Deploy check

Each deploy is checked from when it’s marked; nothing extra runs on your site.

A deploy marker is the line on the board’s charts for a deploy. Any of these starts its check:

• The site’s deploy hook, from Changes on the board: one URL to call from Vercel, Netlify, GitHub, or any CI, no key needed. It answers with the marker’s id.
• The API: POST /sites/{key}/markers with a write key. Put the commit in the text, like a1b2c3d New checkout, and the prompt starts from it.
• Changes on the board, with the kind set to Deploy.

A deploy marked more than a day after it happened isn’t checked: the hours after it are already behind other changes.

Each window is compared with the same hours on each of the 7 days before. PageHum keeps each hour’s counts for 8 days for this; until a site has 3 days of them, it compares with the 7 whole days before instead.

• Worse: the change is also past three times how much that reading usually moves from day to day, and for a rate, how much it moves by chance in a window that size. A site whose errors swing a lot needs a bigger jump.
• Watch: past twice that, and half the size. A new error seen 10 times or more also makes a fine deploy one to watch.
• Looks fine: everything within its usual range.
• Too quiet to tell: no reading had enough traffic in the window, and at least 3 of the 7 days before with half that. A quiet site never gets a false Worse.

A checked page or flow that started failing after the deploy makes it worse whatever the counts say. Errors and failing endpoints that none of the 7 days before had are listed by page, at 3 or more.

• Since the last deploy, near the top of the board for a day after each deploy: each window, what moved with its numbers before and after, and Copy prompt to fix it when it’s worse or one to watch.
• A lamp on the deploy’s marker in every chart: green, amber, or red. Hover the marker for the verdict.
• When it’s worse, once per deploy: a notification, browser push, and your Slack, Discord, Teams, Telegram, or webhook channels (the webhook event is deploy.worse; see payloads).

GET /sites/{key}/deploys/{id} answers for a deploy marker’s id, or latest, with a read key. A pipeline can wait for the window it cares about and fail on regressed. Or ask your assistant: the MCP tool is get_deploy_check.

{
  "ok": true,
  "data": {
    "marker": { "id": "m_8f2k1c", "at": 1791986700000, "text": "a1b2c3d New checkout", "source": "api" },
    "verdict": "worse",
    "regressed": true,
    "settled": false,
    "nextAt": 1791994500000,
    "windows": [
      {
        "key": "1h", "label": "First hour", "ready": true, "verdict": "worse", "pageviews": 412, "baseline": "hours",
        "findings": [
          { "metric": "errors", "label": "Client errors per pageview", "kind": "rate", "before": 0.011, "after": 0.136, "verdict": "worse" }
        ]
      },
      { "key": "3h", "label": "First 3 hours", "ready": false, "verdict": null, "pageviews": 780, "findings": [] },
      { "key": "24h", "label": "First day", "ready": false, "verdict": null, "pageviews": 780, "findings": [] }
    ],
    "newErrors": [{ "path": "/cart", "message": "TypeError: cart.items is undefined", "count": 57 }],
    "newRequests": [],
    "failing": [],
    "prompt": "My website \"Lumen\" got worse after a deploy…"
  }
}

This GitHub Actions workflow marks each deploy, then waits for the first hour’s verdict and fails the run if it’s worse, with the prompt in the run’s summary. Add PAGEHUM_KEY (a key with write access, from your account page) to the repository’s secrets.

name: Deploy
on:
  push:
    branches: [main]
env:
  PAGEHUM_SITE: s_your_site_key
jobs:
  deploy:
    runs-on: ubuntu-latest
    outputs:
      marker: ${{ steps.mark.outputs.id }}
    steps:
      - uses: actions/checkout@v4
      - run: ./scripts/deploy.sh # your deploy
      - id: mark
        name: Mark the deploy on PageHum
        run: |
          id=$(curl -fsS -X POST "https://api.pagehum.com/v1/sites/$PAGEHUM_SITE/markers" \
            -H "Authorization: Bearer ${{ secrets.PAGEHUM_KEY }}" -H "Content-Type: application/json" \
            -d "$(jq -n --arg text "$(git log -1 --pretty='%h %s')" '{text: $text}')" | jq -r .data.id)
          echo "id=$id" >> "$GITHUB_OUTPUT"

  deploy-check:
    needs: deploy
    runs-on: ubuntu-latest
    timeout-minutes: 150
    steps:
      - name: Fail if the first hour after the deploy got worse
        run: |
          url="https://api.pagehum.com/v1/sites/$PAGEHUM_SITE/deploys/${{ needs.deploy.outputs.marker }}"
          while :; do
            check=$(curl -fsS "$url" -H "Authorization: Bearer ${{ secrets.PAGEHUM_KEY }}")
            [ "$(jq '.data.windows[0].ready' <<< "$check")" = "true" ] && break
            sleep 300
          done
          echo "Deploy check: $(jq -r '.data.verdict' <<< "$check")"
          if [ "$(jq '.data.regressed' <<< "$check")" = "true" ]; then
            jq -r '.data.prompt' <<< "$check" >> "$GITHUB_STEP_SUMMARY"
            exit 1
          fi