detach

Stop the managed-mode watcher attached to the current Claude Code session.

1. Locate the compact-manager CLI. The path in the later steps normally reads as an absolute path (script installs write it in). If it instead still says ${CLAUDE_PLUGIN_ROOT} literally — plugin installs may not substitute command markdown — find the CLI file itself: find ~/.claude/plugins -maxdepth 8 -type f -name compact-manager -path "*/bin/*" 2>/dev/null (this handles versioned cache layouts, where a directory merely NAMED compact-manager has no bin/). If several paths match, use the most recently modified (ls -t).
2. Identify this session's id. If your environment exposes one directly (e.g. a CLAUDE_SESSION_ID env var, or a session id in a path you already know, like your transcript or scratchpad path), prefer that — it is authoritative. Otherwise use the overview heuristic in step 3.
3. Heuristic fallback. State files are written by a PostToolUse hook AFTER each tool result, so on your first tool call of a session your own row does not exist yet. Therefore: (a) first run any trivial command (e.g. echo priming) — completing it writes your state file; (b) then run "${CLAUDE_PLUGIN_ROOT}/bin/compact-manager" overview and read the updated ages. Expect YOUR row marked >, seconds old. (The overview prints session ids shortened to 8 characters; stop accepts an unambiguous prefix, so the short id is usable as-is.) Do NOT proceed on the heuristic alone — show the row and ask the user to confirm — if ANY of: the top row is not seconds-old, its age says FUTURE-MTIME, two rows are under ~2 minutes old (a parallel active session), or only a single session row exists and you could not confirm the id from your environment in step 2.
4. Run "${CLAUDE_PLUGIN_ROOT}/bin/compact-manager" stop <session-id> and report the result. If no watcher held this session, say so — that is a normal outcome, not an error.