[Bug] Intelligence always returns "check intelligence settings" error despite valid API key (400 Bad Request)

[Akv2021]

Describe the bug
Cashew Intelligence fails on every prompt with a popup error: "Intelligence error - check intelligence settings and try again." The Gemini API dashboard confirms the requests are reaching Google but returning 400 Bad Request.

Setup

• Created a fresh Gemini API key (no restrictions) and added it in Intelligence settings
• Key validation ("Test") shows success
• AI Model: "Default" followed by Gemini 2.5 Flash, 2.5 Flash Lite, and 3 Flash individually, same error on all
• Category Names: enabled
• All other Intelligence settings are default (default prompts, no custom rules, Account Names disabled, Recommend Category disabled, Review Before Adding enabled)

How to reproduce

1. Enable Intelligence in Experimental Features
2. Add a valid Gemini API key → Test succeeds
3. On the home page, swipe up on the (+) button to open Intelligence prompt
4. Enter: Purchase Snacks from The ABC store for Rs 300 today
5. Error popup appears immediately

Observations

• Gemini API usage dashboard shows 100% success rate for requests, but the API Errors section lists 400 Bad Request entries corresponding to each attempt
• Reproduced on two different devices with the same result
• Intelligence - Invalid API Key #939 (deprecated model) was the closest match but was resolved in 6.2.4+466 with the model selector, which does not fix this

Ask

1. How can this be resolved? Is there a specific model or setting combination that works?
2. Could the error surface be improved? The current generic "check intelligence settings" message gives no actionable information. Surfacing the actual API error response (e.g., the 400 reason) or adding debug/verbose logging would help users and bug reporters diagnose issues.

Details

• Device: Android (reproduced on 2 devices)
• App Version: 6.2.8+471 (beta)

[jameskokoska]

Unfortunately I wasn't able to replicate the issue. It seems though that Google has been quietly nerfing the functionality of the free tier for Google Gemini (i.e. giving lower priority, slower responses etc).

The reason the error message says to "check intelligence settings and try again" is because the settings page will provide the full error message if the API check fails.

The error you see in the console (Bad request 400) indicates an issue with the input data. This may be fixed in the next beta release for you (as some requests that contain missing information would be a malformed request).

For the issue you are experiencing, it is strange it wouldn't fail on the API check, it indicates your connection is fine. All responses are logged and can be accessed in the About -> Debug Logging page. You can add the log outputs here so I can further investigate the issue. I am going to close this issue for now until more details are provided/it can be replicated reliably.

[Akv2021]

Hi @jameskokoska ,
Here's the error - https://gist.github.com/Akv2021/221af10127316c8868148d46f46671a5

1. I suspect the no of categories is causing the problem here. Although I'd derived from Gemini dashboard that per request average token consumption was 2.37K for Gemini 2.5 models which is very tiny as compared to the average payload of 50K tokens per request on the same models I see at my job.
2. I couldn't find a way to update the default prompts (Rules field would be an addendum to the prompt, I'm seeking for prompt replacement). Is there a plan to make it client configurable in later releases (Atleast the prompt, if not the LLM config like reasoning, temperature etc)

[jameskokoska]

Thanks for the details, it seems that if there are too many enums passed into the properties, Gemini can't process the prompt. I guess the best course of action isn't to use the enums as a selection method but rather to let the LLM return any string and match to an existing category.