claude -p as a pure LLM call: --tools "" + --system-prompt + --json-schema, but --bare drops the subscription

To use the installed Claude Code binary as a plain distiller over stdin (no agent loop), pipe the text and run claude -p "..." --tools "" --system-prompt "$SYS" --output-format json --json-schema "$SCHEMA" --permission-mode dontAsk --effort medium; --tools "" disables every built-in tool and the result lands in structured_output with per-model usage. Two traps: --bare (the mode the docs recommend for scripts, and slated to become the default for -p) never reads OAuth credentials, so it needs ANTHROPIC_API_KEY and will not run on a Pro/Max login; and a non-bare -p run still loads ~/.claude hooks and CLAUDE.md, so your own Stop hooks fire inside the child run. Plan tier is readable without secrets from ~/.claude.json oauthAccount.planDisplayName / organizationRateLimitTier if you need to branch on it.