[katerina20]

Part of updating the SDK to the current Crowdin API docs. This PR covers the AI API.

Crowdin routes live under /users/{userId}/…; for Enterprise, pass userId = null, the same convention as the existing AI methods.

New methods (AIApi)

Methods	Route
listAiSnippets, addAiSnippet, getAiSnippet, editAiSnippet, deleteAiSnippet	ai/settings/snippets[/{aiSnippetId}]
exportAiRequestLogs, checkAiRequestLogsExportStatus, downloadAiRequestLogsExport	ai/request-logs/exports[/{exportId}[/download]]
listAiUsageMembers, getAiUsageMember	ai/usage/members[/{memberId}]
listAllAiProviderModels	GET ai/providers/models
getProjectAiSettings	GET /projects/{projectId}/ai/settings
aiGatewayGetRaw / PostRaw / PutRaw / PatchRaw / DeleteRaw	ai/providers/{aiProviderId}/gateway/{path}

The snippets reuse the existing CustomPlaceholder models, because the schema is identical.

Updates

• AiSetting: added the pre-translation, QA-check, context-review and alignment (Enterprise) prompt ids, plus dailyCostLimit, monthlyCostLimit, userDailyCostLimit, userMonthlyCostLimit, isLimitingActive and perUserOverrides.
• AI provider Credentials: now holds the fields of every documented provider type: base URL and headers, Azure, Vertex AI, watsonx, Microsoft Foundry deployments, and custom AI.
• AiPrompt: added createdBy, updatedBy, lastUsedBy, lastUsedAt, usageCount.
• finishedAt: added to prompt completions and fine-tuning datasets.
• AiProviderModel: added the model capability fields.
• AI reports: languageIds added.
• Prompt completion request:
  • tools is sent as the documented array (aiTools).
  • tool_choice accepts a string or an object (toolChoice).
• Fine-tuning: learningRateMultiplier is now a number (learningRateMultiplierValue).
• Visibility: AiSetting.ShortCut and AiFileTranslation.FileTranslationError are now public; they were exposed by public getters but couldn't be used.
• Javadoc: the AI prompt methods now link the Crowdin operations, and the gateway methods have Enterprise links.

Deprecated (still working)

• Custom placeholder methods (list/add/get/edit/deleteCustomPlaceholder): the API says to use AI snippets instead.
• aiGatewayGet/Post/Put/Patch/Delete: the API returns the provider's response without a data wrapper, so these returned null or failed. Use the *Raw methods.
• Replaced fields: the tools/tool_choice maps and the Long learningRateMultiplier.
• Fields the API deprecates: AiPromptAddRequest.isEnabled and AiProviderRequest.config.
• Fields no longer in the API:
  • AiSetting.showSuggestion/shortcuts.
  • AiPromptResource.filteredStringsIds/projectDescription.

Backward compatibility

No public method, field or class was removed, renamed or retyped. Behaviour corrections:

• setTools(Map): the value is no longer sent, because the API expects an array and there is no sensible conversion. Use setAiTools.
• setTool_choice(Map): fills toolChoice.
• learningRateMultiplier: the Long setter fills the number field, so 2 is sent as 2.0. The Long getter returns the value truncated to an integer (1.8 reads as 1).

Open questions for review

1. New models carry fields the API marks deprecated:

   • AiPromptCompletionData.eta.
   • Project-level ProjectAiSettings.preTranslationAiPromptId.

   Both are annotated @Deprecated. Decide whether to keep them (they are returned by the API) or drop them. The user-level AiSetting.preTranslationAiPromptId is not deprecated.

Testing

• ./gradlew build: 660 tests, 0 failures, on both JDK 17 and JDK 8.
• ./gradlew javadoc on JDK 8: no errors.
• New and updated mock tests and fixtures cover:
  • every new method;
  • the new response fields;
  • the tools/tool_choice request JSON (promptCompletionRequestWireFormatTest);
  • fractional learningRateMultiplier values.

🤖 Generated with Claude Code