Capability packs

Document version: 1.3

Source docs/en/site/mech-capability-pack.md

Document version: 1.3 Status: Matches current backend (capability, workspace_capabilities API, conversation tool gating) Voice: For product, operations, and admins — what workspace “assistant capability packs” are and how they map to assistant tools. Implementation names: Implementation mapping.

Related:

产品规格.md §3.5 (runtime tools), §3.6 (MCP and extensions), §4.2.4 (platform plugins)
技能组成规范.md (skill permissions.tools vs capability packs)
机制索引.md
User-side notes: help/product-features/data-integration.md (data connections and predefined queries)

1. Up front

An assistant capability pack is a switch workspace admins turn on for members: tool-class capabilities the assistant may call in Messages. It is not equal to one tool function: one pack may map to several tools, or only extend how existing tools behave.

Cadau uses three layers:

Layer	User language	Duty
Capability pack	Assistant capability pack	Turn on, validity window, audit and billing grain
Tool	(not user-visible)	Function-call primitives exposed to the model
Implementation	(not user-visible)	Built-in module / MCP plugin / business adapter

Hard rules:

New capabilities prefer hanging on existing tools; add a new tool only when the governance boundary is independent.
Pipeline capabilities (vision, knowledge-document injection, skill recall, memory retrieval) do not enter capability packs by default, unless they need a separate turn-on or a compliance gate.

2. Existing capability packs

2.1 `basic_documents` — basic document tools

Dimension	Convention
User value	Members can have the assistant read/organize Markdown, plain text, and similar, and export downloadable files
Default	New workspaces on by default
Mapped tools	`file_read` + `file_write`
Mapped implementation	Built-in `agentfile` + `attachmenttext` (text types)
Write formats	`.md` `.txt` `.csv` `.json`; one file ≤ 512 KiB
Off / expired	Tools do not appear in the schema; the assistant should explain the limit

2.2 `office_documents` — office document tools

Dimension	Convention
User value	The assistant can read PDF, Word, Excel, PPT attachments, create or edit .docx/.xlsx/.pptx, and generate laid-out .pdf
Default	Off by default
Mapped tools	No new tool name (extends `file_read`); `office_document` (read/write Office; `build_pdf` generates PDF)
Mapped implementation	Read: `attachmenttext` + OfficeCLI `view` (Excel pages by row so whole-sheet JSON is not cut in half); write/edit: OfficeCLI (`create`/`merge`/`set`/`add` and similar); PDF generate: built-in `gopdf` (`build_pdf`, not via OfficeCLI)
Prerequisite	`basic_documents` must be in force at the same time
Platform switch	`AGENT_OFFICECLI_TOOL_ENABLED`; the server must have OfficeCLI installed (`MINDLINK_OFFICECLI` or PATH)
When off	`file_read` returns a clear error for office extensions; `office_document` does not appear in the schema

Turning off basic document tools automatically cascades office document read off.

2.3 `http_integration` — network requests

Dimension	Convention
User value	The assistant can call business APIs as skill notes describe
Default	Off by default
Mapped tools	`http_request`
Mapped implementation	Built-in `agenthttp` (site policy, SSRF protection)
Dual gate	Workspace capability pack and server `AGENT_HTTP_TOOL_ENABLED`
Site policy	Workspace allowlist / platform filed list / public internet (needs platform permission)
Member override	Admins may set network scope per member (inherit or tighten)
Collaboration	Workspace skills, `permissions.tools: [http_request]`

Do not split a tool per business API; business path and auth are described in the skill operations body.

2.4 `code_execution` — script execution

Dimension	Convention
User value	The assistant can run Python under skill `scripts/`; when built-in tools are not enough, after admin allow, install libraries, write code, and deposit a skill for reuse
Default	Off by default
Mapped tools	`run_script`, `skill_script_write`
Mapped implementation	Built-in `agentscript` (sandbox, timeout, outputs registered as uploads)
Prerequisite	`basic_documents` must be in force at the same time; editing existing Office attachments should also have `office_documents` on
Member policy	`workspace_members.script_policy_json`: inherit / forbid / grant separately (skill slug, temp scripts, and similar)
Platform switch	`AGENT_SCRIPT_TOOL_ENABLED` / `agent_script_tool.enabled`
Collaboration	Workspace skills, `permissions.tools: [run_script]` or body `## 脚本执行`
Install libraries	Libraries the platform already allows can install directly. Others: the assistant must say “library name / what it can do / which current problem it solves”: the admin allows on the spot, or a member applies to an admin. After approve, write into this workspace’s allowlist. In-script `pip install` is forbidden. Do not generate `.xlsx` / `.docx` / `.pptx` / `.pdf` with Python (use `office_document`).
Reuse	After a successful run, deposit a workspace skill (`scripts/` + dependencies); later similar conversations prefer running the existing script

2.5 `data_integration` — data connections (data integration)

Dimension	Convention
User value	Members look up MySQL, PostgreSQL, SQL Server and similar business DBs already authorized in the workspace; admins configure connections, predefined queries, and controlled mutations
Default	Off by default
Mapped tools	`data_source_invoke` (read-only + controlled mutation share one tool name)
Mapped implementation	Built-in `datasource` + `store` (encrypted connections, data resources, member grants, mutation approval)
Read-only actions	`ping`, `schema.tables`, `schema.columns`, `table.preview`, `query.list`, `query.run`
Mutation actions	`mutation.list`, `mutation.submit`, `mutation.approve` / `mutation.reject`
SQL constraints	No ad-hoc SQL; read-only uses predefined queries or table preview; writes use mutation definitions
Table preview rows	Default 200 rows, hard cap 500 per call (`mindlink.json` → `data_source_tool.default_preview_max_rows` / `absolute_preview_max_rows`)
Member policy	Data resources + who can use: tick readable tables/queries and CUD rights per member
Dual gate	Workspace capability pack and server `AGENT_DATA_SOURCE_TOOL_ENABLED` / `data_source_tool.enabled`
Host allowlist	Platform `data_source_tool.allow_hosts` (empty means no limit in development)
Collaboration	Workspace skills, `permissions.tools: [data_source_invoke]`; after saving a connection, generate a skill to sync query notes
Binary/photos	Image BLOBs in query results land as attachments, returning `upload_id` / `download_url`; conversation shows `mindlink://upload/{id}`, can embed in Office
Off / expired	Tool does not appear in the schema; the assistant should explain the limit

Admin sub-features (user-side entry is the whole Data connections page, same shell as the knowledge base): Data connections, Data resources, Who can use, Mutation definitions, Mutation approval.

2.6 `data_visualization` — data charts

Dimension	Convention
User value	The assistant generates bar/line/pie PNG charts in conversation, downloadable or embeddable in Office
Default	Off by default
Mapped tools	`chart`
Mapped implementation	Built-in `agentchart` (matplotlib render → register as upload)
Prerequisite	`basic_documents` must be in force at the same time (`DataVisualizationActive`)
Platform switch	No separate platform section; depends on workspace turn-on and a Python/matplotlib runtime (Docker image already includes Python 3)
Collaboration	Workspace skills, `permissions.tools: [chart]`; embedding in PPT/Word uses `office_document` + `image_upload_id`
Numbers vs charts	When the workspace also has data integration on: business-report numbers must come from `data_source_invoke`; user-attached images default to layout/leader-line reference only; copying numbers off the screenshot is forbidden (except the user clearly asks to “digitize from the screenshot”)
When off	`chart` does not appear in the schema

Turning off basic document tools automatically cascades data charts off.

2.7 `wecom_integration` — WeCom

Dimension	Convention
User value	Service notices and scheduled tasks can push to personal WeCom; the assistant can operate contacts, app messages, to-dos, calendar, meetings, docs, smart sheets, WeDrive, mail, and read-only query approvals and leave
Default	Off by default
Mapped tools	`wecom_list_methods`, `wecom_invoke`
Mapped implementation	Built-in `wecom` (self-built app CorpID + AgentID + Secret, stored encrypted)
Dual gate	Workspace capability pack and server `AGENT_WECOM_TOOL_ENABLED` / `agent_wecom_tool.enabled`
Push	In-site notices also send an app card; after a scheduled auto-run succeeds, an extra summary may be pushed (task switch `notify_wecom`, on by default when the member is bound)
Not done	Conversation archive (colleague DMs), submitting approvals on someone’s behalf, changing leave balances, DingTalk/Feishu/email channels, changing the WeCom “app admin” console
Approvals/leave	In WeCom App admin → Approvals detail, tap “Apps that can call APIs” → “Settings” to authorize this app (do not tap the API next to the title); query window at most 31 days
Off / expired	Tools do not appear in the schema; conversation does not treat WeCom as an optional push channel

3. Capability-pack relations

flowchart LR
  subgraph packs [Workspace capability packs]
    B[basic_documents]
    O[office_documents]
    H[http_integration]
    CE[code_execution]
    DI[data_integration]
    DV[data_visualization]
    WC[wecom_integration]
  end
  subgraph tools [Model tools]
    FR[file_read]
    FW[file_write]
    OD[office_document]
    HTTP[http_request]
    RS[run_script]
    DS[data_source_invoke]
    CH[chart]
    W1[wecom_list_methods]
    W2[wecom_invoke]
  end
  B --> FR
  B --> FW
  O --> FR
  O --> OD
  H --> HTTP
  CE --> RS
  B --> CE
  DI --> DS
  B --> DV
  DV --> CH
  WC --> W1
  WC --> W2

Capability pack	Prerequisite
`basic_documents`	None
`office_documents`	`basic_documents` (in force together)
`http_integration`	None (strongly related to skills/site config)
`code_execution`	`basic_documents` (in force together)
`data_integration`	None (strongly related to data-connection / data-resource config)
`data_visualization`	`basic_documents` (in force together)
`wecom_integration`	None (must configure self-built app credentials and bind members to push/operate)

4. Baseline capabilities that are not capability packs

User agents in Messages have these by default (except help-assistant mode):

Capability	Type	Note
Structured memory	Tool	`memory_search` / `memory_get`
Memos	Tool	`memo_save` / `memo_search` and similar; not auto-injected; called only when the user clearly wants to look up/save
Multimodal vision	Pipeline	Message `image_url` injection, not a function call
Knowledge base	Pipeline	Retrieve and inject by agent/workspace config
Workspace skill recall	Pipeline	Inject operations material after match
Conversation-deposited skills	Pipeline + API	Triggered by user wording; product lives in the Skills center

Capability packs mainly cover behavior that may go outbound or needs a separate compliance approval.

5. MCP and plugin access rules

Scene	Suggestion
Replace an HTTP backend	Adapt MCP to the same `http_request` schema; the workspace still uses `http_integration`
Replace Office parse	Hang on the `file_read` implementation chain; the workspace still uses `office_documents`
New business API	Prefer skill + `http_request`, no new tool
Must add a new tool	Also add a capability pack, audit, and `permissions.tools` declaration
Platform plugin modules (§4.2.4)	Usually app-desktop iframe business apps (see 平台插件SDK.md ); need not 1:1 with conversation-tool capability packs. Pure conversation API capabilities do not register `platform_plugins`; follow this table “new business API → skill + `http_request`”

6. Phase-2 candidates (pre-registered)

Candidate ID	User language	Suggested mapping	Why turn on separately
`wiki_knowledge`	Structured knowledge base	`wiki_search` + `wiki_get`	Separate from free memory, needs provenance
`enterprise_search`	Enterprise search	Reuse `http_request` or a dedicated search tool	Split only when schema differs a lot from HTTP

7. Admin UI convention

Item	Copy
Section title	Assistant capability packs
Admin note	Turn on assistant tool capabilities for this workspace; set validity and network sites
Member note	See what is on; apply when off or expired
Technical master switches	`AGENT_HTTP_TOOL_ENABLED`, `AGENT_DATA_SOURCE_TOOL_ENABLED`, `AGENT_SCRIPT_TOOL_ENABLED`, `AGENT_WECOM_TOOL_ENABLED`, and similar live in admin docs, not on members’ first screen

8. Implementation mapping

User concept	Implementation
Capability-pack JSON	`workspaces.capabilities_json`
Parse and dependencies	`capability.ParseWorkspace`, `ValidateWorkspaceDependencies`, `NormalizeWorkspaceDependencies`
Office runtime	`Workspace.OfficeDocumentsActive`
Tool gating	`chat.go` → `fileToolContext`, `dataSourceToolContext`, `workspaceCapsForAgent`; `agenttool.RunAgentToolLoop`
Member script policy	`script_policy_json`, `ResolveEffectiveScriptPolicy`, `GET/PATCH .../script-policy`
Duty grants	`workspace_duties` / `workspace_duty_pack_grants`; once a sensitive pack has duty grants, ordinary members without a grant no longer get it by default (admins excepted; personal “disable” still wins)
Data connections and grants	`workspace_data_sources`, `workspace_data_datasets`, `workspace_member_data_source_grants`; `datasource.Invoke`, `store.ResolveUserDataSourceAccess`
Controlled mutations	`workspace_data_mutations`, `data_mutation_requests`; `datasource.PreviewMutation` / `ExecuteMutation`
Data charts	`agentchart.Render`; gate `wsCaps.DataVisualizationActive`
WeCom	`wecom.Invoke` / `wecom.AfterNotificationInserted`; gate `wecom_integration` + `AGENT_WECOM_TOOL_ENABLED`
Admin API	`GET/PATCH /api/v1/workspaces/:id/capabilities`; `.../data-sources`, data-resource and mutation REST
Capability applications	`workspace_capability_request` handler
Front end	`WorkspaceCapabilitiesPanel.tsx`, `WorkspaceDataIntegrationWorkspace.tsx`, `DataIntegrationPanel.tsx`, `capabilityPackDefs.ts`

9. Decision checklist when extending a pack

Does an admin need to turn it on separately? No → pipeline capability.
Can it reuse an existing tool? Yes → extend the implementation only.
Is it high-risk / outbound? Yes → independent capability pack + audit.
Is it only one business API? Yes → skill + http_request.
After MCP is connected, does the user-visible tool name stay the same? Prefer yes.

1. Up front

2. Existing capability packs

2.1 basic_documents — basic document tools

2.2 office_documents — office document tools

2.3 http_integration — network requests

2.4 code_execution — script execution

2.5 data_integration — data connections (data integration)

2.6 data_visualization — data charts

2.7 wecom_integration — WeCom