Fixing Gemini JSON Parsing, Token Limits, and Calendar Font Scale in Tagmind

What Was Built

A focused polish pass fixing a persistent Gemini response bug, improving resilience for all LLM providers, and fixing a UI regression at maximum system font size.

The Gemini JSON Bug (Root Cause Found)

For several sessions, Gemini was displaying {"answer":"Based on your notes... as the literal answer text instead of the parsed value. Multiple attempts to fix Gson deserialization didn't resolve it.

The actual root cause was maxOutputTokens: 512. Gemini's structured JSON response was getting cut off before the closing }, producing malformed JSON. JSONObject threw, lastIndexOf('}') returned -1, and the raw string fell through to the answer field.

Fixes applied:

1. Raised token limits — All providers now use 4096 output tokens. Gemini omits maxOutputTokens entirely for search calls, using the model's default (8192).

2. Replaced Gson with JSONObject — Android's built-in JSONObject(json).optString("answer", "") has no generic type-erasure issues and handles malformed input gracefully. Removed the LlmJsonResponse data class entirely.

3. Truncation fallback — Even if the JSON is malformed (no closing }), a string walk extracts whatever answer text was generated before the cutoff:

   val answerKey  = cleaned.indexOf("\"answer\"")
   val colonIdx   = cleaned.indexOf(':', answerKey + 8)
   val quoteOpen  = cleaned.indexOf('"', colonIdx + 1)
   val quoteClose = cleaned.indexOf('"', quoteOpen + 1)
   val answer = if (quoteClose == -1) cleaned.substring(quoteOpen + 1).trim()
                else cleaned.substring(quoteOpen + 1, quoteClose).trim()

Token Limit Warnings

Added a yellow warning box to both Settings → LLM and the onboarding tutorial (Ask AI page):

"AI responses are limited to 4,096 tokens. Very long responses may be cut off mid-sentence."

The right place for this information is at the moment the user is configuring the feature — not in a README.

Calendar Heatmap: Large Font Fix

At maximum system font size, the month labels in the Note Activity heatmap (labelMedium — sp-based) grew wider than the 13dp cell boxes, making each column wider and overflowing the card.

Fix: Wrap month labels in CompositionLocalProvider with fontScale = 1f:

val d = LocalDensity.current
CompositionLocalProvider(
    LocalDensity provides Density(density = d.density, fontScale = 1f)
) {
    Text(monthName, style = MaterialTheme.typography.labelMedium, ...)
}

The labels are decorative (heatmap axis ticks), so locking them to font scale 1.0 is appropriate. The rest of the screen still respects the user's accessibility setting.

Files Changed