dsh-repo-atlas

The biggest risk an AI coding agent faces in an unfamiliar repository is not that it cannot read the code, but that it fabricates understanding, presenting speculation as fact and unverified assumptions as conclusions. dsh-repo-atlas (RepoAtlas) is a DeepSeek Harness plugin built for exactly that gap: before any action is taken, it grounds what is in the repository in file-level evidence.

The plugin works along two lines. The understand line scans a confirmed workspace under path, content, sensitive-file, and budget limits, producing structured evidence that covers technology stack, entry points, key directories, tests, and configuration, along with a recommended reading order. Conclusions are strictly separated into those directly supported by files and those inferred from text, and uncertain parts are explicitly labelled partial or unknown rather than silently filled in. The review line lets the agent define explicit targets and intent for a session-only change proposal, and as the lifecycle reaches the relevant states, review, export, and verify a bounded patch. Worktree state, commit readiness, landing relation, recovery guidance, and release readiness are kept as separate observations, so that ready to review is never conflated with already applied.

Safety is a first-class design constraint. The default path is read-only: no patches are generated or applied, no commits, merges, pushes, dependency installs, network calls, or code uploads happen without explicit opt-in. Controlled actions require fixed recipes, explicit approval, sandbox enforcement, and bounded redacted output. All evidence caches, proposal registries, and preflight assessments are session-scoped and disappear when the session ends. The plugin suits developers and maintainers who want their coding agent to read before it acts and who keep human approval in the loop for every code change.