[FEATURE] DeepSeek Harness (dsh): document the mcp-client patch now, ship a bundle in the npm package once dsh's bundle contract is documented

[debugmcpdev]

🚀 Feature Request

Is your feature request related to a problem?

DeepSeek Harness (dsh, developer preview) is a plugin-only agent runtime on the Cordis kernel. It already consumes MCP servers through its first-party @deepseek-ai/dsh-mcp-client plugin, but nothing in our docs tells a dsh user how to register mcp-debugger, and the plugin bridges tools only (MCP resources and prompts are explicitly deferred), which affects what they should expect.

Describe the solution you'd like

Two steps, in the same shape as the pi integration (#714 / #715: the published package carries its harness glue, the repo root stays clean):

1. Docs now. A "For DeepSeek Harness" block next to the Codex/Claude Code/pi ones with the patch snippet for $DSH_HOME/cordis.patch.yml (or a per-profile patch):

   - insert:
       - id: mcp-debugger
         name: '@deepseek-ai/dsh-mcp-client'
         config:
           serverName: mcp_debugger
           transport: stdio
           command: npx
           args: ['-y', '@debugmcp/mcp-debugger', 'stdio']

   plus one honest sentence: tools surface as mcp__mcp_debugger__<tool>; the debug://sessions/{id}/output resource and the debugging-workflow prompt do not (tools-only bridge) — use get_output instead. The skill installs via the existing cross-agent copy into ~/.agents/skills/ (dsh's dsh-skill-filesystem scans it by default) or from the npm package's skills/debugging once 0.25.0 ships.

2. Later, a bundle in the published package. dsh's architecture doc describes a dsh.bundle manifest field for third-party packages. Once that contract is documented well enough to build on, @debugmcp/mcp-debugger can carry a bundle that mounts the mcp-client with our server entry and points dsh-skill-filesystem at the shipped skills/, so a dsh user gets the server and the skill from one package — the same one-install shape pi gets from the pi manifest.

Describe alternatives you've considered

• A dsh-* directory at the repository root — rejected by the CONTRIBUTING rule added in feat(pi): ship the agent skill and a pi manifest in the published package (#714) #715: harness integrations live in docs or inside the published package.
• Waiting for dsh to leave developer preview before documenting anything — the docs snippet is cheap and the mcp-client config shape is stable enough to publish now.

📊 Impact

• Who would benefit: DeepSeek Harness users
• Priority: Low
• Complexity: Simple for step 1 (docs); step 2 depends on dsh's bundle contract

📝 Additional Context

Verified against the dsh repo at the time of filing: packages/mcp/mcp-client (config keys serverName, transport: stdio|streamable-http, command, args, env, cwd; "Tools are the only bridged MCP capability — Resources and Prompts have no harness consumer mechanism and are deferred"), packages/skill/skill-filesystem (customSkillDirs; default roots include ~/.agents/skills), docs/user/guide/mcp-memory.md (the patch-file workflow, applied with dsh web --patch).