onboard

name: onboard description: "Use when first-touch harness setup is needed: user is new, asks where or how to start, invokes `/onboard`, opens with an empty or unclear prompt, or has no configured harness environment."

First-touch triggered setup. Clarify what the user is considering, infer the tools the task depends on, confirm the install plan, then route to the right harness skill.

Audience

The user may not know the harness vocabulary and may only have a vague first goal. Keep each question warm, short, and concrete. Do not ask them to name support tools.

Workflow

1. Intent Gate

Before installing, configuring, or routing, get the first intent.

If the opening prompt already makes the intent concrete, use it and skip this gate. Otherwise ask one single-choice question. Use an interactive choice UI when the host supports it; otherwise show the same options as a numbered list and wait.

Tell me what you’re considering first, so I can set up the harness properly. If none of these fit, describe the direction briefly.

Options:

Run a calculation
Reproduce a paper
Explore a model or physics question
Set up for later

If the user answers outside these options or uses a freeform "Other" field, ask at most one short follow-up, then classify triggers from that answer.

Completion criterion: one intent is selected, a freeform intent is understood, or the opening prompt already supplies an equivalent intent.

2. Task Detail

Ask only the next missing detail needed to decide triggers and routing:

Run a calculation: ask for the model/problem and target quantity if absent.
Reproduce a paper: ask for the paper, figure, or reproduction target if absent.
Explore a model or physics question: ask for the model/topic/question if absent.
Set up for later: ask whether the user wants minimal setup, common support tools, or a specific stack.

Completion criterion: the task is concrete enough to decide whether it needs source material, software docs, compute resources, and a method/software route.

3. Triggered Setup Survey

Inspect; do not install yet. Add a setup item only when a trigger below is active and the item is not already available.

Trigger	Setup item	Availability check
Ion-managed skills are missing or stale	`make skills`	`ion add`/skill state requires sync
Task may need source material beyond local KB: arXiv, DOI, PDF, captions, figures, methodology details, benchmark values, or external reference text	`pdf-render`	`.venv/bin/python -c 'import pymupdf4llm'`
Task may need precise/current package usage beyond local refs: package APIs, examples, setup, migrations, capabilities, or version-sensitive docs	`node`	Node.js 18+, `npm`, and `npx`
Task likely exceeds local compute, needs arrays/scans, or user wants reusable remote compute setup	`/setup-cluster`	`skills/using-slurm/profiles/active.toml`
A method/software route has been selected	matching stack install	Makefile `INSTALLABLE` plus `skills/<stack>/stack.toml`

Source-material support is not a paper-only mode. It is triggered whenever the task may need material outside .knowledge/ or the local skill references.

Completion criterion: every active trigger is either added to the pending install/config plan, marked already available, or explicitly skipped as irrelevant.

4. Route Software Before Installing Stacks

If the task involves a calculation, code, package use, or paper reproduction that needs a method/software route, consult the relevant local sources before proposing stack installs:

/model or /physics cards for problem routing.
/method-* skills for method choice.
/using-* skills and skills/<stack>/stack.toml for software choice, install target, smoke test, and runtime profile.
find-docs only when package/API details are harness-relevant, hard to find, or version-sensitive.

Present a single-choice fork with 2-3 real software/method options when there is a genuine choice. Each option gets a short reason and tradeoff. Include an Other / preferred stack escape hatch when the user may already have a preference.

Completion criterion: either a software route is selected, no route is needed yet, or the task is routed to /model, /physics, or /reproduce-paper to decide the route there.

5. Confirm Install Plan

Before running any install or configuration command, show one compact table:

Benefit	Item	Action
cleaner PDF extraction and extracted figures	`pdf-render`	`make install pdf-render`
more precise package usage	`node`	`make install node`
selected calculation stack	`itensors`	`make install itensors`
remote jobs without re-asking	cluster profile	`/setup-cluster`

Only include rows that apply. Then ask one yes/no question:

Install/configure these now?

If the user says no, skip the plan and continue only along paths that do not require the skipped setup.

Completion criterion: the user approves the plan, rejects it, or asks to change it.

6. Execute Approved Setup

For each approved Makefile target, confirm the target appears in INSTALLABLE, then run make install <tool>. For skill sync, run make skills. For cluster setup, hand off to /setup-cluster.

Classify failures:

Support tools (pdf-render, node): say what failed, offer to debug, and continue if the task can proceed without them.
Domain stacks: block compute on that stack until fixed.
Cluster setup: continue local-only only if the task remains feasible locally.

Do not inline package-manager commands in chat. The Makefile and stack contracts own install details.

Completion criterion: every approved setup item is installed, skipped by the user, or failed with its effect on the task stated.

7. Route or Exit

Run a calculation: route to /model for a specific Hamiltonian/model, or /physics for a cross-model question.
Reproduce a paper: route to /reproduce-paper.
Explore a model or physics question: route to /model or /physics; install nothing unless computation becomes necessary.
Set up for later: exit after the approved setup plan is handled.

Completion criterion: the next skill owns the task, or setup-only onboarding ends with one line: Harness setup handled. Bring a model, paper, or calculation when you are ready.

Rules

No installs before intent, trigger survey, and install-plan confirmation.
No support-tool questions before they are triggered by the task.
No preinstalling all method stacks.
One decision per question.
Use interactive single-choice gates when available; otherwise use concise numbered options.
Setup gates follow the project decision-anchor rule in AGENTS.md.
Do not lecture about the harness, list every skill, or ask the user to read docs.