Greptile Summary

This PR adds support for Opus 4.7/4.8 and GPT-5.4+ models by introducing two new reasoning effort levels (xhigh, max), a TaskBudgetConfig for agentic loops, adaptive-thinking model tiers (adaptive, strict, mid_conversation_system), and a phase field on OutputMessage for GPT-5.4 conversation chaining.

• New effort levels: XHigh and Max are added to both ReasoningEffort and ResponsesReasoningEffort. Non-supporting providers (Bedrock, Vertex, non-strict Anthropic) clamp them to High; only strict Anthropic models (Opus 4.7/4.8) pass xhigh/max through.
• Model-tier system: Three new configurable lists (adaptive_thinking_models, strict_thinking_models, mid_conversation_system_models) gate per-model behaviors: adaptive thinking with effort-based output config, sampling-param suppression, and inline system messages.
• Task budget: CreateResponsesPayload gains a task_budget field (Opus 4.7/4.8 only). apply_task_budget correctly skips it when thinking is explicitly disabled and synthesizes adaptive thinking when none was specified. The task-budgets-2026-03-13 beta header is included via the updated compute_beta_header.

Confidence Score: 5/5

Safe to merge — no blocking defects found in the changed paths.

All three issues raised in the prior review round have been resolved: top_p is now correctly suppressed whenever thinking is active or the model is strict; the None effort variants use unreachable!() in the effort-translation helpers; and apply_task_budget guards against forwarding a task budget when thinking is explicitly disabled. The new model-tier system is driven by configurable lists rather than hardcoded substrings, and the effort-clamping logic is clearly separated. Test coverage for the new behaviors is thorough.

No files require special attention.

Important Files Changed

_{Reviews (11): Last reviewed commit: "Fix mid-conversation system gating and m..." | Re-trigger Greptile}

@greptile-apps

@greptile-apps

@greptile-apps

@greptile-apps