Summary

The native Codex route uses an openai/gpt-* model ref plus agents.defaults.agentRuntime.id: "codex", but the Codex OAuth/runtime context metadata and user overrides are exposed under the openai-codex provider.

That creates a confusing split for context configuration. A user can configure:

{
  models: {
    providers: {
      "openai-codex": {
        models: [
          {
            id: "gpt-5.5",
            contextWindow: 1050000,
            contextTokens: 1000000,
          },
        ],
      },
    },
  },
  agents: {
    defaults: {
      model: { primary: "openai/gpt-5.5" },
      agentRuntime: { id: "codex" },
    },
  },
}

but the native harness context calculation still resolves against provider openai, so it can keep the runtime model's default contextTokens value such as 272000. In practice, /status may show a larger configured/catalog window while agentMeta.contextTokens reports the smaller default runtime cap.

Expected behavior

For openai/gpt-* running through the codex harness, context-window resolution should use openai-codex as the context config provider. That lets models.providers.openai-codex.models[].contextTokens override the Codex runtime cap consistently for the actual run and for /status.

Impact

Users following the recommended native Codex route can otherwise see contradictory context numbers and have no obvious way to make their configured Codex OAuth context cap affect the actual run.