- Thunderbird compatibility: Removed legacy Thunderbird-internal fallback code so Sortana now relies entirely on documented MailExtension/WebExtension APIs.
- Manifest cleanup: Dropped `strict_max_version` from the manifest, which avoids unnecessary version pinning for newer Thunderbird releases.
- Maintenance: Internal hashing and template-loading paths were simplified as part of the API cleanup, reducing legacy code in the classifier.

- Model selection: Load `/v1/models` from the configured endpoint to populate the Model dropdown; pick a model or choose **None** to omit the `model` field in requests.
- Experimental OpenAI auth: Optional API key plus `OpenAI-Organization` and `OpenAI-Project` headers for authenticated endpoints (used for model loading and classification requests).
- Maintenance: New “Reset Timing Stats” button to clear classification timing metrics without wiping cached results.

- Endpoint handling: Store a base URL, auto-append `/v1/completions`, default to `https://`, and preview the resolved endpoint on the
Options page.
- AI parsing: Extract the last JSON object in responses, require `match`/`matched` plus a short `reason`, and surface parse errors in-
session.
- Error visibility: New Errors tab with a session error log; toolbar error icon is now transient instead of sticky.
- Debugging UX: Prompt diff view now shows a token-reduction badge when reduction features are active.
- Prompt templates: Updated OpenAI/Qwen/Mistral prompts to include `reason`; added Harmony (gpt-oss) template option.
- Build tooling: New `build-xpi.sh` bash script alongside the existing PowerShell packager.

**v2.2.0**
- Compatibility: Add support for Thunderbird 140+ (manifest `strict_max_version` 140.*).
- Debugging: New Debug tab on the Options page (enable via Advanced settings) with:
- Payload viewer: Inspect the exact request body sent for classification.
- Diff view: Visualize differences between current and previous payloads.
- Live refresh: Updates as you tweak settings.
- Layout polish for clarity and readability.
- Optimization: Token-reduction strategies to keep payloads lean:
- Optional HTML→Markdown conversion, URL param stripping, image alt-text replacement, whitespace collapsing, and aggressive token reduction mode.
- Replace “context length” control with a simpler “Max tokens” cap and reduction strategies.
- Reliability: Endpoint changes now take effect without requiring a restart.
- Permissions: Add `notifications` to surface errors and persist an error indicator until cleared.
- UI/Icons: Numerous icon updates (including SVG controls), plus improved active/hover states and light/dark mode alignment.
- Rule Editor: Better condition toggles and visual states; minor styling and save-logic fixes.

**v2.1.0**
- Message Handling: Improved details view loading (handles preview pane and displayed-message contexts), robust message lookup, and better fallbacks.
- HTML/Content Options: New sanitization and formatting options (HTML→Markdown, URL param stripping, whitespace collapsing, base64/attachment handling).
- Data Portability: Export/Import of settings for easy migration.
- Theming: Light/Dark theme support with dynamic icons; fixes for context-menu icon visibility in dark mode.
- Internal: Added `normalizeRules` helper, shared default AI parameter module, and message processing refactors for cleaner logic and future features.
- Manifest: Targeted updates for TB 128+ while staying on manifest v2.

**Earlier Highlights (2.0.0)**
- Architecture: Refactor into a modular `AiClassifier`, shared message utilities, and removal of legacy experiment APIs/content scripts.
- Rules & Actions: Multi-action rules UI; apply rules from context menus; improved automatic classification wiring.
- Cache & Performance: Safer handling of large messages by stripping inline base64 and attachment content from AI payloads.
- Permissions/Docs: Updated permissions, README, and contributor guidelines to match the new architecture.