a11-kit 0.6.2__tar.gz → 0.7.0__tar.gz

{a11_kit-0.6.2 → a11_kit-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: a11-kit
-Version: 0.6.2
+Version: 0.7.0
 Summary: A streaming action runtime for AI agents, model serving, and multimodal APIs
 Keywords: agents,generative-ai,llm,model-serving,streaming
 Author-Email: Curiosity Stack <hello@curiositystack.dev>
@@ -61,12 +61,16 @@ Description-Content-Type: text/markdown
 
 # A11
 
-**A streaming action runtime for AI agents, model serving, and multimodal
-APIs.**
+**Build AI applications like normal software.**
 
-A11 gives ordinary asynchronous code a stable boundary for tools, model calls,
-pipelines, and services. Its Python API runs on a native C++20 runtime, with
-TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
+A11 is the open-source runtime for connecting models, tools and interfaces
+while keeping your architecture. It gives ordinary asynchronous code named
+streams, concurrency, tools and local-to-remote execution, and exposes the
+running application to Studio.
+
+The Python API runs on a native C++20 runtime. TypeScript, Kotlin, C++ and Flow
+use the same action boundary where an application needs another language or
+process.
 
 - **Named streams instead of mixed provider events.** Read model text,
   reasoning, tool activity, progress, and durable interaction state from
@@ -77,7 +81,7 @@ TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
   at runtime with defined concurrency, draining, deadlines, cancellation, and
   error propagation.
 
-**[Install in five minutes](#five-minute-quickstart)** ·
+**[Run the first action](#run-your-first-streaming-action)** ·
 **[Try the live research agent][live-demo]** ·
 **[Read the documentation](https://docs.a11.to/)** ·
 **[Star A11](https://github.com/hpnkv/a11)**
@@ -88,10 +92,11 @@ TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
 
 The hosted [parallel research
 agent](https://docs.a11.to/guides/deep-research.html#try-the-deployed-agent)
-plans a topic, runs several investigations concurrently, and streams one report
-to the browser. Its default Ollama backend needs no account or API key.
+plans a topic, searches the web for each brief, reviews the returned evidence
+concurrently, and streams a cited report to the browser. Its default model and
+search backends need no visitor account or API key.
 
-## Five-minute quickstart
+## Run your first streaming action
 
 Install the core runtime on Python 3.11 or later:
 
