LoongSuite WildToolBench Instrumentation

OpenTelemetry instrumentation for the WildToolBench benchmark framework.

Installation

WildToolBench is not available on PyPI. Install it from source:

pip install -e /path/to/WildToolBench/wild-tool-bench
pip install loongsuite-instrumentation-wildtool

Requirements

• OpenAI provider instrumentation: To produce LLM spans, you must also enable an OpenAI provider instrumentation (e.g., opentelemetry-instrumentation-openai or LoongSuite's equivalent). This plugin creates ENTRY/AGENT/CHAIN/STEP/TOOL spans but does not create LLM spans itself.

Usage

from opentelemetry.instrumentation.wildtool import WildToolInstrumentor

WildToolInstrumentor().instrument()

# Run WildToolBench as usual — spans are automatically generated.

Span Topology

ENTRY (enter_ai_application_system)
└── AGENT (invoke_agent wildtool)
    └── CHAIN (workflow task_{idx})
        └── STEP (react step)
            ├── [LLM span — provider instrumentation]
            └── TOOL (execute_tool {tool_name})

Patch Points

Round 2 fixes (see llm-dev/execute.md § "修订记录 (Round 2 fix)")

• H1: TOOL span is now parented on STEP, not CHAIN. Strategy A enhanced — the chain wrapper holds a round → STEP span map and uses trace.set_span_in_context(step_span) to anchor each post-hoc TOOL span on the matching STEP. STEP SpanContexts remain valid parents even after end().
• H2 (provider-name fallback): opentelemetry-instrumentation-openai-v2 == 0.62b1 only emits the legacy gen_ai.system attribute on its LLM span; the new gen_ai.provider.name attribute is missing. As a pure fallback the wildtool plugin writes both gen_ai.system="openai" and gen_ai.provider.name="openai" on the STEP span (not on the LLM span — that is owned by the OpenAI v2 instrumentation and we do not patch it). Once the OpenAI v2 instrumentation upstream emits gen_ai.provider.name natively this fallback can be removed.
• M1: CHAIN span now carries input.value (last user message in inference_data["messages"], truncated to 4096 chars) and output.value (JSON of action_name_label/task_idx/is_optimal).
• M2: STEP span now carries gen_ai.react.finish_reason on error paths. Mapping table is in execute.md § "M2: gen_ai.react.finish_reason 取值映射".
• M3: TOOL span explicitly writes gen_ai.tool.call.arguments / gen_ai.tool.call.result / gen_ai.tool.description, bypassing OTEL_INSTRUMENTATION_GENAI_CAPTURE_* gating in opentelemetry-util-genai. The custom wildtool.tool.execution_mode = "ground_truth_replay" is preserved.