fabrix-ai 1.7.2

Graph-based agent framework powered by oauth-codex

Fabrix

Language: English | 한국어
API Guides: English | 한국어

Overview

Fabrix is a graph-based agent framework built on top of oauth-codex>=2.3.0. It provides a structured execution graph with streaming events for tool-driven workflows.

Key Features

• Graph-based 3-state execution: reasoning, tool_call, response
• Structured state outputs powered by Pydantic models
• Sequential tool execution with strict payload validation
• Async streaming event API for step-by-step observability
• Multimodal input with explicit message models: TextMessage, ImageMessage

Installation

pip install fabrix-ai

Quickstart

import asyncio

from pydantic import BaseModel

from fabrix import Agent
from fabrix.events import (
    ReasoningEvent,
    ResponseEvent,
    TaskFailedEvent,
    ToolEvent,
)
from fabrix.messages import TextMessage
from fabrix.tools import ToolOutput


class AddInput(BaseModel):
    a: int
    b: int


def add_numbers(payload: AddInput) -> ToolOutput:
    return ToolOutput.json({"sum": payload.a + payload.b})


async def main() -> None:
    agent = Agent(
        instructions="You are a precise assistant.",
        state_models={"reasoning": "gpt-5.3-codex"},
        tools=[add_numbers],
    )

    messages = [TextMessage(text="Use add_numbers to compute 3 + 9")]
    async for event in agent.run_stream(messages=messages):
        print(f"[step={event.step}] {event.event_type}")

        if isinstance(event, ReasoningEvent):
            print("reasoning:", event.reasoning)
            print("focus:", event.focus)
        elif isinstance(event, ToolEvent):
            if event.phase == "start":
                print("tool call:", event.tool_name, event.arguments)
            elif event.result is not None:
                print("tool result:", event.result.model_dump())
        elif isinstance(event, ResponseEvent):
            if event.response is not None:
                print("response:", event.response)
            if event.parts is not None:
                print("parts:", [part.model_dump(mode="json") for part in event.parts])
            if event.response is None and event.parts is None:
                print("response: <empty>")
        elif isinstance(event, TaskFailedEvent):
            print("failed:", event.error_code, event.message)


asyncio.run(main())

State Models

Use state_models to override model per graph state (reasoning, tool_call, response). Keys accept NextState or exact string names. Any state not configured in state_models falls back to default_model (which defaults to gpt-5.3-codex).

Example:

agent = Agent(
    instructions="You are a precise assistant.",
    default_model="gpt-5.3-codex",
    state_models={
        "reasoning": "gpt-5.3-codex",
        "tool_call": "gpt-5.3-codex",
        "response": "gpt-5.3-codex",
    },
    tools=[add_numbers],
)

Message Models

Fabrix input is now list[TextMessage | ImageMessage].

• TextMessage(role: str = "user", text: str)
• ImageMessage(role: str = "user", image: str | Path | bytes, text: str | None = None)
• Unknown message fields are rejected at construction time.

ImageMessage.image accepts:

• remote URL (https://...)
• local path (Path or string path)
• raw bytes (bytes), normalized to a data URL for model calls

Multimodal Input

from fabrix.messages import ImageMessage, TextMessage

messages = [
    TextMessage(text="Describe this screenshot"),
    ImageMessage(image="https://example.com/screenshot.png"),
    TextMessage(text="Focus on errors"),
]

async for event in agent.run_stream(messages=messages):
    ...

Tool Contract

Fabrix accepts tools in this shape:

def tool(payload: BaseModel) -> ToolOutput: ...

• The tool must accept exactly one parameter.
• The parameter type must be a Pydantic BaseModel.
• The return type must be ToolOutput (breaking in v1.2.0).
• Runtime arguments must be a JSON object matching payload fields.
• Extra argument keys are rejected.
• Both sync and async tools are supported.
• ToolOutput.image(...) keeps http(s)/data: values as-is.
• ToolOutput.image(...) normalizes file://, local paths, and bytes to local absolute file references.
• Tool-call argument strictness is enforced by model output_schema with strict_output=True.
• Runtime context is embedded in the system message, and transition/tool constraints are enforced by schema validation.
• During LLM history serialization, reasoning/tool_call/response (and legacy tool_result) records are preserved, and local image references are re-normalized to data URLs.

Event Stream

run_stream(...) yields these event types:

• reasoning
• tool (phase="start" / phase="finish")
• response
• task_failed

reasoning is a step-level decision trace summary, not raw internal chain-of-thought. response events now support both response: str | None and parts (structured text/image/json parts); both fields may be None for an empty response event. Terminate by setting next_state=null in response state.

Migration (Breaking)

run_task_stream(task, images, context) has been removed.

• Before: agent.run_task_stream(task=..., images=..., context=...)
• After: agent.run_stream(messages=[...])

Agent(..., model="...") has been removed.

• Before: Agent(instructions=..., model="gpt-5.3-codex", tools=[...])
• After: Agent(instructions=..., state_models={"reasoning": "gpt-5.3-codex"}, tools=[...])

Mapping:

• task text -> TextMessage(text="...")
• images -> ImageMessage(image="..." | Path(...) | b"...")
• context -> include serialized context in TextMessage.text

Tool migration:

• Before: tool returns str / dict / scalar / arbitrary JSON-like objects
• After: tool must return ToolOutput (for example ToolOutput.text(...), ToolOutput.json(...), ToolOutput.image(...))

Documentation

• API usage guide (English): docs/api.md
• API 사용 가이드 (한국어): docs/api.ko.md
• Korean README: README.ko.md

Examples

• Examples guide: examples/README.md
• Minimal quickstart: examples/minimal/quickstart.py
• Multimodal vision: examples/minimal/multimodal.py
• Data workflow: examples/advanced/data_workflow.py
• Incident response workflow: examples/advanced/incident_response.py
• Complex reasoning workflow: examples/advanced/complex_reasoning_workflow.py

Notes

• Public runtime entry point is fabrix.Agent.
• Execution defaults are fixed internally: max_steps=128 and no public per-tool timeout option.
• On successful completion, the stream ends right after the final response event (next_state=null in response state).
• If max_steps is reached, the stream ends without emitting an additional terminal event.