Skip to content

main #3482

Closed
Closed

Description

@juliopedro1802-wq

Summary

The built-in Hugging Face connector in ChatGPT still declares retired tool IDs (model_search, dataset_search, …) instead of the current hub_repo_search, so every search call fails: the shim that translated legacy calls was removed in the same commit that removed the legacy tools.

Affected layer — not a conversation-level cache

Reconnecting, reinstalling and starting new conversations do not change the tool surface. That is expected: OpenAI stores a published app's tool metadata as a reviewed snapshot, locked for safety — only a new submitted and approved app version can change it.

"published tool metadata is stored as a reviewed snapshot, so reconnecting or reinstalling won't refresh … For changes to tool descriptions, names, schemas, or the tool list, create a new draft version … resubmit … and publish it after approval." — community.openai.com, 2026-07-23

"Breaking changes to the MCP server contract inside a published plugin aren't currently supported. … If a deployment breaks the published version, roll back the server change rather than waiting for a new version to complete review." — developers.openai.com/plugins/deploy/app-review

So the connector is pinned to a pre-2026-07-26 snapshot of this server's schema. Resolving it needs either a server-side compatibility path here, or a republished app version.

Reproduction

  1. Connect the built-in Hugging Face connector in ChatGPT.
  2. Ask it to list its Hugging Face tools → model_search is declared; hub_repo_search is absent.
  3. Call model_search(limit=3, sort="trendingScore") → Resource not found: Hugging_Face.model_search (client-side) / Tool model_search not found (server-side).
  4. Reconnect, reinstall, new conversation → unchanged.

Server-side verification (direct tools/call against https://huggingface.co/mcp, 2026-09-15)

call server response
model_search Tool model_search not found
dataset_search Tool dataset_search is disabled by server configuration
hub_repo_search works
hf_jobs (implemented, but absent from this connection's tools/list) server validates its parameters and returns valid enum values

The last row is the control: an unadvertised tool is still recognized and callable, so "absent from tools/list" ≠ "not callable" — model_search genuinely no longer exists server-side. 9 endpoint variants probed (incl. ?bouquet=chatgpt): none advertise any of the 12 retired IDs.

Evidence that the connector's schema is the pre-2026-07-26 schema

It declares model_search with exactly query, author, task, library, sort ∈ {trendingScore, downloads, likes, createdAt, lastModified}, limit (default 20) — verbatim the zod schema in packages/mcp/src/model-search.ts at the commit before m`. That is a pinned snapshot of this server, not a tool the client invented.

Timeline

| Date | Event |

Activity

  1. changed the title [-]## Summary[/-] [+]## l[/+] on Oct 12, 2026
  2. changed the title [-]## l[/-] [+]main[/+] on Oct 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions