Codex

The codex adapter drives the Codex CLI through its non-interactive codex exec --json mode, detected as the codex binary on PATH; import codex from anyagent-js/codex.

What you can ask for

readOnly, model, effort, resume, and attachments map to the CLI’s own flags; systemPrompt and schema are provided by AnyAgent. mcp is not declared; reach Codex’s -c config overrides via extraArgs. authStatus() asks codex login status (a ChatGPT login asserts billing: "subscription"), models() asks codex debug models, and usageStatus() asks the CLI’s app-server for live rate-limit standing. The full grid is on the support matrix.

Good to know

• effort has no closed vocabulary; accepted values are per-model, listed by models() on ModelInfo.reasoningEfforts.
• The thread.started id lands on result.sessionId; pass it back as resume.
• Cache reads fold into the input total and reasoning into the output total; cacheReadTokens and reasoningTokens report those shares, not additions. No cost is reported.
• Codex meters separately per feature, so usageStatus() can return more than one set of windows. Labels are prefixed with the feature they belong to (codex:primary), and the same answer carries the credit balance on credits.
• Read windowMinutes, never the label, to tell a 5-hour window from a weekly one. OpenAI has shipped the weekly number under the label primary with no second window at all, so position says nothing about length.
• The limits come from codex app-server, a separate process from the one that runs prompts. Asking costs a short handshake and no model call.

Under the hood

Every run executes codex exec --json --sandbox danger-full-access with the prompt piped over stdin through the - positional. Run options change the sandbox and append flags: