dots

Work with stacked GitHub PRs — creating stacks, restacking after changes anywhere in the stack, batch-rebasing 9+ deep stacks in one command, merging bottom-up, and tracing/visualizing an existing stack's topology and CI status. Use when the user mentions stacked PRs, dependent PRs, PR chains, rebasing a stack, restacking, merging a stack, editing a commit deep in a stack, tracing or inspecting a stack, summarizing the state of a stack, or when you detect a PR whose base branch is another PR's head branch. Also triggers for questions about --update-refs, git absorb with stacks, force-pushing an entire stack, or normalizing GitHub statusCheckRollup across CheckRun and StatusContext entries.

Attach the right visual proof of a change to a GitHub PR — a before/after screenshot, a GIF, or a video — rendered inline, hosted by GitHub (no third-party host, no dangling branches). Use when a PR touches UI or visible behavior and a reviewer should see it: "add a screenshot", "before/after", "show the change", "record a demo", "add a video/GIF to the PR", "attach visual evidence". Picks the lightest medium, captures it with agent-browser, and embeds it — native user-attachments upload by default (png/gif/mp4), commit+raw fallback for images and GIFs when there's no browser login.

Discover agent skills from the open skills.sh ecosystem and install them the dots way — vendored into this repo and committed, never imperatively. Use when the user asks "how do I do X", "find a skill for X", "is there a skill for X", "can you do X", wants to add/install a new skill, or wants to extend agent capabilities.

Create and maintain a long-lived tracking doc (progress doc, build doc, worklog, artifact, dossier) so work survives across sessions, context compactions, and machines. Trigger when starting or resuming substantial multi-session work — a large feature, migration, multi-PR/ changelist stack, or deep investigation — or when the user mentions a tracking/progress doc, worklog, scratchpad, or "keep notes as you go," or wants to avoid losing context across compactions. NOT for user-facing docs (READMEs, API guides) or build/CI artifacts.

Monitor a PR through CI and AI-reviewer cycles — triage every comment (most AI feedback is noise), fix only what's real, reply to false positives with evidence, and keep looping until the PR is green and quiet. Use whenever the user says "handle reviews", "review loop", "address reviewers", "wait for CI", or after creating a PR with CI checks or AI reviewers like CodeRabbit, Devin, Copilot, Greptile, Bito. Also use when a cloud agent or human is pushing concurrently to the same PR, when comments keep bouncing across many rounds, or when an AI reviewer is generating false positives.

Use the AXLE (Axiom Lean Engine) API to verify, check, analyze, and transform Lean 4 theorem proofs via the axiom-axle Python SDK, CLI, or HTTP API. Trigger this skill whenever the user works with Lean 4 formal proofs, asks to verify/check theorems, mentions AXLE/axle, wants to manipulate Lean code (simplify, repair, extract, merge, rename, disprove), or needs to validate candidate proofs against formal statements. Also trigger when the user mentions sorry, proof verification, Mathlib, or formal mathematics tooling.

Submit GPU jobs to UVA's Rivanna/Afton HPC cluster using the rv CLI. Use this skill whenever the user mentions rv, Rivanna, Afton, HPC GPU jobs, SLURM on Rivanna, training on the cluster, or running Python/ML workloads on remote GPUs. Also trigger when users discuss multi-GPU training with torchrun on HPC, distributed training setup for Rivanna, debugging failed GPU jobs on the cluster, checkpoint-restart for long training runs, or port forwarding from compute nodes. Even if the user just says "submit this to the cluster" or "run this on GPUs", use this skill.

Use when building Python 3.11+ applications requiring type safety, async programming, or robust error handling. Generates type-annotated Python code, configures mypy in strict mode, writes pytest test suites with fixtures and mocking, and validates code with black and ruff. Invoke for type hints, async/await patterns, dataclasses, dependency injection, logging configuration, and structured error handling.

Monitor, diagnose, and compare Weights & Biases training runs. Pulls metrics via the wandb Python API, generates matplotlib dashboards, runs anomaly detection, and provides actionable health checks. Use when the user asks to check training progress, debug loss curves, compare experiments, or analyze wandb runs.