Why coding agents lose context
A repository tells an agent what the code is, but not necessarily what the project is trying to achieve, which files are sensitive, which tests are mandatory or what failed in the last attempt. Those details often live in temporary chat history or in a human's head.
BranchSlate turns that knowledge into a project-level contract. The context is versioned and available through the product and API, so different agents and later sessions can begin from the same current state.
What belongs in persistent project context
BranchSlate context includes the current objective, architecture, important files, coding conventions, protected areas, required tests, active work, recent changes, known failures and deployment expectations.
That is intentionally different from dumping an entire conversation into a prompt. The goal is a small, durable set of facts that helps an agent make better decisions while keeping the repository and Git history authoritative for source code.
Context that stays connected to the work
Tasks carry acceptance criteria, relevant files and required tests. Proposed changes remain attributable to the agent that made them, verification records attach to exact revisions and accepted work becomes normal Git history.
The result is a handoff model that works across AI sessions, devices and collaborators without making the AI conversation itself the system of record.
Build with the AI you already use.
Start with a private Git repository, connect your coding agent and keep the project context, tasks, verification and review history together.
Private Git repositories · scoped agent access · configurable approval · reversible history.