WeaveScope

Debug traces

Find a trace and inspect its model calls, tools, payloads, errors, and timings.

Use Tracing to understand one agent run. Start here when a user reports a bad response or when one request is unexpectedly slow or expensive.

Use Monitoring when you need to know whether the same pattern affects many runs.

Find a trace

Open Tracing, select the project that received the data, and choose a time range. You can then search or apply filters.

Quick views include:

View Shows
All Every trace in the selected time range.
Errors Traces with an error status.
Slow > 30s Traces with at least 30 seconds of latency.

The trace table includes status, trace ID, name, user, model, span count, tags, tokens, cost, total time, and start time.

Search matches trace IDs, names, payload previews, errors, user IDs, providers, models, and custom field values. Filters are available for status, type, environment, name, model, tags, user, release, and custom fields.

Inspect a trace

Select a trace to open its inspector. The trace header summarizes its status, kind, name, ID, start time, latency, token usage, cost, user, environment, release, and model when those values are available.

The inspector has two parts:

Area Use
Observation list Move between the steps in the trace.
Detail panel Inspect the selected observation's messages, tools, metadata, payloads, and errors.

Choose tree or waterfall

Use Tree to inspect parent and child relationships:

  • Which model call came from this agent step?

  • Which tool call failed?

  • Did a child observation attach to the expected parent?

Use Waterfall to inspect timing:

  • Which observation consumed most of the latency?

  • Did model and tool calls run sequentially or overlap?

  • Did a child start before or after its parent finished?

For very large traces, WeaveScope shows the first 500 matching rows. Narrow the observation search to jump deeper into the trace.

Read observation details

The detail panel can show:

Tab Contents
Messages Input and output messages, plus final outputs.
Tools Tool calls, arguments, and returned values.
Details IDs, timestamps, metadata, custom fields, provider data, tokens, and cost.
Error Recorded error text for a failed observation.

Tabs appear only when the selected observation contains that type of data.

Payloads can be viewed as text, JSON, XML, or Markdown. Values longer than 65,536 characters are collapsed until you expand them.

Common debugging workflows

Find a failed tool call

  1. Open Tracing and select Errors.

  2. Open the failed trace.

  3. Find the first failed child observation in the tree.

  4. Open Tools to inspect its arguments and output.

  5. Open Error to read the recorded failure.

Explain a slow trace

  1. Open Tracing and select Slow > 30s.

  2. Open the trace and switch to Waterfall.

  3. Find the widest model, tool, retriever, or agent observation.

  4. Open Details to confirm its provider, model, tokens, and metadata.

Verify cost and token usage

  1. Open the model observation.

  2. In Details, check the provider, model, token usage, cached input tokens, service tier, inference geography, and explicit cost fields.

  3. If cost is zero, follow Model pricing .

Reconnect a split trace

If a child run appears as a separate root trace, asynchronous work probably lost the active tracing context. BeamWeaver propagates context through its graph nodes. For application-owned tasks, use the helpers described in BeamWeaver integration .

Allow for processing time

Immediately after ingestion, one view may update before another. If the request was just accepted, wait briefly and refresh. If the trace still does not appear, record the project ID and trace ID, then follow Troubleshooting .