Manual testing: multimodal image pipeline (Slack inbound, outbound, modality gate, TUI)

[Aaronontheweb]

Context

The multimodal image pipeline has been implemented and has automated test coverage for serialization round-trips, ChatMessageConverter media handling, modality gating, and the attach_file tool. However, several integration paths require manual testing against real infrastructure before we can consider this feature complete.

Branch: claude-wt-multi-modal-models

Manual Test Plan

1. Slack Inbound — File Download + Content Scanning

• Send a message with an image attachment to a Slack channel where Netclaw is active
• Verify the image is downloaded (check session media directory under /tmp/netclaw-sessions/)
• Verify IContentScanner.ScanAsync() is invoked on the downloaded bytes
• Verify the image reaches the LLM as DataContent in the chat history
• Send a non-image file (e.g., PDF, ZIP) — verify it is skipped with a debug log

2. Slack Outbound — File Upload

• Trigger a FileOutput (e.g., via the attach_file tool or an LLM that produces file output)
• Verify the file is uploaded to the correct Slack thread via files.uploadV2
• Verify upload failures are handled gracefully (logged, don't crash the thread actor)

3. Modality Gate — Real Model Behavior

• Configure a text-only model (e.g., qwen3:30b) and send a message with an image attachment
  • Verify the user sees: [Images removed — the current model does not support vision input]
  • Verify the LLM still receives and responds to the text portion
• Send an image-only message (no text) to a text-only model
  • Verify the user sees the explanation message and no LLM call is made
• Configure a vision-capable model and send a message with an image
  • Verify the image passes through without stripping and the model responds to the image content

4. TUI Rendering

• Trigger a FileOutput and verify ChatPage renders the file path in cyan
• Verify HeadlessChannel prints the file path to console/log

5. SignalR Wire Format

• Verify SessionOutputMapper.ToDto correctly maps FileOutput to type "file" with FilePath, FileName, MimeType
• Verify DaemonClient.FromDto correctly reconstructs FileOutput from the DTO

Automated Coverage (already passing)

For reference, these areas are covered by the 20 new automated tests:

• Protobuf round-trip: SerializableChatMessage and SendUserMessage with MediaReferences (3 tests)
• ChatMessageConverter: DataContent write/read, full round-trip, missing file handling, empty content, MIME mapping (7 tests)
• Modality gate: text-only strips images, image-only skips LLM, vision model passes through (3 tests)
• AttachFileTool: valid path, path traversal, dotdot traversal, missing file, empty path, no session dir, display name (7 tests)

[Aaronontheweb]

Completed