ai-code-review-catalog

<reviewers>

• /cross-harness-sub-agent with a short broad review prompt written by the main agent
• native sub-agent reading and following /thermo-nuclear-code-quality-review
• native sub-agent using prompts/ponytail-review.md
• native sub-agent using prompts/code-review.md
• harness's native code review agent
• codex cli native review through the shared launcher:

  bin/codex-native-review origin/develop
  # optional evidence and output location:
  bin/codex-native-review origin/develop playground/review-handoffs playground/checks.md
  

  • when using exact-commit verification evidence, run from a clean, committed checkout; output is a fresh directory containing review.md and review.log
  • when the parent harness is sandboxed, launch the helper outside that parent sandbox with an escalated shell command; the reviewer itself remains --sandbox read-only
  • the helper supplies source-inspection instructions through developer_instructions, since codex review --base rejects a positional prompt
  • verification belongs to the caller or CI. supply existing evidence with its exact commit, commands or CI links, scope and pass/fail/pending status; the reviewer reports gaps instead of running write-dependent checks
  • retain the report and log on failure; launcher permission failures need the correct parent launch context, not another identical attempt

</reviewers> <prompts>

• code review prompt: prompts/code-review.md
• ponytail review prompt: prompts/ponytail-review.md
• bug-class review prompt: prompts/bug-class-review.md
• data leak review prompt: /data-leak-review

</prompts>

native sub-agent model routing:

harness	model	reasoning
codex	gpt-6.1-sol	high
claude code	claude-opus-5-5	high
other	what the harness recommends	what the harness recommends

don't use fast mode if a command fails, try similar variants if a cli tool is not installed/authenticated, skip it