Two Claude Code Config Tricks You Probably Missed: /config from the Prompt and Tool(param:value) Scoping

TL;DR: /config thinking=false flips any Claude Code setting mid-session without restarting; Tool(param:value) in permissions scopes allow/deny by what gets passed to a tool, not just its name.

Both shipped quietly in the June 15–17 releases and are worth knowing.

/config key=value from the prompt (v2.1.181)

Any Claude Code setting can now be changed from the chat interface — no menu, no editing settings.json, no session restart:

/config thinking=false
/config maxTokens=32000
/config wheelScrollAccelerationEnabled=true

Works in interactive mode, in -p (print/non-interactive), and in Remote Control. This matters most in scripted pipelines where you want per-invocation settings without rewriting config files — pass /config maxTokens=8000 before a bulk summarization loop, then /config maxTokens=64000 for the analysis step, all in one script.

Bonus from v2.1.181: Set CLAUDE_CLIENT_PRESENCE_FILE=/tmp/claude-here to suppress mobile push notifications while you're at the machine. Touch the file on connect, delete on disconnect — notifications only fire when it doesn't exist.

Tool(param:value) permission scoping (v2.1.178)

Permission rules now match on tool input parameters, using * as a wildcard:

{
  "permissions": {
    "allow": ["Agent(model:sonnet)"],
    "deny": ["Agent(model:opus)"]
  }
}

This lets Sonnet subagents run automatically while requiring a confirmation prompt for Opus subagents — useful for cost control in long unattended runs where a subagent might reach for a heavier model unnecessarily.

Bonus from v2.1.183: Add "attribution": { "sessionUrl": false } to your settings.json to omit the claude.ai session URL from commit messages and PRs — useful if you don't want session links persisted in public repos.

Sources: Claude Code changelog — code.claude.com, v2.1.181 full release notes — DevelopersIO