> Use this skill when the user is doing hands-on DOCA RDMI (RDMA Initiator) programming — picking doca-rdmi vs doca-rdma for an accelerator-initiated one-sided RDMA flow, standing up a doca_rdmi_connection or doca_rdmi_poster, attaching a doca_dpa_completion or doca_verbs_cq before doca_ctx_start(), retrieving the DPA-side handle for a DPA kernel, auditing whether a doca_rdmi_* symbol is EXPERIMENTAL on this DOCA, or debugging DOCA_ERROR_* returns from RDMI calls. Trigger even when the user does not say "DOCA RDMI" or "initiator" — implicit phrasings include "my DPA kernel needs to post RDMA writes to a remote responder", "DPA kernel sees no completions", "function not found on doca_rdmi_* at link time", "DOCA_ERROR_BAD_STATE from completion attach", or "the DPA posted but the work request never arrived". Refuse and route elsewhere for two-sided or host-CPU RDMA, the DPA programming model, GPU-side RDMA initiation, or general RDMA/IB/RoCE concepts — those belong to other skills.
npx skills add https://github.com/NVIDIA/skills --skill doca-rdmi
Where to start: This skill assumes DOCA is already installed and
the user is doing hands-on RDMI work on a host or BlueField with
the DOCA package set that ships the doca-rdmi library. Open
TASKS.md if the user wants to *do* something
(install / configure / build / modify / run / test / debug / use);
open CAPABILITIES.md when the question is
*what can RDMI express on this version* — the object model, the
DPA-side handle types, the relationship to doca-rdma, the
EXPERIMENTAL-tag policy, and the safety overlay.
**End-to-end "walk me through doca-rdmi" questions are answerable
entirely from this skill.** Go straight to
TASKS.md ## end-to-end (quickref),
which carries the self-contained install-check → device/cap
discovery → sample → pkg-config build → run → debug walkthrough
with the exact commands. You do not need to open doca-setup or
doca-programming-guide to answer an RDMI build/run/debug question.
Route to doca-setup when the required
DOCA prerequisites are absent, partial, or version-mismatched.
The CLASSES of RDMI questions this skill is built to answer, each
with one worked example. The agent should treat the *class* as the
load-bearing piece — the worked example is a single instance.
doca-rdmi or doca-rdma for this case?" —worked example: *"I have a DPA kernel that needs to fire 1 MB
RDMA writes at a remote responder; which library?"*. Answered by
the *initiator-side vs general-purpose* selection rule in
CAPABILITIES.md ## Capabilities and modes
surface-selection table + the routing back to
doca-rdma when the use case is
two-sided or host-CPU initiated.
worked example: *"create a doca_rdmi_connection, attach a DPA
completion context, hand the DPA-side handle to my kernel"*.
Answered by the connection-object lifecycle in
CAPABILITIES.md ## Capabilities and modes
+ the configure walk in
TASKS.md ## configure.
— worked example: *"my application receives work requests AND
posts RDMA writes; do I need a doca_rdmi_connection plus a
doca_rdmi_poster, or one of them?"*. Answered by the
two-object model in
CAPABILITIES.md ## Capabilities and modes
+ the modify-from-sample slot table in
TASKS.md ## modify.
example: *"hook the connection to a doca_dpa_completion so my
kernel polls completions directly"*. Answered by the DPA-side
completion-attach pattern in
CAPABILITIES.md ## Capabilities and modes
+ the run-side wiring in
TASKS.md ## run, cross-linked into
doca-dpa for the DPA programming
surface itself.
ship?"** — worked example: *"is doca_rdmi_poster_post GA on
my installed DOCA, or still EXPERIMENTAL?"*. Answered by the
EXPERIMENTAL-tag policy in
CAPABILITIES.md ## Version compatibility
+ the version-discovery rule
(pkg-config --modversion doca-rdmi) pinned in
TASKS.md ## configure.
DOCA_ERROR_* from a doca_rdmi_* callmean?"** — worked example: *"DOCA_ERROR_BAD_STATE from
doca_rdmi_connection_dpa_completion_attach"*. Answered by the
RDMI overlay on the cross-library taxonomy in
CAPABILITIES.md ## Error taxonomy
+ the layered ladder in TASKS.md ## debug
that escalates to doca-debug.
This skill serves **external developers building DPA-resident DOCA
applications that need to *initiate* one-sided RDMA operations
against a remote responder** — i.e., users whose accelerator-side
code wants to post sends, writes, or reads directly from the
accelerator without round-tripping through the host CPU. The
canonical caller is a DPA kernel that has been compiled with
doca-dpacc-compiler and runs on the BlueField DPA datapath; a
GPU-side caller that drives the DPU's RDMA queues is the
sister case routed to doca-gpi. This
skill is *not* for NVIDIA developers contributing to DOCA RDMI
itself, and it is not the right surface for general host-CPU
two-sided RDMA — that belongs to
doca-rdma.
