Independent Multi-Agent Review
- Select the execution profile, shared DelegationBudget, Reviewer effort, and model profile independently. Reviewer state does not charge the total budget twice.
- Start every assignment at Luna Low and one point. Add budget-mode and valid-evidence points, then dispatch once. Registered Reviewers have ten combinations up to Astra High; workers/explorers retain four. No xhigh/max/ultra, direct premium default, or preliminary calls merely to climb tiers. Read reviewer-model-routing.md for the frozen scoring rules.
- Bind every Review Packet to Project ID, Task ID, Git baseline, Task Envelope, packet hash, validation summary, assigned scope, and a matching Reviewer permit in
delegation-budget.py when unified accounting is active. For scoped review, also use scripts/scoped_review.py to fingerprint targets, statically discovered dependencies, configuration, and authority files. Unknown dynamic dependencies are INCOMPLETE, relevant changes are STALE, and scoped PASS never means release-wide PASS.
- Read summary and statistics first, then assigned diff and direct dependencies, then expand only when evidence remains insufficient.
- Collect one round before deduplication, root-cause clustering, conflict resolution, and centralized repair.
- After repair, refresh only affected evidence and packet content. Stop repeated review when the packet is unchanged or no new information exists.
- Defaults: parallel <=3, cumulative <=6, post-implementation rounds <=2, repair rounds <=2, Terra High reviewers <=1; Sol/Astra attempts <=2, premium parallelism <=1, Astra High <=1. Apply the stricter root-budget and controller limits.
Independent context is not system read-only. A writable parent without sandbox-denial evidence supports only logical-readonly. Review evidence never authorizes modification, commit, push, deployment, restart, or production operation.
Controlled-Evolution Boundary
This Skill produces independent-review evidence and attribution input; it does not maintain the cross-task evolution contract. Route long-window Reviewer yield, model-cost, or routing-deviation governance to controlled-evolution-governance. A single review must not load evolution rules automatically.
1---2name: multi-agent-independent-review-23description: Use for risk-based design gates and independent functional, compatibility, security, performance, contract, concurrency, and delivery reviews.4---56# Independent Multi-Agent Review781. Select the execution profile, shared DelegationBudget, Reviewer effort, and model profile independently. Reviewer state does not charge the total budget twice.92. Start every assignment at Luna Low and one point. Add budget-mode and valid-evidence points, then dispatch once. Registered Reviewers have ten combinations up to Astra High; workers/explorers retain four. No xhigh/max/ultra, direct premium default, or preliminary calls merely to climb tiers. Read reviewer-model-routing.md for the frozen scoring rules.103. Bind every Review Packet to Project ID, Task ID, Git baseline, Task Envelope, packet hash, validation summary, assigned scope, and a matching Reviewer permit in `delegation-budget.py` when unified accounting is active. For scoped review, also use `scripts/scoped_review.py` to fingerprint targets, statically discovered dependencies, configuration, and authority files. Unknown dynamic dependencies are `INCOMPLETE`, relevant changes are `STALE`, and scoped PASS never means release-wide PASS.114. Read summary and statistics first, then assigned diff and direct dependencies, then expand only when evidence remains insufficient.125. Collect one round before deduplication, root-cause clustering, conflict resolution, and centralized repair.136. After repair, refresh only affected evidence and packet content. Stop repeated review when the packet is unchanged or no new information exists.147. Defaults: parallel <=3, cumulative <=6, post-implementation rounds <=2, repair rounds <=2, Terra High reviewers <=1; Sol/Astra attempts <=2, premium parallelism <=1, Astra High <=1. Apply the stricter root-budget and controller limits.1516Independent context is not system read-only. A writable parent without sandbox-denial evidence supports only `logical-readonly`. Review evidence never authorizes modification, commit, push, deployment, restart, or production operation.1718## Controlled-Evolution Boundary1920This Skill produces independent-review evidence and attribution input; it does not maintain the cross-task evolution contract. Route long-window Reviewer yield, model-cost, or routing-deviation governance to `controlled-evolution-governance`. A single review must not load evolution rules automatically.
JimmyVGDY