Introduction

When developing agents in DeepSeek Harness (DSH) by routing through GPT/OpenAI Responses routes, if steps like compression, Replay, and search do not align with the provider’s native capabilities, issues such as switching implementations mid-route or repeated cache invalidation can easily occur. DSH’s philosophy is “everything is a plugin,” and these gaps are perfect for plugins to fill.

The dsh-lcx-codex introduced below does exactly this: it intercepts the DSH GPT Responses path, filling in Native V2 Compact/Replay, Hosted Search, Prompt Cache, and continuous web operations.

Overview

dsh-lcx-codex is a DSH plugin maintained by kk3ya03-star, released under the MIT license. The current stable version is 0.4.2, and npm latest points to this version. One-sentence description (from the original README):

Adds Native V2 Compact/Replay, Hosted Search, Prompt Cache, and continuous web operations to the GPT/OpenAI Responses route of DeepSeek Harness.

Note: This is an independent community plugin and has no affiliation or official endorsement with OpenAI, DeepSeek, Sub2API, or NewAPI.

Core Features

Intercepting Responses Requests from the First Round

When LCX is enabled, it takes over GPT Responses requests starting from the first round of normal requests. Tool continuation, Compact, and Replay no longer switch implementations mid-route.

The plugin does not override the DSH host: Agent, Session, tool execution, and compaction policy are still managed by DSH. The plugin’s own Pi version (0.84.3) is isolated from the DSH host’s Pi (0.82.1) and will not upgrade the host’s Pi via override.

Native V2 Compact / Replay

The timing of compression is still decided by DSH; LCX is only responsible for sending the compression request to Responses Native V2 and receiving the result back into the existing Session. Default pressure policy:

< 90%   Normal operation
90%     Prefer Native V2 Compact
95%     Allow DSH emergency prune

Manual /compact continues to use DSH’s own compaction transaction. Native V2 Compact supports Replay, restart recovery, and GPT model switching.

Native state is only reused on compatible same session/route; when switching to an incompatible GPT route, it automatically falls back to portable history instead of hard-stuffing the old opaque state.

Prompt Cache

On GPT-5.6 compatible routes, LCX uses prompt_cache_options and maintains a stable cache identity. Normal continuous conversation and tool-heavy tasks can continuously hit the cache as long as the request prefix remains unchanged.

Search: Retain Native Entry, Enable Advanced Capabilities on Demand

Normal search still goes through DSH’s native web_search, with the pipeline:

DSH web_search → LCX SearchProvider → GPT Hosted Search

When more control is needed, two capabilities can be enabled separately:

  • websearch_gpt_advanced: Hosted Search parameters such as domain filtering, location, search context, image search, etc.
  • websearch_alpha: Continuous search / open / find / click / screenshot operations; disabled by default, only registered after the current route’s capability probe passes.

Installation and Usage

First, confirm your environment, then execute the installation command. You need:

  • DSH 0.1.1-rc.2
  • Node.js ^22.19.0 || >=24.0.0
  • A working GPT Responses route

Installation commands:

dsh plugin --profile web add dsh-lcx-codex
dsh web

The first command installs the plugin into DSH’s web profile, and the second starts DSH Web.

After the steps above, you can see the LCX switch in the plugin settings of DSH Web.

Typical Usage

In DSH Web’s plugin settings, turn on LCX:

LCX OFF  → Fully use DSH native LLM flow
LCX ON   → Current GPT session follows the LCX Responses path from the first round

Two reminders:

  1. LCX is only for GPT/OpenAI Responses. Turn off LCX before switching to non-GPT models like Claude, Gemini, or DeepSeek; no need to restart DSH.
  2. It is recommended to use default values for configuration and only change options that are truly needed; see the repository README for the full list of configuration items.

Compatibility and Known Limitations

Official compatibility scope of 0.4.2:

  • Plugin 0.4.2
  • DSH 0.1.1-rc.2
  • DSH host Pi 0.82.1
  • Plugin Pi 0.84.3 (@earendil-works/pi-ai, used only by the plugin)
  • Node.js ^22.19.0 || >=24

DSH 0.1.2-alpha.1 is not yet included in the official compatibility scope of 0.4.2.

Known limitations:

  • Current commonly used routes do not accept content-level prompt_cache_breakpoint, and the plugin does not expose explicit breakpoint settings.
  • When dynamically loading skills/plugins that change the top-level tool set, a Prompt Cache reset may occur once; tool functionality is not affected.
  • reasoning.context / reasoning.mode depends on whether DSH / Pi actually exposes these fields.
  • 1.05M long context and Programmatic Tool Calling are not officially supported items in 0.4.2 yet.

Use Cases and Notes

Suitable for: Developers using GPT/OpenAI Responses routes in DSH who need stable compression for long sessions, Hosted Search, or Prompt Cache. If you only use non-GPT models, this plugin is not useful.

Two notes before installation:

  1. The plugin runs with the permissions of the current dsh process. Once installed in a profile, it has the permissions of the dsh process. It is recommended to check the source code and license before enabling to ensure there are no issues.
  2. Confirm versions against the compatibility scope above, especially the DSH version.

Summary

dsh-lcx-codex is very focused: it completes the DSH GPT Responses path, ensuring compression, caching, and search use provider native capabilities as much as possible, while not touching the DSH host. If you run GPT routes in DSH, you can try the steps above.

  • Plugin directory page: https://www.skillhub.cn/plugins/kk3ya03-star/dsh-lcx-codex
  • GitHub repository: https://github.com/kk3ya03-star/dsh-lcx-codex

Final reminder: skillhub.cn is an independent community plugin directory with no official affiliation to DeepSeek/Huanshan; plugin information is based on the repository README.