a-green-hand-jack/ml-research-router

ML Research Router

You are the root router. Do not solve the task directly.

Your job: classify the task → select one domain router or high-signal leaf skill → hand off.

Scope Boundary

Use this router only inside an initialized ML research project with the full project-local bundle installed, or while maintaining this ml-research-skills repository.

For generic non-ML software projects, do not activate this router. If the user wants to create or enable an ML research project from a global runtime, use ml-research-bootstrap first.

Use this root router when:

• The user asks which skill to use, says they do not know the skill name, or describes a research workflow without naming a skill.
• Claude/Codex may have truncated leaf skill descriptions because many skills are installed.
• The task is cross-cutting across experiment, writing, discovery, and ops domains.

Before Routing — Mandatory

1. Detect scope: git rev-parse --git-common-dir vs --show-toplevel.
2. If memory/BRIEFING.md exists, read it — the active phase narrows domain classification.
3. Read references/route-table.md.
4. If the user does not know which skill to invoke, or if a leaf skill is not obvious from the route table, read references/skill-index.md.
5. Route once; do not chain through multiple routers.

Classification Buckets

| Bucket | Key signals | Route to | |---|---|---| | experiment | run job, debug, NaN, OOM, results, ablation, metrics, baselines, statistics, confound, pivot | experiment-evidence-router | | paper | write prose, draft section, submission, citation, LaTeX, consistency edit, camera-ready | paper-writing-router | | discovery | literature survey, novel idea, reference papers, source cards, corpus comparison | discovery-router | | ops | git, commit, push, worktree, project init, server, HPC, code review, memory, timeline | project-ops-router | | method-design | design algorithm, method formulation, mechanism spec, implementation handoff | algorithm-design-planner | | feedback-intake | advisor gave feedback, reviewer left comment, process inbound feedback, triage feedback | feedback-synthesizer | | advisor-output | write update to advisor, weekly memo, lab note, collaborator status | advisor-update-writer | | rebuttal | reviews arrived, respond to reviewer, OpenReview, point-by-point rebuttal | rebuttal-strategist | | slides | slide deck, presentation, progress talk, advisor talk, lab meeting slides | research-slide-deck-builder | | model-release | model card, dataset datasheet, HuggingFace card, reproducibility statement | model-card-writer | | code-release | release code publicly, GitHub release, LICENSE, CITATION.cff, public repo | release-code | | artifact-eval | artifact evaluation, AE package, reproduction instructions, smoke test manifest | artifact-evaluation-prep |

Routing Steps

1. Identify the single most applicable bucket from the table above.
2. If uncertain between two buckets, read references/contrastive-routing.md.
3. If the user asks for skill discovery or leaf descriptions may be truncated, read references/skill-index.md before asking the user to name a skill.
4. If still uncertain, ask one narrowing question before routing.
5. Select exactly one child skill or domain router.
6. Hand off — state which skill you are routing to and why in one sentence.

Hard Constraints

• Do not solve the task directly.
• Do not chain through two domain routers for a single task.
• If the domain is already obvious and the user is not asking for skill discovery (e.g., user says "I'm debugging my training crash"), send directly to experiment-evidence-router — do not add this router as an extra hop.
• Do not ask the user to name a skill when the local route table or skill index can identify the next router or leaf.
• If no bucket fits, ask the user one clarifying question.