Claude Code System Prompts

@internal Subagents started through the Agent tool in this session, as running totals (forked skills, workflows, teammates and other internal agents are not counted; two sessions in one process keep separate counts). Cumulative like modelUsage: read the latest result rather than summing across results; a resumed session starts fresh and a mid-session /clear zeroes it, though a background subagent that outlives the /clear still records its outcome, so completed, failed and killed can then exceed spawned. A result that was held back while background subagents finished carries the counts as of when it is written to the stream, as do its total_cost_usd, duration_api_ms and modelUsage (and usage where that is a running total; a per-turn main-loop usage keeps its turn-end value). Omitted by hosts without a local Agent tool and on crash/startup-error results; all zeros until the Agent tool starts or refuses a subagent. A remote (isolation: remote) launch counts as spawned but never reports an outcome. Per-subagent detail is on the task_started / task_notification events.