Custom Framework Integration

Integrate FluxLoop with any agent framework safely and reliably.

Decorator Ordering and Safe Instrumentation

When integrating FluxLoop with external agent frameworks (e.g., ChatKit, custom tools, LangChain), follow these rules to avoid type conflicts and ensure observations are captured reliably.

Core Principle

Framework decorator MUST be outermost, FluxLoop instrumentation inside or within function body.

• If a framework provides its own decorator/wrapper that transforms a plain function into a framework-specific object (e.g., a Tool), that decorator must remain the outermost (top) decorator.
• FluxLoop decorators should be placed inside (below) the framework decorator, or instrument from within the function body using the SDK context APIs.

Two Safe Patterns

Pattern A: Manual Instrumentation (Safest, Framework-Agnostic)

Use get_current_context() and manually push/pop an ObservationData inside the function body. This keeps signatures and framework typing unchanged.

Example: Tool Function

from fluxloop import get_current_context
from fluxloop.models import ObservationData, ObservationType

async def my_tool(param: str) -> dict:
    """Tool function with manual FluxLoop instrumentation."""
    fl_ctx = get_current_context()
    obs = None
    if fl_ctx and fl_ctx.is_enabled():
        obs = ObservationData(
            type=ObservationType.TOOL,
            name="tool.my_tool",
            input={"args": {"param": param}},
        )
        fl_ctx.push_observation(obs)

    try:
        result = {"result": do_work(param)}
        if obs:
            obs.output = result
        return result
    except Exception as e:
        if obs:
            obs.error = str(e)
        raise
    finally:
        if fl_ctx and obs:
            fl_ctx.pop_observation()

Why this works:

• No decorator conflicts
• Framework type system untouched
• Works with any framework
• Complete control over observation lifecycle

Pattern B: Stacked Decorators (Framework Outermost)

If you prefer decorator syntax, stack them with the framework decorator on top.

Example: Framework Tool with FluxLoop

@framework_tool_decorator(description="My tool")
@fluxloop.tool(name="tool.my_tool")
async def my_tool(param: str) -> dict:
    """Tool with stacked decorators."""
    return {"result": do_work(param)}

Important: Order matters!

# ✅ CORRECT - Framework outermost
@framework_tool_decorator(...)
@fluxloop.tool(...)
async def my_tool(...):
    ...

# ❌ WRONG - FluxLoop outermost breaks framework type
@fluxloop.tool(...)
@framework_tool_decorator(...)
async def my_tool(...):
    ...  # Framework sees plain function, raises "Unknown tool type"

LLM and Streaming Calls

For LLM calls (including async generators/streams), use one of these approaches:

Approach 1: Helper Function with Decorator

@fluxloop.prompt(name="prompt.agent", model="gpt-4")
async def _run_llm_prompt(agent, prompt, context):
    """Instrumented LLM call."""
    return await agent.generate(prompt, context=context)

# Use in your code
async def handle_request(request):
    result = await _run_llm_prompt(my_agent, request, ctx)
    async for chunk in result:
        yield chunk

Approach 2: Context Manager

async def handle_request(request):
    """Handle request with instrumented LLM call."""
    with fluxloop.instrument("prompt.agent"):
        result = await agent.generate(request)
        async for chunk in result:
            yield chunk

Real-World Example: ChatKit Integration

Here's a complete example integrating FluxLoop with OpenAI's ChatKit framework:

from agents import function_tool, RunContextWrapper
from chatkit.agents import AgentContext
from fluxloop import get_current_context
from fluxloop.models import ObservationData, ObservationType

@function_tool(
    description_override="Cancel the traveller's upcoming trip and note the refund.",
)
async def cancel_trip(ctx: RunContextWrapper[AgentContext]) -> dict:
    """ChatKit tool with FluxLoop instrumentation."""
    fl_ctx = get_current_context()
    obs = None
    if fl_ctx and fl_ctx.is_enabled():
        obs = ObservationData(
            type=ObservationType.TOOL,
            name="tool.cancel_trip",
            input={"args": {}},
        )
        fl_ctx.push_observation(obs)

    try:
        message = state_manager.cancel_trip(get_thread_id(ctx))
        result = {"result": message}
        if obs:
            obs.output = result
        return result
    except Exception as e:
        if obs:
            obs.error = str(e)
        raise
    finally:
        if fl_ctx and obs:
            fl_ctx.pop_observation()

Common Errors and Solutions

Error: "Unknown tool type: <class 'function'>"

Cause: FluxLoop decorator placed outside framework decorator.

Solution: Reverse decorator order or use Pattern A (manual instrumentation).

# ❌ Wrong
@fluxloop.tool(...)
@framework_tool(...)
async def my_tool(...):
    ...

# ✅ Correct
@framework_tool(...)
@fluxloop.tool(...)
async def my_tool(...):
    ...

# ✅ Also correct (safest)
@framework_tool(...)
async def my_tool(...):
    fl_ctx = get_current_context()
    # manual instrumentation...

Error: Observations Not Captured

Cause: No FluxLoop context active when observation is created.

Solution: Ensure the parent function is wrapped with fluxloop.instrument() or @fluxloop.agent().

# In experiment runner or test harness
with fluxloop.instrument("my_agent_run"):
    result = await agent.run(input_data)  # Tools inside will be captured

Best Practices

1. Default to Pattern A for framework integrations unless you control both decorators.
2. Test decorator order if using Pattern B—verify the framework still recognizes the function type.
3. Always check context before creating observations: if fl_ctx and fl_ctx.is_enabled():
4. Use try/finally to ensure observations are popped even on errors.
5. Capture errors in obs.error for failed operations.

Observation Types

Choose the appropriate observation type:

• ObservationType.TOOL — Function/tool calls
• ObservationType.GENERATION — LLM completions
• ObservationType.AGENT — High-level agent operations
• ObservationType.SPAN — Generic timing spans
• ObservationType.EVENT — Point-in-time events

Verification

After integration, verify observations are captured:

1. Run your agent with FluxLoop enabled
2. Check offline store: <output_dir>/artifacts/observations.jsonl
3. Confirm tool/generation observations appear with correct trace_id
4. Use fluxloop parse experiment to verify structured output

fluxloop test --config config.yaml
# Check artifacts directory
ls results/<experiment_name>/artifacts/
# Should see observations.jsonl

# Parse and analyze
fluxloop parse experiment results/<experiment_name>

Need Help?

If you encounter integration issues:

1. Check decorator ordering
2. Verify parent context exists
3. Enable debug mode: fluxloop.configure(debug=True)
4. Review observation capture in logs
5. Open an issue with framework details and error messages