@@ -152,7 +157,68 @@ pip install "a11-kit[llm]"
 a11 chat --provider ollama --no-voice --no-shell-tools
 ```
 
-## Continue from a working example
+## Add capabilities incrementally
+
+**Streaming.** An asynchronous iterator yields values before it finishes.
+Named outputs give model text, reasoning, progress and structured values their
+own interfaces. The caller reads those streams and checks the action's final
+status separately.
+
+**Concurrency.** Use ordinary Python tasks or C++ runtime primitives for
+independent work. Use nested A11 actions when the runtime should understand a
+parent/child relationship and propagate cancellation.
+
+**Tools.** Expose registered actions to model calls through the SDK, or serve
+the same registry through MCP. The application still owns validation,
+authorization and policy.
+
+**Remote execution.** Serve a registry from another process or machine. The
+caller configures a connection and calls the action; the handler's input and
+output contract stays stable.
+
+Flow is available when an application needs a composition expressed as data.
+It is optional: model APIs, MCP, LangGraph and existing application code can
+remain part of the stack.
+
+## Inspect the work with Studio
+
+[Open Studio](https://a11.to/studio) to explore an endpoint's action schemas,
+run calls and inspect streamed outputs, logs, timings, failures and
+cancellation.
+
+![A11 Studio showing an action and streamed results](https://a11.to/examples/studio-poster.jpg)
+
+## Explore complete applications
+
+- [Deep research](https://docs.a11.to/guides/deep-research.html) plans a topic,
+  searches the web in parallel and streams a cited synthesis.
+- [Model interactions](https://docs.a11.to/guides/llm.html) present one
+  streaming interface across Claude, Gemini, Ollama and OpenAI-compatible
+  providers.
+- The `a11 chat --voice-model ...` path keeps audio capture, transcription and
+  model interaction on separate streams.
+- [Generative media](https://docs.a11.to/guides/generative-media.html) streams
+  progress independently from generated image bytes.
+- [Browser tools](https://docs.a11.to/guides/browser-tools.html) let a model
+  call typed actions implemented inside an existing page.
+
+Each guide names its provider, model and hardware prerequisites. Local actions
+and remote transports do not require a model.
+
+## Install only what the application needs
+
+The Python distribution is `a11-kit`, imported as `a11`. Provider integrations
+are optional extras such as `a11-kit[ollama]`, `a11-kit[claude]` and
+`a11-kit[openai]`. Browser and TypeScript applications use:
+
+```sh
+npm install @curiositystack/a11
+```
+
+The standalone C++20 runtime needs no Python installation. Build and link it
+through its exported CMake targets using the instructions below.
+
+## Go deeper
 
 - [Build a parallel research agent in
   Python](https://docs.a11.to/guides/deep-research.html).

{a11_kit-0.6.2 → a11_kit-0.7.0}/README.md

@@ -5,12 +5,16 @@
 
 # A11
 
-**A streaming action runtime for AI agents, model serving, and multimodal
-APIs.**
+**Build AI applications like normal software.**
 
-A11 gives ordinary asynchronous code a stable boundary for tools, model calls,
-pipelines, and services. Its Python API runs on a native C++20 runtime, with
-TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
+A11 is the open-source runtime for connecting models, tools and interfaces
+while keeping your architecture. It gives ordinary asynchronous code named
+streams, concurrency, tools and local-to-remote execution, and exposes the
+running application to Studio.
+
+The Python API runs on a native C++20 runtime. TypeScript, Kotlin, C++ and Flow
+use the same action boundary where an application needs another language or
+process.
 
 - **Named streams instead of mixed provider events.** Read model text,
   reasoning, tool activity, progress, and durable interaction state from
@@ -21,7 +25,7 @@ TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
   at runtime with defined concurrency, draining, deadlines, cancellation, and
   error propagation.
 
-**[Install in five minutes](#five-minute-quickstart)** ·
+**[Run the first action](#run-your-first-streaming-action)** ·
 **[Try the live research agent][live-demo]** ·
 **[Read the documentation](https://docs.a11.to/)** ·
 **[Star A11](https://github.com/hpnkv/a11)**
@@ -32,10 +36,11 @@ TypeScript, Kotlin, C++, and Flow interfaces for the boundaries that need them.
 
 The hosted [parallel research
 agent](https://docs.a11.to/guides/deep-research.html#try-the-deployed-agent)
-plans a topic, runs several investigations concurrently, and streams one report
-to the browser. Its default Ollama backend needs no account or API key.
+plans a topic, searches the web for each brief, reviews the returned evidence
+concurrently, and streams a cited report to the browser. Its default model and
+search backends need no visitor account or API key.
 
-## Five-minute quickstart
+## Run your first streaming action
 
 Install the core runtime on Python 3.11 or later:
 
@@ -96,7 +101,68 @@ pip install "a11-kit[llm]"
 a11 chat --provider ollama --no-voice --no-shell-tools
 ```
 
-## Continue from a working example
+## Add capabilities incrementally
+
+**Streaming.** An asynchronous iterator yields values before it finishes.
+Named outputs give model text, reasoning, progress and structured values their
+own interfaces. The caller reads those streams and checks the action's final
+status separately.
+
+**Concurrency.** Use ordinary Python tasks or C++ runtime primitives for
+independent work. Use nested A11 actions when the runtime should understand a
+parent/child relationship and propagate cancellation.
+
+**Tools.** Expose registered actions to model calls through the SDK, or serve
+the same registry through MCP. The application still owns validation,
+authorization and policy.
+
+**Remote execution.** Serve a registry from another process or machine. The
+caller configures a connection and calls the action; the handler's input and
+output contract stays stable.
+
+Flow is available when an application needs a composition expressed as data.
+It is optional: model APIs, MCP, LangGraph and existing application code can
+remain part of the stack.
+
+## Inspect the work with Studio
+
+[Open Studio](https://a11.to/studio) to explore an endpoint's action schemas,
+run calls and inspect streamed outputs, logs, timings, failures and
+cancellation.
+
+![A11 Studio showing an action and streamed results](https://a11.to/examples/studio-poster.jpg)
+
+## Explore complete applications
+
+- [Deep research](https://docs.a11.to/guides/deep-research.html) plans a topic,
+  searches the web in parallel and streams a cited synthesis.
+- [Model interactions](https://docs.a11.to/guides/llm.html) present one
+  streaming interface across Claude, Gemini, Ollama and OpenAI-compatible
+  providers.
+- The `a11 chat --voice-model ...` path keeps audio capture, transcription and
+  model interaction on separate streams.
+- [Generative media](https://docs.a11.to/guides/generative-media.html) streams
+  progress independently from generated image bytes.
+- [Browser tools](https://docs.a11.to/guides/browser-tools.html) let a model
+  call typed actions implemented inside an existing page.
+
+Each guide names its provider, model and hardware prerequisites. Local actions
+and remote transports do not require a model.
+
+## Install only what the application needs
+
+The Python distribution is `a11-kit`, imported as `a11`. Provider integrations
+are optional extras such as `a11-kit[ollama]`, `a11-kit[claude]` and
+`a11-kit[openai]`. Browser and TypeScript applications use:
+
+```sh
+npm install @curiositystack/a11
+```
+
+The standalone C++20 runtime needs no Python installation. Build and link it
+through its exported CMake targets using the instructions below.
+
+## Go deeper
 
