Superseded runs
SUPERSEDED
A newer commit on the same branch or PR is already running. The older run is dead weight, so Watchdog stops it.
Watchdog cancels superseded and stale GitHub Actions runs, flags runaways before they eat your budget, and explains every failed build in plain English. Safe by default, with a kill switch always in reach.
acme/worker · test.yml · #990
41 min elapsed, historical p95 is 12 min
acme/web · e2e.yml · #2207
PR #318 merged 40s ago, 3 jobs still running
acme/payments · ci.yml · #4812
Newer SHA a91f3c2 already running on this PR
Every push starts a fresh run while the old ones keep billing, minute by minute.
The PR is in. Its 20-minute e2e suite doesn't know and doesn't stop.
A hung job sits until the 6-hour timeout. Nobody notices until the invoice.
Watchdog runs beside your CI, not inside it. Nothing changes in your workflows until you decide it should.
Pick the repos. No workflow edits, no YAML in every repo. Webhooks start flowing in seconds.
For 48 hours Watchdog only logs what it would cancel, with the minutes and dollars attached.
Flip repos live one at a time. First cancels can need approval, and the kill switch is always one click away.
Rules are pure, tested functions. Only two ever cancel on their own. Everything else is a signal for a human.
SUPERSEDED
A newer commit on the same branch or PR is already running. The older run is dead weight, so Watchdog stops it.
STALE_ON_CLOSE
The PR was merged or closed, but its CI keeps burning minutes. Watchdog cancels it the moment GitHub says so.
RUNAWAY_DURATION
A job runs far past its historical p95. You get an alert; it only cancels past a hard timeout you set.
NO_RELEVANT_CHANGES
Path filters say this workflow should not have run for the files that changed. Flagged for a human.
MISSING_CONCURRENCY
Workflows with no concurrency block are listed in the digest, and Watchdog can open the PR that adds one.
INVESTIGATION
Failed jobs get a structured diagnosis from an open-weight LLM, built from redacted, bounded log evidence.
Watchdog pulls a bounded, redacted slice of the failed job and asks an open-weight model for a structured diagnosis: what broke, the evidence, and the next safe step. It posts to Slack and the dashboard. It never runs a command.
Dependency install failed: private registry rejected the token.
npm ERR! 401 Unauthorized — GET https://npm.acme.dev/@acme%2fui
Rotate NPM_TOKEN in repo secrets, then re-run the failed job.
Illustrative example. Secrets redacted before the model sees anything.
Drag the sliders to your numbers. Duplicate pushes and merged-PR runs add up faster than most teams expect.
Linux ubuntu-latest at $0.006/min. Windows and macOS runners cost more, so savings scale up. Estimates, not invoice math. Watchdog shows your real numbers after 48 hours of dry-run.
A tool that cancels CI has to earn trust. Every guardrail is on by default.
Every repo is observed for 48 hours before a single real cancel.
main, master and release/* are never cancelled unless you opt in.
Freeze every auto-cancel from the dashboard without a redeploy.
Approve the first live cancels from Slack or the dashboard.
A PR comment says why it stopped and how to re-run.
The investigator sees redacted, bounded evidence only.
One plan per GitHub App installation. Every plan starts in dry-run. Control pays for itself once it stops $39 of wasted CI a month.
$0 free forever
Dry-run only. Up to five repos. Alerts and the audit log, no live cancel.
$39 / month
Go-live wizard, first-N approvals, Slack buttons, concurrency fixer PRs.
$129 / month
Everything in Control plus invoice-backed usage and spend-spike alerts.
Not unless you explicitly allow it. Protected branches (main, master, release/*) are excluded by default, on every plan.
Only if every workflow in every repo has it, and most don't. Watchdog covers the gaps across the org, handles merged-PR waste concurrency can't, and opens PRs to add the block where it's missing.
A bounded, redacted excerpt of the failed job. Secrets and tokens are scrubbed first, raw logs are never stored, and findings are advice only. The AI never cancels or runs anything.
They're estimates from elapsed time × jobs × runner rate. On the Org plan, Watchdog compares them with GitHub's own Actions billing data.
Yes. It's a Node service plus SQLite, open on GitHub, with a Next.js dashboard.
Install Watchdog, watch it for 48 hours, and see exactly how many minutes it would have saved.