dotfiles

Run general-review and deep-code-review together as one consolidated review. Use when the user asks for a full code review, a combined general and deep review, or a comprehensive review of code, a diff, branch, or pull request.

Clone/checkout an ai-dynamo/dynamo PR, build it (maturin + editable Python), bring up etcd/NATS, run the end-to-end aggregated server via the repo's examples/backends/sglang/launch/agg.sh, drive load with uvx aiperf, verify the change empirically, then post a scoped evidence-backed GitHub review. Use when asked to test/try/review a Dynamo PR (e.g. "test ai-dynamo/dynamo#10254", "review this dynamo PR").

First-pass workflow for GitHub PR comments and review threads: pull feedback into a simple actionable markdown table before deciding what to fix. Use before github:gh-address-comments when the user asks to review, triage, address, fix, summarize, classify, ledger, or selectively handle PR comments/review feedback.

Check out an SGLang PR in a throwaway worktree, build it in a fresh venv, run it with sglang.launch_server, drive load with uvx aiperf, verify the change empirically, then post a scoped evidence-backed GitHub review. Use when asked to test/try/review an sglang PR (e.g. "test sgl-project/sglang#12345", "review this sglang PR").

Generate a bi-weekly/monthly engineering status update ("T5T") for a date range by mining the ~/memory work-vault and cross-referencing GitHub PR activity via gh. First emit a full evidence-backed mega dump of active work, then a separately labeled recommended T5T for approval. Use when asked to "write my T5T", "status update", "what did I do from X to Y", "biweekly update", "T5T mega dump", or "T5T for approval".

Create or update GitHub pull request descriptions without deleting existing content or adding long narrative slop. Use when the user asks to write, refresh, clean up, standardize, or update a PR body/description for a current branch or GitHub PR.

MANUAL-ONLY. Use only when the user explicitly invokes `mh-draw-the-owl` or asks to use the draw-the-owl workflow. Never auto-invoke it based on feature or diff size. Explore a feature end to end, then decompose oversized work into atomic, incremental, reviewable tasks.

Create a self-contained local HTML walkthrough of a pull request, branch, commit, or diff. Use when the user asks for a visual recap, HTML report, implementation walkthrough, architecture explainer, or an HTML document showing how a code change works. Ground every claim in the actual diff, code, tests, benchmark artifacts, and review evidence. Do not trigger for an ordinary PR review unless the user also wants the visual HTML artifact.

Generate a shareable HTML usage report from local Codex session logs. Use when the user asks for an insights page like Claude Code's `/insights`, wants a summary of how they use Codex over time, wants session/tool/git/subagent/friction metrics, or wants machine-readable per-session metadata derived from the active Codex home.

Run a deep, adversarial code review focused on correctness, hot-path performance, maintainability, abstraction quality, and codebase health. Use when the user asks for deep-code-review, a deep or overall code review, a thermo-nuclear or thermonuclear review, a code-judo review, a strict performance review, or an especially harsh maintainability audit. Do not use for ordinary bug-focused review unless the user asks for this stricter rubric.

Analyze conversation history, find gaps and drift in AGENTS/CLAUDE instructions and skills, propose and apply targeted improvements.

Ask one or more fresh Claude Code processes for parallel, independent, read-only verification, design adjudication, or adversarial review while Codex retains ownership of the task. Use when the user explicitly asks for a Claude second opinion, cross-model verification, parallel independent review, or when a consequential code/design/benchmark decision benefits from an unprimed verifier.

Investigate benchmark telemetry from Tachometer Parquet, Prometheus snapshots, SGLang and Dynamo logs, traces, and cache/storage metrics. Use when diagnosing throughput, latency, routing, KV-cache, HiCache, or storage-tier behavior; validating a benchmark; or explaining why an observed metric changed.

MANUAL-ONLY. Build a Codex fork/branch from source on Apple Silicon macOS and install it as both the CLI (~/.local/bin/codex) and the desktop app (/Applications/Codex.app) by swapping the app's embedded binary and re-signing the bundle. ONLY run when the user explicitly types /local-mac-codex-install-from-patch. NEVER auto-invoke this skill, never select it for a model-initiated task, never run it as a side effect of any other request.

Review any requested artifact for material concerns, likely regressions, ambiguity, unnecessary complexity, and performance overhead. Use whenever the user asks for a review of anything, including a conversation, plan, design, implementation, patch, diff, pull request, document, or scoped code surface.

Structure and draft Linear issue bodies, updates, and comments with strict anti-slop rules. Use when the user wants to write a Linear issue/update/comment or decide the shape of one; do not use for general Linear lookup, status checks, project browsing, or routine reads.

Capture real browser UI screenshots and visual evidence for webpages, local web apps, dashboards, Perfetto traces, and other browser-rendered artifacts. Use when the user asks to use the browser, inspect a visual UI state, verify layout/rendering, produce screenshots, or attach browser/Perfetto snapshots as evidence.

Manage inference server lifecycle through NVIDIA srt-slurm: author benchmark YAMLs, render lifecycle bash with `srtctl apply --bash`, run SGLang/Dynamo/AIPerf jobs, configure first-class telemetry/observability, validate artifacts, and improve the srtctl lifecycle renderer. Use for server lifecycle tasks where srt-slurm is the unified control plane.

Babysit a PR through CI to green. Merge main into the PR branch, push, wait for CI, and ask the user before retrying on failure. Ping the user when CI is green. Use after a PR has been approved for merge.

Build and run NVIDIA Dynamo with SGLang backend from source on a fresh Ubuntu VM — includes all dependency troubleshooting.

Use when the user asks to access, inventory, troubleshoot, or refresh auth for Teleport-managed clusters, NVIDIA DGX Cloud clusters, tsh login, Kubernetes cluster contexts, SLURM clusters behind Teleport, or asks which clusters are available. Reads ~/memory/clusters first, checks Teleport certificate state, and prompts the user for MFA/port-forward refresh when needed.

Bump Dynamo's SGLang backend to a new SGLang version. Sets up a clean dynamo main checkout, target SGLang tag, fresh venv, then walks every launch script and fixes API breakage as it surfaces. Use when the user asks to upgrade/bump SGLang in Dynamo (e.g. "update dynamo for sglang 0.5.X").

Log a milestone, finding, or result to ~/memory/. Invoke explicitly with /memory-log or proactively when a session produces something worth remembering.

Diagnose hung or stalled Python processes using py-spy. Captures live Python and native (C/C++) stack traces from one or more PIDs without restarting the process — works for local processes, processes inside containers, and processes inside SLURM jobs. Use when a Python process has stopped emitting output but is still alive, when distributed workers appear deadlocked, or when you need to attribute a stall to NCCL / mutexes / pure Python work for an upstream bug report.