 - [Build a parallel research agent in
   Python](https://docs.a11.to/guides/deep-research.html).

a11_kit-0.7.0/VERSION

@@ -0,0 +1 @@
+0.7.0

{a11_kit-0.6.2 → a11_kit-0.7.0}/a11/_native/__init__.pyi

@@ -191,6 +191,7 @@ __all__: list[str] = [
     "UDP",
     "VerifiedAuthorization",
     "WEB_FETCH_ACTION",
+    "WEB_RENDER_ACTION",
     "WHOLE_JSON",
     "WIRE_MESSAGE_VERSION",
     "WIRE_STREAM_ABORT_STATUS_HEADER",
@@ -9942,7 +9943,7 @@ def http_actions() -> (
     list[tuple[str, ActionSchema, ActionHandler | NativeActionHandler | None]]
 ):
     """
-    Return the HTTP Actions as (name, schema, handler) triples: make_http_request and web-fetch, in that order.
+    Return the HTTP Actions as (name, schema, handler) triples: make_http_request, web-fetch, and web-render, in that order.
     """
 
 def install_fiber_dump_signal_handler(
@@ -10071,7 +10072,7 @@ def register_audio_actions(registry: ActionRegistry | None) -> None:
 
 def register_http_actions(registry: ActionRegistry | None) -> None:
     """
-    Register make_http_request and web-fetch on `registry`.
+    Register make_http_request, web-fetch, and web-render on `registry`.
     """
 
 def registry_to_json(
@@ -10263,6 +10264,7 @@ TEXT_MIMETYPE: str = "text/plain"
 TLS: TurnRelayType
 UDP: TurnRelayType
 WEB_FETCH_ACTION: str = "web-fetch"
+WEB_RENDER_ACTION: str = "web-render"
 WHOLE_JSON: str = "$"
 WIRE_MESSAGE_VERSION: int = 1
 WIRE_STREAM_ABORT_STATUS_HEADER: str = "x-a11-abort-status"

{a11_kit-0.6.2 → a11_kit-0.7.0}/a11/_native/flow.pyi

@@ -42,6 +42,7 @@ __all__: list[str] = [
     "highlight",
     "parse",
     "plan",
+    "register_standard_actions",
     "request",
     "run_program",
     "scan",
@@ -311,6 +312,26 @@ def plan(source: str, source_name: str = "-") -> dict[str, typing.Any]:
     whole truth would be misled.
     """
 
+def register_standard_actions(
+    registry: ActionRegistry | None,
+    roots: collections.abc.Sequence[str],
+    allow_write: bool = False,
+    allow_run: bool = False,
+    require_sandbox: bool = True,
+    inherit_environment: bool = False,
+    max_seconds: typing.SupportsInt | typing.SupportsIndex = 600,
+    current_directory: str = ".",
+) -> None:
+    """
+    Register the native filesystem and process actions on a registry.
+
+    The policy is captured by the handlers at registration. A caller may narrow one
+    run through action options but cannot widen the roots, write access, process
+    access, environment inheritance, deadline ceiling, or kernel-sandbox requirement.
+    Relative filesystem paths resolve against ``current_directory``. Spawned
+    processes may make outbound network connections by default.
+    """
+
 def request(request: dict[str, typing.Any]) -> dict[str, typing.Any]:
     """
     One request to the language service, answered.