MadAgents

Install the MadAgents agent system into a folder on this machine, to be run without a container. Use when the user wants MadAgents outside Apptainer — in a project folder, on a login node, or anywhere they already have MadGraph — or wants to refresh an existing install.

Find documentation problems that caused agent mistakes — gaps, inaccuracies, or ambiguities. Uses grade context to guide diagnosis.

Edit MadGraph documentation with parallel review: style, quality, and factual verification. Revises until all checks pass or max rounds reached.

Generate evaluation questions about MadGraph and related tools with verified reference answers, using web research for real-world use cases.

Get a writable copy of the MadGraph documentation.

Run one iteration of the doc improvement loop: generate or accept questions, answer them in parallel, verify, grade, diagnose, fix the docs, and re-evaluate until convergence.

Extract factual claims from text and verify each one using execution, source inspection, or physics reasoning.

Install default-mode MadAgents (the multi-agent HEP setup — orchestrator, workers, reviewers) into a target folder/repo, for either Claude Code or Codex. The agent runs in the user's repo directly (bare mode). If MadAgents is already installed there, refers the user to update-madagents. Use when the user wants to set up MadAgents in a specific directory.

Update an existing MadAgents install in a repo to the current source version, preserving the user's edits via a 3-way merge (base = the version they installed, current = their files, new = the current version). Reconstructs the original from git history — no local snapshot. Use when MadAgents is already installed and the user wants to update it.

Audit this agent system's own harness — its agents, skills, rules, and learned memory — against your past sessions and a corpus of agent-design principles, then recommend structural changes (add / remove / adjust a consultant, skill, or rule). Human-gated: recommends and discusses first, snapshots before editing, applies only on explicit approval, and runs a reviewer panel on the change. Use when asked to audit, doctor, tune, or improve the setup itself — not for a physics or MadGraph task.

Harden the agent system against past mistakes — point it at evidence of a failure (`/ma-reflect <previous runs | good and bad example pairs | a recurring mistake you describe>`) and it tightens its own behaviour so the mistake does not recur. The lead reads the evidence as a whole, abstracts the failures to general mechanisms, and installs a behavioural discipline for each — on the lead, the owning consultants, the runtime probe, or a combination — in whatever form fits (a sharpened operating principle, a check, a routing refinement, a wiki page, a new skill, and so on). Every discipline must change behaviour rather than restate what the surface already carries, is scoped to the condition under which it fires, and is confirmed real before it is written. User-invoked, not auto-invoked.

Periodic tidy pass on the wiki/MEMORY tree at `.madagents/wiki/` — structural checks, merge near-duplicates, split overgrown pages, balance the MEMORY/wiki tiers, prune; plus a frontmatter scan sub-operation. Reorganizes already-recorded content; it does not source-walk, run MadGraph, or author new principles. User-invoked; do not auto-invoke at session start or end.

Record a finding into your own wiki/MEMORY, or consolidate your notes. Invoke to cache something you source-walked for THIS input that is non-obvious and worth a future dispatch (signal = genuine learning, NOT "might be useful"); or to create / extend / supersede a page, merge / split / generalize across pages, or move a lesson between MEMORY.md and the wiki. Writes your OWN subtree only (per-agent: consultants/<your-name>/; lead: lead/).

Heavy adversarial verification of an assembled MadGraph simulation-spec — an independent, slice-routed claim panel plus a runtime probe that checks the setup by running it, with bounded revision. Use ONLY when the user explicitly asks for deep verification ("verify deeply", "go deep", "check carefully"). Do not auto-invoke.

Adversarial runtime probe of a candidate MadGraph setup — write well-formed expectations about what should be true if the spec is correct, dispatch the `ma-probe` agent, read per-expectation results. Use any time you're about to run MadGraph to test a candidate — `/mg-deep-verify` Stage 4, a sanity check after `/mg-setup` builds, a user-asked "does this run cleanly", or debugging a failed launch. Skip when no runtime question exists (pure source-walk questions belong to the consultants).

Set up MG5_aMC for a physics request end-to-end — the default path for any task that builds or configures a MadGraph process or run. Use whenever the deliverable is an MG5 setup: the process line, model, parameters, cuts, scales, LO/NLO mode, decay chains, EFT orders, and so on. Frames the request, classifies the physics regime (which surface keywords routinely under-specify), builds the candidate via consultant dispatches, and reconciles against the physics-spec. Skip only for a pure source-mechanics lookup (dispatch the consultant directly) or a quick factual answer.

Warm-up by phased source scan — study a topic of MadGraph (`/mg-study NLO`, `/mg-study decays`) or, bare, each consultant's core area, ahead of any task. Dispatches the relevant consultants to scan their slice in four phases (explore → physics → gaps → cautions), caching facts-with-provenance and flagging probe-candidates; cheap probes confirm comprehension inline, expensive ones are proposed for confirmation. Refines per-slice routing in the agent descriptions and cross-slice seams in the lead wiki. User-invoked warm-up between tasks; not auto-invoked, does not answer physics questions (that is `/mg-setup`).