Skip to main content
Version: 2.20

Plugin upgrade guide

How to adjust for changes made to the Plugin API.


Deprecated some Request types in favor of Get with only one arg

For several APIs like pants.core.util_rules.system_binaries, we had an eager and lazy version of the same API. You could do either of these two:

from pants.core.util_rules.system_binaries import ZipBinary, ZipBinaryRequest
from pants.engine.rules import Get, rule

class MyOutput:

def my_rule(zip_binary: ZipBinary) -> MyOutput:
return MyOutput()

async def my_rule_lazy() -> MyOutput:
zip_binary = await Get(ZipBinary, ZipBinaryRequest())
return MyOutput()

The lazy API is useful, for example, when you only want to Get that output type inside an if branch.

We added syntax in 2.17 to now use Get(OutputType), whereas before you had to do Get(OutputType, OutputTypeRequest) or (as of 2.15) Get(OutputType, {}). So, these OutputTypeRequest types are now redundant and deprecated in favor of simply using Get(OutputType).

EnvironmentBehavior.UNMIGRATED is no longer available

Following the deprecation cycle in 2.15, all Goals need to set EnvironmentBehavior.LOCAL_ONLY or EnvironmentBehavior.USES_ENVIRONMENTS.


RunFieldSet and TestRequest now have a .rules() method

These methods should be used to register your run/test plugins:

def rules():
return [

Additionally, these types now by-default register the implementations for the rules used for --debug/--debug-adapter. If your plugin doesn't support these flags, simply remove the rules you've declared and let the default ones handle erroring. If your plugin does support these, set the class property(s) supports_debug = True/supports_debug_adapter = True, respectively.

RunFieldSet can be used to run targets in the sandbox as part of a build rule

With the new experimental_run_in_sandbox target type, targets that implement RunFieldSet can be run as a build rule for their side effects.

Many rules that create RunRequests can be used verbatim, but others may make assumptions that they will not be run hermetically. You will need set run_in_sandbox_behavior to one of the following values to generate a rule that allows your targets to be run in the sandbox:

  • RunInSandboxBehavior.RUN_REQUEST_HERMETIC: Use the existing RunRequest-generating rule, and enable cacheing. Use this if you are confident the behaviour of the rule relies only on state that is captured by pants (e.g. binary paths are found using EnvironmentVarsRequest), and that the rule only refers to files in the sandbox.
  • RunInSandboxBehavior.RUN_REQUEST_NOT_HERMETIC: Use the existing RunRequest-generating rule, and do not enable cacheing. Use this if your existing rule is mostly suitable for use in the sandbox, but you cannot guarantee reproducible behavior.
  • RunInSandboxBehavior.CUSTOM: Opt to write your own rule that returns RunInSandboxRequest.
  • RunInSandboxBehavior.NOT_SUPPORTED: Opt out of being usable in experimental_run_in_sandbox. Attempting to use such a target will result in a runtime exception.

We expect to deprecate RUN_REQUEST_NOT_HERMETIC and NOT_SUPPORTED in a few versions time: these options are provided to give you some time to make your existing rules match the semantics of RUN_REQUEST_HERMETIC, or to add a CUSTOM rule.

BinaryShimsRequest no longer accepts output_directory

BinaryShims now produces all of its shim scripts in the root of its digest, and provides helper methods for use with immutable_input_digests and the PATH environment variable. It also produces a unique directory name so that multiple rules can be called to populate PATH.

Consider using these helper methods in favor of the old behavior:

process = Process(
env={"PATH": binary_shims.path_component},

You can replicate the previous behavior using AddDigest:

new_digest = await Get(Digest, AddDigest(binary_shims.digest, output_directory))


lint and fmt schema changes

In order to accomplish several goals (namely targetless formatters and unifying the implementation of lint) lint and fmt have undergone a drastic change of their plugin API.

1. Lint<Targets|Files>Request and FmtTargetsRequest now require a tool_subsystem class attribute.

Instead of the name class attribute, Lint<Targets|Files>Request and FmtTargetsRequest require subclasses to provide a tool_subsystem class attribute with a value of your tool's Subsystem subclass.

