Agent skills for engineers who won't lower their standards to use AI.

fusion-runner 0.2.12, a new plugin

Two skills for a Windows GitHub Actions runner in VMware Fusion on a Mac. fusion-runner-setup walks the Broadcom download and Windows installation, verifies the Windows login, records each setup checkpoint, keeps a reusable Windows baseline distinct from the working VM, and ships hosted-compatibility smoke tests that compare the VM with GitHub-hosted Windows. fusion-runner-run starts, checks, and gracefully stops the configured VM, verifying GitHub readiness separately from VM power and preserving the working VM's files.

repo-hygiene 0.12.0 — ci-fix reproduces Linux jobs in whatever is already on the machine

A new scripts/detect_runners.py surveys podman, docker, lima, act and the devcontainer CLI and writes ~/.config/ci-fix/runners.toml. The survey is read-only — it starts no VM, pulls no image, installs nothing — and ranks by what the machine already has: installed beats absent, since starting a stopped runner is one command while installing is a download; an image or VM already on disk beats one that must be fetched, so a stopped lima holding a VM outranks a running but empty podman; and only then does the order podman → lima → docker decide, with podman leading as the easiest (rootless, licence-free, docker's CLI) and lima the right answer when a job needs a full VM. act sits above them but is a driver, ready only when a runtime is. A machine with nothing installed is told the easiest thing to get, not the best, and a hand-written pinned value overrides the ranking and survives --refresh. Architecturally this adds no rung: a Linux runner changes *where* rungs 0 and 1 execute, so every floor and cap from 0.11.0 holds, and CI is the lab now means no local path exists at all. ladder_plan.py gains --local-env container, which doubles a step's CI-proxy estimate for its first run in a container and defers to the ledger afterwards. references/local-runners.md carries the podman/docker, act and lima recipes, the toolchain-to-image map and the architecture rule. Every start, pull and install is one AskUserQuestion, and a decline is recorded.

repo-hygiene 0.11.0 — ci-fix isolates only where it pays, with a floor per level

A new scripts/ladder_plan.py computes the entry rung and the remote mode from the duration profile. Locally (isolating costs seconds): a failed step expected over --isolate-local (30s) runs its failing IDs first, a shorter one runs whole; the whole step (rung 1) always runs when expected within a 10-minute cap and is asked about once above it. In CI (isolating costs a marked commit, a dispatch, and a second cycle): a plain push is both rung 2 and rung 3 unless CI is the only place the failure can be seen and the job is expected over --isolate-remote (5m); then the failing workflow is dispatched alone with only the failing tests through a workflow_dispatch filter input, with a one-time offer to add the input (references/filter-input.md, newline-split through env:, never ${{ }} in run:, empty input runs the whole suite). Whole-workflow dispatch without a filter is gone from the fix path (runner minutes, not time). Local durations are recorded by scripts/local_ledger.py in a machine-level ledger so later fixes decide from local numbers. The profile reports runs per workflow and tops up starved workflows from their own listing; two workflow files with one name: are keyed by path and an unattributable run is reported as an ambiguous join. extract_failures.py reads vitest and cargo-nextest logs and recovers a pytest ID with an unmatched [. --update-versions and the pin audit name a replacement tag only after git/refs/tags/<tag> returns it (the setup-uv@v10 lesson).

repo-hygiene 0.10.0 — ci-audit is now ci-fix, and it fixes CI

Every run measures first: the last successful runs are profiled through the REST jobs endpoint into per-job medians, queue wait, slowest step, a slow/fast/unmeasured class against --slow-threshold (default 2m), and a data-derived CI wait bound. Fix mode triages each red job, extracts the failing test IDs from the job log (pytest, jest, cargo, go; shell-safe quoted), reproduces before editing, and verifies up a ladder: the isolated tests locally, the failed step locally, a local sweep of every other fast job, then a push that dispatches only the failing workflow through workflow_dispatch with the fix commit carrying [skip ci] (the isolated tests again when the workflow declares a filter input), then every run on the pushed commit. Done means every job green on the last pushed commit. Then it offers once to add the failed check as a lefthook or pre-commit hook, staged by its measured duration. --audit reports only; --optimize dispatches a read-only sub-agent that returns ranked, evidence-backed speed changes without applying them. Three stdlib scripts (ci_profile.py, extract_failures.py, workflow_inventory.py), 55 tests with real fixtures, an end-to-end run on a scratch PR. docs/skills/ci-audit.md is a redirect stub. This repo's CI workflow now declares workflow_dispatch. (PROJECTS P42, P45 PR-1; PRs #49, #55)

clear-decision-communication plugin (0.1.2)

Decision requests from an agent to a human during a run, as inline ASCII text: a gate that refuses to ask for facts or for anything already authorized; seven diagnostic axes sized into three tiers (Confirm, Compact brief, Full brief) with length signals and caps; representation chosen from the change type out of a 21-form ASCII catalog (F1–F21, including the span chart, layer stack, containment boxes, decision tree, threshold on a scale, and a general tree); a six-question brief rendered decision-first with verified, inferred, assumed, and not-verified evidence told apart; numbered questions and lettered options (Q1.A) with the recommended option always A and a runner-up condition; a silence default that is always the most reversible option; plain text canonical with a derived dialog rendering behind the two-turn gate; decision records keyed by question ID. Self-contained: synthesizes the owner's decision-communication framework with the clarity rules and the two standards (ISO 24495-1, ASD-STE100) behind clear-technical-communication and the text forms of show-me; names no other skill; never produces HTML. Marketplace goes from 13 plugins / 30 skills to 14 / 31. (PRs #50, #51, #52; PROJECTS P43)