dsh-mode

An agent that plans a one-line answer with full todo-list ceremony, or improvises a multi-step refactor without structure, wastes turns on every turn. Pushing methodology into the system prompt fails in the opposite direction: thousands of tokens of standing law eventually fade into wallpaper the model stops reading. dsh-mode cuts both extremes. The methodology is no longer resident in the prompt; it arrives as an ordinary tool result at the exact moment you switch phase.

The plugin ships five modes — base, planner, analyst, explorer, executor — each carrying a complete mindset text that codifies phase-specific discipline: the analyst must reproduce before concluding and tag every claim with evidence level, the executor must stop and report before any irreversible action. One switch_mode call enters the mode; it writes zero session events, never touches the system prompt, and leaves no residue on later turns. After a context compaction, re-switching to the same mode restores the methodology idempotently. Beneath the variable mode layer, a static core-behavior section provides mode-independent working principles — root cause before action, honesty about uncertainty, scope discipline, conclusions first — as a fixed floor shared by all five modes.

Built for developers in the dsh-agent ecosystem who want their agent to shift working posture as the task phase changes, without stacking rule lists into the system prompt or manually narrating phase changes. One explicit tool call, the methodology arrives, it guides the phase, and it walks away without sticking.