3 min read
headroom
Generated source view for the actual executable
engineering/headroomskill. The durable routing article is Agent Operations Skills. Source: skills/engineering/headroom/SKILL.md
Runtime Source
| Field | Value |
|---|---|
| Category | engineering |
| Origin | unknown |
| Slug | headroom |
| Source slug | headroom |
| Family | Agent Operations Skills |
| Source | skills/engineering/headroom/SKILL.md |
Bundled Resources
No bundled resource files.
Description
Kevin's macOS app that shows this Mac's memory levels and history; agents only read it. Use when checking how much memory a large job would leave, or when old instructions mention Headroom notes, requests, hooks or guards, which no longer exist.
Skill Source
---
name: headroom
description: >-
Kevin's macOS app that shows this Mac's memory levels and history; agents only read it. Use when
checking how much memory a large job would leave, or when old instructions mention Headroom
notes, requests, hooks or guards, which no longer exist.
---
# Headroom
Headroom is the person's memory viewer and history tracker: a menu bar item and an app window that show how much memory the Claude and Codex apps and their sessions use, and a record of it over days and months. It is for the person and takes no messages from agents. The one thing it does on its own is restart a bloated desktop Claude Code session once the restart loses nothing; agents need not do anything about that. It also lists leftovers, the processes that outlived the session, shell or script that started them, and its Kill button stops one only when the person clicks it twice.
## What an agent may do
Read, never write:
- `headroom status` prints the level (ok, warning or critical), the agents' totals, free memory, the leftovers' count, sessions and heavy jobs. `--json` prints the same document as `state.json`.
- `headroom can-start --gb N` prints how much room a job of N GB would leave and always exits 0; `headroom wait` returns at once. Nothing stops agents from starting work (Kevin, 2026-10-21: "never stop agents from starting new stuff"): never hold off, queue or wait because of memory.
- `headroom history`, `headroom days` and `headroom episodes` print the record, if the person asks what happened to memory.
- `headroom leftovers` prints the leftovers: browsers, searches, dev servers and scripts whose parent ended, with their pids, age, memory, last activity, folder and command. `--json` prints the same items. They are for the person to kill. Never kill or signal one yourself; when they hold a lot of memory, say so in your reply.
## What no longer exists
Version 2 removed everything agents used to talk to Headroom with. Do not run or configure any of these, even when older notes, memories or resume files mention them:
- `headroom note` and `headroom request` (announcing jobs, asking the person to approve actions);
- `headroom hook`, `headroom mcp` and `headroom install-hook`;
- `headroom sessions`, `headroom session` and `headroom rescue`;
- `headroom guard` and the runaway guard's LaunchAgent;
- the HTTP API on 127.0.0.1:7341.
If a process you did not start is using too much memory, tell the person in your reply; do not stop it.
Timeline
- 2026-11-02 | Generated a browseable source page from the actual executable skill file. Source: skills/engineering/headroom/SKILL.md