fix(opencode): support proxied MCP OAuth callbacks

[rev-rwasilewski]

Issue for this PR

Closes #31014.

Related to #7377, #18955, and #23787.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This fixes MCP OAuth for environments where the browser reaches OpenCode through a public/proxied URL, but OpenCode still needs to listen on a local callback server. One example is OpenCode running inside a JupyterHub user server or another proxied container environment, where the browser sees a public URL that forwards back to the local callback listener.

Before this change, oauth.redirectUri was used for both the provider-facing OAuth redirect URI and for choosing the local callback listener port/path. That works for simple localhost flows, but breaks proxied setups.

This separates the two concerns:

• redirectUri stays as the OAuth provider redirect URI, and is used by the MCP SDK for authorization and token exchange.
• callbackPort / callbackPath control the local callback listener when using a public/non-local redirect URI.
• Existing localhost and existing non-local redirectUri behavior is preserved unless callbackPath is set.

How did you verify your code works?

• Ran bun test test/mcp/oauth-provider.test.ts test/mcp/oauth-callback.test.ts test/mcp/oauth-auto-connect.test.ts --timeout 30000 from packages/opencode
• Ran bun typecheck from packages/opencode
• Pushed the branch, which ran the workspace pre-push bun turbo typecheck hook successfully
• Manually tested Slack MCP OAuth through a proxied JupyterHub environment

Screenshots / recordings

N/A, CLI/OAuth behavior only.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Related PR Found:

PR #30022: fix(mcp): bind oauth callback to IPv4 loopback

• URL: fix(mcp): bind oauth callback to IPv4 loopback #30022
• Why it might be related: This PR also addresses MCP OAuth callback issues. While PR fix(opencode): support proxied MCP OAuth callbacks #31013 focuses on supporting proxied setups with callbackPort/callbackPath, PR fix(mcp): bind oauth callback to IPv4 loopback #30022 addresses IPv4 loopback binding for callbacks. These could be related efforts addressing different aspects of MCP OAuth callback handling.

[remysaissy]

Worth noting: MCP OAuth servers deployed behind AWS WAF with AWSManagedRulesCommonRuleSet will reject requests containing http://127.0.0.1 URLs in the body or query parameters. The GenericRFI_BODY and GenericRFI_QUERYARGUMENTS rules flag IP-literal URLs as potential Remote File Inclusion attacks and return 403 Forbidden.

This affects both Dynamic Client Registration (redirect_uris in the JSON body) and the authorization redirect (redirect_uri query parameter) when the redirect URI uses the 127.0.0.1 form.

RFC 8252 §7.3 recommends using localhost rather than IP literals for loopback redirect URIs in native OAuth apps. Using http://localhost:... avoids triggering these WAF rules entirely.

Relevant AWS docs: AWS Managed Rules - Common Rule Set

It would be good to make sure the default redirect URI in this refactor uses localhost rather than 127.0.0.1 to avoid this class of issues.