAI SDK and Pi Agent no longer fall back to default credentials.
* **Dynamic selection options**: User selection Nodes and single- or multi-select form fields now keep only the first `200` valid variable-bound options at runtime, with each option truncated to at most `100` UTF-16 code units. Workflows that depend on longer option values or later candidates must adjust their candidate data and review downstream matching logic.
* **API identity parameters**: API sessions no longer allow switching user identity; existing sessions use their saved identity. Continue with the same `customUid`; omitting, clearing, or changing a bound value, or switching from the default to a custom identity, is rejected. Integrations that reuse a session across users must allocate separate `chatId` values for different users. A nonempty `customUid` requires an API Key. It is mutually exclusive with `authProxy`, even when empty; integrations that send both must choose one identity.
* **Session ID length**: Session management, file, citation, and Sandbox APIs now limit `chatId` to `1024` characters. Integrations that use longer custom IDs must adjust their generation rules and review existing oversized session IDs and their linked records.
* **Concurrent generation responses**: Non-streaming `chat/completions` requests now return HTTP `409` when the same session already has a generation in progress. Custom clients that depend on the previous error status code must handle this conflict response.
* **Log CSV exports**: Scripts that parse Team members from the “User” column must support `ID:name` instead of the previous member name; when a member has no name, the column contains only the ID. Channel columns are now split into “Call channel” and “Channel name,” so scripts that parse column names must also be updated.
* **Collection deletion API**: Select exactly one request-body form: `id`, `collectionIds`, or `datasetId + apiFileIds`. Legacy `query.id` remains supported but is deprecated; new integrations should use the request body.

## 🚀 New features

1. **\[Tools and channels]** Added App API Keys bound to a single App, managed separately from personal keys on the App publishing page. App managers share control of these keys for that App’s chats and chat management, independently of the owning member’s App permissions; the Commercial edition transfers key ownership through resource handover, and deleting an App also removes its bound keys.
2. **\[Tools and channels]** Chat and related management APIs now share support for external user IDs and Team member proxies. Use `customUid` to narrow access to a user’s history, messages, files, citations, and Sandbox resources; personal keys with proxy permission can use `authProxy` to act with the represented member’s permissions.
3. **\[Tools and channels]** Chat logs now include an “Authorizer” column, hidden by default, separate from “User” to distinguish your users from the authenticated Team member. API `customUid` values appear unchanged as the user, and CSV exports follow the selected columns. Log columns now size to their contents so long UIDs do not squeeze member names.
4. **\[Tools and channels]** Added FastGPT branding to public share pages across desktop, mobile, and embedded layouts; when enabled, new links show it by default, and desktop visitors can dismiss it for the current visit. Persistent hiding requires a valid plan entitlement, configurable per plan or individual subscription in the Commercial edition; legacy links without a saved preference remain exempt, while links with a saved hiding preference show branding again if the entitlement becomes invalid and hide it again when the entitlement is restored.
5. **\[Tools and channels]** Added Plugin-based content moderation, disabled by default. When enabled, policies can independently check user input, model context, and final output, including streamed output, with configurable actions for suspected risks and moderation service errors. Admins can configure Plugins, test moderation results, and view global moderation records; chat details also show the records for that chat.
6. **\[Workflows and Agents]** Workflow Agent Nodes can now be added from the template panel and used as Tool Nodes by other Agents. Dataset inputs support variable bindings; standalone ChatAgent Apps can also be referenced as Workflow child Apps or Agent Tools, with file inputs and resumable interactions for composing multi-Agent Workflows.
7. **\[Workflows and Agents]** Added MCP and HTTP Toolkit Nodes that you can configure directly in a Workflow without first creating a standalone Tool App. Customize child Tool descriptions and authentication, debug each Tool, and let Tool call Nodes use the latest remote Tool definitions at runtime.
8. **\[Workflows and Agents]** Commercial edition Workflow Builder can build, validate, and repair Workflows directly from clear creation or modification requests, while still waiting for preview confirmation when you ask to review a proposal first. Before delivering a candidate, the same Agent checks requirement coverage in its existing context; you still apply the candidate to the canvas manually.
9. **\[Workflows and Agents]** Commercial edition Workflow Builder supports switching between independent tasks within a session, saving each task’s requirements, proposals, and draft separately so work can continue from saved data after context compression or an interrupted run. Recovery still checks task status and the canvas baseline; unsaved drafts may be lost if the sandbox is lost.
10. **\[Workflows and Agents]** App templates now support experience links, with Commercial edition management for links on custom and Plugin system templates, so template integrations can offer a preview before copying.
11. **\[Models and billing]** System models now support an “Upcoming deprecation” flag in model management and model and speech selectors; Commercial edition admins can create notification plans with multiple rounds for Teams whose currently published Apps or Datasets statically use these models, sending in-app messages with optional SMS or email. Plans support failure inspection, run recovery, retries, and termination, and stop new deliveries at the deadline; model removal remains a separate action.
12. **\[Models and billing]** Commercial edition Team models can now use custom request URLs and authentication through configuration APIs or templates as a transitional option for direct chat, Agent, embedding, reranking, and speech requests, retaining endpoint settings after edits. Endpoints undergo safety checks when saved and requested, including redirect restrictions; model lists hide endpoint and authentication details.
13. **\[Models and billing]** OpenAI and Azure Model providers now offer Auto, Responses, and Chat Completions as upstream chat protocols, with new providers defaulting to automatic selection based on the model. This affects only chat models, and the upstream must support the selected protocol.
14. **\[Models and billing]** Model management now supports dragging models into a custom order and canceling before saving; the root account can reorder system models, and with team models enabled in the Commercial edition, members with the “Install model” permission can reorder their own models. “Available models” groups your own team models, shared team models, and system models in that order and applies the custom order within each group; JSON configuration updates preserve this order, which the client also uses for its final system-model fallback. Newly created or newly imported models appear at the top of their group without changing the relative order of existing models.
15. **\[Models and billing]** Model tables with permission configuration now show creators and identify system models separately. Member display pages can load current creator details, while shared-link catalogs do not expose member information.
16. **\[Models and billing]** Usage in the Commercial edition now supports mutually exclusive member, department (including subdepartments), or group filters, with the same scope applied to details, charts, and CSV exports. Selections use current organization and group memberships; members with Team management permission can view usage for the selected members, while other members can view only their own usage.
17. **\[Datasets and files]** With Dataset enhancement enabled in the Commercial edition, the template FAQ import API now supports optional automatic index generation using both questions and answers as input. Automatic indexing tasks that exceed the processing length limit fall back to standard vector indexing, retaining each entry and its input order.
18. **\[Datasets and files]** Dataset search adds a rerank candidate limit of `10–100` chunks, with a default of `100`. Candidates are sent to the rerank model in text recall order after fusion and deduplication, helping balance search quality, latency, and model load; after successful reranking, text chunks beyond the candidate limit are excluded from weighted fusion so they cannot bypass the similarity filter.
19. **\[Datasets and files]** API file Datasets now support batch deletion by external file or directory IDs through `datasetId + apiFileIds`, including the corresponding derived folders and child documents, without first looking up internal Collection IDs. Deletion still requires ownership of each deletion root.
20. **\[Teams and permissions]** Commercial edition Team audit logs now combine date, member, and event filters, defaulting to the latest `7` local calendar days including today, with older dates and inactive members available. Logged-in members with Team management permission can export all matching records as localized CSV, up to `50000` records per export; narrow the filters above that limit. Log timestamps in the CSV use the query start offset, and each export request creates an audit record.
21. **\[Teams and permissions]** Commercial edition user management APIs now include Team memberships, member names, and department paths. Team member lists also return login usernames to distinguish accounts from member identities.
22. **\[System and operations]** Chats now record the channel used by the latest run. Logs show channel types and names separately, with channel IDs available for inspection; filters, CSV exports, and Commercial edition statistics also use the new channel classification.

## ⚙️ Improvements

1. **\[Workflows and Agents]** Fixed internal Node memory being overwritten when multiple child Apps or Workflow Tools call the same target. Content extraction, question classification, and Pi Agent state are isolated by the full call path, including cloned Apps that share descendant Tools and recursive calls back into the root App.
2. **\[Workflows and Agents]** Fixed missing memory updates from Agents, Tool calls, and nested Workflows, loss of memory from branches that did not run, and concurrent child Workflows overwriting other Nodes' new state with stale state. Only each child Workflow's updates are merged, preserving other valid memory.
3. **\[Workflows and Agents]** Fixed memory propagation and failure isolation in loops and parallel runs. Loops pass memory between successful iterations and preserve checkpoints when paused for interaction; failed iterations do not commit updates. Parallel runs merge only successful task results, keep failed attempts from affecting retries, and merge in input order, with later successful tasks taking precedence for the same memory key.
4. **\[Workflows and Agents]** Unified Tool, Skill, and Dataset cards and status checks across Agent forms and Workflow Nodes. Running, publishing, and single-Node Debug use the same rules, report invalid, offline, or unconfigured resources, and preserve existing configuration during asynchronous selection and saves.
5. **\[Workflows and Agents]** Commercial edition Workflow Builder previews now show a title, a short summary, and a pending-confirmation status. Restored proposals awaiting confirmation open at the top of the card and preserve the reading position during subsequent rendering; confirmation, revision, and cancellation use the shared interaction controls.
6. **\[Workflows and Agents]** Raised the option limit for single- and multi-select configuration from `50` to `200`. Manual options in User selection Nodes are capped at `200`, and each option text input accepts up to `100` UTF-16 code units.
7. **\[Workflows and Agents]** Standardized button and model selector sizes across App, Workflow, and Tool editors, and aligned help icons with text for a more consistent layout.
8. **\[Datasets and files]** PostgreSQL now maintains FP16 partial HNSW indexes per Dataset, building them on demand above `10000` vector rows and removing them below `5000`; the existing state is retained between these thresholds. Datasets without a ready index use exact search, and candidates from all groups are ranked by FP32 distance without changing the precision of stored vectors. This substantially reduces compute and storage overhead.
9. **\[Datasets and files]** In the Commercial edition, moved tag filtering for simple Apps, ChatAgents, and Workflow Agents to the “Search filter” tab in Dataset search settings and added a dedicated toggle. Turning it off in manual input mode preserves conditions for later editing, and the summary reflects its state; variable-bound Datasets or filters preserve their configuration so tags that cannot be resolved while editing are not removed.
10. **\[Datasets and files]** Tripled the per-minute file upload rate limit for each upload identity, reducing “Too many requests” errors during repeated or batch uploads.
11. **\[Datasets and files]** Fixed website sync in the Commercial edition missing child links on directory pages without valid content and incorrectly deleting existing Collections after invalid content, write failures, or page-limit truncation; link deduplication and relative resource resolution after redirects are improved, and incomplete syncs preserve existing data. Early stops and timeouts now clean up in order with bounded waits, preventing subsequent writes and retries after cancellation while database operations already in progress finish through their normal mechanisms.
12. **\[Models and billing]** Consistently place models in test mode last in the mobile model selector, speech playback, channel bindings, and model log and dashboard filters, preserving catalog order within each group.
13. **\[Models and billing]** Aligned header and body column widths in the model management table and isolated edit dialog state by row, reducing full-table rerenders when opening a dialog in a large list.
14. **\[Models and billing]** Improved truncation and hover text for long names in model and speech selectors, keeping names from stretching controls or crowding the voice column.
15. **\[Chat experience]** Improved chat history sidebar refreshes with batched queries for long lists and retries while a chat is running or a request has failed. Hidden tabs pause periodic refreshes, resource switches discard late responses, and failed refreshes preserve the previous status and read flags.
16. **\[Chat experience]** Fixed inaccurate generation status and resume connections waiting indefinitely after long periods without output or client disconnects. Persisted chats now use independently renewed Redis leases, which expire within 60 seconds of the last renewal if the process exits; acquisition or read errors trigger fallback behavior, so mutual exclusion and accurate status cannot be guaranteed during Redis failures.
17. **\[Teams and permissions]** Fixed App creation, JSON imports, copies, and conversions to new Workflow copies failing when they include unreadable or missing resources. The initial version snapshots only readable resources; any remaining unusable Node references are marked as errors in the detail view for later correction.
18. **\[Tools and channels]** API documentation now groups parameters by resource, identity, and pagination mode, highlights fields for the selected branch, and corrects required-field markers and request examples to clarify each integration option.

## 🐛 Fixes

1. **\[Workflows and Agents]** Fixed canvas errors or stale parent output types when a Loop end Node has an empty or incomplete reference, or its source Node has been deleted; loop and parallel containers now fall back to a generic array type.
2. **\[Workflows and Agents]** Fixed context, plans, or answers being lost when Agents resume questions as child Tools or Apps, and unfinished interaction history affecting downstream Nodes or the next loop iteration. Pause state is saved per Node, and resumes produce the complete answer.
3. **\[Workflows and Agents]** Fixed root sessions being unable to browse or download files after a nested Agent successfully prepared a sandbox. Readiness now records actual sandbox use for the session; later rounds without sandbox use or errors after readiness do not clear this state, and system and Team permissions still apply.
4. **\[Workflows and Agents]** Fixed parent, child, or sibling Agent and Tool call Nodes overwriting same-named files from different sources in a shared sandbox during the same run. Unique names and model-visible paths are assigned consistently; cross-Node uploads with the same name and source reuse the name allocated in that run.
5. **\[Workflows and Agents]** Fixed Nodes selecting different Skills clearing one another’s deployed versions and initialization state. Published Skill versions and successful initialization state remain cached in the sandbox; redeploying a missing directory clears its old marker so initialization runs again.
6. **\[Workflows and Agents]** Fixed concurrent child Apps failing immediately when competing for a shared sandbox’s initialization lock. A busy lock now triggers bounded retries; timeouts and lease errors still return an initialization error.
7. **\[Workflows and Agents]** Fixed initialization clearing stop requests received during preparation for Workflows and Commercial edition auxiliary generation. Workflow runs without SSE, including MCP calls and scheduled triggers, now also respond to explicit stop signals.
8. **\[Workflows and Agents]** Fixed saves failing when legacy Agents fall back to manual Dataset selection but retain old variable references. Manual Dataset selections in Agents and Dataset search Nodes now filter malformed entries while preserving the remaining selections.
9. **\[Workflows and Agents]** Fixed Skill exports failing when the edit sandbox was not running: when initialization is allowed, the server starts or resumes the sandbox and waits until it is ready before packaging. Members with write permission must wait for the sandbox to be ready; read-only collaborators can access export regardless of their local edit sandbox state.
10. **\[Workflows and Agents]** Fixed Agents dropping the configured time granularity and start/end limits when passing interactive form fields for a date/time or time range. The original field settings are now preserved.
11. **\[Workflows and Agents]** Fixed cursor resets in multiline text fields that disrupted input method composition and candidate confirmation.
12. **\[Workflows and Agents]** Fixed scrolling in Tool and form parameter configuration dialogs when their content exceeds the available space. Menus in dialogs such as Collaborator permissions and HTTP authentication settings no longer get clipped by the dialog body or expand its scroll area; long menus scroll internally so you can reach the bottom options.
13. **\[Workflows and Agents]** Fixed Agent maximum output tokens and answer visibility settings not taking full effect. Hiding the answer preserves required interactions and Node results; with error capture enabled, errors go to the error branch instead of being appended to the visible answer.
14. **\[Workflows and Agents]** Fixed Pi Agent requests being rejected by model providers when Tool parameter schemas contained legacy required flags or internal extension fields. Tools, Sandbox Tools, and built-in Tools now normalize schemas before model calls.
15. **\[Workflows and Agents]** Fixed untranslated field hints for text editing, Dataset citations, and custom feedback during assisted Workflow generation across Simplified Chinese, Traditional Chinese, English, and Korean. Text editing hints now include the complete reference syntax for upstream outputs and global variables.
16. **\[Datasets and files]** Fixed incomplete scope constraints in Collection subtree queries and permission synchronization. Traversal and updates now stay within the current Team and Dataset, preventing invalid parent-child links from affecting other Datasets.
17. **\[Datasets and files]** Fixed Alibaba Cloud OSS rejecting stream uploads when a known length was supplied but chunked encoding was also added, conflicting with `Content-Length`.
18. **\[Datasets and files]** Fixed filenames containing escaped quotes being truncated when parsing `Content-Disposition`. Quotes and backslashes in the filename are now restored correctly.
19. **\[Chat experience]** Fixed input suggestions stopping during continuous typing or after moving the cursor, and replacing the wrong text. Use the up and down keys to select a suggestion and Enter or Tab to fill it while preserving surrounding text; stale suggestions and input method confirmation keys no longer trigger unintended completion.
20. **\[Chat experience]** Fixed concurrent sends being rejected and then waiting for a new reply instead of recovering the original round. The rejected text and attachments are restored to the input, and the existing reply is refreshed and recovered, including interactive resumes and replies that have already finished.
21. **\[Chat experience]** Fixed members with read-only App access being unable to mark their own sessions as read; session access restrictions still apply.
22. **\[Chat experience]** Fixed admin feedback and closing custom feedback updating the user message when it shares a `dataId` with the AI reply in the same round. These operations now update only the corresponding AI reply.
23. **\[Chat experience]** Fixed images being missed when their Markdown descriptions contain paired square brackets, such as `![Image [1]](...)`; image extraction and replacement now handle these descriptions correctly.
24. **\[Chat experience]** Added missing Simplified Chinese, Traditional Chinese, English, and Korean translations for sandbox connection and file errors, Skill generation status, and Tool version labels, fixing untranslated text or internal keys appearing in some screens. Also added Korean text for account deletion, password setup, related audit events, Dataset tag errors, and rate-limit messages; Question classification Nodes now use Korean default options.
25. **\[Chat experience]** Fixed refreshing after sending a message being mistaken for a send failure when the browser closed the stream, causing submitted text and attachments to reappear in the draft.
26. **\[Models and billing]** Fixed the OpenAI SDK and Pi Agent using AI Proxy or host environment credentials when a custom model endpoint has no configured authentication, preventing default credentials from being sent to that address.
27. **\[Teams and permissions]** Fixed API Key expiration and quota errors including the plaintext key. Error messages no longer expose credentials; health checks now read keys from the `Authorization` header, reducing exposure in URLs and access logs.
28. **\[Teams and permissions]** Fixed members with only read or write access being able to move root-level resources, covering Apps, Tools, Datasets, Skills, their folders, and root-level Collections within a Dataset.
29. **\[Teams and permissions]** Fixed an incorrect permission check when creating HTTP Tools inside a tool folder. Collaborators with folder write permission can now create them, while members with only log read permission are rejected.
30. **\[Teams and permissions]** Fixed audit details treating some names, URLs, or changed values containing colons as translation keys. System labels are still translated, while user content is displayed unchanged.
31. **\[Teams and permissions]** Fixed the “Cancel” button in the account deletion confirmation dialog not following the current interface language.
32. **\[Teams and permissions]** Fixed share pages loading public App information through identifiers for other publishing channels. Only share channels are now accepted.
33. **\[System and operations]** Fixed CSV formula escaping that could be bypassed by fields starting with whitespace, a BOM, tabs, or line breaks, strengthening log, usage, and audit exports.
34. **\[System and operations]** Fixed io_uring crashes during heavy output or backpressure in JavaScript code workers on Linux with seccomp enabled. Workers fall back to epoll before startup while retaining the existing native isolation.
35. **\[System and operations]** Fixed MongoDB 4.x failures caused by empty update fields when clearing a team's AI account or external Workflow variables, moving Dataset Collections, and updating or clearing chat feedback in some cases.
36. **\[System and operations]** Fixed OpenSandbox treating the transitional `Pending` and `Allocated` states as errors. Starting or reusing a sandbox now waits until it is running before connecting.
37. **\[System and operations]** Fixed sandbox file metadata queries treating permission, connection, or command errors as missing files. Query failures now surface as errors, and metadata for empty files correctly retains size and permissions.
38. **\[System and operations]** Fixed OpenSandbox applying only the final replacement when a batch edits the same file multiple times. All edits to that file now take effect in order.
39. **\[System and operations]** Fixed the Commercial edition's admin configuration page showing values that differ from the running configuration when the database has no system configuration or an older record lacks `systemEnv`. Valid `config.json` settings now initialize the running configuration, and the page fills only missing sections while preserving existing site and plan settings, preventing defaults from overwriting them on save.
40. **\[System and operations]** Fixed inconsistent filter boundaries caused by the date picker mutating original dates, requests dropping milliseconds, and charts querying an extra day; page selections, lists, statistics, and exports now preserve the complete selected date range, and the model dashboard defaults to yesterday and today. Commercial edition App charts no longer lose time buckets across daylight saving transitions, and log and usage CSV exports use the request’s start offset to avoid server daylight saving effects.
41. **\[System and operations]** Fixed legacy source names or invalid values causing failed log reads and inconsistent filtering or source counts. Unrecognized channel types are classified as “Unknown” consistently across lists, CSV exports, and Commercial edition source charts.
42. **\[System and operations]** Fixed user, feedback, and session search filters overwriting one another in chat logs; the list and CSV export now apply all selected conditions together.
43. **\[System and operations]** Fixed Langfuse Node attributes leaking into parent traces and missing Agent generation records and usage. Concurrent Nodes now collect usage separately without including other Nodes' data.
44. **\[Tools and channels]** Fixed audio transcription input validation errors being wrapped as generic service errors. Invalid inputs now return the standard HTTP `400` parameter error.
45. **\[Tools and channels]** Fixed localized README files being omitted when uploading Tool packages to the Marketplace. Language-specific files are now uploaded, and their documentation URLs are retained for both Toolkits and child Tools.
46. **\[Tools and channels]** Fixed blank text in some Tool names, descriptions, and tags when their Traditional Chinese translation is empty. These values fall back to Simplified Chinese, then English if no usable text is available.

## 🛠️ Code improvements

1. **\[Workflows and Agents]** Unified execution for canvas child Apps, Agent child App Tools, and legacy App call Nodes, centralizing file and variable filtering, memory isolation, interaction resumes, streaming controls, and usage aggregation to reduce duplicate code. Plugin runtimes and call tokens now also receive the usage record ID so the backend can associate calls with the current run.
2. **\[System and operations]** Added a shared server-side translation loader for output localization, with resource loading and caching by request language and namespace. It falls back to English or the original key for missing text, supports retries after load failures, and keeps each request’s translation language fixed.
3. **\[System and operations]** Added skills for authoring and running integration test cases organized by business area. Workflow editor regression cases use scenario matrices with single-page browsing and filtering, plus manually triggered Node and headless Playwright scripts; AI runs uncovered cases or reviews script failures, while temporary HTML reports retain the executor, results, evidence, and retest history for each item.
4. **\[System and operations]** Added a manually triggered GitHub Actions workflow for backporting merged PRs; its target branch defaults to `v4.17.x` and can be changed, and it creates a draft PR when complete. The repository's `fastgpt-backport-pr` skill supports batch backports to an explicitly specified maintenance branch, resolving mechanical conflicts automatically and requiring confirmation for semantic conflicts. After targeted checks, it pushes code directly, aligns Pro with the matching maintenance branch for `v4.17.x`, and submits Chinese and English release notes for the corresponding unreleased version in a separate PR based on `upstream/main`.
5. **\[System and operations]** Updated the repository's `pr-review` skill to assess functional correctness, duplication, redundancy, maintainability, and current repository rules independently; quality issues and clear rule violations can also require changes, with evidence from searches, comparable implementations, and specific rules. Reviews cover implementation plans, business constraints, security, performance, PR comments, review summaries, and inline discussions; re-reviews use previous reports and the full current diff to check earlier findings and new risks.
6. **\[System and operations]** Added per-language draft control to the documentation site, with consistent filtering across routes, navigation, search, and exports. `llms.txt` now exports published pages in the requested language, while missing translations can still fall back to published content.
7. **\[System and operations]** Removed standalone evaluation workspace configuration and related Pro scripts, commands, and historical assets to simplify development tooling.
8. **\[System and operations]** The Marketplace’s default type-check scope now excludes the Workflow core and CLI packages, reducing interference from type checks across workspaces.
9. **\[System and operations]** Upgraded Mongoose to `8.24.8` and the Plugin SDK to `1.2.0-beta.2`, updating Plugin streaming response types.
