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

0190. Catalog GETs share bounded HTTP/2 transport

Status: accepted 2026-09-23

Context

Dynamic model catalogs are startup inputs. Some providers paginate them, and a later page can fail after earlier pages were valid. The old HTTP/1.1 fetch followed redirects and accepted unbounded response bodies. The small response limit for optional JSON POSTs in 0189 cannot accommodate complete catalogs with verbose model metadata.

Decision

Use the same exclusive HTTP/2 pool and joined deadline for catalog GETs. Keep the page walk and its valid-prefix/cache fallback in the catalog layer. Bound each page at 16 MiB and reject larger responses before parsing. A page larger than the picker response is valid; the transport limit must not silently turn it into a missing catalog. Read the status before the body on both HTTP/2 and HTTP/1.1, including failed responses whose bodies stall or exceed the cap.

Follow up to three GET redirects within the same deadline on either protocol. Keep catalog headers on same-origin redirects; send only non-credential Accept on a scheme, host, or port change. HTTP/1.1 fallback is allowed only before an HTTP/2 request is sent. The transport allocator owns pooled sessions until shutdown; the completed page is copied to the result allocator.

Consequences

Catalog discovery can reuse an HTTP/2 connection across pages without making startup depend on a partial response or unbounded body. Each concurrent fetch may temporarily reserve its page limit. A response beyond 16 MiB falls back to the existing cached or baked catalog rather than activating a truncated list. The offline TLS fixture covers redirects, credentials, pagination, large valid pages, limits, status-first failures, and cancellation.