oh-my-qemu - QEMU Modeling Skills

A collection of AI agent skills for QEMU hardware modeling, source builds, image packaging, boot testing, qtest coverage, debugging, RST documentation, and TCG workflows.

The skills give an agent structured, evidence-driven workflows for turning a hardware target into a runnable QEMU model. They are designed to be used inside a QEMU source tree and keep all agent-created artifacts under:

.oh-my-qemu/<task-slug>/

The runtime adds .oh-my-qemu/ to the checkout's local .git/info/exclude when the QEMU tree is a Git worktree. This keeps QEMU source trees free of .plan/, .humanize/, root-level scratch notes, logs, temporary scripts, and other agent working files.

---

Example prompts

Once installed, start Codex from a QEMU checkout and give the agent a concrete target and workload. For example:

git clone https://github.com/processmission/qemu.git
cd qemu
codex

Model a board or peripheral until a workload runs

/goal Help me model the xxx hardware model in QEMU. Proactively collect relevant datasheets, technical blogs, and driver source code. Follow the oh-my-qemu workflow end to end, including source provenance, register extraction, peripheral or board modeling, qtest/model verification, build, boot/run evidence, and RST documentation where appropriate. Continue until the xxx workload can run under QEMU, such as a bare-metal program, firmware payload, RTOS image, or Linux boot.

Extract a register contract before modeling

Use the oh-my-qemu workflow to research this UART/SPI/PCIe block. Compare the datasheet, Linux driver, U-Boot driver, device tree bindings, and firmware initialization sequence, then produce a register-extraction handoff for QEMU modeling.

Debug a stuck firmware boot

The firmware boot hangs after early MMIO setup. Use QEMU logs, trace events, gdbstub, and instruction-window analysis to identify the blocking device or missing register behavior.

Create QEMU board documentation

Write the QEMU RST documentation for this board model. Include supported devices, direct Linux boot, firmware boot, test commands, and known limitations, and update the relevant toctree.

Replace xxx with the SoC, board, peripheral, ISA feature, or test program you want to model and run.

---

Contents

Directory	Purpose
.agents/skills/	Agent skills, one SKILL.md per flow primitive or workflow
scripts/	Skill validation, Codex registration, portable install, task initialization, and methodology issue helpers
commands/	Slash command definitions such as /qemu-init-task
hooks/	Artifact-policy hooks that keep agent scratch files under .oh-my-qemu/
src/	Oh My Pi plugin runtime glue and shared hook logic
.claude-plugin/	Claude Code plugin and marketplace metadata
skills	Symlink to .agents/skills/ for plugin-compatible discovery

---

Skills

Skill schema

Oh My QEMU uses two skill schemas:

Schema	Naming	Role
Flow primitive	qemu-<capability>	Owns one bounded action, record, or evidence contract. It does not orchestrate other primitives or workflows.
Workflow	qemu-workflow-<name>	Composes primitives to complete a larger QEMU task path.

Flow primitives are the stable building blocks. Workflows choose and order those primitives based on the task input.

Flow primitive list

Skill	Purpose
qemu-agent-feedback	Summarize sanitized oh-my-qemu agent workflow problems and file improvement issues with gh when approved
qemu-boot-run	Build reproducible QEMU run commands, logs, timeout markers, and result classification
qemu-build	Configure, build, reuse, and diagnose QEMU build directories
qemu-debug	Use gdbstub, host debugger, QEMU logs, trace events, replay, and instruction-window analysis
qemu-image-layout	Describe and verify boot media formats, partitions, offsets, writes, and hashes
qemu-image-packaging	Package kernels, firmware, DTBs, rootfs, modules, and boot media images
qemu-kernel-build	Build Linux kernel artifacts for QEMU boot tests
qemu-model-verification	Produce PASS/FAIL/INCONCLUSIVE evidence for model, board, device, and runtime behavior
qemu-plan	Define task scope, acceptance criteria, artifact root, and verification gates
qemu-qtest	Design, register, run, and debug qtest coverage
qemu-register-extraction	Extract register maps, bitfields, side effects, IRQ/DMA behavior, and driver sequences
qemu-rlcr-loop	Iterate implementation/debugging rounds with verification, review, local checkpoints, and final-series draft preparation
qemu-rst-documentation	Write QEMU RST docs, board pages, boot examples, Sphinx index updates, and docs validation
qemu-source-provenance	Record source trees, revisions, configs, toolchains, outputs, and hashes
qemu-uboot-build	Build U-Boot, SPL/TPL, FIT/ITB, and firmware-chain artifacts

