explore-signal-patterns

Explore signal patterns

If this connection lists call_platform_tool instead of the tools this workflow names, find each named tool with search_platform_tools, describe it with describe_platform_tool, then run it with call_platform_tool and the same arguments. Whether a tool exists for you is decided by describe_platform_tool with its exact name, not by tools/list or one page of search results: a name returned in not_found is unavailable to you, so say so rather than guess. The call itself still checks your access. When execute is also listed, a multi-step read may run in execute with tools.search, tools.describe and tools.call on platform--<tool name> paths. End each execute run at a confirmation point and ask the user before any write.

Use authenticated Speakeasy AI Control Plane tools to answer a bounded question about captured activity. First successfully call list_projects through this client's own connection and select the exact project requested by the user.

1. Call describe_analytics_catalog with that project ID. signal_metrics describes the entitled signal-measurement contract and supplies current sensor names, modes, ordered signal membership and signal names. Availability requires Signals Intelligence and project read access; Explore's rollout is independent.
2. Select the sensor answering the user's question and its signal. Shared signal identity does not make different sensors equivalent. State the time window and explicit match rule: a multi-label probability threshold, the exclusive selected signal, or cumulative probability at/above a zero-based ordered level. For an exploratory threshold use 0.5 only when stated clearly; it is not a universal quality or severity standard. Never threshold or average expected_index.
3. Discover labels with list_analytics_dimension_values: use dataset: signals, signal_scope containing the sensor and trigger/run, and dimension: attribute_keys. To discover values, use dimension: attribute and the exact attribute key. present: false means absent; present empty strings are a different group. Names such as department, model_family, inference_provider and model_name are starting labels, not a closed vocabulary.
4. Call run_analytics_query with dataset: signals, explicit from/to, grain: auto (or an admitted UTC interval) and signal_query:
   • selection: sensor_id, signal_id or minimum_level, and threshold;
   • trigger: stream, or manual with one exact run_id;
   • group_by: structured {attribute: key} or {field: name} references;
   • filters: ANDed dimension/operator/value objects; exists and missing take no values;
   • optional groups and points budgets. Use this semantic request rather than composed measures/dimensions. If a physical read is refused, narrow the window. Coarser output alone does not reduce the canonical input scanned.
5. Explain the result using signal_result.metadata, signal_result.summary, signal_result.groups and the complete top-level rows (joined by series_id). Distinguish matching messages, exact distinct sessions, share among measured answers, failed/not_applicable/draft/unmeasured companions, and unknown census coverage. Gaps are unknown, not supported zeroes. Report partial edge buckets and top-group selection. A session can belong to several attribute groups, so distinct-session counts are not additive. Terminal event-level failures such as invalid_evaluation_context emit no sensor outcome. They contribute to unknown census coverage, not to the failed counter or any other sensor-outcome companion.
6. When definitions are mixed or an interpretation needs inspection, call get_signal_definition for an exact contributing sensor and fingerprint. Historical snapshots may be unavailable; current configuration is not a reconstruction of a past measurement. Treat definition, classifier and input changes as possible explanations for a trend.
7. For cross-session questions call find_signal_sessions with the same window, source and predicate, omitting group_by. Presence and breadth count exact conversations; concentration divides top-session hits by all hits linked to conversations. A peer predicate produces separate same-message and same-session joint-answer denominators. Do not interchange those units. compare_previous: true retains the previous values, descriptive share difference and failed comparability gates. Recent windows, sparse observations or changed definitions cannot support an unqualified trend claim. Unknown census coverage remains a limitation even when those gates pass.
8. Request evidence: true only when supporting conversation references help. Results contain bounded permitted examples, never transcripts or an exhaustive session list. Open the referenced conversation in the dashboard for content; an absent example does not establish absence of matching activity.

Return the scope, chosen rule, numerator/denominator, supporting measures, query time and limitations together. These are descriptive observations, not task counts, time worked, verified agent errors or causal findings. Live reads can change with corrections and imports; retain the actual result and request when reproducibility matters. Do not ask for credentials or propose raw SQL access.