elizaos--eliza/packages/benchmarks/codex-adapter/AGENTS.md

Codex Adapter — Agent Guide

Skeleton harness bridge that lets the benchmark orchestrator run benchmarks against the Codex CLI agent, authenticated per OpenAI-Codex account (#10193/#10199). API-shaped like smithers-adapter; select it with --adapters codex and iterate accounts with --accounts <n|list>. Not registered as a standalone benchmark — it wraps other benchmarks.

Each turn spawns a one-shot codex exec subprocess authenticated as a selected account by pointing CODEX_HOME at that account's materialized home (<stateDir>/auth/_codex-home/<accountId>/, written by the TS runtime's coding-account-bridge.ts). The orchestrator process imports no Codex/Node dependency — it only needs the codex binary (or CODEX_BIN) and at least one materialized home.

Layout

Path	Role
codex_adapter/accounts.py	CODEX_HOME discovery + --accounts parsing + round-robin (fully offline-testable)
codex_adapter/client.py	CodexClient — one-shot codex exec turn as the selected account
tests/	Offline pytest suite (account selection, rotation, failure modes)

Run + test

• Live run (credential-gated): see ../docs/HITL_MULTI_CODEX_RUNBOOK.md.
• Offline tests: pytest codex-adapter/tests/ -v.

Notes

• --accounts: integer N = first N accounts; a,b = exactly those ids; omitted = all. Turns round-robin: turn i → accounts[i % len].
• The client never fabricates a response — a missing binary, an unauthenticated account (auth.json absent), or a nonzero codex exec raises.
• A live gpt-5.5 run needs real authenticated Codex homes + entitlement; the offline suite proves the account-selection/iteration logic without a model.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing here is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. "Tests pass" is not proof.

• Record AND read model trajectories from a live LLM (here: a real gpt-5.5 Codex account), not a mock. A captured-but-unread trajectory is not evidence.
• Real, full-featured E2E — no larp. Drive the real path end to end incl. error/edge/concurrency/adversarial input. A test asserting against a mock standing in for the thing under test does not count.
• No residuals, no shortcuts. Clear blockers by the hard path. The live multi-Codex run is credential-gated: mark it N/A with the reason (no Codex creds in this env) — never fake it.