diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0bcb8037..26049574 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,76 @@
All notable changes to this project will be documented in this file.
+## [0.8.0] - 2026-09-03
+
+### New Features
+
+- **NEW**: **Skill Sandbox Runtime (Docker / E2B / Cube)** — the headline of this release. Agent skills now run in a **session-persistent sandbox**: one sandbox per chat session, with `shell_exec`, file read/write/edit, attachment staging, and generated-file collection all landing in the same workspace. Three backends share one `RemoteSandboxClient` protocol: **Docker** (single-host / self-hosted, talking to the Engine API rather than `docker run --rm`), **E2B** (cloud or any E2B-compatible control plane, including self-hosted), and **CubeSandbox**. Per-workspace configs cover image, CPU/memory, TTL, DNS, templates and snapshots (migrations `000082_tenant_sandbox_config`, `000083_session_sandbox_config`). Admins can set a **network policy** per config (default-deny egress, allow/deny lists, Cube L7 rules, E2B host rules) (#2995). The old **Local host-process backend is removed**. The Docker backend is **opt-in** (`WEKNORA_SANDBOX_DOCKER_ENABLED` or System Settings → Network Security) because a mounted `docker.sock` is host root (#2936). See [`docs/sandbox-docker-backend.md`](./docs/sandbox-docker-backend.md) and [`docs/sandbox-protocol.md`](./docs/sandbox-protocol.md).
+- **NEW**: **Tenant Skill Catalog** — skills are first-class workspace resources, not files dropped next to the process (migrations `000086_tenant_skills`, `000087_skill_install_transcript`, `000088_skill_snapshot_planned_name`, `000090_skill_catalog`). Install from **ClawHub**, **SkillHub / skills.sh**, **GitHub/GitLab URLs**, or a zip upload; each install is a snapshot on the chosen sandbox config, with live circular progress, install transcripts, stop/reinstall, and uninstall that keeps catalog archives. Browse and edit installed skill files from Settings; agents get `write_skill_file` / `edit_skill_file` plus `shell_exec` that no longer requires a skill to be installed. Personal and workspace **skill env vars** inject at execution without ever being read back (migration `000089_env_vars`). `@mention` no longer silently shrinks the skill whitelist.
+- **NEW**: **Cross-Session Long-Term Memory** — a new memory product (migration `000084_memory`), independent of the Neo4j conversation-memory that 0.7.1 removed. Workspaces opt in; each user can further turn it off. Memories are typed (`profile` / `preference` / `fact` / `task` / `interest`), written either explicitly or by a background extractor, and **inferred items stay pending until the user confirms**. Resident profile/preference blocks ride in every turn; situational facts are recalled lexically and (optionally) semantically; `search_memory` looks up on demand. Document affinity conditions retrieval toward sources the user keeps citing. Settings, items, topics, document affinity, confirm/reject, export and forced consolidation are exposed as `/api/v1/memory/*` (full-access API keys only; the subject is always the caller).
+- **NEW**: **In-Process anydoc Parser** — Office documents (doc/docx/ppt/pptx and the types anydoc converts) can be parsed **inside the Go app** via `third_party/anydoc-go`, without a round-trip to docreader. When the binding is linked, anydoc is preferred for every type it converts; PPT/PPTX still default to MarkItDown when no engine rule is set. The app image links anydoc by default.
+- **NEW**: **DeepSeek Harness Plugin** — official npm package [`@wxg-prc-cpg/dsh-weknora`](https://www.npmjs.com/package/@wxg-prc-cpg/dsh-weknora) exposes four read-only tools (`weknora_search`, `weknora_read_document`, `weknora_ask`, `weknora_list_knowledge_bases`) so a DeepSeek Harness coding agent can retrieve from a WeKnora deployment (#2759).
+- **NEW**: **GitLab & Tencent IMA Data Sources** — GitLab projects sync as a knowledge source (#2656); Tencent IMA notes sync through the note OpenAPI, with retries on transient failures.
+- **NEW**: **LiteLLM Model Provider** — LiteLLM is a first-class chat/embedding/rerank provider, so one gateway URL can front many upstream models (#2923).
+- **NEW**: **Exa & Metaso Web Search** — two new web-search providers join the existing set (#2617).
+- **NEW**: **XMind Outline Parsing** — `.xmind` files are parsed into an outline and indexed like other documents (#2713).
+- **NEW**: **Chat Artifacts, Question Outline & Timestamps** — sandbox-generated files collect into a per-message artifacts drawer (migration `000081_message_artifacts`) with download and a pulse on the folder when new files arrive; long overflowing sessions get a Codex-style question outline / minimap (#2839); messages show timestamps as conversation-flow dividers.
+- **NEW**: **Document Auto-Tagging** — after parse, a chat model may pick matching tags from the knowledge base's existing tag set and attach them incrementally, without creating tags or overwriting manual ones (migration `000080_knowledge_base_auto_tag_config`).
+- **NEW**: **Context Compaction & Prompt-Cache Markers** — long sandbox turns compact tool history instead of truncating or rewriting files; stable prompt prefixes plus provider cache markers improve cache hit rate on Anthropic/OpenAI-compatible backends. Per-turn LLM token usage is attributed and stored (migration `000085_message_usage`).
+- **NEW**: **OIDC JWKS Verification & `/auth/oidc/start`** — ID tokens are signature-verified via JWKS (#2799); a direct 302 start endpoint supports login without a pre-rendered SPA handshake.
+- **NEW**: **Optional Complex Passwords** — registration, self-service change-password and admin reset can require mixed-case, digits and special characters (`WEKNORA_AUTH_COMPLEX_PASSWORD_ENABLED` / system setting) (#2929).
+- **NEW**: **System-Admin User Creation** — platform admins can create users from the console without an invite link (#2722).
+- **NEW**: **Invitation Auto-Accept & Invite-Only Join** — inviting an already-registered user can auto-accept; in invite-only mode, share links send the visitor through login before joining the workspace.
+
+### Improvements
+
+- **IMPROVED**: **Sandbox security** — all Docker execs (scripts, `shell_exec`, file ops, artifact bootstrap) run as the sandbox user (uid 1000), not root; symlink-following `chown`/file ops that could escape the workspace are closed; zombie processes from cancelled execs are reaped; in-use snapshot deletes retry as conflicts; idle sandboxes detach and sweep; skill image pointers stay out of config PUT.
+- **IMPROVED**: **Skill install UX** — catalog cards, avatar shortcut, live circular progress that keeps the drawer open, batch install verification, stop-in-flight, ClawHub `@owner/slug` and skills.sh catalog pages, GitHub subtree counting, and sandbox uninstall moved into the manage-drawer header.
+- **IMPROVED**: **Chat / Agent UX** — `shell_exec` renders as a command + stream card; sandbox file lists and skill names show on tool cards; generated-files drawer restyled to match settings lists; Mermaid diagrams stay rendered while streaming; message timestamps align with persisted `created_at`.
+- **IMPROVED**: **Knowledge search** — MMR selection is incremental; hybrid-search honors `resource_urls=public`; MatchCount=0 no longer truncates every result; parent-child embeddings survive chunking; deleted images in a chunk are no longer retrieved.
+- **IMPROVED**: **Docreader / anydoc** — DOCX vertical merges and table text preserved; PDF embedded-image SMask applied (no more all-black figures); MHTML header objects stripped correctly; URL scrapes fail closed instead of indexing error pages; PPT/PPTX attachments get a parser engine.
+- **IMPROVED**: **IM** — Feishu websocket long-connection reverse-proxy support (#2550); Feishu streaming cards finalize and show progress in full-output mode; WeCom 32-byte PKCS7 padding; DingTalk rich-text follow-ups; Yunzhijia thread sessions; IM file attachments in QA.
+- **IMPROVED**: **Auth / tenancy** — `SwitchTenant` records last-active-tenant preference; self-service change password in the profile; folder drag-and-drop upload preserves directory structure; document download action; editable document summaries; paginated document chunks; FAQ batch actions restored.
+- **IMPROVED**: **Observability** — sandbox operations emit product-level Langfuse spans; follow-up completions nest under the parent chat trace; prompt-cache markers stay stable across turns.
+- **IMPROVED**: **Deployment capabilities** — menus and settings hide modules the binary did not register (`GET /system/capabilities`), so Lite / trimmed builds stop advertising missing features (#2674).
+- **IMPROVED**: **CI** — golangci-lint workflow gates new PR code; path-filtered checks avoid triggering anydoc/unrelated lints; git pre-commit / pre-push hooks; Playwright WebKit for docreader URL tests; dsh-weknora e2e cache.
+
+### Bug Fixes
+
+- **FIXED**: OIDC ID-token signature is verified through JWKS, including explicit-endpoint configs; JWKS response bodies are closed (#2799).
+- **FIXED**: High-risk system settings and agent skill pickers read the latest stored value instead of a stale form snapshot.
+- **FIXED**: Sandbox tools are scoped to the workspace they actually serve; agents are told which workspace they are in; file tools can read staged attachments without hanging empty skill tools; skill-script stdout is kept on failure.
+- **FIXED**: Tool-call JSON with a stray trailing fragment is recovered; stringified JSON arrays in skill args parse; `@mention` no longer narrows the skill whitelist.
+- **FIXED**: Feishu streaming cards no longer stick in an in-progress state; full-output mode shows the progress card (#2911, #2909).
+- **FIXED**: Model deletion blocked by usage now shows the usage details (#2975).
+- **FIXED**: `nginx-api-proxy.conf` is included in the UI image build, restoring API proxying in the frontend container.
+- **FIXED**: Postgres `docker-compose` image no longer fails the default security check (#2704); S3-compatible endpoints disable optional checksums; MinerU results resolve by upload filename stem and preserve the extension.
+- **FIXED**: Wiki housekeeping no longer force-fails documents that still have a durable Wiki backlog; concurrent duplicate wiki page identities are prevented; wiki generate-with-template calls set `MaxTokens`.
+- **FIXED**: Shared knowledge bases no longer report `doc_count=0` in agent `runtime_context`; data-source deletion sync is scoped per connector and actually persisted.
+- **FIXED**: Chunk merge no longer drops content when contiguous chunks repeat text; context headers are sanitized before persistence; markdown table regex catastrophic backtracking is stopped (#2770).
+- **FIXED**: Built-in agent names/descriptions follow the UI locale (#2828); default locale can be set at runtime via `DEFAULT_LOCALE`.
+- **FIXED**: Direct invites are reconciled after a share-link join; stale home tenant is cleared after member removal.
+
+### Breaking Changes
+
+- **BREAKING**: The **Local host-process sandbox backend is removed**. Existing Local configs must be recreated against Docker (opt-in), E2B, or Cube.
+- **BREAKING**: The Docker sandbox backend is **off by default**. Enable it in System Settings → Network Security, or set `WEKNORA_SANDBOX_DOCKER_ENABLED=true`.
+- **BREAKING**: Complex-password mode, when enabled, rejects registration / password-change payloads that do not mix upper, lower, digit and special characters.
+
+### Infrastructure & Build
+
+- **BUILD**: Migrations `000080`–`000090` (auto-tag, message artifacts, tenant/session sandbox config, memory, message usage, tenant skills, install transcript, snapshot planned name, env vars, skill catalog); matching SQLite migrations `000003`–`000012`.
+- **BUILD**: New `internal/sandbox` remote-client stack (Docker Engine API, E2B, Cube), `internal/agent/compaction`, `internal/application/service/memory`, `internal/ipclass` (shared SSRF + sandbox URL classification), `third_party/anydoc-go`.
+- **BUILD**: Go client: sandbox skill install/stop/files, personal env-var APIs, long-term memory APIs.
+- **BUILD**: `wechatopenai/weknora-sandbox` image; Docker backend requires `appuser` in the `docker.sock` group inside the app container.
+- **BUILD**: golangci-lint, anydoc, and dsh-plugin GitHub Actions workflows.
+
+### Documentation
+
+- **DOC**: Sandbox protocol / Docker backend / cluster guides under `docs/sandbox-*.md`; skill env-var surface documented in `docs/api/skill.md`; long-term memory in `docs/api/memory.md`.
+- **DOC**: `docs/QA.md` extended for Docker opt-in, Local-backend removal, skill catalog, long-term memory, anydoc, and OIDC JWKS.
+- **DOC**: Architecture diagram updated for sandbox backends, skill catalog, long-term memory, anydoc, GitLab/IMA, LiteLLM, and the DeepSeek Harness plugin.
+
## [0.7.2] - 2026-08-07
### New Features
diff --git a/README.md b/README.md
index c732ddb3..04816334 100644
--- a/README.md
+++ b/README.md
@@ -31,7 +31,7 @@
-
+
@@ -53,13 +53,14 @@
[**WeKnora**](https://weknora.weixin.qq.com) is an open-source, LLM-powered knowledge framework built for enterprise-grade document understanding, semantic retrieval, and autonomous reasoning.
-It is organized around three core capabilities: **RAG-based Quick Q&A** for everyday lookups, a **ReAct Agent** that autonomously orchestrates retrieval, MCP tools and web search to handle complex multi-step tasks, and a brand-new **Wiki Mode** in which agents distill raw documents into a self-maintaining, interlinked markdown knowledge base with an interactive knowledge graph, complete with manual editing, revision history and one-click rollback. Knowledge curation is equally hands-on: a **tree-structured folder view** preserves the directory layout of uploads, and **chunk editing with revision history** lets retrieval chunks be edited, diffed and reverted like documents. Combined with multi-source ingestion (Feishu wiki / Feishu Drive / Notion / Yuque / RSS, and growing), **website embed widgets** for publishing agents to external sites, **scoped API keys with a principal model** for programmatic integrations, **multi-instance storage backends** per workspace for flexible data placement, 20+ LLM provider integrations, full Langfuse observability plus a **runtime task-queue dashboard with worker-pool governance**, **enterprise-ready multi-workspace RBAC** (4-tier role matrix + per-resource ownership + per-workspace audit log), and a fully self-hostable modular architecture, WeKnora turns scattered documents into a queryable, reasoning-capable, continuously evolving knowledge asset.
+It is organized around three core capabilities: **RAG-based Quick Q&A** for everyday lookups, a **ReAct Agent** that autonomously orchestrates retrieval, MCP tools, a **tenant skill catalog**, session-persistent **Docker / E2B / Cube sandboxes** and web search to handle complex multi-step tasks, and a brand-new **Wiki Mode** in which agents distill raw documents into a self-maintaining, interlinked markdown knowledge base with an interactive knowledge graph, complete with manual editing, revision history and one-click rollback. **Cross-session long-term memory** remembers who you are and what you keep asking about. Knowledge curation is equally hands-on: a **tree-structured folder view** preserves the directory layout of uploads, and **chunk editing with revision history** lets retrieval chunks be edited, diffed and reverted like documents. Combined with multi-source ingestion (Feishu wiki / Feishu Drive / GitLab / Tencent IMA / Notion / Yuque / RSS, and growing), **website embed widgets** for publishing agents to external sites, **scoped API keys with a principal model** for programmatic integrations, **multi-instance storage backends** per workspace for flexible data placement, 20+ LLM provider integrations (including LiteLLM), full Langfuse observability plus a **runtime task-queue dashboard with worker-pool governance**, **enterprise-ready multi-workspace RBAC** (4-tier role matrix + per-resource ownership + per-workspace audit log), and a fully self-hostable modular architecture, WeKnora turns scattered documents into a queryable, reasoning-capable, continuously evolving knowledge asset.
-The framework supports auto-syncing knowledge from Feishu, Notion, and Yuque (more data sources coming soon), handles 10+ document formats including PDF, Word, images, and Excel, and can serve Q&A directly through IM channels like WeCom, Feishu, Slack, and Telegram. It is compatible with major LLM providers including OpenAI, DeepSeek, Qwen (Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, and Ollama. Its fully modular design allows swapping LLMs, vector databases, and storage backends, with support for local and private cloud deployment ensuring complete data sovereignty. WeKnora also integrates with **Langfuse** for comprehensive observability into agent reasoning, token usage, and pipeline tracing.
+The framework supports auto-syncing knowledge from Feishu, GitLab, Tencent IMA, Notion, and Yuque (more data sources coming soon), handles 10+ document formats including PDF, Word, images, Excel and XMind, and can serve Q&A directly through IM channels like WeCom, Feishu, Slack, and Telegram. It is compatible with major LLM providers including OpenAI, DeepSeek, Qwen (Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, LiteLLM, and Ollama. Office files can be parsed in-process with **anydoc**. Its fully modular design allows swapping LLMs, vector databases, and storage backends, with support for local and private cloud deployment ensuring complete data sovereignty. WeKnora also integrates with **Langfuse** for comprehensive observability into agent reasoning, token usage, and pipeline tracing.
## ✨ Latest Updates
+- **v0.8.0** — **Skill sandbox runtime** (session-persistent Docker / E2B / Cube backends with per-tenant network policy; Local host-process backend removed; Docker opt-in); **tenant skill catalog** (install from ClawHub / SkillHub / git / zip, per-sandbox snapshots, live progress, file browse/edit, personal and workspace env vars); **cross-session long-term memory** (profile / preference / fact / task / interest, auto-extract with confirm, `search_memory`); **in-process anydoc office parser**; official **DeepSeek Harness plugin** `@wxg-prc-cpg/dsh-weknora`; GitLab and Tencent IMA data sources; LiteLLM; Exa and Metaso web search; XMind parsing; chat artifacts, question outline and timestamps; context compaction and provider prompt-cache markers. Plus OIDC JWKS verification, optional complex passwords, document auto-tagging, and broad sandbox/security hardening. See [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.2** — Launched the **official product documentation site** (VitePress; six sections, ~50 pages covering ~360 API endpoints and ~150 environment variables, with standalone Docker/Nginx deployment, quickstart sample data and a local MCP demo); **knowledge base folder tree** (upload paths stored as first-class data, browse/rename/re-file documents like a file manager); **chunk editing with revision history** (edit retrieval chunks in the UI, per-version diff and rollback, automatic reindexing, plus custom document metadata); **Wiki page revision history** (snapshots + line-level diff + one-click rollback + in-browser manual editing); **directly loadable file URLs** via `resource_urls=public` / `RESOURCE_URL_MODE` (third-party apps render images and files without a second authenticated proxy call); **Feishu Drive data source** and docx sync through the blocks API; batch document tagging; **MCP Server 1.1.x** (migrated to the mcp 2.x high-level API, official PyPI package `tencent-weknora-mcp`, new `create_knowledge_from_text` and `list_shared_knowledge_bases` for 29 tools total); AWS S3 default credential chain (IAM Role / IRSA); local HTML upload parsing; QQBot markdown replies; new PR CI checks for app / frontend / docreader / mcp-server. Plus large-scale router and `modelcontext` refactors, rerank and chunking quality work, and broad stability fixes. See [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.1** — New **Yunzhijia (云之家) IM integration** (WebSocket + image messages + markdown replies); **Volcengine rerank** provider (with request batching) and **Zhipu AI web search** provider; **platform-scoped API keys** for control-plane automation (tenant management, system settings, runtime queues, audit logs); **per-KB activity audit trail**; FAQ management enhancements (filtering, tagging, export, import tracking); **Langfuse OTLP/OTel tracing** migration with W3C traceparent propagation; chat header actions with one-click **Markdown export** and wiki tool results in the references drawer; prompt-cache observability; session channel governance (admin-scoped IM/embed/API sessions); resilient Feishu large-wiki sync; and removal of the legacy Neo4j conversation-memory dependency. Plus broad slug-integrity, SSRF-transport, and state-sync hardening. See [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.0** — Fine-grained **scoped API keys & principal model** (capability-level grants + per-KB restriction + API integration playground); **runtime task-queue observability dashboard & worker-pool governance** (per-stage pools + per-model concurrency governors + failed-task inspection/retry); **multi-instance storage backends** (multiple storage instances per workspace, per-KB binding, default instance); **session-scoped temporary attachments** (async image/doc parsing + combined limits); question & follow-up suggestions; stable resource registry with LLM-context alias compaction; `@Skill / @MCP` mentions with scoped agent runtime; mid-conversation MCP OAuth; QQBot & Lark (Feishu International) IM integration; Redis TLS; Requesty model provider + Keenable web search; tenantless provisioning & gated self-service workspaces; admin password reset; knowledge base duplicate flow; `weknora` CLI v0.10. Plus broad security hardening (SSRF, secret redaction, SQL validation, IDOR). See [`CHANGELOG.md`](./CHANGELOG.md).
