Roadmap: native attachments from Twilight to Cloud

[chen-ran]

Problem

Uploaded files are stored correctly, but most non-image attachments become workspace paths before the model call. The agent then tries to install tools or write scripts to read them, even when the model API can read the original file.

Twilight starts the problem: FilePart becomes text in OpenAI and Anthropic adapters. Memoh only sends images as native input. Cloud Relay does not yet route file inputs across providers.

Rules

• Keep attachments as first-class message parts.
• Use native provider input when both the model and channel support it.
• Otherwise use a fixed, observable processor. Never ask the agent to install a reader.

Roadmap

flowchart LR
    T["1. Twilight<br/>Define FilePart sources<br/>Encode native provider requests"]
    O["2. Memoh OSS<br/>Preserve attachment refs<br/>Build model input and fallback"]
    C["3. Memoh Cloud<br/>Relay Files API<br/>Route by model + channel capability"]
    R["4. Release<br/>Sync submodule<br/>Migrate and roll out"]
    V["5. Verify<br/>Provider matrix<br/>Follow-up and retry tests"]

    T --> O --> C --> R --> V

Loading

Done when

A user can upload a PDF containing text, a scanned page, a table, and a chart. Every supported model receives it through the best available input path, follow-up questions still work, and the agent never needs to install software to read it.

References: OpenAI, Anthropic, Gemini, Kimi, DeepSeek.