hermes - 💡(How to fix) Fix DingTalk adapter does not support MEDIA/local file attachments

Summary

The DingTalk gateway adapter currently cannot deliver local files referenced through Hermes MEDIA:<path> messages. This affects PDFs and other document/image attachments generated by tools or scheduled jobs.

Current behavior

When a response contains a local media attachment, the shared gateway/media path eventually calls the platform adapter's document/image send method. For DingTalk, the adapter returns an unsupported result for local files, e.g.:

DingTalk session webhook replies do not support local file attachments. Only markdown/text replies are supported without OpenAPI message send.

In practice this means that a message body may be delivered, but the actual MEDIA:/path/to/file.pdf attachment is not delivered to DingTalk.

Expected behavior

DingTalk should either:

1. support local MEDIA: attachments through the DingTalk OpenAPI media-upload + robot-message flow; or
2. clearly document DingTalk as text/markdown-only and make the failure visible before users rely on MEDIA: delivery for scheduled reports.

Ideally the behavior should be consistent with other Hermes messaging platforms where tool output can include local MEDIA: attachments.

Reproduction outline

1. Configure DingTalk gateway/stream mode.

2. Send or schedule an agent response containing a local file attachment, for example:

   Here is the report.
   
   MEDIA:/tmp/example-report.pdf

3. Observe that DingTalk does not receive the file attachment and the adapter reports that local file attachments are not supported.

Notes from local investigation

• The DingTalk adapter supports markdown/text replies, but local file/image send methods are not implemented for actual attachment delivery.
• A local experiment confirmed that DingTalk OpenAPI can obtain an access token and upload a PDF via the media upload API, so the missing piece appears to be adapter-level support for selecting the correct DingTalk robot message API/message template and routing context.
• A robust implementation likely needs to handle at least:
  • OpenAPI access token acquisition/reuse;
  • media upload for file/image types;
  • sending the uploaded media_id via the appropriate DingTalk robot/group or one-to-one message API;
  • clear handling of robotCode, openConversationId, and private-vs-group conversation context;
  • safe fallback/error messages when the tenant/app does not have the required DingTalk permissions or message templates.

No secrets, chat IDs, tenant identifiers, or customer data are included here.

Environment

• Hermes repository: NousResearch/hermes-agent
• Local commit tested: ec31390a9
• Platform: Linux
• Messaging platform: DingTalk gateway