Safety Gates

Destructive CLI commands display an impact summary and require explicit confirmation before any mutations occur. The confirmation method depends on the command's risk level.

---

Confirmation policies

Policy	Interactive TTY	CI / non-interactive
FLAG	Must re-run with the force flag	Must pass the force flag
TYPED	Must type an exact phrase at the prompt	Must pass --force
COUNTDOWN	5-second countdown — Ctrl-C to abort	Must pass --force
NONE	Proceeds silently (audit entry only)	Proceeds silently

FLAG is the most common policy. The CLI prints a full impact summary then exits 1 with a hint like "Re-run with --force to confirm." Passing the flag on the same invocation is the only way to proceed — interactive terminals do not get a yes/no prompt.

TYPED is used for CRITICAL operations. The CLI prompts you to type a specific phrase (shown in the impact summary). Typos abort. Pass --force to skip in CI.

COUNTDOWN is used for admin overrides where a 5-second window to change your mind is appropriate (e.g. force-transition). Ctrl-C cancels. Pass --force to skip in CI.

---

Risk levels

Level	Color	Typical use
LOW	Blue	Informational / audited-only
MEDIUM	Yellow	Cancellations, deletions of recoverable state
HIGH	Red	Irreversible deletes, credential revocations, kill switches
CRITICAL	Red background	System-wide halts, full local reset

---

Affected commands

Command	Risk	Policy	Flag	Reversible
lf reset	CRITICAL	TYPED (RESET)	--force	No
lf kill-switch platform on --scope system	CRITICAL	TYPED (PLATFORM DOWN)	--force	Yes
lf battle delete	HIGH	FLAG	--confirm	No
lf battle byok-key revoke	HIGH	FLAG	--force	No
lf battle force-transition	HIGH	COUNTDOWN (5s)	--force	Partial
lf communities delete	HIGH	FLAG	--confirm	No
lf kill-switch on	HIGH	FLAG	--confirm	Yes
lf kill-switch platform on (scoped)	HIGH	FLAG	--force	Yes
lf execution cancel	MEDIUM	FLAG	--force	Partial
lf schedule delete	MEDIUM	FLAG	--force	No

---

CI usage

In any CI environment (detected via CI, GITHUB_ACTIONS, GITLAB_CI, BUILDKITE, DRONE, CODEBUILD_BUILD_ID, TF_BUILD, CIRCLECI, TRAVIS, JENKINS_URL), destructive commands block unless the force flag is supplied:

# CI-safe invocations
lf reset --force
lf schedule delete <id> --force
lf execution cancel <run> --force
lf kill-switch platform on --scope battle --target <id> --reason "…" --force

Non-interactive shells (no TTY) are treated the same as CI regardless of environment variables.

---

Audit log

Every safety gate check — whether confirmed or aborted — appends one JSONL line to ~/.lenserfight/audit.log:

{"ts":"2026-05-09T14:23:01.000Z","risk":"HIGH","reversibility":"IRREVERSIBLE","env":"remote","confirmed":true,"description":"Permanently delete battle abc123 (creator only).","resources":[{"type":"battle","name":"abc123","scope":"remote"}]}

Fields: ts, risk, reversibility, env, confirmed, description, resources.

The file is append-only and never rotated automatically. Write failures are silent — audit logging never blocks the operation.

---

Auth auto-recovery

When a command fails authentication (no stored token or a 401 response), the CLI attempts recovery before surfacing an error:

1. Silent token refresh — uses the stored refresh token. Transparent when it works.
2. Browser device-login — auto-opens your browser, displays an approval code, and resumes the original command once approved. Interactive TTY only.
3. CI / non-interactive fallback — prints Set LF_API_KEY or run \lf auth login`` and exits.

In headless environments, set LF_API_KEY (or LENSERFIGHT_API_KEY) to a developer token to avoid interactive auth entirely.

---

See also

• Auth Commands — lf auth login, token management
• Global Flags — --local, --debug
• Kill Switch — per-agent and platform-wide kill switches
• lf reset — full local reset