Skip to content

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

Python
3.1 KB
Raw Blame

v0.0.253 (2026-08-14)

This release makes long-running OpenAI sessions safer, cuts paid-provider and startup latency, ships a production-embeddable TypeScript SDK, and replaces the desktop agent rail with focused agent control.

  • Manual /compact now uses first-party OpenAI compaction when available. Official OpenAI API-key sessions call the standalone Responses compact endpoint; Codex subscriptions request the supported in-stream compaction directive and retain only the returned opaque compaction state. History is replaced transactionally only after a valid response. A rejected or malformed result falls back to the existing local summarizer without losing history, while non-OpenAI providers remain local-only because the wire contract is OpenAI-specific.
  • Kimi K3 and startup work cut measured interactive startup from 5.85s to 2.76s. That is 53% less waiting, or about 2.1x faster. Kimi uses a durable per-project prompt-cache partition, lookup-shaped turns can route to low effort without overriding an explicit user setting, and oversized tool-result handles expose their cost in the REPL. Authenticated Kimi/router catalogs and MCP servers now fan out concurrently; interactive yolo startup can defer the join until the first request, /mcp, or shutdown. MCP cache and lifecycle state preserve deterministic config-order results despite the concurrency.
  • @codegraff/sdk is ready for production TypeScript embedding. The npm package includes the platform harness binary instead of requiring a separate graff install, and its persistent Harness/Remote APIs cover live events, inbox controls, cancellation, request correlation, and process ownership. Package-install fixtures and JSON cancellation/live tests guard the shipped artifact rather than only the source checkout.
  • Normal REPL output is quieter and diagnostics are intentional. Recovered detached recap retries stay in traces instead of appearing as raw worker noise; normal tool rows omit arguments/results while debug mode retains human- readable details. /debug shows a content-free session HUD and /usage aliases cost reporting. Shared-worktree collisions render as a structured callout with the active owner and isolation action.
  • Exact-key file lookups avoid an unnecessary full read. read_file accepts contains, streams literal matching lines with line numbers under the normal output cap, and has a behavior eval ensuring a known-file lookup uses one tool call.
  • The desktop app makes agents visible without a permanent third column. A compact agent-control pane surfaces active, waiting, and recent work from the conversation projections, with focused tests for status and navigation. The app, favicon, and generated desktop/mobile icons now share the Codegraff emblem.
  • Release reliability gets several smaller fixes. The run budget reserves the narration turn instead of exhausting itself on landing work, Codex response failures use bounded overload retries, xAI login requests match grok-build's scope/referrer contract, and benchmark fixtures validate their generated public tests before scoring.