dsh-ios

When debugging iOS apps inside an AI conversation, the agent's biggest problem is that it can neither see nor touch: it gets no screen pixels, so it falls back to guessing coordinates or reading screenshot files—slow and unreliable. dsh-ios puts a real iOS Simulator inside a DeepSeek Harness conversation: tool calls on the left, a live device panel on the right, where you can even tap, drag, rotate, and press Home directly on the video. The agent can boot a device, build and run an Xcode project or Swift package, drive the UI by accessibility identity or OCR text, read unified logs, and inspect processes, backtraces, and leaks—while visual bytes reach the UI only through signed, expiring URLs, never as image blocks or recording files.

Core capabilities are solid: 22 agent tools cover device management, screenshots, interaction, build & run, AXe-backed accessibility tree with tap-by-element, Vision OCR find/tap, and SwiftUI preview hot reload (~2–5 s hot swap without relaunching the app). For list/feed apps, deep snapshots turn each row into indexed entries with heuristically parsed counters (e.g. "57 replies"); row taps are then verified by the counter moving exactly ±1—the only reliable confirmation a list app offers. With a real iPhone over USB, the plugin builds and launches WebDriverAgent, tunnels control and screen ports over loopback, and drives the phone with the same panel and tools, while every real-device tap is gated by identify-before-tap rules so a physical device is never used as a probe.

Who is this for? iOS developers who want an AI agent to build, launch, traverse, and reproduce issues inside a conversation; QA and automation engineers who want semantic tools instead of brittle coordinate scripts; and teams exploring conversational coding workflows that need real "agent operates iOS app" infrastructure. Simulator capabilities require macOS; on non-macOS hosts the tools stay registered but fail with a clear explanation rather than failing silently.