Skip to main content
Time-Travel Replay lets you step back inside any recorded run and examine what your agent actually saw and did at each decision point. Because Veritrix captures the complete input and output of every span — not just metadata — you can reconstruct the exact state of a conversation or tool call without rerunning anything. This is especially powerful for the class of bugs that only surface one run in seven, where your logs show a bad outcome but nothing in your code explains why.

​
Why Time-Travel Replay Matters

Non-deterministic AI systems are notoriously hard to debug. The same prompt can produce different outputs across runs due to temperature, model updates, or upstream data changes. Traditional logging catches that something went wrong, but not what the model was actually given or what it returned. Replay solves this by recording the full payload at every span boundary.

Reproduce intermittent bugs

If a failure only happens one in seven runs, Replay lets you open the one run where it did fail and inspect it frame by frame — no need to reproduce it live.

Understand model behavior

See the verbatim prompt the model received, including injected context, retrieved documents, and conversation history, not a reconstructed approximation.

Debug tool call errors

Inspect the exact JSON your agent sent to a tool and the exact response it got back — essential when an API returns an unexpected schema.

Compare divergent outcomes

Load two runs of the same agent side by side and see precisely where their inputs or outputs diverged.

​
What You Can Inspect

Every recorded span exposes the following in the Replay view:
  • Exact prompt sent to the model — the full message array, including system prompt, conversation history, retrieved context, and any injected instructions
  • Exact completion returned — the verbatim model output, before any post-processing your application code applied
  • Tool input and output — the arguments passed to a tool.exec span and the response it received
  • Retry attempts — every llm.retry sub-span, with its own prompt and completion, so you can see how the input changed between attempts
  • Token counts and cost — input_tokens, output_tokens, and calculated cost for that span

​
How to Use Replay

1

Open a trace

Go to the Traces tab in the Veritrix dashboard and click any trace to open its waterfall.
2

Select a span

Click any span row — a llm.call, tool.exec, agent.handoff, or llm.retry — to open the Span Detail drawer.
3

Switch to Replay view

Click the Replay tab inside the drawer. You’ll see the full input on the left and the full output on the right, exactly as they were at that moment in the run.
4

Step through adjacent spans

Use the ← Previous and Next → arrows at the top of the drawer to walk forward or backward through the trace one span at a time, in chronological order.
For llm.retry spans, Replay renders each attempt as a numbered tab (Attempt 1, Attempt 2, …). You can click between them to see how the prompt was modified between retries and whether the model’s output changed.

​
Comparing Two Runs

When the same agent produces different outcomes on different runs, use the Compare feature to find the divergence point.
1

Open the first trace

Open a trace in the waterfall and copy its permalink from the Share menu.
2

Open the comparison panel

In the second trace’s waterfall, click Compare in the top-right toolbar and paste the first trace’s permalink.
3

Read the diff

Veritrix aligns matching spans by name and highlights differences in prompt content, tool arguments, and model output side by side. Spans that exist in one run but not the other are marked as missing.
Comparison is available for any two traces in the same project, regardless of when they ran — as long as both are within your plan’s retention window.
Every span inside a trace has its own stable permalink. To copy one:
  1. Open the Span Detail drawer for the span you want to share.
  2. Click the Share icon at the top of the drawer.
  3. Copy the URL.
Your teammate can open that link and land directly in the Replay view for that specific span, with the full prompt and completion already loaded. This makes async debugging and incident post-mortems significantly faster — paste a span link in Slack and your colleague sees exactly what you’re looking at.
Span permalinks include the full prompt payload in an encrypted reference, but the contents are only accessible to users with access to the project. Do not share span links in public channels if your prompts contain sensitive data.