Skip to content

Data flows

This document describes the intended CloudDev AI single-user deployment. A live Railway deployment has not been tested from this checkout. Railway hosting is in Europe West only when every service and volume is actually provisioned in europe-west4-drams3a.

Components and stored data

Component Data handled Persistent location
Shared workspace Source files, shell history, editor settings, CLI configuration, SSH host identity, Git identity, CloudCLI session index, task status One Railway volume mounted at /workspace
Forgejo Accounts, repositories, issues, pull requests, Git SSH host identity, application configuration Forgejo volume at /var/lib/gitea and PostgreSQL database
PostgreSQL Forgejo relational state and authentication metadata Railway managed PostgreSQL storage
OmniRoute Provider connections, encrypted provider keys, inference keys, dashboard settings, request and usage metadata OmniRoute volume at /app/data; encryption key must also be backed up
Railway Service configuration, environment variables, deployment events, resource metrics, network metadata Railway control plane
AI providers Prompts, relevant project context, tool outputs, and provider request metadata Provider-controlled systems under the selected account and contract

Main flows

  1. A browser connects to the shared workspace over Railway HTTPS. Caddy calls the launchpad's loopback auth verifier to check a signed HttpOnly workspace cookie before forwarding to CloudCLI, the launchpad, or code-server. The login form uses a per-deployment password and sets the cookie Secure on production HTTPS. The workspace TLS boundary ends at Railway's HTTP edge. CloudCLI's Bearer Authorization header passes through unchanged.
  2. Native VS Code Remote-SSH connects through the workspace TCP proxy to sshd on internal port 2222. sshd accepts operator-provisioned Ed25519 public keys only. The workspace login is coder; the private client key stays on the developer's computer. The public host-key fingerprint must be checked at first connection against the value shown in the launchpad.
  3. VS Code Desktop, browser VS Code, and CloudCLI operate on the same /workspace/projects tree and the same persistent HOME. CloudCLI stores session metadata in its SQLite database and scans provider-native session artifacts. The durable task registry stores metadata, not prompts or file contents. A Railway restart terminates processes; it does not terminate provider history files or the SQLite index when the volume remains intact.
  4. The terminal wrappers and CloudCLI send model requests to OmniRoute over Railway private networking. The workspace receives only a scoped inference key. OmniRoute stores provider master credentials and routes prompts, source snippets, tool outputs, and model identifiers to the operator-selected provider.
  5. Git HTTPS and Git SSH clients connect to Forgejo. Git SSH uses a dedicated ~/.ssh/id_ed25519_clouddev workspace key registered to a regular Forgejo account. The key is distinct from the workspace Remote-SSH key. Forgejo uses PostgreSQL over its private network and has public registration disabled.
  6. The OmniRoute dashboard is privately hosted and reached through the workspace's cookie-protected /omniroute/ path. The user then signs in to OmniRoute with its separate native account. The loopback-only CLI adapter sends the health probe and OpenAI-compatible API routes to private dashboard port 20128. It does not forward the OmniRoute dashboard cookie. The launchpad shows service links and aggregate status only; it does not proxy an administrator session or store provider credentials.

External processing and telemetry

Railway EU West does not imply that provider processing stays in the EU. OpenAI, Anthropic, and other providers may process data outside Railway's deployment region under their current product and account terms. The operator chooses the provider, model, account, retention settings, and applicable contract.

OmniRoute stores configuration and usage/audit metadata in SQLite. The template sets short call-log retention, disables file logs, detailed request logs, stream-chunk capture, and selected background catalogue syncs. These settings do not prove that no prompt-derived information is stored in every upstream component, provider account, crash report, or platform log. Review the upstream version's logging behavior and the provider's data terms before using sensitive or regulated source data.

Railway may process deployment logs, service metrics, network metadata, and variable values according to its platform behavior. Sealed variables are not visible in the dashboard or retrievable by the API, but running services still receive them.

Public documentation crawlers

The documentation is public. Its robots.txt allows general search crawlers and dedicated answer-retrieval crawlers, including OAI-SearchBot, PerplexityBot, Claude-SearchBot, and Google-Extended. It disallows GPTBot and ClaudeBot. This blocks the training-specific crawlers whose search counterparts have separate controls. Googlebot remains allowed for ordinary search indexing. A robots.txt rule is a request, not access control. Anyone can still open or download a public documentation page.

The user-agent names and effects are documented by OpenAI, Perplexity, Anthropic, and Google.

The docs site allows Google-Extended so Google may use it for Gemini grounding. Google's token also covers using crawled content for future Gemini training. This is an intentional discoverability choice for public product documentation, not permission to publish private source code, workspace files, credentials, or user data.

Deletion and retention

Deleting a Git repository in Forgejo does not delete clones or copies in the workspace, backups, forks, provider logs, or external provider records. Deleting a workspace file does not purge the same content from an agent transcript or a provider. Rotate exposed browser, SSH, Forgejo, OmniRoute, and provider credentials at their respective source. To retire an instance, stop services, export required records, revoke credentials, and remove each persistent volume only after backups are verified. Railway volume deletion may become permanent after its documented recovery window.

Evidence status

  • Documented upstream behavior: CloudCLI 1.37.5 stores session data in SQLite and provider-native files; code-server supports password authentication; Railway private networking does not constitute outbound egress filtering.
  • Implemented locally: to be confirmed from the merged runtime, Compose stack, and executed tests.
  • Live Railway behavior: not verified. Do not describe the running service, region, proxy, volumes, or domain configuration as deployed until a fresh project has been checked.