feat(cloud): expose connectProvider() in @agent-relay/cloud SDK

[khaliqgant]

Summary

Lifts the Daytona-brokered provider auth flow out of the relay CLI and into the published @agent-relay/cloud package so other CLIs (ricky, future tools) can drive it directly via import { connectProvider } from '@agent-relay/cloud' instead of shelling out to agent-relay cloud connect.

• New SDK exports: connectProvider, getProviderHelpText, normalizeProvider, runInteractiveSession, formatShellInvocation, wrapWithLaunchCheckpoint, loadSSH2, createAskpassScript, buildSystemSshArgs, DEFAULT_SSH_RUNTIME, plus SSH-related types
• ssh2 is now an optionalDependency of @agent-relay/cloud — when missing, the existing system-ssh fallback path is used (behavior unchanged, just newly visible to consumers)
• @agent-relay/cloud bumped 6.0.2 → 6.1.0 (additive)
• The agent-relay cloud connect command body shrinks from ~150 LOC to a thin wrapper around connectProvider() that keeps telemetry attribution
• src/cli/lib/{ssh-interactive,auth-ssh}.ts now re-export the SDK helpers (rather than redefining them) so the existing auth command path is unchanged
• ssh-interactive.test.ts moved to packages/cloud/src/lib/ alongside the implementation

Net diff: ~1000 lines deleted (deduped to the SDK), ~600 added in the new package files.

Why

Other tools in the AgentWorkforce family (ricky in particular) want a <tool> connect <provider> UX that uses the same Daytona-brokered auth flow. Today they'd have to either shell out to agent-relay cloud connect or copy the SSH/PTY code. Extracting it into the published cloud package gives them a clean import boundary.

Test plan

• vitest run in packages/cloud (32 passed — includes the moved ssh-interactive tests + new connect unit tests for normalizeProvider/getProviderHelpText)
• vitest run src/cli/commands/{cloud,auth}.test.ts src/cli/bootstrap.test.ts (14 passed — confirms CLI wiring still resolves through the SDK)
• tsc --noEmit at repo root — clean
• Smoke agent-relay cloud connect claude end-to-end against staging cloud (not run from this branch — needs a cloud account)
• Once merged + published, ricky imports connectProvider and ships ricky connect <claude|codex|opencode> (follow-up PR in the ricky repo)

Compatibility

• @agent-relay/cloud 6.0.x consumers see only additive surface — no removed exports
• Relay CLI's auth command (agent-relay auth <provider>) is unchanged — same SSH path, just imported from the cloud package
• ssh2 moves from a regular dep (at the relay repo root, where it stays) to an optionalDependency of @agent-relay/cloud so SDK consumers get it on a clean install but can npm install --no-optional if they want to skip the native build

🤖 Generated with Claude Code

---

[barryollama]

👋 Review from Barry (barryollama)

Thanks for the review request! I took a look at this refactoring PR and overall it looks like a solid architectural improvement. Here are my thoughts:

What I like:

1. Clean SDK extraction — Moving connectProvider(), runInteractiveSession(), and the SSH runtime helpers into @agent-relay/cloud (6.1.0) creates a proper import boundary for other CLIs like ricky. The cloud package now owns the SSH implementation details.

2. Backward compatibility — The existing CLI commands still work by re-exporting from the SDK, which is important for existing users.

3. Optional dependency — Making ssh2 an optionalDependency is the right call. It allows consumers without the native module to fall back to system ssh seamlessly.

4. Type safety — The exported types give consumers a clear API contract.

A couple of observations:

• The CLI command shrinks from ~150 LOC to a thin wrapper — excellent deduplication.
• The moved ssh-interactive.test.ts plus new connect.test.ts shows the test coverage is being preserved.
• The CHANGELOG.md entry is comprehensive and follows the project's format.

Overall assessment:

• This is a well-structured refactoring.
• The extraction reduces duplication while maintaining backward compatibility.
• The test plan in the PR description covers the right surface area.

LGTM! 🚀

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 5 additional findings.