Google Gemini

The gemini provider plans with Google Gemini models on the Gemini Developer API, through Google's official Gen AI Java SDK.

Set up

1. Create an API key in Google AI Studio. The key belongs to a Google Cloud project, which must be active.

2. Export it with a model name. Gemini model names change often, so there is no default:

   export GEMINI_API_KEY=...
   export NEXORA_LLM_MODEL="your-gemini-model-id"   # replace with a model your project can use

List the models your key can use:

curl -s "https://generativelanguage.googleapis.com/v1beta/models" -H "x-goog-api-key: $GEMINI_API_KEY" \
  | jq -r '.models[] | select(.supportedGenerationMethods | index("generateContent")) | .name'

Use the name without the models/ prefix. With only GEMINI_API_KEY set, the planner selects Gemini; with several keys set, add NEXORA_LLM_PROVIDER=gemini.

Settings

Variable	Default	Notes
GEMINI_API_KEY		Or NEXORA_LLM_API_KEY_FILE
NEXORA_LLM_MODEL	required	A Gemini model that supports generateContent
NEXORA_LLM_MAX_TOKENS	16000	Sent as maxOutputTokens; includes thinking on thinking models
NEXORA_LLM_TEMPERATURE	the model's default	0 to 2
NEXORA_LLM_EFFORT		Not accepted; the plugin refuses to start with it set
NEXORA_LLM_BASE_URL	Google's API	For a proxy or gateway

The overview lists the settings every provider shares.

How it behaves

• Structured output. The plan is requested as application/json constrained by the plan schema (responseJsonSchema).
• Safety. A blocked prompt, or an answer stopped by a safety filter (SAFETY, PROHIBITED_CONTENT, BLOCKLIST, SPII, RECITATION), falls back with REFUSED.
• Token limit. An answer stopped at MAX_TOKENS falls back with TRUNCATED. Thinking models spend part of the limit on reasoning.
• Retries. The SDK retries transient failures up to NEXORA_LLM_MAX_RETRIES times.

Gemini directly or through OpenRouter

Gemini is also reachable through OpenRouter (for example google/gemini-3.8-flash). Call Google directly when you already have a Google Cloud project and billing, want requests to go only to Google, or want Google's own quotas and data terms. Use OpenRouter to compare Gemini against other vendors behind one key.

Verified

• The Gemini client is tested against a local fake of the Gemini API: request shape, JSON schema, safety blocks, token-limit stops, and error mapping.
• Against the real API, a request reached Google through the SDK, and Google's 403 PERMISSION_DENIED fell back to the rule-based plan with REJECTED and Google's message.
• A real Gemini model (google/gemini-3.8-flash) planned a five-step goal correctly through OpenRouter.

Not yet verified: a successful plan on the native Gemini API. In particular, whether Gemini accepts every part of the plan schema (anyOf for nullable fields and additionalProperties: false). If it rejects them, calls fail with REJECTED and the engine plans with its rules, so nothing breaks, but Gemini would not plan. Report the result if you run it.

Troubleshooting

You see	Do this
REJECTED ... HTTP 403 PERMISSION_DENIED "has been suspended"	The key's Google Cloud project is suspended. Reactivate it in the Cloud console, or create a key in a new project
REJECTED ... HTTP 403 other	The Generative Language API is not enabled for the project, or the key is restricted
REJECTED ... HTTP 400 INVALID_ARGUMENT about the schema	Gemini rejected part of the plan schema; please report it
REJECTED ... HTTP 404	The model name is wrong; list models and drop the models/ prefix
REFUSED	A safety filter blocked the goal or answer; reword the goal or capability descriptions
TRUNCATED	Raise NEXORA_LLM_MAX_TOKENS