Patch release: headless JSON output can report usage and cost, bundled multi-agent skills keep temp artifacts in a private per-user scratch directory, and rg is treated as a read-only search command.
Changelog
[new] Headless json results (and stream-json end / error events when spend was recorded) can include usage and cost when the prompt reached the model: usage, num_turns, modelUsage, total_cost_usd, and total_cost_usd_ticks. Docs cover uncached vs cache-read token fields, how subagents and turns are counted, when cost is omitted or incomplete, and that missing cost means unreported—not free.
[fix] Bundled skills design, implement, execute-plan, and review no longer write orchestration artifacts straight under shared /tmp. They create a mode-700 per-user directory under $TMPDIR (falling back to /tmp) and place state, summaries, reviews, diffs, and helper temps there so multi-user hosts do not leak skill files or ignore a configured TMPDIR. Resume paths recompute the same scratch directory. pr-babysit uses $TMPDIR for its review-thread temp files as well.
[changed] Permissions treat rg as a read-only search/inspection command alongside grep. rg --pre / rg --pre=… stay blocked because they spawn a preprocessor per file.
[changed] User-guide pages for headless mode and permissions/safety updated for the spend fields and rg policy.
Under the hood
Bundled content pack version advanced to public-2026-07-10-r2.
Remote settings may include an auto_interject_on_task_wait flag for client behavior.
Upgrade notes
Headless automations that parse json or stream-json should tolerate optional spend fields on success (and sometimes on prompt-level errors). Prefer the documented headless token fields for spend tooling; use total_cost_usd_ticks when you need exact integer reconciliation.
Anything that assumed fixed skill artifact paths under /tmp/grok-* should follow the new scratch layout (or stop depending on those paths).
