Compatibility flags
Compatibility flags enable specific features. They can be useful if you want to help the Workers team test upcoming changes that are not yet enabled by default, or if you need to hold back a change that your code depends on but still want to apply other compatibility changes.
Compatibility flags will often have a date in which they are enabled by default, and so, by specifying a compatibility_date for your Worker, you can quickly enable all of these various compatibility flags up to, and including, that date.
Setting compatibility flags
You may provide a list of compatibility_flags, which enable or disable specific changes.
Via Wrangler
Compatibility flags can be set in a Worker’s wrangler.toml file.
This example enables the specific flag formdata_parser_supports_files, which is described below. As of the specified date, 2021-09-14, this particular flag was not yet enabled by default, but, by specifying it in compatibility_flags, we can enable it anyway. compatibility_flags can also be used to disable changes that became the default in the past.
# Opt into backwards-incompatible changes through September 14, 2021.compatibility_date = "2021-09-14"# Also opt into an upcoming fix to the FormData API.compatibility_flags = [ "formdata_parser_supports_files" ]Via the Cloudflare Dashboard
Compatibility flags cannot be set or updated on the Cloudflare Dashboard at this time.
Via the Cloudflare API
Compatibility flags can be set when uploading a Worker using the Workers Script API or Workers Versions API in the request body’s metadata field.
Node.js compatibility flag
A growing subset of Node.js APIs are available directly as Runtime APIs, with no need to add polyfills to your own code. To enable these APIs in your Worker, add the nodejs_compat compatibility flag to your wrangler.toml:
compatibility_flags = [ "nodejs_compat" ]As additional Node.js APIs are added, they will be made available under the nodejs_compat compatibility flag. Unlike most other compatibility flags, we do not expect the nodejs_compat to become active by default at a future date.
The Node.js AsyncLocalStorage API is a particularly useful feature for Workers. To enable only the AsyncLocalStorage API, use the nodejs_als compatibility flag.
compatibility_flags = [ "nodejs_als" ]Change history
Newest changes are listed first.
| Default as of | 2024-09-02 |
| Flag to enable | allow_custom_ports |
| Flag to disable | ignore_custom_ports |
When this flag is enabled, and you specify a port when making a subrequest with the fetch() API, the port number you specify will be used.
When you make a subrequest to a website that uses Cloudflare (“Orange Clouded”) — only ports supported by Cloudflare’s reverse proxy can be specified. If you attempt to specify an unsupported port, it will be ignored.
When you make a subrequest to a website that doesn’t use Cloudflare (“Grey Clouded”) - any port can be specified.
For example:
const response = await fetch("https://example.com:8000");With allow_custom_ports the above example would fetch https://example.com:8000 rather than
https://example.com:443.
Note that creating a WebSocket client with a call to new WebSocket(url) will also obey this flag.
| Default as of | 2024-06-03 |
| Flag to enable | blob_standard_mime_type |
| Flag to disable | blob_legacy_mime_type |
When calling response.blob.type(), the MIME type will now be properly extracted from content-type headers, per the WHATWG spec ↗.
| Default as of | 2024-04-29 |
| Flag to enable | brotli_content_encoding |
| Flag to disable | no_brotli_content_encoding |
When the brotli_content_encoding compatibility flag is enabled, Workers supports the br content encoding and can request and respond with data encoded using the Brotli ↗ compression algorithm. This reduces the amount of data that needs to be fetched and can be used to pass through the original compressed data to the client. See the Fetch API documentation for details.
| Flag to enable | cache_option_enabled |
| Flag to disable | cache_option_disabled |
When you enable the cache_option_enabled compatibility flag, you can specify a value for the cache property of the Request interface.
When this compatibility flag is not enabled, or cache_option_disabled is set, the Workers runtime will throw an Error saying The 'cache' field on 'RequestInitializerDict' is not implemented.
When this flag is enabled you can instruct Cloudflare not to cache the response from a subrequest you make from your Worker using the fetch() API:
The only cache option enabled with cache_option_enabled is 'no-store'.
Specifying any other value will cause the Workers runtime to throw a TypeError with the message Unsupported cache mode: <the-mode-you-specified>.
When no-store is specified:
-
All requests have the headers
Pragma: no-cacheandCache-Control: no-cacheare set on them. -
Subrequests to origins not hosted by Cloudflare bypasses Cloudflare’s cache.