DOCA RDMI ships as a C library with the pkg-config module name
doca-rdmi. The library's *host-side* surface
(doca_rdmi_connection_*, doca_rdmi_poster_*) is C; the *DPA-side*
surface that the accelerator kernel uses is also C, compiled against
the DOCA DPA toolchain documented in
doca-dpa. Other-language consumers (Rust,
Go, Python, …) consume the host-side *.so through FFI; the skill's
contribution in that case is to keep the connection / poster
lifecycle, the EXPERIMENTAL-tag policy, the DPA-side handoff rules,
and the safety overlay language-neutral, and to route the agent to
the public C ABI as the authoritative surface that any wrapper will
eventually call. The DPA-side surface is *not* wrappable in another
language — it is compiled and linked into the DPA binary itself.
Load this skill when the user is doing hands-on DOCA RDMI work
on a host or BlueField with DOCA installed. Concretely:
doca-rdmi and doca-rdma for a newone-sided RDMA workload that runs from the DPA datapath.
doca_rdmi_connection or doca_rdmi_poster,attaching a doca_dpa_completion or a doca_verbs_cq, and
starting the context on the DPA datapath.
doca_rdmi_connection_get_dpa_handle /
doca_rdmi_poster_get_dpa_handle into a DPA kernel that calls
the matching device-side header
(doca_rdmi_dev_connection.h, doca_rdmi_dev_poster.h,
doca_rdmi_dev_cqe.h).
doca_rdmi_connection_recv_ack after the DPA kernel consumed
them.
DOCA version, before declaring an RDMI-using component
production-stable.
DOCA_ERROR_* returned by a doca_rdmi_* call anddeciding whether the cause is a configuration mistake, a
lifecycle ordering bug, an unsupported capability on this device,
or a layer below DOCA.
Do not load this skill for general DOCA orientation, install of
DOCA itself, or two-sided host-CPU RDMA questions. For those, use
doca-public-knowledge-map,
doca-setup, and
doca-rdma respectively.
This is a thin loader. The body keeps only the orientation
needed to pick the right next file. The substantive RDMI-specific
material lives in two companion files:
CAPABILITIES.md — what RDMI can express on this version: thedoca_rdmi_connection and doca_rdmi_poster object model, the
DPA-side completion-attach pattern, the relationship to
doca-verbs (RDMI builds on a doca_verbs_context) and to
doca-dpa (the accelerator-side datapath), the EXPERIMENTAL-tag
rule for version handling, the RDMI overlay on the cross-library
DOCA_ERROR_* taxonomy, the observability surface (completion
events on the PE / DPA-side completions), and the safety policy
that gates posting work from an accelerator kernel into a remote
responder's memory.
TASKS.md — step-by-step workflows for the eight in-scope verbs:install, configure, build, modify, run, test,
debug, use. Plus a ## rollback overlay (RDMI-specific
five-step teardown for the verbs / connection / DPA-attach /
MR stack) and the 5-phase universal debug-loop instantiation
appended to ## debug. Plus a Deferred task verbs block
that points out-of-scope questions at the right next skill.
The skill assumes a host or BlueField where DOCA is already
installed at the standard location and the user has the privileges
their public install profile expects. It does not cover installing
DOCA itself — that path goes through
doca-setup.
This skill is agent guidance, not a samples or templates
bundle. To keep the boundary clean, it deliberately does not
contain — and pull requests should not add:
language.** The agent's job is to route the user to verified
reference code on the user's installed DOCA and to prescribe a
minimum-diff modification via the universal modify-a-sample
workflow in
doca-programming-guide,
layered with the RDMI-specific overrides in
TASKS.md ## modify. Because every RDMI
symbol is EXPERIMENTAL at the time of writing, the skill
refuses to author RDMI source code from documentation prose —
the API can change between releases and the resulting code may
not even compile.
meson.build, CMakeLists.txt,Cargo.toml, …) parked inside the skill. The agent constructs
the build manifest *in the user's project directory* against the
user's installed DOCA, where pkg-config --modversion doca-rdmi
is the source of truth.
doca-dpa; RDMI's DPA-side headers are
*consumed by* the DPA programming model documented there. This
skill names the RDMI-specific handoff (the DPA handle type, the
completion-attach call) but does not author DPA kernels.
samples/, bindings/, or reference/ subtree of anykind. A mock or incomplete artifact in this skill's tree, even
one labeled "reference", is misleading: users will read it as
buildable.
