Lightweight Harness on top of DPSK-Harness / ClaudeCode / Codex using DPSK-Harness

[XiaoSongXS]

Hi DPSK Team and the Community,

I want to know is it possible to build a lightweight harness using dpsk-harness (for some fixed pattern work, let' say gpu kernel debug) and make that work on top of DPSK-Harness / ClaudeCode / Codex, such that i can (a) have the lightweight harness suit for my customized day to day work and (b) reuse the existing agentic flow in DPSK-Harness / ClaudeCode / Codex as it include quite good system prompt, rules, requirement, etc.

The reason I ask about this is b.c. it can be quite hard for developers to build a agentic flow from 0 to something that can match both their (a) customized need, and (b) quality of claude code / codex. A more practical method is build on top of claude code / codex, some additional lightweight "customized" layer (e.g. make sure agent always run some ptx unit test before change code, make sure agent always check official doc, some agent dedicated for compute sanitier run).

Wish someone can help explain if it possible to use dpsk-harness for this purpose.

Best,
Xiao

[denial123789]

Yes — but I would separate two different designs, because they have different reuse boundaries.

1. Make the GPU workflow a DSH specialization

This is the most direct path. Treat the workflow as a small product assembled from DSH primitives:

• a profile / overlay chooses the composition;
• AGENTS.md and a skill encode the PTX-debugging procedure and official-doc requirements;
• dedicated tools wrap compute-sanitizer, unit tests, profiling, and artifact capture;
• approval and sandbox plugins enforce what those tools may change;
• an optional subagent preset owns the sanitizer or documentation-checking role.

For example, make the success contract observable rather than prompt-only: the final answer is not accepted until the PTX unit test and sanitizer tool have returned structured passing results. A system-prompt sentence is useful guidance, but a tool/workflow gate is the reliable enforcement layer.

The shipped architecture supports this shape: profiles layer bundles and cordis.patch.yml; the base bundle already composes models, tools, persistence, sandbox/approval policy, skills, and host-level subagent providers. Use dsh --profile <name> --dump-config to inspect the resolved product before running it.

2. Use DSH with Claude Code or Codex

DSH has dormant Claude Code and Codex subagent providers in the base composition, so a DSH agent can delegate bounded work to those products when an Agent Preset enables the corresponding delegation tool. This is the clean route when DSH should remain the coordinator and Claude Code/Codex should be workers.

The inverse direction is possible through a protocol boundary rather than by inheriting another product's hidden prompt. DSH ships an ACP automation-server example; an ACP-capable parent can spawn it, create sessions, prompt them, cancel them, and receive streamed messages. That gives composition reuse, not automatic reuse of every Claude Code/Codex rule or internal agent loop.

So I would not design around “stacking system prompts.” Prompts, tools, permissions, context loading, persistence, and lifecycle are a single runtime contract. Recreate the behavior you need as explicit DSH skills/tools/policies, and delegate only tasks with a narrow input/output contract.

A practical first milestone:

1. Fork the headless or ACP example into a gpu-debug profile.
2. Add one gpu-kernel-debug skill.
3. Expose run-ptx-tests, run-compute-sanitizer, and lookup-official-docs as bounded tools.
4. Add a workflow gate requiring test + sanitizer evidence before completion.
5. Only then add a Claude Code/Codex subagent for difficult code edits.

The independent DeepSeek Harness Handbook has source-backed architecture and “designing an agent” guides that map profiles, skills, tools, policy, and subagents; it may help while turning this into a small reproducible profile.

Relevant upstream references: docs/architecture.md, the ACP package, and the ACP demo composition.

[XiaoSongXS]

thanks for the detailed explain. let me dig to ur reference material and see wether we should use a DSH native agentic flow or a lightweight add on.