@@ -115,9 +116,11 @@ Fully modular pipeline from document parsing, vectorization, and retrieval to LL
| Capability | Details |
|------------|---------|
-| Intelligent Reasoning | ReACT progressive multi-step reasoning, autonomously orchestrating knowledge retrieval, MCP tools, and web search |
+| Intelligent Reasoning | ReACT progressive multi-step reasoning, autonomously orchestrating knowledge retrieval, MCP tools, skill sandboxes, and web search |
| Quick Q&A | RAG-based Q&A over knowledge bases for fast and accurate answers |
| Wiki Mode | Agent-driven auto-generation of structured, interlinked markdown Wiki pages from raw documents; in-browser manual editing, page revision history, line-level diff and one-click rollback |
+| Skill Catalog & Sandbox | Workspace skill catalog (ClawHub / SkillHub / git / zip) installed onto session-persistent Docker / E2B / Cube sandboxes; `shell_exec`, file tools, artifacts, per-config network policy; Local host-process backend removed |
+| Long-term Memory | Cross-session memory (profile / preference / fact / task / interest) with auto-extract, user confirm, and on-demand `search_memory` |
| Tool Calling | Built-in tools, MCP tools (incl. OAuth2 remote services, mid-conversation OAuth), web search; `@Skill / @MCP` mentions to scope the agent runtime per turn |
| Conversation Strategy | Online Prompt editing, retrieval threshold tuning, multi-turn context awareness, per-agent citation output toggle |
| Suggested Questions | Auto-generated question suggestions and after-answer follow-ups based on knowledge base content |
@@ -134,8 +137,9 @@ Fully modular pipeline from document parsing, vectorization, and retrieval to LL
| Chunk Editing & Revisions | Edit retrieval chunks directly in the UI with per-version snapshots, diff and one-click rollback, and automatic reindexing after an edit; generated questions can be added, edited, deleted and regenerated; custom document metadata supported |
| Per-Upload Process Config | Override parser, chunking, multimodal (VLM / ASR), graph extraction, and question generation per upload batch via upload-confirm dialog or `process_config` API; reparse with new settings |
| Batch Reparse | Re-queue parsing for multiple documents at once with optional per-batch `process_config` |
-| Data Source Import | Auto-sync from Feishu wiki / Feishu Drive / Lark / Notion / Yuque / RSS feeds (more data sources coming soon); incremental and full sync |
-| Document Formats | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / Images / CSV / Excel / PPT / JSON |
+| Data Source Import | Auto-sync from Feishu wiki / Feishu Drive / Lark / GitLab / Tencent IMA / Notion / Yuque / RSS feeds (more data sources coming soon); incremental and full sync |
+| Document Formats | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / Images / CSV / Excel / PPT / JSON / XMind |
+| Auto-Tagging | After parse, pick matching tags from the knowledge base's existing set without creating tags or overwriting manual ones |
| Retrieval Strategies | BM25 sparse / Dense retrieval / GraphRAG / parent-child chunking / HNSW-accelerated pgvector (1024-dim) / multi-dimensional indexing |
| Batch Selection & Tagging | Marquee drag-select multiple documents in the KB list for batch reparse and batch tagging (common tags pre-selected) |
| E2E Testing | Full-pipeline visualization with recall hit rate, BLEU / ROUGE metric evaluation |
@@ -144,13 +148,13 @@ Fully modular pipeline from document parsing, vectorization, and retrieval to LL
| Capability | Details |
|------------|---------|
-| LLMs | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / Ollama |
+| LLMs | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / LiteLLM / Ollama |
| Embeddings | Ollama / BGE / GTE / Zhipu / OpenAI-compatible APIs |
| Vector DBs | PostgreSQL (pgvector) / Elasticsearch / OpenSearch / Milvus / Weaviate / Qdrant / Apache Doris / Tencent VectorDB |
| Object Storage | Local / MinIO / AWS S3 (IAM Role / IRSA default credential chain) / Volcengine TOS / Alibaba Cloud OSS / Kingsoft Cloud KS3 / Huawei Cloud OBS; **multiple storage instances per workspace** with per-KB binding and a default instance |
| IM Channels | WeCom / Feishu / Lark (Feishu International) / QQBot / Slack / Telegram / DingTalk / Mattermost / WeChat / Yunzhijia |
| Website Embed | Publish agents via embed widget with domain allowlists, rate limits, and secure-mode token exchange |
-| Web Search | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI |
+| Web Search | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI / Exa / Metaso |
| API Integration | Scoped API keys (capability-level grants + per-KB restriction + throttled last-used tracking) with an API integration playground; MCP OAuth and embed sessions isolated per principal; `resource_urls=public` returns directly loadable file/image URLs, removing the second authenticated proxy call |
| MCP Server | Official PyPI package `tencent-weknora-mcp` with 29 tools over stdio / SSE / HTTP transports |
@@ -161,7 +165,7 @@ Fully modular pipeline from document parsing, vectorization, and retrieval to LL
| Deployment | Local / Docker / Kubernetes (Helm) with private and offline support |
| UI | Web UI / RESTful API / CLI (`weknora`) / Chrome Extension / Website Embed Widget / WeChat Mini Program |
| Access Control | Workspace RBAC with 4-tier role matrix (Owner / Admin / Contributor / Viewer), per-KB resource ownership, per-workspace audit log, invite-only workspaces, tenantless provisioning & gated self-service workspace creation, admin password reset (session revocation), cross-workspace superuser, scoped API keys |
-| Security | AES-256-GCM at-rest encryption for API keys and MCP / data-source credentials with graceful key rotation; gRPC TLS + Token between app and docreader; Redis TLS; SSRF-safe HTTP client (data sources, URL import, redirect chains); secret redaction in responses; sandbox isolation for agent skills |
+| Security | AES-256-GCM at-rest encryption for API keys and MCP / data-source credentials with graceful key rotation; gRPC TLS + Token between app and docreader; Redis TLS; SSRF-safe HTTP client (data sources, URL import, redirect chains); secret redaction in responses; skill sandbox isolation (Docker opt-in / E2B / Cube) with per-config network policy; OIDC ID-token JWKS verification; optional complex-password policy |
| Observability | Integrated Langfuse (sole tracing backend) for ReAct loops, token tracking, tool calls, and pipeline tracing; built-in Langfuse-style document parsing trace timeline with stage-by-stage progress; system-admin runtime task-queue dashboard (queue depth, per-model concurrency, failed-task inspection & manual retry) |
| Task Management | MQ async tasks with per-stage worker-pool governance (core / post-process / enrichment / maintenance + elastic shared pool, plus an independent Wiki pool) and per-model background concurrency governors; automatic database migration on version upgrade |
| Model Management | Centralized config, declarative built-in models via YAML, per-knowledge-base model selection, per-model thinking-mode and embedding-dimension overrides, interactive model test debugger, multi-workspace built-in model sharing, WeKnora Cloud hosted models and parsing |
diff --git a/README_CN.md b/README_CN.md
index e95ef838..8e6d9c42 100644
--- a/README_CN.md
+++ b/README_CN.md
@@ -31,7 +31,7 @@
-
+
@@ -53,12 +53,13 @@
**[WeKnora(维娜拉)](https://weknora.weixin.qq.com)** 是一款开源的、基于大语言模型(LLM)的知识管理框架,专为企业级文档理解、语义检索与智能推理场景打造。
-框架围绕三大核心能力构建:**RAG 快速问答**适合日常知识查询,**ReAct Agent 智能推理**自主编排知识检索、MCP 工具与网络搜索完成复杂多步任务,全新的 **Wiki 模式**则让 Agent 从原始文档中自治生成相互链接的 Markdown 知识库与可视化知识图谱,并支持人工编辑、版本历史与一键回滚。知识加工环节也完全可控:**树形文件夹**保留上传目录结构,**分块编辑与版本历史**让检索片段可以像文档一样被修改、比对与回滚。结合多源数据接入(飞书知识库 / 飞书云盘 / Notion / 语雀 / RSS,更多持续接入中)、**网站嵌入 Widget** 将智能体发布到外部站点、**权限范围 API Key 与 Principal 模型**面向程序化集成、**每空间多实例存储后端**灵活编排数据落地、二十余家主流模型厂商集成、Langfuse 全链路可观测性与**运行时任务队列面板 + Worker 池治理**、**企业级多空间 RBAC(四级角色矩阵 + 资源归属 + 空间审计日志)**,以及完全可私有化部署的模块化架构,WeKnora 帮助团队把分散文档沉淀为可查询、可推理、可持续演进的专属知识资产。
+框架围绕三大核心能力构建:**RAG 快速问答**适合日常知识查询,**ReAct Agent 智能推理**自主编排知识检索、MCP 工具、**技能目录**、会话级 **Docker / E2B / Cube 沙箱**与网络搜索完成复杂多步任务,全新的 **Wiki 模式**则让 Agent 从原始文档中自治生成相互链接的 Markdown 知识库与可视化知识图谱,并支持人工编辑、版本历史与一键回滚。**跨会话长期记忆**会记住你是谁、你常问什么。知识加工环节也完全可控:**树形文件夹**保留上传目录结构,**分块编辑与版本历史**让检索片段可以像文档一样被修改、比对与回滚。结合多源数据接入(飞书知识库 / 飞书云盘 / GitLab / 腾讯 IMA / Notion / 语雀 / RSS,更多持续接入中)、**网站嵌入 Widget** 将智能体发布到外部站点、**权限范围 API Key 与 Principal 模型**面向程序化集成、**每空间多实例存储后端**灵活编排数据落地、二十余家主流模型厂商集成(含 LiteLLM)、Langfuse 全链路可观测性与**运行时任务队列面板 + Worker 池治理**、**企业级多空间 RBAC(四级角色矩阵 + 资源归属 + 空间审计日志)**,以及完全可私有化部署的模块化架构,WeKnora 帮助团队把分散文档沉淀为可查询、可推理、可持续演进的专属知识资产。
-框架支持从飞书、Notion 及语雀等外部平台自动同步知识(更多数据源持续接入中),覆盖 PDF、Word、图片、Excel 等十余种文档格式,并可通过企业微信、飞书、Slack、Telegram 等 IM 频道直接提供问答服务。模型层面兼容 OpenAI、DeepSeek、Qwen(阿里云)、智谱、混元、Gemini、MiniMax、NVIDIA、Ollama 等主流厂商。全流程模块化设计,大模型、向量数据库、存储等组件均可灵活替换,支持本地与私有云部署,数据完全自主可控。WeKnora 还无缝集成了 **Langfuse**,为 Agent 运行、Token 使用及任务流水线提供了全面的可观测性追踪。
+框架支持从飞书、GitLab、腾讯 IMA、Notion 及语雀等外部平台自动同步知识(更多数据源持续接入中),覆盖 PDF、Word、图片、Excel、XMind 等十余种文档格式,并可通过企业微信、飞书、Slack、Telegram 等 IM 频道直接提供问答服务。模型层面兼容 OpenAI、DeepSeek、Qwen(阿里云)、智谱、混元、Gemini、MiniMax、NVIDIA、LiteLLM、Ollama 等主流厂商。Office 文档可由 **anydoc** 在 Go 进程内解析。全流程模块化设计,大模型、向量数据库、存储等组件均可灵活替换,支持本地与私有云部署,数据完全自主可控。WeKnora 还无缝集成了 **Langfuse**,为 Agent 运行、Token 使用及任务流水线提供了全面的可观测性追踪。
## ✨ 最新更新
+- **v0.8.0** —— **技能沙箱运行时**(会话级常驻 Docker / E2B / Cube 后端,按空间配置网络策略;移除 Local 宿主机进程后端;Docker 需显式开启);**空间技能目录**(从 ClawHub / SkillHub / git / zip 安装,按沙箱快照、实时进度、文件浏览/编辑、个人与空间环境变量);**跨会话长期记忆**(profile / preference / fact / task / interest,自动抽取需确认,`search_memory`);进程内 **anydoc** Office 解析;官方 **DeepSeek Harness 插件** `@wxg-prc-cpg/dsh-weknora`;GitLab 与腾讯 IMA 数据源;LiteLLM;Exa 与 Metaso 网络搜索;XMind 解析;对话产物、问题大纲与时间戳;上下文压缩与供应商 Prompt Cache 标记。另有 OIDC JWKS 验签、可选复杂密码、文档自动打标签,以及大范围沙箱/安全加固。详见 [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.2** —— 上线**官方产品文档站**(VitePress,六大板块约 50 篇,覆盖约 360 个 API 端点与约 150 个环境变量,含独立 Docker/Nginx 部署、快速上手样例数据与本地 MCP demo);**知识库文件夹树**(文件夹路径独立入库,可像文件管理器一样浏览、重命名与重新归档);**分块编辑与版本历史**(可视化编辑检索分块、逐版本 diff 与回滚、自动重建索引、文档自定义元数据);**Wiki 页面版本历史**(快照 + 行级 diff + 一键回滚 + 浏览器内手动编辑);**API 文件直链模式** `resource_urls=public` / `RESOURCE_URL_MODE`(第三方 App 无需二次调用鉴权代理即可加载图片与文件);**飞书云盘数据源**与 docx blocks 逐类型下钻同步;文档批量打标签;**MCP Server 1.1.x**(迁移到 mcp 2.x 高级 API,官方 PyPI 包 `tencent-weknora-mcp`,新增 `create_knowledge_from_text` 与 `list_shared_knowledge_bases`,共 29 个工具);AWS S3 默认凭据链(IAM Role / IRSA);本地 HTML 上传解析;QQBot Markdown 回复;新增 app / frontend / docreader / mcp-server 的 PR CI 检查。另有 router 与 modelcontext 大规模重构、重排与分块质量优化,以及大量稳定性修复。详见 [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.1** —— 新增**云之家 IM 集成**(WebSocket + 图片消息 + Markdown 回复);**火山引擎 Rerank** 供应商(自动分批请求)与**智谱 AI 网络搜索**供应商;**平台级 API Key**,用于控制面自动化(空间管理、系统设置、运行时队列、审计日志);**按知识库的活动审计追踪**;FAQ 管理增强(筛选、打标签、导出、导入结果追踪);**Langfuse OTLP/OTel 追踪**迁移,支持 W3C traceparent 跨服务传播;对话头部操作栏,支持一键 **Markdown 导出**,并在引用抽屉中展示 Wiki 工具结果;Prompt 缓存可观测性;会话渠道治理(IM/嵌入/API 会话按管理员范围隔离);飞书大型 Wiki 同步韧性增强;移除旧版 Neo4j 会话记忆依赖。另有大范围的 slug 完整性、SSRF 传输与状态同步加固。详见 [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.0** —— 细粒度**权限范围 API Key 与 Principal 模型**(能力级授权 + 按 KB 限制 + API 集成调试台);**运行时任务队列可观测面板与 Worker 池治理**(分阶段独立池 + 按模型并发治理 + 失败任务排查/重试);**多实例存储后端**(每空间多存储实例、按 KB 绑定、默认实例);**会话级临时附件**(图片/文档异步解析 + 合并限额);推荐问题与追问;稳定资源注册表与 LLM 上下文别名压缩;`@Skill / @MCP` 提及范围化 Agent 运行时;会话内 MCP OAuth 授权;QQBot 与 Lark(飞书国际版)IM 集成;Redis TLS;Requesty 模型厂商 + Keenable 网络搜索;无空间预置与受控自助创建工作区;管理员密码重置;知识库复制流程;`weknora` CLI v0.10。同时完成大范围安全加固(SSRF、密钥脱敏、SQL 校验、越权)。详见 [`CHANGELOG.md`](./CHANGELOG.md)。
@@ -114,9 +115,11 @@
| 能力 | 详情 |
|------|------|
-| 智能推理 | ReACT 渐进式多步推理,自主编排知识检索、MCP 工具与网络搜索 |
+| 智能推理 | ReACT 渐进式多步推理,自主编排知识检索、MCP 工具、技能沙箱与网络搜索 |
| 快速问答 | 基于知识库的 RAG 问答,快速准确地回答问题 |
| Wiki 模式 | Agent 驱动从原始文档中自动生成并维护结构化、相互链接的 Markdown Wiki 知识页面;支持浏览器内人工编辑、页面版本历史、行级 diff 与一键回滚 |
+| 技能目录与沙箱 | 空间技能目录(ClawHub / SkillHub / git / zip)安装到会话级 Docker / E2B / Cube 沙箱;`shell_exec`、文件工具、产物收集、按配置网络策略;已移除 Local 宿主机进程后端 |
+| 长期记忆 | 跨会话记忆(profile / preference / fact / task / interest),支持自动抽取、用户确认与按需 `search_memory` |
| 工具调用 | 内置工具、MCP 工具(含 OAuth2 远程服务、会话内 OAuth 授权)、网络搜索;支持 `@Skill / @MCP` 提及以按轮次范围化 Agent 运行时 |
| 对话策略 | 在线 Prompt 编辑、检索阈值调节、多轮上下文感知、按 Agent 引用输出开关 |
| 推荐问题 | 基于知识库内容自动生成推荐问题与答后追问 |
@@ -133,8 +136,9 @@
| 分块编辑与版本历史 | 在界面直接编辑检索分块,保留逐版本快照,支持 diff 与一键回滚,编辑后自动重建索引;生成问题可增删改与重新生成;支持文档自定义元数据 |
