fusion-skills: Instructions file for GitHub Copilot

Skill scripts safety instructions

Apply when authoring, reviewing, or changing skills/**/scripts/**.

Guardrails for script authoring for downstream users. Scripts are blueprint artifacts — consuming agents may adapt execution, so encode safety in intent, constraints, documentation.

Mandatory rules

Guard-dog mode

• Assume script changes are risky until proven safe for downstream execution.
• Block unsafe patterns; do not approve patterns that lack explicit validation or documentation of safety.
• If safety is unclear, ask for clarification or require safer alternatives.
• Keep recommendations minimal, explicit, and security-first.

Authoring-time threat model

• Evaluate scripts as distributed artifacts, not local one-off helpers.
• Assume scripts may run in unknown environments with real credentials present.
• Assume consuming agents may modify command composition, sequencing, or parameters.
• Minimize blast radius if a user runs the script with elevated permissions.
• Require explicit documentation of risks, side effects, and prerequisites.

Safety baseline

• Assume hostile or malformed input by default.
• Validate and sanitize all external inputs before use.
• Never request, read, print, or persist secrets/credentials/tokens.
• Never execute remote code directly (no download-and-run patterns).
• Never embed telemetry, tracking, or opaque outbound calls.
• Avoid hidden network access and undocumented side effects.
• Encode non-negotiable safety constraints in comments/docs so adapted execution preserves intent.

Defensive checks

• Prefer safe defaults and fail-closed behavior.
• Fail fast with actionable, non-sensitive error messages.
• Use explicit guards for file paths, environment inputs, and command arguments.
• Avoid destructive operations unless explicitly requested and confirmed.
• Keep behavior deterministic and auditable.
• Prefer idempotent and bounded operations so adapted execution reduces risk.

Distribution and install safety