1. Install and open RepoOps, then choose your repository in first-run setup.
2. Review the detected stack, capture choices, and exact before-and-after file changes. Inspection does not run repository scripts.
3. Apply the reviewed setup. A changed file invalidates its preview, so review again if another session edits the repository.
4. For an existing installation, open Settings and review its memory setup or integration maintenance plan.

The terminal adoption flow uses the same integration engine. Run npx repoops adopt --dry-run from the repository to inspect the capture integration, then run npx repoops adopt to review and confirm. This capture scope is broader than the memory-only setup in Settings.

The npm package is a launcher for the compatible app or command bundle. It does not add a project runtime dependency. Desktop installation also supports teams whose stack does not use Node.

The minimal brain contains an index, decisions, lessons, and source/version metadata under .claude/brain/. RepoOps merges its managed sections into CLAUDE.md and AGENTS.md. Existing text outside those sections is preserved.

Capture integration can also configure supported assistant hooks and settings. Read the plan for the selected scope; creating brain files alone does not activate capture.

Credentials, transcript buffers, upload checkpoints, setup journals, and backups belong in local application state outside the repository. Use the authenticated publisher to share enrolled-repository evidence with your team.

Local integration

The planned files are readable and owned changes match the installed manifest.

Assistant capture

A supported session appears with its developer and repository identity. Check prompt, response, tool, model, and usage coverage; unsupported fields remain unknown.

Team delivery

Bind the device to the team, enable the required publishers, and confirm the hosted account received the evidence. A queued upload is not a receipt.

Production source

Use a supported production collector or provider connection and verify a runtime signal reached the hosted service. Its delivery must continue while developer laptops are asleep.

Attribution

Follow the runtime signal through service, release, commit, developer, session, prompt, and response. A missing link is partial attribution, not an inferred match.

Connect the local app to your team. Owners manage enrolled repositories and evidence policy. Repository collaborators and cloud or model-provider permissions are separate observations; access alone does not prove who caused an incident.

In Settings, select the repository, integration scope, and maintenance action. Inspect the plan before applying repair, rollback, or removal. The history identifies local runs; rollback requires a specific run.

Only unchanged RepoOps-owned content can be reversed automatically. If you edited an owned section, the plan shows a conflict and preserves it. A process lock protects concurrent setup; recovery is offered only when the recorded process is no longer running.

Deterministic setup needs no model call. An AI assistant can help explain architecture or propose lessons later, using source files and revisions. Review that work before adopting it as repository guidance, and choose a cost limit for any model-powered task.

The legacy bootstrap prompt and kernel command remain available for existing workflows. They are optional tools, not prerequisites for the local app to create its initial brain.

Resolve incidents through Attribution, AI Security, Remediation, LLM Cost Metrics and Memory, the primary pages in the sidebar. Preserve the verified fix and lesson, then measure recurrence rather than treating a written lesson as proof of prevention.

Inspect brain health or return to installation.