Welcome to TrapArena’s Documentation!¶
TrapArena is an adversarial code-repair benchmark for coding agents (Adversarial Repair under Deceptive Environmental Feedback).
It asks one question: without a reviewer, when public tests and environmental feedback are incomplete or even misleading, can a coding agent find and fix the root cause of a bug, rather than just turning the visible tests green?
Key Features¶
Three-layer environment: a visible repository and public tests, an induction layer of misleading feedback, and hidden tests used only for evaluation.
Overt and covert bugs: public tests either expose the bug or let a superficial patch pass.
Pluggable distractors: rule strategies, five collaborator profiles, and a self-evolving attacker with a model-specific skill library.
Root-cause evaluation: hidden tests, false-fix detection, and change-focused coverage instead of “public tests pass”.
Standard-library core: runs end to end with no third-party packages.
Quick Example¶
python -m traparena.run --task examples/tasks/clamp_overt --agent gold
[clamp_overt] agent=gold distractor=none public=✅ hidden=✅ overfit=False susceptible=False false_fix=[] coverage=1.0
Note
This documentation is under active development.
Contents¶
Getting Started