Build Gemini API integrations with multi-turn chat and agents

A skill for writing code against Gemini's Interactions API, covering text generation, multi-turn chat, multimodal understanding, image and video generation, streaming, background research, function calling, structured output, and migration from the legacy generateContent API. It documents current models to use: gemini-3.5-flash (1M tokens, fast, multimodal), gemini-3.1-pro-preview (1M tokens, complex reasoning/coding/research), gemini-3.1-flash-lite (cost-efficient, high-frequency lightweight tasks), gemini-3-pro-image/gemini-3.1-flash-image/gemini-3.1-flash-lite-image (Nano Banana Pro/2/2-Lite image generation and editing at 65k/32k tokens), gemini-3.1-flash-tts-preview (expressive TTS with Director's Chair prompting), gemini-omni-flash-preview (video generation and editing), and Gemma 4 open models - explicitly flagging gemini-2.5/2.0/1.5 as deprecated and to be substituted with gemini-3.5-flash. Current agents include antigravity-preview-05-2026 (general-purpose sandboxed Linux agent with code execution and web access), deep-research-preview-04-2026 and deep-research-max-preview-04-2026 (fast vs. maximally exhaustive research), plus custom agents via client.agents.create(). Current SDKs are google-genai >= 2.3.0 (Python) and @google/genai >= 2.3.0 (JS/TS); the legacy google-generativeai and @google/generative-ai packages are deprecated. Key rules: interactions are stored by default (55 days paid tier, 1 day free tier, disable with store=false, which also disables previous_interaction_id and background=true); tools/system_instruction/generation_config are interaction-scoped and must be re-specified every turn; managed agents require environment="remote" to provision a sandbox; and migration guidance lives in references/migration.md, to be consulted (and scope confirmed with the user) before editing. The quick-start pattern is client.interactions.create(model=..., input=...), returning an Interaction object with convenience accessors output_text (the trailing text run), output_image (base64 data plus mime_type), and output_audio (same shape). Stateful multi-turn conversation is achieved by passing previous_interaction_id from one interaction into the next, letting the server retain context server-side rather than resending history. Deep Research agents require background=True and are polled via client.interactions.get(id) until status is "completed" or "failed"/"cancelled".

Use it whenever writing code against the Gemini API - the skill explicitly states its rules override the model's training data since that knowledge is outdated, and that the hosted docs (not this skill's minimal examples) should be fetched before writing code for the full API surface.

Input is a model or agent name plus an input prompt (and optionally previous_interaction_id for continued context, or background=True for async agents). Output is an Interaction object with output_text/output_image/output_audio accessors, or a polled completion status for background agent runs.

pip install -U google-genai
npm install @google/genai

Covers the Gemini Interactions API across Python and JavaScript/TypeScript SDKs, with managed agents requiring a remote sandbox environment.

Developers building on the Gemini API who need current, non-deprecated model names, SDK versions, and interaction patterns - stateful chat via previous_interaction_id, background research agents, and the store/tools/system_instruction scoping rules that differ from the legacy generateContent API.