Introduction

In DSH (deepseek-harness, referred to as dsh), agent developers frequently repeat similar multi-tool tasks: reading project files, summarizing content, organizing results, and then proceeding based on context. Existing approaches usually involve manually maintaining prompts, scripts, or fixed processes; DeepJIT addresses a different problem: identifying repeated hot workflows from agent execution traces, then using LLMs to compile them into reusable markdown skills or flow step templates, and hot-loading them into the running dsh.

What is it

DeepJIT is a DSH plugin maintained by fly3366, licensed under MIT. Its positioning is to consolidate repeated agent workflows into reusable assets, reducing the cost for developers to repeatedly organize similar tool call sequences.

It is aimed at developers using DSH to run multi-tool agent workflows, especially those who need to reuse similar processes in headless or web profiles.

Core Features

The following introduces several capabilities of DeepJIT.

  1. Monitor agent execution traces to mine repeated hot workflows.
    This step focuses on “which workflows repeat frequently” rather than the tasks themselves.

  2. Use LLMs to compile workflows into reusable markdown skills or flow step templates.
    The compiled artifacts are skills or flow templates that dsh can directly reuse later.

  3. Hot-load compiled artifacts into the running dsh without restarting.
    This means newly discovered skills/flow templates can enter the current plugin context rather than staying in local files only.

  4. Provide a compiled lifecycle: AOT verification, tiering/promotion, deoptimization, and GC.
    These mechanisms correspond to the process of artifacts from validation, promotion, to degradation and cleanup.

  5. Record metrics and output GenAI semantic-convention spans for LLM calls.
    This provides structured information for observing LLM call quality later.

Installation and Usage

DeepJIT declares compatibility with Node ^22.19 || >=24 and supports headless and web profiles. DSH is currently in pre-release, and the API might drift, so it is recommended to confirm the current DSH version and environment before enabling.

Installation commands can choose between npm package names or git repository forms based on project needs.

Installing from an npm package name:

dsh plugin --profile web add @fly3366/deepjit

Installing from a git repository:

dsh plugin --profile web add github:fly3366/DeepJIT

If you mainly run headless processes, you can also replace --profile web with --profile headless, for example:

dsh plugin --profile headless add github:fly3366/DeepJIT

Typical Usage

The following is a reproducible headless usage workflow. First, do the installation, then task execution, then check the compiled artifacts.

Install the plugin:

dsh plugin --profile headless add github:fly3366/DeepJIT

Run a task to read project configuration files and summarize:

DEEPSEEK_API_KEY=... dsh --profile headless "read package.json and tsconfig.json, then summarize"

After repeatedly executing similar tasks, DeepJIT will mine and compile hot workflows based on traces.

View compiled artifacts:

dsh plugin --profile headless "use deepjit_status to list compiled artifacts"

If you need to clean up local data, you can use the following command:

rm -rf ~/.dsh/deepjit

Applicable Scenarios and Notes

DeepJIT is more suitable for the following scenarios:

  • You frequently run similar multi-tool workflows;
  • You want to consolidate high-frequency workflows into markdown skills or flow step templates;
  • You want to hot-load new artifacts during dsh runtime rather than restarting every time;
  • You need to retain metrics and GenAI semantic-convention spans for LLM compiled calls.

Pre-use considerations:

  • The plugin runs with the permissions of the current dsh process; please check the source code, dependencies, and license before installing;
  • File writing is limited to ~/.dsh/deepjit/;
  • Aside from LLM calls via dsh ctx.llm, there is no other network access;
  • It does not store keys; you can rely on dsh credential services or the DEEPSEEK_API_KEY environment variable;
  • DSH is pre-release, and the API might drift;
  • The lifecycle mechanism handles AOT verification, tiering/promotion, deoptimization, and GC; the availability status of artifacts may change based on usage feedback.

Conclusion

The core value of DeepJIT is to transform repetitive agent workflows in DSH from “doing it manually every time” into “monitored, compiled, reusable, and degradable” assets. For DSH developers who frequently run similar multi-tool workflows, it provides a relatively complete entry point for live compilation and lifecycle management.

Directory Page: https://www.skillhub.cn/plugins/fly3366/DeepJIT
GitHub: https://github.com/fly3366/DeepJIT