dsh-cloudflare

DeepSeek Harness (DSH) plugin providing 1:1 parity with the Codex Cloudflare plugin, as published in openai/plugins plugins/cloudflare.

It bundles the Codex plugin's complete skill surface and its two slash commands, adds two further Cloudflare skills (think and flue), and ships a ready-to-apply cordis.yml for its single MCP server.

What it provides

Skills

Eleven Cloudflare skill bundles, discovered as the cloudflare provider. The first nine are vendored verbatim from the Codex plugin; think and flue are authored additions covering Cloudflare technologies the Codex plugin does not yet ship (see PROVENANCE.md).

Skill	Purpose
agents-sdk	Build AI agents on Cloudflare Workers using the Agents SDK
building-ai-agent-on-cloudflare	End-to-end agent building workflow
building-mcp-server-on-cloudflare	Build remote MCP servers on Workers
cloudflare	Comprehensive Cloudflare platform skill
durable-objects	Stateful coordination with Durable Objects
flue	Build/deploy Flue agents on Cloudflare (Flue Durable Objects, Workers AI)
sandbox-sdk	Secure sandboxed code execution
think	Durable chat agents with @cloudflare/think
web-perf	Web performance auditing via Chrome DevTools MCP
workers-best-practices	Workers production best-practice review
wrangler	Wrangler CLI usage and configuration

The catalog is generated from the vendored SKILL.md frontmatter by scripts/gen-manifest.mjs, so the DSH catalog stays a single source of truth for every skill directory (see PROVENANCE.md).

Slash commands

The Codex commands /cloudflare:build-agent and /cloudflare:build-mcp map to the DSH command-name grammar (which forbids :), as:

• /cloudflare-build-agent
• /cloudflare-build-mcp

Each handler expands the verbatim Codex command brief (substituting $ARGUMENTS) and queues it as a follow-up turn on the receiving agent, then acknowledges — preserving Codex's invoke→model-does-the-work behavior.

MCP server

The Codex plugin declares one MCP server. DSH models one MCP server per @deepseek-ai/dsh-mcp-client instance, so the shipped patch layer inserts one cloudflare-api mcp-client row beside the plugin row. The cloudflare-api server mirrors .mcp.json (https://mcp.cloudflare.com/mcp).

Install

Published to npm as @try-works/dsh-cloudflare.

@try-works/dsh-cloudflare is a DSH bundle as well as a plugin: its package.json declares dsh.bundle.patch pointing at the shipped cordis.patch.yml, which inserts the plugin row plus the MCP client row. Installing it as a plugin therefore auto-mounts the whole layer:

dsh plugin --profile <profile> add @try-works/dsh-cloudflare

dsh plugin forwards to pnpm in the profile directory, then reconciles the package into dsh.profile.bundles because it declares dsh.bundle; on the next boot the profile composes the shipped cordis.patch.yml over its layers. (Installing it as a plain dependency — e.g. pnpm add @try-works/dsh-cloudflare — stays possible but only warns and does not mount the layer; apply the repo-root cordis.yml overlay in that case.)

For a manual --patch install, the repo-root cordis.yml carries the same two rows:

dsh --profile <profile> --patch ./cordis.yml

Either path requires the profile to already provide ctx.skills, ctx.commands, and ctx.tools (the shipped @deepseek-ai/dsh-base bundle does), so the skills, slash commands, and MCP tools mount cleanly.

How DSH becomes aware of the plugin

DSH has no plugin auto-discovery: a plugin becomes known to the runtime when a loader entry row names it, i.e. when its name appears in the composed entry tree. The bundle dsh.plugin add path and the --patch overlay are both just ways of adding that row; once the row is composed, the plugin's apply registers its skills (ctx.skills) and commands (ctx.commands), which is what makes them available to the model — the skill catalog (@deepseek-ai/dsh-tool-skill) and the command registry surface registered skills and commands to the agent.

AGENTS.md is not a plugin registration mechanism and does not need to be edited for the plugin to work. DSH's @deepseek-ai/dsh-agent-instructions loads AGENTS.md/CLAUDE.md only from the project tree and $DSH_HOME; it has no plugin-inventory input. Adding a note to an AGENTS.md is an optional per-workspace convention, not something this plugin requires or performs.

Development

node scripts/gen-manifest.mjs   # regenerate src/manifest.ts from vendored SKILL.md
node scripts/typecheck.mjs      # tsc -b packages/dsh-cloudflare (builds lib/)
node scripts/parity-check.mjs   # assert 11 skills / 2 commands / 1 MCP server

The package builds standalone: npm install inside packages/dsh-cloudflare (devDependencies include typescript and @types/node, and the peers resolve from npm) then npm run build / npm run prepack emits lib/. scripts/typecheck.mjs first looks for a package-local typescript, and otherwise falls back to the DeepSeek Harness checkout at DSH_HARNESS (default D:/deepseek-harness).

Provenance & license

Vendored skills/commands originate from openai/plugins at commit 11c74d6ba24d3a6d48f54a194cd00ef3beea18f9 (plugin version 0.1.2). See PROVENANCE.md for the full record. The package and plugin code are Apache-2.0 (see the repo-root LICENSE); vendored skill content carries its own licenses, Apache-2.0 for the Cloudflare and Flue bundles (see packages/dsh-cloudflare/skills/*/LICENSE.txt).