Agent Claude Code
Engage after every install or upgrade, once the manifest has been written — the installer dispatches you to check the result before reporting it to the user. In slice: run installer.py verify, read its output, then judge what a checksum cannot — whether the folder is a working install for this user's setup, whether…
Agent Claude Code
Extracts individual verifiable factual claims from text. Outputs a JSON array of claim objects.
Agent Claude Code
Selects which newly verified claims are worth caching in the persistent claim database. Filters out inconclusive and already-known claims. Designed for haiku.
Agent Claude Code
Matches newly extracted claims against a known-claims database. Returns IDs of relevant known claims to skip redundant verification. Designed for haiku.
Agent Claude Code
Edits MadGraph documentation files based on instructions. Knows the doc layout, design principles, and file conventions. Does not review or verify — only edits.
Agent Claude Code
Reviews MadGraph documentation changes for quality: right file, duplication, scope, generality, and coherence. Checks against the flat doc layout and design principles.
Agent Claude Code
Reviews MadGraph documentation changes for style and formatting consistency. Checks markdown structure, terminology, tone, naming conventions, and cross-reference style.
Agent Claude Code
Grades an agent's answer using two-level grading: a primary grade (CORRECT/INCORRECT/INCONCLUSIVE) plus zero or more tags (hasmistakes, inefficient).
Agent Claude Code
The primary expert on MadGraph and related tools (e.g. Pythia8, Delphes, MadSpin), with access to authoritative local documentation. Specialized in MadGraph5aMC@NLO: process definition, event generation, shower/hadronisation, detector simulation setup, generation/prediction-level studies.
Agent Claude Code
Full MadAgents orchestrator — delegates all work to specialist agents (madgraph-operator, physics-expert, etc.) and ensures quality via reviewers. Use as agenttype when spawning teammates that should behave as independent MadAgents instances.
Agent Claude Code
Specialized in reading, summarizing, and extracting information from PDF files.
Agent Claude Code
Particle physics reasoning — e.g., theory, phenomenology, simulation setup validation.
Agent Claude Code
Specialized in generating clear, well-formatted plots and plotting scripts for a physics audience. Iteratively refines plots until they meet quality standards.
Agent Claude Code
Evaluates whether deliverables are well-presented and ready for the intended audience. Has built-in quality standards.
Agent Claude Code
Deep web research: finding, cross-checking, and synthesizing information across multiple sources. Returns structured factual overviews with explicit source citations.
Agent Claude Code
Specialized in bash and Python. Can inspect/modify the environment and create/edit/execute scripts.
Agent Claude Code
Critically evaluates worker outputs for correctness, completeness, and soundness of reasoning and evidence. Can create and run verification scripts. Supports two review intensities (specify in instruction): quick check (default; plausibility) or thorough (active verification).
Agent Claude Code
MadAgents orchestrator specialized for claim verification — extracts factual claims and verifies each one using execution, source inspection, or physics reasoning. Restricted to verification-relevant subagents only.
Agent Claude Code
Engage when the symptom is in the GENERATED helicity routines — ALOHA fails on a model's Lorentz structure (exotic/high-spin, custom lorentz.py), Source/DHELAS will not compile, or an amplitude value looks wrong; when you must READ or CALL exported ME routines — arg layout, ±IC(i) nsf/nsv signs; when the EMITTED…
Agent Claude Code
Engage when an NLO run's ACCEPTANCE, NORMALISATION, BEAMS, SHOWER, ACCURACY or WEIGHTS are at stake. Co-dispatch nlo-syntax: it owns the bracket, I own the run. In slice: the NLO run card — a DIFFERENT class from the LO card, own names/defaults: IR-safe cutting (jets clustered BEFORE cuts, dressed leptons); the…
Agent Claude Code
In slice: scan the assembled spec + per-consultant summary against MadGraph's pipeline. Identify code paths / pipeline stages the consultant trail did not cover for this input. Return a structured list of detected blind spots, each with a recommended consultant to dispatch. Common redirects (non-exhaustive): actually…
Agent Claude Code
Engage when a propagator is forced OFF its mass pole and the on-shell window decides whether events survive: a parent below its daughters' mass sum, a daughter pinned off-peak by a forced decay — the default +-bwcutoffGamma window then silently cuts sigma to zero, so bwcutoff must be DERIVED per leg…
Agent Claude Code
Engage when the decay is written INTO the process line. X > a b inside generate restricts the amplitude to that resonant topology — not the inclusive same-final-state ME — and forces the mother's BW. Also: a 2+-level cascade whose sub-decay is SILENTLY DISCARDED unless parenthesised ("it parsed" != "it bound"); an…
Agent Claude Code
Engage when the COLOUR STRUCTURE itself is the question — the number of independent colour structures, the colour matrix / factors of the squared ME, leading- vs full-colour; when a model carries an exotic colour rep (sextet, epsilon/diquark, colour-octet); when a colourless/EW-only process should collapse to a…