Workflow list

Skill	Purpose
qemu-workflow-linux-boot	Boot Linux directly or through firmware based on task inputs and preserve handoff evidence
qemu-workflow-board-modeling	Model boards, SoCs, memory maps, boot paths, firmware handoff, FDT, and IRQ topology
qemu-workflow-peripheral-modeling	Model MMIO, qdev, SysBus, IRQ, timer, DMA, and register-bank devices
qemu-workflow-tcg-frontend-instruction	Add, review, or debug guest ISA decode and TCG translation
qemu-workflow-tcg-backend-adaptation	Adapt TCG host backend ops, constraints, register allocation, emission, atomics, and vectors

---

Installation

This repository follows the open Agent Skills standard. Each skill directory contains a SKILL.md file that compatible agents can load on demand. The same skill content is exposed through portable skills, Codex local registration, Oh My Pi, and Claude Code plugin packaging.

Agent/runtime	Project-level path	User-level path
OpenAI Codex	.agents/skills/	${CODEX_HOME:-$HOME/.codex}/skills
Portable npx skills canonical install	.agents/skills/	~/.agents/skills/
Claude Code plugin	plugin marketplace/cache	plugin marketplace/cache
Oh My Pi plugin	plugin marketplace/cache	plugin marketplace/cache

Portable skills

List available skills:

npx skills add https://github.com/processmission/oh-my-qemu -l

From a local checkout, install globally for agents that support global skill paths:

git clone https://github.com/processmission/oh-my-qemu.git
cd oh-my-qemu
npm run portable:skills:install-global

The install script wraps npx skills add with an explicit list of global-capable agents. It intentionally excludes PromptScript because PromptScript is project-only in current skills CLI releases. The script also refreshes Codex links under ${CODEX_HOME:-$HOME/.codex}/skills from this checkout.

Verify the completed install:

npx skills list -g

Install into the current project instead:

npx skills add https://github.com/processmission/oh-my-qemu --all

OpenAI Codex

For Codex development against this checkout, validate and register symlinks into ${CODEX_HOME:-$HOME/.codex}/skills:

cd /path/to/oh-my-qemu
npm run codex:skills:validate
npm run codex:skills:register:dry-run
npm run codex:skills:register

Use a copy instead of symlinks, or target another directory:

node scripts/register-codex-skills.mjs --copy --target /path/to/codex/skills

Start a new Codex session after registering so the new skill metadata is loaded.

Oh My Pi plugin

The Oh My Pi plugin exposes the same skills plus OMP-specific helpers:

• qemu_init_task tool;
• /qemu-init-task slash command;
• artifact-policy hook that redirects root-level scratch artifacts to .oh-my-qemu/<task-slug>/.

Install from the self-hosted marketplace:

omp plugin marketplace add processmission/oh-my-qemu
omp plugin install oh-my-qemu@processmission

Local development link:

omp plugin link /path/to/oh-my-qemu

Claude Code plugin

The Claude Code plugin exposes the same skills, /qemu-init-task, and the artifact-policy PreToolUse hook.

Install from the self-hosted marketplace:

claude plugin marketplace add processmission/oh-my-qemu
claude plugin install oh-my-qemu@processmission

Install into the current project only with --scope project, or keep it local with --scope local.

Local development against a checkout:

claude plugin marketplace add /path/to/oh-my-qemu
claude plugin install oh-my-qemu@processmission

Verify the plugin and its skills:

claude plugin list
claude plugin details oh-my-qemu@processmission

Installed plugin skills are namespaced as oh-my-qemu:<skill>, for example oh-my-qemu:qemu-plan.

