The Workspace

nxs init leaves state in three places. This page says what each one is, so you can tell what is yours to edit, what is generated, and what never leaves this machine.

.nxs/ — the workspace itself

.nxs/
  config.toml    which blocks are active, plus each block's own settings
  db.sqlite      the one store every active block writes into
  replica.toml   this machine's identity: the id prefix and site id
  .gitignore     a single `*` — the directory ignores itself

config.toml is the only file here meant for human eyes:

active_modules = ["flow", "memory", "chat"]

[flow]
plugin = "issue-tracker"

active_modules is what nxs prime and nxs guide fan out over, and each block contributes its own [<block>] table. You can edit it, but nxs init is the safer route — it validates the names and runs each block's own setup.

The store is device-local and git-ignored. .nxs/.gitignore is a single *, so the workspace never lands in a commit: your board does not travel with your code. That is a deliberate choice, not an oversight — the way to share state is a sync stream (nxs sync), and the way to share a decision is the pull request. If you came expecting the tracker in git, this is the thing to know first.

A pre-existing .nexusflow/ directory from an older version is renamed to .nxs/ when the workspace is opened — idempotently, and only where the workspace actually resolves, so nothing is stranded and opening an old workspace repeatedly is harmless.

The shared agent file

Each active block contributes a managed section to AGENTS.md, assembled into one document rather than three competing ones. That file is committed: it is how a contributor — or an agent on another machine — learns the conventions of this project.

CLAUDE.md is deliberately not a second copy. It belongs to the host that runs the session hooks below, so a block there would be a second delivery of what the hook has already handed the session — and the memory binding in it would arrive as every memory's full body instead of the budgeted index. nxs init therefore only ever takes a retired block back out of CLAUDE.md; it never writes one.

The session hooks

nxs init wires one SessionStart hook per active block into .claude/settings.json, each running that block's own prime. Memory's entry — and only memory's — carries a || cat NEXUS_MEMORY.md tail, so that a contributor who has not installed the suite still gets the project's memory: the settings file is committed, and on their machine the block binaries are simply not commands. It hangs on one entry rather than all of them because three entries carrying it would deliver the same file into the same session three times.

A failing hook is left to fail loudly rather than being papered over with an echo. A hook that exits zero while delivering nothing is worse than one that says the tool is missing.

Checking it is healthy

nxs doctor

doctor — nxs status is the same command — reports the active modules, the schema version and its standing, the replica identity, sync state, the op count, and a database integrity check. It adds one warning line if a half-finished migration from beads is still wired; a clean workspace prints nothing there. It is foundation-only: it works whatever blocks are active, and it opens no product view, so it still answers when a block is unhappy.

nxs migrate

The schema is raised on open by default; migrate is the explicit lever for CI and repair. It says which of the two things happened rather than succeeding silently — schema v3 → v4 when it really migrated, and workspace db already current (schema v4) when there was nothing to do, which is the usual answer in CI.

Going further

• Sharing a workspace between machines or people: nxs sync, and running a relay for the server half.
• What each block does with the store: modules.