🤖 adk-agentic-logging

adk-agentic-logging transforms the observability of AI agents from a fragmented stream of text into a structured, queryable dataset.

Its core philosophy is the "Wide Event" model: capturing the entire lifecycle of an agentic interaction—including HTTP context, OpenTelemetry trace headers, and deep ADK internal state—into a single, high-cardinality JSON artifact per request.

By seamlessly bridging web frameworks with Google Cloud’s ecosystem, it enforces a "zero-config" standard that eliminates scattered debug prints, turning complex AI behavior into transparent insights from day one.

✨ Features

• � One Request = One Log Line: Aggregates all context into a single JSON object emitted at the end of the request.
• 🧠 Deep ADK Observability: Automatically captures token usage, tool calls, and session metadata.
• ☁️ Zero-Config GCP: Automatic detection of project ID and environment metadata.
• 🔍 Trace-Aware: Injects OpenTelemetry trace_id and span_id for seamless Cloud Trace linking.
• �🔌 Framework Agnostic: First-class support for FastAPI, Flask, and Django.
• 🧹 Vertex AI Trace Sanitization: Surgically removes large JSON payloads (prompts, tool definitions) from Cloud Trace spans to reduce noise and cost.
• ⚡ Supercharged Tracing: One-line enablement of full Google Cloud Trace with automatic correlation to internal ADK spans.

📦 Installation

# Core library
uv pip install adk-agentic-logging

# With Google Cloud Tracing capabilities
uv pip install "adk-agentic-logging[google-adk]"

⚡ Supercharged Tracing (Unified Observability)

The configure_logging function is your unified entry point for observability. It activates full Google Cloud Tracing or Console Tracing with a single line of code, handling OTel configuration and GCP project detection automatically.

Crucially, adk-agentic-logging is lazy. It will not touch OpenTelemetry or initialize any SDK providers until you explicitly call configure_logging().

Usage Examples

1. Local Development (Console Only)

By default, console tracing is enabled with beautiful Rich formatting.

from adk_agentic_logging import configure_logging

# Console tracing is ON by default
configure_logging()

2. Production (Google Cloud Trace)

Enable GCP export for production. If project_id is omitted, it is auto-detected from the environment.

configure_logging(enable_google_tracing=True)

3. Mixed Mode / Advanced

You can toggle exporters independently and specify an explicit project ID.

configure_logging(
    enable_google_tracing=True,
    enable_console_tracing=False, # Silence console logs in production
    project_id="my-custom-gcp-project"
)

Configuration Parameters

🚀 Quick Start

1. Integrate with your Web Framework

FastAPI

from fastapi import FastAPI
from adk_agentic_logging.integrations.fastapi import AgenticLoggingMiddleware

app = FastAPI()
app.add_middleware(AgenticLoggingMiddleware)

Flask

from flask import Flask
from adk_agentic_logging.integrations.flask import AgenticLogging

app = Flask(__name__)
agentic_logging = AgenticLogging(app)

Django

Add to your MIDDLEWARE in settings.py:

MIDDLEWARE = [
    ...,
    "adk_agentic_logging.integrations.django.AgenticLoggingMiddleware",
    ...,
]

2. Instrument your ADK Agent

Use @instrument_runner to automatically capture internal state from your ADK runners.

from adk_agentic_logging.adk.instrumentation import instrument_runner

class MyAgentRunner:
    @instrument_runner
    def run(self, runner_input, **kwargs):
        # Your ADK agent logic
        return result

3. Enrich with Agent Context

You can explicitly enrich the wide event with any context, such as the user's question and the agent's response, using log_ctx.

from adk_agentic_logging import log_ctx

# In your agent logic or route handler
log_ctx.enrich(
    question="What is the weather in Paris?",
    response="The weather in Paris is sunny with 22°C."
)

Adding these fields makes your logs easily queryable for specific interactions.

Cost Optimization & Sanitization

Traces can become expensive if they capture massive LLM prompts and responses. adk-agentic-logging includes a Vertex AI Sanitizer that automatically strips these large payloads from your traces (Waterfall UI) while keeping them safely in your structured logs.

This gives you the best of both worlds: lightweight, low-cost Waterfall traces and full-fidelity logs for debugging.

The sanitizer is enabled by default when you use configure_logging().

# Everything is handled: OTel, Exporters, and Sanitization
configure_logging(enable_google_tracing=True)

🛠️ How It Works

1. Context Initialization: When a request starts, the middleware initializes a thread-local log context and resolves GCP/OTel metadata.
2. Execution: As your code runs, arbitrary metadata is attached or captured automatically via ADK instrumentation.
3. Aggregation: Content is held in a "bucket" bound to the current request.
4. Emission: On request completion, a single structured JSON blob (HTTP info, performance metrics, ADK stats, errors) is emitted to stdout.

📄 License

This project is licensed under the terms of the MIT license.