v1.2.6

Changed Prompt API now uses structured output ( responseConstraint ) for reliable JSON. Following Chrome's built-in AI do's and don'ts, content.js requests a JSON Schema constraint at session creation and parses the model's full JSON response directly, instead of relying on prompt instructions and a fragile "sentiment": prefill. The code falls back through all older option shapes ( initialPrompts + language options, initialPrompts alone, legacy systemPrompt ) so every supported Chrome build keeps working. Pre-warm the model as soon as the user focuses a comment field. This matches the recommendation to start the session once user intent is clear, reducing cold-start latency before the first analysis fires after the debounce. Cache recent analyses for identical text. A short in-memory cache (5-minute TTL) avoids re-running the model when the user re-types, edits back to a previous phrase, or switches between similar inputs. A new ↻ refresh button in the tooltip lets users force a fresh