English | 中文

dsh-kernel-mesh

DSH runs on one simple idea: everything is a plugin. Models, tools, subagents — plug them together however you like.

So we did exactly that: we turned four coding harnesses — Kimi Code, Grok Build, OpenAI Codex, and MiniMax Mini-Agent — into DSH plugins, bundled in this one package.

The payoff is simple: switch to a Kimi / Grok / Codex / MiniMax model inside DSH, and it's no different from opening each one's own CLI. Every model stays in the environment it knows best — main agent or subagent, it feels like coming home.

What's inside

• L1 kernel routes — kimi-kernel / grok-kernel / codex-kernel / minimax-kernel, registered as DSH model routes your main agent can switch to.
• L2 subagent recipes — kimi-agent / kimi-explore / kimi-plan, grok-agent / grok-explore / grok-plan, codex-explore / codex-worker (each kernel's own subagent types; minimax has none upstream). Each recipe carries the vendor's own subagent prompt and tool whitelist, so a kernel subagent sees and uses exactly what that harness's subagent would — independent of the parent's preset.
• Three kernel tools — kernel_status, kernel_run, kernel_switch.
• Vendor search tools, offered only when opted in — kimi_search / kimi_fetch appear only if dsh-kernel-kimi is installed and a Moonshot credential exists; grok_search / grok_fetch appear only if dsh-kernel-grok is installed and a Grok OAuth credential exists. They stay separate tools (different corpora). Official web_search is DeepSeek's own search and is not a wrapper.

Kernel matrix

Kernel	Wire	Endpoint
kimi-kernel	Anthropic Messages	https://api.kimi.com/coding/v1/messages (kimi-cli 1.49.0: max_tokens clamped to remaining context; catalog includes k3)
grok-kernel	Responses (proxy)	https://cli-chat-proxy.grok.com/v1/responses
codex-kernel	Responses (custom)	your codex base_url + /responses
minimax-kernel	Anthropic Messages (CN)	https://api.minimaxi.com/anthropic/v1/messages

System prompts

Each kernel plugin registers the vendor's own upstream system prompt as the agent's sole system-prompt section (complete: true + suppressRuntimeContext()), so a session on a kernel sees ONLY that harness's prompt — not DSH's.

Fallback routes

When a kernel's own API is out of quota, set DSH_KERNEL_USE_FALLBACK=1 (default on) to route the kernels through your existing subscriptions:

Kernel	Fallback route	Model
kimi-kernel	ollama	kimi-k2.7-code
grok-kernel	ollama	gpt-oss:120b
codex-kernel	opencode-go	gpt-5.6-luna
minimax-kernel	ollama	minimax-m3

Keys are read from ~/.dsh/.credentials.yaml (OLLAMA_API_KEY, MY_OPENCODE_GO_API_KEY). Set DSH_KERNEL_USE_FALLBACK=0 to force the official kernel APIs.

Install

Copy this directory into your profile's node_modules and register it as a bundle:

cp -r dsh-kernel-mesh ~/.dsh/profiles/node_modules/dsh-kernel-mesh

Then add "dsh-kernel-mesh" to dsh.profile.bundles in ~/.dsh/profiles/web/package.json and restart:

dsh web

Once pnpm works in your environment, you can install straight from GitHub:

dsh plugin --profile web add github:oppnc/dsh-kernel-mesh

Usage

kernel_status                          # which kernels, L2 types, transports are registered
kernel_run(kernel, type, task)         # fan a task out to a foreign kernel
kernel_switch('kimi')                  # set the default model route for future sessions

Kernel	type values
kimi	coder, explore, plan
grok	general, explore, plan
codex	explore, worker
minimax	(none — no subagent tool upstream)

Screenshots

The whole family on GitHub — every README has a one-tap language switch:

Full-size images live in docs/screenshots/.

License

MIT — see LICENSE.