Skip to content

Latest commit

 

History

History
1184 lines (918 loc) · 62.4 KB

a2ui_agent.blueprint.md

File metadata and controls

1184 lines (918 loc) · 62.4 KB
name a2ui_agent
type module
description Agent SDK specification for catalog management, prompt engineering, parsing, and payload validation.

Agent SDK Development Guide

This document describes the architecture of an A2UI Agent SDK. The design separates concerns into distinct layers to follow a similar structure for consistency across languages, providing a streamlined developer experience for building AI agents that generate rich UI.

The Agent SDK is responsible for:

  • Catalog management
  • Capability negotiation
  • Prompt engineering
  • Response parsing
  • Payload validation
  • Transport packaging

It enables Large Language Models (LLMs) and autonomous agents to understand available UI capabilities and ensures that generated UI payloads conform strictly to negotiated specification contracts before transmission to client renderers.


1. Unified Architecture Overview

The Agent SDK architecture introduces a clear separation between low-level single-responsibility primitives and a high-level application manager facade:

graph TD
    Facade["High-Level Application Facade<br/>(A2uiGenerator / A2uiRequestProcessor)"]
    Transformers["Catalog Transformers<br/>(Component / Function Allowlist Rules)"]
    Formats["Inference Formats<br/>(DirectJson / Express / More to be added)"]
    Parsers["Parser Engine<br/>(wrap, unwrap, compile, decompile, parse_response, open_stream)"]
    CoreModels["Core Data Models & Validation<br/>(a2ui_core Catalog, MessageProcessor & PayloadValidator)"]

    Facade --> Transformers
    Facade --> Formats
    Facade --> Parsers
    Transformers --> CoreModels
    Formats --> CoreModels
    Parsers --> CoreModels
Loading
  1. Decoupled Primitive Layer:
    • Catalog Representation: Directly uses canonical Catalog models from a2ui_core, as the schema-only CatalogApi.
    • Catalog Transformers: Standalone rule sets (CatalogTransformer, ComponentPruningTransformer, FunctionPruningTransformer) for filtering component definitions and function signatures from pristine catalogs.
    • Inference Formats: Strategy facades (InferenceFormat, InferenceFormatFactory) pairing format-specific prompt generators (PromptGenerator) and parsers (Parser). Supported strategies include DirectJsonFormat and ExpressFormat.
    • Prompt Generators: Format builders consuming transformed catalogs and prompt examples to generate system instruction snippets.
    • Parsers: Response extraction engines performing tag detection (has_format_content), tag unwrapping (unwrap) and wrapping (wrap), syntax compilation (compile) and decompilation (decompile). A format whose notation can be read incrementally also opens a ResponseStream that reads a response chunk by chunk (open_stream).
    • Validation Layer: Reaches a2ui_core validation through MessageProcessor, which holds every active catalog, resolves each item to the catalog it belongs to, and delegates the item to that catalog's PayloadValidator. Protocol version branching (v0_8, v0_9, v0_9_1, v1_0) is handled by the core version adapters.
  2. Encapsulated Application Processor:
    • CatalogConfig: Configuration dataclass pairing a resolved Catalog with the custom transformers to apply to it. A CatalogProvider (FileSystemCatalogProvider, InMemoryCatalogProvider) is how a catalog document becomes that Catalog, but the config holds the result rather than the provider.
    • A2uiGenerator: Agent-level lifecycle manager holding supported CatalogConfigs, generating pre-negotiated A2uiRequestProcessor instances per renderer capability signature.
    • A2uiRequestProcessor: Central processor facade object unifying multi-catalog capability resolution (resolve_catalogs), system prompt snippet rendering, turn-scoped parser creation, and response validation.

2. Directory & Package Structure

All SDK implementations of a2ui_agent must maintain a standardized directory layout similar to the Python SDK under python/a2ui_agent/.

a2ui_agent/
├── processor/                 # High-level application facade package
│   ├── catalog_config         # CatalogConfig structure for catalog registration
│   ├── processor              # A2uiRequestProcessor facade implementation
│   ├── generator              # A2uiGenerator class
│   └── catalog_providers      # Catalog provider classes
├── inference_format           # Abstract InferenceFormat & InferenceFormatFactory facades
├── inference_formats/         # Concrete inference format strategy implementations
│   ├── direct_json/           # Self-contained Direct JSON format package
│   │   ├── format             # DirectJsonFormat, DirectJsonFormatFactory
│   │   ├── prompt_generator   # DirectJsonPromptGenerator
│   │   └── parser             # DirectJsonParser class, including open_stream streaming
│   └── express/               # Self-contained Express DSL format package
│       ├── format             # ExpressFormat, ExpressFormatFactory
│       ├── compiler           # ExpressCompiler class
│       ├── decompiler         # ExpressDecompiler class
│       ├── parser             # ExpressParser class
│       └── prompt_generator   # ExpressPromptGenerator class
├── parser/                    # Common Parser contracts and data structures
│   ├── parser                 # Abstract Parser base class
│   └── response_part          # RawResponsePart, RawA2uiPart, TextPart, A2uiPart data structures
├── prompt/                    # Prompt Generation contracts
│   └── generator              # Abstract PromptGenerator base class
├── catalog_transformers/      # Catalog and Protocol Transformers
│   ├── base                   # Abstract CatalogTransformer class
│   └── pruning                # ComponentPruningTransformer, FunctionPruningTransformer
└── utils/                     # Utility helpers layer
    ├── catalog_resolver       # resolve_catalogs capability resolution function
    └── renderer_capabilities  # read_renderer_capabilities transport metadata helper

3. Interface Specification

A. Catalog Representation & Catalog Transformers

The Agent SDK uses a2ui.core.Catalog directly as the canonical model representing component definitions, function signatures, and theme schemas.

An agent describes and validates renderer functions without running them, so every catalog it loads, negotiates, prompts with, or validates against is a CatalogApi, the core name for Catalog[ComponentApi, FunctionApi] (whose functions may include AgentFunction entries with agent-side handlers when the optional agent_functions feature is implemented). Catalog transformers are the exception: they stay generic over the component and function types, so the same rules can also prune a catalog that carries implementations.

CatalogTransformer

Abstract base interface for transformation rules applied to catalog schemas prior to prompt engineering and payload validation.

TComponent = TypeVar("TComponent", bound=ComponentApi)
TFunction = TypeVar("TFunction", bound=FunctionApi)

class CatalogTransformer(ABC):
    """Abstract base interface for transformation rules applied to catalog schemas."""

    @abstractmethod
    def transform(
        self, catalog: Catalog[TComponent, TFunction]
    ) -> Catalog[TComponent, TFunction]:
        """Transforms a Catalog into a modified Catalog of the same component and function types."""
        pass

ComponentPruningTransformer

Prunes catalog component definitions to an allowlist of allowed components.

class ComponentPruningTransformer(CatalogTransformer):
    """Prunes catalog component definitions to an allowlist of allowed components."""

    def __init__(self, allowed_components: Sequence[str]):
        self.allowed_components = set(allowed_components)

    def transform(
        self, catalog: Catalog[TComponent, TFunction]
    ) -> Catalog[TComponent, TFunction]:
        """Returns a new Catalog filtered to only include components in allowed_components."""
        pass

FunctionPruningTransformer

Prunes catalog function definitions to an allowlist of allowed renderer-side validation rules and logic functions.

class FunctionPruningTransformer(CatalogTransformer):
    """Prunes catalog function definitions to an allowlist of allowed functions."""

    def __init__(self, allowed_functions: Sequence[str]):
        self.allowed_functions = set(allowed_functions)

    def transform(
        self, catalog: Catalog[TComponent, TFunction]
    ) -> Catalog[TComponent, TFunction]:
        """Returns a new Catalog filtered to only include functions in allowed_functions."""
        pass

B. Prompt Generation Layer (a2ui.prompt)

PromptGenerator

Abstract base class for constructing system prompt instruction snippets across inference formats.

Application code gets a prompt generator from InferenceFormat.prompt_generator and renders it with generate(). What a generator is built from (catalogs, examples, format-specific options) belongs to each format's constructor rather than to the shared interface.

A format implements the abstract catalogs property and three hooks, and never overrides generate():

  • generate_base_rules(): what every catalog shares, rendered once: the syntax contract, grammar and sentinel tags, plus any protocol schemas the format shows the model (Direct JSON's agent-to-renderer and common types schemas).
  • generate_catalog_instructions(include_schema, catalog): the component and function signatures or schemas of one catalog, or of every catalog in catalogs when catalog is omitted.
  • generate_examples(catalog, validate): the formatted example turns for one catalog, or for every catalog when catalog is omitted.

generate() is final and returns assemble_instructions(self), the same function SkillGenerator builds a monolithic skill's body with, so a prompt snippet and a monolithic skill for the same catalogs have the same body. The per-format generator classes are not part of the public API: callers reach them through InferenceFormat.prompt_generator, and a third-party format subclasses PromptGenerator directly.

Every format's generator renders the prompt example turns it is given in the order given. Each turn is the list of AgentToRendererMessage objects making it up. An example carries no label of its own: what the model learns from it is the payload, so a description would be the prompt author's prose rather than part of the contract.

A generator lists only agent-callable functions as callRendererFunction targets (see Agent-callable functions).

EXAMPLES_HEADING = "### Examples:\n\n"

class PromptGenerator(ABC):
    """Abstract base class for format-specific prompt generators."""

    @property
    @abstractmethod
    def catalogs(self) -> list[CatalogApi]:
        """The catalogs the generator describes, in order."""

    def generate_base_rules(self) -> str:
        """Renders what every catalog shares: syntax rules, sentinel tags, shared schemas."""
        return ""

    def generate_catalog_instructions(
        self, include_schema: bool = True, catalog: CatalogApi | None = None
    ) -> str:
        """Renders one catalog's signatures or schemas, or every catalog's when omitted."""
        return ""

    def generate_examples(
        self, catalog: CatalogApi | None = None, validate: bool = False
    ) -> str:
        """Renders one catalog's example turns, or every catalog's when omitted."""
        return ""

    @final
    def generate(self) -> str:
        """Renders the prompt snippet. The caller adds its own role and workflow text."""
        return assemble_instructions(self)


def assemble_instructions(
    generator: PromptGenerator, catalogs: Sequence[CatalogApi] | None = None
) -> str:
    """Joins the hooks' output for some catalogs (default: the generator's own).

    The text is the base rules, then the catalog instructions, then the examples
    under EXAMPLES_HEADING, separated by blank lines; sections that render nothing
    are left out.

    - For the generator's own catalogs, the catalog instructions and examples come
      from one call each with no catalog, so a format can render what its catalogs
      share once. The result is cached per generator, which is treated as
      immutable once built.
    - For any other catalogs, each catalog's instructions and examples are rendered
      on their own and joined, and examples a catalog repeats are added once.
    """

C. Common Parser Package (a2ui.parser)

Response Part Structures

@dataclass
class TextPart:
    """Represents extracted conversational text from an LLM response.

    Attributes:
        text: The conversational text content intended for user display.
    """
    text: str

@dataclass
class RawA2uiPart:
    """Represents an uncompiled A2UI format content block extracted from an LLM response.

    Attributes:
        a2ui_raw: The raw uncompiled format content string (e.g., raw XML/DSL/JSON).
    """
    a2ui_raw: str

@dataclass
class RawResponsePart:
    """Represents an uncompiled token from an LLM response stream.

    Attributes:
        part: The underlying content, either conversational TextPart or uncompiled RawA2uiPart.
        is_final: Whether this part is complete/closed (not truncated during streaming).
    """
    part: Union[TextPart, RawA2uiPart]
    is_final: bool = True

@dataclass
class A2uiPart:
    """Represents extracted and compiled A2UI payload messages.

    Attributes:
        a2ui: List of validated AgentToRendererMessage objects to deliver to client renderers.
    """
    a2ui: list[AgentToRendererMessage]

ResponsePart = Union[TextPart, A2uiPart]

Parser

Base interface for response parsers across all inference format strategies.

A parser knows only its own format. It reads its own sentinel tags and treats everything else in a response as conversational text, including a block in another format's tags: a model that shows the other notation in prose has written text, not a payload, and adding a format must not require changing the parsers of the existing ones.

class Parser(ABC):
    """Abstract base class for response parsers.

    Responsible for tokenizing LLM output streams, unwrapping format tags, and compiling raw format
    expressions into standard A2UI payload messages.
    """

    @abstractmethod
    def has_format_content(self, content: str, complete: bool = False) -> bool:
        """Reports whether the content carries a block written in this format.

        A caller uses this to decide whether a response is this format's business at all,
        without paying for a parse. It tokenizes blocks without compiling. Another format's
        tags do not count. A block recovered by `unwrap` when its opening sentinel tag was
        omitted (closed by this format's closing tag) also counts as a complete block.

