Pasted images fail on text-only models — should offload to file + MCP path

Problem

When a user pastes an image (Cmd-V) while using a text-only model (GLM-5.2, DeepSeek, Qwen, etc.), kimi-code sends the image as image_url content directly to the model API. The model rejects it with:

Error: Current model does not support image input.

The paste fails entirely — the image never reaches the conversation, and there is no fallback path.

Expected behavior

When the active model lacks image_in capability, kimi-code should offload the pasted image to a temp file and insert a text reference (file path) into the conversation instead of rejecting it outright. The agent can then process the image through a vision MCP server (e.g. @z_ai/mcp-server, @playwright/mcp) if one is configured.

This matters because many users pair a text-only coding model with a separate vision MCP server. The current behavior makes paste-to-chat unusable for this setup.

Suggested flow

1. User pastes image (Cmd-V)
2. kimi-code checks if the active model supports image_in (from the model capabilities in config.toml)
3. If yes: send as image_url (current behavior)
4. If no: save to /tmp/kilo-shots/pasted-{timestamp}.png, insert the file path as text in the input box (or as a system note), and let the agent decide how to process it (ReadMediaFile, vision MCP tool, etc.)

Current workaround

I run a clipboard watcher daemon that intercepts images, saves them to /tmp/kilo-shots/, and replaces the clipboard with the file path text. This works but is external to kimi-code — the fix should be native.

Environment

• kimi-code v0.18.0
• macOS (but the issue is cross-platform)
• Model: zai-coding-plan/glm-5.2 (text-only, capabilities: ["thinking", "tool_use"] — no image_in)
• Vision MCP: @z_ai/mcp-server (provides analyze_image, extract_text_from_screenshot, etc.)

Config

[models."zai-coding-plan/glm-5.2"]
provider = "zai-coding-plan"
model = "glm-5.2"
max_context_size = 1000000
max_output_size = 131072
capabilities = [ "thinking", "tool_use" ]  # no image_in
display_name = "GLM-5.2"

Start with the Cmd-V paste flow and the model capability settings shown in config.toml, especially the absence of image_in. Trace how image_url content is created for text-only models and how the file path should reach the conversation or a vision MCP tool. Done means image-capable models retain the current behavior, while text-only models save the paste and provide its path without rejecting the conversation.