2. Your tool subsystem should have a skip option.

Although not explicitly not required by the engine to function correctly, mypy will complain if the subsystem type provided to tool_subsystem doesn't have a skip: SkipOption option registered.

Otherwise, you can # type: ignore[assignment] on your tool_subsystem declaration.

3. The core goals now use a 2-rule approach


In order to support targetless formatters, fmt needs to know which files you'll be operating on. Therefore the plugin API for fmt has forked into 2 rules:

  1. A rule taking <RequestType>.PartitionRequest and returning a Partitions object. This is sometimes referred to as the "partitioner" rule.
  2. A rule taking <RequestType>.Batch and returning a FmtResult. This is sometimes referred to as the "runner" rule.

This way fmt can serialize tool runs that operate on the same file(s) while parallelizing tool runs that don't overlap.

(Why are targetless formatters something we want to support? This allows us to have BUILD file formatters, formatters like Prettier running on your codebase without boilerplate targets, as well as Pants doing interesting deprecation fixers on its own files)

The partitioner rule gives you all the matching files (or FieldSets depending on which class you subclassed) and you'll return a mapping from <key> to files (called a Partition). The <key> can be anything passable at the rule boundary and is given back to you in your runner rule. The partitioner rule gives you an opportunity to perform expensive Gets once for the entire run, to partition the inputs based on metadata to simplify your runner, and to have a place for easily skipping your tool if requested.

The runner rule will mostly remain unchanged, aside from the request type (<RequestType>.Batch), which now has a .files property.

If you don't require any Gets or metadata for your tool in your partitioner rule, Pants has a way to provide a "default" implementation. In your FmtRequest subclass, set the partitioner_type class variable to PartitionerType.DEFAULT_SINGLE_PARTITION and only provide a runner rule.


Lint plugins are almost identical to format plugins, except in 2 ways:

  1. Your partitioner rule still returns a Partitions object, but the element type can be anything.
  2. <RequestType>.Batch has a .elements field instead of .files.

As always, taking a look at Pants' own plugins can also be very enlightening.

test schema changes

To enable running tests in batches, the plugin API for test has significantly changed. The new API largely resembles the lint/fmt API described above.

1. Test plugins must now define a skip-able Subsystem.

To hook into the new API, a test runner must declare a subclass of Subsystem with a skip: SkipOption option. Add skip = SkipOption("test") to your existing (or new) subsystems.

2. Test plugins must define a subclass of TestRequest.

To define the rules expected by the new test API, you will need to define a TestRequest subclass. This new type will point at your plugin-specific TestFieldSet and Subsystem subclasses:

class CustomTestRequest(TestRequest):
field_set_type = CustomTestFieldSet
tool_subsystem = CustomSubsystem

After declaring your new type, register its rules:

def rules():
return [
# Add to what you already have:

3. Test execution now uses a 2-rule approach

The plugin API for test has forked into 2 rules:

  1. A rule taking <TestRequestSubclass>.PartitionRequest and returning a Partitions object. This is sometimes referred to as the "partitioner" rule.
  2. A rule taking <TestRequestSubclass>.Batch and returning a TestResult. This is sometimes referred to as the "runner" rule.

The "partitioner" rule was introduced to allow plugins to group tests into "compatible" batches, to be executed as a batch within the "runner" rule. The "runner" rule is a replacement for the previous API which took TestFieldSet instances as input.

By default, registering <TestRequestSubclass>.rules() will register a "partitioner" rule that creates a single-element partition per input TestFieldSet, replicating the behavior from before Pants 2.15. You can then upgrade your existing "runner" rule to take the new input type.


async def run_test(field_set: CustomTestFieldSet) -> TestResult:


async def run_tests(batch: CustomTestRequest.Batch) -> TestResult:
field_set = batch.single_element

If you would like to make use of the new support for batched testing, override the partitioner_type field in your TestRequest subclass:

class CustomTestRequest(TestRequest):
field_set_type = CustomTestFieldSet
tool_subsystem = CustomSubsystem
partitioner_type = PartitionerType.CUSTOM

This will prevent registration of the default "partitioner" rule, allowing you to implement any partitioning logic you'd like. You'll then need to update your "runner" rule to handle a multi-element batch.

