wastegate

Dollar-denominated waste forensics for GitHub Actions.

Not "what does CI cost" — which specific waste, in dollars, and the one-line YAML change that stops it. Local, read-only, uses the gh auth you already have. Nothing leaves your machine.

Every CI cost dashboard answers "where do the minutes go?" None of them answer the question your team actually has: which minutes were waste, what caused it, and what's the fix? wastegate replays your repo's Actions run history through counterfactual analyses — what would a concurrency block have cancelled? what did failed runs burn? what did re-runs of flaky jobs cost? — and prints a ledger where every line item is a cause, a dollar figure with its arithmetic shown, and a copy-pasteable fix.

npx wastegate                      # audit the repo you're in
npx wastegate vitejs/vite --days 1 # audit any public repo, right now

Real output (vitejs/vite, one day of history):

wastegate · vitejs/vite · last 1 days · 128 completed runs
Compute spend $9.34 (at hosted-runner rates) · identified waste $8.07 (86%)

$7.45  Spend on failed runs (16)
  16 failed runs × their billable minutes
  ·    $0.73  CI on fix/deep-import-error-message — failed after 6 min
  ·    $0.72  CI on renovate/pnpm-11.x — failed after 6 min
  fix:
    Order jobs cheapest-first and fail fast: run lint/typecheck/unit before slow
    integration/e2e jobs, add `timeout-minutes:` to long jobs...

$0.61  Spend on re-run (flaky) attempts (2)
  2 runs with >1 attempt, charging (attempt−1)/attempt of cost as waste
  ·    $0.33  CI on feat/hmr-add-runtimeErrors-option — re-run 2× before passing
  fix:
    Re-runs are flaky tests with a dollar sign. Quarantine the flakers...

$0.02  Runs a concurrency block would have cancelled (2)
  2 runs that overlapped a newer run on the same workflow+branch
  fix:
    concurrency:
      group: ${{ github.workflow }}-${{ github.ref }}
      cancel-in-progress: true

Every figure is billable-minutes × GitHub's per-minute rate, rounded per job as GitHub bills.
Read-only: wastegate changed nothing. Apply the fixes yourself.

Install

npx wastegate              # nothing to install
# or
npm install -g wastegate

Requires the GitHub CLI (gh auth login) and Node ≥ 18. Zero npm dependencies.

Usage

wastegate                            # current repo, last 30 days
wastegate owner/repo --days 14       # any repo your gh token can read
wastegate --md  > waste-report.md    # paste into an issue/PR for your team
wastegate --json                     # pipe into your own tooling
wastegate --rate-ubuntu 0            # self-hosted runners? price them at $0

| Flag | | | --- | --- | | owner/repo | Repo to audit. Default: the repo you're standing in. | | -d, --days <n> | History window. Default: 30. | | --json / --md | Machine-readable / markdown output. | | --rate-ubuntu/-windows/-macos <usd> | Per-minute rates. Defaults: GitHub's standard-runner prices (0.008 / 0.016 / 0.08). |

The analyses

| Cause | What it computes | | --- | --- | | Superseded runs | Runs that kept burning minutes after a newer push made them irrelevant — i.e., exactly what concurrency.cancel-in-progress would have cancelled, computed counterfactually from your real run timestamps. Static linters can tell you the block is missing; wastegate tells you what its absence cost last month. | | Failed-run spend | Every minute of a failed run produced nothing. Ranked by cost, surfacing the runs that died late after long builds — the classic "e2e fails in minute 22 because lint would have failed in minute 1" pattern. | | Re-run (flake) spend | Runs with attempt > 1 — someone clicked re-run because something flaked. Earlier attempts are charged as waste, conservatively. |

Costs are computed from per-job wall-clock durations × GitHub's per-minute runner rates, rounded up per job exactly as GitHub bills — which also means it works on public repos (where GitHub's own billable field is always $0) and answers "what would this cost on metered minutes."

Honesty rules: every finding prints its arithmetic; estimates say "estimate"; the tool is structurally read-only (it shells out to gh api GETs and nothing else).

Why not ___?

• Hosted CI-cost dashboards (Kleore, CostOps, Mergify…): require sending your data to a third party, and stop at where the money goes. wastegate is local and answers why, with the fix.
• Local usage CLIs (actions-usage, gh-workflow-stats…): minutes and durations, no dollars, no causes, no counterfactuals.
• Static workflow linters (ci-doctor-style): they predict ("you're missing a concurrency block"); wastegate measures ("its absence cost $412 in the last 30 days, here are the runs").

The counterfactual analyses — superseded-run cost, and the planned matrix-redundancy and draft-PR analyses — exist in no other tool, hosted or local, as of mid-2026.

Roadmap

Matrix entries that never caught a unique failure · draft-PR run waste · org-wide rollup · gh extension install. The analyses are pure functions over run records — see CONTRIBUTING.md if you want to add one.

License

MIT