Dvalin Security plugin

One portable plugin payload for Codex and Claude Code. It shares the same autonomous security-gate skill across both ecosystems, while the verifier—not the model-specific client—remains the source of the verdict.

The package uses the current portable root plugin.json format and retains .codex-plugin/plugin.json, .claude-plugin/plugin.json, and .mcp.json as compatibility files. The repository's .agents/plugins/marketplace.json exposes it to Codex desktop as an available local plugin; it is not enabled merely by cloning the repository. The portable Codex package is intentionally skill-only: Codex runs plugin-owned stdio processes from the installed plugin directory, so bundling --workspace . would scan the plugin cache instead of the user's project. .mcp.json remains the Claude compatibility connection.

For the lightest Codex setup, install only the MCP server:

npx dvalincode mcp-install codex
codex mcp list

To add the contextual security skill from this checkout as well:

codex plugin marketplace add /absolute/path/to/dvalincode
codex plugin add dvalin-security@dvalincode-local

Start a new Codex session after MCP installation. The server initialization response teaches Codex the scan → inspect → verify sequence even without the packaged skill. Installing the plugin additionally makes that workflow discoverable from security-sensitive task context.

For development validation:

python3 ~/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py integrations/dvalin-security
claude plugin validate integrations/dvalin-security

The MCP server starts with:

npx -y dvalincode mcp-serve --workspace .

Scanning and offline fix-record verification need no model, API key, or cloud account. dvalin_scan is read-only and persists no state, so clients can allow it by default. A local workflow is created only when dvalin_begin_verification is explicitly called; project commands run only when dvalin_verify_findings is subsequently called.

Client permission behavior

• Codex consumes the MCP readOnlyHint, so dvalin_scan works with approval=never; workflow creation remains a separate state-changing tool.
• Claude Code deliberately requires explicit permission for every MCP tool, including read-only tools. Approve the scan once interactively, add only mcp__plugin_dvalin-security_dvalin__dvalin_scan to the permission allowlist, or pass it through --allowedTools. Do not use bypassPermissions for this.

The plugin cannot silently grant that Claude permission: plugin-shipped settings.json supports only agent/status-line defaults. This preserves Claude's trust boundary while keeping the grant as narrow as one read-only tool. See Anthropic's MCP permission documentation.

Verified from this source tree on macOS on 2026-09-03:

The packaged extension was also installed into a clean VS Code 1.134.0 profile. The real editor result below shows the shared fixture finding and partial coverage status from the same local scanner contract.

imageDvalin finding and coverage inside VS Code