# DeepSeek Harness Tool Reference

This file maps Superpowers concepts to native DSH tools. Read it when a skill references a Claude Code tool or hook.

## Skill invocation

- Load a skill with the DSH `skill` tool before acting when the skill applies.
- Skill names are bare names such as `brainstorming`, `writing-plans`, and `systematic-debugging`.
- The `using-superpowers` skill is bootstrapped automatically by the preset template.

## Claude Code → DSH

| Claude Code | DSH | Notes |
| --- | --- | --- |
| `Bash` | `pwsh` on Windows / `bash` on POSIX | Use native shell semantics. |
| `Read` / `Write` / `Edit` | `read` / `write` / `edit` | Omit `sandbox_permissions` unless a genuine wider escalation is needed. |
| `Glob` / `Grep` | `glob` / `grep` | `glob` returns files only; `grep` uses ripgrep syntax. |
| `TodoWrite` | `todo_write` | Each call replaces the complete list. |
| `Task` | `subagent` / `subagent_fork` | Background by default. |
| `AskUserQuestion` | `ask_user_question` | Stable question ids are echoed in answers. |
| `WebSearch` | `web_search` | Returns summaries and source URLs when configured. |
| Plan mode | `exit_plan_mode` | Submit the complete plan for approval. |
| Background jobs | `run_in_background: true` | Use `job_output`, `job_list`, and `job_kill`. |

## DSH-native capabilities

- Persisted objectives use `create_goal`, `get_goal`, and `update_goal`.
- Large fan-out work uses `workflow` with phases and structured results.
- Fresh-agent iteration uses `ralph` only when the human explicitly requests it.

## Windows sandbox rule

When the current effective mode is already `danger-full-access`, do not send `sandbox_permissions: "danger-full-access"` or a `justification`. The preset compatibility wrapper also cleans legacy same-mode fields if a model emits them anyway. Genuine wider requests in lower modes remain subject to the normal approval policy.