demo

name	demo
description	Constraint-first interactive front door to the Profumo blind-spot demo (arXiv:2506.19062). Presents a model picker (Singlet-Doublet, 2HDM+a, Dark SU(3)) and delegates to the chosen per-model skill. Invoke when the user says "run the demo", "show me HEPhaestus", "reproduce the blind spot", or on fresh-install evaluation.

Demo

Thin front door to the Profumo blind-spot demo. Presents a paper intro, lets the user pick a model, and hands off to the per-model skill (/singlet-doublet, /2hdm-a, or /dark-su3). Each per-model skill owns its own constraint interview, time-estimate gate, and execution chain.

This skill must never fall back to analytic Python for an observable. All physics observables come from real tools driven by the per-model skills.

Flow

Step 0 — Preflight

Read ${XDG_CONFIG_HOME:-~/.config}/hephaestus/config.json. Required keys: madgraph_path, sarah_path, spheno_path, wolfram_engine_path.

For each key, confirm the executable responds (--version or equivalent no-op). Build the missing-list: every required key absent from the config or whose executable does not respond. If it is empty, continue to Step 1. Otherwise, present an AskUserQuestion gate offering to install the missing tools. The pick_subset option is included only when the missing-list has more than one entry.

{
  "question": "<N> of 4 tools needed for /demo are missing: <comma-separated missing-list>. Install them now in their default locations?",
  "options": [
    {"id": "install_now",  "label": "Install missing tools now", "description": "Hand off to /install and let it install the missing tools in their defaults. /demo resumes once install completes."},
    {"id": "pick_subset",  "label": "Let me pick which to install", "description": "Hand off to /install in interactive mode so you can choose. Re-run /demo afterwards."},
    {"id": "cancel",       "label": "Cancel — I'll run /install myself later", "description": "Stop /demo cleanly with no install attempt."}
  ],
  "allowMultiple": false,
  "required": true
}

Dispatch per answer:

• install_now — for each tool in the missing-list, invoke /install --tool <tool> (mapping madgraph_path → mg5, sarah_path → sarah, spheno_path → spheno, wolfram_engine_path → wolfram). After all installs return, re-read the config and re-probe the four keys. If the missing-list is now empty, continue to Step 1. Otherwise stop and print:

  "Install finished but these tools are still missing/unresponsive: <list>. Inspect the install logs, then re-run /demo."

• pick_subset — invoke /install with no arguments (its bundle/per-tool picker takes over). Then stop and print:

  "Pick the tools you need in /install, then re-run /demo."

• cancel — stop and print:

  "The demo needs MadGraph, SARAH, SPheno, and Wolfram Engine configured. Missing or unresponsive: <list>. Run /install to set them up, then re-run /demo."

Do not probe common install paths from /demo itself. /install owns installation; /demo only orchestrates the gate above.

Completion: all four keys present and responsive (continue to Step 1), or the skill has stopped with one of the printed messages above.

Step 1 — Paper introduction

Print verbatim:

Arcadi & Profumo ask: where can dark matter hide from direct detection? They identify blind spots — regions of parameter space where the tree-level DM-nucleon coupling vanishes by cancellation, so the direct-detection signal is suppressed far below naive expectations. Blind spots matter because they weaken "direct detection rules out WIMPs" arguments: a model can evade current limits not by tuning the DM mass, but by tuning the couplings to a cancellation.

This demo walks the full pipeline for one of three paper-benchmark models — Lagrangian → SARAH → SPheno → MadGraph/MadDM → a figure — with constraint selection (relic, direct, indirect) driving which sub-skills run. Some prereq skills (the multi-component DM combiner and ID pull-computation) are on the roadmap. For 2HDM+a the loop-only direct-detection chain (FeynArts → FormCalc → /looptools eval → /ddcalc) is now wired end-to-end and yields a real, EW-anchor-validated σ_SI at one benchmark point (box not folded into f_N (≤1.7× upward), SI only; not an experimental exclusion; Tier-3 smoke gated off by default, green on a tooled box); for the other models the per-model loop-DD chain is not yet integrated end-to-end. Pending constraints will surface as [COMING SOON] and you can choose to run only the ready subset.

Step 2 — Gate: continue?

AskUserQuestion:

{
  "question": "Ready to begin?",
  "options": [
    {"id": "continue", "label": "Continue"},
    {"id": "not_now",  "label": "Not now"}
  ],
  "allowMultiple": false,
  "required": true
}

On not_now, stop cleanly with no output. On continue, proceed to Step 3.

