agent-rules

CodeInspectus agent rules

Drop-in rules that make your AI coding agent run CodeInspectus's scan → surface → consent → fix → rescan loop (PRD §12): the tool reports the findings, your agent surfaces them to you and asks before changing code, and fixes only what you approve. CodeInspectus never edits or deletes your source code or repository — the only file it writes is an optional SBOM (to a managed directory by default, or a path you choose), and engine data + scan history stay under ~/.codeinspectus; your agent applies any fix, with your consent. Each scan also returns a read-only git-safety advisory: if you have uncommitted work — or no git repo — your agent will offer, with your approval, to checkpoint first so any fix can be rolled back cleanly (CodeInspectus itself never runs git). All clients use the same MCP server; only the rule file location differs. Remediation is deliberately one finding at a time: a triage state never authorizes reproduction or edits, and test evidence stays separate from CodeInspectus rescan evidence.

1. Register the MCP server (once per machine)

The server command is shared across clients, but the configuration format is client-specific. JSON-based MCP clients use:

{
  "mcpServers": {
    "codeinspectus": { "command": "npx", "args": ["-y", "codeinspectus"] }
  }
}

Client	Where
Claude Code	claude mcp add-json codeinspectus '{"command":"npx","args":["-y","codeinspectus"]}'
Cursor	~/.cursor/mcp.json (or project .cursor/mcp.json)
VS Code	code --add-mcp '{"name":"codeinspectus","command":"npx","args":["-y","codeinspectus"]}'
Codex	codex mcp add codeinspectus -- npx -y codeinspectus, or use Codex settings / TOML below
Windsurf / Cline / Aider	that client's JSON MCP configuration

Codex can also be configured through Settings → MCP servers → Add server (STDIO; command npx; arguments -y, codeinspectus) or with:

[mcp_servers.codeinspectus]
command = "npx"
args = ["-y", "codeinspectus"]
tool_timeout_sec = 600

Codex defaults MCP tool calls to 60 seconds. CodeInspectus allows each security engine up to five minutes, so 600 seconds prevents premature client timeouts on larger repositories. This Codex-only setting does not alter other clients.

First run: npx codeinspectus repair-engines once to fetch + verify the required engine binaries and offline Trivy DB. Later scans expose structured engine_setup state; if it is not ready, the agent must explain the coverage impact and ask before running this networked repair.

2. Install the rule

Client	File
Claude Code	append claude-code.md to your project CLAUDE.md
Cursor	copy cursor.mdc to .cursor/rules/codeinspectus.mdc
Windsurf	append windsurf.md to .windsurfrules
Codex	append codex-AGENTS.md to your AGENTS.md
Cline	add cline.md to your Cline custom instructions

Skill-capable agents can install the self-contained codeinspectus-fix-one/ directory using that client's local skill installation mechanism. The folder includes SKILL.md and agents/openai.yaml; copying it is a client configuration action, not part of a scan.

Two other installable skills are deliberately opt-in:

• codeinspectus-threat-model/ uses project documentation as untrusted explanatory context while preserving the raw finding set unchanged.
• codeinspectus-multi-review/ orchestrates a bounded review whose agent output stays outside the deterministic result set.

All three skill directories ship in the npm artifact. Installing or invoking an optional skill is an agent-client action; CodeInspectus never invokes an LLM during a normal scan.

You can also just say "use codeinspectus to check my code" — it resolves to the same codeinspectus_scan tool.

See the one-finding remediation workflow for approval gates, outcome boundaries, and the separate investigation/regression/test/rescan evidence contract. See also the threat-model workflow and bounded multi-agent review.