Skip to content

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

Python
1.8 KB
Raw Blame

0003. Codegraff gateway wire follows model capability

Status: accepted 2026-08-17

Context

The Codegraff gateway has one OpenAI-shaped public contract but two transport capability sets. Chat Completions accepts every alias and translates Claude and Gemini at the edge. Responses (HTTP or WebSocket) is native for OpenAI GPT-5.6+ (including Codex gpt-5.6-* and GPT-6 gpt-6-astra) and grok-4.6; Claude and Gemini return 400 there. A frozen 5.6 alias list left later GPT generations on Chat Completions, so /compact fell through to a client summarizer instead of /v1/responses/compact. A single static provider wire therefore either leaves native Responses features unused or breaks translated models.

Decision

Resolve the Codegraff provider's wire per model. GPT-5.6 and every later GPT generation (gpt-5.6, gpt-5.6-sol/terra/luna, gpt-6-astra, …) plus grok-4.6 use /v1/responses and may use its WebSocket transport. /compact on Codegraff GPT uses the standalone /responses/compact endpoint, same as direct OpenAI. Grok compaction stays on the client summary (xAI's blob path when seated on xAI). Every other gateway alias uses /v1/chat/completions.

Consequences

  • GPT-5.6+ and GPT-6 receive native Responses request shapes, prompt cache keys, streaming events, the WS-to-SSE fallback ladder, and OpenAI server compaction. Grok-4.6 gets the same Responses wire, not OpenAI compact.
  • Claude and Gemini remain on the edge-translated Chat Completions path.
  • Switching models on one Codegraff credential must rebuild both kind and endpoint. New GPT generations matching gpt-5.6* or gpt-{N>=6}* join Responses without another allowlist edit; Claude/Gemini still must not.

Evidence: Codegraff gateway contract.