dsh-vision-bench

Firmware development for embedded targets is still scattered across a dozen windows: Keil for building, OpenOCD for flashing, a serial terminal for bus traffic, a separate Modbus polling script, and yet another canvas tool for dashboards. Context switching is constant, and AI agents essentially cannot reach into this hardware chain. dsh-vision-bench folds that entire pipeline into a single DSH conversation workbench with three native workspaces—Debug, HMI (host), and Monitor—so engineers can compile, burn, read/write registers, watch bus frames, and tune alarm thresholds all inside one session, while an AI agent drives hardware operations under explicit user approval.

Core capabilities include: a pure-Node Keil UV4 integration (project scan, target enumeration, AST-based source mapping, and build) with no Python dependency; Modbus RTU/TCP point I/O across FC01/02/03/04/05/06/15/16 with automatic read-back verification after writes, where any agent-initiated register or coil write requires a confirmation card in the UI; OpenOCD flashing gated by SHA-256 hash check and a server-side requestId approval flow; a GridStack-based visualization layer backed by ECharts, uPlot, or a CSS fallback, with 600-sample ring-buffer trends, configurable alarm thresholds, and a shared operation log; and RTU serial packet inspection with frame filtering, hex copy, and JSONL export. The agent interacts through the vision_bench tool (status, build, map, read, write, visualization, manual, etc.), every mutating call carries a configVersion to prevent drift, and failures return explicit error codes rather than silent degradation.

It is aimed at embedded engineers who rely on Keil plus STM32 or NRF52 targets and need day-to-day Modbus communication debugging—especially teams that want to unify serial monitoring, point-table editing, dashboard prototyping, and alarm replay into one surface, and are open to letting an AI assist with build diagnostics and register-level operations while humans retain the final say on every write and flash. No Python install is required; it runs on Windows and macOS, with automated tests passing on macOS and a formal Windows acceptance checklist in place.