You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Date: 2026-02-13
Reviewer: furiosa (parallel code-reviewer agents, Opus 4.6)
Methodology: Each fork PR reviewed independently in parallel. Three questions per PR:
PR 1226 MT-3 Guardrails Plan — Refinery formula safety for auto_land
MT-3 Fix: Integration Branch Landing Guardrails
Context
During Julian's manual testing of PR #1226, the refinery AI (Sonnet 4.5) bypassed the auto_land=false setting and landed an integration branch to main using raw git merge + git push. This is MT-3 — a formula guardrail gap, not a code bug. The AI agent ignored the formula's "nothing to do" instruction and acted autonomously.
This plan implements a three-layer defense-in-depth approach:
Layer 1 (Formula + Role): Explicit FORBIDDEN directives in both the formula and role template
Layer 2 (Git Hook): Pre-push hook that blocks integration branch content landing on the default branch
Layer 3 (Code): PushWithEnv() method that only the legitimate gt mq integration land code path can use to bypass the hook
Manual validation tests for PR #1226: make integration branches end-to-end across the pipeline.
Each phase is designed to be run sequentially by a crew member in the gastown rig.
Review note: This is a revised version of the original test plan from crew/max/.
Changes from the original are marked with [REVISED] or [NEW] tags.
A comprehensive guide to understanding and using Steve Yegge's multi-agent orchestrator for Claude Code
Generated based on the current gt repo (as of 20/1/26 and the various articles found online as of the same date)