Abstract:
Python has become a popular language of beginners due to its simplicity. However, syntax
errors remain as a hurdle to productivity and learning. Although automated program repair
(APR) has improved, in the majority of cases the existing solutions have difficulties in trading
the deterministic accuracy of rule-based systems with the flexibility of AI-guided models. This
paper deploys a rule-based repair agent in the form of a self-healing agent as the initial step
in a proposed hybrid repair system, concerned with the automatic detection and correction of
common Python syntax errors, including missing colons and indentation errors. The agent is
modeled upon the MAPE-K (Monitor, Analyze, Plan, Execute, Knowledge) control loop and
is tested on 37,639 real-world Python snippets with raw syntax errors that are mined on the
public GitHub repositories. The agent based on rules is able to fix 14,149 snippets with a repair
accuracy of 37.59% and an average fix time of around 0.0008 seconds. These results indicate
that a rulebased layer can efficiently handle most low-complexity syntax errors with minimal
delay, and represents a promising approach to rapid preprocessing. Meanwhile, the 62.41% of
the unresolved cases also show the constraints of the use of static heuristics in context-sensitive
or structurally complicated errors. This long abstract provides the empirical base to the rulebased
stage and gives the impetus to the subsequent stage of the project, where a learning-based
part will be added to create a hybrid repair architecture.
