NanoResearch AGENTS.md

NanoResearch — Codex Integration Mode

NanoResearch is an end-to-end autonomous AI research engine. In Codex integration mode, Codex should drive the existing pipeline rather than inventing a second workflow.

Core Goal

Given a research topic, NanoResearch should produce a resumable research workspace containing:

• literature artifacts
• planning artifacts
• runnable experiment code when needed
• execution or literature-analysis evidence
• figures
• a LaTeX paper draft
• review output and final exported assets

Pipeline

NanoResearch uses a 9-stage pipeline:

IDEATION -> PLANNING -> SETUP -> CODING -> EXECUTION -> ANALYSIS -> FIGURE_GEN -> WRITING -> REVIEW

Stage meanings:

• ideation: literature search, gap finding, hypothesis or theme extraction
• planning: experiment blueprint or survey blueprint generation
• setup: environment and resource preparation
• coding: runnable experiment generation
• execution: local or SLURM-backed experiment execution
• analysis: structured evidence extraction from outputs
• figure_gen: figure generation for paper assets
• writing: LaTeX paper drafting
• review: critique, verification, and revision

Workspace Convention

Workspaces live under ~/.nanoresearch/workspace/research/. A typical workspace contains:

{session_dir}/
├── manifest.json
├── papers/
├── plans/
├── experiment/
├── drafts/
├── figures/
├── output/
└── logs/

Reuse an existing workspace when the user asks to continue, inspect status, resume, or revise a prior run.

Paper Modes

Topic prefixes:

• original: Topic -> original_research
• survey:short: Topic -> survey_short
• survey:standard: Topic -> survey_standard
• survey:long: Topic -> survey_long

Behavior:

• original research follows the full 9-stage pipeline
• survey modes skip experiment-heavy stages and use literature-grounded planning, writing, and review
• the prefix is parsed by the existing CLI and manifest logic; Codex should reuse that behavior