OpenAI Codex

AGENTS.md is the documented repository instruction contract, with hierarchical scope for more specific directories.

The reviewed AGENTS.md source does not establish one general web-surface loading contract outside the named Codex clients and cloud workflows.

Codex exposes first-party local lifecycle hooks. OpenAI explicitly describes hooks as guardrails, not complete policy enforcement.

The reviewed official hooks reference documents the local Codex hook boundary, not an equivalent cloud hook runtime.

The reviewed hook reference does not establish a browser-hosted hook runtime.

Codex documents reusable skills as the first-party packaging mechanism for repeatable workflows and supporting resources.

The reviewed build-skills source is insufficient to classify skill discovery and execution across web and cloud as equivalent to local clients.

Codex enables built-in and custom subagents by default and surfaces their parallel work as coordinator-managed threads.

Hosted subagent workflows are documented, but availability and controls differ from configurable local Codex agents.

Codex runs parallel specialized subagents and collects results through the main coordinator. The reviewed docs do not describe a peer mailbox or shared task-ledger topology.

Hosted parallel subagents are documented as coordinator-managed workflows; the source does not describe a peer team with shared task state.

Codex documents first-party MCP client configuration for extending local agent tool and context access.

The reviewed MCP source is insufficient to assert one hosted capability, transport, and authentication contract.

Codex CLI documents an interactive slash-command surface, while repeatable custom workflows are packaged as skills.

The reviewed slash-command source is CLI-specific and does not establish identical commands on IDE, web, or cloud surfaces.

Codex exposes a first-party configuration file and layered controls for models, sandboxing, approvals, and extensions.

The reviewed config-file source does not establish an equivalent hosted configuration contract.

Local Codex clients use a local memory store and per-chat controls; required team rules still belong in AGENTS.md or checked-in documentation.

ChatGPT and hosted work use separate account or workspace memory behavior rather than the local Codex memory store and controls.