Control tables

WIRE keeps its configuration and history in the app's control schema. Anyone with the wire_user application role can read these tables. Change configuration through the Configuration app or the procedures rather than writing to the tables directly. The procedures validate what you save.

Table	Contents
control.use_case	Registered use cases.
control.agent_config	Agent definitions: type, prompt, model, input view, output table, and version.
control.agent_type	The available agent types, with their defaults.
control.class_category	Categories for triage agents, grouped by class name and version.
control.agent_execution	One row per agent run.
control.work_item_feedback	Feedback people leave on individual work items.
control.app_feedback	Product feedback submitted from the Configuration app.

agent_execution

Every run of run_work_item_agent writes one row. Use it to monitor runs, investigate errors, and attribute cost.

Column	Description
id	Execution ID. Output rows carry it in agent_execution_id.
use_case, agent_name, version	What ran.
execution_status	RUNNING, COMPLETED, COMPLETED_WITH_ERRORS, or FAILED.
execution_start_dtm, execution_end_dtm, execution_duration_sec	When and how long.
inserted_rows, updated_rows, total_upsert_rows	Rows written.
errored_rows	Rows the model returned no result for.
error_message	The distinct row-level errors, or the error that stopped the run.
prompt_tokens, completion_tokens, total_tokens	Token usage.
agent_query_ids	Query IDs of the AI statements, for cost and warehouse lookups.
param_*	The options the run was called with.
system_prompt, llm_model, parameters, response_format, input_view, output_table, ...	The configuration the run used, as it was at the time.

Find the warehouse a run used

Run from a worksheet, not from inside a procedure:

SELECT e.id, h.query_id, h.warehouse_name
  FROM wire_native_app.control.agent_execution AS e,
       LATERAL FLATTEN(input => e.agent_query_ids) AS q
  JOIN TABLE(INFORMATION_SCHEMA.QUERY_HISTORY()) AS h
    ON h.query_id = q.value::VARCHAR
 WHERE e.id = <execution_id>;

Find the credits a run used

Cortex usage reaches ACCOUNT_USAGE after a delay of up to several hours:

WITH exec_queries AS (
    SELECT e.id AS execution_id, q.value::VARCHAR AS query_id
      FROM wire_native_app.control.agent_execution AS e,
           LATERAL FLATTEN(input => e.agent_query_ids) AS q
)
SELECT eq.execution_id, SUM(u.token_credits) AS credits
  FROM exec_queries AS eq
  JOIN SNOWFLAKE.ACCOUNT_USAGE.CORTEX_AISQL_USAGE_HISTORY AS u
    ON u.query_id = eq.query_id
 GROUP BY eq.execution_id;