We’re following the Import skills from the MCP server guide for our plugin submission, and the skills import silently produces nothing — even though the scanner’s own requests against our server all succeed.
Setup
- Plugin draft: app
asdk_app_6a0dcc1413f88191ba2dd68c73cb841e, versionasdk_app_v_6a706af263f481918abe73478bc6baf5 - MCP server:
https://alpix.alpic.ai/(public, No Auth, domain verified) - The server implements the current SEP-2640 revision:
io.modelcontextprotocol/skillsdeclared undercapabilities.extensions, paginatedskills/list,skills/get, andresources/readfor every listed URI.
What we verified server-side (all pass)
skills/listreturns one skill:{uri: "skill://paint-canvas/SKILL.md", frontmatter: {name, description}, resources: [{uri, digest}]}— nonextCursorskills/getreturns the identical entry as{skill: ...}resources/readon the SKILL.md returns exactly one text content item whoseurimatches the request; SHA-256 ofcontent.textequals the listed digest (sha256:473d4f306e310715445b68036b65c659f5ee569512bedd1097994aecd0348515)- Fetched frontmatter is identical to the catalog entry
- Well under all documented import limits (1 skill, 1 file, ~2 KB)
The bug
When we click Scan Tools, our server logs show the scanner executing the full verification chain, every request answered 200 in ~10 ms:
12:53:14 initialize x2
12:53:14 tools/list
12:53:15 resources/read
12:53:20 skills/list
12:53:20 resources/read
12:53:20 skills/get
12:53:21 resources/read
Yet the corresponding platform API call — POST /v1/dashapi/versions/.../mcp/inspect (HTTP 200, openai-processing-ms: 12783) — returns the tools correctly, policy_info.safety_status: "SCANNED_OK", and skills: null. The draft keeps skills: [], the Skills tab stays empty, and no error is surfaced anywhere. Waiting more than 20 minutes and re-scanning changes nothing.
Per the docs, a failed entry should mean “Scan Tools still returns the server’s tools but does not update the draft’s imported skills” — but nothing identifies any failed check here, and the scanner demonstrably fetched and could verify everything.
Trace IDs for one failing scan (2026-08-04 12:53:21 UTC)
- inspect request:
x-request-id: req_dee1bfced7c546129aad5ab226d2a8cc - link:
link_6a71e0b3ecd88191aef9350d74eb2170
Is the MCP skills import currently gated to specific orgs (in which case, could the docs say so), or is this an importer bug? Happy to provide anything else — the server is public, so the scan is reproducible at will.
1 Like