Preface¶

In the plugin-based usage scenarios of DeepSeek Harness, some requirements do not involve adding a new agent runtime, but rather connecting the existing longgraph prompt library to the conversation’s skill seam. dsh-plugin-longgraph belongs to this category of community plugins: it registers three authoring skills—longgraph, loop-graph, and loop-converge—into ctx.skills.

What This Is¶

dsh-plugin-longgraph is a community plugin maintained by levi-qiao, licensed under MIT. It serves only DeepSeek Harness and does not install or start nodes on other products. According to the documentation, it has no official affiliation or maintenance relationship with DeepSeek AI, nor is it an official Harness bundle.

This plugin does not add an agent runtime. It exposes the longgraph prompt library through DeepSeek Harness’s skill seam and uses resourceBase to point to the packaged skill directory, allowing templates and references to be resolved.

Core Capabilities¶

The plugin provides three skills:

1. longgraph: Performs a fit check first, then delegates to loop-graph.

2. loop-graph: Conducts interviews, compiles to .longgraph/<date-slug>/, then launches or prints DSH prompts.

3. loop-converge: Applies the code-convergence preset, then performs the same compilation.

In implementation, it has no runtime config and injects skills via ctx.skills.registerProvider.

It also declares dsh.bundle.patch and ships with cordis.patch.yml, which inserts a line named dsh-plugin-longgraph.

dsh.bundle.patch
cordis.patch.yml
dsh-plugin-longgraph

Installation and Enablement¶

DeepSeek Harness is currently in developer preview. The installation commands provided in the documentation are as follows:

dsh plugin --profile web add github:levi-qiao/dsh-plugin-longgraph
dsh --profile web

After installation, the following skill can be invoked:

invoke skill({ name: "longgraph" })

You can also invoke loop-graph or loop-converge.

To pin a specific commit, use:

dsh plugin --profile web add github:levi-qiao/dsh-plugin-longgraph#<commit>

It should be noted that the installation commands in the documentation use --profile web; the documentation does not specify whether this profile is required in all environments.

Git installation fetches the source code, and prepare runs tsdown. If you are using pnpm 10+ and pnpm blocks prepare, you can add the package to the profile’s pnpm-workspace.yaml build allowlist and then retry.

The documentation also provides a Direct Cordis row:

- id: longgraph-skills
  name: dsh-plugin-longgraph

Typical Usage¶

First, choose an entry point based on your goal:

1. Use longgraph: Perform a fit check first, then delegate to loop-graph.

2. Use loop-graph: Enter the interview flow, compile to .longgraph/<date-slug>/, then launch or print DSH prompts.

3. Use loop-converge: Apply the code-convergence preset first, then perform the same compilation.

If timers are involved, the documentation provides the following constraints:

• Use two separate root sessions.
• The every_seconds lower bound is 300.
• The schedule is session-local, and both sessions must remain live.
• If a session is busy, it will wait until idle before followup(); the timer will not interrupt firing.
• Do not followup() or inject() into another session.

Dependencies and Compatibility¶

The peerDependencies ranges provided in the documentation are as follows:

@deepseek-ai/cordis >=4.0.1 <5
@deepseek-ai/dsh-skill >=0.1.0-rc.5 <0.2.0

Since DeepSeek Harness is still in developer preview, breaking compatibility changes will require a new plugin version.

Development Commands¶

For local verification, the documentation provides the following commands:

pnpm install
pnpm typecheck
pnpm test
pnpm build
pnpm integration

Applicable Scenarios and Notes¶

This is suitable for developers who want to use the longgraph prompt library within DeepSeek Harness, especially in scenarios where only skill registration is desired without introducing an additional agent runtime.

It is not suitable as a replacement for:

• dsh-agent-loop
• Schedule
• Jobs
• Web UI

It is also not a multi-host skill pack.

Before installation, note the following: the plugin runs with the current dsh process permissions and registers skills that can be invoked by sessions. It is recommended to review the GitHub source code, the MIT license, the peerDependencies ranges, and the line inserted into cordis.patch.yml.

Links¶

• GitHub: https://github.com/levi-qiao/dsh-plugin-longgraph
• Directory page: Not provided in the documentation.

Conclusion¶

The value of dsh-plugin-longgraph is quite specific: it connects the authoring skills longgraph, loop-graph, and loop-converge to DeepSeek Harness’s ctx.skills, allowing models to use the corresponding instructions through the skill seam. For scenarios where you want to use the longgraph prompt library in DSH, it serves as a lightweight community integration point.