graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
v0.0.250 (2026-08-12)
This release makes OpenAI server-side compaction safe across both first-party host and authentication paths, while restoring useful tool visibility in the REPL.
- Canonical compact history. Graff installs the entire validated
/responses/compactoutputarray, rather than retaining only its newest opaque item. Installation is transactional: missing, malformed, or empty output preserves the original history and uses readable local compaction as the fallback. - Strict first-party boundary. Server compaction is limited to Responses
requests from the exact
codexandopenaiproviders. Non-OpenAI providers continue to use local compaction. - Official Platform Responses route. The
openaiprovider now targetshttps://api.openai.com/v1/responseswith ordinary bearer authentication. ChatGPT backend identity headers, Codex WebSocket chaining, and Codex-only priority behavior remain exclusive tocodex. - Readable REPL tool rows. Tool starts and results appear as stable terse rows with short command, path, URL, description, or result previews. This avoids raw JSON noise and the output corruption risk of cursor-up rewriting during parallel work.
The implementation was exercised with controlled continuation replays on both
first-party OpenAI routes: 9/9 perfect trials through the ChatGPT/Codex backend
and 12/12 through the official Platform across gpt-5.6-sol,
gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6. The long fixture preserved 28
paired function calls and outputs; invalid remote responses were separately
verified to retain the original local history.