Versioned Deploy Archives

Status: landed (archived 2026-05-18 — v1+v2 hosts live, careervector-web → careervector-ui rename complete).

Updated: 2026-05-14

Decision

CareerVector cutover uses versioned living archives instead of compatibility paths inside the new app.

The archive rule is strict: v1 and v2 get their own deployment/resource surfaces. v3 must not carry old-version routes, fallback behavior, or runtime switches just to keep older versions alive.

Existing Config Seen Locally

Current branch config has a single production surface:

• ui/wrangler.toml
  • worker: careervector-ui (renamed from careervector-web on 2026-05-17)
  • D1 binding CV_DB: careervector
  • R2 binding CV_BUCKET: careervector
  • service binding REALTIME: careervector-realtime
  • production JOBCACHE_SERVICE_URL: https://jobcache.corbet.ch
• realtime/wrangler.toml
  • worker: careervector-realtime
  • durable object class: WorkspaceDoc
  • D1 binding CV_DB: careervector
• mcp/wrangler.toml
  • worker: careervector-mcp
  • D1 binding CV_DB: careervector
  • service bindings to careervector-ui and careervector-realtime
  • WEB_URL: https://careervector.corbet.ch
• wiki/content/runbooks/STAGING.md
  • deployed staging is intentionally retired.

Known D1 names observed during the cutover work:

• careervector
• careervector-old
• careervector-cutover-20260513
• careervector-v2-archive-20260514

Do not infer source/target roles from names alone. Verify each binding before any write or DNS flip.

Archive Inventory Captured On 2026-05-14

The v2 D1 export contained BLOB rows too large for one-shot SQL import into D1. The completed import used normal SQL for every table except workspace_sub_doc, then a temporary authenticated Worker bound to the archive D1 inserted the snapshot BLOBs through prepared-statement parameters. The temporary loader Worker was deleted after count and byte-length verification.

Archive Requirements

Each living archive needs an inventory row before it can be called safe:

• repository or frozen branch/commit;
• deployed worker/app name;
• hostname route;
• D1 database name/id;
• R2 bucket name if files can be downloaded;
• realtime worker/DO surface if the archive needs live editing;
• secrets/provider bindings supplied by platform secret stores;
• smoke command proving it boots;
• rollback or removal command.

Archive DBs are read-mostly. If an archive remains editable temporarily, it must be labeled as such and must not be treated as champion data after v3 wins.

Cutover Order

1. Preserve v1:
   • point v1.careervector.corbet.ch at the old app currently served by old.careervector.corbet.ch;
   • keep its D1/R2 resources intact;
   • smoke at least dashboard and CVL view.
2. Preserve v2:
   • freeze current production before v3 promotion;
   • copy or bind isolated D1/R2 resources as needed;
   • point v2.careervector.corbet.ch at that frozen surface;
   • smoke the same minimal archive paths.
3. Promote v3:
   • deploy the clean main candidate to the champion resources;
   • run migration verification and browser E2E;
   • point careervector.corbet.ch at v3 only after proof is recorded.
4. Stand up quality:
   • quality.careervector.corbet.ch is internal and protected;
   • runners talk outbound to it;
   • provider accounts are optional plugins, never design requirements.

Open Risks

• The current branch history contained and reverted a product-repo Cloudflare helper. Main should receive a clean squash/cherry-pick candidate, not this raw branch history.
• careervector-cvl is a worktree attached to the sibling careervector checkout. Desktop/Claude work must verify Git worktree state before editing.
• Existing infra docs still contain old staging language and account-specific historical notes. Clean publication hygiene is lower priority than cutover, but it remains necessary before a public repo release.