Retry Loop

chimera.core.loops.retry wraps any inner loop with retry + scoring semantics. After each attempt the result is scored; if the score meets a threshold the loop stops early, otherwise a fresh attempt is started with feedback from the previous failure. The best attempt (by score) is returned.

Inspired by SWE-Agent’s AbstractRetryLoop / ScoreRetryLoop.

Key Classes

Class	Description
RetryLoop	Retry wrapper — runs an inner loop up to N times, selects the best attempt
RetryAttempt	Record of a single attempt with attempt, result, and score fields

Quick Start

from chimera.core.loop import ReAct
from chimera.core.loops.retry import RetryLoop

loop = RetryLoop(
    inner=ReAct(max_steps=20),
    max_retries=3,
    success_threshold=1.0,
)

result = loop.run(provider, tools, context, env)
print(f"Best score: {max(a.score for a in loop.attempts)}")

Custom Scorer

By default, the scorer returns 1.0 for success and 0.0 otherwise. Pass a custom scorer for finer-grained evaluation:

def test_score(result):
    """Score based on how many tests pass."""
    if "5 passed" in result.output:
        return 1.0
    elif "3 passed" in result.output:
        return 0.6
    return 0.0

loop = RetryLoop(inner=ReAct(), scorer=test_score, max_retries=5)

Retry Context

On each retry, RetryLoop builds a fresh context containing the original user message plus feedback from the previous failure. This gives the inner loop a chance to try a different approach without being influenced by the previous conversation history.

Integration with LoopConfig

Pass a LoopConfig to forward permissions, events, and detection to the inner loop:

config = chimera.LoopConfig(permissions=policy, events=bus)
loop = RetryLoop(inner=ReAct(max_steps=20, config=config))

Import Reference

from chimera.core.loops.retry import RetryLoop, RetryAttempt

Related

• Loops — overview of all loop types
• Plan/Act Loop — two-phase plan then execute
• Lint Feedback Loop — linter-driven iteration