Agent Guardrails

This file is written for AI agents using a concrete stack pack. Follow it before making stack recommendations, editing app code, or claiming that an app is ready.

Source Order

Use sources in this order:

1. docs-index.md official docs, release notes, policy docs, and official samples.
2. Local-docs indexed official docs when available.
3. Direct local repo reads for exact source, examples, package files, and command definitions.
4. Maintainer issues, changelogs, and discussions for current breakage.
5. Community examples only after cross-checking with official sources.

Do not use model memory for setup commands, signing, release, platform permissions, or policy claims.

Default Allowed Actions

These actions are normally safe during stack-pack research:

• Read pack files and source notes.
• Search local docs and existing cloned repos.
• Read official docs through approved fetch tools.
• Run non-mutating inspections such as listing files or reading config.
• Record unknowns instead of guessing.

Actions That Need Approval

Ask before you do any of these:

• Clone, pull, install, build, or test third-party repositories.
• Run setup scripts, package-manager scripts, demos, MCP servers, hooks, or generated code from third-party sources.
• Modify _repos/ mirrors.
• Create public endpoints, publish packages, submit stores, sign releases, or mutate cloud accounts.
• Add a live .agents/skills/<stack>/SKILL.md from a learn pack.
• Route global proxy settings or secrets through a tool.

Command Discipline

• Use exact commands from commands.md.
• If a command is missing, stop and fill the command entry from official docs before proceeding.
• Capture proof from quality-gates.md, not just terminal success.
• Mark device, package, signing, and release gates separately.
• Never claim release readiness from build success alone.

Release Claim Discipline

Use precise language:

Stop And Ask

Stop when the user’s goal depends on a gate that is missing, stale, or blocked by credentials. Ask one short question with the needed decision or approval.

Examples:

• The pack has build evidence but no signing gate. Should I stop at package proof, or do you want me to investigate signing?
• The official docs require an account login for release validation. Should I treat release as manual for this pack?
• This starter runs an unreviewed setup script. Can I inspect it only, or do you approve running it in a sandbox?

Red Flags

• The agent says should work without command output.
• The agent treats a tutorial as official docs.
• The agent skips device or browser proof for a platform-specific claim.
• The agent runs a template script before reading it.
• The agent creates a live skill before the pack has repeated successful use.