Agent Safety Preflight

Overview

Agent Safety Preflight helps Claude inspect a local repository before making AI-agent-assisted edits. It produces a compact Green, Yellow, or Red risk receipt from local files only, with special attention to destructive shell patterns, credential-writing automation, agent authority surfaces, uncommitted changes, and unavailable git state.

The skill pairs with the /agent-preflight command in this plugin. Prefer the bundled lightweight scanner when it is available from a trusted Claude install root, then explain the decision and next safe action before editing.

Prerequisites

• A local repository or workspace path to inspect.
• Python 3 available for the bundled agentpreflightlite.py scanner.
• Git available when the workspace is a git repository.
• Permission to read the target workspace locally.

Do not send source, secrets, tokens, private repository contents, payment credentials, or environment variables to external services while using this skill.

Instructions

1. Confirm the target repository path. Use the current workspace unless the user supplied a path.
2. Use Read only for local documentation or configuration files that explain the workspace; do not copy private source into the response.
3. Run /agent-preflight when the command is installed, or invoke the bundled scanner from a trusted Claude install root.
4. If the scanner is unavailable, apply the same manual decision rules: destructive commands, credential-writing automation, risky agent authority, git state, and high-risk shell or network install chains.
5. Classify the workspace as Green, Yellow, or Red.
6. Report the decision, evidence, and next safe action before making edits.
7. For Yellow results, ask for a checkpoint or narrower scope before broad edits.
8. For Red results, stop and request human review before changing files.

Output

Return a short receipt in Markdown that is safe to paste into the local work log or pull request notes:

## Agent Safety Preflight Receipt

- Decision: Green | Yellow | Red
- Repository: <path>
- Git state: clean | dirty | unavailable
- Risk buckets: <bullets>
- Evidence: <file/path markers>
- Next safe action: proceed | checkpoint | stop

Error Handling

• If the scanner cannot be found from a trusted install root, say that the automated scanner did not run and use the manual rules.
• If git status is unavailable, classify the workspace as Yellow rather than Green.
• If a file cannot be decoded, scan the readable prefix with tolerant UTF-8 decoding and include a note if relevant evidence may be incomplete.
• If a destructive or credential-writing marker appears in an agent-controlled path, classi