Examples using cache: 'no-store':
const response = await fetch("https://example.com", { cache: 'no-store'});The cache value can also be set on a Request object.
const request = new Request("https://example.com", { cache: 'no-store'});const response = await fetch(request);| Default as of | 2022-10-31 |
| Flag to enable | export_commonjs_default |
| Flag to disable | export_commonjs_namespace |
CommonJS modules were previously exporting a module namespace (an object like { default: module.exports }) rather than exporting only the module.exports. When this flag is enabled, the export is fixed.
| Default as of | 2023-12-01 |
| Flag to enable | crypto_preserve_public_exponent |
| Flag to disable | no_crypto_preserve_public_exponent |
In the WebCrypto API, the publicExponent field of the algorithm of RSA keys would previously be an ArrayBuffer. Using this flag, publicExponent is a Uint8Array as mandated by the specification.
| Default as of | 2022-06-01 |
| Flag to enable | dont_substitute_null_on_type_error |
| Flag to disable | substitute_null_on_type_error |
There was a bug in the runtime that meant that when being passed into built-in APIs, invalid values were sometimes mistakenly coalesced with null. Instead, a TypeError should have been thrown. The dont_substitute_null_on_type_error fixes this behavior so that an error is correctly thrown in these circumstances.
| Default as of | 2022-10-31 |
| Flag to enable | capture_async_api_throws |
| Flag to disable | do_not_capture_async_api_throws |
The capture_async_api_throws compatibility flag will ensure that, in conformity with the standards API, async functions will only ever reject if they throw an error. The inverse do_not_capture_async_api_throws flag means that async functions which contain an error may throw that error synchronously rather than rejecting.
| Default as of | 2022-03-08 |
| Flag to enable | no_cots_on_external_fetch |
| Flag to disable | cots_on_external_fetch |
The no_cots_on_external_fetch flag disables the use of the Custom Origin Trust Store when making external (grey-clouded) subrequests from a Cloudflare Worker.
| Default as of | 2021-11-10 |
| Flag to enable | durable_object_fetch_requires_full_url |
| Flag to disable | durable_object_fetch_allows_relative_url |
Originally, when making a request to a Durable Object by calling stub.fetch(url), a relative URL was accepted as an input. The URL would be interpreted relative to the dummy URL http://fake-host, and the resulting absolute URL was delivered to the destination object’s fetch() handler. This was a mistake — full URLs were meant to be required. This flag makes full URLs required.
| Default as of | 2023-03-01 |
| Flag to enable | dynamic_dispatch_tunnel_exceptions |
| Flag to disable | dynamic_dispatch_treat_exceptions_as_500 |
Previously, when using Workers for Platforms’ dynamic dispatch API to send an HTTP request to a user Worker, if the user Worker threw an exception, the dynamic dispatch Worker would receive an HTTP 500 error with no body. When the dynamic_dispatch_tunnel_exceptions compatibility flag is enabled, the exception will instead propagate back to the dynamic dispatch Worker. The fetch() call in the dynamic dispatch Worker will throw the same exception. This matches the similar behavior of service bindings and Durable Objects.
| Default as of | 2021-11-10 |
| Flag to enable | fetch_refuses_unknown_protocols |
| Flag to disable | fetch_treats_unknown_protocols_as_http |
Originally, if the fetch() function was passed a URL specifying any protocol other than http: or https:, it would silently treat it as if it were http:. For example, fetch() would appear to accept ftp: URLs, but it was actually making HTTP requests instead.
Note that Cloudflare Workers supports a non-standard extension to fetch() to make it support WebSockets. However, when making an HTTP request that is intended to initiate a WebSocket handshake, you should still use http: or https: as the protocol, not ws: nor wss:.
The ws: and wss: URL schemes are intended to be used together with the new WebSocket() constructor, which exclusively supports WebSocket. The extension to fetch() is designed to support HTTP and WebSocket in the same request (the response may or may not choose to initiate a WebSocket), and so all requests are considered to be HTTP.
| Default as of | 2024-06-03 |
| Flag to enable | fetch_standard_url |
| Flag to disable | fetch_legacy_url |
The fetch_standard_url flag makes fetch() use WHATWG URL Standard ↗ parsing rules. The original implementation would throw TypeError: Fetch API cannot load errors with some URLs where standard parsing does not, for instance with the inclusion of whitespace before the URL. URL errors will now be thrown immediately upon calling new Request() with an improper URL. Previously, URL errors were thrown only once fetch() was called.
| Default as of | 2024-03-26 |
| Flag to enable | fetcher_no_get_put_delete |
| Flag to disable | fetcher_has_get_put_delete |
Durable Object stubs and Service Bindings both implement a fetch() method which behaves similarly to the global fetch() method, but requests are instead sent to the destination represented by the object, rather than being routed based on the URL.
