mrvladd-d/mb-review

mb-review — Multi-expert Memory Bank review

• What it does: runs independent reviewers over architecture, scope, task planning, security, and Memory Bank discipline.
• Use it when: you need a clean review gate before execution, after bootstrap, or before trusting the docs.
• Input: an existing .memory-bank/.
• Output: reviewer reports in .tasks/TASK-MB-REVIEW/ plus a synthesized list of fixes and review verdicts.

Goal

Detect gaps, contradictions, broken traceability, and non-compliance early.

This is not the same as per-task semantic verification. Use mb-red-verify when the question is “did this task solve the right problem in substance?”

Preconditions

• .memory-bank/ exists.

Process

1) Create review task folder

Create:

• .tasks/TASK-MB-REVIEW/

2) Spawn reviewers (fresh contexts)

Spawn these subagents in parallel (max 5–7):

1. Architect — ./agents/shared-review-architect.md
2. Scope/RTM — ./agents/shared-review-scope.md
3. Task planning — ./agents/shared-review-plan.md
4. Security — ./agents/shared-review-security.md
5. MBB compliance — ./agents/shared-mb-reviewer.md
6. Code quality (conditional: if repo has code) — ./agents/shared-review-code.md

Each reviewer must:

• write a detailed report to .tasks/TASK-MB-REVIEW/
• return only a short summary + verdict to the orchestrator

3) Synthesize and decide

As orchestrator:

• combine findings
• deduplicate
• rank issues P0–P3
• produce a concrete fix plan

If the repo is preparing for /autonomous:

• treat unresolved P0/P1 issues as blocking
• do not allow batch execution until the final verdict is APPROVE

4) Gate

If any reviewer returns REJECT:

• fix MB
• re-run mb-review

Definition of done

• .tasks/TASK-MB-REVIEW/ contains the reviewer reports.
• Orchestrator produced an actionable prioritized fix list.
• Final verdict: APPROVE.