CrossVendorAudit runs codex at reasoning effort "none" under the isolated CODEX_HOME #2276

Thanks for the isolated CODEX_HOME in CrossVendorAudit.ts. Keeping a LifeOS-for-Codex persona out of the audit fixed a real problem here. It has one side effect I ran into, reported in case it helps.

What happens

buildIsolatedCodexHome() creates a fresh home containing only auth.json. Without a config.toml, nothing sets a reasoning effort, and the exec args at LifeOS/install/LIFEOS/TOOLS/CrossVendorAudit.ts:413 don't pass one. On codex-cli 0.159.3, codex then runs the audit at no reasoning effort. Its own session header says so:

model: gpt-5.6-sol
provider: openai
reasoning effort: none

So the cross-vendor pass that is supposed to catch what the Claude reviewers missed runs as a no-reasoning call.

Repro

H=$(mktemp -d); ln -s ~/.codex/auth.json "$H/auth.json"
echo "Reply with the single word ok." | CODEX_HOME="$H" codex exec \
  --sandbox read-only --skip-git-repo-check --ephemeral --model gpt-5.6-sol - 2>&1 | grep -i "reasoning effort"
# -> reasoning effort: none

With -c 'model_reasoning_effort="high"' added, the same command prints reasoning effort: high.

Suggested fix

Pass the effort explicitly in invokeCodex(), since the isolated home will never carry a config:

["exec", "--sandbox", "read-only", "--skip-git-repo-check", "--ephemeral",
 "-c", 'model_reasoning_effort="high"',
 "--output-schema", schemaPath, "--model", CROSS_VENDOR.forge, "-"]

I'm running that locally, and an end-to-end audit through the tool now reports high.

Optional: also record what codex's header says in the verdict. The model's self-reported model_used came back as "gpt-5" while the header said gpt-5.6-sol, so the header is the reliable source. Locally I parse ^model: and ^reasoning effort: from codex's stdout/stderr into codex_model / codex_effort and log both to cato-findings.jsonl. Then an audit that ran without reasoning shows up in the record instead of passing silently.

Happy to adjust if you'd rather set this in models.ts alongside the model ID.