Historically, API objects that had such a fetch() method also had methods get(), put(), and delete(). These methods were thin wrappers around fetch() which would perform the corresponding HTTP method and automatically handle writing/reading the request/response bodies as needed.
These methods were a very early idea from many years ago, but were never actually documented, and therefore rarely (if ever) used. Enabling the fetcher_no_get_put_delete, or setting a compatibility date on or after 2024-03-26 disables these methods for your Worker.
This change paves a future path for you to be able to define your own custom methods using these names. Without this change, you would be unable to define your own get, put, and delete methods, since they would conflict with these built-in helper methods.
| Default as of | 2021-11-03 |
| Flag to enable | formdata_parser_supports_files |
| Flag to disable | formdata_parser_converts_files_to_strings |
The FormData API ↗ is used to parse data (especially HTTP request bodies) in multipart/form-data format.
Originally, the Workers runtime’s implementation of the FormData API incorrectly converted uploaded files to strings. Therefore, formData.get("filename") would return a string containing the file contents instead of a File object. This change fixes the problem, causing files to be represented using File as specified in the standard.
| Default as of | 2022-03-21 |
| Flag to enable | global_navigator |
| Flag to disable | no_global_navigator |
With the global_navigator flag set, a new global navigator property is available from within Workers. Currently, it exposes only a single navigator.userAgent property whose value is set to 'Cloudflare-Workers'. This property can be used to reliably determine whether code is running within the Workers environment.
| Flag to enable | html_rewriter_treats_esi_include_as_void_tag |
The HTML5 standard defines a fixed set of elements as void elements, meaning they do not use an end tag: <area>, <base>, <br>, <col>, <command>, <embed>, <hr>, <img>, <input>, <keygen>, <link>, <meta>, <param>, <source>, <track>, and <wbr>.
HTML5 does not recognize XML self-closing tag syntax. For example, <script src="foo.js" /> does not specify a script element with no body. A </script> ending tag is still required. The /> syntax simply is not recognized by HTML5 at all and it is treated the same as >. However, many developers still like to use this syntax, as a holdover from XHTML, a standard which failed to gain traction in the early 2000’s.
<esi:include> and <esi:comment> are two tags that are not part of the HTML5 standard, but are instead used as part of Edge Side Includes ↗, a technology for server-side HTML modification. These tags are not expected to contain any body and are commonly written with XML self-closing syntax.
HTMLRewriter was designed to parse standard HTML5, not ESI. However, it would be useful to be able to implement some parts of ESI using HTMLRewriter. To that end, this compatibility flag causes HTMLRewriter to treat <esi:include> and <esi:comment> as void tags, so that they can be parsed and handled properly.
| Default as of | 2023-03-01 |
| Flag to enable | http_headers_getsetcookie |
| Flag to disable | no_http_headers_getsetcookie |
Adds the getSetCookie() ↗ method to the Headers ↗ API in Workers.
const response = await fetch("https://example.com");let cookieValues = response.headers.getSetCookie();| Default as of | 2024-05-13 |
| Flag to enable | internal_stream_byob_return_view |
| Flag to disable | internal_stream_byob_return_undefined |
In the original implementation of BYOB (“Bring your own buffer”) ReadableStreams, the read() method would return undefined when the stream was closed and there was no more data to read. This behavior was inconsistent with the standard ReadableStream behavior, which returns an empty Uint8Array when the stream is closed.
