Openrind Shell

Supported launch flow

1. Configure PostgreSQL and the upstream provider credential in Desktop Settings.
2. In Sandboxes, choose New sandbox → Claude Code (or OpenClaw).
3. Desktop registers the scoped Haloop provider, creates the agent-home named volume, initializes FUSE, and issues a signed conversation context.
4. Select an existing sandbox to return to its live session. After /exit, use Reconnect for a new signed launch. Keep the agent-home volume for history.

Do not invent a provider ID for a standalone CLI launch. Never use --auto-providers, a direct Anthropic provider, or an unsigned claude command as a substitute. Missing Haloop readiness blocks launch; recovery uses Desktop's restart/reconnect controls, never a direct-provider fallback.

Diagnostics

Read README.md and openrind-desktop/apps/desktop/OPENRIND_SHELL.md for the current flow and BUILD.md for source builds. Do not rebuild NVIDIA's base. Never print database URLs or provider keys, or pass provider keys through --env.

Use the paired patched CLI and gateway for FUSE diagnostics. Run sandbox exec -n <sandbox-name> -- openrind-shell-fused health through that CLI with the managed gateway endpoint. The filesystem state must be writable. Ready alone does not prove initialization succeeded. A manual sandbox connect opens a diagnostic shell, not a signed agent launch.

Persistence and safety

• /sandbox/work is PostgreSQL-backed FUSE, not a watched or mirrored folder.
• Claude uses /sandbox/claude-home, on a separate persistent named volume.
• One writable sandbox per workspace; do not create a duplicate to resume.
• FUSE daemon exit or lease loss ends the session. Never grant Claude mount capabilities, /dev/fuse, or a PostgreSQL/TLS-bypass fallback.
• Never delete a sandbox or volume without the user's explicit request.
• Compatibility is a separately requested legacy runtime, not recovery for a failed mandatory Haloop launch. Do not reuse its old provider recipe.
• Desktop captures traces only. Export them to the w8-haloop web app for analysis and Harbor task generation. Post-start capture is best-effort; a successful inference response does not prove its trace was stored.