dsh-llm-agent-bridge

DeepSeek Harness ships its own inference engine, yet many developers want Claude to handle the thinking step while keeping DSH tool sandbox, approval policy, and conversation trajectory intact. dsh-llm-agent-bridge registers Claude as a selectable provider inside DSH via the Anthropic official Claude Agent SDK, usable from Standard, PTC, Creator, Minimal, or any custom preset without swapping out the entire workflow.

The core design is a clean separation of duties: DSH owns the system prompt, tool definitions, sandboxed execution, and approval chain; Claude only reasons and decides which tool to invoke. Every tool call routes back through the DSH agent loop, the result is fed into the same Claude query, and the shared conversation history is preserved across provider switches. The plugin also projects adaptive thinking summaries, per-turn token usage, and the context window into DSH native stream vocabulary; auto-prepends missed turns when switching models mid-session; and actively shuts down the built-in OTLP telemetry exporter at load time so no Claude output leaks into telemetry uploads.

Built for developers who already hold a Claude subscription and want Claude reasoning inside every DSH mode without losing tool trajectories or approval guarantees. The plugin never reads, stores, or intermediates credentials; all authentication is handled by the Claude Code binary itself.