EnvironmentName is now required to run processes, get environment variables, etc

Pants 2.15 introduces the concept of "Target Environments", which allow Pants to execute processes in remote or local containerized environments (using Docker), and to specify configuration values for those environments.

In order to support the new environments feature, an EnvironmentName parameter is now required in order to:

  • Run a Process
  • Get environment variables
  • Inspect the current Platform

This parameter is often provided automatically from a transitive value provided earlier in the call graph. The choice of whether to use a local or alternative environment must be made at a @goal_rule level.

In many cases, the local execution environment is sufficient. If so, your rules will not require significant work to migrate, and execution will behave similarly to pre-2.15 versions of Pants.

In cases where the environment needs to be factored into to rule execution, you'll need to do some work.

2.15 adds a deprecation warning for all goals that have not considered whether they need to use the execution environment.


2.15 adds the environment_behavior property to the Goal class, which controls whether an EnvironmentName is automatically injected when a @goal_rule runs.

When environment_behavior=Goal.EnvironmentBehavior.UNMIGRATED (the default), the QueryRule that is installed for a @goal_rule will include an EnvironmentName and will raise a deprecation warning.

If your Goal only ever needs to use the local target environment, use environment_behavior=Goal.EnvironmentBehavior.LOCAL_ONLY. The QueryRule installed for the @goal_rule will include an EnvironmentName that refers to a local environment, and will silence the deprecation warning. No further migration work needs to be done for your Goal.

For goals that need to respect EnvironmentFields

If your goal needs to select the target's specified environment when running underlying rules, set environment_behavior=Goal.EnvironmentBehavior.USES_ENVIRONMENTS, which will silence the deprecation. Unlike for the LOCAL_ONLY behavior, any rules that require an EnvironmentName will need to specify that name directly.

In general, Goals should use EnvironmentNameRequest to get EnvironmentNames for the targets that they will be operating on.


Then, the EnvironmentName should be used at Get callsites which require an environment:

Get(TestResult, {field_set: TestFieldSet, environment_name: EnvironmentName})

The multi-parameter Get syntax provides the value transitively, and so will need to be used in many Get callsites in @goal_rules which transitively run processes, consume the platform, etc. One exception is that (most of) the APIs provided by are pinned to running in the __local__ environment, and so do not require an EnvironmentName to use.


To reduce the number of changes necessary in tests, the RuleRunner.inherent_environment argument defaults to injecting an EnvironmentName when running @rules in tests.

platform kwarg for Process deprecated

Previously, we assumed processes were platform-agnostic, i.e. they had identical output on all platforms (OS x CPU architecture). You had to opt into platform awareness by setting the kwargplatform on the Process; otherwise, remote caching could incorrectly use results from a different platform.

This was not a safe default, and this behavior also breaks the new Docker support. So, now all processes automatically are marked as platform-specific. proposes how you will eventually be able to mark a Process as platform-agnostic.

To fix this deprecation, simply delete the platform kwarg.

Environment, EnvironmentRequest, and CompleteEnvironment renamed and moved

The types were moved from pants.engine.environment to pants.engine.env_vars, and now have Vars in their names:

Before: pants.engine.environment.{Environment,EnvironmentRequest,CompleteEnvironment} After: pants.engine.env_vars.{EnvironmentVars,EnvironmentVarsRequest,CompleteEnvironmentVars}

The old names still exist until Pants 2.16 as deprecated aliases.

This rename was to avoid ambiguity with the new "environments" mechanism, which lets users specify different options for environments like Linux vs. macOS and running in Docker images.

