Add Pollinations.AI image description + global max-tokens + language-aware default prompt

[keyang556]

Summary

• Pollinations.AI provider: New image description backend using gen.pollinations.ai/v1/chat/completions (OpenAI-compatible, authenticated tier). Bundled API key is encrypted with the existing PBKDF2-HMAC scheme. Four models with friendly labels: claude-fast → Claude Haiku 4.5, openai → GPT-5.4 Nano, openai-fast → GPT-5 Nano, openai-large → GPT-5.4. Requests use User-Agent: curl/8.4.0 to avoid the 403 the Pollinations edge layer returns for Python's default UA.
• Default prompt: Changed from Chinese to the AI Content Describer standard English prompt. When using the default prompt, a language hint (Respond in <Language>.) is automatically appended based on NVDA's UI language (zh_TW/zh_HK → Chinese Traditional, zh_CN → Chinese Simplified, ja → Japanese, and 20+ other locales).
• Max-tokens setting: Global user-configurable max output tokens (50–4000, default 700) exposed as a SpinCtrl in the settings panel. Applied to all backends: Google (generationConfig.maxOutputTokens), NVIDIA, and Pollinations (max_tokens). NVIDIA previously had this hardcoded at 512.

Technical notes

• Endpoint discovery: text.pollinations.ai/openai is the legacy API — it only exposes one text-only model and returns 404/routing errors for authenticated users requesting models like claude-fast. The correct authenticated endpoint is gen.pollinations.ai/v1/chat/completions (confirmed working with live vision test).
• All new storage files follow the existing pattern under NVDA's config dir: line_desktop_pollinations_api_key.dat, line_desktop_pollinations_model.txt, line_desktop_image_max_tokens.txt.
• Model dropdown for Pollinations shows user-friendly labels; internal model IDs are still sent to the API.
• _nvdaLanguageHint() reads NVDA's languageHandler.getLanguage() and maps locale codes to English language names. Language hint is only appended when the user has not customised the prompt.

Test plan

• Open Settings → LINE Desktop, verify "Pollinations.AI" appears in the provider dropdown
• Select Pollinations.AI, confirm model dropdown shows Claude Haiku 4.5 / GPT-5.4 Nano / GPT-5 Nano / GPT-5.4
• Verify max-tokens SpinCtrl appears (range 50–4000)
• Focus an image message in LINE and press NVDA+Windows+I with Pollinations selected
• With NVDA set to zh_TW language, confirm description prompt appends "Respond in Chinese Traditional."
• Set a custom prompt and confirm the language hint is NOT added
• Confirm Google and NVIDIA still work and now respect the max-tokens setting

🤖 Generated with Claude Code

Greptile Summary

此 PR 新增 Pollinations.AI 圖片描述後端（OpenAI 相容端點），引入全域使用者可設定的 max-tokens SpinCtrl（50–4000，預設 700），並將預設提示語從中文改為英文加上基於 NVDA UI 語言的自動語言提示（_nvdaLanguageHint）。

• Pollinations.AI 後端：新增完整的 API 金鑰管理、模型選擇（4 個視覺模型附友善標籤）、OpenAI 相容請求函式，以及 User-Agent: curl/8.4.0 繞過邊緣層 403 的機制。
• 全域 max-tokens：NVIDIA 後端從硬編碼 512 改為 _getEffectiveImageMaxTokens()（預設 700）；Google 後端沿用 getUserImageMaxTokens()，僅在使用者明確設定時才送出 maxOutputTokens；Ollama 後端目前未套用此設定。
• 語言感知提示：_getRuntimeImagePrompt() 僅對預設提示語附加 Respond in <Language>.，自訂提示保持不變。

Confidence Score: 5/5

所有新功能均有完善的例外處理與回退機制，滞次 PR 可安全合併。

新後端遵循現有的金鑰加密與快取模式，_geminiContentsToPollinationsMessages 的 OpenAI 格式轉換邏輯正確，_nvdaLanguageHint 對未知語言回傳 None 不會影響既有行為；唯一發現的小問題為 _loadCurrentMaxTokens 例外分支使用硬編碼常數，不影響執行期正確性。

無需特別關注的檔案。

Important Files Changed

Filename	Overview
addon/appModules/line.py	新增 Pollinations.AI 後端、_nvdaLanguageHint()、_getRuntimeImagePrompt()，並將 Google/NVIDIA 套用全域 max-tokens 設定
addon/globalPlugins/lineDesktopHelper.py	設定面板新增 Pollinations.AI 支援與 max-tokens SpinCtrl；_loadCurrentMaxTokens except 分支有硬編碼 700 的細微問題

Comments Outside Diff (1)

1. addon/appModules/line.py, line 3276-3277 (link)

   Ollama 後端不套用全域最大 Token 設定

   _callOllamaImageDescriptionApi 的請求 body 未包含 max_tokens 欄位，但其他三個後端（Google、NVIDIA、Pollinations）都已套用 _getEffectiveImageMaxTokens()。使用者在設定面板調整最大 Token 後切換到 Ollama，輸出長度不受控制。Ollama /api/chat 支援 options.num_predict 來限制輸出，或可在 body 最上層傳入 num_predict，視服務端版本而定。

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: addon/appModules/line.py
   Line: 3276-3277
   
   Comment:
   **Ollama 後端不套用全域最大 Token 設定**
   
   `_callOllamaImageDescriptionApi` 的請求 body 未包含 `max_tokens` 欄位，但其他三個後端（Google、NVIDIA、Pollinations）都已套用 `_getEffectiveImageMaxTokens()`。使用者在設定面板調整最大 Token 後切換到 Ollama，輸出長度不受控制。Ollama `/api/chat` 支援 `options.num_predict` 來限制輸出，或可在 body 最上層傳入 `num_predict`，視服務端版本而定。
   
   How can I resolve this? If you propose a fix, please make it concise.

   

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
addon/globalPlugins/lineDesktopHelper.py:369-379
`_loadCurrentMaxTokens` 的 `except` 分支直接回傳硬編碼的 `700`，而非使用 `_IMAGE_DESCRIPTION_DEFAULT_MAX_TOKENS` 常數。若日後預設值調整，這個例外分支的回傳值不會跟著更新，導致設定面板顯示錯誤的初始值。

```suggestion
	def _loadCurrentMaxTokens(self):
		_defaultMaxTokens = 700
		try:
			from appModules.line import (
				_IMAGE_DESCRIPTION_DEFAULT_MAX_TOKENS,
				getUserImageMaxTokens,
			)

			_defaultMaxTokens = _IMAGE_DESCRIPTION_DEFAULT_MAX_TOKENS
			return getUserImageMaxTokens() or _IMAGE_DESCRIPTION_DEFAULT_MAX_TOKENS
		except Exception:
			log.debug("LINE: cannot load image max tokens", exc_info=True)
			return _defaultMaxTokens
```

_{Reviews (2): Last reviewed commit: "Fix Google backend: only send maxOutputT..." | Re-trigger Greptile}