Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.
Gemini Agent SDK / Gemini Workflow Adapter
Use this skill when implementing a Gemini-backed runner for a workflow that also exists for Cursor, OpenCode, Claude, or OpenAI. This repo currently packages Gemini CLI skills/commands, so a Gemini workflow adapter should be added deliberately and kept compatible with the shared workflow contract.
Local references
docs/gemini-cli-setup.md documents current Gemini CLI packaging.agents/research-runner/ shows the multi-runtime runner pattern to mirror.agents/research-runner/shared/ should hold reusable prompts, templates, and output code.
Workflow adapter contract
Shared types: agents/research-runner/shared/workflow-contract.ts (WorkflowInput, WorkflowResult, optional safetyNotes).
No Gemini runner exists yet. Add agents/research-runner/gemini/runner.ts (or a new workflow folder) implementing this contract; keep .gemini/skills as CLI distribution only.
Implementation rules
- Use official Gemini docs for current SDK names and APIs. Do not rely on memory for package names, auth, tool-calling, or streaming semantics.
- Mirror the existing runner layout. Add
agents/<workflow>/gemini/runner.ts or equivalent beside the other harness variants.
- Keep CLI packaging separate.
.gemini/skills and .gemini/commands are distribution assets; SDK workflow runners are executable code.
- Use structured responses for machine decisions. Parse JSON for handoffs, verdicts, and next actions.
- Treat safety blocks as workflow states. Return
blocked with the exact model/safety reason instead of retrying blindly.
- Verify claims outside the model. File checks, git checks, and test commands belong to deterministic runner code.
- Keep prompts shared. Only put Gemini-specific prompt adaptations in the Gemini adapter.
Good fits
- Research and synthesis variants
- Provider comparison workflows
- Workflows where Gemini is selected by model policy
Poor fits
- Repo mutation without a local verification harness
- Assuming Gemini CLI skills are the same thing as a Gemini SDK runner
1---2name: gemini-agent-sdk3description: Build Gemini-backed agent workflows and adapters. Use for Google Gemini API/ADK-style runners, Gemini CLI workflow parity, tool calling, structured outputs, or adding a Gemini variant beside Cursor/OpenCode/Claude runners.4metadata:5 category: user-invoked6 version: 1.0.07 tags: agent-sdk, gemini, google, workflow-runner8disable-model-invocation: true9---1011Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.1213# Gemini Agent SDK / Gemini Workflow Adapter1415Use this skill when implementing a Gemini-backed runner for a workflow that also exists for Cursor, OpenCode, Claude, or OpenAI. This repo currently packages Gemini CLI skills/commands, so a Gemini workflow adapter should be added deliberately and kept compatible with the shared workflow contract.1617## Local references1819- `docs/gemini-cli-setup.md` documents current Gemini CLI packaging.20- `agents/research-runner/` shows the multi-runtime runner pattern to mirror.21- `agents/research-runner/shared/` should hold reusable prompts, templates, and output code.2223## Workflow adapter contract2425Shared types: `agents/research-runner/shared/workflow-contract.ts` (`WorkflowInput`, `WorkflowResult`, optional `safetyNotes`).2627No Gemini runner exists yet. Add `agents/research-runner/gemini/runner.ts` (or a new workflow folder) implementing this contract; keep `.gemini/skills` as CLI distribution only.2829## Implementation rules30311. **Use official Gemini docs for current SDK names and APIs.** Do not rely on memory for package names, auth, tool-calling, or streaming semantics.322. **Mirror the existing runner layout.** Add `agents/<workflow>/gemini/runner.ts` or equivalent beside the other harness variants.333. **Keep CLI packaging separate.** `.gemini/skills` and `.gemini/commands` are distribution assets; SDK workflow runners are executable code.344. **Use structured responses for machine decisions.** Parse JSON for handoffs, verdicts, and next actions.355. **Treat safety blocks as workflow states.** Return `blocked` with the exact model/safety reason instead of retrying blindly.366. **Verify claims outside the model.** File checks, git checks, and test commands belong to deterministic runner code.377. **Keep prompts shared.** Only put Gemini-specific prompt adaptations in the Gemini adapter.3839## Good fits4041- Research and synthesis variants42- Provider comparison workflows43- Workflows where Gemini is selected by model policy4445## Poor fits4647- Repo mutation without a local verification harness48- Assuming Gemini CLI skills are the same thing as a Gemini SDK runner
v1truv1us