fix: honor gen_ai payload capture disable regardless of OTel env vars

[shreyas70]

Problem

gen_ai_input.message keeps showing up in UDP for services that set GenAI payload capture to false (HARNESS_GEN_AI_PAYLOAD_CAPTURE_ENABLED=false / gen_ai.payload_capture_enabled: false).

The flag only ever worked in the enable direction. Three leaks:

1. genai_env.maybe_set_genai_payload_capture_env_vars() checked pre-existing OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT / OTEL_SEMCONV_STABILITY_OPT_IN before reading config and bailed out. On disable it also just returned without forcing anything off. Any deployment that sets those OTel vars itself (shared base image, another auto-instrumentation layer) keeps capturing content, so OpenAI / Anthropic / Google GenAI wrappers emit gen_ai.input.messages.
2. The LiteLLM wrapper set gen_ai.input.messages gated only on litellm.turn_off_message_logging / otel_logger.message_logging — it never consulted Harness config.
3. The MCP mirror had the same "already set, skip" hole for TRACELOOP_TRACE_CONTENT.

Fix

Payload capture is a privacy control, so false must mean false: config now wins over the environment in the disable direction, while the enable path keeps supplying defaults only (user-set env still wins).

• genai_env.py — config checked first; disabling forces OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=NO_CONTENT and patches the cached semconv stability mode for GEN_AI to DEFAULT, overwriting whatever the environment had.
• litellm/__init__.py — explicit payload_capture_enabled gate before setting gen_ai.input.messages.
• mcp/gen_ai_mirror.py — force TRACELOOP_TRACE_CONTENT=false when disabled, even if preset.
• New GenAiPayloadScrubSpanProcessor, wired as the outermost layer of the default pipeline. When capture is disabled it strips gen_ai.input.messages, gen_ai.output.messages, gen_ai.system_instruction, gen_ai.prompt*, gen_ai.completion*, traceloop.entity.input/output at on_end. This is the backstop for third-party OTel contrib instrumentations the SDK does not wrap. Cheap no-op when capture is enabled.

Test plan

• ./scripts/run-unit-tests.sh — 219 passed, 1 skipped. The 2 gRPC failures (test_grpc_1, test_grpc_2) reproduce on unmodified main in the same sandbox (local port binding), unrelated to this change.
• New coverage: capture disabled with OTel vars preset forces NO_CONTENT; LiteLLM emits no input messages when disabled; MCP forces TRACELOOP_TRACE_CONTENT=false; enabled path leaves user env untouched; scrub processor strips attributes when disabled and no-ops when enabled.
• Verify on a service with HARNESS_GEN_AI_PAYLOAD_CAPTURE_ENABLED=false that gen_ai_input.message no longer reaches UDP.

Note

test/conftest.py's reset_singletons does not reset the module-level _applied flag in genai_env.py, so in a full run that function's logic executes only once. The new tests reset it locally. Worth a separate cleanup.

Made with Cursor

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}