feat(ge-demo-generator): Managed Autonomous Agent delegation, Workspace Authorization mode, and deliverable skills

[ryotat7]

Description

Major feature update for search/gemini-enterprise/ge-demo-generator (APP_VERSION v11.5-public).

What's new

• Managed Autonomous Agent (Antigravity) — the generated demo agent can now delegate long-horizon autonomous tasks (live web research, code execution, professional deliverables) to a Pre-GA Managed Agents API agent running in its own cloud sandbox, with live progress streamed into the Gemini Enterprise chat. The setup script provisions it in two phases so the ~8-10 min creation is hidden behind the rest of the setup. Enabled by default in the UI; requires only the Vertex AI API (no allowlist).
• agent_template/managed_agent/ — new provisioning helpers (create_managed_agent.py, warmup_managed_agent.py) as real template files, copied by the setup script (consistent with the template-fetch architecture from Feat/add ge demo generator #2963).
• agent_template/demo_skills/ — deliverable craft skills (SKILL.md packs for documents, presentation decks, and web reports) uploaded to the demo's dashboards bucket and mounted into the Managed Agent sandbox.
• Workspace Authorization (No MCP) — a new lightweight toggle that adds Google sign-in for the demo user without the Developer-Preview Workspace MCP allowlist. Combined with the Managed Agent, deliverables are saved to the user's Drive as native Google Slides / Docs / Sheets and the sandbox can act on Gmail / Chat / Calendar via the open-source Workspace CLI.
• Runtime templates (tools.py, agent.py, fast_api_app.py) gain env-gated feature blocks (ENABLE_MANAGED_AGENT, ENABLE_WORKSPACE_AUTH) following the existing ENABLE_WORKSPACE_MCP / ENABLE_COMPUTER_USE pattern; per-demo variation stays in .env / generated_instruction.md / mcp_config.json.
• Runtime robustness — A2UI healing fixes and a stub-recovery ladder (re-execute with tools when a turn stalls before gathering data).
• GEBE subproject — renamed to GE Demo Generator Lite (v1.2-public) with demo output language pinning (fast flash-lite pre-detection), plus a fix for the Apache header comment style in Code.gs (// instead of #, which is invalid JavaScript and broke clasp push — same fix as previously applied to app/Code.gs).
• Docs — README and AGENTS.md updated with the new architecture, env vars, and directory layout.

Template pinning

As during the #2963 review, TEMPLATE_REPO / TEMPLATE_REF are temporarily pinned to this PR branch's fork commit so generated setup scripts work before merge. After merge I will open a small follow-up PR re-pointing them at this repository (same as #2970).

Verification

• python3 validate_examples.py — 33/33 template files (24 JSON, 24 checked with json.loads, 9 Python compiled).
• Generated setup scripts across the feature matrix (all-off / MCP+Computer-Use / Managed-Agent+Workspace-Auth / all-on with varied scalars): bash -n clean, no embedded-content leaks, template fetch ordered before first use.
• Byte-level equivalence tests of the resolved agent instruction chain and runtime feature blocks against the internal reference implementation (both flag states).
• node --check on all Apps Script files and inline HTML scripts.
• Live sparse-fetch test of the pinned template ref (exact commands the setup script runs), then py_compile of the fetched files.