| 按批次解析配置 | 上传确认对话框或 `process_config` API 覆盖解析引擎、分块、多模态(VLM / ASR)、图谱抽取与问题生成;支持 reparse 时调整配置 |
| 批量重新解析 | 一次为多篇文档重新排队解析,可携带批次级 `process_config` |
-| 数据源导入 | 飞书知识库 / 飞书云盘 / Lark / Notion / 语雀 / RSS 订阅自动同步(更多数据源开发中),支持增量与全量同步 |
-| 文档格式 | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 图片 / CSV / Excel / PPT / JSON |
+| 数据源导入 | 飞书知识库 / 飞书云盘 / Lark / GitLab / 腾讯 IMA / Notion / 语雀 / RSS 订阅自动同步(更多数据源开发中),支持增量与全量同步 |
+| 文档格式 | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 图片 / CSV / Excel / PPT / JSON / XMind |
+| 自动打标签 | 解析完成后从知识库已有标签中选择匹配项增量关联,不创建新标签、不覆盖人工标签 |
| 检索策略 | BM25 稀疏召回 / Dense 稠密召回 / GraphRAG 图谱增强 / 父子分块 / pgvector HNSW 加速(1024 维)/ 多维度索引 |
| 批量选择与打标签 | 知识库文档列表支持框选(marquee)多选,可批量重新解析、批量打标签(自动预选公共标签) |
| 端到端测试 | 检索+生成全链路可视化,评估召回命中率、BLEU / ROUGE 等指标 |
@@ -143,13 +147,13 @@
| 能力 | 详情 |
|------|------|
-| 模型厂商 | OpenAI / Azure OpenAI / Anthropic(Claude)/ DeepSeek / Qwen(阿里云)/ 智谱 / 混元 / 豆包(火山引擎)/ Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / Ollama |
+| 模型厂商 | OpenAI / Azure OpenAI / Anthropic(Claude)/ DeepSeek / Qwen(阿里云)/ 智谱 / 混元 / 豆包(火山引擎)/ Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / LiteLLM / Ollama |
| 向量数据库 | PostgreSQL (pgvector) / Elasticsearch / OpenSearch / Milvus / Weaviate / Qdrant / Apache Doris / 腾讯云 VectorDB |
| Embedding | Ollama / BGE / GTE / 智谱 / OpenAI 兼容接口 |
| 对象存储 | 本地 / 腾讯云COS / 火山引擎 TOS / MinIO / AWS S3(支持 IAM Role / IRSA 默认凭据链)/ 阿里云 OSS / 金山云 KS3 / 华为云 OBS;支持**每空间多实例存储后端**,不同知识库可绑定不同实例并设置默认实例 |
| IM 集成 | 企业微信 / 飞书 / Lark(飞书国际版)/ QQBot / Slack / Telegram / 钉钉 / Mattermost / 微信 / 云之家 |
| 网站嵌入 | 通过嵌入 Widget 发布智能体,支持域名白名单、限流与安全模式 Token 交换 |
-| 网络搜索 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / 智谱 AI |
+| 网络搜索 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / 智谱 AI / Exa / Metaso |
| API 集成 | 权限范围 API Key(能力级授权 + 按 KB 限制 + 节流的 last_used 追踪)与 API 集成调试台;MCP OAuth 与嵌入会话按 Principal 隔离;`resource_urls=public` 直接返回可加载的文件/图片直链,免去二次鉴权代理调用 |
| MCP Server | 官方 PyPI 包 `tencent-weknora-mcp`,29 个工具,支持 stdio / SSE / HTTP 三种传输 |
@@ -161,7 +165,7 @@
| 部署 | 本地 / Docker / Kubernetes (Helm),支持私有化离线部署 |
| 界面 | Web UI / RESTful API / 命令行(`weknora`)/ Chrome Extension / 网站嵌入 Widget / 微信小程序 |
| 权限控制 | 空间 RBAC 四级角色矩阵(Owner / Admin / Contributor / Viewer),按知识库的资源归属,每空间审计日志,invite-only 准入,无空间预置与受控自助创建工作区,管理员密码重置(会话吊销),跨空间超级管理员,权限范围 API Key |
-| 安全 | API Key 与 MCP / 数据源凭据 AES-256-GCM 静态加密、支持平滑密钥轮换;app ↔ docreader gRPC TLS + Token;Redis TLS;防 SSRF HTTP 客户端(覆盖数据源、URL 导入、重定向链等);密钥响应脱敏;Agent 技能沙箱隔离 |
+| 安全 | API Key 与 MCP / 数据源凭据 AES-256-GCM 静态加密、支持平滑密钥轮换;app ↔ docreader gRPC TLS + Token;Redis TLS;防 SSRF HTTP 客户端(覆盖数据源、URL 导入、重定向链等);密钥响应脱敏;技能沙箱隔离(Docker 需开启 / E2B / Cube)与按配置网络策略;OIDC ID Token JWKS 验签;可选复杂密码策略 |
| 可观测性 | 集成 Langfuse(唯一追踪后端)以追踪 ReAct 循环、Token 消耗、工具调用和任务流水线;内置 Langfuse 风格的文档解析追踪时间线,逐阶段展示解析进度;系统管理员运行时任务队列面板(队列深度、按模型并发、失败任务排查与手动重试) |
| 任务管理 | MQ 异步任务,分阶段独立 Worker 池治理(core / 后处理 / enrichment / maintenance + 弹性共享池,Wiki 独立池)与按模型后台并发治理;版本升级自动数据库迁移 |
| 模型管理 | 集中配置,YAML 声明式内置模型配置,知识库级别模型选择,按模型思考模式与 Embedding 维度覆盖,交互式模型调试器,多空间共享内置模型,WeKnora Cloud 托管模型与文档解析 |
diff --git a/README_JA.md b/README_JA.md
index ccfa3b60..08bf38b8 100644
--- a/README_JA.md
+++ b/README_JA.md
@@ -31,7 +31,7 @@
-
+
@@ -53,12 +53,13 @@
[**WeKnora(ウィーノラ)**](https://weknora.weixin.qq.com) は、大規模言語モデル(LLM)をベースとしたオープンソースのナレッジフレームワークで、エンタープライズ級の文書理解、セマンティック検索、自律推論シナリオ向けに設計されています。
-本フレームワークは **3 つのコア能力** を中心に構築されています:日常的な検索に最適な **RAG ベースのクイック Q&A**、ナレッジ検索・MCP ツール・Web 検索を自律的にオーケストレーションし複雑なマルチステップタスクを処理する **ReAct Agent 推論**、そして Agent が生のドキュメントから相互リンクされた Markdown ナレッジベースとインタラクティブなナレッジグラフを自律生成・維持する全く新しい **Wiki モード**(手動編集・バージョン履歴・ワンクリックロールバック対応)。ナレッジの整備も細かく制御可能で、**ツリー型フォルダー**がアップロード時のディレクトリ構造を保持し、**チャンク編集とバージョン履歴**により検索チャンクをドキュメントと同様に編集・差分比較・ロールバックできます。さらに、多様なデータソース連携(Feishu ナレッジベース / Feishu クラウドドライブ / Notion / Yuque / RSS、随時拡充中)、**ウェブサイト埋め込み Widget** による外部サイトへのエージェント公開、プログラム連携向けの**スコープ付き API キーと Principal モデル**、ワークスペースごとの**マルチインスタンスストレージバックエンド**、20 以上の LLM プロバイダー統合、Langfuse による全体可観測性と**ランタイムタスクキューダッシュボード + Worker プール統治**、**エンタープライズ向けマルチテナント RBAC(4 階層ロールマトリクス + リソース所有権 + テナント監査ログ)**、完全セルフホスト可能なモジュラーアーキテクチャと組み合わせることで、WeKnora は散在する文書を「検索可能・推論可能・継続的に進化する」専用ナレッジ資産へと昇華させます。
+本フレームワークは **3 つのコア能力** を中心に構築されています:日常的な検索に最適な **RAG ベースのクイック Q&A**、ナレッジ検索・MCP ツール・**スキルカタログ**・セッション永続の **Docker / E2B / Cube サンドボックス**・Web 検索を自律的にオーケストレーションし複雑なマルチステップタスクを処理する **ReAct Agent 推論**、そして Agent が生のドキュメントから相互リンクされた Markdown ナレッジベースとインタラクティブなナレッジグラフを自律生成・維持する全く新しい **Wiki モード**(手動編集・バージョン履歴・ワンクリックロールバック対応)。**クロスセッション長期メモリ**は「あなたが誰で、何を繰り返し尋ねるか」を覚えます。ナレッジの整備も細かく制御可能で、**ツリー型フォルダー**がアップロード時のディレクトリ構造を保持し、**チャンク編集とバージョン履歴**により検索チャンクをドキュメントと同様に編集・差分比較・ロールバックできます。さらに、多様なデータソース連携(Feishu ナレッジベース / Feishu クラウドドライブ / GitLab / Tencent IMA / Notion / Yuque / RSS、随時拡充中)、**ウェブサイト埋め込み Widget** による外部サイトへのエージェント公開、プログラム連携向けの**スコープ付き API キーと Principal モデル**、ワークスペースごとの**マルチインスタンスストレージバックエンド**、20 以上の LLM プロバイダー統合(LiteLLM 含む)、Langfuse による全体可観測性と**ランタイムタスクキューダッシュボード + Worker プール統治**、**エンタープライズ向けマルチテナント RBAC(4 階層ロールマトリクス + リソース所有権 + テナント監査ログ)**、完全セルフホスト可能なモジュラーアーキテクチャと組み合わせることで、WeKnora は散在する文書を「検索可能・推論可能・継続的に進化する」専用ナレッジ資産へと昇華させます。
-Feishu、Notion、Yuqueなどの外部プラットフォームからのナレッジ自動同期(他のデータソースも順次対応中)に対応し、PDF、Word、画像、Excelなど10以上の文書フォーマットをサポート。WeChat Work、Feishu、Slack、TelegramなどのIMチャネルから直接Q&Aサービスを提供できます。モデル層ではOpenAI、DeepSeek、Qwen(Alibaba Cloud)、Zhipu、Hunyuan、Gemini、MiniMax、NVIDIA、Ollamaなど主要プロバイダーに対応。全プロセスをモジュラー設計し、大規模モデル、ベクトルデータベース、ストレージなどのコンポーネントを柔軟に差し替え可能。ローカルおよびプライベートクラウドデプロイに対応し、データは完全に自己管理可能です。さらにWeKnoraは **Langfuse** とシームレスに統合され、Agentの推論、トークン消費、パイプラインに対する包括的な可観測性(オブザーバビリティ)を提供します。
+Feishu、GitLab、Tencent IMA、Notion、Yuqueなどの外部プラットフォームからのナレッジ自動同期(他のデータソースも順次対応中)に対応し、PDF、Word、画像、Excel、XMindなど10以上の文書フォーマットをサポート。WeChat Work、Feishu、Slack、TelegramなどのIMチャネルから直接Q&Aサービスを提供できます。モデル層ではOpenAI、DeepSeek、Qwen(Alibaba Cloud)、Zhipu、Hunyuan、Gemini、MiniMax、NVIDIA、LiteLLM、Ollamaなど主要プロバイダーに対応。Office 文書は **anydoc** でプロセス内解析できます。全プロセスをモジュラー設計し、大規模モデル、ベクトルデータベース、ストレージなどのコンポーネントを柔軟に差し替え可能。ローカルおよびプライベートクラウドデプロイに対応し、データは完全に自己管理可能です。さらにWeKnoraは **Langfuse** とシームレスに統合され、Agentの推論、トークン消費、パイプラインに対する包括的な可観測性(オブザーバビリティ)を提供します。
## ✨ 最新アップデート
+- **v0.8.0** — **スキルサンドボックス実行環境**(セッション永続の Docker / E2B / Cube バックエンド、テナント単位のネットワークポリシー;Local ホストプロセスバックエンドを削除;Docker はオプトイン);**テナントスキルカタログ**(ClawHub / SkillHub / git / zip からインストール、サンドボックス単位のスナップショット、ライブ進捗、ファイル閲覧/編集、個人・ワークスペース環境変数);**クロスセッション長期メモリ**(profile / preference / fact / task / interest、確認付き自動抽出、`search_memory`);プロセス内 **anydoc** Office パーサー;公式 **DeepSeek Harness プラグイン** `@wxg-prc-cpg/dsh-weknora`;GitLab と Tencent IMA データソース;LiteLLM;Exa と Metaso Web 検索;XMind 解析;チャット成果物、質問アウトライン、タイムスタンプ;コンテキスト圧縮とプロバイダー Prompt Cache マーカー。加えて OIDC JWKS 検証、任意の複雑パスワード、ドキュメント自動タグ付け、広範なサンドボックス/セキュリティ強化。詳細は [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.2** — **公式製品ドキュメントサイト**を公開(VitePress、6 セクション約 50 ページで約 360 の API エンドポイントと約 150 の環境変数を網羅、独立した Docker/Nginx デプロイ・クイックスタート用サンプルデータ・ローカル MCP デモを同梱);**ナレッジベースのフォルダーツリー**(アップロードパスを独立したデータとして保存し、ファイルマネージャーのように参照・リネーム・再配置);**チャンク編集とバージョン履歴**(UI で検索チャンクを編集、バージョン単位の差分とロールバック、編集後のインデックス自動再構築、ドキュメントのカスタムメタデータ);**Wiki ページのバージョン履歴**(スナップショット + 行単位差分 + ワンクリックロールバック + ブラウザ内手動編集);**ファイル直リンクモード** `resource_urls=public` / `RESOURCE_URL_MODE`(サードパーティアプリが認証プロキシへの二次リクエストなしで画像とファイルを表示);**Feishu クラウドドライブデータソース**と docx の blocks API 同期;ドキュメントの一括タグ付け;**MCP Server 1.1.x**(mcp 2.x の高レベル API へ移行、公式 PyPI パッケージ `tencent-weknora-mcp`、`create_knowledge_from_text` と `list_shared_knowledge_bases` を追加し計 29 ツール);AWS S3 デフォルト資格情報チェーン(IAM Role / IRSA);ローカル HTML アップロード解析;QQBot の Markdown 返信;app / frontend / docreader / mcp-server の PR CI チェック追加。加えて router と `modelcontext` の大規模リファクタリング、リランクとチャンキングの品質改善、広範な安定性修正。詳細は [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.1** — 新しい**Yunzhijia(云之家)IM 連携**(WebSocket + 画像メッセージ + Markdown 返信);**Volcengine Rerank** プロバイダー(リクエストの自動分割)と **Zhipu AI Web 検索**プロバイダー;コントロールプレーン自動化向けの**プラットフォームスコープ API キー**(テナント管理、システム設定、ランタイムキュー、監査ログ);**KB 単位のアクティビティ監査証跡**;FAQ 管理の強化(フィルタリング、タグ付け、エクスポート、インポート結果追跡);**Langfuse OTLP/OTel トレーシング**への移行と W3C traceparent 伝播;チャットヘッダーアクションによるワンクリック **Markdown エクスポート**、参照ドロワーへの Wiki ツール結果表示;プロンプトキャッシュ可観測性;セッションチャネルガバナンス(IM/埋め込み/API セッションを管理者スコープで分離);Feishu 大規模 Wiki 同期の堅牢化;レガシー Neo4j 会話メモリ依存の削除。加えて広範な slug 整合性・SSRF トランスポート・状態同期の強化。詳細は [`CHANGELOG.md`](./CHANGELOG.md)。
- **v0.7.0** — きめ細かい**スコープ付き API キーと Principal モデル**(能力単位の付与 + KB 単位の制限 + API 連携プレイグラウンド);**ランタイムタスクキュー可観測ダッシュボードと Worker プール統治**(ステージ別プール + モデル別並行度ガバナー + 失敗タスクの調査/再試行);**マルチインスタンスストレージバックエンド**(ワークスペースごとに複数のストレージインスタンス、KB 単位のバインド、デフォルトインスタンス);**セッションスコープの一時添付**(画像/ドキュメントの非同期解析 + 合算上限);推奨質問とフォローアップ;安定リソースレジストリと LLM コンテキストのエイリアス圧縮;`@Skill / @MCP` メンションによるスコープ化 Agent ランタイム;会話中の MCP OAuth;QQBot と Lark(Feishu 国際版)IM 連携;Redis TLS;Requesty モデルプロバイダー + Keenable Web 検索;テナントレスプロビジョニングと制御されたセルフサービスワークスペース;管理者パスワードリセット;ナレッジベース複製フロー;`weknora` CLI v0.10。加えて大規模なセキュリティ強化(SSRF、シークレットのマスキング、SQL 検証、IDOR)。詳細は [`CHANGELOG.md`](./CHANGELOG.md)。
@@ -115,9 +116,11 @@ Feishu、Notion、Yuqueなどの外部プラットフォームからのナレッ
| 機能 | 詳細 |
|------|------|
-| インテリジェント推論 | ReACT プログレッシブ・マルチステップ推論、ナレッジ検索・MCP ツール・Web 検索を自律的にオーケストレーション |
+| インテリジェント推論 | ReACT プログレッシブ・マルチステップ推論、ナレッジ検索・MCP ツール・スキルサンドボックス・Web 検索を自律的にオーケストレーション |
| クイック Q&A | ナレッジベースベースの RAG Q&A、迅速かつ正確な回答 |
| Wiki モード | Agent主導で生のドキュメントから構造化された相互リンク済みMarkdown Wikiページを自動生成・保守;ブラウザ内手動編集、ページのバージョン履歴、行単位差分とワンクリックロールバック |
+| スキルカタログとサンドボックス | ワークスペースのスキルカタログ(ClawHub / SkillHub / git / zip)をセッション永続の Docker / E2B / Cube サンドボックスへインストール;`shell_exec`、ファイルツール、成果物、設定単位のネットワークポリシー;Local ホストプロセスバックエンドは削除 |
+| 長期メモリ | クロスセッションメモリ(profile / preference / fact / task / interest)。自動抽出、ユーザー確認、オンデマンド `search_memory` |
| ツール呼び出し | 組み込みツール、MCP ツール(OAuth2 リモートサービス・会話中 OAuth 含む)、Web 検索;`@Skill / @MCP` メンションでターン単位に Agent ランタイムを範囲化 |
| 対話戦略 | オンライン Prompt 編集、検索閾値チューニング、マルチターン文脈認識、Agent 単位の引用出力トグル |
| 推奨質問 | ナレッジベースの内容に基づく質問の自動生成と回答後のフォローアップ |
@@ -134,8 +137,9 @@ Feishu、Notion、Yuqueなどの外部プラットフォームからのナレッ
| チャンク編集とバージョン | UI から検索チャンクを直接編集、バージョン単位のスナップショット・差分・ワンクリックロールバック、編集後のインデックス自動再構築;生成質問の追加・編集・削除・再生成;ドキュメントのカスタムメタデータ対応 |
| アップロード単位の解析設定 | アップロード確認ダイアログまたは `process_config` API でパーサー・チャンキング・マルチモーダル(VLM / ASR)・グラフ抽出・質問生成をバッチ単位で上書き;reparse 時も設定変更可能 |
| 一括 reparse | 複数ドキュメントの解析を一度に再キュー、バッチ単位の `process_config` 対応 |
-| データソースインポート | Feishu ナレッジベース / Feishu クラウドドライブ / Lark / Notion / Yuque / RSS フィードの自動同期(他のデータソースも開発中)、増分・全量同期対応 |
-| 文書フォーマット | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 画像 / CSV / Excel / PPT / JSON |
+| データソースインポート | Feishu ナレッジベース / Feishu クラウドドライブ / Lark / GitLab / Tencent IMA / Notion / Yuque / RSS フィードの自動同期(他のデータソースも開発中)、増分・全量同期対応 |
+| 文書フォーマット | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 画像 / CSV / Excel / PPT / JSON / XMind |
+| 自動タグ付け | 解析後、ナレッジベース既存タグから一致するものを増分付与(新規タグ作成や手動タグの上書きはしない) |
| 検索戦略 | BM25 疎検索 / Dense 密検索 / GraphRAG グラフ強化 / 親子チャンキング / pgvector HNSW 加速(1024 次元)/ 多次元インデックス |
| 一括選択とタグ付け | KB リストでマーキー(ドラッグ)複数選択し、一括 reparse と一括タグ付け(共通タグを自動プリセット)を実行 |
| E2E テスト | 検索+生成の全パイプライン可視化、リコール的中率・BLEU / ROUGE 指標評価 |
@@ -144,13 +148,13 @@ Feishu、Notion、Yuqueなどの外部プラットフォームからのナレッ
| 機能 | 詳細 |
|------|------|
-| 大規模モデル | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / Ollama |
+| 大規模モデル | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / LiteLLM / Ollama |
| Embedding | Ollama / BGE / GTE / OpenAI 互換 API |
| ベクトル DB | PostgreSQL (pgvector) / Elasticsearch / OpenSearch / Milvus / Weaviate / Qdrant / Apache Doris / Tencent VectorDB |
| オブジェクトストレージ | ローカル / MinIO / AWS S3(IAM Role / IRSA のデフォルト資格情報チェーン対応)/ 火山引擎 TOS / Alibaba Cloud OSS / 金山雲 KS3 / 華為雲 OBS;**ワークスペースごとに複数のストレージインスタンス**、KB 単位のバインドとデフォルトインスタンス |
| IM 統合 | WeChat Work / Feishu / Lark(Feishu 国際版)/ QQBot / Slack / Telegram / DingTalk / Mattermost / WeChat / Yunzhijia |
| ウェブ埋め込み | 埋め込み Widget でエージェントを公開、ドメイン許可リスト・レート制限・セキュアモード Token 交換 |
-| Web 検索 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI |
+| Web 検索 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI / Exa / Metaso |
| API 連携 | スコープ付き API キー(能力単位の付与 + KB 単位の制限 + 節流付き last_used 追跡)と API 連携プレイグラウンド;MCP OAuth と埋め込みセッションを Principal 単位で分離;`resource_urls=public` で直接読み込み可能なファイル / 画像 URL を返却し、認証プロキシへの二次リクエストを不要に |
| MCP Server | 公式 PyPI パッケージ `tencent-weknora-mcp`、29 ツール、stdio / SSE / HTTP の 3 トランスポート対応 |
diff --git a/README_KO.md b/README_KO.md
index 0aea1458..7164a0c3 100644
--- a/README_KO.md
+++ b/README_KO.md
@@ -31,7 +31,7 @@
-
+
@@ -53,12 +53,13 @@
[**WeKnora**](https://weknora.weixin.qq.com)는 엔터프라이즈급 문서 이해, 시맨틱 검색, 자율 추론 시나리오를 위해 설계된 오픈소스 LLM 기반 지식 프레임워크입니다.
