Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python
922 B
Raw Blame

v0.0.248 (2026-08-11)

A focused exact-key lookup latency hotfix.

  • One targeted file read instead of a whole-file loop. read_file now accepts an optional contains literal and returns only matching numbered lines, so keyed lookups can avoid reading a large file and then searching it in a later model round.
  • Fewer provider round trips. The root policy steers read-only exact-key requests to one targeted read while preserving byte-exact reads before edits. In five matched GPT-5.6 Luna runs, Graff moved from three model calls and two tools to two model calls and one tool, reducing its measured median from 10.10s to 5.12s.
  • Bounded and regression-tested. Literal matching streams through the file, keeps output under the existing read cap, rejects incompatible range/compact combinations, and has both unit and model-backed behavior coverage.

No GUI or general REPL behavior changes are included.