ADR-0007: Retire stage.propflowai.co

• Status: Accepted
• Date: 2026-04-17
• Deciders: Fede
• Related: ADR-0004 (Lambda PR preview deploys), Phase 1 send-lock, Phase 3 CI/CD

Context

stage.propflowai.co was a PM2-managed Next.js instance on EC2 35.172.106.96 (port 3001) fronted by a Cloudflare Tunnel. It existed so we could build once, deploy the artifact to stage, run integration + Playwright tests against a live URL, and then promote the same build to prod.

Since Phase 3 landed (PR #154), every GitHub PR now gets an automatic Vercel preview deploy that:

• Runs the same Next.js build as prod.
• Points at the propflow-stage DynamoDB table + propflow-inbound-processor-stage Lambda (same backing infra stage has always used).
• Gets a stable https://propflowai-git-<branch>-propflow.vercel.app URL the CI api-smoke job already targets.

That makes the EC2 stage instance redundant. Keeping it around costs:

• PM2 process on EC2 + start-with-ssm.sh complexity.
• Cloudflare Tunnel (com.cloudflare.propflow-stage launchd unit) + DNS record.
• Two deploy scripts (deploy-stage.sh, deploy-promote.sh) nobody uses now that merging to main auto-deploys prod.
• Cognitive overhead: new agents reach for npm run deploy:promote.

Decision

Retire stage.propflowai.co. Specifically:

1. Drop stage.propflowai.co as a default target from every caller. Replace with a mandatory SMOKE_TARGET_URL env var (a Vercel preview URL) or --target local.
2. Delete scripts/deploy-stage.sh and scripts/deploy-promote.sh.
3. Remove the propflowai-stage PM2 app from ecosystem.config.js.
4. Kill the PM2 process on EC2 (pm2 stop propflowai-stage && pm2 delete propflowai-stage && pm2 save) — see scripts/retire-stage-pm2.sh.
5. Delete the Cloudflare Tunnel route for stage.propflowai.co and the DNS record — see scripts/retire-stage-cloudflare.sh.
6. Remove the Cloudflare launchd unit on Fede's Mac (manual step once the tunnel is gone).

Explicitly kept:

• propflow-stage DynamoDB table — Vercel previews still use it.
• propflow-inbound-processor-stage Lambda + propflow-inbound-stage.fifo SQS queue — previews still use them.
• seed-stage.sh — still the way to backfill stage DynamoDB from prod for local dev + preview data hygiene. (Retired 2026-07-16 by ADR-0097: raw prod→stage copies are forbidden; use npm run seed:stage for fixtures and the scripts/anon-backfill/ pipeline for anonymized bulk data.)
• EC2 instance 35.172.106.96 itself — still hosts the prod app. Phase 7 will evaluate retiring the EC2 prod host now that Vercel hosts prod.

Migration

• Smoke/integration callers: SMOKE_TARGET_URL env var required; all defaults removed.
• Playwright: TARGET=stage path removed; SMOKE_TARGET_URL is the only way to point at a hosted target.
• Better Auth trusted origins: stage.propflowai.co removed; widened the Vercel preview regex to match *.vercel.app.
• GitHub Actions: STAGE_BASE_URL → SMOKE_TARGET_URL, --target stage → --target preview.

Consequences

Positive:

• One less hosted environment to keep healthy.
• Per-PR isolation: previews can't step on each other like a single stage instance does.
• No more "did someone deploy over my stage" debugging.

Negative / trade-offs:

• CI now depends on Vercel preview availability. Vercel outages block preview-targeted tests. Mitigation: api-smoke is already continue-on-error: true, and local dev + unit tests remain unaffected.
• Agents that pasted stage.propflowai.co URLs into debugging flows need to retrain on preview URLs. Mitigation: every error message that used to default to stage now points at ADR-0007.

Rollback

If we need stage back:

1. git revert the Phase 6 commits on main.
2. pm2 start ecosystem.config.js --only propflowai-stage on EC2 (after restoring the entry in ecosystem.config.js).
3. Re-add the Cloudflare Tunnel route + DNS record.
4. Restore the Cloudflare launchd unit on Fede's Mac.

The propflow-stage DynamoDB table + stage Lambda are preserved throughout, so rollback is a config-only restore.