-본 프레임워크는 **세 가지 핵심 역량**을 중심으로 구성됩니다. 일상 검색에 최적화된 **RAG 기반 빠른 Q&A**, 지식 검색·MCP 도구·웹 검색을 자율적으로 오케스트레이션하여 복잡한 다단계 작업을 처리하는 **ReAct Agent 추론**, 그리고 Agent가 원본 문서에서 상호 연결된 마크다운 지식베이스와 인터랙티브 지식 그래프를 스스로 생성·유지하는 완전히 새로운 **Wiki 모드**(수동 편집·버전 이력·원클릭 롤백 지원)입니다. 지식 가공 과정도 세밀하게 제어할 수 있습니다. **트리형 폴더**가 업로드 시의 디렉터리 구조를 그대로 유지하고, **청크 편집 및 버전 이력**을 통해 검색 청크를 문서처럼 수정·비교·롤백할 수 있습니다. 다양한 데이터 소스 연동(Feishu 지식베이스 / Feishu 클라우드 드라이브 / Notion / Yuque / RSS, 지속 확장 중), **웹사이트 임베드 Widget**으로 외부 사이트에 에이전트 게시, 프로그램 연동을 위한 **범위 지정 API 키 및 Principal 모델**, 워크스페이스별 **다중 인스턴스 스토리지 백엔드**, 20개 이상의 LLM 프로바이더 통합, Langfuse 기반 풀스택 관측 가능성과 **런타임 작업 큐 대시보드 + Worker 풀 거버넌스**, **엔터프라이즈 멀티 테넌트 RBAC(4단계 역할 매트릭스 + 리소스 소유권 + 테넌트 감사 로그)**, 완전 셀프호스팅이 가능한 모듈형 아키텍처를 결합하여, WeKnora는 흩어진 문서를 검색·추론 가능하며 지속적으로 진화하는 전용 지식 자산으로 탈바꿈시킵니다.
+본 프레임워크는 **세 가지 핵심 역량**을 중심으로 구성됩니다. 일상 검색에 최적화된 **RAG 기반 빠른 Q&A**, 지식 검색·MCP 도구·**스킬 카탈로그**·세션 지속 **Docker / E2B / Cube 샌드박스**·웹 검색을 자율적으로 오케스트레이션하여 복잡한 다단계 작업을 처리하는 **ReAct Agent 추론**, 그리고 Agent가 원본 문서에서 상호 연결된 마크다운 지식베이스와 인터랙티브 지식 그래프를 스스로 생성·유지하는 완전히 새로운 **Wiki 모드**(수동 편집·버전 이력·원클릭 롤백 지원)입니다. **크로스 세션 장기 메모리**는 당신이 누구이며 무엇을 반복해서 묻는지 기억합니다. 지식 가공 과정도 세밀하게 제어할 수 있습니다. **트리형 폴더**가 업로드 시의 디렉터리 구조를 그대로 유지하고, **청크 편집 및 버전 이력**을 통해 검색 청크를 문서처럼 수정·비교·롤백할 수 있습니다. 다양한 데이터 소스 연동(Feishu 지식베이스 / Feishu 클라우드 드라이브 / GitLab / Tencent IMA / Notion / Yuque / RSS, 지속 확장 중), **웹사이트 임베드 Widget**으로 외부 사이트에 에이전트 게시, 프로그램 연동을 위한 **범위 지정 API 키 및 Principal 모델**, 워크스페이스별 **다중 인스턴스 스토리지 백엔드**, 20개 이상의 LLM 프로바이더 통합(LiteLLM 포함), Langfuse 기반 풀스택 관측 가능성과 **런타임 작업 큐 대시보드 + Worker 풀 거버넌스**, **엔터프라이즈 멀티 테넌트 RBAC(4단계 역할 매트릭스 + 리소스 소유권 + 테넌트 감사 로그)**, 완전 셀프호스팅이 가능한 모듈형 아키텍처를 결합하여, WeKnora는 흩어진 문서를 검색·추론 가능하며 지속적으로 진화하는 전용 지식 자산으로 탈바꿈시킵니다.
-Feishu, Notion, Yuque 등 외부 플랫폼에서 지식 자동 동기화를 지원하며(추가 데이터 소스 개발 중), PDF, Word, 이미지, Excel 등 10가지 이상의 문서 포맷을 처리합니다. WeChat Work, Feishu, Slack, Telegram 등의 IM 채널을 통해 Q&A 서비스를 직접 제공할 수 있습니다. 모델 레이어에서 OpenAI, DeepSeek, Qwen(Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, Ollama 등 주요 프로바이더를 지원합니다. 전체 프로세스가 모듈화 설계되어 LLM, 벡터 DB, 스토리지 등 구성 요소를 유연하게 교체 가능하며, 로컬 및 프라이빗 클라우드 배포를 지원하여 데이터 완전 자체 관리가 가능합니다. 또한 WeKnora는 **Langfuse**와 원활하게 통합되어 Agent 추론, 토큰 사용량 및 파이프라인에 대한 포괄적인 관측 가능성(Observability)을 제공합니다.
+Feishu, GitLab, Tencent IMA, Notion, Yuque 등 외부 플랫폼에서 지식 자동 동기화를 지원하며(추가 데이터 소스 개발 중), PDF, Word, 이미지, Excel, XMind 등 10가지 이상의 문서 포맷을 처리합니다. WeChat Work, Feishu, Slack, Telegram 등의 IM 채널을 통해 Q&A 서비스를 직접 제공할 수 있습니다. 모델 레이어에서 OpenAI, DeepSeek, Qwen(Alibaba Cloud), Zhipu, Hunyuan, Gemini, MiniMax, NVIDIA, LiteLLM, Ollama 등 주요 프로바이더를 지원합니다. Office 문서는 **anydoc**으로 프로세스 내 파싱할 수 있습니다. 전체 프로세스가 모듈화 설계되어 LLM, 벡터 DB, 스토리지 등 구성 요소를 유연하게 교체 가능하며, 로컬 및 프라이빗 클라우드 배포를 지원하여 데이터 완전 자체 관리가 가능합니다. 또한 WeKnora는 **Langfuse**와 원활하게 통합되어 Agent 추론, 토큰 사용량 및 파이프라인에 대한 포괄적인 관측 가능성(Observability)을 제공합니다.
## ✨ 최신 업데이트
+- **v0.8.0** — **스킬 샌드박스 런타임**(세션 지속 Docker / E2B / Cube 백엔드, 테넌트 단위 네트워크 정책; Local 호스트 프로세스 백엔드 제거; Docker는 옵트인); **테넌트 스킬 카탈로그**(ClawHub / SkillHub / git / zip에서 설치, 샌드박스별 스냅샷, 실시간 진행률, 파일 탐색/편집, 개인·워크스페이스 환경 변수); **크로스 세션 장기 메모리**(profile / preference / fact / task / interest, 확인이 필요한 자동 추출, `search_memory`); 프로세스 내 **anydoc** Office 파서; 공식 **DeepSeek Harness 플러그인** `@wxg-prc-cpg/dsh-weknora`; GitLab 및 Tencent IMA 데이터 소스; LiteLLM; Exa 및 Metaso 웹 검색; XMind 파싱; 채팅 산출물, 질문 개요, 타임스탬프; 컨텍스트 압축 및 프로바이더 Prompt Cache 마커. 또한 OIDC JWKS 검증, 선택적 복잡한 비밀번호, 문서 자동 태깅, 광범위한 샌드박스/보안 강화. 자세한 내용은 [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.2** — **공식 제품 문서 사이트** 공개(VitePress, 6개 섹션 약 50편으로 약 360개 API 엔드포인트와 약 150개 환경 변수를 다루며, 독립 Docker/Nginx 배포·빠른 시작 샘플 데이터·로컬 MCP 데모 포함); **지식베이스 폴더 트리**(업로드 경로를 독립 데이터로 저장하여 파일 관리자처럼 탐색·이름 변경·문서 재배치); **청크 편집 및 버전 이력**(UI에서 검색 청크 직접 편집, 버전별 diff와 롤백, 편집 후 인덱스 자동 재구축, 문서 커스텀 메타데이터); **Wiki 페이지 버전 이력**(스냅샷 + 라인 단위 diff + 원클릭 롤백 + 브라우저 내 수동 편집); **파일 직접 링크 모드** `resource_urls=public` / `RESOURCE_URL_MODE`(서드파티 앱이 인증 프록시를 다시 호출하지 않고 이미지와 파일을 렌더링); **Feishu 클라우드 드라이브 데이터 소스** 및 docx blocks API 동기화; 문서 일괄 태깅; **MCP Server 1.1.x**(mcp 2.x 고수준 API로 마이그레이션, 공식 PyPI 패키지 `tencent-weknora-mcp`, `create_knowledge_from_text`와 `list_shared_knowledge_bases` 추가로 총 29개 도구); AWS S3 기본 자격 증명 체인(IAM Role / IRSA); 로컬 HTML 업로드 파싱; QQBot Markdown 응답; app / frontend / docreader / mcp-server PR CI 검사 추가. 또한 router와 `modelcontext` 대규모 리팩터링, 리랭크·청킹 품질 개선 및 광범위한 안정성 수정. 자세한 내용은 [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.1** — 새로운 **Yunzhijia(云之家) IM 통합**(WebSocket + 이미지 메시지 + Markdown 응답); **Volcengine Rerank** 제공자(요청 자동 분할)와 **Zhipu AI 웹 검색** 제공자; 컨트롤 플레인 자동화를 위한 **플랫폼 범위 API 키**(테넌트 관리, 시스템 설정, 런타임 큐, 감사 로그); **KB 단위 활동 감사 추적**; FAQ 관리 강화(필터링, 태깅, 내보내기, 가져오기 결과 추적); **Langfuse OTLP/OTel 트레이싱** 마이그레이션 및 W3C traceparent 전파; 채팅 헤더 액션을 통한 원클릭 **Markdown 내보내기** 및 참조 드로어의 Wiki 도구 결과 표시; 프롬프트 캐시 가시성; 세션 채널 거버넌스(IM/임베드/API 세션을 관리자 범위로 분리); Feishu 대규모 Wiki 동기화 견고화; 레거시 Neo4j 대화 메모리 의존성 제거. 또한 광범위한 slug 무결성, SSRF 전송, 상태 동기화 강화. 자세한 내용은 [`CHANGELOG.md`](./CHANGELOG.md).
- **v0.7.0** — 세분화된 **범위 지정 API 키 및 Principal 모델**(능력 단위 권한 + KB 단위 제한 + API 통합 플레이그라운드); **런타임 작업 큐 가시성 대시보드 및 Worker 풀 거버넌스**(단계별 풀 + 모델별 동시성 거버너 + 실패 작업 조사/재시도); **다중 인스턴스 스토리지 백엔드**(워크스페이스당 여러 스토리지 인스턴스, KB 단위 바인딩, 기본 인스턴스); **세션 범위 임시 첨부**(이미지/문서 비동기 파싱 + 통합 한도); 추천 질문 및 후속 질문; 안정적인 리소스 레지스트리 및 LLM 컨텍스트 별칭 압축; `@Skill / @MCP` 멘션 기반 범위 지정 Agent 런타임; 대화 중 MCP OAuth; QQBot 및 Lark(Feishu 국제판) IM 통합; Redis TLS; Requesty 모델 제공자 + Keenable 웹 검색; 테넌트리스 프로비저닝 및 제어된 셀프서비스 워크스페이스; 관리자 비밀번호 재설정; 지식 베이스 복제 플로우; `weknora` CLI v0.10. 또한 대규모 보안 강화(SSRF, 비밀 마스킹, SQL 검증, IDOR). 자세한 내용은 [`CHANGELOG.md`](./CHANGELOG.md).
@@ -124,9 +125,11 @@ Feishu, Notion, Yuque 등 외부 플랫폼에서 지식 자동 동기화를 지
| 기능 | 상세 |
|------|------|
-| 지능형 추론 | ReACT 점진적 멀티스텝 추론, 지식 검색·MCP 도구·웹 검색을 자율 오케스트레이션 |
+| 지능형 추론 | ReACT 점진적 멀티스텝 추론, 지식 검색·MCP 도구·스킬 샌드박스·웹 검색을 자율 오케스트레이션 |
| 빠른 Q&A | 지식베이스 기반 RAG Q&A, 빠르고 정확한 답변 |
| Wiki 모드 | Agent가 주도하여 원본 문서에서 구조화된 마크다운 Wiki 페이지를 자동 생성 및 유지 관리; 브라우저 내 수동 편집, 페이지 버전 이력, 라인 단위 diff 및 원클릭 롤백 |
+| 스킬 카탈로그와 샌드박스 | 워크스페이스 스킬 카탈로그(ClawHub / SkillHub / git / zip)를 세션 지속 Docker / E2B / Cube 샌드박스에 설치; `shell_exec`, 파일 도구, 산출물, 설정 단위 네트워크 정책; Local 호스트 프로세스 백엔드 제거 |
+| 장기 메모리 | 크로스 세션 메모리(profile / preference / fact / task / interest), 자동 추출, 사용자 확인, 온디맨드 `search_memory` |
| 도구 호출 | 내장 도구, MCP 도구(OAuth2 원격 서비스·대화 중 OAuth 포함), 웹 검색; `@Skill / @MCP` 멘션으로 턴 단위 Agent 런타임 범위 지정 |
| 대화 전략 | 온라인 프롬프트 편집, 검색 임계값 조정, 멀티턴 문맥 인식, Agent별 인용 출력 토글 |
| 추천 질문 | 지식베이스 콘텐츠 기반 질문 자동 생성 및 답변 후 후속 질문 |
@@ -143,8 +146,9 @@ Feishu, Notion, Yuque 등 외부 플랫폼에서 지식 자동 동기화를 지
| 청크 편집 및 버전 | UI에서 검색 청크를 직접 편집하고 버전별 스냅샷·diff·원클릭 롤백, 편집 후 인덱스 자동 재구축; 생성 질문 추가·수정·삭제·재생성; 문서 커스텀 메타데이터 지원 |
| 업로드 단위 파싱 설정 | 업로드 확인 대화상자 또는 `process_config` API로 파서·청킹·멀티모달(VLM / ASR)·그래프 추출·질문 생성을 배치 단위로 덮어쓰기; reparse 시 설정 변경 지원 |
| 일괄 reparse | 여러 문서의 파싱을 한 번에 재큐잉, 배치 단위 `process_config` 지원 |
-| 데이터 소스 임포트 | Feishu 지식베이스 / Feishu 클라우드 드라이브 / Lark / Notion / Yuque / RSS 피드 자동 동기화(추가 데이터 소스 개발 중), 증분·전체 동기화 지원 |
-| 문서 포맷 | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 이미지 / CSV / Excel / PPT / JSON |
+| 데이터 소스 임포트 | Feishu 지식베이스 / Feishu 클라우드 드라이브 / Lark / GitLab / Tencent IMA / Notion / Yuque / RSS 피드 자동 동기화(추가 데이터 소스 개발 중), 증분·전체 동기화 지원 |
+| 문서 포맷 | PDF / Word / Txt / Markdown / HTML / EPUB / MHTML / 이미지 / CSV / Excel / PPT / JSON / XMind |
+| 자동 태깅 | 파싱 후 지식베이스 기존 태그에서 일치 항목을 증분 연결(새 태그 생성·수동 태그 덮어쓰기 없음) |
| 검색 전략 | BM25 희소 / Dense 밀집 / GraphRAG 그래프 강화 / 부모-자식 청킹 / pgvector HNSW 가속(1024차원) / 다차원 인덱싱 |
| 일괄 선택 및 태깅 | KB 목록에서 마키(드래그) 다중 선택으로 일괄 reparse 및 일괄 태깅(공통 태그 자동 선택) |
| E2E 테스트 | 전체 파이프라인 시각화, 리콜 적중률·BLEU / ROUGE 지표 평가 |
@@ -153,13 +157,13 @@ Feishu, Notion, Yuque 등 외부 플랫폼에서 지식 자동 동기화를 지
| 기능 | 상세 |
|------|------|
-| LLM | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / Ollama |
+| LLM | OpenAI / Azure OpenAI / Anthropic (Claude) / DeepSeek / Qwen (Alibaba Cloud) / Zhipu / Hunyuan / Doubao (Volcengine) / Gemini / MiniMax / NVIDIA / Novita AI / SiliconFlow / OpenRouter / Requesty / LiteLLM / Ollama |
| Embedding | Ollama / BGE / GTE / OpenAI 호환 API |
| 벡터 DB | PostgreSQL (pgvector) / Elasticsearch / OpenSearch / Milvus / Weaviate / Qdrant / Apache Doris / Tencent VectorDB |
| 오브젝트 스토리지 | 로컬 / MinIO / AWS S3(IAM Role / IRSA 기본 자격 증명 체인 지원) / Volcengine TOS / Alibaba Cloud OSS / Kingsoft Cloud KS3 / Huawei Cloud OBS; **워크스페이스당 여러 스토리지 인스턴스**, KB 단위 바인딩 및 기본 인스턴스 |
| IM 통합 | WeChat Work / Feishu / Lark(Feishu 국제판) / QQBot / Slack / Telegram / DingTalk / Mattermost / WeChat / Yunzhijia |
| 웹사이트 임베드 | 임베드 Widget으로 에이전트 게시, 도메인 허용 목록·속도 제한·보안 모드 Token 교환 |
-| 웹 검색 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI |
+| 웹 검색 | DuckDuckGo / Bing / Google / Tavily / Baidu / Ollama / SearXNG / Keenable / Zhipu AI / Exa / Metaso |
| API 통합 | 범위 지정 API 키(능력 단위 권한 + KB 단위 제한 + 스로틀링된 last_used 추적)와 API 통합 플레이그라운드; MCP OAuth 및 임베드 세션을 Principal 단위로 격리; `resource_urls=public`으로 바로 로드 가능한 파일 / 이미지 URL을 반환하여 인증 프록시 재호출 제거 |
| MCP Server | 공식 PyPI 패키지 `tencent-weknora-mcp`, 29개 도구, stdio / SSE / HTTP 3가지 트랜스포트 지원 |
diff --git a/VERSION b/VERSION
index 7486fdbc..a3df0a69 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-0.7.2
+0.8.0
diff --git a/client/README.md b/client/README.md
index d3a36f69..a0b7be99 100644
--- a/client/README.md
+++ b/client/README.md
@@ -16,7 +16,8 @@
8. **消息管理**:获取和删除会话消息
9. **模型管理**:创建、获取、更新和删除模型
10. **沙箱技能**:向沙箱配置安装技能(zip 上传,或从 ClawHub / SkillHub / GitHub 等来源),并配置技能所需的环境变量
-11. **认证**:登录、刷新令牌、切换激活空间(`SwitchTenant` 会写入最近活跃租户偏好)
+11. **长期记忆**:当前用户的跨会话记忆(设置开关、条目增删改、确认/否决、主题、文档亲和度、导出、立刻整理)
+12. **认证**:登录、刷新令牌、切换激活空间(`SwitchTenant` 会写入最近活跃租户偏好)
## 使用方法
diff --git a/client/README_EN.md b/client/README_EN.md
index 99e9fd41..7c75e154 100644
--- a/client/README_EN.md
+++ b/client/README_EN.md
@@ -16,7 +16,8 @@ The client includes the following main functional modules:
8. **Model Management**: Create, retrieve, update, and delete models
9. **Evaluation Function**: Start evaluation tasks and get evaluation results
10. **Sandbox skills**: Install a skill onto a sandbox config (zip upload, or ClawHub / SkillHub / GitHub source) and configure the environment variables it needs
-11. **Auth**: Login, refresh tokens, and switch the active workspace (`SwitchTenant` records the last-active-tenant preference)
+11. **Long-term memory**: The caller's cross-session memories (settings, items, confirm/reject, topics, document affinity, export, consolidate)
+12. **Auth**: Login, refresh tokens, and switch the active workspace (`SwitchTenant` records the last-active-tenant preference)
## Usage
diff --git a/client/memory.go b/client/memory.go
new file mode 100644
index 00000000..548f6512
--- /dev/null
+++ b/client/memory.go
@@ -0,0 +1,376 @@
+package client
+
+import (
+ "context"
+ "fmt"
+ "net/http"
+ "net/url"
+ "strconv"
+ "time"
+)
+
+// Memory settings, items, topics and document affinity all operate on the
+// caller's own memory space. There is no subject id on the wire: the server
+// derives identity from the credentials, so a scoped API key cannot inherit
+// another person's memories. Full-access API keys or a Bearer session are
+// required.
