Respan Helicone instrumentation

Trace calls made through Helicone's Python manual logger with Respan. This package instruments the published helicone-helpers HeliconeManualLogger surface without replacing or bypassing Helicone's own logging behavior.

The instrumentation covers:

• HeliconeManualLogger.log_request() and direct send_log() calls
• HeliconeLogBuilder, including streamed chunks and error logs
• OpenAI-, Anthropic-, and custom-model request/response payloads
• Helicone _type=tool, _type=vector_db, and _type=data custom logs
• sync manual logging and the builder's async send_log() lifecycle

Install

pip install respan-ai respan-instrumentation-helicone "helicone-helpers~=1.2.1"

Usage

import os

from helicone_helpers import HeliconeManualLogger
from respan import Respan, workflow
from respan_instrumentation_helicone import HeliconeInstrumentor

respan = Respan(
    api_key=os.environ["RESPAN_API_KEY"],
    instrumentations=[HeliconeInstrumentor()],
)
helicone = HeliconeManualLogger(api_key=os.environ["HELICONE_API_KEY"])


@workflow(name="helicone_manual_log")
def run(prompt: str) -> str:
    request = {
        "model": "custom-chat-model",
        "messages": [{"role": "user", "content": prompt}],
    }

    def operation(recorder):
        response = {
            "model": "custom-chat-model",
            "choices": [{"message": {"role": "assistant", "content": "Hello!"}}],
            "usage": {
                "prompt_tokens": 4,
                "completion_tokens": 2,
                "total_tokens": 6,
            },
        }
        recorder.append_results(response)
        return "Hello!"

    return helicone.log_request(request, operation, provider="openai")


try:
    print(run("Say hello."))
finally:
    respan.flush()
    respan.shutdown()

HeliconeInstrumentor(capture_content=False) keeps span identity, model, provider, timing, usage, and errors while omitting request and response content. Authentication and Helicone transport headers are never copied into spans. Safe constructor/request correlation headers are normalized into canonical customer/session fields and one JSON respan.metadata attribute.

Activation and overlap policy

Helicone manual logging is explicit instrumentation. Activate HeliconeInstrumentor() only in applications that use HeliconeManualLogger; it is not a direct-provider auto-instrumentor. The package patches Helicone's logger and builder methods only—it does not patch OpenAI, Anthropic, or another provider SDK.

Do not combine this package with instrumentation for the provider operation that the same manual log describes unless you intentionally want both records. For example, wrapping an OpenAI call in HeliconeManualLogger.log_request() while also activating OpenAIInstrumentor() produces one provider span and one Helicone manual-log span for the same logical model call.

The instrumentor is reference-counted and restores the exact SDK methods it patched when the final owner deactivates. It also exposes instrument() and uninstrument() aliases for compatibility with OpenTelemetry-style lifecycle code.

See the Helicone example set in respan-example-projects/python/tracing/helicone for deterministic sync, async-builder, streaming, error, tool, vector database, and custom data coverage.