Summary

• Route GPT-5 and o-series OpenAI Chat Completions requests through max_completion_tokens instead of max_tokens.
• Keep existing max_tokens behavior for older OpenAI chat models such as gpt-4o-mini.
• Add focused unit coverage for GPT-5, GPT-5 chat alias, o-series, and legacy chat token limit parameters.

Root Cause

om doctor --validate-key failed when OM_LLM_MODEL=gpt-5.4 because the OpenAI adapter always sent max_tokens. GPT-5-class models reject that parameter and require max_completion_tokens.

Validation

• uv run pytest tests/test_llm.py
• uv run pytest
• uv run ruff check src tests
• uv run om doctor --validate-key confirmed the OpenAI call succeeds with gpt-5.4; unrelated QMD/launchd environment checks were later restored to green in the installed OM environment with om install --both --scheduler launchd --provider openai --llm-model gpt-5.4 --non-interactive.