Antigravity SDK End-to-End Testing

Gotchas (these cost real time)

1. config.yaml's server: defaults to a remote server. Omitting --server sends your turn to that remote deploy — which may be stale and reject the antigravity harness with executor.config.harness: must be one of […], got 'antigravity'. Always pass --server http://127.0.0.1:<port> for local testing. (That allowlist is omnigent/spec/_omnigent_compat.py; if a local server rejects antigravity, it's running stale code — restart it from your checkout.)
2. A spec with spec_version must be a directory + config.yaml, never a single .yaml file.
3. Antigravity needs a Gemini key (no login). Resolution precedence: spec executor.auth (api_key) > stored antigravity: config block (omni setup)

   ambient GEMINI_API_KEY / ANTIGRAVITY_API_KEY. Vertex AI is opt-in via executor.config vertex/project/location.

4. No OpenAI gateway / Databricks. The SDK has no base_url; a databricks or generic-provider auth is warned and ignored, and the run falls back to ambient Gemini creds. Don't expect databricks-* models to route through the AI Gateway like claude-sdk/codex/pi.
5. Model ids are Gemini ids. Default gemini-3.5-flash. gemini-3-pro 404s on a plain AI-Studio key — use gemini-2.5-flash / gemini-3.5-flash unless your key has Pro access.
6. The native binary needs glibc ≥ ~2.36 (see Prereq 4). This is the most common "it won't even start" cause; check it before assuming a harness bug.
7. Turns take ~10–60s — always wrap in timeout 280.
8. Local-runner topology: omni run <bundle> --server <url> runs the harness from your current checkout; the server only holds state. The managed omni server --background server runs from whatever venv launched it.
9. Never print/echo the Gemini key in logs or commands.

Code & tests

• Executor (SDK driver): omnigent/inner/antigravity_executor.py
• Wrap (HARNESS_ANTIGRAVITY_ env → executor):* omnigent/inner/antigravity_harness.py
• Auth / key resolution: omnigent/onboarding/antigravity_auth.py
• Spawn env: _build_antigravity_spawn_env in omnigent/runtime/workflow.py

# Unit tests (use --frozen; the cwsandbox extra is unsatisfiable on public PyPI here)
uv run --frozen --group test python -m pytest \
  tests/inner/test_antigravity_executor.py \
  tests/inner/test_antigravity_harness.py \
  tests/runtime/test_antigravity_spawn_env.py \
  tests/onboarding/test_antigravity_auth.py -q
# (or, if uv re-resolve is blocked on your host: .venv/bin/python -m pytest <same paths> -q)

There is no gated per-harness antigravity e2e test yet (it is deliberately excluded from the live no-AGENT harness matrix in tests/e2e/omnigent/test_run_harness_without_agent_e2e.py, because that matrix authenticates through the Databricks gateway and antigravity is Gemini-native). This skill IS the live coverage.

Bug-bash (fan out)

To stress the harness, run several scenario probes in parallel — each builds a bundle and runs real turns against the same $SERVER, then reports what broke. Highest-value targets: the custom_tools bridge (hangs / lost tool results / errors reported as success), model routing, policy enforcement, streamed-output rendering, history retention across turns, and orphaned localharness processes after teardown.

Known sharp edges (found via the merge review — "as of this writing")

Several were merged as-is and have fix PRs in flight (#276–#281) — verify against your checkout:

• Native/built-in tools bypass the TOOL_CALL policy. Only a PostToolCallHook (post-execution, can't block) was installed at merge, so a DENY/ASK guardrail doesn't gate the SDK's native shell/file tools before they run. Bridged sys_* tools route through the server. (Fix: policy-enforcement PR.)
• LLM_REQUEST / LLM_RESPONSE policies aren't evaluated in run_turn (prompt- deny / output-block silently ignored). (Fix: policy-enforcement PR.)
• History on a fresh/rebuilt session. The SDK has no history-injection API, so prior turns are replayed as a plain-text "Conversation so far: …" prefix (user/assistant text only; tool calls aren't reconstructed). (PR #278.)
• sys_list_models can over-report OpenAI-family models for antigravity (it was mapped to the openai family for shared lookups); the worker only runs Gemini. (Fix: openai-family-cleanup PR.)
• Per-session /model override was rejected with a false "no plumbing" error. (PR #276.) Global auth: (an OpenAI key) could be adopted as a Gemini key. (PR #277.) Tool parameter schemas were dropped (model flew blind on arg shapes). (PR #279.)
• A failed turn (e.g. the glibc error, a bad model) surfaces as a failed session + an error item — if a turn returns little, check GET /v1/sessions/{id} status and …/items rather than assuming success.

Cleanup

.venv/bin/omni server stop      # stop the managed background server
rm -rf /tmp/agy-dev             # remove scratch bundles
pgrep -af "localharness"        # confirm no orphaned native subprocesses linger