Getting started

Pick one path. Both run the same workflow; they differ only in surface.

Requirements: bb desktop ≥ 0.43, a coding-agent CLI installed and authenticated (bb drives the CLI you already have — it does not replace it), and a normal bb project backed by a local workspace source. The plugin vendors all 32 skills, so there is no separate skills install step.

curl -fsSL https://calionauta.github.io/stelow/install.sh | bash

Or from a shell:

bb plugin install stelow

Or in bb: Extensions → Plugins → marketplace, install Stelow. Full instructions: plugin/install-bb.md.

Then:

  1. Open Stelow in bb's left navigation and select a project with Stelow

state.

  1. Choose run knobs (quality, supervision, exploration) and Review mode (default Auto) —

labeled Quality, Supervision, Exploration and Pause for my review on the board — enter a product request in the composer, and submit. The card is created in Triage and the agent begins there.

  1. When the agent needs a decision, a structured question appears — answer

in the form, in the thread, or from the card. The agent waits instead of guessing.

  1. Approve each gate only after review; the plugin records the portable

receipt.

Path B — skills-only (any Agent Skills-compatible host)

Install the skills into ~/.agents/skills/ — either:

git clone https://github.com/calionauta/stelow.git
cd stelow
./install.sh
npx skills add calionauta/stelow -g

./install.sh variants: --minimal (skills only, no optional tooling), update (re-copy + prune retired), remove, --help. Non-interactive CI: ASSUME_YES=1 ./install.sh. The installer never modifies your AGENTS.md / CLAUDE.md.

Then, in any compatible agent:

/sw-start "Here's what I want to build"
/sw-status

Advanced installer options (--minimal, update, remove, non-interactive CI, third-party skill registry, why Git-based distribution): INSTALLATION.md.

/sw-* commands are skill-provided and work on every compatible host — there is no host command registry. Setup asks for run knobs and Review mode explicitly; gates without a native review UI fall back to portable approval receipts under .stelow/approvals/.

Check it worked

scripts/stelow status [--json]   # from a workflow checkout (Path B)
scripts/stelow doctor [--json]   # detects drift: stale locks, missing dirs
bb stelow status --project <id>  # Path A equivalent

If status reports no workflow, the scaffold step (/sw-start or the board composer) has not run yet — scripts/stelow never creates state, it only validates and advances it.

Common problems

SymptomLikely cause
Skills not picked up after installAgent reads another skills directory, not ~/.agents/skills/ (agentskills.io standard)
advance refuses a transitionThe move is not in transitions.md — the stage graph is enforced, not advisory
Worker asks the same question twiceA previous ask timed out before persisting; answer once, the worker re-asks per protocol
Optional tooling missing (cymbal, sem, ctx7)All optional with documented fallbacks; the workflow runs without them