Md Rubayet Islam

Coding agents now produce plausible code patches faster than software teams can review them. The industry's primary bottleneck is no longer generating fixes, but deciding which ones are safe to merge without human review. Keel is an autonomous code-repair agent that knows when it should not be trusted. Built as a zero-infrastructure GitHub Action and cloud platform (keel.boraqs.com), Keel monitors CI failures, samples candidate fixes using Meta's Muse models, and provides a mathematically grounded statistical guarantee on autonomous pull-request merges. How Keel works: 1. Multi-Model Generation: When CI fails, Meta's Muse Spark (via Meta Model API) and Muse Spark Open (open-weights) independently propose candidate patches. 2. Adversarial Cross-Examination: Muse Spark generates adversarial unit tests targeted specifically at boundary conditions where candidate diffs disagree. 3. Code-Aware Nonconformity Scoring: Keel extracts 20 code-level signals, including test outcomes, adversarial survival, diff size and AST locality, static-analysis deltas, cross-model consensus, and token-level logprob uncertainty. 4. Conformal Risk Control (CRC & Learn-then-Test): Calibrated on held-out bug-fix data, Keel calculates a distribution-free threshold. At a user-configured risk tolerance α (e.g., 10%), the empirical rate of incorrect auto-merges is mathematically guaranteed to stay below α. Patches meeting the bound receive an AUTO-MERGE label; ambiguous patches receive a NEEDS REVIEW label with an explanation pinpointing why the agent abstained. Across SWE-bench Lite and QuixBugs, Keel strictly honors its risk bounds while naive heuristics fail.