The Codex reference adapter translates an explicitly supplied codex exec --json
trace into provider-attributed AI SAFE² evidence. It is offline, requires no
credentials, and does not read Codex session storage.
codex exec --json "run the approved task" > codex-trace.jsonl
safe2 adapter codex-jsonl codex-trace.jsonl \
--codex-version YOUR_CODEX_VERSION \
--output codex-evidence.json
safe2 adapter codex-descriptor \
--codex-version YOUR_CODEX_VERSION \
--output codex-adapter.json
safe2 adapter conformance codex-adapter.json codex-evidence.json \
--output codex-conformance.json
The official OpenAI documentation
documents codex exec --json as an ordered JSONL event stream, including
item.started, item.completed, command-execution items, and token usage in
turn.completed events. This adapter consumes only that documented surface.
The output contains aggregate event, item, command-status, and token counts. It never copies prompts, messages, reasoning, command strings, command output, or arbitrary item bodies. The source filename and SHA-256 digest bind the result to exact bytes without retaining an absolute local path.
Unknown event types or documented events with unmodeled shapes are counted as
explicit coverage gaps and produce status: partial; malformed, duplicate-key,
non-finite, oversized, blank-line, symlinked, and non-object inputs fail closed.
The adapter does not execute Codex, validate task success, independently verify
provider-reported usage, authorize an action, or claim AI SAFE² conformance.