SuperAPI AI Directory

gemini-runner

Agents & models

View repository

Updated 10-16 01:26

About this skill

Gemini Runner

Use shared/references/model-preview.md to choose the branch first. Deterministic checks need no model selection. Actual dispatch uses one concrete approved snapshot; nested calls consume it without re-resolution or extra workers.

Execute the caller's scoped prompt through the local CLI. Shared roles, envelope keys, and result handling live in shared/references/runner-common.md. Preserve seat fidelity: name the provider that actually answered, and report unavailable or unverified seats honestly.

Native routing

Read shared/references/host-model-execution.md before choosing an external route. If the current host can dispatch the exact selected Gemini model with the required effort, isolation, and receipt policy in a native subagent or task thread, use that route. agents/openai.yaml is host UI metadata; it does not dispatch a native model.

Invocation

Set SKILL_DIR to this loaded skill directory in the same shell call:

SKILL_DIR="<absolute path of this skill directory>";
python3 "$SKILL_DIR/scripts/run_gemini.py" "<scoped prompt>" \
  --model <approved-model-id> --disable-fallback --json

Copy the selected model label, role, tool mode, receipt policy, and supported limits from the snapshot. This runner has no effort flag or native effort control. A route requiring unsupported control stays blocked. Nested calls use the saved options without resolving defaults again. The example does not grant source sharing or tool authority.

Runtime boundaries

  • Use the local authenticated Antigravity CLI, agy, in print mode.
  • The user configures the model in agy. --model and output-format flags are request labels or prompt hints, not native model or format controls.
  • Permission checks follow local configuration. Prompt and file context may be sent to the configured provider.
  • Analysis roles use a read-only prompt overlay, not an enforced sandbox. Isolate the workspace for untrusted input; --restrict-tools alone is not a security boundary.
  • Missing-CLI fallback is labelled and described in the runtime reference. Approved routes use --disable-fallback; an unverified model requires the caller's explicit receipt policy.
  • Native output has no session ID. --agy-continue resumes the latest shared conversation and must not run concurrently against the same home.

Load by need

Result

Read agent_message and the wrapper envelope. Check success, serving-model receipt, effective runner/provider, and any fallback or malformed-output status. A completed process alone does not prove the requested task or model was used. Report touched files when the caller authorized edits.