+
+// MemorySettings is the merged workspace + personal memory switch.
+type MemorySettings struct {
+ WorkspaceEnabled bool `json:"workspace_enabled"`
+ UserEnabled bool `json:"user_enabled"`
+ Effective bool `json:"effective"`
+ WriteMode string `json:"write_mode"`
+ ItemCount int `json:"item_count"`
+ MaxItems int `json:"max_items"`
+}
+
+// MemoryItem is one long-term memory row as returned by the API.
+type MemoryItem struct {
+ ID string `json:"id"`
+ Kind string `json:"kind"`
+ Content string `json:"content"`
+ Topic string `json:"topic,omitempty"`
+ Importance int `json:"importance"`
+ Origin string `json:"origin"`
+ Status string `json:"status"`
+ SourceSessionID string `json:"source_session_id,omitempty"`
+ SourceMessageID string `json:"source_message_id,omitempty"`
+ ValidFrom time.Time `json:"valid_from"`
+ InvalidAt *time.Time `json:"invalid_at,omitempty"`
+ ExpiresAt *time.Time `json:"expires_at,omitempty"`
+ SupersededBy string `json:"superseded_by,omitempty"`
+ LastUsedAt *time.Time `json:"last_used_at,omitempty"`
+ UseCount int `json:"use_count"`
+ CreatedAt time.Time `json:"created_at"`
+ UpdatedAt time.Time `json:"updated_at"`
+}
+
+// MemoryTopic is a subject the extractor is still counting before promoting
+// it to a long-term interest.
+type MemoryTopic struct {
+ ID string `json:"id"`
+ Topic string `json:"topic"`
+ Aliases []string `json:"aliases"`
+ Hits int `json:"hits"`
+ Threshold int `json:"threshold"`
+ LastSeenAt time.Time `json:"last_seen_at"`
+}
+
+// MemoryDocument is a knowledge entry this person keeps drawing answers from.
+type MemoryDocument struct {
+ ID string `json:"id"`
+ KnowledgeID string `json:"knowledge_id"`
+ KnowledgeBaseID string `json:"knowledge_base_id"`
+ Title string `json:"title"`
+ Hits int `json:"hits"`
+ LastUsedAt time.Time `json:"last_used_at"`
+}
+
+// MemoryConsolidationResult is what POST /memory/consolidate returns.
+type MemoryConsolidationResult struct {
+ Merged int `json:"merged"`
+ Demoted int `json:"demoted"`
+ Expired int `json:"expired"`
+ Reviewed int `json:"reviewed"`
+ Candidates int `json:"candidates"`
+ Skipped string `json:"skipped,omitempty"`
+}
+
+// MemoryExport is the JSON snapshot from GET /memory/export.
+type MemoryExport struct {
+ Total int64 `json:"total"`
+ Truncated bool `json:"truncated"`
+ Items []*MemoryItem `json:"data"`
+}
+
+type memorySettingsResponse struct {
+ Success bool `json:"success"`
+ Data *MemorySettings `json:"data"`
+}
+
+type memoryItemResponse struct {
+ Success bool `json:"success"`
+ Data *MemoryItem `json:"data"`
+}
+
+type memoryListResponse struct {
+ Success bool `json:"success"`
+ Data []*MemoryItem `json:"data"`
+ Total int64 `json:"total"`
+}
+
+type memoryTopicListResponse struct {
+ Success bool `json:"success"`
+ Data []*MemoryTopic `json:"data"`
+ Total int64 `json:"total"`
+}
+
+type memoryDocumentListResponse struct {
+ Success bool `json:"success"`
+ Data []*MemoryDocument `json:"data"`
+ Total int64 `json:"total"`
+}
+
+type memoryClearResponse struct {
+ Success bool `json:"success"`
+ Removed int64 `json:"removed"`
+}
+
+type memoryConsolidateResponse struct {
+ Success bool `json:"success"`
+ Data *MemoryConsolidationResult `json:"data"`
+}
+
+func memoryListQuery(status string, limit, offset int) url.Values {
+ q := url.Values{}
+ if status != "" {
+ q.Set("status", status)
+ }
+ if limit > 0 {
+ q.Set("limit", strconv.Itoa(limit))
+ }
+ if offset > 0 {
+ q.Set("offset", strconv.Itoa(offset))
+ }
+ return q
+}
+
+// GetMemorySettings returns the merged workspace + personal memory switch.
+func (c *Client) GetMemorySettings(ctx context.Context) (*MemorySettings, error) {
+ resp, err := c.doRequest(ctx, http.MethodGet, "/api/v1/memory/settings", nil, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memorySettingsResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// UpdateMemorySettings turns the caller's own long-term memory on or off.
+func (c *Client) UpdateMemorySettings(ctx context.Context, enabled bool) (*MemorySettings, error) {
+ body := map[string]bool{"enabled": enabled}
+ resp, err := c.doRequest(ctx, http.MethodPut, "/api/v1/memory/settings", body, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memorySettingsResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// ListMemoryItems pages through the caller's memories. status may be empty
+// (all) or one of active / superseded / archived / pending.
+func (c *Client) ListMemoryItems(ctx context.Context, status string, limit, offset int) ([]*MemoryItem, int64, error) {
+ resp, err := c.doRequest(ctx, http.MethodGet, "/api/v1/memory/items", nil, memoryListQuery(status, limit, offset))
+ if err != nil {
+ return nil, 0, err
+ }
+ var out memoryListResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, 0, err
+ }
+ return out.Data, out.Total, nil
+}
+
+// CreateMemoryItem manually adds a long-term memory. kind is one of
+// profile / preference / fact / task / interest.
+func (c *Client) CreateMemoryItem(ctx context.Context, kind, content string, importance int) (*MemoryItem, error) {
+ body := map[string]interface{}{
+ "kind": kind,
+ "content": content,
+ "importance": importance,
+ }
+ resp, err := c.doRequest(ctx, http.MethodPost, "/api/v1/memory/items", body, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memoryItemResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// UpdateMemoryItem changes content and importance. After an edit the extractor
+// will not overwrite the row.
+func (c *Client) UpdateMemoryItem(ctx context.Context, id, content string, importance int) (*MemoryItem, error) {
+ if id == "" {
+ return nil, fmt.Errorf("memory id is required")
+ }
+ body := map[string]interface{}{
+ "content": content,
+ "importance": importance,
+ }
+ path := "/api/v1/memory/items/" + url.PathEscape(id)
+ resp, err := c.doRequest(ctx, http.MethodPut, path, body, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memoryItemResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// DeleteMemoryItem permanently removes one memory.
+func (c *Client) DeleteMemoryItem(ctx context.Context, id string) error {
+ if id == "" {
+ return fmt.Errorf("memory id is required")
+ }
+ path := "/api/v1/memory/items/" + url.PathEscape(id)
+ resp, err := c.doRequest(ctx, http.MethodDelete, path, nil, nil)
+ if err != nil {
+ return err
+ }
+ return parseResponse(resp, nil)
+}
+
+// ConfirmMemoryItem accepts an inferred (pending) memory so it starts taking
+// effect.
+func (c *Client) ConfirmMemoryItem(ctx context.Context, id string) (*MemoryItem, error) {
+ if id == "" {
+ return nil, fmt.Errorf("memory id is required")
+ }
+ path := "/api/v1/memory/items/" + url.PathEscape(id) + "/confirm"
+ resp, err := c.doRequest(ctx, http.MethodPost, path, nil, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memoryItemResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// RejectMemoryItem declines an inferred memory and records the rejection so
+// the extractor does not silently re-add it.
+func (c *Client) RejectMemoryItem(ctx context.Context, id string) error {
+ if id == "" {
+ return fmt.Errorf("memory id is required")
+ }
+ path := "/api/v1/memory/items/" + url.PathEscape(id) + "/reject"
+ resp, err := c.doRequest(ctx, http.MethodPost, path, nil, nil)
+ if err != nil {
+ return err
+ }
+ return parseResponse(resp, nil)
+}
+
+// ClearMemoryItems permanently deletes every memory belonging to the caller.
+func (c *Client) ClearMemoryItems(ctx context.Context) (int64, error) {
+ resp, err := c.doRequest(ctx, http.MethodDelete, "/api/v1/memory/items", nil, nil)
+ if err != nil {
+ return 0, err
+ }
+ var out memoryClearResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return 0, err
+ }
+ return out.Removed, nil
+}
+
+// ListMemoryTopics pages through subjects the extractor is still counting.
+func (c *Client) ListMemoryTopics(ctx context.Context, limit, offset int) ([]*MemoryTopic, int64, error) {
+ resp, err := c.doRequest(ctx, http.MethodGet, "/api/v1/memory/topics", nil, memoryListQuery("", limit, offset))
+ if err != nil {
+ return nil, 0, err
+ }
+ var out memoryTopicListResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, 0, err
+ }
+ return out.Data, out.Total, nil
+}
+
+// PromoteMemoryTopic turns a counted topic into a long-term interest without
+// waiting for the remaining hits.
+func (c *Client) PromoteMemoryTopic(ctx context.Context, id string) (*MemoryItem, error) {
+ if id == "" {
+ return nil, fmt.Errorf("topic id is required")
+ }
+ path := "/api/v1/memory/topics/" + url.PathEscape(id) + "/promote"
+ resp, err := c.doRequest(ctx, http.MethodPost, path, nil, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memoryItemResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
+
+// DeleteMemoryTopic stops tracking a topic that has not been promoted yet.
+func (c *Client) DeleteMemoryTopic(ctx context.Context, id string) error {
+ if id == "" {
+ return fmt.Errorf("topic id is required")
+ }
+ path := "/api/v1/memory/topics/" + url.PathEscape(id)
+ resp, err := c.doRequest(ctx, http.MethodDelete, path, nil, nil)
+ if err != nil {
+ return err
+ }
+ return parseResponse(resp, nil)
+}
+
+// ListMemoryDocuments pages through documents this person keeps citing.
+func (c *Client) ListMemoryDocuments(ctx context.Context, limit, offset int) ([]*MemoryDocument, int64, error) {
+ resp, err := c.doRequest(ctx, http.MethodGet, "/api/v1/memory/documents", nil, memoryListQuery("", limit, offset))
+ if err != nil {
+ return nil, 0, err
+ }
+ var out memoryDocumentListResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, 0, err
+ }
+ return out.Data, out.Total, nil
+}
+
+// DeleteMemoryDocument stops using one document for personalized retrieval.
+func (c *Client) DeleteMemoryDocument(ctx context.Context, id string) error {
+ if id == "" {
+ return fmt.Errorf("document affinity id is required")
+ }
+ path := "/api/v1/memory/documents/" + url.PathEscape(id)
+ resp, err := c.doRequest(ctx, http.MethodDelete, path, nil, nil)
+ if err != nil {
+ return err
+ }
+ return parseResponse(resp, nil)
+}
+
+// ExportMemory downloads a JSON snapshot of every memory belonging to the
+// caller. Truncated is true only if the safety ceiling clipped the file.
+func (c *Client) ExportMemory(ctx context.Context) (*MemoryExport, error) {
+ resp, err := c.doRequest(ctx, http.MethodGet, "/api/v1/memory/export", nil, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out MemoryExport
+ out.Items = nil
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return &out, nil
+}
+
+// ConsolidateMemory merges near-duplicate items and archives expired ones
+// without waiting for the daily background pass.
+func (c *Client) ConsolidateMemory(ctx context.Context) (*MemoryConsolidationResult, error) {
+ resp, err := c.doRequest(ctx, http.MethodPost, "/api/v1/memory/consolidate", nil, nil)
+ if err != nil {
+ return nil, err
+ }
+ var out memoryConsolidateResponse
+ if err := parseResponse(resp, &out); err != nil {
+ return nil, err
+ }
+ return out.Data, nil
+}
diff --git a/docs/QA.md b/docs/QA.md
index 7f89d961..974b50d8 100644
--- a/docs/QA.md
+++ b/docs/QA.md
@@ -445,5 +445,49 @@ docker run -d -p 8081:8081 weknora-docs
行为变化提醒:工具执行失败时,MCPServer 2.x 返回 `CallToolResult(isError=True)`,不再像旧版低层 API 那样以成功响应返回 `"Error executing …"` 文本前缀。只解析 `content[0].text` 的客户端通常无感,依赖 `isError` 标志的集成方行为会更符合 MCP 规范。
+## 45. 升级到 0.8.0 后技能沙箱起不来 / 找不到 Local 后端?
+
+0.8.0 **移除了 Local 宿主机进程沙箱**。技能执行改为会话级常驻沙箱,三个后端共用同一套协议:
+
+- **Docker**(单机 / 私有化):默认**关闭**。本机 `docker.sock` 等同宿主机 root,需系统管理员在 **设置 → 系统设置 → 网络安全** 打开,或设置 `WEKNORA_SANDBOX_DOCKER_ENABLED=true`。打开后才会出现「添加 Docker 后端」入口;已有配置仍可查看/删除。
+- **E2B**:E2B Cloud,或任意 E2B 兼容控制面(含自托管)。
+- **CubeSandbox**:集群模板 + 网络策略。
+
+原 Local 配置需要按上面任一后端重建。每个空间可配多个沙箱实例,并可为每个配置设置**网络策略**(默认放行出站、关闭公网入站;可改成默认拒绝出站再写允许名单)。详见 [`docs/sandbox-docker-backend.md`](./sandbox-docker-backend.md) 与 [`docs/sandbox-protocol.md`](./sandbox-protocol.md)。
+
+## 46. 技能目录和沙箱配置是什么关系?安装一直转圈怎么办?
+
+0.8.0 把技能做成空间级目录(迁移 `000086_tenant_skills` / `000090_skill_catalog`),再**按沙箱配置安装成快照**:
+
+1. 在 **设置 → 技能沙箱** 建好后端配置;
+2. 从 ClawHub(`@owner/slug`)、SkillHub / skills.sh、GitHub/GitLab URL 或 zip 上传安装;
+3. 安装抽屉会保持打开并显示环形进度;卡住时用「停止安装」,再用「重新安装」走已保存的安装包。
+
+环境变量分两层:**空间级**(Admin,该空间所有人共用)和**个人级**(`/api/v1/me/env-vars`,值永远不会读回)。技能声明的 `WEKNORA_*` 凭据可以按人填写。卸载沙箱里的技能不会删掉目录里的安装包。
+
+## 47. 0.7.1 删了 Neo4j 会话记忆,0.8.0 的「长期记忆」是一回事吗?
+
+不是。0.7.1 去掉的是旧版 **Neo4j episodic conversation memory**;**知识图谱(GraphRAG)仍然用 Neo4j**。0.8.0 的长期记忆是全新产品(迁移 `000084_memory`):
+
+- 空间管理员先打开,用户还可以再关掉自己的;
+- 类型:`profile` / `preference` / `fact` / `task` / `interest`;
+- 自动抽取的条目先停在「待确认」,不会静默写进提示词;
+- 常驻画像每轮注入,情境记忆按需召回,Agent 也可用 `search_memory`;
+- 反复引用的文档会形成亲和度,检索时加权。
+
+接口在 `/api/v1/memory/*`,只操作当前调用者自己的记忆,需要 full-access API Key 或登录会话。
+
+## 48. 文档解析能否不经过 docreader?anydoc 是什么?
