Developer shares three hard-learned bugs from building a phone-controlled desktop AI agent

A developer building DeepSeek Phone Harness, a mobile remote-control interface for a desktop AI agent, encountered three critical bugs that caused tasks to silently hang or display empty data. The first issue arose because the relay only listened to an approval channel, missing a separate channel used when the agent asks the user a question, causing tasks to deadlock indefinitely. The second trap involved deeply nested JSON payloads for tool results, which meant tool cards on the phone showed names and file paths but always had empty results until recursive parsing was implemented. A third problem stemmed from missing cache-control headers, causing the phone browser to persistently serve an outdated version of the web app regardless of refreshes. The project, now released as open-source under the MIT license, forwards every agent event — approvals, questions, and tool results — as interactive cards on the phone in real time.