An execution plan composed of ordered steps.

Output of a single :class:`Scorer` invocation. ``scorer`` is the registered name of the scorer that produced this result — distinct from ``Sc

Per-task operational metrics derived from a trajectory.

`with Timer() as t: ...; t.ms` — live millisecond wall time for the metadata block.

Executes plan steps by routing tool calls to MCP servers.

Result of executing a single plan step.

Records every generate() prompt and returns a canned response.

Bearing physical dimensions.

Agent runner that delegates to the claude-agent-sdk agentic loop. The sdk handles tool discovery, invocation, and multi-turn conversation aga

Agent runner that delegates to LangChain's deep-agents framework. The deep agent handles planning, tool invocation, and multi-turn conversati

Run a batch of scenarios against their saved trajectories.

Agent runner that delegates to the OpenAI Agents SDK agentic loop. The SDK handles tool discovery, invocation, and multi-turn conversation ag

One assistant turn: text output, tool calls, and token usage.

Entry-point for plan-and-execute workflows using MCP servers as tool providers. Usage:: from agent import PlanExecuteRunner from

Result returned by any AgentRunner.

Result of automated fault classification.

Closure-style scorer that holds an :class:`LLMBackend`.

Configuration for one benchmark method.

A single tool invocation made by the agent.

Full execution trace across all agent turns.

Decomposes a question into a structured execution plan using an LLM.

Return type for :meth:`LLMBackend.generate_with_usage`. ``input_tokens`` / ``output_tokens`` are ``0`` when the backend can't report usage (e

A single step in an execution plan.

Per-key comparison between gold and model output.

Append each span batch to *path* as one OTLP-JSON line. Thread-safe via a write lock so concurrent `export()` calls from the `BatchSpanProces

LLM backend using the native ``openai`` SDK against a compatible router. Args: model_id: prefixed model string, e.g. ``"tokenrouter/MiniM

Agent runner that delegates the agentic loop to ``opencode run``.

A simple model-only runner with no MCP tool calls.

Implements the small CouchClient interface against a dict, including the subset of Mango operators the tools use ($in, $gte, $lte, $elemMatch).

Evaluation scorer wrapper for the trajectory-based pipeline.

Calculated bearing characteristic frequencies at a given RPM.

A stored signal with metadata.

Simple in-memory store for vibration signals.

LLM backend using the litellm library. Args: model_id: litellm model string with provider prefix, e.g.: ``"watsonx/meta

Deterministic LLM that returns a canned response — no network calls.

Trajectory plus raw OpenCode JSON events for debugging parser drift.

Final result from the plan-execute orchestrator.

Resolved endpoint + key for a proxy router.

Returns responses in order across successive generate() calls.

Spectral amplitude at shaft-frequency harmonics.

Structured score for one gold/model answer pair.

Run a question through a Stirrup agent against the MCP servers. Args: llm: Unused; accepted for :class:`AgentRunner` interface parity.

ToolUniverse-style facade over the AssetOpsBench MCP servers.

Wraps an :class:`LLMBackend` and sums token usage across calls. ``Planner`` / ``Executor`` call ``generate()`` and only need a string; this w