---

Recommended workflow in a QEMU tree

Initialize a task workspace:

/qemu-init-task k230-uart-model

This creates:

.oh-my-qemu/k230-uart-model/
  plan.md
  evidence.md
  commands.md
  source-provenance.md
  image-layout.md
  boot-run.md
  methodology-feedback.md
  register-extraction.md
  source-inventory.md
  conflicts.md
  logs/
  reviews/
  scratch/
  rlcr/

For a complete modeling run, keep the request in the task workspace and compose the workflow skills from there. In a QEMU tree, create .oh-my-qemu/k230-uart-model/task.md:

mkdir -p .oh-my-qemu/k230-uart-model
cat > .oh-my-qemu/k230-uart-model/task.md <<'EOF'
---
slug: k230-uart-model
workstream: peripheral-modeling
---

Model the K230 UART and prove it with qtest.
EOF

Then ask the agent to use the appropriate QEMU workflow skill from that QEMU tree. The workflow creates the task root, records a provenance snapshot, plans the work, runs implementation/verification/review loops, prepares human-owned final-series drafts, and writes .oh-my-qemu/k230-uart-model/rlcr/final-summary.md.

Typical composition:

1. qemu-plan - define goal, scope, acceptance criteria, and evidence.
2. qemu-source-provenance, qemu-image-layout, and qemu-boot-run - record sources, artifacts, media layout, and run commands.
3. qemu-register-extraction, qemu-workflow-peripheral-modeling, or qemu-workflow-board-modeling - build the model from verified hardware facts.
4. qemu-build, qemu-qtest, qemu-debug, and qemu-model-verification - prove the model with build, tests, traces, and boot evidence.
5. qemu-rst-documentation - document supported hardware, boot commands, tests, and known limitations.
6. qemu-rlcr-loop - iterate until the acceptance criteria pass.

Common workflows:

kernel build -> linux boot (direct path) -> verification
uboot build -> image packaging -> linux boot (firmware path) -> debug if needed -> verification
register extraction -> peripheral modeling -> qtest -> model verification
board modeling -> qtest -> linux boot (selected path) -> RST documentation

---

Policy and provenance

These skills follow QEMU provenance constraints:

• agents do not generate code intended for QEMU upstream submission;
• agents do not add DCO/review trailers to repository commits or sign on behalf of humans;
• final-series DCO trailers are drafted only for human review and ownership;
• QEMU DCO is the standard Signed-off-by: Name <email> trailer; the tool-user git identity normally satisfies it unless a distinct second signer is recorded;
• AI-used-for: is only a proposed qemu-devel scope-disclosure trailer for human-enabled final-series drafts; it is not an AI-agent DCO sign-off;
• local research, debugging, verification, and workflow guidance are allowed.

The workflow design is based on PolyArch/humanize: plan with explicit acceptance criteria, iterate in reviewed rounds, and keep evidence attached to the work.

---

Methodology feedback

At final completion, pause, blocked state, or max-iteration exit of an RLCR or composed workflow, the agent can write a one-time methodology feedback record. It summarizes reusable workflow problems or improvements into:

.oh-my-qemu/<task-slug>/methodology-feedback.md

The report must be sanitized before it is shown or filed: no private paths, branch names, commit hashes, proprietary logs, code snippets, project-specific URLs, image paths, or board/customer/product identifiers unless explicitly approved.

If reusable improvements exist, use qemu-agent-feedback to draft the public issue from the sanitized report and file it with the local GitHub CLI when approved. The default target is processmission/oh-my-qemu, overrideable with QEMU_METHODOLOGY_ISSUE_REPO=owner/repo.

Draft an improvement issue from a sanitized report:

npm run methodology:issue -- .oh-my-qemu/<task-slug>

This writes:

.oh-my-qemu/<task-slug>/scratch/methodology-issue-title.txt
.oh-my-qemu/<task-slug>/scratch/methodology-issue.md

---

Practice demo

Practice/demo branch using these ideas for downstream QEMU modeling work:

• RK3588/RK3588S Rockchip machines and shared IP models: https://github.com/processmission/qemu/tree/devel

---

Update

Portable skills:

npx skills update -g

Oh My Pi plugin:

omp plugin upgrade oh-my-qemu@processmission

Claude Code plugin:

claude plugin install oh-my-qemu@processmission

Re-add the marketplace to pick up new commits, or manage versions through the interactive plugin menu.

---

License

MIT. Copyright (c) 2026 Process Mission. See LICENSE.