collect

name	collect
description	Collect and structure research papers, projects, and methods related to coding agent training. Use when user says "collect papers", "找论文", "search methods", or wants to build a method registry.
argument-hint	["research-query"]
allowed-tools	Bash(*), Read, Write, Grep, Glob, WebSearch, WebFetch, Agent

Collect — Research Paper & Method Harvester

Collect and structure research on: $ARGUMENTS

Pipeline

1. Search — Query arXiv, Semantic Scholar, and GitHub for recent papers/projects on the given topic
2. Filter — Keep only papers relevant to coding agent training (SFT, RL, trajectory data, evaluation)
3. Extract — For each paper, extract a structured method card:
   • Paper metadata (title, authors, venue, date, arXiv ID)
   • Method name and category (data, training, reward, eval, infrastructure)
   • Key innovation (1-2 sentences)
   • Composable atoms (what building blocks does this method introduce?)
   • Dependencies (what does it require? models, data, compute)
   • Reported results (benchmarks, metrics, baselines)
4. Register — Append method cards to recipes/registry/method_atoms.json
5. Report — Output summary of collected methods with novelty assessment

Output Format

Write method cards to recipes/registry/method_atoms.json following the atom schema. Print a summary table to the user showing: paper | method | category | key innovation.

Constraints

• Focus on 2025-2026 papers unless user specifies otherwise
• Prioritize: SWE-bench results, agentic coding, trajectory data, RL for code
• Skip survey papers unless specifically requested
• Verify arXiv IDs are real (use arxiv_fetch.py tool if available)