agent-skills

Compact the current conversation into a handoff document so a fresh agent can continue the work without losing context. Use when the user says "handoff", "hand off", "write a handoff", or otherwise asks to checkpoint or wrap up the current work so a new session can pick up seamlessly.

Trigger Jenkins CI builds for Composable Kernel pull requests or branches with optional parameter overrides. Use this skill whenever the user wants to trigger a CI build, start a build, kick off CI, rebuild, re-run CI, or launch a Jenkins build for a CK or Composable Kernel PR or branch. Also trigger when the user asks to set CI build parameters, enable/disable specific tests (like FMHA, AITER, FA, pytorch), or wants to check what CI parameters are available for a PR or branch.

Launch and manage a multi-agent team for HIP/GPU kernel development, debugging, and optimization. Configurable roles (Lead, Implementer, Profiler, Researcher) with optional recall-plugin integration and long-running context rotation. Use this skill whenever the user wants to: start a kernel optimization team, launch agents to work on a HIP kernel, create a dev team for GPU kernel work, set up a multi-agent workflow for kernel performance tuning, or resume a previously saved team config. Also trigger when the user says things like "start a team to optimize kernel X", "launch agents for my HIP kernel", "hip-kernel-team load", "resume my kernel team", or "set up a dev team for this GPU kernel".

Analyze CI build failures for Composable Kernel pull requests or branches on the Jenkins dashboard. Use this skill whenever the user asks about CI failures, build errors, Jenkins build status, or wants to investigate why a CK PR or branch build failed. Trigger on phrases like "check CI", "why did the build fail", "CI report for PR", "CI report for branch", "check Jenkins", "what failed in PR-XXXX", "what failed on branch X", "build status", "CI dashboard", or "analyze build failure".

Validate and post inline PR comments from a prior /review output. Spawns a subagent to verify each claim against actual code, drops wrong items, formats confirmed findings as inline comment drafts, and optionally posts them to GitHub. Works with both PR and local branch reviews. Trigger on: "post review", "post comments", "post inline comments", "format review as comments", or "/post-review".

List open pull requests from ROCm/rocm-libraries that are labeled "project: composablekernel" and focused on fused multi-head attention (FMHA) — including kernel code, tests, and CI/infrastructure. Use this skill whenever the user asks about open FMHA PRs, CK attention PRs, composable kernel flash attention pull requests, FMHA test PRs, FMHA CI changes, or wants to see what FMHA-related work is in progress on the rocm-libraries repo. Also trigger when the user asks things like "what PRs touch the attention kernel", "show me open CK FMHA changes", "list MHA pull requests", "what's being worked on for flash attention in CK", "FMHA test PRs", or "FMHA CI changes". Pass `--full-scan` to ignore saved state and re-scan all PRs from scratch.

Bidirectional commit transfer between monorepos and standalone submodule repos. Use when the user wants to export, split, sync, reset, or update commits between a monorepo and a submodule, or when they ask how the monorepo/submodule workflow works. Also trigger on mentions of "monorepo bridge", "subtree bridge", "subtree split", "update the monorepo branch", "sync from monorepo", "pull monorepo changes", "reset bridge", "undo export", porting commits between mono and sub repos, or setting up a bridge.

Report GPU usage on shared servers — shows which processes occupy each GPU, VRAM percentage, GPU utilization, process owner, elapsed time, and Docker container name. Use this skill when the user asks: "who's using the GPU", "check GPU usage", "is the GPU free", "can I profile", "GPU contention", "VRAM usage", "show GPU processes", "gpu-usage", or any question about GPU availability before profiling.

Switch between Perforce (P4) servers and download files from P4 depots. Use this skill whenever the user wants to switch P4 servers, change Perforce connection, or mentions "p4 switch", "p4 server", "switch perforce", "change p4 port". Also trigger when the user asks "which P4 server am I on" or wants to check/view current P4 settings. ALSO trigger when the user provides a Perforce depot path (starting with "//") and wants to download, sync, print, or view a file from P4 — e.g., "download //depot/project/doc/file.pdf" or "get //depot/path/to/file".

Generate comprehensive FMHA kernel selection guides by dynamically analyzing CK codegen source code. Produces filename structure documentation, field references with dispatch semantics, recommended minimal kernel sets, and ready-to-use filter commands for generate.py. Use this skill whenever the user asks about: kernel selection for any FMHA API, minimal kernel sets, codegen filters, what fields are in a kernel filename, how to reduce kernel count, understanding fwd/bwd/splitkv/appendkv/batch_prefill/pagedkv_prefill kernel variants, "which kernels do I need", "how do I filter kernels", "generate a kernel guide", "explain the filename structure", or any question about reducing or selecting FMHA codegen output. Also trigger when the user mentions "kernel selection guide", "minimal set", "filter pattern", or wants to understand the codegen options for a specific FMHA direction/API.

Analyze GPU kernel resource usage — VGPRs, AGPRs, SGPRs, occupancy, scratch memory, LDS, and register spills — by building a CK target with -Rpass-analysis=kernel-resource-usage and parsing the build log. Use this skill whenever the user wants to: check kernel register allocation, find scratch memory or spill usage, analyze VGPR/SGPR/AGPR counts, identify occupancy issues in a CK build target, or run a "resource usage analysis" / "kernel resource report" on a composable kernel target. Also trigger when the user says things like "check register pressure for <target>", "which kernels have spills", "build with resource usage remarks", "kernel resource report", "check occupancy", "how many VGPRs does <target> use", or "ck resource usage".