fix: max_tokens limits

[Lavriz]

Issue

API requests were failing with: max_tokens: 128000 > 32000, which is the maximum allowed number of output tokens for claude-opus-4-20250514, #241 (comment)

see models limits for input/output, we don't have 128k output anymore.

Claude may not use the entire budget allocated, especially at ranges above 32k.

budget_tokens must be set to a value less than max_tokens.

Changes

• Set correct max_tokens limits: 32k for Opus, 64k for Sonnet
• Opus thinking: 20k thinking budget, 12k output
• Sonnet thinking: 32k thinking budget, 32k output

[Lavriz]

@claude review this PR

[claude]

Claude finished @Lavriz's task —— View job

---

PR Review Complete ✅

This PR correctly fixes the token limits issue for Claude models. Here's my detailed review:

Review Summary

LGTM - This change properly addresses the API failures by setting correct token limits according to Anthropic's current model specifications.

Code Analysis

Fixes Applied (src/providers/anthropic.ts:65-83):

• ✅ Opus limits: Set to 32k max_tokens (down from 128k)
• ✅ Sonnet limits: Set to 64k max_tokens
• ✅ Thinking budgets: Properly allocated within max_tokens constraints
  • Opus thinking: 20k budget + 12k output = 32k total
  • Sonnet thinking: 32k budget + 32k output = 64k total

Technical Review

Strengths:

1. Correct token allocation: The math works correctly - thinking budget + output never exceeds max_tokens
2. Model mapping preserved: The logic correctly maps thinking variants to base models (src/providers/anthropic.ts:44-48)
3. Temperature override: Properly sets temperature: 1 for thinking models (src/providers/anthropic.ts:78)
4. Clean conditional logic: Well-structured if/else blocks handle all model variants

Code Quality:

• Clean, readable implementation
• Consistent formatting and naming
• No breaking changes to the API interface

Verification Against Anthropic Docs

The token limits align with Anthropic's model specifications:

• Claude Opus: 32k output limit ✅
• Claude Sonnet: 64k output limit ✅
• Thinking budgets must be < max_tokens ✅

Minor Observations

• The default max_tokens: 8_192 at line 52 gets overridden for all Claude 4 models, which is correct behavior
• No concerns with the anthropic-beta header handling for thinking models

Recommendation

Approve and merge - This fix resolves the immediate API failures and implements correct token limits according to current Anthropic model specifications.

Todo List

• Read repository setup instructions (CLAUDE.md) - No CLAUDE.md found
• Analyze the changed file src/providers/anthropic.ts
• Review token limits against Anthropic documentation
• Check thinking budget implementation
• Provide comprehensive review feedback

---

[Lavriz]

@claude explicitly configure max_tokens for each supported Claude 4 model without using a default value (8_192) since we don't support any other claude models.

[claude]

Claude finished @Lavriz's task —— View job

---

Todo List

• Read current anthropic.ts implementation
• Remove default max_tokens value (8_192)
• Set explicit max_tokens for each Claude 4 model
• Verify all model variants are covered
• Commit changes to PR branch

Changes Made ✅

• Removed the default max_tokens: 8_192 from the query object
• Created explicit max_tokens configuration for each Claude 4 model variant:
  • Claude_4_Opus_thinking: 32,000 tokens
  • Claude_4_Sonnet_thinking: 64,000 tokens
  • Claude_4_Opus: 32,000 tokens
  • Claude_4_Sonnet: 64,000 tokens
• Added error handling for unsupported models with clear error message
• Simplified the thinking mode configuration logic

The code now explicitly configures max_tokens for each supported model without relying on any default value. Each Claude 4 model variant has its own explicit configuration as requested.

Commit: f7b3bbf

---

| Branch: fix-claude-tokens

[Lavriz]

@Kseymur fixed, thanks!!