dsh-coding-tools

In the DeepSeek Harness ecosystem, developers often need a set of safe, practical coding tools for code reading, precise editing, structural search, and language-server queries—without touching the host's security model. dsh-coding-tools fills that gap: it mounts four bounded tool groups behind a compact Settings card, every operation gated by explicit approval prompts and sandbox policies, with zero modifications to DSH source code.

The core capabilities span four dimensions: code_read returns UTF-8 code windows tagged with exact-byte version tokens so downstream edits are traceable; edit_ranges performs atomic, whole-line, observed-range edits with independently togglable edit and sandbox-escalation approvals; ast_grep drives a pinned ast-grep process over a single file snapshot under forced read-only confinement, with pagination caps and no directory-level scans; lsp maintains a persistent read-only session to a language server, supporting hover, go-to-definition, references, symbols, and diagnostics—every response deduplicated, deterministically sorted, and byte-bounded before paging. Fresh diagnostic summaries are attached after plugin-owned edits when the matching LSP server is already active. Notably, ast_edit and DAP debugging are deliberately omitted from the secure initial release.

It is built for DSH operators and developers who want programmatic access to code reading, structured mutation, AST pattern matching, and LSP introspection while insisting that every step carries explicit approval, processes remain read-only, and outputs stay strictly bounded—all without patching the host or loosening sandbox policies.