MockGet expects input_types kwarg, not input_type

It's now possible in Pants 2.15 to use zero arguments or multiple arguments in a Get. To support this change, MockGet from run_run_with_mocks() now expects the kwarg input_types: tuple[type, ...] rather than input_type: type.


mock=lambda _: LintResult(...),


mock=lambda _: LintResult(...),

Deprecated Platform.current

The Platform to use will soon become dependent on a @rule's position in the @rule graph. To get the correct Platform, a @rule should request a Platform as a positional argument.

Deprecated convert_dir_literal_to_address_literal kwarg

The convert_dir_literal_to_address_literal keyword argument for RawSpecs.create() and SpecsParser.parse_specs() no longer does anything. It should be deleted.


See for the changelog.

Removed second type parameter from Get

Get now takes only a single type parameter for the output type: Get[_Output]. The input type parameter was unused.

FmtRequest -> FmtTargetsRequest

In order to support non-target formatting (like BUILD files) we'll be introducing additional fmt request types. Therefore FmtRequest has been renamed to FmtTargetsRequest to reflect the behavior.

This change also matches lint, which uses LintTargetsRequest.

Optional Option flag name

Pants 2.14 adds support for deducing the flag name from the attribute name when declaring XOptions. You can still provide the flag name in case the generated one shouldn't match the attribute name.


my_version = StrOption("--my-version", ...)
_run = BoolOption("--run", ...)


my_version = StrOption(...)  # Still uses --my-version
_run = BoolOption(...) # Still uses --run

InjectDependencies -> InferDependencies, with InferDependencies using a FieldSet

InjectDependenciesRequest has been folded into InferDependenciesRequest, which has also been changed to receive a FieldSet.

If you have an InjectDependenciesRequest type/rule, those should be renamed to Infer....

Then for each InferDependenciesRequest, the infer_from class variable should now point to a relevant FieldSet subclass type. If you had an Inject... request, the required_fields will likely include the relevant Dependencies subclass. Likewise for pre-2.14 Infer... request, therequired_fields will include the relevant SourcesField subclass.

Note that in most cases, you no longer need to request the target in your rule code, and should rely on FieldSet's mechanisms for matching targets and getting field values.

GenerateToolLockfileSentinel encouraged to use language-specific subclasses

Rather than directly subclassing GenerateToolLockfileSentinel, we encourage you to subclass GeneratePythonToolLockfileSentinel and GenerateJvmToolLockfileSentinel. This is so that we can distinguish what language a tool belongs to, which is used for options like [python].resolves_to_constraints_file to validate which resolve names are recognized.

Things will still work if you do not make this change, other than the new options not recognizing your tool.

However, keep the UnionRule the same, i.e. with the first argument still GenerateToolLockfileSentinel.

matches_filespec() replaced by FilespecMatcher

Instead, use FilespecMatcher(includes=[], excludes=[]).matches(paths: Sequence[str]) from pants.source.filespec.

The functionality is the same, but can have better performance because we don't need to parse the same globs each time .matches() is called. When possible, reuse the same FilespecMatcher object to get these performance benefits.


See for the changelog.

AddressInput and UnparsedAddressInputs require description_of_origin

Both types now require the keyword argument description_of_origin: str, which is used to make better error messages when the address cannot be found.

If the address is hardcoded in your rules, then use a description like "the rule find_my_binary". If the address comes from user inputs, it is helpful to mention where the user defines the value, for example f"the dependencies field from the target {my_tgt.address}".

You can now also set UnparsedAddressInputs(skip_invalid_addresses=True, ...), which will not error when addresses are invalid.

WrappedTarget requires WrappedTargetRequest


from pants.engine.addresses import Address
from import WrappedTarget

await Get(WrappedTarget, Address, my_address)


from import WrappedTarget, WrappedTargetRequest

await Get(WrappedTarget, WrappedTargetRequest(my_address, description_of_origin="my rule"))

Redesign of Specs

Specs, aka command line arguments, were redesigned in Pants 2.13:

  • The globs :: and : now match all files, even if there are no owning targets.
  • Directory args like my_dir/ can be set to match everything in the current directory, rather than the default target my_dir:my_dir.
  • Ignore globs were added with a - prefix, like :: -ignore_me::

To support these changes, we redesigned the class Specs and its sibling classes like AddressSpecs.

