dsh-debug-mode

When an AI agent diagnoses a bug, the weakest link is evidence. Without breakpoints, runtime probes, or a reproducible trace, the agent defaults to guessing. dsh-debug-mode turns the DeepSeek Harness Web composer into a three-mode workspace (Normal, Plan, Debug) and, in Debug mode, gives the agent actual forensic power: it instruments a frontend project with browser probes that capture line-numbered runtime events and deliver them over CORS to a local listener; it sets breakpoints in Node (CDP) and Python (DAP) backends and runs the full wait, evaluate, next, continue debug loop, finishing with automatic cleanup and a persistent trace.jsonl for post-mortem review.

The plugin also handles the messier, real-world workflows: the first debug_start asks for user confirmation, listing the target process pid and restart instructions, then stops the existing service so the agent can debug in place and restores normal service on debug_finish. The reproductionScope parameter (local, lan, auto) lets the agent decide whether to advertise a loopback or a LAN address, making it practical to reproduce and collect cross-origin probe events on a physical or emulated device over the local network. Every step—from the agent issuing a debug request to breakpoint hits, expression evaluation, root-cause analysis, and scene restoration—is driven programmatically through the debug_control tool, so the agent no longer needs a human to paste stack traces.

Built for engineers who use DeepSeek Harness to develop full-stack web apps (frontend plus Node/Python backends) or Flutter mobile applications. If you want your agent to produce verifiable, line-numbered, timestamped evidence before drawing a conclusion rather than merely mimicking a debug conversation, flip the Debug switch and let it work with real breakpoints.