Skip to main content

Deterministic replay — re-run yesterday's event

Advanced ⏱ 15 min replay · trajectories · debugging

You will build: nothing — you'll interrogate something you already built. Every event that crosses a pipeline leaves a recorded trajectory (each hop: node, engine, input, output, LLM/MCP calls). You'll list them, replay one for real (each hop re-executes the deployed agent through the live runtime), replay it --mocked (the recorded walk-through — byte-deterministic, zero external calls), and diff the runs to see exactly where behavior changed.

event (yesterday) → trajectory recorded → replay (real re-execution, default) · replay --mocked (recorded, identical) → diff

Prerequisites​

  • Any deployed app that has processed at least one event — the quickstart's hello works.
  • pulse replay requires Pulse ≥ the build carrying B-198 (check: pulse replay --help).

1 · Find a recorded trajectory​

pulse replay list --pipeline hello
hello — 3 trajectorie(s), newest first:
  t-9f2…  OK  2 hop(s)  2026-07-14T08:12:03Z
  …

2 · Inspect the trip, hop by hop​

pulse replay show t-9f2…

Each hop shows the node, its engine, duration, and the output it produced — the flight recorder for one event.

3 · Replay it — real re-execution, by default​

pulse replay run t-9f2…
pulse replay diff t-9f2… r-58…

The default is live: every hop really re-runs the deployed agent through the runtime — the same reasoning engine real traffic uses (LLM calls, rule evaluation and tool resolution actually happen). The diff highlights any hop where the graph, an LLM answer, or an external API differs from the recorded run:

  hop 1 [judge] DIVERGES
    a: {"likelihood":"high",…}
    b: {"likelihood":"medium",…}

4 · The recorded walk-through — determinism, proven​

pulse replay run t-9f2… --mocked
pulse replay diff t-9f2… r-41…
identical — 2 hop(s), no divergence

--mocked opts into the recorded mode: every hop returns the original captured output — no LLM re-billed, nothing re-sent, byte-deterministic given an unchanged graph. That empty diff is the determinism proof.

What just happened​

You debugged with evidence instead of reproduction: the recorded trajectory is the ground truth of what happened; the live replay + diff isolates precisely which hop behaves differently today — because each hop truly re-executed through the runtime; and the mocked walk-through re-reads the record for free. For LLM stages this is the difference between "it answered differently once" and seeing the divergent hop.

Honest boundaries​

Live replay re-executes side-effecting nodes for real: point sinks at a staging destination before live-replaying anything that acts on the world. Mocked replay is byte-deterministic given an unchanged graph — edit the pipeline and the replay follows the new shape.

Next​