        Args:
            content: Raw string response emitted by the LLM, possibly partial.
            complete: Whether to require a closed block. False matches an opening tag on
                its own, which is what a streaming caller needs to know it has started
                receiving a payload.

        Returns:
            True when the content carries a block belonging to this format.
        """
        pass

    @abstractmethod
    def wrap(self, blocks: Sequence[RawResponsePart]) -> str:
        """Converts a sequence of RawResponseParts to a string, adding enclosing tags or markers around each raw A2UI section and concatenating conversational text parts."""
        pass

    @abstractmethod
    def unwrap(self, content: str) -> list[RawResponsePart]:
        """Tokenizes the LLM response into an ordered list of RawResponsePart objects, extracting raw format content between sentinel tags while preserving chronological order.

        Content outside this format's sentinel tags, including another format's tags, is a TextPart.
        Markdown fences and backticks wrapping inline-code sentinel tags are stripped from block
        edges, and a closing tag on a single-line comment line closes the block.
        If the model omits the opening sentinel tag but emits a closing sentinel tag after a fenced
        block or a line that begins a format payload, at most one such block is recovered in
        one-shot unwrapping (streaming parsers require the opening sentinel tag).

        Args:
            content: Raw string response emitted by the LLM.

        Returns:
            An ordered list of RawResponsePart objects representing alternating slices of
            conversational text and cleaned format payload blocks.
        """
        pass

    @abstractmethod
    def compile(self, format_content: str) -> list[AgentToRendererMessage]:
        """Compiles a raw format content string into a list of validated A2UI message structures.

        Compile-time repairs tolerate stray unescaped line breaks inside object keys,
        identifier/reference properties, at single-line string edges, and in data paths or
        split identifiers when the joined name resolves.

        Args:
            format_content: The uncompiled raw payload string (e.g. raw JSON or DSL expression).

        Returns:
            List of compiled AgentToRendererMessage objects.
        """
        pass

    @abstractmethod
    def decompile(self, a2ui_payload: Sequence[AgentToRendererMessage]) -> str:
        """Decompiles structured A2UI payload messages into this format's raw notation.

        Args:
            a2ui_payload: List of AgentToRendererMessage objects to convert to raw format text.

        Returns:
            Raw format content string representing the messages.
        """
        pass

    def parse_response(self, content: str, wrapped: bool = True) -> list[ResponsePart]:
        """Generic non-streaming response parsing. Unwraps raw LLM text and compiles valid A2UI payloads,
        preserving the exact chronological order of conversational text and A2UI payload blocks.

        Args:
            content: Complete raw text response emitted by the LLM.
            wrapped: Whether the output is expected to be wrapped inside format sentinel tags.

        Returns:
            List of ResponsePart objects (TextPart / A2uiPart).
        """
        if wrapped:
            parts = self.unwrap(content)
            result: list[ResponsePart] = []
            for raw_part in parts:
                if isinstance(raw_part.part, TextPart):
                    result.append(raw_part.part)
                elif isinstance(raw_part.part, RawA2uiPart):
                    compiled = self.compile(raw_part.part.a2ui_raw)
                    result.append(A2uiPart(a2ui=compiled))
            return result
        return [A2uiPart(a2ui=self.compile(content))]

    @property
    def supports_streaming(self) -> bool:
        """Whether this parser can read a response incrementally through open_stream.

        Direct JSON and Express both stream incrementally: Direct JSON heals partial JSON
        objects as they arrive, and Express compiles complete top-level statements and
        heals progressive string literals as chunks arrive. A parser that returns False
        here buffers the response and is parsed whole through parse_response.
        """
        return False

    def open_stream(
        self,
        *,
        wrapped: bool = True,
        state: Optional[RendererState] = None,
        after_rejected_block: Literal["end", "continue"] = "end",
    ) -> ResponseStream:
        """Opens a stream that reads one response chunk by chunk.

        Implemented only by parsers whose supports_streaming is True. The default refuses,
        so that a caller handed a non-streaming parser fails at the call rather than
        silently receiving nothing.

        Args:
            wrapped: False if the whole response is one block, without tags.
            state: The renderer state to check the messages against and apply them to. A
                new, empty state for this stream when None.
            after_rejected_block: What the stream does once it rejects a block. "end" ends
                the stream: feed or close raises. "continue" returns an ErrorPart in place
                of the block and reads on: the messages the block emitted before it was
                rejected stay, the rest of the block is skipped, and the text and blocks
                after it are read as usual. A block is rejected when it cannot be read or
                the state rejects one of its messages; any other error ends the stream
                either way.

        Raises:
            NotImplementedError: If this format does not support streaming.
        """
        raise NotImplementedError

ResponseStream

A stream reads one streamed response. It holds the text buffers of the response and the RendererState its messages are applied to. Each message it emits, whether feed and close return it or a raised error carries it in partial_results, has been applied to the state, in the order emitted, and no other message has, so the state mirrors what a renderer that received the emitted parts holds.

@dataclass
class ErrorPart:
    """A block the stream rejected and read past, in place of the block.

    Only a stream opened with after_rejected_block="continue" returns one.
    """
    error: A2uiError

StreamPart = Union[TextPart, A2uiPart, ErrorPart]

class ResponseStream(ABC):
    """Reads one streamed model response, chunk by chunk."""

    @property
    @abstractmethod
    def state(self) -> RendererState:
        """The state the stream's messages are applied to."""

    @property
    @abstractmethod
    def is_in_block(self) -> bool:
        """Whether the response read so far ends inside a block.

        True after an opening tag and before its closing tag, and, for an unwrapped
        response, until the stream ends. Text held back because it could start a tag is
        not inside a block. False once the stream has ended. A caller that joins the chunks
        of a paused and resumed response reads it to tell whether it may add text between
        them, such as a space, which inside a block could break the block.
        """

    @abstractmethod
    def feed(self, chunk: str) -> list[StreamPart]:
        """Reads the next chunk and returns the parts that became ready.

        If it raises, the stream has ended, and the error's partial_results holds the parts
        the chunk made ready before the failure. They have been applied to the state like
        the parts of any earlier chunk, so the caller shows them too.
        """

    @abstractmethod
    def close(self) -> list[StreamPart]:
        """Ends the response and returns the parts still held back.

        Text that could still have started a tag is held until then, and an unwrapped
        response is checked as a whole only then. A block still open is rejected: the
        stream raises a parse error, or, reading past rejected blocks, returns an ErrorPart
        for it. A raised error carries partial_results, as for feed.
        """

An SDK whose type system can say so returns ErrorPart only from a stream opened with after_rejected_block="continue", so that the code of a caller that never asks for one does not have to handle it.

