conduct Commands

Sequence diagrams for each conduct subcommand — the C4 dynamics level of the conduct architecture. Participants are conduct's internal packages: cmd/ (Cobra commands), topology/ (topology parsing and validation), plan/ (execution plan builder), state/ (deployment state under ~/.local/share/conduct/), executor/ (SSH runner) — plus the externals they talk to (filesystem, the ssh client, and through it the remote nodes where pipeline tools such as lay, enjoin, weaver, and waiter actually run). The shipped binary has exactly four subcommands: deploy, status, exec, and destroy.

conduct deploy

Reads a topology file (-f, required; -f - reads stdin; JSON tried first, then YAML), validates it, topologically sorts nodes by depends_on, and builds one step per node role. --dry-run prints the numbered plan to stderr and executes nothing. A real run executes steps sequentially — never in parallel — over ssh (host-key checking disabled, BatchMode=yes, per-node user/port/key honored), with cross-node vars like {{ db-server.host }} interpolated and injected as export VAR=...; prefixes on the remote command. Progress goes to stderr, remote stdout to stdout, and the final deployment state as JSON on stdout; the first failing step marks the node and deployment failed and exits 1. Note: a role's entities list is parsed from the topology but never used — no entity file is copied to the node, so the remote tool must already have its inputs.

conduct status

With a deployment name, loads its state file and prints the deployment header plus a node table (Node | Host | Status) to stdout; an unknown name is an error, exit 1. Run bare, it lists every *.json in the state directory as a summary table — unreadable or corrupt state files are silently skipped, and an empty or missing state directory prints "No deployments found." on stderr and exits 0. Status is read purely from local state; no node is contacted, so it reflects the last deploy run, not live reality.

conduct exec

Runs an arbitrary command on one node of a recorded deployment: conduct exec <deployment> <node> -- <command...>. Everything after the first -- in the raw argument vector becomes the remote command; without -- any trailing positional words are joined instead, and no command at all is an error, exit 1. Remote stdout/stderr are passed through and the process exits with the remote command's exit code. Gotcha: the connection always uses root@host on port 22 with no identity file — the topology's per-node user, port, and key are not persisted in the state file, so exec ignores them even when deploy used different values.

conduct destroy

Deletes the deployment's state file from ~/.local/share/conduct/ and confirms on stderr; a name with no state file is an error, exit 1. Despite the verb, nothing is torn down: no SSH connection is made and remote nodes are left exactly as deployed — this removes the record, not the infrastructure.