setup

name: setup description: Use when a first-time user needs help setting up inspire-cli configuration, or when troubleshooting config issues. Guides through credentials, project discovery, workspace binding, and verification. allowed-tools: Bash(inspire *), Bash(uv run inspire *), Bash(uv run playwright *), Bash(mkdir *), Bash(ls *), Read, Write, Edit

Inspire CLI Setup

Guide users through configuration by collecting info that can't be auto-detected. Treat CLI help as source of truth (inspire --help, inspire <group> --help, inspire <group> <command> --help) — this doc can lag behind releases.

What's automated vs what needs user input

Setup flow

Phase 1: Credentials

Ask for base URL, username, password. Check env vars and global config first — skip if already present. Recommend INSPIRE_PASSWORD env var in shell profile (never commit passwords).

Phase 2: Discovery

Run inspire init --discover from the user's project working directory. Needs playwright (uv run playwright install chromium if missing). If config files already exist, discovery prompts before rewrite; use --force for non-interactive runs.

After discovery: inspire config show and inspire config check to verify.

Phase 2b: Project preference

Show discovered projects from global config catalog. Ask user to rank by preference. Write project_order (list of project names) to [defaults] in project config.

Sort order when no --project flag: project_order (first match wins) > gpu_unlimited (tiebreaker) > priority > name. Projects not in the list sort after all listed ones.

Phase 2c: Verify paths via CPU notebook

Always do this. The catalog paths come from the API and can be wrong. CPU notebook SSH is guaranteed to work (zero setup needed) — use it as ground truth.

Always use --project explicitly when creating notebooks during setup. At this point project_order isn't configured yet, so auto-selection falls back to gpu_unlimited > priority which often picks the wrong project. Ask the user which project to use.

Create a CPU notebook, SSH in, explore the filesystem. Confirm:

• shared_path_group — usually /inspire/hdd/global_user/<username>, visible across ALL projects.
• target_dir — usually /inspire/hdd/project/<slug>/<username>, project-specific workdir.

The CPU notebook can be stopped after verification — it is not needed for subsequent phases.

Phase 3: GPU SSH bootstrap (skip if CPU/4090 only)

GPU notebooks (H100/H200) have no internet. notebook ssh still works but needs an SSH server package installed on the notebook.

rtunnel is handled automatically. The CLI downloads rtunnel locally (~/.local/bin/rtunnel) and uploads it to each notebook via the Jupyter Contents API. No shared filesystem placement or CPU notebook needed.

SSH server requirement — one of:

• apt_mirror_url — simpler, no pre-placed debs needed. Ask if the platform has an internal Ubuntu mirror (Nexus/Artifactory). Dropbear installed automatically via apt.
• dropbear_deb_dir — pre-place dropbear .deb packages on shared filesystem.

Both go in [ssh] section of project config.

Questions to ask:

1. "Do you need GPU notebook SSH?" — skip if no
2. "Does your platform have an internal APT mirror?" — determines apt_mirror_url vs dropbear_deb_dir

Legacy option: rtunnel_bin in [ssh] config is still respected as highest priority if set. Users who already have rtunnel on shared filesystem do not need to change anything.

Phase 4: Bridge profile

inspire notebook ssh <id> --save-as <name> saves a reusable profile. Ask which notebook should be the default bridge — typically a CPU notebook for code sync/execution.

Bridge profiles break when notebooks restart. bridge exec/ssh auto-reconnect for notebook-backed profiles, but manually-added tunnels need manual recovery.

Troubleshooting (check in this order)

1. inspire config show — look for default source tags (means not configured), placeholder values
2. inspire config check — validates auth, catches stale passwords
3. Missing target_dir — most common cause of sync/bridge failures
4. Wrong workspace — bridge/sync need CPU workspace (internet), jobs need GPU
5. GPU SSH not working — needs either apt_mirror_url or dropbear_deb_dir for SSH server installation. rtunnel is auto-uploaded and should not be the cause; check ~/.local/bin/rtunnel exists locally
6. rtunnel upload failed — check Jupyter Contents API access; fallback: set rtunnel_bin in [ssh] config pointing to a shared filesystem path
7. Stale catalog — re-run inspire init --discover to refresh

Reporting issues

If you encounter a setup problem you can't resolve (discovery failures, unexpected CLI behavior, missing features), suggest the user file an issue:

• Private (Codeberg): https://codeberg.org/cyteena/inspire-cli/issues
• Public (GitHub): https://github.com/EmbodiedForge/Inspire-cli/issues