The optional agent_renderer_app feature adds RendererState.of and StreamedAnswer, for an app that is both the agent and the renderer. Such an app checks and applies each message once, on the renderer's own MessageProcessor, and reads a streamed answer into its text, surfaces and rejected blocks in the order the model wrote them.


D. Validation Layer

Validation belongs to the a2ui_core package, and the Agent SDK does not maintain a redundant validator wrapper. Core splits the work across two objects, and an agent reaches validation through the first of them:

  • MessageProcessor (a2ui.core.processing) is the entry point. It holds every active catalog, resolves each component and function call to the catalog it belongs to (the item's own @catalogId or function call catalogId, else the surface default, else A2uiCatalogError), and hands the item to that catalog's validator. It also owns the checks that need more than one item: surface lifecycle, component uniqueness, root reachability, cycles, recursion depth caps, and data binding JSON Pointer syntax. A surface whose components come from several catalogs can only be checked here, which is why this and not the validator is what an agent calls.
  • PayloadValidator (a2ui.core.validation) is scoped to exactly one catalog and checks exactly one item against it, through validate_component, validate_function and validate_theme. It cannot see a whole payload and cannot know which catalog an item belongs to, so it never decides routing.

Wire envelope structure is checked before either of them, by the core version adapter for the protocol version the message declares. That is also where version branching across v0_8, v0_9, v0_9_1 and v1_0 lives, so neither object above is version-aware.

See a2ui_core.blueprint.md for the full validation contract and its implementation matrix.

Surface state during validation

A stateless check sees one outbound payload at a time, the agent-to-renderer messages the agent is about to send, with nothing else to compare it against. When that payload updates a surface it did not itself create, it carries no component tree, so a reference to a component the agent sent in an earlier payload cannot be resolved and is accepted.

An agent that keeps one MessageProcessor across a session and runs its own outbound messages through it holds that tree. References then resolve against the components the surface already has, and cycles are found across the whole surface instead of one payload at a time. The renderer runs these same checks when the payload arrives, so an agent that runs them first catches a bad payload before sending it rather than after.

Agent-callable functions

A function is agent-callable when its catalog allowedCallers is "agentOnly" or "rendererOrAgent" (default "rendererOnly"). Prompt generators MUST list only agent-callable functions as targets of callRendererFunction. Parsers and ResponseStream, or the core MessageProcessor they delegate to, MUST reject a callRendererFunction that names a function that is not agent-callable in the negotiated catalogs with A2uiValidationError. catalog_definition.json already forbids requiresUserActivation: true on agent-callable functions; an SDK MAY also exclude activation-gated functions as a safeguard, as Dart does in prompt_generator.dart.


E. Inference Format Facades (a2ui.inference_format)

InferenceFormatFactory & InferenceFormat

class InferenceFormatFactory(ABC):
    """Abstract interface for constructing InferenceFormat strategies bound to active catalogs."""

    @abstractmethod
    def create_format(
        self,
        catalogs: Sequence[CatalogApi],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
    ) -> "InferenceFormat":
        """Constructs an InferenceFormat instance bound to the provided active catalogs.

        Args:
            catalogs: List of active Catalog instances.
            examples: Optional list of few-shot example turns, each a list of messages.

        Returns:
            An InferenceFormat strategy instance.
        """
        pass

class InferenceFormat(ABC):
    """Coordinator facade pairing a prompt generator (input) and parser (output) for a format."""

    @property
    @abstractmethod
    def prompt_generator(self) -> PromptGenerator:
        """Returns the format prompt generator instance."""
        pass

    @abstractmethod
    def create_parser(self) -> Parser:
        """Creates a new parser instance bound to this format strategy."""
        pass

F. High-Level Application Facade (a2ui.processor)

Catalog Providers

A provider can be constructed with a catalog id and a protocol version. Each is both a default and an assertion: when the document states the same field, the two must agree or the load fails; when the document states nothing, the provider's value is used, and the resulting catalog is no different from one that carried its own.

Neither field has a fallback default. A load fails when nothing names the catalog, since nothing could then address it in supportedCatalogIds or createSurface, and it fails when nothing states a version, since a document read under the wrong version mis-validates every payload written against it. This is what the notes below about catalog_id being undefined in v0.8 and protocol_version being undefined before v1.0 imply: a provider value would otherwise be useless for exactly the documents that need it most.

Every failure a provider raises is a catalog error, whether it comes from a file that is not there, a document that is not JSON, a declaration conflicting with the document, or an id or version nobody stated.

class CatalogProvider(ABC):
    """Abstract base class for loading catalog definitions."""

    @abstractmethod
    def load(self) -> CatalogApi:
        """Loads and returns a Catalog definition instance."""
        pass

class FileSystemCatalogProvider(CatalogProvider):
    """Loads a catalog definition from a JSON file on the local filesystem."""

    def __init__(
        self,
        path: str,
        protocol_version: Optional[ProtocolVersion] = None,  # protocol_version is not defined before v1.0
        catalog_id: Optional[str] = None,        # catalog_id is not defined in v0.8
    ):
        """Initializes the filesystem catalog provider.

        Args:
            path: Absolute or relative filesystem path to the catalog JSON file.
            protocol_version: Optional expected A2ui protocol version for validation.
            catalog_id: Optional expected catalog ID for validation.
        """
        self.path = path
        self.protocol_version = protocol_version
        self.catalog_id = catalog_id

    def load(self) -> CatalogApi:
        """Reads the catalog JSON file and returns a Catalog instance.

        If self.protocol_version or self.catalog_id are defined and the loaded catalog
        has protocol_version or catalog_id specified, verify they match; if they conflict, raise an error.
        """
        pass

class InMemoryCatalogProvider(CatalogProvider):
    """Loads a catalog definition from an in-memory dictionary schema."""

    def __init__(
        self,
        catalog: Mapping[str, Any],
        protocol_version: Optional[ProtocolVersion] = None,  # protocol_version is not defined before v1.0
        catalog_id: Optional[str] = None,        # catalog_id is not defined in v0.8
    ):
        """Initializes the in-memory provider.

        Args:
            catalog: Raw catalog schema dictionary.
            protocol_version: Optional expected A2ui protocol version for validation.
            catalog_id: Optional expected catalog ID for validation.
        """
        self.catalog = catalog
        self.protocol_version = protocol_version
        self.catalog_id = catalog_id

    def load(self) -> CatalogApi:
        """Constructs and returns a Catalog instance from the raw schema dictionary.

