Skip to content

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

Python
1.4 KB
Raw Blame

0028. Codex session_id is the prompt_cache_key

Status: accepted 2026-08-25

Context

openai/codex ModelClient::prompt_cache_key defaults to CodexResponsesMetadata.session_id. The same string is sent as the session_id header (build_session_headers). A hit needs that key plus an identical instructions/tools prefix.

Graff already sent a durable cwd prompt_cache_key (ADR 0009). The ChatGPT session_id header was a per-process UUIDv4. Trajectories of a live Codex session (gpt-5.6-sol) still recorded multi-million cache_read_tokens on later calls in a turn, and 0/0 on turns whose usage never landed in g_cost. The first-party client never splits those two ids.

Decision

  • Codex session_id (HTTP extra headers and the WS handshake) is the same value as the Responses body's prompt_cache_key (requestCacheKey: project root for main//btw, four role lanes for children; ADR 0223: the account base when the codex provider has an account id).
  • sessionId() stays a per-process UUIDv4 for the persisted graff session record. It is not the ChatGPT cache partition.
  • store: false is unchanged (openai/codex also sends store: false).

Cost

ChatGPT session telemetry now groups by project/lane rather than process. That is the cache contract. Do not restore a random session_id header without measuring a live cached_tokens regression.