dsh-subagent-claude-live

The built-in Claude Code provider in DeepSeek Harness is a one-shot call: the model is pinned in config, the parent sees only the final answer, and the sidebar job pane reads "Failed to read output" while the child runs. dsh-subagent-claude-live turns Claude Code into a genuine streaming subagent. It appears in the subagent catalog as a session-backed child, and its thinking, text, tool calls, and tool results render token by token live, exactly like an in-process child, before being flushed to persistence on completion.

The plugin registers two tools on the host plane. claude_code delegates one self-contained task and accepts per-call parameters: a model (preset name, CLI alias, or full id), an effort level from low to max, a permission mode, and an optional run_in_background flag that returns a job id immediately while the child keeps streaming. claude_models queries the installed CLI for the current supported model list on demand, so a model released tomorrow is usable tomorrow without touching config. The CLI process is spawned through dsh-subprocess, inherits the harness credential-scrubbed environment, and by default uses the claude binary on PATH, letting native CLI auto-updates flow through without a 250 MB bundled payload.

Best for developers building multi-agent workflows in Harness who need to switch Claude models frequently and observe what the subagent is actually doing. If your pipeline orchestrates tasks on the DeepSeek side and delegates concrete coding work to Claude Code, this plugin turns that hand-off from a black-box relay into an observable, tunable, streaming conversation.