        If self.protocol_version or self.catalog_id are defined and the loaded catalog
        has protocol_version or catalog_id specified, verify they match; if they conflict, raise an error.
        """
        pass

CatalogConfig

@dataclass
class CatalogConfig:
    """Configuration model associating a component catalog definition with its transformations.

    Attributes:
        catalog: Base Catalog instance loaded via a CatalogProvider.
        transformers: Optional list of CatalogTransformer rules to apply sequentially.
    """
    catalog: CatalogApi
    transformers: Optional[Sequence[CatalogTransformer]] = None

    @property
    def transformed_catalog(self) -> CatalogApi:
        """Returns the Catalog after applying all configured transformers sequentially."""
        current = self.catalog
        if self.transformers:
            for t in self.transformers:
                current = t.transform(current)
        return current

    @classmethod
    def from_path(
        cls,
        catalog_path: str,
        transformers: Optional[Sequence[CatalogTransformer]] = None,
        protocol_version: Optional[ProtocolVersion] = None,
        catalog_id: Optional[str] = None,
    ) -> "CatalogConfig":
        """Factory method loading a Catalog from disk into a CatalogConfig.

        Args:
            catalog_path: Path to the catalog JSON file.
            transformers: Optional list of catalog transformers.
            protocol_version: Optional expected protocol version for validation.
            catalog_id: Optional expected catalog ID for validation.

        Returns:
            A CatalogConfig instance.
        """
        catalog = FileSystemCatalogProvider(
            catalog_path,
            protocol_version=protocol_version,
            catalog_id=catalog_id,
        ).load()
        return cls(catalog=catalog, transformers=transformers)

A2uiGenerator

class A2uiGenerator:
    """Agent-level generator holding agent-supported catalogs and returning A2uiRequestProcessor instances per request renderer capabilities.

    Attributes:
        catalogs: Master list of CatalogConfig objects supported by the agent.
        examples: Optional list of few-shot example turns shared across sessions.
        factory: Default InferenceFormatFactory used when instantiating processors.
    """

    def __init__(
        self,
        catalogs: Sequence[CatalogConfig],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
        inference_format_factory: Optional[InferenceFormatFactory] = None,
    ):
        """Initializes A2uiGenerator with supported catalog configurations and format factory.

        Args:
            catalogs: List of supported CatalogConfig configurations.
            examples: Optional list of prompt example turns.
            inference_format_factory: Optional default InferenceFormatFactory (defaults to ExpressFormatFactory).
        """
        pass

    def create_processor(
        self,
        renderer_capabilities: A2uiRendererCapabilities,
        inference_format_factory: Optional[InferenceFormatFactory] = None,
    ) -> A2uiRequestProcessor:
        """Creates an A2uiRequestProcessor bound to the specified renderer capabilities.

        Args:
            renderer_capabilities: Capabilities sent by the client renderer. Must not be None:
                a processor is negotiated for one renderer, so a request without capabilities
                is a catalog error here, although resolve_catalogs accepts one. A language
                with non-nullable types enforces this in the signature.
            inference_format_factory: Optional override format factory for this processor.

        Returns:
            Pre-negotiated client-bound A2uiRequestProcessor instance.
        """
        pass

A2uiRequestProcessor

class A2uiRequestProcessor:
    """Central request processor facade unifying multi-catalog capability resolution, prompt rendering, parser creation, and validation."""

    def __init__(
        self,
        catalogs: Sequence[CatalogApi],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
        format_factory: Optional[InferenceFormatFactory] = None,
    ):
        """Initializes A2uiRequestProcessor, resolving active catalogs and instantiating validator and format strategy.

        Args:
            catalogs: List of active Catalog instances.
            examples: Optional list of prompt example turns.
            format_factory: Format factory for instantiating format strategies (defaults to ExpressFormatFactory).
        """
        pass

    @property
    def active_catalogs(self) -> list[CatalogApi]:
        """Returns the list of active negotiated Catalog instances for this processor."""
        pass

    @property
    def examples(self) -> Optional[list[list[AgentToRendererMessage]]]:
        """Returns the ordered list of prompt example turns."""
        pass

    @property
    def prompt_snippet(self) -> str:
        """Format-specific system prompt instruction snippet."""
        pass

    def parse_response(self, content: str) -> list[ResponsePart]:
        """Parses and validates the LLM response into ResponseParts."""
        pass

G. Utility Helpers (a2ui.utils)

resolve_catalogs (a2ui.utils.catalog_resolver)

Negotiates renderer capabilities against a registered sequence of catalogs (Sequence[CatalogConfig]) to select matching active schemas for a session.

A2uiRendererCapabilities comes from a2ui_core and is a multi-version union, the same way AgentToRendererMessage is: it covers each supported protocol version's capabilities shape, including the v0.9 A2uiClientCapabilities. An agent that supports more than one protocol version accepts any member of that union here and must not assume the v1.0 shape.

def resolve_catalogs(
    catalogs: Sequence[CatalogConfig],
    renderer_capabilities: Optional[A2uiRendererCapabilities],
    accepts_inline_catalogs: bool = False,
) -> list[CatalogApi]:
    """Matches renderer capabilities against registered catalogs and returns active transformed Catalog objects.

    Resolution follows these rules:

    - `None`, for a request that carries no capabilities object, activates every
      registered catalog in registration order, since the renderer stated no
      preference. Where the language allows it, the argument stays required
      but nullable, so a caller passes `None` deliberately rather than by
      leaving it out.
    - A present but empty `supportedCatalogIds` with no inline catalogs raises a catalog
      error, because the renderer has said it can render nothing.
    - Active catalogs come back in the renderer's preference order rather than the
      agent's registration order, so the renderer's first choice is the first catalog
      the model reads about.

    The catalogs returned are the transformed ones, so a pruning transformer is visible
    both in the prompt snippet and in what the processor will later accept.
    """
    pass

read_renderer_capabilities (a2ui.utils.renderer_capabilities)

read_renderer_capabilities(metadata, version) (TypeScript and Dart readRendererCapabilities) is a transport-neutral helper that extracts and validates renderer capabilities from an inbound transport metadata mapping. Alongside it, the SDK exports the two transport metadata key constants:

  • A2UI_CLIENT_CAPABILITIES_KEY = "a2uiClientCapabilities" (Dart a2uiClientCapabilitiesKey): transport metadata key for v0.8, v0.9, and v0.9.1 client capabilities.
  • A2UI_RENDERER_CAPABILITIES_KEY = "a2uiRendererCapabilities" (Dart a2uiRendererCapabilitiesKey): transport metadata key for v1.0 and later renderer capabilities.

