[Roger V (rgrvlsk)]

paid_model names one model for every agent. Left empty, Codex runs the plugin's calls (the loop guard's judge and redirect, the micro-skills proposer) on the person's default model and effort. Set to a Codex model, it breaks Claude Code's calls, and the other way round. Someone who works with both agents cannot pick a small model for each.

What changed

• <agent>_paid_model, one per agent with side calls (claude, codex, copilot, pi, opencode, cursor, grok, hermes), wins over paid_model for that agent. The names follow pi_cli and hermes_cli. _host_call reads the agent's key first, so every host's call gets it without further changes. TE_<AGENT>_PAID_MODEL sets it for one session.
• claude_paid_effort reaches claude --effort, but only when this Claude Code's --help lists the option, as claude_flags already does for --no-session-persistence. codex_paid_effort reaches codex exec -c model_reasoning_effort="…". Each accepts its own CLI's words: Claude Code 2.1.319 low … max, Codex 0.167.0 none … max. Whether a model takes a value is still the CLI's call. The other agents' calls take only --model today, so they get no effort key.
• With the new keys empty, every host's argv is unchanged.
• The settings table in plugin/README.md and the config.py docstring list the new keys.

This touches sidecall.py and config.py, as #185 (local_model) and #163 (--safe-mode in claude_call) do. The overlap is a few lines, and either can go first.

Tests

tests/test_live_core.py:

• test_side_calls_take_a_model_per_agent_and_an_effort: the per-agent model wins over paid_model. Each effort reaches its CLI in that CLI's form. A Claude Code whose help lacks --effort gets no --effort, and a value from the other CLI's vocabulary is refused. With nothing set, the Codex argv is upstream's.
• test_every_agent_with_side_calls_has_its_own_paid_model_setting: every host in hosts.HOSTS except NO_SIDE_CALLS has its key, so a new agent cannot be added without one.

Both fail on main. python -m pytest tests gives 852 passed and 16 skipped on Python 3.10 and 3.13 (macOS), and experiments/tests gives 394 passed. Tried on macOS only: claude -p --model haiku --effort high and codex exec -m <model> -c model_reasoning_effort="high" each answered.

[Roger V (rgrvlsk)]

@microsoft-github-policy-service agree