chameleon-disable

name	chameleon-disable
description	Use when the user explicitly invokes /chameleon-disable to suppress chameleon's advisory injections for the rest of the current session

/chameleon-disable

Disable chameleon's per-edit layer for the current session. The hook stack still executes (and fails open), but once the session is disabled the PreToolUse hook early-returns before doing any work: no <chameleon-context> is injected AND the PreToolUse enforcement denies (secret-detected-in-content, eval-call, import-preference-violation) do NOT fire. Disable is a FULL per-edit opt-out, not an advisory-only mute. If you want to keep advisory guidance but stop only the blocking, use CHAMELEON_ENFORCE=0 instead (advisory ON, blocking OFF) — that is the opposite trade-off from disable/pause (everything OFF).

When to use

• User runs /chameleon-disable explicitly
• User expresses frustration with chameleon's latency or pattern advice (the callout-detector hook surfaces this command on detected frustration)
• User is doing experimental / one-off work where conformance pressure is unwelcome

Opt-out hierarchy

Most-permanent →    .chameleon/.skip (per-repo, all users, committed → team-wide)
                ↓   CHAMELEON_DISABLE=1   (per-user globally; in shell rc)
                ↓   /chameleon-disable    (this session only)
                ↓   /chameleon-pause-15m  (next 15 minutes)
Most-temporary

Use the most-temporary option that solves the immediate need. Revert by:

• /chameleon-disable → starts new Claude Code session
• /chameleon-pause-15m → expires automatically
• CHAMELEON_DISABLE=1 → unset the env var
• .chameleon/.skip → remove the file from the repo

Silencing ONE surface durably (keep the rest of chameleon)

The full opt-outs above are usually the wrong tool when the complaint is a single recurring surface. The turn-end review (correctness/duplication/idiom findings) is advisory-only and delivered next turn, so a session touching no idiom-governed files or triggering no real violation is already silent — but a team that wants one lens off entirely (not just quiet on a clean turn) has these .chameleon/config.json keys (per-repo, committed → team-wide, survive new chats):

• "enforcement": {"idiom_review": false} — turn off the idiom lens of the async turn-end review job. Per-edit guidance, denies, and every other turn-end check (correctness, duplication, deterministic advisories) stay live. This is the answer to "stop idiom findings, but keep chameleon".
• "enforcement": {"stop_backstop": false} — turn off the ENTIRE Stop turn-end pipeline (relint block, the whole review job — correctness/ duplication/idiom lenses — and every deterministic turn-end advisory). Per-edit hooks stay live.
• "enforcement": {"stop_block_cap": 0} — never let Stop BLOCK; the review job and its advisories still run and still deliver findings.

When the user asks to disable chameleon because of turn-end idiom findings specifically, offer idiom_review: false first — it solves the recurring annoyance without giving up the per-edit layer. Editing .chameleon/config.json is a repo file change: make the edit, tell the user it applies from the next turn, and let them commit it when they want it team-wide.

Prerequisites

disable_session requires a trust grant. If the repo has no .trust record, the tool returns status: failed with a message to run /chameleon-trust first.

The flow

1. Confirm chameleon is currently active in this session.
2. Call chameleon-mcp::chameleon_lifecycle(action="disable_session", params={"repo": <repo_root>, "session_id": <current session_id>}).
   • If the tool returns session_unknown_to_chameleon: true, it means this session has never invoked another chameleon tool. Retry with "force": true in params if the user explicitly asked for disable.
3. The PreToolUse hook checks for the resulting .session_disabled.<sha256(session_id)[:16]> marker before injecting; if present, skips.
4. Confirm to user: "chameleon disabled for this session. SessionStart primer will re-enable on next session unless you set CHAMELEON_DISABLE=1 globally or .chameleon/.skip in this repo."

Don't suggest disable for the wrong problem

• Pattern advice is wrong → use /chameleon-teach instead
• Latency is too high → run /chameleon-doctor to check health
• One archetype's canonical is bad → edit .chameleon/canonicals.json directly OR use /chameleon-refresh
• Profile drift is causing churn → /chameleon-refresh
• Turn-end idiom findings feel noisy or wrong → "enforcement": {"idiom_review": false} in .chameleon/config.json (see above)

Disable is the escape hatch for situations where chameleon legitimately isn't useful in the moment, not a tool for fixing other problems.