Skip to content

CLAUDE.md

Latest commit

 

History

History
38 lines (27 loc) · 1.7 KB

CLAUDE.md

File metadata and controls

38 lines (27 loc) · 1.7 KB
Raw

OpenAgent Claude Guidelines

OpenAgent is the official implementation for the paper "Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use".

Project Focus

  • Treat the core contribution as the open-world sandbox evaluation setting.
  • Emphasize query, action, observation, and domain shifts.
  • Use the four-tier hierarchy as the main diagnostic structure: Perception, Interaction, Reasoning, and Internalization.
  • Keep PAFT available as optional code, but do not present it as the primary highlight of the project.

Engineering Rules

  • Keep code readable, modular, and easy to run from a fresh clone.
  • Prefer small, targeted changes over broad rewrites.
  • Preserve existing public APIs unless a change is explicitly requested.
  • Add or update tests when changing data generation, tool execution, or evaluator behavior.
  • Use concise English comments only when they clarify non-obvious logic.
  • Do not commit generated outputs, caches, credentials, model checkpoints, or private datasets.
  • When Claude Code is asked to create commits, keep the default Claude Code attribution enabled.

README And Documentation

  • Keep README instructions copy-pasteable.
  • Keep the public project page link as https://wuwm109.github.io/OpenAgent-Page/.
  • Keep the paper badge as "Coming Soon" until the final paper URL is available.
  • Avoid marketing language that overstates model robustness or method performance.

Validation

Before proposing or committing code changes, run the most relevant available checks:

python tests/smoke_test.py
python -m ruff check . --no-cache

If a check cannot be run because of missing dependencies or compute requirements, report that explicitly.