dsh-plugin-codex

dsh-plugin-codex addresses a core challenge: how to directly and flexibly utilize a local Codex App Server as the main model provider within the DeepSeek Harness framework, rather than as a constrained delegated sub-agent. The traditional integration may not fully expose Codex's model catalog and capabilities. By registering Codex as the codex-local provider, this plugin enables users to select and run various models offered by Codex (such as the GPT-5.6 family) directly within Harness's agent loop.

Its core strengths lie in intelligent session and tool management. The plugin maintains the local App Server process, automatically handles session cold/warm reconstruction, manages tool invocation via namespace mapping and execution, validates and converts image attachments, and translates Codex's internal state into standard Harness stream events. All configurations can be adjusted via the plugin settings in the Web UI, requiring no modification to core code.

This plugin is ideal for developers and teams already using DeepSeek Harness who wish to seamlessly integrate the powerful capabilities of Codex models. If you need to leverage Codex's image generation, optional web search takeover, or advanced reasoning and coding abilities within the Harness environment, while maintaining transparency and control over configuration, this is an excellent choice. It is particularly suited for complex application scenarios that require a stable, configurable local model inference service and value comprehensive tool call tracking and session context management.