+
+0.8.0 引入进程内 **anydoc** 引擎(`third_party/anydoc-go`)。当绑定已链接时,anydoc 能转换的格式(含 doc/docx/ppt/pptx)会优先走 Go 进程,不再先打到 docreader。PPT/PPTX 在没有引擎规则时仍默认 MarkItDown。官方 app 镜像默认链接 anydoc。解析失败或格式不在 anydoc 覆盖范围时,仍回退到 docreader / MarkItDown / MinerU 等既有引擎。
+
+## 49. OIDC 登录提示签名无效,或想跳过前端直接 302?
+
+0.8.0 会通过 JWKS **校验 ID Token 签名**(#2799),显式配置的 token endpoint 同样走这条路径。请确认 IdP 的 JWKS URL 可达、密钥已轮换到当前 kid。若要做门户级跳转,可调用 `GET /auth/oidc/start` 直接 302 到 IdP,不必先渲染 SPA 握手页。
+
+## 50. 开启复杂密码后注册 / 改密失败?
+
+系统设置或 `WEKNORA_AUTH_COMPLEX_PASSWORD_ENABLED=true` 打开后,密码必须同时包含大写、小写、数字和特殊字符(`!@#$%^&*()_+-=[]{}|;:,.<>?`),长度 8–32。注册、个人中心改密、管理员重置走同一套规则。未打开时仍只要求长度。
+
## P.S.
如果以上方式未解决问题,请在issue中描述您的问题,并提供必要的日志信息辅助我们进行问题排查
diff --git a/docs/api/README.md b/docs/api/README.md
index 4b368561..6fbe7827 100644
--- a/docs/api/README.md
+++ b/docs/api/README.md
@@ -132,6 +132,7 @@ WeKnora API 按功能分为以下几类:
| MCP 服务 | MCP 工具服务管理 | [mcp-service.md](./mcp-service.md) |
| 组织管理 | 组织、成员、知识库/智能体共享 | [organization.md](./organization.md) |
| Skills | 预装与已安装的智能体技能、技能环境变量 | [skill.md](./skill.md) |
+| 长期记忆 | 当前用户的跨会话记忆(设置、条目、主题、文档亲和度、导出、整理) | [memory.md](./memory.md) |
| 网络搜索 | 网络搜索服务商 | [web-search.md](./web-search.md) |
| 向量存储 | 向量数据库连接管理 | [vector-store.md](./vector-store.md) |
| 存储后端 | 对象/文件存储实例(多实例)管理 | [storage-backend.md](./storage-backend.md) |
diff --git a/docs/api/memory.md b/docs/api/memory.md
new file mode 100644
index 00000000..fb7b4150
--- /dev/null
+++ b/docs/api/memory.md
@@ -0,0 +1,96 @@
+# 长期记忆 API
+
+[返回目录](./README.md)
+
+长期记忆空间始终绑定在**当前调用者**上:路径里没有 subject id,服务端从凭证推导身份。因此 Viewer+ 的会话令牌,或 **full-access API Key** 才能访问;带知识库范围的集成 Key 不能继承某个人的记忆。
+
+工作空间管理员先在设置里打开空间级开关,用户还可以再关掉自己的记忆。推断出的条目会停在 `pending`,确认后才进入提示词。
+
+| 方法 | 路径 | 描述 |
+| ---- | ---- | ---- |
+| GET | `/memory/settings` | 获取合并后的记忆开关(空间级 + 个人级)与条数 |
+| PUT | `/memory/settings` | 开启或关闭当前用户自己的长期记忆 |
+| GET | `/memory/items` | 分页列出记忆,可按状态过滤 |
+| POST | `/memory/items` | 手动新增一条记忆 |
+| PUT | `/memory/items/{id}` | 修改内容与重要度(之后不会被后台抽取覆盖) |
+| DELETE | `/memory/items/{id}` | 永久删除一条记忆 |
+| POST | `/memory/items/{id}/confirm` | 确认一条推断出的记忆 |
+| POST | `/memory/items/{id}/reject` | 否决一条推断出的记忆 |
+| DELETE | `/memory/items` | 清空当前用户的全部记忆 |
+| GET | `/memory/topics` | 列出尚未提升为长期关注的主题计数 |
+| POST | `/memory/topics/{id}/promote` | 立即把主题记为长期关注 |
+| DELETE | `/memory/topics/{id}` | 停止跟踪一个主题 |
+| GET | `/memory/documents` | 列出反复引用的文档(未达习惯门槛的不展示) |
+| DELETE | `/memory/documents/{id}` | 停止用某份文档做个性化检索 |
+| GET | `/memory/export` | 以 JSON 导出全部记忆 |
+| POST | `/memory/consolidate` | 立刻整理(合并近义条目、归档到期事项) |
+
+## GET `/memory/settings`
+
+```curl
+curl --location 'http://localhost:8080/api/v1/memory/settings' \
+--header 'Authorization: Bearer '
+```
+
+**响应**:
+
+```json
+{
+ "success": true,
+ "data": {
+ "workspace_enabled": true,
+ "user_enabled": true,
+ "effective": true,
+ "write_mode": "auto",
+ "item_count": 12,
+ "max_items": 200
+ }
+}
+```
+
+`write_mode` 为 `explicit_only`(只记用户明确要求记住的)或 `auto`(后台从对话蒸馏)。`effective` = 空间开关 ∧ 个人开关。
+
+## PUT `/memory/settings`
+
+```curl
+curl --location --request PUT 'http://localhost:8080/api/v1/memory/settings' \
+--header 'Authorization: Bearer ' \
+--header 'Content-Type: application/json' \
+--data '{"enabled": true}'
+```
+
+`enabled` 必填。只改个人开关,不能用这个接口改空间级配置。
+
+## GET `/memory/items`
+
+查询参数:
+
+| 参数 | 说明 |
+| ---- | ---- |
+| `status` | 可选:`active` / `superseded` / `archived` / `pending`;省略则不过滤 |
+| `limit` | 默认 50,最大 200 |
+| `offset` | 默认 0 |
+
+`kind` 取值:`profile` / `preference` / `fact` / `task` / `interest`。
+
+## POST `/memory/items`
+
+```json
+{
+ "kind": "preference",
+ "content": "回答直接给结论,少铺垫",
+ "importance": 3
+}
+```
+
+## POST `/memory/items/{id}/confirm` / `reject`
+
+推断出的记忆(`status=pending`)确认后才注入提示词;否决会留下 tombstone,避免下一轮蒸馏把同一句话再写回来。
+
+## GET `/memory/export`
+
+返回 `{success, total, truncated, data}`,并带 `Content-Disposition: attachment; filename="weknora-memories.json"`。`truncated` 仅在触达导出上限(2 万条)时为 true。
+
+## POST `/memory/consolidate`
+
+不等待每日后台整理,立刻合并意思接近的条目并归档到期事项。返回 `merged` / `demoted` / `expired` / `reviewed` / `candidates`;若什么都没合并,`skipped` 会说明原因(例如 `too_few_items`)。
diff --git a/docs/docs.go b/docs/docs.go
index 76dd2e98..29dfd25c 100644
--- a/docs/docs.go
+++ b/docs/docs.go
@@ -12376,6 +12376,70 @@ const docTemplate = `{
}
}
},
+ "/sandbox-configs/{id}/skills/{skillId}/stop": {
+ "post": {
+ "security": [
+ {
+ "Bearer": []
+ },
+ {
+ "ApiKeyAuth": []
+ }
+ ],
+ "description": "Abort an in-flight install so the operator can retry or uninstall. After a process restart the row may still say installing with no live process; this rewrites it immediately instead of waiting for the stuck-run reaper. Removal is not stopped.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "SandboxConfig"
+ ],
+ "summary": "Stop a skill install",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Sandbox config ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ },
+ {
+ "type": "string",
+ "description": "Skill ID",
+ "name": "skillId",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "Stopped skill",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ },
+ "400": {
+ "description": "Skill is not installing",
+ "schema": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError"
+ }
+ },
+ "401": {
+ "description": "Unauthorized",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ },
+ "404": {
+ "description": "Skill not found",
+ "schema": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError"
+ }
+ }
+ }
+ }
+ },
"/sandbox-configs/{id}/skills/{skillId}/transcript": {
"get": {
"security": [
@@ -13369,6 +13433,184 @@ const docTemplate = `{
}
}
},
+ "/skills/catalog": {
+ "get": {
+ "security": [
+ {
+ "Bearer": []
+ },
+ {
+ "ApiKeyAuth": []
+ }
+ ],
+ "description": "Returns every skill definition in this workspace and which sandbox configs it is installed on.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "List workspace skills",
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ },
+ "post": {
+ "description": "Records a skill without installing it. Send a zip as multipart field \"file\", or JSON {\"source\":\"...\"}.",
+ "consumes": [
+ "application/json",
+ "multipart/form-data"
+ ],
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Add a skill to the workspace catalog",
+ "responses": {
+ "201": {
+ "description": "Created",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}": {
+ "delete": {
+ "description": "Refused while any sandbox still has an installation of this skill.",
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Delete a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/files": {
+ "get": {
+ "description": "Lists the stored catalog bundle. Files belong to the skill definition, not a sandbox install.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "List files of a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/files/content": {
+ "get": {
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Read one file of a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ },
+ {
+ "type": "string",
+ "description": "Skill-root-relative file path",
+ "name": "path",
+ "in": "query",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/install": {
+ "post": {
+ "description": "Runs the existing snapshot install onto each named sandbox config.",
+ "consumes": [
+ "application/json"
+ ],
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Install a catalog skill onto sandboxes",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "202": {
+ "description": "Accepted",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
"/storage-backends": {
"get": {
"security": [
@@ -17579,7 +17821,7 @@ const docTemplate = `{
"type": "integer"
},
"parser_engine_rules": {
- "description": "ParserEngineRules configures which parser engine to use for each file type.\nWhen empty, the builtin engine is used for all types.",
+ "description": "ParserEngineRules configures which parser engine to use for each file type.\nWhen empty, DefaultParserEngine is used (builtin/simple routing, except\ntypes that only a specific engine can parse: ppt/pptx fall back to\nmarkitdown). A linked anydoc binding is preferred for every type it\nconverts.",
"type": "array",
"items": {
"$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.ParserEngineRule"
@@ -17737,6 +17979,61 @@ const docTemplate = `{
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.CubeEgressRule": {
+ "type": "object",
+ "properties": {
+ "audit": {
+ "description": "Audit is none | metadata | full. Empty uses the server default.",
+ "type": "string"
+ },
+ "deny": {
+ "description": "Deny inverts the action, which defaults to allow. A deny rule still\nneeds Host or SNI: the target has to reach CubeEgress for it to answer\nwith a request-level 403 instead of the network layer dropping it.",
+ "type": "boolean"
+ },
+ "host": {
+ "type": "string"
+ },
+ "inject": {
+ "description": "Inject adds credential headers on allowed HTTPS requests.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.CubeHeaderInject"
+ }
+ },
+ "methods": {
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "name": {
+ "type": "string"
+ },
+ "path": {
+ "type": "string"
+ },
+ "scheme": {
+ "type": "string"
+ },
+ "sni": {
+ "type": "string"
+ }
+ }
+ },
+ "github_com_Tencent_WeKnora_internal_types.CubeHeaderInject": {
+ "type": "object",
+ "properties": {
+ "format": {
+ "type": "string"
+ },
+ "header": {
+ "type": "string"
+ },
+ "secret": {
+ "type": "string"
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.CubeSandboxConfig": {
"type": "object",
"properties": {
@@ -17908,11 +18205,11 @@ const docTemplate = `{
"type": "integer"
},
"max_completion_tokens": {
- "description": "Maximum completion tokens (only for normal mode)",
+ "description": "Maximum completion tokens. Quick-answer uses this for the RAG answer.\nSmart-reasoning ReAct rounds send this value as-is (zero becomes\nDefaultMaxCompletionTokens at call time: 4096, or 24576 with a sandbox).",
"type": "integer"
},
"max_iterations": {
- "description": "===== Agent Mode Settings =====\nMaximum ReAct iterations. -1 is unlimited; 0 is unset (default applied).",
+ "description": "===== Agent Mode Settings =====\nMaximum iterations for the ReAct loop. Zero is unset (filled with a\ndefault). A negative value is unlimited: the loop runs until the model\nstops, the user cancels, or another guard fires.",
"type": "integer"
},
"mcp_auth_wait_timeout": {
@@ -18200,6 +18497,21 @@ const docTemplate = `{
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.E2BHostRule": {
+ "type": "object",
+ "properties": {
+ "headers": {
+ "description": "Headers values are credentials and are encrypted at rest; names stay\nreadable so operators can see which headers are injected.",
+ "type": "object",
+ "additionalProperties": {
+ "type": "string"
+ }
+ },
+ "host": {
+ "type": "string"
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.E2BSandboxConfig": {
"type": "object",
"properties": {
@@ -19861,6 +20173,10 @@ const docTemplate = `{
"base_url": {
"type": "string"
},
+ "context_window": {
+ "description": "ContextWindow is the model's total context window in tokens and\nMaxOutputTokens the most it emits in one response. Both are provider\nfacts the agent cannot discover but has to act on: the context window is\nwhat decides when conversation history gets compacted, and assuming a\nwindow larger than the real one means compaction never fires and the\nprovider rejects the request mid-conversation instead. 0 means unknown,\nwhich falls back to DefaultMaxContextTokens.",
+ "type": "integer"
+ },
"custom_headers": {
"description": "CustomHeaders 允许在调用远程模型 API 时附加自定义 HTTP 请求头,\n用途类似 Python OpenAI SDK 的 extra_headers 参数,\n常见场景包括透传企业网关鉴权信息、追踪 ID、路由标识等。\n保留字段(Authorization、api-key、Content-Type、Accept 等)会在运行期被忽略以避免破坏签名/鉴权流程。",
"type": "object",
@@ -19885,6 +20201,9 @@ const docTemplate = `{
"description": "MaxConcurrency caps concurrent in-flight BACKGROUND (ingestion /\nenrichment) calls to THIS specific model, keyed by model ID and shared\nacross all replicas. 0 (the default) means \"fall back to the\nprocess-wide model.max_concurrency\". Interactive user-facing calls are\nnever gated. Only chat / vlm / embedding honour this (see limiter.Gate).",
"type": "integer"
},
+ "max_output_tokens": {
+ "type": "integer"
+ },
"parameter_size": {
"description": "Ollama model parameter size (e.g., \"7B\", \"13B\", \"70B\")",
"type": "string"
@@ -19929,6 +20248,7 @@ const docTemplate = `{
"ModelSourceGemini": "Gemini model",
"ModelSourceHunyuan": "Hunyuan model",
"ModelSourceJina": "Jina AI model",
+ "ModelSourceLiteLLM": "LiteLLM proxy model",
"ModelSourceLocal": "Local model",
"ModelSourceMimo": "Mimo model",
"ModelSourceMinimax": "Minimax mode",
@@ -19936,7 +20256,6 @@ const docTemplate = `{
"ModelSourceNvidia": "NVIDIA model",
"ModelSourceOpenAI": "OpenAI model",
"ModelSourceOpenRouter": "OpenRouter model",
- "ModelSourceLiteLLM": "LiteLLM proxy model",
"ModelSourceRemote": "Remote model",
"ModelSourceRequesty": "Requesty model",
"ModelSourceSiliconFlow": "SiliconFlow model",
@@ -20871,6 +21190,47 @@ const docTemplate = `{
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy": {
+ "type": "object",
+ "properties": {
+ "allow_out": {
+ "description": "AllowOut accepts IPv4, IPv4 CIDR, a DNS name, or a single-label\nwildcard such as \"*.example.com\". Domain entries are only meaningful\ntogether with a deny-all; see ValidateSandboxNetworkPolicy.",
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "allow_public_inbound": {
+ "description": "AllowPublicInbound exposes the sandbox's public URL without a\ncredential. false requires every inbound request to carry the\nper-sandbox traffic access token.",
+ "type": "boolean"
+ },
+ "cube_rules": {
+ "description": "CubeRules are validated whenever present and consumed only by the Cube\nprovider adapter.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.CubeEgressRule"
+ }
+ },
+ "deny_egress_by_default": {
+ "description": "DenyEgressByDefault installs a 0.0.0.0/0 deny-all, after which only\nAllowOut (and L7 rule targets) can reach the network. false allows\npublic egress, which is what skill installs need.",
+ "type": "boolean"
+ },
+ "deny_out": {
+ "description": "DenyOut accepts IPv4 and IPv4 CIDR only. Neither provider can deny a\ndomain: denial is a pure longest-prefix match on the destination IP.",
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "e2b_host_rules": {
+ "description": "E2BHostRules are validated whenever present and consumed only by the E2B\nprovider adapter.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.E2BHostRule"
+ }
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.SearchParams": {
"type": "object",
"properties": {
@@ -21706,6 +22066,14 @@ const docTemplate = `{
"type": "string"
}
},
+ "network": {
+ "description": "Network is the outbound/inbound network policy applied to every sandbox\ncreated from this config — chat sessions, skill installs and deep\nconnectivity probes alike. nil and the zero value mean the same thing:\noutbound egress allowed, inbound public access closed.",
+ "allOf": [
+ {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy"
+ }
+ ]
+ },
"sandbox_type": {
"description": "SandboxType is cube, e2b, or docker; disabled is the hidden policy row.",
"type": "string"
diff --git a/docs/images/architecture.png b/docs/images/architecture.png
index 89117fae..6cdc21e1 100644
Binary files a/docs/images/architecture.png and b/docs/images/architecture.png differ
diff --git a/docs/swagger.json b/docs/swagger.json
index fec4cbeb..b2d18ef7 100644
--- a/docs/swagger.json
+++ b/docs/swagger.json
@@ -12369,6 +12369,70 @@
}
}
},
+ "/sandbox-configs/{id}/skills/{skillId}/stop": {
+ "post": {
+ "security": [
+ {
+ "Bearer": []
+ },
+ {
+ "ApiKeyAuth": []
+ }
+ ],
+ "description": "Abort an in-flight install so the operator can retry or uninstall. After a process restart the row may still say installing with no live process; this rewrites it immediately instead of waiting for the stuck-run reaper. Removal is not stopped.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "SandboxConfig"
+ ],
+ "summary": "Stop a skill install",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Sandbox config ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ },
+ {
+ "type": "string",
+ "description": "Skill ID",
+ "name": "skillId",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "Stopped skill",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ },
+ "400": {
+ "description": "Skill is not installing",
+ "schema": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError"
+ }
+ },
+ "401": {
+ "description": "Unauthorized",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ },
+ "404": {
+ "description": "Skill not found",
+ "schema": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError"
+ }
+ }
+ }
+ }
+ },
"/sandbox-configs/{id}/skills/{skillId}/transcript": {
"get": {
"security": [
@@ -13362,6 +13426,184 @@
}
}
},
+ "/skills/catalog": {
+ "get": {
+ "security": [
+ {
+ "Bearer": []
+ },
+ {
+ "ApiKeyAuth": []
+ }
+ ],
+ "description": "Returns every skill definition in this workspace and which sandbox configs it is installed on.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "List workspace skills",
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ },
+ "post": {
+ "description": "Records a skill without installing it. Send a zip as multipart field \"file\", or JSON {\"source\":\"...\"}.",
+ "consumes": [
+ "application/json",
+ "multipart/form-data"
+ ],
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Add a skill to the workspace catalog",
+ "responses": {
+ "201": {
+ "description": "Created",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}": {
+ "delete": {
+ "description": "Refused while any sandbox still has an installation of this skill.",
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Delete a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/files": {
