Fault Localization

chimera.training.fault_localization ranks code locations by suspiciousness using test failure analysis. It parses pytest output, extracts traceback references, and scores locations that appear more frequently in failing tests using an Ochiai-inspired approach.

---

Key Classes

SuspiciousLocation

Dataclass representing a code location ranked by suspiciousness.

Field	Type	Description
file	str	Source file path
function	str | None	Enclosing function name (if identified)
line	int	Line number in the source file
score	float	0.0 (not suspicious) to 1.0 (very suspicious)
reason	str	Human-readable explanation of the score

FaultLocalizer

Parses test output and ranks suspicious locations. Filters out test files so only production code is reported.

Key methods:

• localize(test_output) -> list[SuspiciousLocation] — parse and rank
• augment_prompt(prompt, locations, max_locations) -> str — append suspected bug locations to an agent prompt

Usage

from chimera.training.fault_localization import FaultLocalizer

localizer = FaultLocalizer()

# Parse pytest output
test_output = env.run_tests().output
locations = localizer.localize(test_output)

for loc in locations[:3]:
    print(f"{loc.file}:{loc.line} — {loc.score:.0%} suspicious")
    print(f"  {loc.reason}")

# Augment an agent prompt with fault info
prompt = localizer.augment_prompt(
    "Fix the failing tests in the calculator module.",
    locations,
    max_locations=3,
)
result = agent.run(prompt, env=env)

Related

• Impact Analysis — analyze blast radius of changes
• Training concepts — strategies that use test feedback