When the internal_stream_byob_return_view flag is used, the BYOB read() will implement standard behavior.
const resp = await fetch('https://example.org');const reader = resp.body.getReader({ mode: 'byob' });await result = await reader.read(new Uint8Array(10));
if (result.done) { // The result gives us an empty Uint8Array... console.log(result.value.byteLength); // 0
// However, it is backed by the same underlying memory that was passed // into the read call. console.log(result.value.buffer.byteLength); // 10}| Default as of | 2022-04-05 |
| Flag to enable | minimal_subrequests |
| Flag to disable | no_minimal_subrequests |
With the minimal_subrequests flag set, fetch() subrequests sent to endpoints on the Worker’s own zone (also called same-zone subrequests) have a reduced set of features applied to them. In general, these features should not have been initially applied to same-zone subrequests, and very few user-facing behavior changes are anticipated. Specifically, Workers might observe the following behavior changes with the new flag:
- Response bodies will not be opportunistically gzipped before being transmitted to the Workers runtime. If a Worker reads the response body, it will read it in plaintext, as has always been the case, so disabling this prevents unnecessary decompression. Meanwhile, if the Worker passes the response through to the client, Cloudflare’s HTTP proxy will opportunistically gzip the response body on that side of the Workers runtime instead. The behavior change observable by a Worker script should be that some
Content-Encoding: gzipheaders will no longer appear. - Automatic Platform Optimization may previously have been applied on both the Worker’s initiating request and its subrequests in some circumstances. It will now only apply to the initiating request.
- Link prefetching will now only apply to the Worker’s response, not responses to the Worker’s subrequests.
| Default as of | 2022-10-31 |
| Flag to enable | url_standard |
| Flag to disable | url_original |
The original implementation of the URL ↗ API in Workers was not fully compliant with the WHATWG URL Standard ↗, differing in several ways, including:
-
The original implementation collapsed sequences of multiple slashes into a single slash:
new URL("https://example.com/a//b").toString() === "https://example.com/a/b" -
The original implementation would throw
"TypeError: Invalid URL string."if it encountered invalid percent-encoded escape sequences, likehttps://example.com/a%%b. -
The original implementation would percent-encode or percent-decode certain content differently:
new URL("https://example.com/a%40b?c d%20e?f").toString() === "https://example.com/a@b?c+d+e%3Ff" -
The original implementation lacked more recently implemented
URLfeatures, likeURL.canParse()↗.
Set the compatibility date of your Worker to a date after 2022-10-31 or enable the url_standard compatibility flag to opt-in the fully spec compliant URL API implementation.
Refer to the response_redirect_url_standard compatibility flag , which affects the URL implementation used in Response.redirect().
| Default as of | 2023-08-01 |
| Flag to enable | no_cf_botmanagement_default |
| Flag to disable | cf_botmanagement_default |
This flag streamlines Workers requests by reducing unnecessary properties in the request.cf object.
With the flag enabled - either by default after 2023-08-01 or by setting the no_cf_botmanagement_default flag - Cloudflare will only include the Bot Management object in a Worker’s request.cf if the account has access to Bot Management.
With the flag disabled, Cloudflare will include a default Bot Management object, regardless of whether the account is entitled to Bot Management.
| Default as of | 2024-03-04 |
| Flag to enable | no_global_importscripts |
| Flag to disable | global_importscripts |
Suppresses the global importScripts() function. This method was included in the Workers global scope but was marked explicitly as non-implemented. However, the presence of the function could cause issues with some libraries. This compatibility flag removes the function from the global scope.
| Flag to enable | nodejs_als |
| Flag to disable | no_nodejs_als |
Enables the availability of the Node.js AsyncLocalStorage ↗ API in Workers. This API allows you to store data that is accessible to all asynchronous operations within a given execution context. This is useful for storing data that is relevant to the current request, such as request-specific metadata or tracing information.
| Flag to enable | nodejs_compat |
| Flag to disable | no_nodejs_compat |
Enables the full set of available Node.js APIs in the Workers Runtime.
| Default as of | 2024-01-29 |
| Flag to enable | python_workers |
This flag enables first class support for Python. Python Workers implement the majority of Python’s standard library, support all bindings, environment variable, and secrets, and integration with JavaScript objects and functions via a foreign function interface.
| Default as of | 2024-03-18 |
| Flag to enable | queues_json_messages |
| Flag to disable | no_queues_json_messages |
With the queues_json_messages flag set, Queue bindings will serialize values passed to send() or sendBatch() into JSON format by default (when no specific contentType is provided).
| Default as of | 2022-08-04 |
| Flag to enable | r2_list_honor_include |
With the r2_list_honor_include flag set, the include argument to R2 list options is honored. With an older compatibility date and without this flag, the include argument behaves implicitly as include: ["httpMetadata", "customMetadata"].
| Default as of | 2024-04-03 |
| Flag to enable | rpc |
| Flag to disable | no_rpc |
With this flag on, Durable Object stubs and Service Bindings support RPC. This means that these objects now appear as if they define every possible method name. Calling any method name sends an RPC to the remote Durable Object or Worker service.
