auto-research

Turn a completed research run into a venue-targeted paper package with traceable claims, verified citations, reviewer-style self-critique, and publishable LaTeX artifacts. Use after auto-research-execution when `results.csv`, `results_summary.json`, and `run_report.md` exist and the user wants a full paper draft, revision loop, or negative-result framing. Read the target venue profile and use the fetched official LaTeX template first when available. Do NOT use for early-stage ideation, method design, or ad-hoc English polishing detached from experiment artifacts.

Implement and run the experiments specified in a Stage 2 experiment plan, producing logs, results.csv, and a run report that Stage 4 (writing) can consume. Generates a Sakana-AI-Scientist-style template repo (experiment.py + plot.py + configs/), trains/evaluates with mixed precision and proper checkpointing, captures every metric to WandB or local TensorBoard, and runs a self-healing debug loop (capped at 5 attempts) for failed runs. Use when an experiment plan exists and needs to be executed end-to-end. Do NOT use for ad-hoc one-off scripts (use a normal coding session) or to design experiments (route to auto-research-method).

Generate an evidence-anchored author response to peer reviews of a Stage-4 paper (OpenReview / CMT / pasted markdown). Decomposes each review into atomic concerns, classifies each by severity and axis, anchors every claim back to artifacts that already exist under runs/<id>/ (results.csv, results_summary.json, literature_pool.json, run_report.md, paper.tex, lint_report.md), then drafts a rebuttal whose stances are drawn from REBUT-WITH-EVIDENCE, CONCEDE-AND-PATCH, OUT-OF-SCOPE, NEW-EXPERIMENT-NEEDED. CONCEDE stances must produce a synchronized paper.tex patch and a revision_diff.md entry — no silent edits. Use when paper.tex exists and the user has reviewer comments. Do NOT use to write the original paper (route to auto-research-writing) or to design net-new experiments without reviewer pressure (route to auto-research-method).

Orchestrate a fully-autonomous CS/AI research pipeline from a broad topic to a venue-targeted paper draft. Start by asking the user which conference or journal they want to target, fetch the official LaTeX template and call-for-papers requirements from official sources, analyze what kinds of contributions that venue rewards, then run the closed loop: literature mining, hypothesis design, code execution, and LaTeX paper drafting. Delegates each stage to a specialized sub-skill (auto-research-ideation, auto-research-method, auto-research-execution, auto-research-writing) and enforces state hand-off, hallucination guards, and human-in-the-loop checkpoints.

Generate venue-targeted research ideas in CS/AI (LLMs, CV, RL, multimodal, systems-for-ML) by mining real research gaps from OpenReview reviewer comments and recent arXiv/Semantic-Scholar papers, then running a STORM-style multi-perspective debate to surface non-obvious angles. Reads the target venue's CFP and review emphasis first, then outputs 3 candidate ideas with verified citations, a feasibility-under-budget score, a novelty score against the SOTA, and a venue-fit score. Use when the user provides a domain and wants research directions; do NOT use for literature-only summaries (route to a search skill instead) or for picking from already-narrowed ideas (route to auto-research-method).

Convert a research idea (from auto-research-ideation) into a rigorous, mathematically-precise method specification and a reviewer-defensible experiment plan. Outputs (1) method.md with notation, formal definitions, derivations, and algorithm pseudocode, and (2) experiment_plan.yaml with datasets, baselines, metrics, ablations, seeds, and a falsifiable hypothesis. Use after Stage 1 (ideation) for any CS/AI research question that will need top-tier-conference-grade theoretical and empirical scaffolding. Do NOT use to polish an already-existing method (route to research-paper-writing) or for ad-hoc demo experiments.