Skip to main content
Traces are the recording of every agent run. The CLI is the fastest way to query traces at scale — filter by status, date, verdict, or free-text search — and to programmatically trigger verifier passes.

​
Commands

​
traces list

​
Filters

  • --status pending|running|completed|error — run status.
  • --needs-review — only traces flagged for human review.
  • --search <text> — full-text over prompts, tool args, tool responses, answers.
  • --from <YYYY-MM-DD> / --to <YYYY-MM-DD> — date range.
  • --flow <slug> — scope to one Flow.
  • --verdict <verdict> — filter by verification verdict: SUPPORT, PARTIALLY_SUPPORTED, NOT_SUPPORT, unverified.
  • --limit <n> — max rows. Default 50.
  • --page <n> — pagination, 1-based.
Add --json for structured output.

​
traces get

Prints the full trace: every span, every tool call, every retrieval, the model’s outputs, and verification results if any. Always JSON — traces are large and structured. Common uses:
  • Save for offline review: nora traces get tr_abc > trace-tr_abc.json.
  • Extract just the tool calls: nora traces get tr_abc | jq '.spans[] | select(.kind=="tool")'.
  • Diff two runs: diff <(nora traces get tr_a) <(nora traces get tr_b).

​
traces verify

Triggers a verifier pass on the trace. Verifier decomposes the answer into claims, then checks each against retrieved evidence. Async — the command returns immediately with a job ID. Poll traces verification to get the result. Flags:
  • --decompose-model <m> — override the model used to decompose claims.
  • --verify-model <m> — override the model used to check each claim.
Both default to workspace defaults (see Providers).

​
traces verification

Fetches the verifier’s summary — overall verdict, per-claim results, ratio of supported to unsupported claims. Verdicts:
  • SUPPORT — all claims are supported by retrieved evidence.
  • PARTIALLY_SUPPORTED — some claims supported, some not.
  • NOT_SUPPORT — no support found for the answer.
  • unverified — no verification has been triggered.
Returns unverified if the trace hasn’t been verified yet — call traces verify first.

​
traces stats

Workspace-wide counters:
  • Total traces (all time and last 24h).
  • Traces per Flow.
  • Verdict distribution.
  • Error rate.
Good for a quick health check without opening the app. Add --json for the raw numbers.

​
Recipes

​
Find recent failures on one Flow

​
Nightly: run verifier on every trace missing verification

​
Weekly report: NOT_SUPPORT rate

​
Bulk-download traces for offline analysis

​
Attach a trace to a signal

There’s no CLI command for creating signals from traces — use the app for that. But for programmatic feedback, see nora feedback.

​
Trace retention

Traces are retained per Flow settings (default 90 days). Payload retention (raw prompts and tool args) can be shorter for compliance workloads. Once a trace payload is scrubbed, traces get returns metadata only. traces stats respects retention — counters aren’t affected by scrubbing (metadata retains longer than payloads).