For most applications, this change will have no impact unless you use it. However, it’s possible some existing code will be impacted if it explicitly checks for the existence of method names that were previously not defined on these types. For example, we’ve seen code in the wild which iterates over bindings and tries to auto-detect their types based on what methods they implement. Such code will now see service bindings as implementing every method, so may misinterpret service bindings as being some other type. In the cases we’ve seen, the impact was benign (nothing actually broke), but out of caution we are guarding this change behind a flag.
| Default as of | 2022-01-31 |
| Flag to enable | workers_api_getters_setters_on_prototype |
| Flag to disable | workers_api_getters_setters_on_instance |
Originally, properties on Workers API objects were defined as instance properties as opposed to prototype properties. This broke subclassing at the JavaScript layer, preventing a subclass from correctly overriding the superclass getters/setters. This flag controls the breaking change made to set those getters/setters on the prototype template instead.
This changes applies to:
AbortSignalAbortControllerBlobBodyDigestStreamEventFileRequestReadableStreamReadableStreamDefaultReaderReadableStreamBYOBReaderResponseTextDecoderTextEncoderTransformStreamURLWebSocketWritableStreamWritableStreamDefaultWriter
| Default as of | 2023-03-14 |
| Flag to enable | response_redirect_url_standard |
| Flag to disable | response_redirect_url_original |
Change the URL implementation used in Response.redirect() to be spec-compliant (WHATWG URL Standard).
| Default as of | 2021-11-10 |
| Flag to enable | streams_byob_reader_detaches_buffer |
| Flag to disable | streams_byob_reader_does_not_detach_buffer |
Originally, the Workers runtime did not detach the ArrayBuffers from user-provided TypedArrays when using the BYOB reader’s read() method, as required by the Streams spec, meaning it was possible to inadvertently reuse the same buffer for multiple read() calls. This change makes Workers conform to the spec.
User code should never try to reuse an ArrayBuffer that has been passed into a BYOB reader’s read() method. Instead, user code can reuse the ArrayBuffer backing the result of the read() promise, as in the example below.
// Consume and discard `readable` using a single 4KiB buffer.let reader = readable.getReader({ mode: "byob" });let arrayBufferView = new Uint8Array(4096);while (true) { let result = await reader.read(arrayBufferView); if (result.done) break; // Optionally something with `result` here. // Re-use the same memory for the next `read()` by creating // a new Uint8Array backed by the result's ArrayBuffer. arrayBufferView = new Uint8Array(result.value.buffer);}The more recently added extension method readAtLeast() will always detach the ArrayBuffer and is unaffected by this feature flag setting.
| Default as of | 2022-11-30 |
| Flag to enable | streams_enable_constructors |
| Flag to disable | streams_disable_constructors |
Adds the work-in-progress new ReadableStream() and new WritableStream() constructors backed by JavaScript underlying sources and sinks.
| Default as of | 2023-08-01 |
| Flag to enable | strict_compression_checks |
| Flag to disable | no_strict_compression_checks |
Perform additional error checking in the Compression Streams API and throw an error if a DecompressionStream has trailing data or gets closed before the full compressed data has been provided.
| Default as of | 2023-08-01 |
| Flag to enable | strict_crypto_checks |
| Flag to disable | no_strict_crypto_checks |
Perform additional error checking in the Web Crypto API to conform with the specification and reject possibly unsafe key parameters:
- For RSA key generation, key sizes are required to be multiples of 128 bits as boringssl may otherwise truncate the key.
- The size of imported RSA keys must be at least 256 bits and at most 16384 bits, as with newly generated keys.
- The public exponent for imported RSA keys is restricted to the commonly used values
[3, 17, 37, 65537]. - In conformance with the specification, an error will be thrown when trying to import a public ECDH key with non-empty usages.
| Default as of | 2022-11-30 |
| Flag to enable | transformstream_enable_standard_constructor |
| Flag to disable | transformstream_disable_standard_constructor |
Previously, the new TransformStream() constructor was not compliant with the Streams API standard. Use the transformstream_enable_standard_constructor to opt-in to the backwards-incompatible change to make the constructor compliant. Must be used in combination with the streams_enable_constructors flag.
