decision-layer

Decision layer

Typing /decision-layer arms it already — a hook fires before this file loads. Nothing more to do. Just confirm the boundary is on for this session and carry on with the work.

If the skill was triggered by plain words rather than the typed command, no arming event fired. Arm it by hand:

bash "${CLAUDE_PLUGIN_ROOT}/hooks/decision-layer-mode.sh" --arm

If the output style is not selected, arming does nothing. The session-start hook already says so on screen when that happens. The fix is /decision-layer setup: it writes "outputStyle": "decision-layer:Plain" into ~/.claude/settings.json, which covers every project and every surface. The terminal's /config picker covers only the project it was run in, and neither the VS Code extension nor the desktop app has a picker at all. The style is read at session start either way, so it lands on the next session. The selection is remembered, and the boundary still starts off in every new session because the hook owns that part.

Turning it off

• /decision-layer off — off for the session.
• --impl-off on the first or last line of an ordinary message — same thing, handled by the hook, so it works mid-run when the user is busy and not reading carefully.
• --impl on the first or last line of a message — off for that one reply only, then back on automatically. Also handled by the hook, so there is nothing to detect.

Only those two lines count, and only a switch standing there as a word of its own. Quoting one mid-message — pasting a transcript, or this file — is not typing one and changes nothing.

A plain request for code — "show me that function", "what does the error say" — also drops the boundary for that reply. That one is a judgement call, not a hook.

Why the arming matters

Arming is what makes the mode survive. Without the flag file, the rules in the output style are inert: they are written conditionally and apply only to a turn the hook has marked. The flag is keyed by session id, so a new session always starts with the boundary off.