AGENTS.md — RigorPilot Skills

Guidance for any coding or research agent (OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI, Claude Code, and other AGENTS.md-aware tools) working in or with this repository. Skills here follow the Agent Skills open standard; each skills/<slug>/SKILL.md is the canonical, model-agnostic contract.

What this project is

RigorPilot Skills is a research-first Agent Skills repository for deep learning experiments. It keeps AI-assisted research grounded in comparability, reproducible evidence, and auditable changes while an agent reproduces, improves, or explores a research repository. The flagship output is the annotated README: repro_outputs/ANNOTATED_README.md replays the target repository's README verbatim with a color-coded, evidence-linked account of what the agent did in each section.

Using the skills on a research repository

• Default to the trusted lane. Ambiguous requests route to reproduction, setup, conservative execution, training verification, analysis, or safe debugging — never to exploration.
• Read skills/<slug>/SKILL.md before acting; it defines fit, boundaries, workflow, and the output contract for that skill.
• Entrypoints by intent:

• In the explore lane, follow references/research-thinking-loop.md: a greedy, evidence-anchored cycle (observe → ground → hypothesize → single-variable design → bounded run → fair compare → keep or roll back → record).
• Consult ~/.rigorpilot/PERSONAL_RIGOR.md if present, under references/continuous-learning-policy.md — advisory only; it never relaxes rigor gates and the repository always wins on conflict.
• Write evidence bundles to the documented output directories (repro_outputs/, train_outputs/, analysis_outputs/, debug_outputs/, explore_outputs/). Never claim success without recorded evidence.
• Never make silent semantic changes to a target repository; record every assumption, deviation, and blocker.
• Exploration is candidate-only and requires the researcher's explicit authorization plus a durable current_research anchor.

Developing this repository

• Validate before committing:

python scripts/validate_repo.py
python scripts/test_skill_registry.py

• During local iteration, python scripts/run_all_tests.py --core runs the high-signal reproduction/install subset. Before commit/release, the full regression suite is every scripts/test_*.py file; all must pass.
• Conventions: public SKILL.md files stay ≤ 130 lines; the skill registry (references/skill-registry.json) stays in sync with skills/*; scripts run on both Windows PowerShell and Linux shells; human-readable outputs are bilingual (en/zh) where the writers support it.
• Shared bundle writers live in shared/scripts/; five skills load them at runtime, and the installer ships them alongside the skills.

Hard rules

1. Comparability first: if evaluation conditions change, say results are not directly comparable.
2. No score chasing without explanatory value; no novelty claims without evidence.
3. Engineering fixes must not be disguised as research contributions.
4. Important changes must be auditable, reversible, and explainable.