Prune Protocol
Make AI-assisted software work cleaner
Prune gives agents and developers a repeatable contract for secure, schema-first software work — from planning and implementation through review and completion.
Why Prune exists
AI-assisted software work is fast, but speed can hide missing requirements, unsafe assumptions, and changes that are difficult to review. Prune gives the work a small, explicit contract before the first edit is made.
It is harness-agnostic. The same priorities apply whether the work is happening in Codex, Claude, an editor plugin, or a custom agent runtime.
The protocol loop
Every task starts with a compact ChangeSpec. It names the objective, allowed scope, security boundary, existing structure, behavior contract, and verification plan.
security → correctness → structure → reuse → minimal implementation → handoffThe protocol ends with a concise status: complete, blocked, question, review, or error. The handoff states what changed and what was verified.
What Prune covers
Security
Identify trust boundaries, validate inputs, protect secrets, and fail closed when a required dependency is unavailable.
Structure
Define scope, interfaces, invariants, and existing abstractions before implementation begins.
Reviewability
Keep changes narrow, behavior explicit, and completion concise enough for a human to verify.
Adapters
The canonical protocol lives in protocol/. Runtime adapters translate that source into the conventions of each agent environment without redefining it.
Next steps
- Read the README — protocol overview and local setup
- Browse the canonical protocol — schemas, policies, and task modes
- Inspect the examples — valid ChangeSpec documents
Start with protocol/ when adapting Prune to a new runtime.