Glossary

audit chain: A sequence of verifiable actions and decisions that trace the evolution of an agent's work, ensuring accountability and transparency throughout the autonomous coding process.

broker: The kew component responsible for managing git ceremonies, including creating branches, committing changes, and handling pull requests during agent execution.

dispatch: The process of initiating an agent to begin work on a task, typically triggered by a command or event that starts the agent's execution loop.

epic: A large, complex task that is broken down into smaller, manageable subtasks or issues, often tracked through a dedicated branch or workflow in kew.

loop: A continuous execution cycle where an agent performs planning, execution, and evaluation steps until a task is completed or a stopping condition is met.

rescue: When a run hits its max-steps limit with coherent work, kew commits and pushes it and opens a flagged draft pull request instead of discarding it.

runner: The component that executes the actual code changes and agent workflows, typically interfacing with external tools and services to perform the requested tasks.

worktree: A Git concept used by kew to manage multiple working directories for different branches or tasks, enabling parallel development and isolation of changes.

budget cap: A limit on the computational resources or time that can be consumed by an agent's execution, preventing excessive resource usage.

agent PR: A pull request created by an agent to propose changes to the codebase, typically containing the results of its autonomous coding work.