Greptile Summary

This PR fixes a 400 error that occurs when Anthropic's budget_tokens (set via reasoning.effort) exceeds max_tokens — a common mismatch since both are derived independently (e.g., medium effort yields a 16 000-token budget against the 4 096 default). A new max_tokens_with_thinking_headroom helper ensures max_tokens is raised to at least budget_tokens + THINKING_OUTPUT_MARGIN (8 000) whenever fixed-budget thinking is active.

• Introduces max_tokens_with_thinking_headroom that only ever raises max_tokens — generous caller-supplied values, adaptive thinking, and disabled thinking are left untouched — and emits a tracing::debug! log when a raise occurs.
• Applied consistently to both the chat-completions path (create_chat_completion) and the Responses API path (create_responses); the legacy create_completion stub is unaffected as it always returns an error.
• Four unit tests cover the key boundary cases: below budget, above budget+margin (preserved), between budget and budget+margin (still raised), and non-fixed thinking modes (unchanged).

Confidence Score: 5/5

Safe to merge — the change only raises max_tokens when needed and cannot reduce it, so existing behaviour for callers who already set a high value or use adaptive/disabled thinking is unchanged.

The helper is purely additive: it raises max_tokens or leaves it alone, never lowers it. Both active call sites are updated, the legacy stub is correctly unaffected, overflow is guarded with saturating_add, and the four unit tests cover all meaningful boundary conditions including the edge case between budget and budget+margin.

No files require special attention.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Incoming request with reasoning.effort] --> B{Resolve max_tokens from payload / config / default}
    B --> C[convert_*_reasoning_config produces thinking + output_config]
    C --> D{thinking variant?}
    D -- "Enabled { budget_tokens }" --> E["desired = budget_tokens + THINKING_OUTPUT_MARGIN (8 000)"]
    E --> F{"max_tokens >= desired?"}
    F -- "Yes (caller set generous value)" --> G[Keep max_tokens unchanged]
    F -- "No (too low)" --> H[Raise max_tokens to desired, emit tracing::debug!]
    D -- "Adaptive / Disabled / None" --> G
    G --> I[Build AnthropicRequest with final max_tokens]
    H --> I
    I --> J[POST /v1/messages]

_{Reviews (3): Last reviewed commit: "Review fixes" | Re-trigger Greptile}

@greptile-apps