This is a replay of a real run, not a mock-up. The transcript below is read from the runtime's own session log — every tool call, its arguments, and the (redacted) result it got back — then played back at the pacing it originally ran. The same gold set the production pipeline gates re-runs with scores the final answer at the bottom.
Press Replay run to stream the transcript.
dsh --profile headless. The agent decides what to do with it — the
trace below is its choices, not a script.
Replay the run to reveal the answer.
Two answers to the same task, graded by the same gate. Run A is the run recorded above, replayed exactly as it happened. Run B is a re-run of the same task after the prompt was corrected — same tools, same table allowlist, no new capability. Read the verdicts, then read the judge line under each one.
Score both versions to see the gate's verdict and per-criterion evidence.
scripts/record_harness_demo.py from a genuine
session log, and tool payloads are reduced to column facts (names, null coverage, cardinality)
before they are committed — no row contents are published.
AgentSpec over a shared runtime, bounded MCP tool surfaces,
actor-attributed feedback, and a gold-set gate on every re-run.