Reference Harness 0.1 controlled snapshot

Loop

AgentRunner records session start, requests turns from Scripted Model, executes tool calls, appends tool results to the next request, and stops when a turn has no tool calls.

Context

This experiment has fixed system, user, assistant, and tool messages. It performs no compaction and records no hidden reasoning in Trace.

Tools

Arguments are checked against the declared tool schema before the read-only read_file handler runs. The fixture is copied into a temporary workspace, so the canonical input is not mutated.

Safety

No shell, write, or network tool is exposed. The path boundary rejects absolute paths and workspace escape, but it is not operating-system isolation.

Reliability

Both repetitions must pass event-order, tool-result, final-answer, fixture-integrity, and deterministic-replay metrics.

Extensibility

Model and ToolRegistry are deliberately small interfaces. A native adapter can replace the input boundary, but this controlled result cannot be relabeled as native evidence.

Orchestration

This snapshot has no subagent support. The dimension remains unknown.

Interfaces

The public entry point is python3 labs/runner.py reference-tool-roundtrip-v1; --check rebuilds and compares artifacts without writing them.

Observability

Each run emits Trace 0.1 JSONL, a structured Result, and a readable report. The site projects those canonical artifacts into an experiment page and per-event replay.