Target: x1e80100-microsoft-romulus15. Revised 9 September 2026.
Evidence levels. Every step is tagged: [T] tested on X Elite hardware by named people in omacom/omarchy-iso#129. [T-other] tested, but on a different board — Yoga Slim 7x, HP EliteBook Ultra G1q, ASUS Vivobook S15. [I] inferred for Romulus from SL7 sources. Not tested by anyone on this machine.
No one has booted this stack on a Surface Laptop 7. Romulus is absent from the tested set. Do not present this as a working build.
Tasks can be grouped semantically by using name prefixes separated with :s. For example all testing related tasks may begin with test:. Nested grouping can also be used to further refine groups and simplify pattern matching. For example running mise run test:**:local will matchtest:units:local, test:integration:local and test:e2e:happy:local (See Wildcards for more information).
TIP
Since TOML keys can't contain colons without quoting, use quoted keys in mise.toml:
| # >>> mise:activate >>> managed by mise — do not edit between markers | |
| eval "$(mise activate zsh)" | |
| source <(mise completion zsh) | |
| # <<< mise:activate <<< |
A thorough guide to completely removing Podman from your system, including containers, images, volumes, configuration files, and all associated data.
| [env] | |
| PIP_INDEX_URL = "https://packagefeedproxy.microsoft.io/pypi/simple/" | |
| UV_INDEX_URL = "https://packagefeedproxy.microsoft.io/pypi/simple/" | |
| npm_config_registry = "https://packagefeedproxy.microsoft.io/npm/" | |
| [tasks."gpu-offload:cluster-00-list"] | |
| description = "List local Podman-backed kind clusters and node containers" | |
| run = ''' | |
| KIND_EXPERIMENTAL_PROVIDER=podman kind get clusters | |
| podman ps --all --filter label=io.x-k8s.kind.cluster |
| // cache-drop-notifier | |
| // | |
| // In one line: every turn this tells you how many tokens were reused from the prompt | |
| // cache, and warns you when that number suddenly collapses. | |
| // | |
| // Why it matters: every time you send a message, Copilot re-sends your entire | |
| // conversation plus a large block of setup text (its instructions, every tool | |
| // definition, your enabled skills, and so on). That is tens of thousands of tokens | |
| // before you have typed a word. The model provider caches the front of that payload | |
| // so it does not reprocess it each turn. You keep that discount only while the front |
A pattern for building personal knowledge bases using LLMs.
This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.
Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.
| LOCATION=westus3 | |
| WORKLOAD_PROFILE_TYPE="Consumption-GPU-NC24-A100" | |
| WORKLOAD_PROFILE_NAME="nc24-a100-profile" | |
| """ | |
| The most atomic way to train and inference a GPT in pure, dependency-free Python. | |
| This file is the complete algorithm. | |
| Everything else is just efficiency. | |
| @karpathy | |
| """ | |
| import os # os.path.exists | |
| import math # math.log, math.exp |