dsh-improved-inline-edit

Realizing mid-task that your DSH agent is heading in the wrong direction is frustrating. Waiting for it to finish wastes time; stopping it cold throws away the context and progress you already built. dsh-improved-inline-edit turns interruption and continuation into a single gesture: the moment an agent starts running, a quiet input bar slides in above the composer. Type your correction, hit send, and the instruction is injected via the official agent.steer() mechanism into the agent's very next decision step—all context preserved, original goals intact. When the run ends, the bar fades out and the interface is clean again, zero distraction.

Beyond the core steering flow, a 100-entry Deep phrase pool keeps the status line alive, cycling through playful states like Deep diving, Deep simmering, and Deep zigzagging as the agent progresses, with one-click Chinese-English switching. Version 0.2.0 adds message recall, localStorage draft persistence, an optional steering-prefix toggle, and an auto-growing textarea (capped at 140 px) so long corrections are never clipped. The send button smoothly shifts to make room for a checkmark, then slides back after two seconds.

This plugin is built for developers who run long DSH sessions—coding, architecting, multi-step reasoning—and want to course-correct mid-flow the way you would interject with a colleague. It ships as a standard bundle patch, touches no DSH source code, and uninstalls cleanly.