Extraction and validation follow these rules:

  • For v1.0 and later, read_renderer_capabilities reads a2uiRendererCapabilities keyed by the canonical version string ("1.0"). For v0.9 and v0.9.1, it reads a2uiClientCapabilities keyed by "v0.9" (v0.9.1 reuses the "v0.9" capabilities key). Where an SDK supports deprecated v0.8, it reads a2uiClientCapabilities keyed by "v0.8" or normalizes a bare v0.8 entry ({"supportedCatalogIds": [...]}) to {"v0.8": ...}.
  • When metadata is None/null, when the target version's metadata key or version entry is absent, or when the metadata key or version entry holds null, the helper returns None (undefined / null), indicating A2UI is off for the turn. Another protocol version's metadata key is ignored.
  • Where the language allows untyped version arguments, the helper accepts ProtocolVersion or a version string, canonicalizes patch-zero SemVer strings (such as "1.0.0" to "1.0"), and raises A2uiCatalogError on unsupported or invalid version strings. It also raises A2uiCatalogError when metadata, the metadata key value, or the version entry is not a mapping / plain JSON object, when supportedCatalogIds is missing or not a list of strings, when the version entry contains unrecognized keys outside supportedCatalogIds and inlineCatalogs, or when inlineCatalogs is not a list of valid catalog documents with non-empty catalogId strings.
  • The returned version-keyed capabilities structure ({"1.0": ...}, {"v0.9": ...}, or A2uiRendererCapabilities) is accepted unchanged by resolve_catalogs and A2uiGenerator.create_processor.
A2UI_CLIENT_CAPABILITIES_KEY = "a2uiClientCapabilities"
A2UI_RENDERER_CAPABILITIES_KEY = "a2uiRendererCapabilities"

def read_renderer_capabilities(
    metadata: Optional[Mapping[str, Any]],
    version: ProtocolVersion | str = ProtocolVersion.V1_0,
) -> Optional[A2uiRendererCapabilities]:
    """Reads and validates renderer capabilities for the given protocol version from a transport metadata mapping."""
    pass

4. Inference Format Strategy Implementations

Component envelope keys. On v1.0 surfaces, every Agent SDK component writer (format compilers, typed builders, macro expanders, and streaming readers that emit or re-emit components) writes the optional component envelope mixins only as @catalogId, @accessibility, @metadata, and @checks, and never as the bare keys. A bare catalogId, accessibility, metadata, or checks on a v1.0 component is written only when the component schema declares it as its own property. Readers of existing payloads, such as the Express decompiler, also accept a legacy bare envelope key during the transition when the component schema does not declare a property of that name. Pre-1.0 output keeps bare accessibility and checks. See Component Envelope Keys in the core blueprint.

A. DIRECT_JSON Format (a2ui.inference_formats.direct_json)

Standard A2UI JSON payload format enclosed in <a2ui-json> sentinel tags.

class DirectJsonFormatFactory(InferenceFormatFactory):
    """Factory for instantiating DirectJsonFormat strategies bound to active catalogs.

    `create_format` is called by the SDK with the catalogs of each request, so the
    options of the format are given to the factory and passed on to every format
    it creates.
    """

    def __init__(
        self,
        allowed_messages: Optional[Sequence[str]] = None,
        progressive_keys: frozenset[str] = frozenset(),
    ):
        """Initializes DirectJsonFormatFactory.

        Args:
            allowed_messages: Optional list of allowed payload envelope names.
            progressive_keys: String property keys whose partial values the parser
                heals while streaming. Empty turns healing off.
        """
        self.allowed_messages = allowed_messages
        self.progressive_keys = progressive_keys

    def create_format(
        self,
        catalogs: Sequence[CatalogApi],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
    ) -> InferenceFormat:
        """Constructs a DirectJsonFormat instance bound to the provided active catalogs.

        Args:
            catalogs: List of active Catalog instances.
            examples: Optional list of prompt example turns.

        Returns:
            DirectJsonFormat strategy instance.
        """
        return DirectJsonFormat(
            catalogs=catalogs,
            examples=examples,
            allowed_messages=self.allowed_messages,
            progressive_keys=self.progressive_keys,
        )

class DirectJsonFormat(InferenceFormat):
    """Coordinator facade pairing DirectJsonPromptGenerator and DirectJsonParser."""

    def __init__(
        self,
        catalogs: Sequence[CatalogApi],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
        allowed_messages: Optional[Sequence[str]] = None,
        progressive_keys: frozenset[str] = frozenset(),
    ):
        """Initializes DirectJsonFormat with active catalogs, examples, and format options.

        Args:
            catalogs: Active Catalog instances.
            examples: Optional list of prompt example turns.
            allowed_messages: Optional list of allowed payload envelope names.
            progressive_keys: String property keys each parser heals while streaming.
        """
        self._prompt_generator = DirectJsonPromptGenerator(
            catalogs, examples=examples, allowed_messages=allowed_messages
        )
        self._catalogs = catalogs
        self._progressive_keys = progressive_keys

    @property
    def prompt_generator(self) -> DirectJsonPromptGenerator:
        """Returns the DirectJsonPromptGenerator instance."""
        return self._prompt_generator

    def create_parser(self) -> DirectJsonParser:
        """Creates a fresh DirectJsonParser instance bound to active catalogs."""
        return DirectJsonParser(
            catalogs=self._catalogs, progressive_keys=self._progressive_keys
        )

class DirectJsonPromptGenerator(PromptGenerator):
    """Formats standard JSON schema system prompt instructions enclosed in <a2ui-json> tags."""

    def __init__(
        self,
        catalogs: Sequence[CatalogApi],
        examples: Optional[Sequence[Sequence[AgentToRendererMessage]]] = None,
        allowed_messages: Optional[Sequence[str]] = None,
    ):
        """Initializes DirectJsonPromptGenerator.

        Args:
            catalogs: Active Catalog instances.
            examples: Optional list of prompt example turns.
            allowed_messages: Optional list of allowed payload envelope names.
        """
        self.catalogs = catalogs
        self.examples = examples
        self.allowed_messages = allowed_messages

    def generate_base_rules(self) -> str:
        """Renders the <a2ui-json> output rules, then a schema block with the protocol schemas.

