dsh-btw

When using DeepSeek Harness for complex, long-running agent tasks, users often face a dilemma: you may need to quickly confirm a detail—like a configuration name or code logic—without derailing the main workflow. Sending a regular message interrupts the flow and pollutes the context, while opening a new session loses accumulated task information. The dsh-btw plugin is designed to solve this exact friction. Inspired by Claude Code's /btw command, it lets you fire off a one-off question via a bypass channel while the main agent continues its work. It snapshots the parent session's context (including system prompts and tool schemas), generates an independent answer displayed in a temporary panel, and ensures nothing enters the main conversation history, preserving the task's focus and coherence.

Its core strengths lie in being 'non-intrusive' and 'fast'. First, it supports parallel execution: you can submit a question even when the agent is invoking tools or thinking deeply, without pausing the primary task. Second, by leveraging context sharing and cache optimizations, it delivers accurate answers with virtually zero additional background-explanation overhead. All interactions are ephemeral: the Markdown response appears above the input box and can be dismissed via shortcuts, staying fully isolated from regular session logs. It offers three sidechain modes (private JSONL, memory-only, or none) to balance audit needs and privacy.

dsh-btw is ideal for developers, researchers, or technical writers who thrive in immersive, session-driven workflows. If you frequently collaborate with DeepSeek Harness agents on lengthy tasks like code refactoring, project planning, or documentation, this plugin becomes a key tool for maintaining flow. It transforms 'quick checks' from workflow interruptions into a safe, unobtrusive utility. However, for tasks requiring new file reads, command execution, or multi-turn discussion, you should still use regular main-session messages.