dsh-subagent-flowtext

When DeepSeek Harness executes a single Agent instruction, it typically must run through classification, planning, tool invocation, clarification, and approval. Exposing every one of those steps as DSH tool calls adds configuration overhead and makes the execution chain hard to govern in one place. dsh-subagent-flowtext keeps the scope narrow: it routes all DSH Agent traffic through the FlowText loopback gateway, so DSH stays in the lightweight role of receiving user instructions and surfacing the final answer, while FlowText owns the entire execution lifecycle end to end.

Every Agent request DSH receives is forced onto the FlowText gateway. By default, safe-phase summaries, execution plans, and tool digests stream back as reasoning content rather than appearing as DSH tool calls; setting progressMode to off limits the reply to the final answer alone. A given sessionId reuses one persistent FlowText conversation panel, and separate sessions remain fully isolated. Clarification prompts and dangerous-operation approvals pause DSH until the user handles them in the FlowText UI. Model-level retries are explicitly disabled to prevent duplicate write tasks. Credentials live locally only and never enter the Profile, repository, shell history, or model context.

The plugin is aimed at knowledge workers and developers who already run the FlowText Agent Gateway inside Obsidian desktop and use DeepSeek Harness as their day-to-day AI assistant. If you want DSH to stop orchestrating tools directly, delegate the full execution pipeline to FlowText, and still keep tight control over approvals, clarifications, and retry behavior, it fills the clean, single-route bridge between the two systems.