Test Antigravity preview as an optional Agent Factory backend

[Copilot]

Canonical issue

Tracked by the linked canonical issue.

Outcome

This PR tightens the existing Antigravity spike into the contract described in the issue: one bounded build/test interaction through antigravity-preview-05-2026, one read-only EventRelay MCP server, synchronous side-effect gates, and a durable receipt/comparison artifact. The default provider path remains unchanged; this only hardens the optional preview backend.

Scope

• Included:
  • Payload contract
    • Adds code_execution to the managed-agent request.
    • Prefixes the prompt as a bounded build/test task and keeps Video Pack data as normalized text context only.
    • Mounts inline .agents/hooks.json plus a synchronous policy_gate.py script in the remote environment.
  • Safety contract
    • Requires exactly one MCP server, named eventrelay.
    • Tightens MCP server naming to lowercase alphanumeric only.
    • Enforces HTTPS Streamable HTTP plus explicit read-only tool allowlisting.
    • Records hook acknowledgement and synchronous guard policy in the receipt.
  • Comparison artifact
    • Expands the provider-neutral comparison output to cover completion, latency, cost, artifact determinism, provenance, recovery, and provider-independent control-plane ownership.
  • Contract tests
    • Extends mocked unit coverage for payload shape, hook wiring, MCP validation, receipt policy, and comparison output.
  • Example

    payload = backend.build_payload("Run pytest", {"video_pack_id": "pack-1"})
    assert payload["tools"][0] == {"type": "code_execution"}
    assert payload["tools"][1]["name"] == "eventrelay"
    assert payload["environment"]["sources"][0]["target"] == ".agents/hooks.json"

• Explicitly excluded:
  • Live preview execution or billing-incurring runs
  • Any default-provider switch or production rollout
  • Direct video ingestion into Antigravity
  • Additional MCP servers or write-capable tool access

Risk

• Risk level: low
• Failure mode:
  • The optional preview backend could reject misconfigured MCP names/servers or deny undeclared code/filesystem actions more aggressively than before.
  • Comparison consumers may need to handle the richer artifact shape.
• Rollback:
  • Revert this backend/test delta or keep the backend disabled via its existing feature flag and live-execution gates.

Verification

List exact automated and manual checks, tied to the current head SHA.

• Focused tests
  4c0106c6a3a8f31d981ed966d85bf7a3ae6495c3
python -m pytest --no-cov tests/unit/test_antigravity_backend.py tests/unit/test_antigravity_orchestration.py -q
python -m ruff check src/youtube_extension/services/agents/antigravity_backend.py tests/unit/test_antigravity_backend.py
• Required CI
• Review threads resolved

Production evidence

Not applicable. This PR does not deploy or change the default execution provider, and no live Antigravity preview run was performed.

Agent handoff

• One canonical issue is linked
• No competing PR implements the same issue
• Acceptance criteria are satisfied
• Required checks pass on the current head
• Human decision is requested only for product, security, irreversible infrastructure, or production approval

• Fixes Test Google Antigravity as optional Agent Factory execution backend #1630

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
v0-uvai	Ready	Preview, v0	Sep 8, 2026 11:23pm UTC

[vercel]

Policy gate authorizes code_execution via command.startswith(allowed_prefixes), so any shell command beginning with an allowed prefix (e.g. pytest && curl ... | sh) bypasses the guard and runs arbitrary undeclared side effects.