Renames for Spec subclass:

  • SiblingAddresses -> DirGlobSpec (vs. DirLiteralSpec)
  • DescendantAddresses -> RecursiveGlobSpec
  • AscendantAddresses -> AncestorGlobSpec

Those classes now have a keyword arg error_if_no_target_matches, rather than having a distinct class like MaybeDescendantAddresses.

AddressSpecs was renamed to SpecsWithoutFileOwners, and FilesystemSpecs to SpecsWithOnlyFileOwners. But almost always, you should instead use the new RawSpecs class because it is simpler. See Rules API and Target API for how to use Get(Targets, RawSpecs), including its keyword arguments.

If you were directly creating Specs objects before, you likely want to change to RawSpecs. Specs allows us to handle "ignore specs" like -ignore_me/, which is usually not necessary in rules. See the above paragraph for how to use RawSpecs.

SpecsSnapshot is now SpecsPaths

SpecsSnapshot was replaced with the more performant SpecsPaths from pants.engine.fs, which avoids digesting any files into the LMDB store.

Instead of specs_snapshot.snapshot.files, use specs_paths.files to get a list of all matching files.

If you still need the Digest (specs_snapshot.snapshot.digest), use await Get(Digest, PathGlobs(globs=specs_paths.files)).

Removed PutativeTargetsSearchPaths for tailor plugins


all_proto_files = await Get(Paths, PathGlobs, req.search_paths.path_globs("*.proto"))


all_proto_files = await Get(Paths, PathGlobs, req.path_globs("*.proto"))

You can also now specify multiple globs, e.g. req.path_globs("*.py", "*.pyi").

Banned short option names like -x

You must now use a long option name when defining options. You can also now only specify a single option name per option.

(These changes allowed us to introduce ignore specs, like pants list :: -ignore_me::.)


See for the changelog.

Unified formatters

Formatters no longer need to be installed in both the FmtRequest and LintTargetsRequest @unions: instead, installing in the FmtRequest union is sufficient to act as both a linter and formatter.

See Add a formatter for more information.


See for the changelog.

Deprecated Subsystem.register_options()

Pants 2.11 added "concrete" option types which when used as class attributes of your subsystem. These are more declarative, simplify accessing options, and work with MyPy!


class MySubsystem(Subsystem):
options_scope = "example"
help = "..."

def register_options(cls, register):


class MySubsystem(Subsystem):
options_scope = "example"
help = "..."

my_opt = BoolOption(

To access an option in rules, simply use my_subsystem.my_opt rather than my_subsystem.options.my_opt.

See Options and subsystems for more information, including the available types.

Moved BinaryPathRequest to pants.core.util_rules.system_binaries

The new module pants.core.util_rules.system_binaries centralizes all discovery of existing binaries on a user's machines.

The functionality is the same, you only need to change your imports for types like BinaryPathRequest to pants.core.util_rules.system_binaries rather than pants.engine.process.

Deprecated not implementing TargetGenerator in GenerateTargetsRequest implementors

See for an explanation and some examples of how to fix.

Replaced GoalSubsystem.required_union_implementations with GoalSubsystem.activated()

See for an explanation and some examples of how to fix.


See for the changelog.

Rename LintRequest to LintTargetsRequest

Pants 2.10 added a new LintFilesRequest, which allows you to run linters on code without any owning targets!

To improve clarity, we renamed LintRequest to LintTargetsRequest.

FmtRequest, CheckRequest, and LintTargetsRequest must set name

You must set the class property name on these three types.


class MyPyRequest(CheckRequest):
field_set_type = MyPyFieldSet


class MyPyRequest(CheckRequest):
field_set_type = MyPyFieldSet
name = "mypy"

This change is what allowed us to add the lint --only=flake8 feature.

For DRY, it is a good idea to change the formatter_name, linter_name, and checker_name in FmtResult, LintResults, and CheckResults, respectively, to use rather than hardcoding the string again. See for examples.

