Claude Code Custom Skills Loadout

Deliverable AGT-008. Structure per cc02-standards/SYSTEM-STANDARD.md. Every component traces to a design basis (NEED/REQ) and to an acceptance criterion.

1. Purpose and Design Basis

1.1 Scope and Objective

The Claude Code Custom Skills Loadout is a set of read-only custom CLI capabilities and MCP servers designed to enable agents to query the model-based systems engineering schemas and directories without write access to raw files.

1.2 Design Basis

This system maps agent-runner custom tool extensions (MCP servers and skills configs) to read-only database connections and shell wrappers. It blocks arbitrary file modifications while providing programmatic validation capabilities to ensure agents can check system states deterministically.

1.3 Needs

  • NEED-AGT-008-01: Developer/PI (CS) needs specialized CLI tools for agents to inspect the local MBSE state without editing permissions.
    • Source: CS (Developer / Systems Engineer) security tools plan.
    • Context: Agents require access to manifest nodes and schemas to perform validation but should be blocked from raw file modifications.
  • NEED-AGT-008-02: Detail operators (MR/JM) need read-only access to regional package schemas to verify geocoded data against the local lexicon during advance surveys.
    • Source: MR/JM (Detail Operators) advance survey planning.
    • Context: Manual lookups of regional codes or address structures are error-prone; an automated check is required.

1.4 Requirements

  • REQ-AGT-008-01 (Traceability: NEED-AGT-008-01): When invoked by an agent, the Custom Skills Loadout shall provide read-only access to domain manifest schemas [VR-AGT-008-01].
    • Verification Method: Test (T)
  • REQ-AGT-008-02 (Traceability: NEED-AGT-008-01): On command, the Custom Skills Loadout shall output the target schema in standard JSON format within 2 seconds [VR-AGT-008-02].
    • Verification Method: Analysis (A)
  • REQ-AGT-008-03 (Traceability: NEED-AGT-008-02): The custom skill command suite shall block any commands that contain write, delete, or commit arguments [VR-AGT-008-03].
    • Verification Method: Test (T)

2. Architecture As Built

The Claude Code Custom Skills Loadout is a set of read-only capabilities exposed as a Model Context Protocol (MCP) server (Python 3.12, utilizing the FastMCP framework). It hooks directly into the local agent workspace environment and translates model queries into JSON representations, allowing the agent runner to access the systems engineering definitions without touching files directly.

Data flow:

  1. The agent client sends a tool request payload containing the schema type and identifier.
  2. The MCP command wrapper intercepts the tool invocation.
  3. The input analyzer checks the command parameters to confirm no write arguments or shell sequences are present.
  4. The schema reader loads the requested schema definition from the local model repository (such as _model/architecture.toml).
  5. The JSON formatter parses the data and packages it into structured JSON format.
  6. The JSON payload is returned via standard I/O streams back to the agent client.
+-------------------------------------------------------------------------+
| AGT-008 Claude Code Custom Skills Loadout                               |
|                                                                         |
|  [Agent Runner CLI] ---> [Custom Skill Wrapper] ---> [FastMCP Server]   |
|                                                            |            |
+------------------------------------------------------------|------------+
                                                             v
                                                   Local Schema Database

3. Components and Bill of Materials

ComponentSpecVendor / LinkPriceTierStatusTraceability
CLI WrapperCustom Python command line wrapper scriptInternal$0.00-SELECTEDREQ-AGT-008-03
MCP Schema ServerPython fastmcp server moduleOS$0.00-SELECTEDREQ-AGT-008-01
Skills ConfigJSON-formatted definition config file (skills.json)Internal$0.00-SELECTEDREQ-AGT-008-01

4. Build and Deployment

4.1 CLI Wrapper Deployment

  1. Package the custom CLI and MCP servers in the python engine bundle; install project dependencies in the python virtual environment: uv sync.
  2. Link the MCP server execution command to the local agent runner configs.

4.2 Config Installation

  1. Deploy the skills.json configuration file to the agent environment workspace; verify that skills.json is properly populated and copied to the local runtime path.
  2. Run validation queries to ensure schema access operates correctly: uv run mcp-schema-server.

5. Hardening Baseline

  • User Restraints: Restrict command execution permissions strictly to the sandbox agent user account [REQ-AGT-008-01].
  • Input Sanitization: Apply strict regex filters on all command line parser inputs to prevent shell command injection.
  • Data Protection: Enforce read-only flags on all database connection streams utilized by the FastMCP server [REQ-AGT-008-03].
  • Network Isolation: Run the server process on standard input/output streams rather than exposing public network ports.

6. Integration Points

InterfaceTypeDirectionContract
Schema QueryMCPConsumedRead-only schema queries made by the agent runner
Output FormatFile (JSON)ProducedFormatted JSON output payload representing the system state

7. Acceptance Tests

CriterionMethodExpected ResultReference
REQ-AGT-008-01TestQuery a domain schema via MCP tool and verify the return payload matches the architecture.toml schema.VR-AGT-008-01
REQ-AGT-008-02AnalysisBenchmark 100 consecutive schema queries to confirm average response time is under 100 milliseconds.VR-AGT-008-02
REQ-AGT-008-03TestTry running an MCP command with a —write flag and verify it throws a permissions error.VR-AGT-008-03

8. Run and Maintenance Handover

  • Console Tool: Start the server via uv run mcp-schema-server.
  • Update Cadence: Recompile schema maps and configs weekly to capture upstream MBSE updates.
  • Ownership: Systems Engineer (CS).

END OF SYSTEM

Model wiring

Generated from cell frontmatter at publish time.