Skip to main content
A signal is an automatic per-trace detection: expressed user emotion, a failed tool call, an effort shortfall, or a false statement to the user. The Signals page (sidebar: Observe → Signals) is where you browse user emotion and tool failures. Laziness and hallucinations are not on this page; see Laziness and hallucinations for where they appear. Detections run in the background after ingest: emotion analysis once a trace has been quiet for about 30 minutes, and tool-failure classification about an hour after it goes quiet. Tool calls themselves are recorded as they arrive.

Page controls

  • Date range — Last 24 hours, 3, 7, 14, 30, or 90 days, or a custom range.
  • Refresh — reloads the active tab.
  • Tabs — User Emotion, Tool Failures.
The active tab is addressable via the type URL parameter (emotion, tool-failures), and each tab supports its own deep-link parameter: family (User Emotion) and tool (Tool Failures).

User Emotion

Detects emotion expressed in the user’s own messages, classified into 7 families and 17 signals: five negative families, one positive family, and Redirection, which is neither (see Emotion taxonomy). Every analyzed trace counts toward the KPIs; the table lists traces with at least one detection or a negative-family risk score of 0.5 or higher. KPIs: Below the KPIs, a strip counts detections per emotion family, Redirection and Positive included; select a family to filter the list (?family=, which takes the stored key, so Redirection is ?family=disapproval). Rows carry a status of Detected when the trace has a negative-family detection, and Positive otherwise, including traces whose only detection is Redirection; Detected only leaves those out. The Confidence column is the detector’s 0 to 1 confidence for the strongest emotion in the row, judged from the user’s own words; higher means a stronger detection of that emotion, not a happier user (a distrust row at 1.00 is strongly distrustful). Rows with nothing detected have no confidence and show their risk score instead, labelled risk. This is unrelated to the 0 to 100 health score on the Users page. Filter by status (All statuses / Detected only), family, elicitor, and trajectory; sort by confidence, risk, last seen, or message count, in either direction; search by trace ID or cause. Opening a row shows the cause, trajectory, elicitor, and verbatim user quotes — each quote links into the trace’s Timeline at the quoted turn.

Emotion taxonomy

Redirection is a calm rejection of something the agent produced or proposed, with no annoyance in it (“no, not that one”, “I don’t like the ending”). A plain edit request such as “make it shorter” is not Redirection; a rejection with heat is Frustration. The stored keys are disapproval (family) and disapprove (signal), so filters, the Data API and the CLI use those names. Each detection also records:
  • Elicitor — what elicited the emotion: bot, external, self, or ambiguous.
  • Trajectory — how it evolved across the trace: none, building, peak, resolved, or sustained.
  • Confidence (0 to 1, from expressed evidence only) and risk score (0 to 1, likelihood from all signals including observed ones; negative families only, so Redirection and Positive carry none), a primary cause, and supporting quotes.

Tool Failures

“Tool invocations that returned errors or failed to execute.” Tool calls are recorded at ingest and then classified into outcomes and named failure categories per tool. The failure list and moda detections tools count classified failures. The Signals failure-rate KPI is classified failures / recorded tool receipts (side_effect_receipts); moda detections tools divides classified failures by classified events. The Tools catalog and moda tools measure receipt success, so these rates can differ. KPIs: Total Failures, Failure rate (per tool call by default; you can switch it to per turn, task or trace), and Tools Impacted. A Failures over time chart plots failure events (stacked or grouped), and a per-tool drill-down is deep-linkable with ?tool=. Each failure category carries: Individual failure events link into the trace at the failed call. Tool-level success rates and usage live on the Tool Catalog page.

Laziness and hallucinations

Moda still detects laziness and hallucinations, but the Signals page has no tab for them. Laziness is checked on every trace. Hallucinations come from the Problems reader, so in a workspace set to analyze a sample of its traces, only the sampled traces are checked. Emotion analysis and tool-failure classification follow the same sample. You’ll find them here: See Analysis for how each detection works.

Next steps

  • Problems — signals grouped by root cause and ranked by impact.
  • Home — the same quality rates as at-a-glance KPIs.
  • Traces — where every signal’s evidence links land.
  • Data API: signals — query emotions, tool failures, and hallucinations programmatically.
  • Analysis — how each detection works, including laziness and hallucinations.