v0.7.5.

The deployment-workflows release: services can be grouped into environments (deployment lanes), a commit soaked in staging can be promoted to production with one call, and failed builds can be diagnosed by an AI provider under your own API key. Also: operators are now told when scheduled database backups silently stop happening.

Added

• Deployment lanes (environments). A workspace can define named lanes — production, staging, development, anything — and services join a lane from the service create/edit form. The services page gains a lane filter and a Lanes manager (create, rename, delete, with live service counts per lane). Deleting a lane detaches its services rather than deleting anything. Role floors are enforced server-side: member to create or rename, admin to delete. SDK: client.environments.
• Staging → production promotion. POST /v1/services/:id/promote redeploys ANOTHER service at this service's exact running commit — the canonical soak-in-staging-then-promote flow. Both services must track the same repository, the caller needs member on both, the source must have a pinned running deployment, and the target's queued-deploy cap applies. The deploy tab shows a Promote this commit card when same-repo siblings exist. SDK: client.deploys.promote.
• AI failure diagnosis (bring your own key). The operator configures one OpenAI-compatible endpoint, model and API key under Settings → AI Diagnosis — the key is sealed with the same versioned AES-256-GCM envelope as other credentials and is never returned by any route. Failed deployments then offer AI diagnosis: the sanitized tail of the build log (16 KB) goes to the provider with the log framed strictly as data. The sanitizer strips ANSI escapes and masks credential-named assignments, Authorization/Bearer headers and user:password@ URLs before anything leaves the host. Diagnosis needs member on the service, works only on failed deployments, and is audited. Endpoint/model are operator-set by design — pointing the provider at a local Ollama/LM Studio is an intended scenario. SDK: client.ai.
• Missed-backup watchdog. The backup scheduler now notices when a running database's newest scheduled backup is more than 48 hours old — the silent failure mode where a scheduler tick dies and nobody looks — and raises a backup.missed alert once per incident until a successful backup clears it.