        The protocol schemas are the agent-to-renderer schema (pruned to
        `allowed_messages`) and the common types schema (pruned to what the
        catalogs and that schema use). They belong to the base rules, so text that
        combines several catalogs' instructions holds them once.
        """
        pass

    def generate_catalog_instructions(
        self, include_schema: bool = True, catalog: Optional[CatalogApi] = None
    ) -> str:
        """Renders one schema block per catalog, each holding only that catalog's schema."""
        pass

    def generate_examples(
        self, catalog: Optional[CatalogApi] = None, validate: bool = False
    ) -> str:
        """Renders each example turn as JSON inside <a2ui-json> tags."""
        pass

class DirectJsonParser(Parser):
    """Parser for standard A2UI JSON payload envelopes enclosed in <a2ui-json> sentinel tags."""

    def __init__(
        self,
        catalogs: Sequence[CatalogApi],
        progressive_keys: frozenset[str] = frozenset(),
    ):
        """Initializes DirectJsonParser.

        Args:
            catalogs: Active Catalog instances for validation.
            progressive_keys: String property keys whose partial values are healed
                while streaming. Which properties hold prose depends on the catalog,
                so there is no built-in set; empty turns healing off.
        """
        self.catalogs = catalogs
        self.progressive_keys = progressive_keys

    @property
    def supports_streaming(self) -> bool:
        """Direct JSON reads incrementally, so this parser implements open_stream."""
        return True

    def compile(self, format_content: str) -> list[AgentToRendererMessage]:
        """Parses and fixes JSON payload content string into AgentToRendererMessage objects.

        Args:
            format_content: Raw JSON string extracted from <a2ui-json> tags.

        Returns:
            List of compiled AgentToRendererMessage objects.
        """
        pass

    def decompile(self, a2ui_payload: Sequence[AgentToRendererMessage]) -> str:
        """Decompiles AgentToRendererMessage list into standard formatted A2UI JSON string.

        Args:
            a2ui_payload: List of AgentToRendererMessage objects.

        Returns:
            Formatted A2UI JSON string representation.
        """
        pass

    def open_stream(
        self,
        *,
        wrapped: bool = True,
        state: Optional[RendererState] = None,
        after_rejected_block: Literal["end", "continue"] = "end",
    ) -> ResponseStream:
        """Opens a stream that reads one response, auto-healing progressive_keys in real
        time. See Parser.open_stream."""
        pass

B. EXPRESS Format (a2ui.inference_formats.express)

Compact functional DSL format designed to reduce output token consumption, enclosed in <a2ui> sentinel tags. For formal grammar and syntax specification, see Express Specification and Express Grammar.

The Express format package under a2ui/inference_formats/express/ contains:

  • format: ExpressFormatFactory (subclassing InferenceFormatFactory) and ExpressFormat (subclassing InferenceFormat).
  • prompt_generator: ExpressPromptGenerator (subclassing PromptGenerator, internal to the package and reached through ExpressFormat.prompt_generator), rendering compact positional signatures for catalog components and functions into prompt instructions.
  • compiler: ExpressCompiler, lexing and parsing the DSL expressions found between <a2ui> tags into standard AgentToRendererMessage list structures, both in single-shot mode and incrementally across streaming chunks.
  • decompiler: ExpressDecompiler, converting AgentToRendererMessage payload lists back into Express DSL string format.
  • parser: ExpressParser (subclassing Parser), delegating compilation and decompilation to ExpressCompiler and ExpressDecompiler, and opening ExpressResponseStream sessions via open_stream().
  • response_stream: ExpressResponseStream (subclassing ResponseStream), incrementally unwrapping <a2ui> blocks, scanning statement boundaries, healing progressive string properties, and emitting delta surface messages.
DEFAULT_EXPRESS_PROGRESSIVE_KEYS: frozenset[str] = frozenset({"text"})

class ExpressFormatFactory(InferenceFormatFactory):
    """Factory for instantiating ExpressFormat strategies bound to active catalogs."""

    def __init__(
        self,
        allowed_messages: Optional[Sequence[str]] = None,
        progressive_keys: Optional[Set[str]] = DEFAULT_EXPRESS_PROGRESSIVE_KEYS,
    ):
        """Initializes ExpressFormatFactory.

        Args:
            allowed_messages: Optional list of allowed payload envelope names. The
                prompt teaches only the statements that compile to them.
            progressive_keys: String property keys whose partial values are healed
                while streaming (defaults to {"text"}).
        """
        self.allowed_messages = allowed_messages
        self.progressive_keys = progressive_keys

Incremental Streaming (ExpressResponseStream)

ExpressFormat.supports_streaming and ExpressParser.supports_streaming are True. Calling ExpressParser.open_stream(wrapped=True, state=None) returns an ExpressResponseStream that processes incoming token chunks under five rules:

  1. Statement-Level Boundary Scanning: Inside an <a2ui> block, complete statements (assignments, surface(...), deleteSurface(...), and standalone function calls) are split at top-level statement boundaries outside strings, comments, and nested delimiters, and compiled as soon as they arrive.
  2. Progressive String Healing (progressive_keys): When the trailing incomplete fragment of an open <a2ui> block is inside an unclosed string literal for a component argument whose property name is in progressive_keys (such as Text("partial... or Text(text="partial...), or for a data path assignment whose last segment is in progressive_keys (such as $/text = "partial... or $/items/0/text = "partial...), the stream heals the unclosed string literal and parentheses on a scratch copy and emits the updated component or data path value immediately. The healed value is always a string prefix of the final value, and the full value replaces it when the string closes. Any other data path assignment waits for its statement to complete, as a Direct JSON value under a key outside progressive_keys does. URL prefixes (http://, https://, data:) are never closed early.
  3. Pre-root Gating on New Surfaces: For a surface not yet present in RendererState (and without update=true), $/path assignments and bottom-up child component definitions appearing before root = ... are compiled into the scope session and held until root = ... arrives, then flushed together with root in the initial surface creation batch.
  4. Hybrid v1.0 Envelope & Delta Updates: When an <a2ui> block opens and closes within a single feed() call, it emits the single-shot createSurface(components, dataModel) message. When a block spans multiple chunks, the first ready batch on a new surface emits createSurface (or in v0.9: [createSurface, updateComponents, updateDataModel(path="/", value=...)] in that chunk's A2uiPart), and subsequent chunks emit delta updateComponents (containing only newly compiled or mutated components) and per-path updateDataModel messages.
  5. Forward Component References vs. Value Variables: Forward component ID references in ComponentId and ChildList slots emit immediately using the variable name as the provisional ID (re-emitting the parent component if the referenced child later declares an explicit id="..."). Components that reference an undefined value variable (Event, function call, _template, check, map, or primitive) wait until that variable is defined. At </a2ui>, any still-undefined reference or broken topology raises A2uiCompilationValidationError inside a RendererState transaction. Unlike one-shot parse_response, which checks topology after all blocks in the response have been applied and can recover a single block whose opening <a2ui> tag was omitted, open_stream requires the opening <a2ui> sentinel tag in wrapped=True mode, strips Markdown fences and inline backticks placed either inside or immediately outside <a2ui> and </a2ui>, and checks surface topology as each block closes.

