dsh-plugin-amdgpu-inspect

Letting an agent in DeepSeek Harness actually read AMDGPU binaries used to mean piping giant llvm-objdump dumps into context and hoping for the best. dsh-plugin-amdgpu-inspect replaces that with two structured tool calls: amdgpu_object_inspect lists the kernels in a file and reports code-object metadata plus instruction-category counts, while amdgpu_isa_query walks a single kernel's normalized instruction stream page by page, capping each response at 200 records with a deterministic cursor so the full disassembly never floods the model window. The plugin reports static facts only—no source compilation, no HSA runtime loading, no GPU execution, no performance prediction.

On top of the raw API, a bundled Kernel X-Ray Skill gives an agent a structured playbook: compile before/after artifacts, query this plugin, form a hypothesis, and demand a workload measurement before anyone claims a speedup. Under the hood, a single-artifact cache keyed by absolute path, size, and modification time keeps repeated calls cheap; child processes run without a shell and honor tool-call cancellation; inputs are limited to regular files up to 256 MiB.

It fits engineers who already have an HSACO or ELF in hand and want to inspect which kernels it contains, where a specific opcode lives, or how the instruction stream changed between two builds—without standing up a separate disassembly pipeline. The plugin stays a low-level facts layer; the strategy and conclusions live in the Skill or your own prompt, making it easy to slot into larger agent orchestration.