Calls, failures and time per tool in a period

Counts the tool calls whose tool_use event was stored in [since, until), per tool name, the most-called first and at most 200 names. scope, since and until work as in GET /v1/usage/summary; agent_id narrows to the sessions pinned to a release of that agent. Subagent calls count too.

A call is timed from its tool_use to the tool_result paired with it in the same session and attempt by tool_use_id. A call with no result yet, or with either event's time missing, counts in calls but in no _ms figure. failed counts results with is_error: true; a result without is_error is a success, and one whose is_error is not a boolean is neither. median_ms is read from a histogram: the middle call's time, or the mean of the two middle calls' for an even count, each read within 5% or 1 ms, except that a call of about 65 minutes or more reads as about 65 minutes, so a median that takes in such a call reads low. It is never above max_ms. Over whole UTC days max_ms is the largest time a call has had: if an earlier result later replaces a call's time, it does not come down there. Figures come from a table the reconciler fills on each pass, so a call shows up within one pass of being stored, usually seconds. Each finding names a row of tools: most_time the largest total_ms, fails_often the highest failed / calls, slowest the largest median_ms; a kind no tool qualifies for is left out. description is the line an MCP server gave its tool when a session in the same scope started, the newest such session's, redacted, on one line and at most 200 characters; built-in tools, and tools no session in scope described, have null. The response names tools and figures only, never a session. It changes nothing.

Scope: sessions:read (API key or cookie session).