| name | auto-skill-build-mcp-conformance-harness |
| description | Design a repeatable MCP conformance, health-check, and tool-surface regression harness for local or CI use. Use when the user asks for MCP CI, protocol or OAuth conformance, server doctor checks, tool-surface diffing, JUnit/XML reports, or release gates for an MCP server. Requires an MCP endpoint or stdio launch command; skip when the task is only ad-hoc manual inspection where `$mcp-server-lab` is enough.
|
MCP Conformance Harness
Generated by: Codex Supercharge maintenance automation.
Goal: turn MCP server testing into a repeatable harness with clear evidence,
secret handling, and pass/fail gates.
Workflow
- Use
$sandbox-source-intake and $mcp-safety-review before running a fresh
MCP server, unknown CLI, install script, or broad-account tool.
- Identify the target matrix:
- transport: HTTP or stdio
- auth: none, static token, OAuth interactive, OAuth headless, or M2M
- checks: doctor, tool surface, protocol, OAuth, apps UI
- Prefer read-only checks first:
server probe, server doctor, and
server export with --no-telemetry and JSON output.
- Add
protocol, oauth, or apps conformance suites only after auth and
blast radius are understood; emit JUnit/XML or JSON summary artifacts.
- Capture before/after
server export snapshots and sort them before diffing
when release tool-surface drift is the risk.
- Use the command recipes in
references/mcp-conformance-harness.md instead
of expanding long shell blocks into the skill body.
- Store reports as CI artifacts, but keep credentials out of artifacts. Prefer
environment secrets, credentials files with mode
0600, or CI secret stores.
- Summarize failures by phase: transport, auth, protocol, tool schema, apps UI,
or release diff.
Harness Contract
# MCP Harness Plan
Target:
Transport/auth:
Read-only checks:
Conformance suites:
Tool-surface diff:
Secrets:
Artifacts:
Pass/fail gates:
Known skips:
Skip When
- The user only wants to manually explore one server once.
- No MCP endpoint, stdio command, or runnable server target is available.
- The server/tool is untrusted and has not passed source intake or safety review.
- Running the harness would call write/destructive tools without explicit scope.
References
references/mcp-conformance-harness.md- Source inspiration only, code not executed:
sources/mcpjam-inspector
- Support skills:
$mcp-server-lab, $mcp-safety-review, $auto-skill-safety-mcp-secret-guard
Validation
plugins/codex-supercharge/scripts/skill_audit.sh plugins/codex-supercharge/skills
python3 "$HOME/.codex/skills/.system/skill-creator/scripts/quick_validate.py" \
plugins/codex-supercharge/skills/auto-skill-build-mcp-conformance-harness
