docyrus-agent-design

v2026.09.24
docyrus/agent-skills00

Design, build, validate, and test a custom Docyrus AI agent end-to-end using the `docyrus agent` CLI commands and its sub-resources. Use when the user wants to create or configure a custom AI agent (assistant/skill/workflow agent) in a dev app — set its instructions, pick its models, toggle capabilities (tools, data sources, files, knowledge base, web search), attach data sources, attach tools, add knowledge docs, wire MCP servers, connect agents (forward/subagent), add dynamic contexts, build workflow steps, or create deployments. Covers choosing the agent shape, the required vs optional config, the sub-resource attach order, and proving the agent is well-formed. Triggers on "create an AI agent", "build an assistant", "configure an agent", "give the agent a data source / tool / knowledge", "agent workflow steps", "agent deployment", "connect agents", `docyrus agent create`, `docyrus agent tools`, `docyrus agent data-sources`, `docyrus agent workflow-steps`, or any custom-agent design + validation task. For authoring the tools an agent calls, use docyrus-app-ai-tools (define with `apps ai-tools`, then attach here with `agent tools`).

GitHub
安装命令
npx skhub add docyrus/docyrus-agent-design
Markdown
SKILL.md

Docyrus Agent Design

Build a custom AI agent with docyrus agent, then validate its configuration and test that it's well-formed. An agent is a row in tenant_ai_agent (identity + model + capability config) plus a set of sub-resources (tools, data sources, docs, MCPs, connections, dynamic contexts, workflow steps, deployments). This skill is the design workflow; the platform's conceptual capability catalog lives in docyrus-platform (references/ai-capabilities.md), and exhaustive CLI flags are available via docyrus agent <sub> --help (command index: docyrus-cli-app).

Workflow

  1. Confirm app + auth. Agents are dev-app scoped.

    docyrus auth who --json
docyrus apps list --json          # target appSlug / appId

  2. Design the agent. Decide its shape (is it a conversational assistant, a callable skill, or a workflow agent?), its instructions, its default/backup models, and which capabilities it needs (tools? data sources? files? knowledge base? web search?). Toggle a capability on before attaching the matching sub-resource. See references/agent-model.md to map intent → fields.

  3. Create the agent. The only required field is --skillName. Set identity + model + capability toggles in the same call. See Create.

  4. Attach sub-resources in this order (each is its own command group, see Sub-resources):

    • data-sources the agent can read/write (turn on --supportDataSources first),
    • tools the agent can call (turn on --supportTools; author the tool first with docyrus-app-ai-tools, then attach it here),
    • docs (knowledge; turn on --supportKnowledgeBase), mcps, connections (forward/subagent), dynamic-contexts,
    • workflow-steps (only if --hasWorkflow), then deployments (+ deployment-tools / deployment-data-sources).

  5. Validate — re-read the agent and each sub-resource list; confirm config + attachments landed. See Validate.

  6. Test — confirm the agent is well-formed (lists cleanly, references resolve) and, where possible, exercise it. See Test.

A full worked example (assistant with a data source, an attached tool, and a knowledge doc) is in references/workflow-examples.md.

Command cheat-sheet

Selectors: --appId | --appSlug, --agentId. Sub-resource row id is the generic --id (not --toolId/--dataSourceId). Deployment-nested groups also need --deploymentId. Write commands take camelCase flags or --data/--from-file JSON (flags merge over JSON; API keys are snake_case). Append --json.

Create an agent

docyrus agent create --appSlug crm \
  --skillName "deal_coach" --name "Deal Coach" --agentName "deal-coach" \
  --description "Helps reps qualify and advance deals" \
  --instructions "You are a CRM assistant. Be concise. Use the deals data source to answer." \
  --isAssistant true --mode conversation \
  --supportTools true --supportDataSources true --supportKnowledgeBase true --json
# → capture data.id as AGENT_ID

⚠️ Known backend issue (verified on env=dev). agent create currently fails with cannot insert a non-DEFAULT value into column "name". The dev create service (apps/api/src/dev/app/agent.service.ts createAgent) inserts name, but tenant_ai_agent.name is a generated column (GENERATED ALWAYS AS (COALESCE(agent_name, skill_name)) STORED) — Postgres rejects any explicit insert into it, regardless of flags. Until the service stops inserting name, agents can't be created via this endpoint. The flag/usage below is correct and will work once fixed; in the meantime, operate on agents created through another path. The flag values themselves are unaffected — Postgres derives name from agent_name/skill_name.

  • --skillName is the only required field. Everything else is optional with server defaults (mode→conversation, ownership→CUSTOM, status→1, is_agent→true, is_assistant/is_skill→false, memory_token_limit→4000).
  • Identity: --name (display), --agentName (internal), --description, --category, --welcomeMessage.
  • Models: --defaultAiModelId, --backupAiModelId, --cotAiModelId, --defaultReasoningLevel, --temperature, --maxTokens. Model ids are free-form FK strings (no enum) — resolve a real core_ai_model id, don't invent one.
  • Shape toggles: --isAgent, --isAssistant, --isSkill, --hasWorkflow, --multipleDeployment, --jsonOutput.
  • Capability toggles gate the sub-resources: --supportTools, --supportDataSources, --supportFiles, --supportKnowledgeBase, --supportWebSearch. Turn the relevant one on before attaching.
  • --ownership enum is APP | CUSTOM | PRODUCT | SYSTEM | USER (default CUSTOM). ⚠️ The flag help text says SYSTEM | APP | TENANT — that is wrong; TENANT is not valid.
  • JSON-only config (schemas, memory options) goes via --data: instruction_schema, input_form_schema, output_render_schema, memory_options, plus the long tail in references/agent-model.md.

Sub-resources

Each group supports list/get/create/update/delete (a few exceptions noted). Always pass app + --agentId; the row is --id.

docyrus agent data-sources create --appSlug crm --agentId AGENT_ID --tenantDataSourceId <ds-id> --privilege read --json
docyrus agent tools        create --appSlug crm --agentId AGENT_ID --coreAiToolId <tool-id> --json     # define <tool-id> first via apps ai-tools (docyrus-app-ai-tools)
docyrus agent docs         create --appSlug crm --agentId AGENT_ID --title "Pricing FAQ" --content "..." --json
docyrus agent mcps         create --appSlug crm --agentId AGENT_ID --tenantMcpServerId <id> --json
docyrus agent connections  create --appSlug crm --agentId AGENT_ID --connectedAiAgentId <id> --connectionType subagent --json
docyrus agent dynamic-contexts create --appSlug crm --agentId AGENT_ID --name "Today" --systemMessage "..." --json
docyrus agent workflow-steps create --appSlug crm --agentId AGENT_ID --name "Step 1" --type prompt --inputSchema '{}' --outputSchema '{}' --json
docyrus agent deployments  create --appSlug crm --agentId AGENT_ID --name "Production" --json
docyrus agent list --appSlug crm --json
docyrus agent get  --appSlug crm --agentId AGENT_ID --json

The full sub-resource table (required-on-create fields, createOnly fields, enums) is in references/agent-model.md.

Critical rules

  • Only --skillName is required to create an agent. Don't over-specify; rely on server defaults and add what the use case needs.
  • Capability toggle before attachment. Attaching a data source/tool/knowledge doc without turning on the matching --support* flag leaves the agent unable to use it. Set the toggle on create (or update) first.
  • --ownership valid values are APP | CUSTOM | PRODUCT | SYSTEM | USER (default CUSTOM); the flag help is inaccurate.
  • Tools are a two-step, separate flow. docyrus agent tools create --coreAiToolId <id> only attaches an existing app-scoped tool (a tenant_ai_tool/core_ai_tool) to the agent. You define the tool first via docyrus apps ai-tools — see docyrus-app-ai-tools.
  • Sub-resource rows use --id, not --toolId/--dataSourceId (deliberate, to avoid clashing with body flags). createOnly fields (--coreAiToolId on tools, --tenantDataSourceId on data-sources) are rejected on update; --archived is update-only.
  • agent tasks and agent recurring-tasks are non-functional against the current backend (no matching dev routes → 404, verified: Cannot GET …/agents/:id/tasks). Don't design around them; schedule recurring work with an automation recurrence trigger (docyrus-automation-design) instead.
  • Sub-resource create ops need an agent you can see under RLS. Sub-resource creates run inside an RLS-enforced write transaction that re-checks the agent via agentExists (id + tenant + app). An agent owned by another user (or with a null owner) is invisible there and the create returns 404 "resource not found" — even though list/get (broader reads) show it. Operate on agents you own. (Sub-resource list ops are verified working; create happy-path is currently gated by the create bug above.)
  • agent get/list/create/upload etc. don't error on extra JSON keys — sub-resource "required on create" fields (e.g. title, name, core_ai_tool_id, connection_type) are enforced by the backend DTO, so a missing one returns a server error, not a CLI complaint. Read back to confirm.
  • Model/data-source/tool ids are real FKs, not enums — resolve them (e.g. from list-data-sources, apps ai-tools list) before attaching; a fabricated id fails server-side.
  • Validate then test, every time. Re-read the agent + sub-resource lists. Delete throwaway agents you create.

Validate

  1. docyrus agent get --appSlug <app> --agentId <id> --json — confirm identity, models, mode, ownership, and that each --support* toggle is set as intended.
  2. For each attached kind: docyrus agent data-sources list … , docyrus agent tools list … , docyrus agent docs list … , etc. — confirm the attachments exist and reference the right ids.
  3. If --hasWorkflow: docyrus agent workflow-steps list … — confirm steps, their type, and parent/sort_order sequencing.
  4. If deployed: docyrus agent deployments list … and the deployment-tools/data-sources lists.

Per-field "what correct looks like" is in references/workflow-examples.md.

Test

  1. Well-formed check: every list above returns the expected rows and every referenced id resolves (no dangling FK). The agent get shows capability toggles consistent with its attachments.
  2. Exercise it where a path exists — invoke the agent from an automation ai-agent node (docyrus-automation-design) or the app's chat surface, give it a prompt that needs an attached data source/tool, and confirm it uses them.
  3. Clean up: delete throwaway sub-resources, then docyrus agent delete --appSlug <app> --agentId <id> --json.

Full test playbook in references/workflow-examples.md.

References

  • references/agent-model.md — Agent fields grouped by purpose (identity, models, shape, capabilities, render/output, memory, multi-agent), the enum values and server defaults, and the full sub-resource table (groups, required-on-create, createOnly, enums).
  • references/workflow-examples.md — End-to-end worked agent (assistant + data source + tool + knowledge doc), validation checklist, test playbook.
  • docyrus-platform → references/ai-capabilities.md — what the agent builder can do. docyrus-cli-app — CLI command index; docyrus agent <sub> --help for exhaustive flags. docyrus-app-ai-tools — author the AI tools an agent calls (apps ai-tools); attach them to the agent here with agent tools create --coreAiToolId.
发现
标签

此技能尚未发布标签。

版本
最新版本元数据

版本

v2026.09.24

发布时间

Sep 24, 2026

分类

未分类

许可证

未指定

源路径

skills/docyrus-agent-design

默认分支

main

最新提交

5eb990f

Tree SHA

d012754