CausalMerge: Evidence-Driven Merge Repair

Modern Git workflows detect textual conflicts, but they often miss semantic conflicts: two branches merge cleanly while their combined behavior violates a business requirement. These failures are difficult to diagnose because each branch may pass independently, leaving reviewers to guess which interaction caused the regression. CausalMerge is an evidence-driven investigation and repair workflow for software teams. It captures a compact Incident Bundle containing the failure, relevant Git changes, test output, and business contracts. IBM Bob analyzes this bounded evidence, coordinates specialized investigation perspectives, and produces ranked, falsifiable hypotheses. A deterministic replay engine then tests BASE, Change A, Change B, and A+B in isolated Git worktrees. Targeted counterfactual interventions distinguish correlation from causation and produce a machine-readable causal proof. For our MiniCommerce demonstration, the baseline and both branches pass independently, while the merged A+B revision fails Guest Checkout with HTTP 401. Counterfactual replay confirms the precise interaction between global authentication hardening and a newly introduced anonymous checkout route. CausalMerge then proposes the smallest evidence-supported repair, preserves protected endpoints, validates A+B+FIX at runtime, and keeps the historical failure evidence immutable. The target users are developers, reviewers, release engineers, and engineering leads working with parallel feature branches. They interact with a web dashboard that explains the incident, branch lineage, hypotheses, experiments, confirmed mechanism, proposed patch, preserved behaviors, validation results, and provenance. Human approval remains mandatory before repair generation and release. What makes CausalMerge unique is its separation of reasoning, deterministic execution, and approval. It does not present an AI guess as proof. Every conclusion is connected to reproducible Git revisions, executable predicates, counterfactual experiments, artifact digests, and explicit business intent. The deployed web application serves approved, prerecorded capsules and does not depend on IBM Bob at runtime, allowing the production experience to remain reliable after the development session ends.