| Default as of | 2024-04-01 |
| Flag to enable | unwrap_custom_thenables |
| Flag to disable | no_unwrap_custom_thenables |
With the unwrap_custom_thenables flag set, various Workers APIs that accept promises will also
correctly handle custom thenables (objects with a then method) that are not native promises, but
are intended to be treated as such). For example, the waitUntil method of the ExecutionContext
object will correctly handle custom thenables, allowing them to be used in place of native promises.
async fetch(req, env, ctx) { ctx.waitUntil({ then(res) { // Resolve the thenable after 1 second setTimeout(res, 1000); } }); // ...}| Default as of | 2023-07-01 |
| Flag to enable | urlsearchparams_delete_has_value_arg |
| Flag to disable | no_urlsearchparams_delete_has_value_arg |
The WHATWG introduced additional optional arguments to the URLSearchParams object delete() ↗ and has() ↗ methods that allow for more precise control over the removal of query parameters. Because
the arguments are optional and change the behavior of the methods when present there is a risk of
breaking existing code. If your compatibility date is set to July 1, 2023 or after, this compatibility flag will be enabled by default.
For an example of how this change could break existing code, consider code that uses the Array forEach() method to iterate through a number of parameters to delete:
const usp = new URLSearchParams();// ...['abc', 'xyz'].forEach(usp.delete.bind(usp));The forEach() automatically passes multiple parameters to the function that is passed in. Prior to the addition of the new standard parameters, these extra arguments would have been ignored.
Now, however, the additional arguments have meaning and change the behavior of the function. With this flag, the example above would need to be changed to:
const usp = new URLSearchParams();// ...['abc', 'xyz'].forEach((key) => usp.delete(key));| Default as of | 2023-11-08 |
| Flag to enable | vectorize_query_metadata_optional |
| Flag to disable | vectorize_query_original |
A set value on vectorize_query_metadata_optional indicates that the Vectorize query operation should accept newer arguments with returnValues and returnMetadata specified discretely over the older argument returnVectors. This also changes the return format. If the vector values have been indicated for return, the return value is now a flattened vector object with score attached where it previously contained a nested vector object.
| Default as of | 2023-08-15 |
| Flag to enable | web_socket_compression |
| Flag to disable | no_web_socket_compression |
The Workers runtime did not support WebSocket compression when the initial WebSocket implementation was released. Historically, the runtime has stripped or ignored the Sec-WebSocket-Extensions header — but is now capable of fully complying with the WebSocket Compression RFC. Since many clients are likely sending Sec-WebSocket-Extensions: permessage-deflate to their Workers today (new WebSocket(url) automatically sets this in browsers), we have decided to maintain prior behavior if this flag is absent.
If the flag is present, the Workers runtime is capable of using WebSocket Compression on both inbound and outbound WebSocket connections.
Like browsers, calling new WebSocket(url) in a Worker will automatically set the Sec-WebSocket-Extensions: permessage-deflate header. If you are using the non-standard fetch() API to obtain a WebSocket, you can include the Sec-WebSocket-Extensions header with value permessage-deflate and include any of the compression parameters defined in RFC-7692 ↗.
Experimental changes
These changes can be enabled via compatibility_flags, but are not yet scheduled to become default on any particular date.
| Flag to enable | html_rewriter_treats_esi_include_as_void_tag |
The HTML5 standard defines a fixed set of elements as void elements, meaning they do not use an end tag: <area>, <base>, <br>, <col>, <command>, <embed>, <hr>, <img>, <input>, <keygen>, <link>, <meta>, <param>, <source>, <track>, and <wbr>.
HTML5 does not recognize XML self-closing tag syntax. For example, <script src="foo.js" /> does not specify a script element with no body. A </script> ending tag is still required. The /> syntax simply is not recognized by HTML5 at all and it is treated the same as >. However, many developers still like to use this syntax, as a holdover from XHTML, a standard which failed to gain traction in the early 2000’s.
<esi:include> and <esi:comment> are two tags that are not part of the HTML5 standard, but are instead used as part of Edge Side Includes ↗, a technology for server-side HTML modification. These tags are not expected to contain any body and are commonly written with XML self-closing syntax.
HTMLRewriter was designed to parse standard HTML5, not ESI. However, it would be useful to be able to implement some parts of ESI using HTMLRewriter. To that end, this compatibility flag causes HTMLRewriter to treat <esi:include> and <esi:comment> as void tags, so that they can be parsed and handled properly.