dsh-claude-code

DSH handles multi-step orchestration and tool routing, but conversations frequently contain a fully self-contained coding subtask—fixing a bug, writing a new module, refactoring a function. Pushing it turn by turn is slow and loses context; handing the whole thing to a dedicated coding agent that can call tools, iterate, and return a finished result is often far more efficient. dsh-claude-code is built for exactly this: it delegates such tasks to a local Claude Code subscription inside your DSH session, feeds the result back to DSH for the next step—DSH is the leader, Claude Code is the worker.

The plugin drives the local claude CLI through the official Claude Agent SDK, extracts no OAuth tokens, and stays fully compliant with Anthropic subscription and authentication policy. Core capabilities span foreground synchronous execution and background async mode (run_in_background returns a jobId immediately; stream live output via job_output); session resume for multi-turn iteration (pass the previous sessionId to keep context); structured output, cost caps, appended system prompts, and custom subagents for fine-grained control. A built-in Web monitoring panel shows 5-hour and 7-day usage gauges, a horizontal task tab strip, Markdown-rendered real-time output, tool-call cards, collapsible thinking blocks, and one-click cancellation, giving you a clear view of long-running tasks.

Ideally suited for advanced DSH users who already hold a Claude Code subscription and want to offload complete sub-tasks as a whole. Note that this plugin does not turn Claude into a raw model adapter hitting the API directly—that would require extracting OAuth tokens, violates Anthropic policy, and risks account termination. Keep cwd and permissionMode within a scope you trust, because Claude Code executes its own tools outside DSH sandbox constraints.