DA Agent

An AI assistant Cloudflare Worker for the Document Authoring (DA) platform. It exposes a streaming chat API backed by Claude on Amazon Bedrock and integrates directly with the DA Admin API to read, create, update, and manage content.

Features

• Streaming chat — Server-sent event stream via the Vercel AI SDK (POST /chat)
• DA tools — Full set of DA Admin API operations available as LLM tools: list, get, create, update, delete, copy, move sources; versioning; media upload; fragment lookup
• Human-in-the-loop approval — Destructive tools (create, update, delete, move) require explicit user confirmation before executing
• Collab integration — When the user is in edit view, reads and writes go through the live Y.js collaborative session (da-collab service binding) so changes appear in real time
• Page context awareness — The current org, site, path, and view are injected into the system prompt and used as defaults for all tool calls
• Custom skills — Customers define reusable workflow instructions as markdown files stored at /.da/skills/ in their DA repository; the agent loads and injects them automatically on each request

Architecture

Client (DA authoring UI)
  │  POST /chat  (messages, pageContext, imsToken)
  ▼
da-agent (Cloudflare Worker)
  ├── Vercel AI SDK streamText → Amazon Bedrock (claude-sonnet-4-6)
  ├── DA tools → DAADMIN service binding → da-admin Worker
  ├── Collab client → DACOLLAB service binding → da-collab Worker
  └── Skills loader → DAADMIN (reads /.da/skills/*.md)

Project structure

src/
  server.ts            # Worker entry point, chat handler, system prompt, skills loader
  collab-client.ts     # Y.js collab session client
  da-admin/
    client.ts          # DA Admin API client (wraps service binding calls)
    types.ts           # TypeScript types for DA Admin API
  tools/
    tools.ts           # Vercel AI SDK tool definitions wrapping DAAdminClient
    utils.ts           # Shared utilities (path helpers)
samples/
  skills/
    translate-content.md  # Example skill file
test/
  server.test.ts       # Vitest tests

Local development

Prerequisites

• Node.js 22+
• Wrangler CLI
• AWS credentials with Bedrock access (for the Claude model)
• Local instances of da-admin and da-collab Workers running

Setup

npm install

Copy .dev.vars.example to .dev.vars and fill in your credentials:

AWS_BEARER_TOKEN_BEDROCK=<your-aws-bearer-token>

Start the dev server (connects to local da-admin and da-collab via service bindings):

npm run dev

The worker is available at http://localhost:5173.

Chat API

POST /chat
Content-Type: application/json
Authorization: Bearer <ims-token>

{
  "messages": [...],
  "pageContext": {
    "org": "my-org",
    "site": "my-site",
    "path": "/docs/my-page",
    "view": "edit"
  },
  "imsToken": "<ims-token>"
}

Returns a streaming UI message response (Vercel AI SDK format).

HEAD /chat   →  200 OK  (health check / CORS preflight)

Custom skills

Skills let customers extend the agent with site-specific workflow instructions without any code changes.

Create markdown files at /.da/skills/<skill-name>.md in your DA repository. The agent fetches all .md files from that folder on every request and injects their contents into the system prompt.

Skill format

---
name: My Skill
description: What this skill does and when to use it.
triggers:
  - keyword or phrase that activates this skill
---

# My Skill

Instructions for the agent...

## Steps

1. Step one
2. Step two

See samples/skills/translate-content.md for a full example that translates the current page into another language and saves it under a language-specific path (e.g. /de/, /fr/).

Tools

Tool	Description	Requires approval
da_list_sources	List files and folders in a repository path	No
da_get_source	Read a file's content	No
da_create_source	Create a new file	Yes
da_update_source	Update an existing file	Yes
da_delete_source	Delete a file	Yes
da_copy_content	Copy a file to a new path	No
da_move_content	Move a file to a new path	Yes
da_create_version	Snapshot the current state of a file	No
da_get_versions	Get version history for a file	No
da_lookup_media	Look up a media asset	No
da_lookup_fragment	Look up a content fragment	No
da_upload_media	Upload a base64-encoded image or file	No

Deployment

Deployments are managed automatically via semantic-release on every push to main.

The pipeline (main.yaml) runs three jobs:

Job	Trigger	What it does
Test	All branches	npm run lint + npm test
Test Deploy	Non-main branches	Deploys to da-agent-ci (staging services) + semantic-release dry run
Release	main only	Semantic-release: bumps version, updates CHANGELOG.md, deploys to CI then production, creates GitHub release

Manual deploy

npm run deploy

Required GitHub secrets

Secret	Description
CLOUDFLARE_API_TOKEN	Cloudflare API token with Worker deploy permissions
CLOUDFLARE_ACCOUNT_ID	Cloudflare account ID

Environments

Wrangler env	Worker name	Services
dev (local)	da-agent-local	da-admin-local, da-collab-local
ci	da-agent-ci	da-admin-stage, da-collab-stage
production	da-agent	da-admin, da-collab

License

Apache-2.0