Skip to content

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

Python
1.2 KB
Raw Blame

v0.302.48

  • oMLX is a built-in local provider. --model omlx talks to the oMLX app's OpenAI-compatible server on 127.0.0.1:8000 from any directory, with no .graff/.config.router. Store its key with graff key set omlx <key> or set OMLX_API_KEY. graff models lists the models oMLX serves, a bare --model omlx starts on the first one, /effort and ACP thought levels are sent as reasoning_effort, and calls are billed as local (#1695).
  • Local model windows come from the server. Live model catalogs read max_model_len, which oMLX and vLLM-style servers report, so a local model gets its real context window instead of the 200k fallback.
  • A stopped local server is named. When a loopback server refuses the connection, the failed turn ends with the URL it tried.
  • Compaction counts opaque state by the provider's own measure. Compaction state was estimated at bytes / 16, so a large blob could read as several times its real size and keep the context over the compaction point, compacting every turn (#1686). After a full request that carries the state, its share of the provider's reported input replaces the estimate for that state.
  • Retains all changes from v0.302.47.