Skip to content

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

Python
1.7 KB
Raw Blame

Codegraff v0.0.302.25

A compaction hotfix for ChatGPT and OpenAI sessions, plus fewer requests on hosted models.

Long ChatGPT and OpenAI sessions stop compacting on every step

On the ChatGPT-plan sign-in, the OpenAI API and Codegraff's GPT models, a long session could start compacting before every tool call and never stop: each step waited on a full compaction request, and the context meter showed more than the window. The meter was reading the provider's encrypted compaction state as if it were text. That state is far denser than text, so a session anchored on it looked full on every step, and each new compaction produced state of the same size. The meter now counts that state at its own rate, so the session compacts only when the window is actually full (ADR 0259).

These routes now compact only with the provider's own compaction. If it fails, you see the error instead of a silent local summary that would discard state the provider still holds. GRAFF_SERVER_COMPACT=0 still switches back to the local summary.

Fewer requests on hosted models

  • One hosted model family takes fewer requests per task: dependent shell steps chain, a file just written is not read back, and the completion can share the final response.
  • On the stateful hosted chat route, the project instructions and layout are sent once per conversation instead of with every request.
  • An empty tool result is sent as "(no output)", so one empty result no longer makes every later request in a hosted chat conversation fail.

MCP

A slimmed comment list keeps the issue its rows belong to, so results fetched in parallel say which issue they describe.