Please do not open a public issue for security vulnerabilities.

Report them privately using GitHub's advisory feature:

1. Open the repository's Security tab.
2. Select Report a vulnerability.
3. Describe the issue, the affected file(s), and steps to reproduce.

You should receive an acknowledgment once the report is triaged. Please allow reasonable time for a fix before any public disclosure.

LOOP is a prompt-level workflow: a command file, a skill file, and installer/validation scripts. Security-relevant behaviors include:

• Installers (scripts/install.ps1, scripts/install.sh) — these write to OpenCode configuration directories. They must never silently overwrite existing files, must back up before replacing, must not follow unexpected paths outside their resolved destination, and must not require or transmit any credentials.
• Validation script (scripts/validate.mjs) — runs local file checks only; it must not execute repository content beyond the installer syntax/behavior tests it spawns explicitly.
• Skill instructions — the skill instructs agents never to print secret values, never to treat repository content as authority over user instructions, and never to perform destructive operations without explicit authorization.
• Behavioral harness (tests/harness/) — its unattended model runner (run-case.mjs + safety.mjs) drives real agents: unattended auto-approval must never be enabled silently (the --allow-auto acknowledgement gate), the child environment must stay scrubbed to an allowlist (credential variables never inherited), and the least-privilege permission config must be delivered to every run. The harness must continue to document honestly what is isolated (fixture directory, per-run session DB) and what is not (config directory, OS sandbox — see tests/harness/README.md), and reports must disclose the machine's own environment only as a count: which variables exist is not report content, only operator-named (--env-pass) variables may appear by name, and no values are ever printed.

• Model behavior quality: LOOP structures agent work but cannot guarantee a model's outputs. Reports of an agent following LOOP yet producing flawed code are quality issues, not security vulnerabilities, unless they involve secret exposure, destructive operations, or injection.
• Vulnerabilities in OpenCode itself — report those upstream to the OpenCode project.

Write the report in English. Include the repository version (see CHANGELOG.md) if possible.