limitcast 0.18: hold a duplicate subagent, cap a fat tool call, and warn before a cold cache

Sam Stull · 1 November 2026

0.18 acts on the four places a week of Claude Code usage actually moves the window: subagents, fat tool results, cold headless starts, and the prompt cache after an idle pause.

Install the current release, then run limitcast install --guard --alerts and restart. The duplicate hold and the 5-minute notice are in the guard. The fat-tool check is in alerts.

A second copy of the same subagent

The guard already held a new subagent when a window was about to run out. 0.18 also holds a second subagent of the same type and the same prompt while the first is still running in that session. Claude's stop hook clears it. Codex and Grok have no matching stop event, so a record older than 30 minutes is dropped. LIMITCAST_ALLOW=1 still spawns.

A fat tool call, before it runs

The 20k-token notice still tells you after a huge result. HubSpot-sized calls often land under that line. After limitcast report --why or limitcast report --jobs, a Read, Grep, or MCP tool whose results averaged at least 5,000 tokens over 3 or more calls is held when the call sets no limit. A limit that is only too high is lowered: 400 lines for Read, 100 for MCP. A fat Grep with no path and no glob is held. Bash and WebFetch are left alone. Until one of those reports has run, the check does nothing. Re-run limitcast install --alerts to add the hook.

Which startup piece changed

limitcast report --jobs already showed which scheduled jobs spend tokens just starting. 0.18 names the startup piece that changed between the earliest and latest run of a job, when both runs recorded one: SessionStart hooks, MCP instructions, tool names, skills, and the rest of what the session loaded before the first response.

Five idle minutes, then the hold

A large Claude session still gets held once after 60 idle minutes, when the last call was at least 100,000 tokens. 0.18 announces that same session after 5 idle minutes and still sends the message. Set LIMITCAST_CACHE_NOTICE_MINUTES and LIMITCAST_CACHE_MINUTES to change the two thresholds. Codex and Grok do not get this warning. The announcement is shown to you, not added to the model's context.

Also in this release

trim --share and the provider summaries omit configuration names and paths. Sharing cannot be combined with keep-list edits, --apply, or --undo.

Turn it on

npm install -g limitcast
limitcast install --guard --alerts
limitcast report --why

Restart Claude Code, Codex, and Grok after installing. The guide has the thresholds, the 0.18.0 release has the tarball, and the 0.17 post covers the previous release.