auth: refresh_token() always sends `scope`, so refresh fails against Salesforce ("scope parameter not supported")

Aberta
#1,408 3 comentários 0 reações 1 responsável Ver no GitHub

Mantenedores costumam responder em até 3 dias

@gocamille já está trabalhando nisso.

Desde 29/10/2026.

Avaliação

Esta issue ainda não foi avaliada.

Descrição

bug P2 ready for work T-security T-transport

Summary

AuthorizationManager::refresh_token() always adds the stored granted_scopes (plus offline_access when advertised) to the refresh request as scope. Salesforce's token endpoint rejects any refresh request that includes scope:

400 {"error":"invalid_request","error_description":"scope parameter not supported"}

So every refresh fails against a Salesforce-hosted MCP server, and the user has to go through the full authorization flow again each time the access token expires. Clients built on rmcp, including Codex, can't do anything about it, because there's no way to leave scope out of the refresh request.

Where it happens

crates/rmcp/src/transport/auth.rs on main (9e0783d):

let mut refresh_scopes = stored_credentials.granted_scopes.clone();
self.add_offline_access_if_supported(&mut refresh_scopes);
let requested_scopes = refresh_scopes.clone();
for scope in refresh_scopes {
    refresh_request = refresh_request.add_scope(Scope::new(scope));
}

Clearing the stored granted_scopes doesn't help either. Salesforce returns scope in every token response ("scope": "refresh_token mcp_api"), and resolve_granted_scopes stores it again, so the next refresh sends scope once more.

Reproduction

Server: a Salesforce Hosted MCP server (https://api.salesforce.com/platform/mcp/v1/...) with a pre-registered public client (PKCE, no secret). The External Client App has "Require secret for Refresh Token Flow" turned off.

Direct calls to the org's /services/oauth2/token with the same refresh token:

Request Result
grant_type=refresh_token&client_id=…&refresh_token=…&scope=mcp_api refresh_token 400 invalid_request: scope parameter not supported
grant_type=refresh_token&client_id=…&refresh_token=… 200, new access token; response includes "scope":"refresh_token mcp_api"

The same failure through Codex CLI 0.161.0 (rmcp 3.5.1, rev 98a5478), with RUST_LOG=codex_rmcp_client=debug,rmcp=debug:

refresh_transaction{... refresh_reason="expiry"}: rmcp::transport::auth: refresh token present, attempting refresh
WARN ... MCP OAuth provider refresh failed error=OAuth token refresh failed: Server returned error response: invalid_request: scope parameter not supported
WARN ... MCP OAuth access token is expired and refresh failed; reauthorization required

For comparison, the MCP TypeScript SDK's refreshAuthorization sends no scope, only resource, and refreshes silently against the same Salesforce server and client. That also shows Salesforce accepts the RFC 8707 resource parameter that rmcp now adds to refresh requests. Only scope is the problem.

Why this should be configurable

  • RFC 6749 §6 makes scope OPTIONAL on refresh. If it's omitted, the server treats the request as asking for the originally granted scope.
  • Salesforce's OAuth 2.0 Refresh Token Flow documents grant_type, refresh_token, client_id, client_secret, client_assertion, client_assertion_type and format as the request parameters. scope isn't one of them.
  • Some providers do need scope on refresh. Microsoft Entra is one (see mcp-remote#357). So "never send scope" isn't a safe default.

Proposed fix

Either of these would work. They aren't mutually exclusive:

  1. Retry without scope: if the refresh fails with invalid_request (or invalid_scope) and scope was sent, retry once without it. This works without any client changes and doesn't change behavior for providers that accept scope.
  2. Opt-out: a setting on AuthorizationManager / OAuthState, for example omit_scope_on_refresh: bool, that clients can expose in their config.

ToolHive hit the same Salesforce error and fixed it with an opt-out: stacklok/toolhive#6830 (PR #6832).

Related

  • #1330 / #1331: #1330 is about refresh requesting offline_access that was never granted. PR #1331 refreshes with the granted scopes, so it still sends scope and won't fix Salesforce. A fix for this issue should cover the case where the server returns scope in the token response but won't accept it on refresh.
  • openai/codex#58103: the downstream Codex issue.
Linguagem predominante
Rust
Estrelas
4.1k
Forks
694
Merge médio
4d 16h
PRs com merge (30d)
43

Preparar o ambiente

Abrir no Codespaces

Inicia o contêiner de desenvolvimento do projeto no navegador, com a sua própria conta do GitHub.

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de modelcontextprotocol/rust-sdk

Todas as issues de modelcontextprotocol/rust-sdk

Issues semelhantes

Mais issues de Rust

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.