Skip to content

broomva/cross-review

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

cross-review

License: MIT bstack Inspired by

bstack P20 — Cross-Model Adversarial Review Gate. The model that wrote the code cannot be the final judge of the code. Substantive PRs fire an adversarial review (different evaluator, anti-slop ≥7/10) before merge.

Quick start

npx skills add broomva/cross-review

Before pushing a substantive PR (>200 LOC OR public API OR multi-file OR governance-class):

cross-review pre-push --diff-base origin/main

The script auto-detects Codex CLI. If installed → fires Strata A (true cross-vendor). Otherwise → Strata B (fresh subagent). Strata C (composed adversarial-review skills) always runs in parallel.

Why this exists

When the same AI model plans, implements, and reviews, it will not challenge its own assumptions. A different model — trained differently, with different biases — catches what the first one misses. This pattern was systematized by Dallionking/cross-model-agents (May 2026) — 31-agent bidirectional Claude↔Codex review system with anti-slop scoring.

cross-review brings the discipline (cross-model gate is mandatory for substantive PRs) into bstack as primitive P20, while keeping our composition pattern (3 strata so the primitive works whether or not Codex is installed).

The 3 strata

Strata Mechanism Strength
A codex exec -m gpt-5.4 reads diff and scores Strongest — genuinely different model
B Fresh Agent subagent under devil's-advocate brief Strong — fresh context breaks within-session echo
C Composed existing adversarial-review skills (superpowers:constructive-dissent, devils-advocate, pr-review-toolkit:*, critique, premortem) Always — toolkit P20 makes mandatory

The rubric

Anti-slop scoring across 5 dimensions (2 pts each):

  1. No over-engineered abstractions
  2. No template-paste patterns
  3. Correct contracts at boundaries
  4. Failure modes named explicitly
  5. Tests cover the change

PASS at ≥7/10. LOOP if <7 (max 3 fix rounds). ESCALATE on round 3 failure.

Full rubric: references/rubric.md.

Composition with bstack

Primitive Role
P4 PR Pipeline P20 fires before P4 auto-merge
P7 CI Watcher After P20 passes, P7 watches CI
P11 Empirical Feedback Different dim: P11 "does it run", P20 "is it well-built"
P17 Lens-Routed Articulation Lenses become evaluator stances
P18 Format-Follows-Audience Verdict logged as PR comment (markdown for both audiences)
P19 Mechanism Selection P20-gated PRs naturally run as a /goal arc

Reading order

  1. SKILL.md — full skill contract (3 strata, rubric, reflexive trigger, anti-rationalizations)
  2. references/rubric.md — the anti-slop scoring rubric + adversarial brief templates
  3. scripts/cross-review.sh — the entry point (auto-detects strata, structures the gate)
  4. tests/ — verification battery (pressure scenarios + integration tests)

License

MIT — see LICENSE.

About

bstack P20 — Cross-Model Adversarial Review Gate. The writer cannot judge the work. 3 strata: Codex cross-vendor / fresh subagent / composed adversarial-review skills. Anti-slop ≥7/10.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages