github.com/Tencent/WeKnora

New Features

• NEW: Scoped Tenant API Keys & Principal Model — the headline of this release. WeKnora now issues fine-grained, capability-scoped API keys that are first-class principals separate from human users (migrations 000064_principal_model, 000065_tenant_api_keys). Each key carries an explicit role plus capability grants (manage_kbs covering the full KB lifecycle, manage_storage_backends, member/space capabilities, etc.), can be restricted to specific knowledge bases, and updates its last_used_at under a throttle. Route-level guards (DenyAPIKeyPrincipal, api_key_gate) close IDOR/scope gaps, and a new API Integration Playground in the web UI lets owners mint, scope, and test keys interactively. MCP OAuth and embed sessions are now scoped per principal so external integrations stay isolated.
• NEW: Runtime Task Queue Observability & Worker-Pool Governance — a system-admin Runtime Queues dashboard exposes live queue depth, per-model concurrency stats, failed-task inspection, manual retry, and cursor-paginated task listings. The ingestion pipeline moves from a single aggregate worker pool to guaranteed per-stage pools (core / post-process / enrichment / maintenance) plus a shared elastic pool, with per-model background concurrency governors (model.max_concurrency) wrapping chat, embedding, rerank, and VLM calls. Wiki generation runs in its own independently governed pool. See Async task capacity planning.
• NEW: Multi-Instance Storage Backends — each workspace can register multiple object/file storage instances (local / minio / cos / tos / s3 / oss / ks3 / obs) and bind different knowledge bases to different instances, with a workspace-level default (migration 000068_storage_backends). Ships a full CRUD + connectivity-test API and settings UI, credential masking on read, and hardened image-source protection for storage backends. See Infrastructure API.
• NEW: Session-Scoped Temporary Attachments — attach images and documents to a chat session for one-off Q&A with asynchronous parsing (migration 000070_temporary_documents). Enforces a combined image + attachment limit, normalizes temporary attachment IDs, persists attachment content across turns, and adds a chat attachment preview drawer.
• NEW: Question & Follow-up Suggestions — knowledge-grounded suggested questions plus after-answer follow-ups (migration 000067_question_suggestions), with tag-scope aware generation, KB-scope preservation on click, and a dedicated follow-up rendering component.
• NEW: Stable Resource Registry & LLM-Context Alias Compaction — a request-local resource registry (migration 000069_resource_registry) assigns stable res:// aliases to retrieved sources so LLM context stays compact and citations remain consistent; includes orphan-alias detection/logging and Markdown-based image context normalization.
• NEW: @Skill / @MCP Mentions with Scoped Agent Runtime — mention skills and MCP services inline in chat to scope the agent runtime for a single turn, with hardened @mention scope resolution and consolidated per-turn scoping across knowledge tools.
• NEW: Mid-Conversation MCP OAuth — MCP services can prompt for and resume OAuth authorization mid-chat, driven by AuthType with auto-detection on test, an OAuth skip endpoint, and in-chat OAuth interaction cards aligned with the agent tool timeline.
• NEW: QQBot & Lark (Feishu International) IM Integration — new QQBot instant-messaging platform integration and support for Feishu's international edition (Lark), including region-aware routing and reply-in-thread via the Feishu reply-message API.
• NEW: weknora CLI v0.10 (BREAKING) — an agent-first CLI refresh: new model, message, config, and skills command groups; doc reparse / doc update; kb config / kb config set; session resume (renamed from continue) and session tool-approval; agent-first chat and session ask output modes; typed SDK errors/enums and KB model config; hardened SSE reliability; schema and exit-code contracts.
• NEW: Redis TLS Support — TLS connections to Redis with config surfaced at startup and hardened TLS tests (#1930).
• NEW: New Providers — Requesty added as an OpenAI-compatible model provider; Keenable added as a configurable web search provider.
• NEW: Tenantless Provisioning & Gated Self-Service Workspaces — OIDC/login provisioning can create users without a tenant, with gated self-service workspace creation and a workspace onboarding flow, unified under a shared default_tenant_mode policy.
• NEW: Admin Password Reset & System Settings Tabs — system admins can reset user passwords with session revocation and edit builtin model configs; system settings reorganized into tabbed sections. Password fields are sanitized in logs.
• NEW: Knowledge Base Duplicate Flow — clone a knowledge base (config + structure) via a dedicated API and UI; custom instruction fields added to KB configuration with length validation.
• NEW: Per-Agent Citation Output Toggle — agents can enable/disable citation output; retrieval references are still emitted to the references drawer even when in-answer citations are disabled.
• NEW: Chat References Drawer & UX Polish — a dedicated references drawer with web/KB source distinction, inline session-title rename in the sidebar, chat reference links opening in new tabs, and native streaming loading placeholders.