+ "get": {
+ "description": "Lists the stored catalog bundle. Files belong to the skill definition, not a sandbox install.",
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "List files of a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/files/content": {
+ "get": {
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Read one file of a catalog skill",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ },
+ {
+ "type": "string",
+ "description": "Skill-root-relative file path",
+ "name": "path",
+ "in": "query",
+ "required": true
+ }
+ ],
+ "responses": {
+ "200": {
+ "description": "OK",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
+ "/skills/catalog/{id}/install": {
+ "post": {
+ "description": "Runs the existing snapshot install onto each named sandbox config.",
+ "consumes": [
+ "application/json"
+ ],
+ "produces": [
+ "application/json"
+ ],
+ "tags": [
+ "Skills"
+ ],
+ "summary": "Install a catalog skill onto sandboxes",
+ "parameters": [
+ {
+ "type": "string",
+ "description": "Catalog skill ID",
+ "name": "id",
+ "in": "path",
+ "required": true
+ }
+ ],
+ "responses": {
+ "202": {
+ "description": "Accepted",
+ "schema": {
+ "type": "object",
+ "additionalProperties": true
+ }
+ }
+ }
+ }
+ },
"/storage-backends": {
"get": {
"security": [
@@ -17572,7 +17814,7 @@
"type": "integer"
},
"parser_engine_rules": {
- "description": "ParserEngineRules configures which parser engine to use for each file type.\nWhen empty, the builtin engine is used for all types.",
+ "description": "ParserEngineRules configures which parser engine to use for each file type.\nWhen empty, DefaultParserEngine is used (builtin/simple routing, except\ntypes that only a specific engine can parse: ppt/pptx fall back to\nmarkitdown). A linked anydoc binding is preferred for every type it\nconverts.",
"type": "array",
"items": {
"$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.ParserEngineRule"
@@ -17730,6 +17972,61 @@
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.CubeEgressRule": {
+ "type": "object",
+ "properties": {
+ "audit": {
+ "description": "Audit is none | metadata | full. Empty uses the server default.",
+ "type": "string"
+ },
+ "deny": {
+ "description": "Deny inverts the action, which defaults to allow. A deny rule still\nneeds Host or SNI: the target has to reach CubeEgress for it to answer\nwith a request-level 403 instead of the network layer dropping it.",
+ "type": "boolean"
+ },
+ "host": {
+ "type": "string"
+ },
+ "inject": {
+ "description": "Inject adds credential headers on allowed HTTPS requests.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.CubeHeaderInject"
+ }
+ },
+ "methods": {
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "name": {
+ "type": "string"
+ },
+ "path": {
+ "type": "string"
+ },
+ "scheme": {
+ "type": "string"
+ },
+ "sni": {
+ "type": "string"
+ }
+ }
+ },
+ "github_com_Tencent_WeKnora_internal_types.CubeHeaderInject": {
+ "type": "object",
+ "properties": {
+ "format": {
+ "type": "string"
+ },
+ "header": {
+ "type": "string"
+ },
+ "secret": {
+ "type": "string"
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.CubeSandboxConfig": {
"type": "object",
"properties": {
@@ -17901,11 +18198,11 @@
"type": "integer"
},
"max_completion_tokens": {
- "description": "Maximum completion tokens (only for normal mode)",
+ "description": "Maximum completion tokens. Quick-answer uses this for the RAG answer.\nSmart-reasoning ReAct rounds send this value as-is (zero becomes\nDefaultMaxCompletionTokens at call time: 4096, or 24576 with a sandbox).",
"type": "integer"
},
"max_iterations": {
- "description": "===== Agent Mode Settings =====\nMaximum ReAct iterations. -1 is unlimited; 0 is unset (default applied).",
+ "description": "===== Agent Mode Settings =====\nMaximum iterations for the ReAct loop. Zero is unset (filled with a\ndefault). A negative value is unlimited: the loop runs until the model\nstops, the user cancels, or another guard fires.",
"type": "integer"
},
"mcp_auth_wait_timeout": {
@@ -18193,6 +18490,21 @@
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.E2BHostRule": {
+ "type": "object",
+ "properties": {
+ "headers": {
+ "description": "Headers values are credentials and are encrypted at rest; names stay\nreadable so operators can see which headers are injected.",
+ "type": "object",
+ "additionalProperties": {
+ "type": "string"
+ }
+ },
+ "host": {
+ "type": "string"
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.E2BSandboxConfig": {
"type": "object",
"properties": {
@@ -19854,6 +20166,10 @@
"base_url": {
"type": "string"
},
+ "context_window": {
+ "description": "ContextWindow is the model's total context window in tokens and\nMaxOutputTokens the most it emits in one response. Both are provider\nfacts the agent cannot discover but has to act on: the context window is\nwhat decides when conversation history gets compacted, and assuming a\nwindow larger than the real one means compaction never fires and the\nprovider rejects the request mid-conversation instead. 0 means unknown,\nwhich falls back to DefaultMaxContextTokens.",
+ "type": "integer"
+ },
"custom_headers": {
"description": "CustomHeaders 允许在调用远程模型 API 时附加自定义 HTTP 请求头,\n用途类似 Python OpenAI SDK 的 extra_headers 参数,\n常见场景包括透传企业网关鉴权信息、追踪 ID、路由标识等。\n保留字段(Authorization、api-key、Content-Type、Accept 等)会在运行期被忽略以避免破坏签名/鉴权流程。",
"type": "object",
@@ -19878,6 +20194,9 @@
"description": "MaxConcurrency caps concurrent in-flight BACKGROUND (ingestion /\nenrichment) calls to THIS specific model, keyed by model ID and shared\nacross all replicas. 0 (the default) means \"fall back to the\nprocess-wide model.max_concurrency\". Interactive user-facing calls are\nnever gated. Only chat / vlm / embedding honour this (see limiter.Gate).",
"type": "integer"
},
+ "max_output_tokens": {
+ "type": "integer"
+ },
"parameter_size": {
"description": "Ollama model parameter size (e.g., \"7B\", \"13B\", \"70B\")",
"type": "string"
@@ -19922,6 +20241,7 @@
"ModelSourceGemini": "Gemini model",
"ModelSourceHunyuan": "Hunyuan model",
"ModelSourceJina": "Jina AI model",
+ "ModelSourceLiteLLM": "LiteLLM proxy model",
"ModelSourceLocal": "Local model",
"ModelSourceMimo": "Mimo model",
"ModelSourceMinimax": "Minimax mode",
@@ -19929,7 +20249,6 @@
"ModelSourceNvidia": "NVIDIA model",
"ModelSourceOpenAI": "OpenAI model",
"ModelSourceOpenRouter": "OpenRouter model",
- "ModelSourceLiteLLM": "LiteLLM proxy model",
"ModelSourceRemote": "Remote model",
"ModelSourceRequesty": "Requesty model",
"ModelSourceSiliconFlow": "SiliconFlow model",
@@ -20864,6 +21183,47 @@
}
}
},
+ "github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy": {
+ "type": "object",
+ "properties": {
+ "allow_out": {
+ "description": "AllowOut accepts IPv4, IPv4 CIDR, a DNS name, or a single-label\nwildcard such as \"*.example.com\". Domain entries are only meaningful\ntogether with a deny-all; see ValidateSandboxNetworkPolicy.",
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "allow_public_inbound": {
+ "description": "AllowPublicInbound exposes the sandbox's public URL without a\ncredential. false requires every inbound request to carry the\nper-sandbox traffic access token.",
+ "type": "boolean"
+ },
+ "cube_rules": {
+ "description": "CubeRules are validated whenever present and consumed only by the Cube\nprovider adapter.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.CubeEgressRule"
+ }
+ },
+ "deny_egress_by_default": {
+ "description": "DenyEgressByDefault installs a 0.0.0.0/0 deny-all, after which only\nAllowOut (and L7 rule targets) can reach the network. false allows\npublic egress, which is what skill installs need.",
+ "type": "boolean"
+ },
+ "deny_out": {
+ "description": "DenyOut accepts IPv4 and IPv4 CIDR only. Neither provider can deny a\ndomain: denial is a pure longest-prefix match on the destination IP.",
+ "type": "array",
+ "items": {
+ "type": "string"
+ }
+ },
+ "e2b_host_rules": {
+ "description": "E2BHostRules are validated whenever present and consumed only by the E2B\nprovider adapter.",
+ "type": "array",
+ "items": {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.E2BHostRule"
+ }
+ }
+ }
+ },
"github_com_Tencent_WeKnora_internal_types.SearchParams": {
"type": "object",
"properties": {
@@ -21699,6 +22059,14 @@
"type": "string"
}
},
+ "network": {
+ "description": "Network is the outbound/inbound network policy applied to every sandbox\ncreated from this config — chat sessions, skill installs and deep\nconnectivity probes alike. nil and the zero value mean the same thing:\noutbound egress allowed, inbound public access closed.",
+ "allOf": [
+ {
+ "$ref": "#/definitions/github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy"
+ }
+ ]
+ },
"sandbox_type": {
"description": "SandboxType is cube, e2b, or docker; disabled is the hidden policy row.",
"type": "string"
@@ -25062,4 +25430,4 @@
"in": "header"
}
}
-}
+}
\ No newline at end of file
diff --git a/docs/swagger.yaml b/docs/swagger.yaml
index 68de7146..4a6543fb 100644
--- a/docs/swagger.yaml
+++ b/docs/swagger.yaml
@@ -552,7 +552,10 @@ definitions:
parser_engine_rules:
description: |-
ParserEngineRules configures which parser engine to use for each file type.
- When empty, the builtin engine is used for all types.
+ When empty, DefaultParserEngine is used (builtin/simple routing, except
+ types that only a specific engine can parse: ppt/pptx fall back to
+ markitdown). A linked anydoc binding is preferred for every type it
+ converts.
items:
$ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.ParserEngineRule'
type: array
@@ -690,6 +693,46 @@ definitions:
weknoracloud:
$ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.WeKnoraCloudCredentials'
type: object
+ github_com_Tencent_WeKnora_internal_types.CubeEgressRule:
+ properties:
+ audit:
+ description: Audit is none | metadata | full. Empty uses the server default.
+ type: string
+ deny:
+ description: |-
+ Deny inverts the action, which defaults to allow. A deny rule still
+ needs Host or SNI: the target has to reach CubeEgress for it to answer
+ with a request-level 403 instead of the network layer dropping it.
+ type: boolean
+ host:
+ type: string
+ inject:
+ description: Inject adds credential headers on allowed HTTPS requests.
+ items:
+ $ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.CubeHeaderInject'
+ type: array
+ methods:
+ items:
+ type: string
+ type: array
+ name:
+ type: string
+ path:
+ type: string
+ scheme:
+ type: string
+ sni:
+ type: string
+ type: object
+ github_com_Tencent_WeKnora_internal_types.CubeHeaderInject:
+ properties:
+ format:
+ type: string
+ header:
+ type: string
+ secret:
+ type: string
+ type: object
github_com_Tencent_WeKnora_internal_types.CubeSandboxConfig:
properties:
api_key:
@@ -871,12 +914,17 @@ definitions:
description: Timeout for a single LLM call in seconds (0 = use global default)
type: integer
max_completion_tokens:
- description: Maximum completion tokens (only for normal mode)
+ description: |-
+ Maximum completion tokens. Quick-answer uses this for the RAG answer.
+ Smart-reasoning ReAct rounds send this value as-is (zero becomes
+ DefaultMaxCompletionTokens at call time: 4096, or 24576 with a sandbox).
type: integer
max_iterations:
description: |-
===== Agent Mode Settings =====
- Maximum ReAct iterations. -1 is unlimited; 0 is unset (default applied).
+ Maximum iterations for the ReAct loop. Zero is unset (filled with a
+ default). A negative value is unlimited: the loop runs until the model
+ stops, the user cancels, or another guard fires.
type: integer
mcp_auth_wait_timeout:
description: |-
@@ -1148,6 +1196,18 @@ definitions:
cert.pem and key.pem. Required when Host is a TCP endpoint.
type: string
type: object
+ github_com_Tencent_WeKnora_internal_types.E2BHostRule:
+ properties:
+ headers:
+ additionalProperties:
+ type: string
+ description: |-
+ Headers values are credentials and are encrypted at rest; names stay
+ readable so operators can see which headers are injected.
+ type: object
+ host:
+ type: string
+ type: object
github_com_Tencent_WeKnora_internal_types.E2BSandboxConfig:
properties:
api_key:
@@ -2418,6 +2478,16 @@ definitions:
type: string
base_url:
type: string
+ context_window:
+ description: |-
+ ContextWindow is the model's total context window in tokens and
+ MaxOutputTokens the most it emits in one response. Both are provider
+ facts the agent cannot discover but has to act on: the context window is
+ what decides when conversation history gets compacted, and assuming a
+ window larger than the real one means compaction never fires and the
+ provider rejects the request mid-conversation instead. 0 means unknown,
+ which falls back to DefaultMaxContextTokens.
+ type: integer
custom_headers:
additionalProperties:
type: string
@@ -2444,6 +2514,8 @@ definitions:
process-wide model.max_concurrency". Interactive user-facing calls are
never gated. Only chat / vlm / embedding honour this (see limiter.Gate).
type: integer
+ max_output_tokens:
+ type: integer
parameter_size:
description: Ollama model parameter size (e.g., "7B", "13B", "70B")
type: string
@@ -2483,6 +2555,7 @@ definitions:
ModelSourceGemini: Gemini model
ModelSourceHunyuan: Hunyuan model
ModelSourceJina: Jina AI model
+ ModelSourceLiteLLM: LiteLLM proxy model
ModelSourceLocal: Local model
ModelSourceMimo: Mimo model
ModelSourceMinimax: Minimax mode
@@ -2490,7 +2563,6 @@ definitions:
ModelSourceNvidia: NVIDIA model
ModelSourceOpenAI: OpenAI model
ModelSourceOpenRouter: OpenRouter model
- ModelSourceLiteLLM: LiteLLM proxy model
ModelSourceRemote: Remote model
ModelSourceRequesty: Requesty model
ModelSourceSiliconFlow: SiliconFlow model
@@ -3185,6 +3257,50 @@ definitions:
use_ssl:
type: boolean
type: object
+ github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy:
+ properties:
+ allow_out:
+ description: |-
+ AllowOut accepts IPv4, IPv4 CIDR, a DNS name, or a single-label
+ wildcard such as "*.example.com". Domain entries are only meaningful
+ together with a deny-all; see ValidateSandboxNetworkPolicy.
+ items:
+ type: string
+ type: array
+ allow_public_inbound:
+ description: |-
+ AllowPublicInbound exposes the sandbox's public URL without a
+ credential. false requires every inbound request to carry the
+ per-sandbox traffic access token.
+ type: boolean
+ cube_rules:
+ description: |-
+ CubeRules are validated whenever present and consumed only by the Cube
+ provider adapter.
+ items:
+ $ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.CubeEgressRule'
+ type: array
+ deny_egress_by_default:
+ description: |-
+ DenyEgressByDefault installs a 0.0.0.0/0 deny-all, after which only
+ AllowOut (and L7 rule targets) can reach the network. false allows
+ public egress, which is what skill installs need.
+ type: boolean
+ deny_out:
+ description: |-
+ DenyOut accepts IPv4 and IPv4 CIDR only. Neither provider can deny a
+ domain: denial is a pure longest-prefix match on the destination IP.
+ items:
+ type: string
+ type: array
+ e2b_host_rules:
+ description: |-
+ E2BHostRules are validated whenever present and consumed only by the E2B
+ provider adapter.
+ items:
+ $ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.E2BHostRule'
+ type: array
+ type: object
github_com_Tencent_WeKnora_internal_types.SearchParams:
properties:
disable_keywords_match:
@@ -3834,6 +3950,14 @@ definitions:
become visible to all scripts running in the tenant's\nsandboxes — do not
place secrets here that scripts must not access."
type: object
+ network:
+ allOf:
+ - $ref: '#/definitions/github_com_Tencent_WeKnora_internal_types.SandboxNetworkPolicy'
+ description: |-
+ Network is the outbound/inbound network policy applied to every sandbox
+ created from this config — chat sessions, skill installs and deep
+ connectivity probes alike. nil and the zero value mean the same thing:
+ outbound egress allowed, inbound public access closed.
sandbox_type:
description: SandboxType is cube, e2b, or docker; disabled is the hidden policy
row.
@@ -14228,6 +14352,50 @@ paths:
summary: Retry a skill install
tags:
- SandboxConfig
+ /sandbox-configs/{id}/skills/{skillId}/stop:
+ post:
+ description: Abort an in-flight install so the operator can retry or uninstall.
+ After a process restart the row may still say installing with no live process;
+ this rewrites it immediately instead of waiting for the stuck-run reaper.
+ Removal is not stopped.
+ parameters:
+ - description: Sandbox config ID
+ in: path
+ name: id
+ required: true
+ type: string
+ - description: Skill ID
+ in: path
+ name: skillId
+ required: true
+ type: string
+ produces:
+ - application/json
+ responses:
+ "200":
+ description: Stopped skill
+ schema:
+ additionalProperties: true
+ type: object
+ "400":
+ description: Skill is not installing
+ schema:
+ $ref: '#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError'
+ "401":
+ description: Unauthorized
+ schema:
+ additionalProperties: true
+ type: object
+ "404":
+ description: Skill not found
+ schema:
+ $ref: '#/definitions/github_com_Tencent_WeKnora_internal_errors.AppError'
+ security:
+ - Bearer: []
+ - ApiKeyAuth: []
+ summary: Stop a skill install
+ tags:
+ - SandboxConfig
/sandbox-configs/{id}/skills/{skillId}/transcript:
get:
description: |-
@@ -14854,6 +15022,126 @@ paths:
summary: 获取当前沙箱配置上可执行的 Skills
tags:
- Skills
+ /skills/catalog:
+ get:
+ description: Returns every skill definition in this workspace and which sandbox
+ configs it is installed on.
+ produces:
+ - application/json
+ responses:
+ "200":
+ description: OK
+ schema:
+ additionalProperties: true
+ type: object
+ security:
+ - Bearer: []
+ - ApiKeyAuth: []
+ summary: List workspace skills
+ tags:
+ - Skills
+ post:
+ consumes:
+ - application/json
+ - multipart/form-data
+ description: Records a skill without installing it. Send a zip as multipart
+ field "file", or JSON {"source":"..."}.
