cloudflare-orchestrator - SKILL.md Agent Skill

name: cloudflare-orchestrator description: "Route a Cloudflare task to the right spoke among the platform specialists — Workers and Pages compute, the wrangler CLI and binding/config model, storage (KV, R2, D1), Durable Objects for stateful coordination, the Agents SDK and Sandbox SDK, transactional email, and deploy/auth debugging. USE WHEN a user is building, configuring, or shipping something on Cloudflare's edge but hasn't named the specific service." cluster: cloudflare version: 1.0.0

Cloudflare Orchestrator

The single entry skill for Cloudflare edge work. It locates the task on the compute × primitive map and delegates to one of the specialist spokes. The cross-cutting model every Cloudflare app shares — a Worker (or Pages Function) reaching stateful primitives only through bindings declared in wrangler.jsonc under a pinned compatibility_date — lives in cloudflare-core; read it before adding a binding, choosing a storage primitive, or writing config.

Routing map (intent → spoke)

Deploy & operate (the front door)

"Deploy a Worker / MCP server / Pages site", auth/token failures, deploy debugging → cloudflare
Account-level management (deploy + KV + R2 + Pages + DNS/routes via scripts, API token setup) → cloudflare-manager

The CLI & config (everything routes through here)

Any wrangler command, wrangler.jsonc, compatibility_date, wrangler types, secrets, wrangler dev/tail/deploy → wrangler (binding model in cloudflare-core)

Write & review Worker code

New Worker, code review, anti-patterns (streaming, floating promises, global state, secrets, bindings, observability) → workers-best-practices

Stateful coordination

Chat rooms, multiplayer, booking/locking, per-entity state, alarms, WebSockets, SQLite-in-DO, RPC methods → durable-objects (DO is a binding too — see cloudflare-core)

Build AI agents

Stateful agents, durable workflows, scheduled tasks, MCP servers, chat/voice agents, React useAgent hooks, queues/retries/observability → agents-sdk

Run untrusted / arbitrary code

Code interpreters, CI/CD sandboxes, AI code execution, interactive dev environments, preview URLs (container-backed) → sandbox-sdk

Email

Send transactional email (Workers send_email binding or REST API), receive/route email, onEmail()/replyToEmail(), SPF/DKIM/DMARC, wrangler email setup → cloudflare-email-service

Publish & validate static artifacts on R2

NotebookLM/portal artifacts on R2: public URLs, MIME types, cache headers, audio/video playback, inventory/visibility claims → r2-notebooklm-artifact-portal

Standard Operating Flow

Locate the task: which compute surface (Worker / Pages / Durable Object / Agent / Sandbox) and which primitive (KV / R2 / D1 / queue / email) it touches.
If it adds or changes a binding, storage choice, or compatibility_date, pull the model from cloudflare-core first — bindings, config, and the deploy step are interlocking, not independent.
Almost every change ends at the CLI: route config/deploy/secret work through wrangler, and verify deploys via cloudflare (auth-token gotchas live there).
Delegate to the spoke(s). Multi-step asks fan out in build order (e.g. "stateful chat agent" → agents-sdk + durable-objects + wrangler binding + cloudflare deploy).
Return: chosen spoke(s), the bindings/config implied, the compatibility_date to set, and the next action.

Guardrails

See cloudflare-core. In short: declare every dependency as a binding — never hardcode secrets or account IDs in code (wrangler secret put); pin compatibility_date to a recent date and regenerate types after config changes (wrangler types); pick the right primitive (KV = read-heavy cache, R2 = large objects/no egress, D1 = relational, Durable Object = strongly-consistent per-entity state); and prefer retrieval over pre-trained knowledge — these APIs move fast, so fetch the current Cloudflare docs the spokes link before relying on a signature. The auth-token trap is real: env tokens often lack Pages permissions, so unset them for OAuth (cloudflare).

Loading spokes on demand

To keep CLI startup context lean, this cluster's spokes are not separately registered as skills — only this orchestrator and its *-core are enumerated. When you route to a spoke named above, load it on demand by reading its file:

~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md (or skills/<spoke-name>/SKILL.md inside the skill-clusters repo).