feat(agent-connector): fully integrate Gemini CLI and fix stream JSON parsing bugs

[zhoumzh]

Description

This PR fully integrates the Gemini CLI into the OpenAgents ecosystem and resolves several critical issues that previously prevented the Gemini adapter from working correctly in the Desktop Launcher.

🐛 Bug Fixes

• JSON Stream Parsing: Fixed a critical bug in GeminiAdapter where split('\\n') was mistakenly used instead of split('\n'). This prevented the CLI's stream-json output from being parsed correctly, resolving the silent No response generated errors.

✨ Features & Improvements

• Native Authentication Flow: Removed the mandatory GEMINI_API_KEY UI input requirement in registry.json and added login_command: "gemini login". This aligns Gemini with Claude's elegant CLI OAuth flow.
• Aggressive Polling for Lower Latency: Reduced the maximum idle adaptive polling delay in BaseAdapter._pollLoop() from 15 seconds to 3 seconds. This significantly improves responsiveness when chatting with local agents from the remote Workspace.
• Enhanced Debuggability: Added full stack trace (e.stack) logging for joinNetwork failures to make troubleshooting ECONNREFUSED and 404 errors easier.

🛠 Development Environment

• Local Source Prioritization: Modified AgentManager.loadCore() to prioritize loading the local workspace agent-connector code instead of the globally installed @openagents-org/agent-launcher. This makes local development and hot-reloading functional.
• Headless & GPU Options: Added --disable-gpu and --headless startup flags in package.json and main.js to mitigate severe UI lag during local Electron testing.

Documentation

Included a detailed changelog and a technical blog post documenting the debugging journey and infrastructure fixes in changelogs/notes/gemini_integration_fix.md and changelogs/blogs/2026-04-23-gemini_integration_journey.md.

[vercel]

Someone is attempting to deploy a commit to the Raphael's projects Team on Vercel.

A member of the Team first needs to authorize it.

[zomux]

Landed on develop in 8d79ed0 (authored as you), with one follow-up immediately after in 9f5b3f7 that reverts just the polling-rate change in base.js (2s/15s → 1s/3s). The polling change is orthogonal to Gemini integration and warrants its own PR with a load-impact analysis — every connected agent runs that loop, so faster rates multiply across the whole workspace, and we recently tuned the head-cursor cache + Postgres pool for the existing rate.

Conflict resolution note: this PR was branched before e358b64 ("Fix per-agent model config in launcher"), so its agent-manager.js changes would have reverted the per-instance env-var feature. Took develop's version of agent-manager.js entirely; the local agent-connector source priority in loadCore() is already on develop via PR #359 with equivalent logic.

Two small follow-ups worth knowing about:

1. gemini.js:248 — cleanEnv = { ...(this.agentEnv || process.env) } doesn't actually clean. We just merged a fix in claude.js (d9f5449) that strips CLAUDE_* / ANTHROPIC_API_KEY / AI_AGENT_* before spawning the subprocess. Same scrub probably belongs here so a Gemini agent doesn't see Anthropic creds in its env.
2. There's a leftover dev-thinking comment block around line 365 of gemini.js ("Gemini deltas are partial strings, no newline needed... but wait...") — fine, just noise.

Thanks for the bug fix on split('\n') — that's been masked behind the "No response generated" message for a while.