Removed LanguageFmtTargets for fmt

When setting up a new language to be formatted, you used to have to copy and paste a lot of boilerplate like ShellFmtTargets. That's been fixed, thanks to

To fix your code:

  1. If you defined any new languages to be formatted, delete the copy-and-pasted LanguageFmtTargets code.
  2. For every formatter, change the UnionRule to be UnionRule(FmtRequest, BlackRequest), rather than UnionRule(PythonFmtRequest, BlackRequest), for example.

ReplImplementation now passes root targets, not transitive closure

We realized that it's useful to let REPL rules know what was specified vs. what is a transitive dependency:

To adapt to this, you will want to use transitive_targets = await Get(TransitiveTargets, TransitiveTargetsRequest(request.addresses), then operate on transitive_targets.closure.

Removed PexFromTargetsRequest.additional_requirements

Let us know if you were using this, and we can figure out how to add it back:

Removed PexFromTargetsRequest(direct_deps_only: bool)

Let us know if you were using this, and we can figure out how to add it back:

Renamed GenerateToolLockfileSentinel.options_scope to resolve_name

See for more info.

Renamed PythonModule to PythonModuleOwnersRequest

This type was used to determine the owners of a Python module. The new name makes that more clear. See


See for the changelog.

Deprecated RuleRunner.create_files(), .create_file() and .add_to_build_file()

Instead, for your RuleRunner tests, use .write_files(). See for some examples.


See for the changelog.

Target modeling changes

Pants 2.8 cleaned up the modeling of targets. Now, there are targets that describe the atom of each language, like python_test and python_source which correspond to a single file. There are also target generators which exist solely for less boilerplate, like python_tests and python_sources.

We recommend re-reading Targets and BUILD files.


The Sources class was replaced with SourcesField, SingleSourceField, and MultipleSourcesField.

When defining new target types with the Target API, you should choose between subclassing SingleSourceField and MultipleSourcesField, depending on if you want the field to be source: str or sources: list[str].

Wherever you were using Sources in your @rules, simply replace with SourcesField.

Renames of some Sources subclasses

You should update all references to these classes in your @rules.

  • FilesSources -> FileSourceField
  • ResourcesSources -> ResourceSourceField
  • PythonSources -> PythonSourceField


The method OutputPathField.value_or_default() no longer takes Address as an argument.


See for the changelog.

Type hints work properly

Pants was not using PEP 561 properly, which means that MyPy would not enforce type hints when using Pants APIs. Oops! This is now fixed.

Options scopes should not have _

For example, use my-subsystem instead of my_subsystem. This is to avoid ambiguity with target types.


See for the changelog.


ProcessCacheScope.NEVER was renamed to ProcessCacheScope.PER_SESSION to better reflect that a rule never runs more than once in a session (i.e. a single Pants run) given the same inputs.

ProcessCacheScope.PER_RESTART was replaced with ProcessCacheScope.PER_RESTART_ALWAYS and ProcessCacheScope.PER_RESTART_SUCCESSFUL.


Now called InterpreterConstraints and defined in pants.backend.python.util_rules.interpreter_constraints.


See for the changelog.


BoolField.value is no longer bool | None, but simply bool. This means that you must either set required = True or set the default.

Use TriBoolField if you still want to be able to represent a trinary state: False, True, and None.

Added RuleRunner.write_files()

This is a more declarative way to set up files than the older API of RuleRunner.create_file(), .create_files(), and .add_to_build_files(). See Testing plugins.


See for the changelog.

PexRequest changes how entry point is set

See Instead of setting entry_point="pytest" in the PexRequest constructor, now you set main=ConsoleScript("black") or main=EntryPoint("pytest").

Must use EnvironmentRequest for accessing environment variables

See Pants now eagerly purges environment variables from the run, so using os.environ in plugins won't work anymore.

Instead, use await Get(Environment, EnvironmentRequest(["MY_ENV_VAR"]).

For RuleRunner tests, you must now either set env or the new env_inherit arguments for environment variables to be set. Tests are now hermetic.


There were no substantial changes to the Plugin API in 2.3. See for the changelog.


