dsh-subagent-codebuddy

dsh-subagent-codebuddy bridges Tencent CodeBuddy CLI into DeepSeek Harness as an in-process reasoning subagent via an LLM adapter. Unlike a direct ACP integration, context continuity is managed entirely by the dsh session, eliminating the cross-task bleed that occurs when CodeBuddy auto-resumes sessions by working directory. The plugin requires zero source-code changes and works out of the box.

It registers the subagent_codebuddy delegation tool and the codebuddy provider route, enabling parallel agent creation with independent continuation. Each call serializes the agents full conversation history into the prompt while tool steps are bidirectionally translated between the dsh session and the CodeBuddy process. Image inputs are persisted to local paths for CodeBuddy to read, and on Windows the 32K command-line limit is handled by automatically writing oversized prompts to temporary task files. The list_codebuddy_models query tool lets the main agent discover available model IDs and pass one dynamically via the model parameter.

Built for developers running multi-agent dsh workflows who need to delegate inference tasks in parallel-especially with image understanding or long-prompt scenarios-and want a zero-build, drop-in bridge to CodeBuddy.