Claude Code 2.1.251 adds subagent streaming and cache metrics

New hooks gate model switches, resume hooks estimate re-cache cost, file tools get symlink patches, and a --restricted flag locks down untrusted runs.

Nowline AUG 31 6:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Foreground subagents now stream live

    A foreground subagent's tool calls and results now stream to Remote Control clients in real time, instead of only showing status while they run; background subagents stay status-only. You can finally watch a delegated agent work rather than wait for its summary.

  • Model switches are now hookable

    New PreModelSwitch and PostModelSwitch hooks let you block, confirm, or annotate a model change — catch an accidental jump to Opus mid-session, or log every switch. SessionStart resume hooks now also receive session staleness and an estimated re-cache cost.

  • /cost exposes your prompt-cache economics

    /cost adds a per-session prompt-cache line — hit ratio, misses, tokens re-cached, warm vs cold — and the status line gains a prompt_cache object. Cache misses are where token bills quietly balloon, so this is the number to watch on long sessions.

  • File tools stop trusting swapped symlinks

    Read, Write, and Edit no longer follow a symlink swapped in after the permission check, plugin commands can't point outside their own directory, and Grep/Glob apply deny rules through symlinked paths. Worth updating if you run untrusted plugins or repos.

  • New --restricted flag and per-agent cache TTL

    From v2.1.248: a --restricted flag strips command and code execution and confines file access to the working directory — a tighter sandbox for reviewing untrusted code — while agent frontmatter gains experimental.cacheTtl to set prompt-cache TTL per agent.