The failure class to design for is this: your MCP server works against local clients, then stalls in a plugin store review with feedback too vague to act on.
Freeze what you submitted. Pin the manifest, tool list with JSON Schemas, auth scopes, and example transcripts to one commit. Then run the same pre-flight before every resubmit:
- validate all tool input/output schemas and examples with strict validators
- diff tools and required fields against your last submitted commit
- check scope minimization, stable names/descriptions, and consistent error shapes
A credible DIY baseline is plain CI: Ajv or Zod plus a JSON diff and saved fixtures. For infrequent releases and a small tool surface, that catches schema drift and accidental breaking renames without new tooling.
The remaining gap is reviewability. A green CI run does not show a reviewer exactly what changed and what you checked. Keep the commit pair and check logs together so a rejection can be answered with specifics instead of guesses. That bundle matters most when feedback is one vague line.
What pre-submit check has actually caught a store rejection for you before you hit submit?
Top comments (0)