Proposal: lightweight HTTP/SSE channel adapter for mobile web clients

[luvarpt]

Hi OpenACP maintainers,

first of all, thank you for the project. The architecture looks very close to a workflow I have been looking for: controlling local coding agents from a mobile device while the actual agent process runs on my desktop/workstation with access to the project filesystem.

I would like to ask whether a lightweight HTTP/SSE-based channel adapter or mobile web transport would fit the direction of OpenACP.

Use case

My typical workflow is:

• I discuss a technical problem or specification on mobile.
• The outcome eventually becomes a non-trivial document, runbook, spec, ADR, or code change.
• The real source of truth should be a Git repository on my desktop/workstation.
• I would like to send tasks to an agent running in a specific project root.
• The agent should be able to inspect/edit files, propose diffs, ask for permissions, and optionally commit after explicit approval.
• I want to interact with this mostly from a mobile device.

Slack/Discord/Telegram are useful transports, but for my personal workflow I would strongly prefer a very small self-hosted web UI or HTTP transport exposed only over VPN/Tailscale/WireGuard.

Proposed direction

I am considering contributing a proof of concept for an HTTP/SSE channel adapter or lightweight mobile web client.

The rough idea would be:

• REST endpoints for session lifecycle and user messages.
• SSE for server-to-client streaming.
• Typed events
• Reconnect support using Last-Event-ID.
• Idempotency keys for message submission.
• Designed primarily for single-user, self-hosted usage behind a VPN.

Questions

Before I start implementing anything, I would appreciate some guidance:

• Does this fit the intended direction of OpenACP?
• Should this be implemented as a channel adapter, a web client over the existing API, or something else? (Please give me a direction)
• Are there existing internal APIs/events I should build on instead of creating new abstractions?
• Would a small proof of concept be welcome if it is clearly scoped and does not try to solve every UI problem at once?