Complete Codex cost and Claude delegation parity

[hamzamerzic]

Summary

• preserve Codex's cache-write token counter despite generated Python schema lag
• price GPT-5.6 cache writes at 1.25x and apply the >272K long-context multipliers per upstream model call, avoiding aggregate overcharging
• pass connected Claude credentials into Codex turns and seed a bounded non-interactive Claude delegation skill
• keep both delegation directions explicit, scoped, and independently reviewed by the parent agent

Pricing behavior follows OpenAI's current GPT-5.6 model guidance: https://developers.openai.com/api/docs/guides/latest-model

Tests

• MOBIUS_TEST_RUNTIME=1 pytest -q backend/tests/test_usage_metrics.py backend/tests/test_codex_sdk_contract.py backend/tests/test_codex_sdk_runner.py backend/tests/test_provider_agent_browser_env.py backend/tests/test_skills_platform.py backend/tests/test_chat_skills_context.py -q

[hamzamerzic]

Merged to main as part of #304. Codex cost reporting and Claude delegation parity are now included in the integrated default-branch change.