graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
v0.302.69
Peer mail stops rewriting history that was already sent, Kimi international accounts sign in on kimi.ai, workers can use gateway models, and graff harness-tree shows the harness as one value.
- Reading peer mail keeps the conversation intact. After a turn read its inbox, graff removed every peer wake line from the whole history, including ones the model had already been sent. That rewrote history the provider had cached: on Claude Agent every woken turn started a fresh Claude Code process, and on other providers it changed the cached prefix. Only wake lines the model was never sent are removed now. (#1799)
- Kimi sign-in works for international accounts.
graff login kimisigns international accounts in on kimi.ai and mainland China accounts on kimi.com, picking the site from your saved login or your clock and locale;graff login kimi.aiandgraff login kimi.comchoose it outright, and token refresh goes back to the site that issued the login. Earlier logins keep refreshing on kimi.com. (#1733) - Sign-in links are copied for you. In a terminal,
graff loginputs the link on the clipboard (over SSH, on your own terminal's clipboard). Over SSH and withGRAFF_NO_BROWSER, graff no longer opens a browser you can't see. (#1733) - Workers can use gateway models and name their own provider.
--subagent-model <model>with--subagent-provider codegraffaccepts a model the gateway serves even when the built-in catalog does not list it, a persona pinned withmodel: <provider>/<model>runs on that provider, and a worker's calls on another provider report their cost. (#1769) graff harness-tree show|digest|check|apply|diffreads the files graff loads to decide how it works (the instruction file, skills, subagents, scripts and hooks) as one tree you can inspect, compare and apply. (#1770)- MCP over HTTP: cancelling a request retries its notice when the machine is briefly out of local ports, instead of dropping it. (#1796, #1794)
- OpenAI 12-hour prompt cache, opt-in. On the direct OpenAI Responses route for GPT-5.6 and later,
GRAFF_OPENAI_CACHE_TTL=12h(oropenai_cache_ttlin settings) keeps the stable prompt prefix cached for 12 hours. The default is unchanged (30 minutes) until a live check confirms OpenAI accepts and keeps it. Usage now separates 12-hour and 30-minute cache writes. (#1698, ADR 0312) - ADR 0308 records that a cheaper scout worker is an opt-in persona, not a default route. (#1771)
Includes all changes from v0.302.68.