Step 3 — Model picker

AskUserQuestion with three model options:

{
  "question": "Which model do you want to explore?",
  "options": [
    {"id": "singlet-doublet", "label": "Singlet-Doublet (~2–4 hr cold; relic + tree-DD READY, ID COMING SOON)", "description": "3×3 neutralino-like mixing, tree-level blind spot, loop floor. Relic READY via /maddm. Direct detection READY tree-only via MadDM `generate direct_detection` → /ddcalc 90%-CL exclusion (canonical θ=0 benchmark sits far off the blind-spot locus); loop floor near the blind spot coming soon with /looptools eval runtime. Indirect detection COMING SOON with pull computation (v1+)."},
    {"id": "2hdm-a", "label": "2HDM + a (~2–5 hr cold; relic + parser-only ID READY, DD READY-with-caveat)", "description": "Pseudoscalar mediator, CP-forbidden tree SI, loop-only DD. Relic READY via hand-crafted SARAH model (Ω h² ≈ 10.494 at off-resonance benchmark). Indirect detection READY-with-caveat: MadDM `generate indirect_detection` → /gamlike v0 emits ⟨σv⟩ + channels + Fermi-LAT likelihood rows; pull computation deferred to dm-pull v1+. DD READY-with-caveat: loop-only via /feynarts → /formcalc → /looptools eval → /ddcalc yields a real, EW-anchor-validated σ_SI=1.18e-48 cm² at one benchmark point (~40% vs an independent EW box+triangle anchor; box not folded into f_N (≤1.7× upward), SI only; not an experimental exclusion; Tier-3 smoke gated off by default, green on a tooled box). The live /sarah-build renderer UFO backport remains separate, non-blocking debt."},
    {"id": "dark-su3", "label": "Dark SU(3) (~1–3 hr cold, relic only; DD/ID coming soon or vacuous)", "description": "Higgsed SU(3)_D → SU(2)_D dark sector with two DM candidates. Relic READY via analytic-only branch in /dark-matter-constraints. V DD/ID: COMING SOON — MG5_DARK_COLOR_TENSOR_WALL walls MadGraph/MadDM and an analytic DD/ID escape hatch is on the roadmap. Psi DD: PHYSICS-VACUOUS — σ_SI ≈ 0 by exact parameter-independent symmetry (Eq. 29), so the constraint is trivially satisfied rather than tool-blocked. Psi ID: COMING SOON, same MG5 chain."}
  ],
  "allowMultiple": false,
  "required": true
}

Disclosure (dsu3-002): dark-SU(3) relic uses sigma_v approximations (sigmav_approx=True). Paper fidelity (Ω_tot h² ≈ 0.12) is out of reach this run; values reported are regression-anchors, not physics targets.

The cold-hour estimates are total overlap-adjusted estimates assuming all prereqs were implemented; they are hard-coded from each per-model skill's ## Constraints table and constraints.yaml. The user sees which constraints are actually blocked inside the per-model skill's Step 3.

Delegation

Based on the user's choice, read and execute plugins/hep-ph-toolkit/skills/<singlet-doublet|2hdm-a|dark-su3>/SKILL.md. The per-model skill runs its own Steps 1–4 (candidate declaration → constraint interview → time-estimate gate → execution). /demo does not intervene.

Closing block

After the per-model skill returns, read ./demo_output/<model>/summary.json (<model> = the id chosen in Step 3). If it exists, print three lines:

Model run:    <summary.model>
Artifacts:    <summary.artifacts_dir>
Skipped:      <comma-separated list of summary.skipped_constraints[*].id> — <reason> (or "none")

If it does not exist (per-model skill exited via Cancel or errored before writing), print:

<model> interview was cancelled.

Non-goals

• /demo does not own constraint interview, time gates, or execution. These live in each per-model skill.
• /demo does not probe the marketplace for prereq availability at runtime. Prereq status is static — declared in plugins/hep-ph-toolkit/skills/_shared/constraints.yaml.
• /demo does not implement multi-component DM combination. Dark SU(3) per-candidate relic runs through the analytic-only branch of /dark-matter-constraints [EXISTS]; paper-fidelity multi-component combination into Ω_tot h² ≈ 0.12 is on the upgrade roadmap of analytic_models.dark_su3 (per the dsu3-002 regression-anchor banner in dark-su3 SKILL.md), not blocked on /dark-matter-constraints itself.
• /demo does not install tools. /install owns environment setup.
• No collider execution. The collider option in per-model Step 2 is a placeholder only.
• Per-model skills are scoped to Profumo arXiv:2506.19062 benchmark ranges in this iteration.