mg-study

/mg-study [topic]

/mg-study is a warm-up skill: it scans MadGraph source ahead of any task and caches what it finds, so the agent enters later work already knowing its way around — instead of accumulating knowledge only as a byproduct of answering questions. Invoke it deliberately between tasks. It is not auto-invoked.

/mg-study NLO, /mg-study decays, /mg-study the EFT machinery — study a topic. Bare /mg-study — each consultant studies its own core area.

Orchestration

A per-owner parallel dispatch: the relevant consultants in parallel, each into its own consultants/<agent-name>/ subtree; the lead's own pass on lead/; aggregate.

• Topic given — dispatch the consultants that own the topic (NLO → the nlo + loop slices (nlo-syntax, amcatnlo, madloop, fks, ...); decays → the decay + downstream slices (chain-decay, bw-window, madspin-interface, ...)). Route by slice ownership as for any dispatch.
• No topic — dispatch each consultant to study its own core source areas (the ones its card lists under "Source code areas").

Per-consultant dispatch — four phases

Each consultant, dispatched on its own subtree, first reads its existing pages (always gap-aware: add what is missing, extend what is thin, correct what source now contradicts; never duplicate). Then it works the area through four phases, by reading source:

1. Explore — locate the relevant files and functions; record paths and structure.
2. Physics — extract what the code computes and the physics/algorithm it encodes. Understanding, not paraphrase.
3. Gaps — record what cannot be settled from source alone.
4. Cautions — flag source-visible hazards: a default that assumes a regime, a silent fallback, a code path that drops a contribution. A caution is a pointer ("watch out for X here"), not a claim about the runtime outcome.

What each phase records

• Facts — input-independent and statically verifiable: files scanned, verbatim content, file:line citations, what the code computes as readable from source, and cautions grounded in a visible default. Cache at confidence. A study page is a cited source record of your slice, not advice. One exception — a version-dependent numeric answer (a default value, an exact count) is not a cached fact: record its registration site and the one-line recipe to read it; the number is read fresh at need, never stored (see ma-wiki-write's "Cache the lookup, not the value").
• Probe-candidates — any runtime prediction or open question a phase raises (a comprehension prediction "this should produce …"; a gap "can't tell from source whether …"; a caution's runtime consequence "this may silently give …"). A runtime claim is never cached as fact from reading alone.