SKILL.md first to confirm the user's question is inscope.
EXPERIMENTAL-tag policy, the error taxonomy, observability, and
safety policy, see CAPABILITIES.md.**
modify, run, test, debug, use — see TASKS.md.**
Both companion files cross-link to each other and to
doca-public-knowledge-map
whenever the right answer is "look it up in the public docs or the
installed package layout" rather than "RDMI-specific guidance".
doca-rdma — the higher-level RDMAlibrary covering two-sided and host-CPU-initiated RDMA. RDMI is
the focused initiator-side surface; doca-rdma is the right
answer for the majority of RDMA use cases. The selection table
in CAPABILITIES.md ## Capabilities and modes
is the load-bearing decision aid.
doca-dpa — the DOCA DPA programmingsurface. RDMI returns DPA-side handles
(doca_dpa_dev_rdmi_connection_t, doca_rdmi_dev_poster_t)
that the DPA kernel uses through the device-side headers
(doca_rdmi_dev_connection.h, doca_rdmi_dev_poster.h,
doca_rdmi_dev_cqe.h); the DPA toolchain, kernel build, and
execution model are owned by that skill.
doca-gpi — the GPU-side sister ofthis skill. GPI is the channel/queue surface a CUDA kernel
uses to initiate RDMA; RDMI is the DPA-side surface. Both
layer on the same DOCA RDMA / verbs substrate; either may
apply depending on whether the initiator is on the DPA or on
the GPU.
doca-public-knowledge-map — therouting table for every public DOCA documentation source and
the on-disk layout of an installed DOCA package.
doca-setup — env preparation,install verification, and the *I have no install yet* path with
the public NGC DOCA container.
doca-programming-guide —general DOCA programming patterns shared by every library: the
canonical pkg-config + meson build pattern, the universal
modify-a-shipped-sample first-app workflow, the universal
Core-context lifecycle, the cross-library DOCA_ERROR_*
taxonomy. This skill layers RDMI specifics on top.
doca-debug — the cross-cuttingdebug ladder (install / version / build / link / runtime /
program / driver). RDMI-specific debug overlays on top of it.
doca-hardware-safety —the bundle-wide hardware-safety meta-policy. The `## Safety
policy overlay in CAPABILITIES.md` cross-links it.
doca-version — the versiondetection / four-way match rule every per-artifact `##
Version compatibility` anchor builds on. This skill quotes
the RDMI-specific overlay only.
doca-structured-tools-contract —the JSON-schema contracts for the agent-preferred structured
helpers (env probe, capability snapshot, version-matrix
lookup); the ## Command appendix in TASKS.md defers to
them before falling back to the manual chain.
This skill should be used when the user asks to "create a hook", "add a PreToolUse/PostToolUse/Stop hook", "validate tool use", "implement prompt-based hooks", "use ${CLAUDE_PLUGIN_ROOT}", "set up event-driven automation", "block dangerous commands", or mentions hook events (PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, SessionEnd, UserPromptSubmit, PreCompact, Notification). Provides comprehensive guidance for creating and implementing Claude Code plugin hooks with focus on advanced prompt-based hooks API.
This skill should be used when the user asks to "create a hook", "add a PreToolUse/PostToolUse/Stop hook", "validate tool use", "implement prompt-based hooks", "use ${CLAUDE_PLUGIN_ROOT}", "set up event-driven automation", "block dangerous commands", or mentions hook events (PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, SessionEnd, UserPromptSubmit, PreCompact, Notification). Provides comprehensive guidance for creating and implementing Claude Code plugin hooks with focus on advanced prompt-based hooks API.
Build agentic applications with GitHub Copilot SDK. Use when embedding AI agents in apps, creating custom tools, implementing streaming responses, managing sessions, connecting to MCP servers, or creating custom agents. Triggers on Copilot SDK, GitHub SDK, agentic app, embed Copilot, programmable agent, MCP server, custom agent.
Coding Agent Session Search - unified CLI/TUI to index and search local coding agent history from Claude Code, Codex, Gemini, Cursor, Aider, ChatGPT, Pi-Agent, Factory, and more. Purpose-built for AI agent consumption with robot mode.
Destructive Command Guard - High-performance Rust hook for Claude Code that blocks dangerous commands before execution. SIMD-accelerated, modular pack system, whitelist-first architecture. Essential safety layer for agent workflows.
Makepad UI development skills for Rust apps: setup, patterns, shaders, packaging, and troubleshooting.
Secure environment variable management ensuring secrets are never exposed in Claude sessions, terminals, logs, or git commits
Prompt for generating an AGENTS.md file for a repository
Take nvidia/doca-rdmi from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.