+ produces:
+ - application/json
+ responses:
+ "201":
+ description: Created
+ schema:
+ additionalProperties: true
+ type: object
+ summary: Add a skill to the workspace catalog
+ tags:
+ - Skills
+ /skills/catalog/{id}:
+ delete:
+ description: Refused while any sandbox still has an installation of this skill.
+ parameters:
+ - description: Catalog skill ID
+ in: path
+ name: id
+ required: true
+ type: string
+ responses:
+ "200":
+ description: OK
+ schema:
+ additionalProperties: true
+ type: object
+ summary: Delete a catalog skill
+ tags:
+ - Skills
+ /skills/catalog/{id}/files:
+ get:
+ description: Lists the stored catalog bundle. Files belong to the skill definition,
+ not a sandbox install.
+ parameters:
+ - description: Catalog skill ID
+ in: path
+ name: id
+ required: true
+ type: string
+ produces:
+ - application/json
+ responses:
+ "200":
+ description: OK
+ schema:
+ additionalProperties: true
+ type: object
+ summary: List files of a catalog skill
+ tags:
+ - Skills
+ /skills/catalog/{id}/files/content:
+ get:
+ parameters:
+ - description: Catalog skill ID
+ in: path
+ name: id
+ required: true
+ type: string
+ - description: Skill-root-relative file path
+ in: query
+ name: path
+ required: true
+ type: string
+ produces:
+ - application/json
+ responses:
+ "200":
+ description: OK
+ schema:
+ additionalProperties: true
+ type: object
+ summary: Read one file of a catalog skill
+ tags:
+ - Skills
+ /skills/catalog/{id}/install:
+ post:
+ consumes:
+ - application/json
+ description: Runs the existing snapshot install onto each named sandbox config.
+ parameters:
+ - description: Catalog skill ID
+ in: path
+ name: id
+ required: true
+ type: string
+ produces:
+ - application/json
+ responses:
+ "202":
+ description: Accepted
+ schema:
+ additionalProperties: true
+ type: object
+ summary: Install a catalog skill onto sandboxes
+ tags:
+ - Skills
/storage-backends:
get:
description: List all storage backend instances for the current workspace, with
diff --git a/frontend/package-lock.json b/frontend/package-lock.json
index 0a8b3ce7..4c467ea7 100644
--- a/frontend/package-lock.json
+++ b/frontend/package-lock.json
@@ -1,12 +1,12 @@
{
"name": "knowledage-base",
- "version": "0.7.2",
+ "version": "0.8.0",
"lockfileVersion": 3,
"requires": true,
"packages": {
"": {
"name": "knowledage-base",
- "version": "0.7.2",
+ "version": "0.8.0",
"dependencies": {
"@microsoft/fetch-event-source": "^2.0.1",
"@types/dompurify": "^3.2.0",
diff --git a/frontend/package.json b/frontend/package.json
index 8d030668..954eba56 100644
--- a/frontend/package.json
+++ b/frontend/package.json
@@ -1,6 +1,6 @@
{
"name": "knowledage-base",
- "version": "0.7.2",
+ "version": "0.8.0",
"private": true,
"type": "module",
"scripts": {
diff --git a/helm/Chart.yaml b/helm/Chart.yaml
index b549b244..c7df804d 100644
--- a/helm/Chart.yaml
+++ b/helm/Chart.yaml
@@ -5,7 +5,7 @@ description: |
with document parsing, vector search, and LLM integration.
type: application
version: 0.1.0
-appVersion: "v0.7.2"
+appVersion: "v0.8.0"
kubeVersion: ">=1.25.0-0"
home: https://github.com/Tencent/WeKnora
icon: https://raw.githubusercontent.com/Tencent/WeKnora/main/docs/images/logo.png
diff --git a/packages/dsh-weknora/README.md b/packages/dsh-weknora/README.md
index 5b556bff..aca9f5a9 100644
--- a/packages/dsh-weknora/README.md
+++ b/packages/dsh-weknora/README.md
@@ -167,7 +167,7 @@ types still accept and serve them — so a rename on either side fails CI instea
## Compatibility
-Verified against dsh `0.1.0-rc.8` and WeKnora `0.7.2`. dsh is in developer preview and states that breaking changes
+Verified against dsh `0.1.0-rc.8` and WeKnora `0.8.0`. dsh is in developer preview and states that breaking changes
will happen; this package deliberately has **no runtime dependencies** and hands `ctx.tools.register()` a plain object,
so it does not pin any harness package version. If a harness release changes the tool-definition contract, open an issue
against this repository.
diff --git a/packages/dsh-weknora/README_CN.md b/packages/dsh-weknora/README_CN.md
index 6875b18e..d3c63251 100644
--- a/packages/dsh-weknora/README_CN.md
+++ b/packages/dsh-weknora/README_CN.md
@@ -155,7 +155,7 @@ profile,而它写出的 profile 把设置放在 `pnpm-workspace.yaml` 里。
## 兼容性
-已针对 dsh `0.1.0-rc.8` 与 WeKnora `0.7.2` 验证。dsh 处于 developer preview 并明确会有破坏性变更;本包因此**没有任何运行
+已针对 dsh `0.1.0-rc.8` 与 WeKnora `0.8.0` 验证。dsh 处于 developer preview 并明确会有破坏性变更;本包因此**没有任何运行
时依赖**,交给 `ctx.tools.register()` 的只是一个普通对象,不锁定任何 harness 包版本。如果 harness 的工具定义契约发生变化,
请在本仓库提 issue。
diff --git a/website-docs/01-getting-started/01-introduction.md b/website-docs/01-getting-started/01-introduction.md
index d0fca60c..896b5ba6 100644
--- a/website-docs/01-getting-started/01-introduction.md
+++ b/website-docs/01-getting-started/01-introduction.md
@@ -47,7 +47,7 @@ WeKnora(维娜拉)是腾讯开源的知识库问答系统,做的事情是
| FAQ | FAQ 型知识库中的问答对,存于 Chunk 的 Metadata:标准问 `StandardQuestion`、相似问、反例问、多答案与答案策略 |
| Wiki 页面 WikiPage | Wiki 型索引产物:由 LLM 从文档生成的结构化百科页面,最多三级分类路径,可被 Agent 以 `wiki_search` / `wiki_read_page` 工具导航 |
| 知识图谱 Entity / Relationship | 从分块中抽取的实体与关系(强度 1-10),存储在 Neo4j(`NEO4J_ENABLE=true` 时),用于 GraphRAG 增强检索 |
-| 数据源 DataSource | 外部内容连接器。**当前可用 5 个**:`feishu`、`lark`(与飞书同一适配器,域名不同)、`notion`、`yuque`、`rss`,支持 Cron 定时同步(增量/全量)与冲突策略。`internal/types/datasource.go` 里还声明了 `confluence`、`github`、`imap` 等类型常量,但对应实现尚未接入(`initConnectorRegistry()` 中相关注册被注释),选不到 |
+| 数据源 DataSource | 外部内容连接器。**当前可用 7 个**:`feishu`、`lark`(与飞书同一适配器,域名不同)、`gitlab`、`ima`(腾讯 IMA 笔记)、`notion`、`yuque`、`rss`,支持 Cron 定时同步(增量/全量)与冲突策略。`internal/types/datasource.go` 里还声明了 `confluence`、`github`、`imap` 等类型常量,但对应实现尚未接入(`initConnectorRegistry()` 中相关注册被注释),选不到 |
| 检索配置 RetrievalConfig | 租户级检索参数:`EmbeddingTopK`(默认 50)、`VectorThreshold`(0.15)、`KeywordThreshold`(0.3)、`RerankTopK`(10)、`RerankThreshold`(0.2)、RRF 融合参数(`RRFK`=60,向量权重 0.7 / 关键词权重 0.3) |
### 对话与智能体
diff --git a/website-docs/01-getting-started/02-installation.md b/website-docs/01-getting-started/02-installation.md
index c87f03f2..42c8c68f 100644
--- a/website-docs/01-getting-started/02-installation.md
+++ b/website-docs/01-getting-started/02-installation.md
@@ -175,7 +175,7 @@ make docker-build-frontend
## 六、Helm 部署(helm/)
-`helm/Chart.yaml`:apiVersion v2,chart 名 `weknora`,appVersion 跟随版本(如 v0.7.2),要求 Kubernetes >= 1.25.0。
+`helm/Chart.yaml`:apiVersion v2,chart 名 `weknora`,appVersion 跟随版本(如 v0.8.0),要求 Kubernetes >= 1.25.0。
Chart 内包含五个组件:`app`(`wechatopenai/weknora-app`)、`frontend`(`wechatopenai/weknora-ui`)、`docreader`、`postgresql`(ParadeDB 镜像)、`redis`(`redis:7-alpine`),并可选启用 `minio` 与 `neo4j`。
diff --git a/website-docs/03-features/03-document-parsing.md b/website-docs/03-features/03-document-parsing.md
index 69ea1e4f..f105b923 100644
--- a/website-docs/03-features/03-document-parsing.md
+++ b/website-docs/03-features/03-document-parsing.md
@@ -6,7 +6,7 @@
| 类别 | 格式 |
| --- | --- |
-| 文档 | PDF、Word(doc/docx)、PPT(ppt/pptx)、Excel(xls/xlsx)、EPUB |
+| 文档 | PDF、Word(doc/docx)、PPT(ppt/pptx)、Excel(xls/xlsx)、EPUB、XMind |
| 文本 | txt、Markdown、CSV、JSON |
| 网页 | 在线 URL 抓取、本地 HTML / MHTML 归档 |
| 图片 | jpg、png、gif、bmp、tiff、webp(需配置视觉模型才能理解内容) |
diff --git a/website-docs/03-features/07-agent.md b/website-docs/03-features/07-agent.md
index d0bc4f3d..e59f8a2a 100644
--- a/website-docs/03-features/07-agent.md
+++ b/website-docs/03-features/07-agent.md
@@ -32,10 +32,10 @@ WeKnora 提供两种模式,在对话框顶部切换:
| `AgentEngine` | `internal/agent/engine.go` | ReAct 主循环的驱动者,持有配置、工具注册表、Chat 模型、事件总线等 |
| `ToolRegistry` | `internal/agent/tools/registry.go` | 工具注册、查找、参数校验、执行、输出截断、资源清理 |
| 内置工具集 | `internal/agent/tools/*.go` | 24 个内置工具 + 动态注册的 MCP 工具 |
-| Token 估算与压缩 | `internal/agent/token/` | `Estimator`(BPE 估算)与 `CompressContext`(滑动裁剪) |
-| 记忆整合 | `internal/agent/memory/consolidator.go` | LLM 驱动的历史摘要(Memory Consolidation) |
+| Token 估算与压缩 | `internal/agent/token/` + `internal/agent/compaction/` | `Estimator`(BPE 估算)与长轮次上下文压缩(sandbox 工具历史) |
+| 记忆整合 | `internal/application/service/memory/` | 跨会话长期记忆:抽取、召回、主题提升、文档亲和度、整理 |
| 技能系统 | `internal/agent/skills/` | SKILL.md 的发现、加载与脚本执行(Progressive Disclosure) |
-| 执行沙箱 | `internal/sandbox/` | 技能脚本的 Docker / Cube / E2B 隔离执行与安全校验 |
+| 执行沙箱 | `internal/sandbox/` | 技能脚本与 `shell_exec` 的 Docker / Cube / E2B 会话级隔离执行与安全校验 |
| 工具审批 | `internal/agent/approval/gate.go` | MCP 危险工具的人工审批(HITL)与会话内 OAuth 授权 |
| Agent 服务层 | `internal/application/service/agent_service.go` | 组装引擎:注册工具、解析 KB 元信息、初始化技能/沙箱/VLM |
| 会话问答入口 | `internal/application/service/session_agent_qa.go` | 从 `CustomAgent` 构建运行时 `AgentConfig` 并执行 |
@@ -425,13 +425,7 @@ Agent 侧的启停在 `configureSkillsFromAgent`(`internal/application/service
3. **stdin**:内嵌 shell 命令检测;
4. 合并入口 `ValidateAll`。
-**Docker 沙箱**(`docker.go`,`docker run --rm` 隔离):
-
-- `--user 1000:1000` 非 root、`--cap-drop ALL`、`--security-opt no-new-privileges`、`--pids-limit 100`;
-- 默认 `--network none`(除非 `AllowNetwork`);
-- 资源限额:内存默认 `DefaultMemoryLimit = 256MB`(`--memory` + `--memory-swap` 同值禁 swap)、CPU 默认 `DefaultCPULimit = 1.0` 核;
-- 技能目录以只读挂载到 `/workspace`;可选 `--read-only` 根文件系统 + 64MB noexec tmpfs;
-- 按扩展名选择解释器(`.py`→`python3` 等)。
+**Docker 沙箱**(会话级长驻容器,`internal/sandbox/docker_engine.go` / `docker_remote_client.go`):一个会话一个容器,PID 1 为 `sleep infinity`,脚本、`shell_exec`、附件暂存与产物收集都在同一容器里 exec。默认**关闭**(`WEKNORA_SANDBOX_DOCKER_ENABLED` 或系统设置「网络安全」打开)。exec 一律以沙箱账号 `user`(uid 1000) 运行;超时由容器内 `timeout(1)` 执行,空闲回收读活跃标记 mtime。详细能力、网络策略与安全边界见 [`docs/sandbox-docker-backend.md`](../../docs/sandbox-docker-backend.md)。旧的 `docker run --rm` + 只读 bind mount 模型已移除。
Manager 初始化时:`disabled` 模式的 `disabledSandbox` 拒绝一切执行。
diff --git a/website-docs/04-api/01-api-overview.md b/website-docs/04-api/01-api-overview.md
index c0387b26..bfe4d8c1 100644
--- a/website-docs/04-api/01-api-overview.md
+++ b/website-docs/04-api/01-api-overview.md
@@ -224,5 +224,5 @@ X-Accel-Buffering: no
| 模型与初始化 | [02-api-model-system.md](./02-api-model-system.md) | `/models`、`/initialization`、`/evaluation`、`/weknoracloud` |
| 系统与平台管理 | [02-api-system.md](./02-api-system.md) | `/system`、`/system/admin` |
| 基础设施与数据源 | [02-api-infra.md](./02-api-infra.md) | `/vector-stores`、`/storage-backends`、`/web-search-providers`、`/datasource` |
-| Agent、MCP 与技能 | [02-api-agent-mcp.md](./02-api-agent-mcp.md) | `/agents`、`/mcp-services`、`/agent`、`/skills`、`/user/favorites` |
+| Agent、MCP、技能与记忆 | [02-api-agent-mcp.md](./02-api-agent-mcp.md) | `/agents`、`/mcp-services`、`/agent`、`/skills`、`/sandbox-configs`、`/memory`、`/user/favorites` |
| IM、Embed 与文件服务 | [02-api-channels.md](./02-api-channels.md) | `/im`、`/im-channels`、`/wechat`、`/embed-channels`、`/embed`、`/files`、`/r/:token` |
diff --git a/website-docs/04-api/02-api-agent-mcp.md b/website-docs/04-api/02-api-agent-mcp.md
index 447ec4b6..5e4dc190 100644
--- a/website-docs/04-api/02-api-agent-mcp.md
+++ b/website-docs/04-api/02-api-agent-mcp.md
@@ -346,6 +346,24 @@ curl -X POST $BASE/api/v1/agent/mcp-oauth-resolutions/p-1/cancel -H "Authorizati
curl $BASE/api/v1/skills -H "Authorization: Bearer $TOKEN"
```
+安装、停用、文件浏览与环境变量走沙箱配置前缀 `/api/v1/sandbox-configs/{id}/skills*` 以及 `/api/v1/me/env-vars*`,完整示例见仓库 [`docs/api/skill.md`](../../docs/api/skill.md)。
+
+## 长期记忆(/api/v1/memory)
+
+跨会话记忆始终绑定当前调用者,路径里没有 subject id。权限:Viewer+;API Key 必须 full-access。Handler: `internal/handler/memory.go`
+
+| 方法 | 路径 | 用途 |
+| --- | --- | --- |
+| GET/PUT | `/memory/settings` | 个人开关(空间级开关由管理员在租户设置里改) |
+| GET/POST/DELETE | `/memory/items` | 列出 / 手动新增 / 清空 |
+| PUT/DELETE | `/memory/items/{id}` | 修改 / 删除 |
+| POST | `/memory/items/{id}/confirm` `/reject` | 确认或否决推断出的记忆 |
+| GET | `/memory/topics` `/memory/documents` | 主题计数、文档亲和度 |
+| GET | `/memory/export` | JSON 导出 |
+| POST | `/memory/consolidate` | 立刻整理 |
+
+完整字段与 curl 见仓库 [`docs/api/memory.md`](../../docs/api/memory.md)。
+
## 用户收藏(/api/v1/user/favorites)
按用户维度存储(非资源创建者维度);权限均 Viewer+,仅 JWT(API key 默认拒绝)。Handler: `internal/handler/user_resource_favorite.go`
diff --git a/website-docs/05-clients/01-frontend.md b/website-docs/05-clients/01-frontend.md
index fd8095e9..f9f57803 100644
--- a/website-docs/05-clients/01-frontend.md
+++ b/website-docs/05-clients/01-frontend.md
@@ -8,7 +8,7 @@ WeKnora 的 Web 前端是一个基于 **Vue 3 + TypeScript + Vite** 的单页应
## 技术栈总览
-依据 `frontend/package.json`(版本 0.7.2):
+依据 `frontend/package.json`(版本 0.8.0):
| 类别 | 选型 | 版本 | 说明 |
| --- | --- | --- | --- |
diff --git a/website-docs/06-development/01-dev-guide.md b/website-docs/06-development/01-dev-guide.md
index f56eddcf..f4793462 100644
--- a/website-docs/06-development/01-dev-guide.md
+++ b/website-docs/06-development/01-dev-guide.md
@@ -10,7 +10,7 @@ WeKnora 由三个可独立开发的进程组成:
| --- | --- | --- | --- |
| 主后端 `app` | `cmd/server` + `internal/` | Go | **Go 1.26.0**(`go.mod` 中 `go 1.26.0`),需 CGO(DuckDB、sqlite-vec 绑定) |
| 文档解析服务 `docreader` | `docreader/` | Python + gRPC | **Python >= 3.10.18**(`docreader/pyproject.toml` 中 `requires-python`),依赖用 **uv** 管理(仓库含 `uv.lock`,Docker 内 `uv sync --locked`) |
-| 前端 `frontend` | `frontend/` | Node.js + Vue 3 | Node 22 系(`devDependencies` 含 `@tsconfig/node22`、`@types/node ^22`),Vite 7 + TypeScript ~6.0 + Vue 3.5 + TDesign,版本号 `0.7.2` |
+| 前端 `frontend` | `frontend/` | Node.js + Vue 3 | Node 22 系(`devDependencies` 含 `@tsconfig/node22`、`@types/node ^22`),Vite 7 + TypeScript ~6.0 + Vue 3.5 + TDesign,版本号 `0.8.0` |
| CLI | `cli/`(独立 Go module) | Go | Go 1.26(`.github/workflows/cli.yml` 矩阵 `go: ['1.26']`) |
推荐额外安装的开发工具: