Reproducible trace evidence · MIT + CC BY 4.0

Langfuse agent and tool tracing fixture

Inspect one deterministic agent task with retrieval, generations, tool retries, a terminal child error, source-side redaction, and an unavailable telemetry exporter. The bundle includes the code, tests, sanitized result, and diagram.

The evidence covers Python SDK 4.14.5 with one self-hosted Langfuse 3.205.0 instance. Recorded model and tool responses make the trace shape repeatable; they are not a provider or latency benchmark.

Files

workflow.py Deterministic agent workflow and safe trace sink langfuse_adapter.py Langfuse v4 adapter with explicit parentage test_workflow.py Workflow, retry, redaction, and isolation tests test_langfuse_adapter.py Adapter contract tests run_integration.py Real-server runner with minimized output test_sanitized_evidence.py Stored-evidence and credential checks sanitized-trace-shapes-2026-08-26.json Reviewed trace shapes from three scenarios sanitized-trace-shape-2026-08-26.svg Diagram generated from the sanitized JSON render_trace_diagram.py Deterministic SVG renderer evidence-run-2026-08-26.md Versions, assertions, defects, commands, and limits README.md File map and independent rerun instructions

Separate attempts

A timeout and its successful retry remain two tool observations, so recovery does not erase the original failure.

Bounded evidence

The runner removes inputs and outputs, copies metadata through an allowlist, and rejects synthetic identifiers and Langfuse keys before writing JSON.

Failure isolation

The unavailable-exporter case verifies that this fixture returns the same application result even when trace delivery fails.

Verified scope

Fifteen standard-library tests check parentage, retry shape, child and root levels, redaction, repeated runs, exporter isolation, the stored JSON, and the SVG. The remote evidence run covered happy, retry, and terminal scenarios.

MIT license CC BY 4.0 license Read the tested tracing pattern