dsh-ultracode

When working on substantive tasks with DeepSeek Harness, reasoning effort, parallel orchestration, and verification policy often require individual configuration and are easy to overlook. dsh-ultracode collapses all of this into a single /ultra toggle: when activated, every model request is pinned to the deepest reasoning effort the serving model declares, and a standing orchestration policy rides the system prompt—workflow fan-out by default, adversarial verification before reporting, and context hygiene. Toggle it off and everything returns to normal.

The design borrows from Claude Code's ultracode tier: the substance is not simply a deeper parameter but a standing orchestration policy. The policy text is a deployment-owned configuration; the default covers workflow opt-in, adversarial verification, and concurrency backoff retries. Spawned subagents inherit the ultra state through the delegation chain automatically; fork branches inherit through their seeded prefix and freeze there, requiring no extra setup. A ULTRA chip appears in the composer area, and while active the card carries a rainbow border for instant visual confirmation.

Ideal for moments when a task warrants the full reasoning budget—auditing integration points, cross-module refactoring, multi-file consistency checks—where you want depth locked in and the model working under an orchestration discipline you approve of, rather than hand-writing prompt instructions each time. It works with DeepSeek, GLM, Kimi, and any OpenAI-compatible route that declares efforts from the shared vocabulary.