Single-call replay
Replay re-runs one recorded llm span against Anthropic or OpenAI. Pick another model, edit
the prompt, paste your provider key for that request only, and compare recorded vs replayed output
side by side.
This is a re-run of the recorded prompt — not the same call, not a regression test, and not a Morse verdict or quality score. Fidelity flags above the diff say what was and was not reproducible (redacted input, tools not replayed, unknown sampling params, cross-provider, nondeterminism).
Your provider API key is held in the browser tab for that request, sent once over TLS, and never written to Morse storage, logs, or analytics. Closing the panel clears it.
When a span cannot be replayed
- The span is not type
llm - Payloads were shed over quota (
payload_shed) - No recoverable messages/prompt in the stored input
Isolation
Replay results live in their own table. They do not create traces, do not add to agent spend or cost rollups, and do not affect quality verdicts, eval pass rates, or judge budgets. The cost shown is an estimate from the price catalog, borne by the key owner.
Demo mode
In demo mode the Replay opener is write-guarded. A seeded comparison may be available to view — no provider call is ever made from demo.