Hermod.

Problem.

GenericAgent already ran on my Windows desktop, but checking its work from a phone meant returning to a terminal or fighting a desktop-sized remote interface. I wanted mobile access to the agent already running, with its conversation intact and an explicit access boundary.

Implementation.

1. 01

   Built an authenticated web gateway around upstream GenericAgent, with expiring single-use pairing, rate limits, session revocation, and Origin/CSRF checks. Kept the integration separate from the agent’s existing runtime and configuration.

2. 02

   Bridged phone requests into the running terminal agent’s event loop and streamed progress with SSE. Added guarded conversation resume and new-agent launch flows, plus clear status and Stop controls.

3. 03

   Integrated noVNC for a phone-sized desktop view, with trackpad, keyboard, click, drag, and scroll controls. Viewing is the default; switching modes releases held input so a gesture cannot leave a remote button stuck down.

4. 04

   Reworked verbose agent output into named progress and safe Markdown built with DOM elements. Explicit public-summary boundaries keep native reasoning and raw tool arguments out of the displayed transcript; uncertain output waits instead of guessing.

Result.

Installed the gateway and mobile interface on the existing Windows environment. Verified paired API access, delivered a real desktop framebuffer, and confirmed an idle terminal agent reported phone readiness. Local browser fixtures exercise chat, resume, desktop controls, and transcript rendering.

What I learned.

A remote interface has to respect the state it connects to. Busy agents, saved conversations, held input, and disconnected desktop sessions all need explicit handling. Public progress also needs a reliable boundary in the source stream; formatting alone cannot decide what is private reasoning.