feat: add thinking mode to claude4

[Lavriz]

@1712n/dni-nlp-backend do you want to have separate claude4 models with and without thinking? what i added now always adds this thinking mode to the models

[Lavriz]

@claude review this PR

[claude]

Claude finished @Lavriz's task —— View job

---

Code Review Analysis

• Read and analyze the changed file
• Review code quality and implementation
• Provide feedback on architectural decision
• Post comprehensive review

Review Findings

✅ Code Quality & Implementation

The implementation looks clean and follows good practices:

• Clear conditional logic: The thinking mode is properly scoped to only Claude 4 models (src/providers/anthropic.ts:55-62)
• Appropriate configuration: The thinking setup includes reasonable defaults:
  • type: 'enabled'
  • budget_tokens: 32_000 (appropriate for the increased max_tokens: 128_000)
  • temperature: 1 (required constraint when thinking is enabled)

🤔 Architectural Decision: Separate Models vs Always-On

Regarding your question about having separate claude4 models with/without thinking:

Current approach (always-on) is recommended for these reasons:

1. User Experience: Users get the benefit of thinking mode automatically without needing to choose between model variants
2. Simplicity: Avoids model proliferation and reduces configuration complexity
3. Performance: Thinking mode generally improves response quality with minimal downsides
4. Consistency: Matches how other providers handle similar features (e.g., Gemini's thinking models)

Alternative approach (separate models) would only be beneficial if:

• There are significant latency concerns (thinking adds processing time)
• Token budget is very limited (thinking consumes budget tokens)
• You want to offer a "faster" option for simple queries

📝 Minor Suggestions

1. Documentation: Consider adding a comment explaining why temperature is forced to 1:

   query.temperature = 1; // Required by Anthropic API when thinking is enabled

2. Future consideration: If you later want both options, you could add model variants like:

   • claude-sonnet-4-20250514 (with thinking, current)
   • claude-sonnet-4-20250514-fast (without thinking)

✅ Verdict

The current implementation is solid. Recommend merging - the always-on thinking approach provides better user experience without significant drawbacks.

[evgenydmitriev]

@claude enable thinking mode for Claude_4_Opus model by exposing another model called Claude_4_Opus_thinking to the user. Make sure the regular Claude_4_Opus model is still available. Update the documentation.

[claude]

Claude finished @evgenydmitriev's task —— View job

---

Task: Enable thinking mode for Claude_4_Opus with separate model variants

• Read current implementation in src/providers/anthropic.ts
• Check existing documentation and model listings
• Modify provider to expose both Claude_4_Opus and Claude_4_Opus_thinking
• Update documentation to reflect new model
• Test implementation and commit changes

✅ Implementation Complete

Successfully implemented separate Claude 4 Opus model variants as requested:

Changes Made

1. Added new model: claude-opus-4-20250514-thinking to ModelName enum
2. Updated provider mapping: Added to Anthropic models list in providers/index.ts
3. Modified request logic in anthropic.ts:
   • claude-opus-4-20250514-thinking → maps to base model with thinking enabled (128k tokens, temperature=1)
   • claude-opus-4-20250514 → base model without thinking (128k tokens, user temperature)
   • claude-sonnet-4-20250514 → unchanged (still has thinking enabled)
4. Updated README.md: Added new model to available models list with description
5. Added test coverage: Test ensures thinking variant properly maps to base model with thinking config

Result

Users can now choose between:

• claude-opus-4-20250514 - Regular Claude 4 Opus without thinking
• claude-opus-4-20250514-thinking - Claude 4 Opus with thinking mode enabled

Commit: b5c5e6c

---

[evgenydmitriev]

@claude make similar modifications for Claude_4_Sonnet and Claude_4_Sonnet_thinking models

[claude]

Claude finished @evgenydmitriev's task —— View job

---

✅ Task Complete: Claude_4_Sonnet and Claude_4_Sonnet_thinking model variants

Successfully implemented separate Claude 4 Sonnet model variants following the same pattern as Claude_4_Opus models.

Todo List:

• Analyze current implementation pattern
• Read existing code to understand the structure
• Create Claude_4_Sonnet_thinking model enum entry
• Modify anthropic.ts provider logic for both sonnet variants
• Update providers/index.ts with new model
• Update README.md documentation
• Add test coverage for new model variants
• Commit all changes

Implementation Complete:

Changes Made:

1. Added new model: claude-sonnet-4-20250514-thinking to ModelName enum
2. Updated provider mapping: Added to Anthropic models list in providers/index.ts
3. Modified request logic in anthropic.ts:
   • claude-sonnet-4-20250514-thinking → maps to base model with thinking enabled (128k tokens, temperature=1)
   • claude-sonnet-4-20250514 → base model without thinking (128k tokens, user temperature)
4. Updated README.md: Added new model to available models list with description
5. Added test coverage: Tests ensure both sonnet variants work correctly

Result:
Users can now choose between:

• claude-sonnet-4-20250514 - Claude 4 Sonnet without thinking
• claude-sonnet-4-20250514-thinking - Claude 4 Sonnet with thinking mode enabled

Commit: ddba634

---