agentic-jujutsu

Create a separate JjWrapper instance per agent and run their operations concurrently with Promise.all. The underlying jj-based engine is lock-free, so agents commit, branch, and rebase simultaneously without waiting on locks, and most conflicts resolve automatically.

Call startTrajectory with a task description, perform operations, call addToTrajectory, then finalizeTrajectory with a success score between 0.0 and 1.0 and an optional critique. Later, getSuggestion returns recommended operations with confidence scores for similar tasks.

Agentic Jujutsu is lock-free and designed for concurrent agents, reporting roughly 23x higher concurrent commit throughput and 87% automatic conflict resolution versus Git's lock-based model. It also adds operation learning and quantum-resistant fingerprints that Git does not provide.

Validation requires a non-empty task description under 10KB, at least one recorded operation, and a finite success score between 0.0 and 1.0. Empty contexts or out-of-range scores throw validation errors, so clamp scores and record operations before finalizing.

Low confidence means the ReasoningBank lacks sufficient data for that task type. Check getLearningStats for total trajectories and record at least 5-10 trajectories, including several successful ones, before expecting high-confidence suggestions or discovered patterns.