See for the changelog.

PrimitiveField and AsyncField are removed (2.2.0.dev0)

Rather than subclassing PrimitiveField, subclass Field directly. Field now behaves like PrimitiveField used to, and PrimitiveField was removed for simplicity.

Rather than subclassing AsyncField or AsyncStringSequenceField, subclass Field or a template like StringField and also subclass AsyncFieldMixin:

from import AsyncFieldMixin, StringField)

class MyField(StringField, AsyncFieldMixin):
alias = "my_field"
help = "Description."

Async fields now access the raw value with the property .value, rather than .sanitized_raw_value. To override the eager validation, override compute_value(), rather than sanitize_raw_value(). Both these changes bring async fields into alignment with non-async fields.

Set the property help with Subsystems, Targets, and Fields (2.2.0.dev3)

Previously, you were supposed to set the class's docstring for the pants help message. Instead, now set a class property help, like this:

class MyField(StringField):
alias = "my_field"
help = "A summary.\n\nOptional extra information."

Pants will now properly wrap strings and preserve newlines. You may want to run pants help ${target/subsystem} to verify things render properly.


See for the changelog.

SourcesSnapshot is now SpecsSnapshot (2.1.0rc0)

The type was renamed for clarity. Still import it from pants.engine.fs.


See for the changelog.

Use TransitiveTargetsRequest as input for resolving TransitiveTargets (2.0.0rc0)

Rather than await Get(TransitiveTargets, Addresses([addr1])), use await Get(TransitiveTargets, TransitiveTargetsRequest([addr1])), from

It's no longer possible to include TransitiveTargets in your @rule signature in order to get the transitive closure of what the user specified on the command. Instead, put Addresses in your rule's signature, and use await Get(TransitiveTargets, TransitiveTargetsRequest(addresses)).

Codegen implementations: use DependenciesRequestLite and TransitiveTargetsLite (2.0.0rc0)

Due to a new cycle in the rule graph, for any codegen implementations, you must use DependenciesRequestLite instead of DependenciesRequest, and TransitiveTargetsLite instead of TransitiveTargetsRequest. Both imports are still from

These behave identically, except that they do not include dependency inference in the results. Unless you are generating for input = PythonSources, this should be fine, as dependency inference is currently only used with Python.

This is tracked by

Dependencies-like fields have more robust support (2.0.0rc0)

If you have any custom fields that act like the dependencies field, but do not subclass Dependencies, there are two new mechanisms for better support.

  1. Instead of subclassing StringSequenceField, subclass SpecialCasedDependencies from This will ensure that the dependencies show up with pants dependencies and pants dependents.
  2. You can use UnparsedAddressInputs from pants.engine.addresses to resolve the addresses:
from pants.engine.addresses import Address, Addresses, UnparsedAddressedInputs
from import Targets


addresses = await Get(Addresses, UnparsedAddressedInputs(["//:addr1", "project/addr2"], owning_address=None)

# Or, use this.
targets = await Get(
UnparsedAddressedInputs(["//:addr1", "project/addr2"], owning_address=Address("project", target_name="original")

If you defined a subclass of SpecialCasedDependencies, you can use await Get(Addresses | Targets, UnparsedAddressInputs, my_tgt[MyField].to_unparsed_address_inputs()).

(Why would you ever do this? If you have dependencies that you don't treat like normal—e.g. that you will call the equivalent of pants package on those deps—it's often helpful to call out this magic through a dedicated field. For example, Pants's archive target type has the fields files and packages, rather than dependencies.)

package implementations may want to add the field output_path (2.0.0rc0)

All of Pants's target types that can be built via pants package now have an output_path field, which allows the user to override the path used for the created asset.

You optionally may want to add this output_path field to your custom target type for consistency:

  1. Include OutputPathField from pants.core.goals.package in your target's core_fields class property.
  2. In your PackageFieldSet subclass, include output_path: OutputPathField.
  3. When computing the filename in your rule, use my_package_field_set.output_path.value_or_default(field_set.address, file_ending="my_ext").