Inspect a run

The main Trace view — a timeline of everything that happened.

Open any Trace to see the full timeline of a single run. Every step is expandable; anything with a payload has a copy-to-clipboard button.

Layout

The Trace view has three columns:

• Left: timeline — every step in order, with duration bars.
• Middle: step details — inputs, outputs, metadata for the selected step.
• Right: context — variables in scope, retrieved chunks, memory ops, verification results.

Resize any column. Collapse the context column when you don’t need it.

Step types

Steps you’ll see on the timeline:

Step	What it shows
Trigger	The incoming payload and scope.
Retrieval	Query + top-K chunks returned. Expand for full ranking breakdown.
Memory read	Query + returned items, with scope.
Memory write	Content + scope + importance.
Model call	Full prompt (system + messages), model output, token counts.
Tool call	Tool name, args the model chose, response, error if any.
Router	Which branch was taken, and why.
Action	Which Actions fired, in order, with results.
Verification	Claims extracted, graph checks, verdict.

Icons and colors let you scan the timeline visually for problems (red = error, yellow = warning, blue = normal).

Expanding a step

Click any step to focus it in the middle column. Details include:

• Timing — wall clock start, duration, wait time (waiting on external services).
• Cost — dollar amount, broken into model/tool/embedding.
• Inputs — everything that went in.
• Outputs — everything that came out.
• Metadata — provider, region, deployment ID.

Long text fields (prompts, tool responses) have a “view full” button that opens them in a resizable modal.

Navigating between steps

• ↓ / ↑ — next / previous step.
• ⇧↓ / ⇧↑ — jump to next / previous step of the same type.
• Cmd+F — search within the Trace (matches steps, payloads, and metadata).

The final answer

At the bottom of the timeline, the final answer sent to the user. Highlighted with:

• Citations — hover over any citation number to see the chunk it references.
• Warnings — if verification demoted or flagged part of the answer.
• Version — which published Flow version handled this run.

Reproducing a run

Right-click a Trace → Replay. Nora re-runs the exact same input against the current Draft (or a specified version). Useful for testing whether a change fixes the issue you’re investigating.

Replay produces a new Trace. Original is unchanged.

Errors and partial failures

Runs that partially succeeded are marked with a mixed status. The failed step is highlighted red; downstream steps show whether they proceeded or were skipped.

Common patterns:

• Tool call failed → Agent retried with a different arg → succeeded.
• Retrieval returned zero results → Agent apologized and finished.
• Model timeout → Flow rolled to fallback model → succeeded.

The Trace makes each visible so you can tell “recovered” from “silently degraded.”