Graft is an industrial-grade context governor for coding agents where code structure and causal activity are unified.
mindmap
root((Graft))
Policy Integrity
Read Governance
Structural Outlines
Reason-Coded Refusals
Structural Memory
WARP Graph
AST Evolution
Commit-Level History
Causal Provenance
Strand-Scoped History
Activity Explanations
Admitted Observations
Agent-First Legibility
Versioned Schemas
Machine-Readable Decision
Deterministic Replay
A strict read-governance layer that ensures agents consume only what they need. Large files are degraded to structural maps; banned files are refused with explicit, actionable reasons.
A rendering foundation built on WARP (Structural Worldline Memory). Graft tracks the evolution of symbols, files, and relationships across the repository's entire history. Graph reads must never assume the full graph fits in memory — all queries are bounded by traversal scope, tick receipts, or substrate-side filtering. As the graph grows, read cost must stay proportional to the question, not the graph.
The "why" behind structural changes. Graft tracks observations, stages, and transitions to explain the evolution of the codebase between hard Git checkpoints.
The system is designed to be codable and inspectable by both humans and AI. All outputs are versioned, machine-readable JSON payloads with embedded receipts and schemas.
Session management and budgets are substrate properties. Graft tightens read caps as the context window fills, ensuring agents remain focused and cost-effective.
The goal is not just smaller reads. It is the geometric lawfulness of the repository as a provenance-aware medium for agentic work.