Improvements

• IMPROVED: Security hardening (broad triage) — closed numerous open GHSA findings and security-triage gaps: SSRF protection across web_fetch, datasource connectors, knowledge URL import (including redirect chains), MinerU, embedders, and model clients; secret redaction in login, tenant KV, integration-list, and initialization/parser-check responses; SQL-validator bypass closed in agent database tools; refresh-token validation hardened (blocked as bearer); MCP upload sandbox enforced across transports; wiki/IDOR KB-access enforcement; nginx iframe X-Frame-Options.
• IMPROVED: Wiki ingestion — dedicated worker pool with concurrent claim-based batching; recovers stranded claims, prevents finalize double-run and cross-batch document races, and strips internal chunk aliases from page content.
• IMPROVED: Knowledge processing — safe task recovery after server restart, chunking strategy propagated to parent-child splitter configs, richer error handling in knowledge spans, rune-aware span-name fitting with expanded name column, and processing-timeline status aligned with the knowledge row.
• IMPROVED: Terminology — user-facing "tenant" labels renamed to "workspace" across the UI and i18n.
• IMPROVED: Chat streaming — unified streaming wait indicators across chat and embed, follow-up suggestion loading and answer-toolbar timing polished, references drawer closed on session switch, and enforced retrieved-image output in answers.
• IMPROVED: Frontend resilience — hardened settings localStorage load to prevent white-screen from corrupted state; shared document action menu and card-view components extracted; Wiki badge on KB cards; responsive doc filter bar.
• IMPROVED: Infrastructure config — infrastructure host/port configurable via env vars in docker-compose; remote infrastructure supported via .env.local; WEKNORA_MODEL_MAX_CONCURRENCY defaulted to 32.
• IMPROVED: docreader — SSRF utility and safe HTTP client added; legacy doc-payload detection; parser routing tests.

Bug Fixes

• FIXED: Chat now emits retrieval references even when citation output is disabled.
• FIXED: /api/v1/tenants/kv/storage-engine-config returned no Huawei OBS config, causing the KB-creation dialog to omit the OBS unavailable marker.
• FIXED: Empty tenant default_provider fell back to local even when local was not in STORAGE_ALLOW_LIST, breaking KB creation.
• FIXED: DingTalk file/image messages now ingest into the knowledge base with SSRF-checked downloads and pictureDownloadCode fallback (#1771).
• FIXED: Feishu WebSocket long connection now actually closes when the channel is stopped; replies land in the original thread.
• FIXED: Tag filters correctly applied to knowledge search; tag-scope resolution hardened (#1907).
• FIXED: Scoped API keys can poll FAQ import progress; KB-scoped file proxy restricted to exports/ paths while still serving shared KB images.
• FIXED: Web-search controls gated by provider readiness; agent selector capability status chips polished.
• FIXED: Manual knowledge editor publish flow; doc list reloads after tag rename in the manage drawer.
• FIXED: Division-by-zero guard in FAQ import timing logs; VLM temperature made configurable.

Infrastructure & Build

• BUILD: Migrations 000064_principal_model, 000065_tenant_api_keys, 000066_expand_knowledge_span_name, 000067_question_suggestions, 000068_storage_backends, 000069_resource_registry, 000070_temporary_documents.
• BUILD: Per-pool task-queue governance (core / post-process / enrichment / maintenance / shared / wiki); model concurrency limiter/governor package.
• BUILD: Go client extended for scoped API keys, storage backends, message suggestions, KB duplicate, and streaming error types; resolved open Dependabot alerts across Go, npm, and pip.
• BUILD: Swagger / API docs regenerated; docs/api/storage-backend.md added.

Documentation

• DOC: docs/worker-pool-governance.md added; docs/api/storage-backend.md added.
• DOC: docs/QA.md, wiki docs, and per-feature guides extended for scoped API keys, storage backends, runtime queues, temporary attachments, and MCP OAuth.
• DOC: Architecture diagram updated for scoped API keys, multi-instance storage, and worker-pool governance.