Choosing the Security run's Harness, with Codex adjustments #546

Parent

#427

What to build

The Harness. A Security run's Harness is the first of:

1. the one its command names;
2. harness under [security] in the User config;
3. the User config's default Harness;
4. Claude Code.

Its Model and Effort come from that Harness's own section.

On Codex, security sessions raise agents.max_concurrent_threads_per_session to 8, and their prompts ask for fresh sub-agents (fork_turns: "none"), so that the skill's verifiers stay independent. #427's trial found Codex running at most three sub-agents at once, and its sub-agents sharing their parent's conversation.

Acceptance criteria

• harness claude on the command wins over the setting, the setting wins over the default Harness, and the Model and Effort follow the chosen Harness.
• Codex security sessions are started with the raised thread cap, and their prompts carry the fresh sub-agent line. Claude Code's don't.
• A blank or missing harness setting changes nothing.