💡 Stale-Base Semantic Revert Detection Guard

[github-actions[bot]]

Summary

Add a pre-merge CI check that detects when a PR's merge would silently revert changes from a recently merged commit. This guards against the class of bug documented in #655, where a stale-based merge silently rolled back a critical initiative-planner fix, causing idea:approved planning to stop working entirely.

Market Signal

GitHub's merge queue and Stacked PRs features (April 2026) address merge ordering but not semantic reversions. Mergify and similar tools focus on CI-passing guarantees, not content-level revert detection. This is an unserved gap in the CI/CD toolchain that is particularly dangerous for autonomous agent pipelines where no human inspects every merge diff.

User Signal

• Bug bug: stale-base merge (#606) silently reverted the initiative-planner discussion→dispatch fix (#618); idea:approved stopped planning #655: stale-base merge silently reverted the initiative-planner discussion→dispatch fix — idea:approved stopped planning
• Bug Self-host deadlock: force_review bypasses the advisory gate but not the CI gate, so CI-gate fixes can't merge through ring-0 #619: Self-host deadlock reveals related merge-order risks in the CI gate
• The auto-rebase infrastructure (Cut auto-rebase fan-out thrash: restrict to review-ready PRs + instrument the agentic-conflict-resolution rate #736, [Phase 1] Instrument auto-rebase: agentic-conflict-resolution rate + fan-out CI-run volume report #737) mitigates stale bases but cannot catch semantic reversions where both old and new code are syntactically valid

Technical Opportunity

The auto-rebase health instrumentation (scripts/auto_rebase_health.sh) and the downstream-impact mapper (scripts/lib/downstream-impact.sh) already provide scaffolding. A revert-detection check could diff the merge result against the union of individually merged diffs, flagging any hunks that disappear. This is a deterministic, no-LLM script — fast and cheap.

Assessment

Dimension	Score	Rationale
Feasibility	high	Pure bash/git diff script; no model calls or external dependencies
Impact	high	Prevents a documented class of production incident that broke planning silently
Urgency	high	The attack surface grows with agent-generated PR volume; Bug #655 was recent

Adversarial Review

Strongest objection: False positives — legitimate refactors that replace recently merged code would trigger the guard, creating noisy CI failures that slow down the pipeline.

Rebuttal: The guard can be tuned to only flag exact reversions (identical hunk removal within N commits), not refactors. It can be advisory (warning comment) rather than blocking, following the existing advisory review gate pattern. The cost of a false positive (one extra review click) is far lower than the cost of a silent reversion (#655 broke planning for an unknown duration).

Suggested Next Step

Prototype a bash script that compares git diff main...HEAD against recently merged PRs' diffs to detect hunk reversions, and add it as a non-blocking CI check.