See Every Subagent Tool Call From Your Phone — and Know If Your Cache Is Working: Claude Code v2.1.251

TL;DR: Claude Code v2.1.251 (August 28) streams foreground subagent tool calls live to Remote Control clients, and /cost now shows per-session cache hit ratio, misses, and re-cached tokens — the two signals you need to monitor a long agentic run from anywhere.

Foreground subagent live streaming. When a subagent runs in the foreground, its tool calls and results — file reads, bash executions, writes — now stream in real time to connected Remote Control clients (phone or desktop companion). Background subagents (the default) still show status only.

This gives you per-step visibility on a running multi-agent session without staying at the terminal. It's most useful for tasks where you want to be able to interrupt mid-operation: security-sensitive file changes, anything involving credentials, or workflows where a bad tool call deep in the chain would be expensive to roll back.

Cache visibility in /cost. A new prompt_cache line shows: hit ratio, misses, tokens re-cached from the last run, and whether the cache is warm or cold for this session. If you're relying on prompt caching to keep session costs down and haven't been able to confirm it's working, this is the output that tells you.

/cost now includes:
  Cache:  82% hit ratio · 4.2K misses · 38K tokens re-cached · warm

Also new: a spend limit bar in /usage and a rate_limits.spend_limit field for status line scripts, for environments behind a Claude apps gateway with a monthly spend ceiling.

When to use foreground vs. background subagents. Foreground streaming is best when you need per-step accountability — security-sensitive operations, anything involving credentials, or workflows where you want to be able to interrupt cleanly. Background subagents remain the right default for parallelized, fire-and-forget work (batch analysis, test runs, research pipelines).

Real limit. Foreground subagent streaming relays through the Remote Control service — each tool call result waits for that relay. On tight agentic loops this adds measurable latency. Don't switch to foreground just for visibility unless you actually need per-step control; the overhead is real, not just theoretical.

Sources: Claude Code v2.1.251 release — github.com/anthropics/claude-code · Claude Code v2.1.250 to v2.1.251 key updates — dev.classmethod.jp · Claude Code sub-agents — code.claude.com