For protocol v1.0 and later, surface(...) and every component call accept one reserved keyword argument, extensions=, that compiles to the envelope's metadata container: createSurface.metadata: {extensions: ...} for a surface and @metadata: {extensions: ...} (ComponentCommon.@metadata) for a component. The compiler validates each key as a UAX #31 identifier and rejects the block otherwise, and the decompiler emits extensions= back for any envelope that carries extensions, reading a component's @metadata (or, during the transition, a legacy bare metadata that the component schema does not declare). The keyword is catalog-agnostic and is never listed in the default prompt instructions, so an agent that does not use extensions spends no prompt tokens on it. There is no metadata= keyword for the envelope: a v1.0 component may declare its own metadata property, so metadata= is that property's ordinary argument when the schema declares it.

Data Model Compilation and Decompilation

ExpressCompiler distinguishes between surface creation scopes and surface update scopes when compiling data path assignments ($/... = ...):

  • Surface creation scopes (root is assigned and update=true is not set): all $/... assignments in the scope are folded into createSurface.dataModel (or a single initial root updateDataModel message with path: "/" in v0.9). Path folding traverses existing lists and auto-vivifies lists when an unassigned segment is followed by a canonical non-negative integer segment (0 through 10000), so overriding a list element or nested item property within the same scope mutates the list in place rather than replacing it with a map. When a path segment already holds a map, a numeric next segment sets a string key on that map, and assigning past the end of a list pads intermediate slots with null.
  • Surface update scopes (root is omitted or surface(..., update=true) is set): each $/path = value assignment emits its own updateDataModel message targeting path: "/path" in statement order, while $/ = {...} (or $ = {...}) targets path: "/" to replace the root data model. Assigning a non-map value to $/ raises a validation error in both creation and update scopes. Re-assigning the same path within a scope replaces the earlier value at the later statement's position.
  • Decompilation: ExpressDecompiler flattens createSurface.dataModel (and the v0.9 initial root updateDataModel coalesced into createSurface) into sorted per-leaf $/... = ... assignments, while standalone updateDataModel messages decompile directly as {path} = {value} ($/ = {...} at root) to preserve replace semantics. Numeric list index segments are accepted in data paths ($/items/0 = ...).

Grammar and parser generation

Express.g4 is the canonical grammar, and every SDK accepts exactly the language it defines. Generating the lexer and parser from it with ANTLR is preferred, so that a change to the grammar reaches each SDK by regenerating rather than by reimplementing. The Python, TypeScript and Dart SDKs do this.

An SDK may write its parser by hand instead, for example when the language has no maintained ANTLR runtime or the SDK avoids third-party dependencies. It then says so where the parser is defined, and a change to the grammar has to be made in that parser too. <format>/compiler.yaml and the round-trip test below check that the parser still accepts the grammar's language.


5. Agent Workflow & Execution Walkthroughs

Code Example

# 1. Agent Startup: Initialize long-lived A2uiGenerator with agent catalogs.
#    Note: Prompt examples passed here are validated internally during processor creation
#    (create_processor) against active negotiated catalogs, raising A2uiValidationError if
#    any example uses components or structures not supported by the active catalog.
generator = A2uiGenerator(
    catalogs=[
        CatalogConfig(BasicCatalog("v1.0")),
        CatalogConfig.from_path("./catalogs/custom_catalog.json"),
    ],
    examples=load_examples("./prompts/examples/**")
)

# 2. In Request Handler: Retrieve pre-negotiated A2uiRequestProcessor matching renderer capabilities
processor = generator.create_processor(renderer_capabilities)

# 3. Invoke LLM to generate the output
llm_output_text = myagent.call_llm(processor.prompt_snippet, request_context)

# 4. Parse and validate output using the processor
response_parts = processor.parse_response(llm_output_text)

# 5. Deliver A2UI payloads to the renderer

6. Conformance Test Plan

To ensure behavioral parity across all SDK implementations (Python, Kotlin, etc.), the project maintains a language-agnostic conformance suite.

The suites covering this document live under conformance/agent/:

Suite Covers
<format>/compiler.yaml Parser.compile: what a block of that format means
<format>/decompiler.yaml Parser.decompile: how messages are written back into that format
<format>/response_parser.yaml Parser.wrap, unwrap and parse_response: where a payload begins and ends
<format>/prompt_generator.yaml PromptGenerator.generate: what the system prompt snippet has to tell the model
direct_json/response_streaming.yaml Parser.open_stream and ResponseStream, for incremental Direct JSON streaming
express/response_streaming.yaml Parser.open_stream and ResponseStream, for incremental Express DSL streaming
catalog_provider.yaml CatalogProvider.load and the checks it makes on a document
catalog_resolution.yaml resolve_catalogs: capability matching, preference order, and inline catalog negotiation
catalog_transformer.yaml ComponentPruningTransformer and FunctionPruningTransformer allowlist filtering
renderer_capabilities.yaml read_renderer_capabilities: transport metadata extraction and entry validation across versions
request_processor.yaml InferenceFormatFactory.create_format, A2uiGenerator.create_processor and resolve_catalogs

Suites under conformance/agent/legacy/ describe the earlier interface of kotlin/agent_sdk_legacy, which runs them. python/a2ui_agent implements the current interface above and runs none of them. The TypeScript agent harness runs the subset of legacy/parser.yaml and legacy/inference_format.yaml that the current suites do not supersede (loader.ts).

Where a suite fixes something this document leaves open, the decision is stated in that suite's header rather than left for a reader to infer from the cases.

For complete setup instructions, test harness requirements, suite descriptions, and schema definitions, see Conformance README.

Format Round-Trip and Specification Example Testing

In addition to the per-format YAML suites, implementations of non-JSON inference formats (such as Express) should implement a systematic round-trip test against the full corpus of catalog golden example files (catalogs/basic/v1/examples/*.json).

Because Express notation is a compact, catalog-driven DSL, decompile-then-recompile round-trips against the full set of specification examples verify that:

  1. Every standard component, property, data-binding, and check rule decompiles into clean notation without error.
  2. Recompiling the decompiled DSL reproduces the original semantic surface envelope, components, and data model.

Refer to python/a2ui_agent/tests/test_specification_roundtrip.py for a reference implementation of this test.